Import coverage by company and product
Reviewed 2026-09-29. This inventory accounts for 100 registered source selectors, 152 source formats, 18 browser providers, and 229 product or surface entries across 76 company or project groups.
Implemented means that an import route exists in the code. It does not mean every client version, field, account plan, or operating system has passed a native trial. The verification column records this catalog’s evidence level; older validation elsewhere is not promoted without a reviewed evidence record.
Product boundaries
Section titled “Product boundaries”ChatGPT and Codex cloud are different products. Claude.ai chat, Claude Code cloud, and Cowork cloud have distinct import sources. Anthropic is rolling out a unified chat/task experience; that branding change does not establish shared export coverage or execution identity. Gemini chat, Gemini Notebook (formerly NotebookLM), and Gemini Spark need separate coverage. Grok chat, Grok Build, and Grok Bot are different products. Muse Code support does not imply support for Meta’s consumer assistant. Microsoft Copilot and GitHub Copilot have separate histories. GitHub-hosted Claude and Codex agents are distinct from their vendor-owned cloud products. Copilot SDK Remote Control shares local execution; Copilot SDK cloud sessions execute on GitHub. Qwen Chat is separate from Qwen Code; its desktop, daemon, editor, SDK, channel and batch surfaces require independent acquisition evidence.
Cowork cloud has account-wide tagged-task enumeration through claude-cowork-cloud and retained desktop cache capture through claude-cowork. The shared Claude browser route captures explicitly linked workspaces and bounded current files. Shared extension activation is confirmed. Native linked-workspace and current-file acquisition remain unverified. Remote Control describes transport to a local Claude Code runtime and is not evidence of a hosted Claude Code run.
Choosing an import route
Section titled “Choosing an import route”Local-store routes read retained files and can be watched. Cloud routes are explicit network imports. Manual exports remain useful alongside browser or cloud capture. A desktop cache only covers data the application retained. A product opened in a desktop wrapper does not become a different product, and a model provider does not identify the harness that ran it.
Use the linked guide for exact commands, authentication, paths, and limitations. The route names on unsupported rows describe investigation targets, not working commands. Missing metadata, unknown models, and unmeasured usage must remain explicit; a source’s replay support is a separate property from its importer.
Implemented and partial routes
Section titled “Implemented and partial routes”Agent Client Protocol
Section titled “Agent Client Protocol”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| ACP observed conversations (partial) | acp-runtime |
hansard-runtime | Captures only turns Hansard observes through a configured ACP provider instance. Preserves prompts, replies, thinking, tools, metadata and reported turn usage, with namespaced instance and native session identity and raw replay. No provider-native history import, original attachment bytes, unreported usage or response model attribution. Vendor-specific imports remain independent. Installed desktop and external native-agent acceptance remain unverified. | fixture-tested; 3 open gates. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Z.ai chat (partial) | zai-chat |
browser | Experimental browser history capture enumerates normal, pinned and archived chats. It reads every referenced message body in resumable batches, selects the active branch, retains alternate branches, and normalizes text, reasoning, recorded models and tool calls. Preserved captures support offline replay. Native Arc access stops at Sign in. Original file bytes, workspaces, subagents, scheduled tasks and account exports remain incomplete. Developer API support does not establish consumer history access. Source and fixture evidence do not establish installed or automatic native acceptance. | fixture-tested; 8 open gates. 2026-10-08 native trial (blocked): Arc opened the consumer chat at chat.z.ai and displayed Sign in. Public client prod-fe-1.1.98 exposes history, message-batch and JSON export routes. No authenticated conversation or export was acquired. This is an access gate, not evidence of provider-side impossibility. |
Agentic AI Foundation
Section titled “Agentic AI Foundation”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Goose | goose |
local-store | Goose shared SQLite sessions and legacy JSONL transcripts, including tools, reasoning, usage, and child sessions. Shared local CLI, desktop, and API storage does not reliably identify the producing interface. Historical inference and ledger models remain separate from current configuration. Missing counters remain unknown; explicit zero stays in metadata. Direct raw parser replay passed, but full SQLite raw rebuild remains unsupported. | fixture-tested; 8 open gates. 2026-09-29 native trial (partial): Official CLI 1.52.0 with a local synthetic endpoint: initial/resumed conversation retained sixteen messages, two human prompts, seven ledger rows, five tool calls, controlled edit/shell errors and a byte-verified image. Initial transcript prefix, repeat import, source bytes and direct raw parser replay passed. All 127 observed field paths audited. Full raw rebuild, hosted usage and viewer acceptance remain open. 2026-10-09 native trial (blocked): No local data at ~/.local/share/goose/sessions or ~/Library/Application Support/Block/goose/sessions. Upstream v1.54.0 remains schema 16; synthetic coverage adds document bytes, redacted thinking and notifications. Installing and creating a conversation, or supplying a store, would permit a native trial. Full raw rebuild and existing native gates remain open. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Aider | aider |
local-store | Aider repo-local .aider.chat.history.md transcripts with optional LLM/input history raw backup. Chat roles and Markdown are captured. Optional LLM history adds matching response clocks, not model, token, cost, or request metadata. Input history stays raw-only. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no .aider.chat.history.md in the bounded project scan or home fallback, and no configured alternative. Audited v0.86.2. Installation alone creates no history; retained history or separately authorized provider activity is required. |
Alibaba
Section titled “Alibaba”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Qwen Chat (partial) | qwen-chat |
browser, export | Experimental browser capture enumerates normal, pinned, archived and project histories. It follows message cursors, validates a stable newest page, normalizes the selected branch, text, reasoning and recorded tools, and supports offline replay. The single-chat JSON export imports by native account and conversation identity with retained raw snapshots and offline replay. Native Arc access requires sign-in. Hidden branches, temporary chats, original files, project instructions and generated revisions remain uncollected or unverified. Only the observed single-chat JSON export is accepted. Bulk and archived exports remain unverified. The provider removes some citation data. Browser history takes precedence. Only chat.qwen.ai is implemented. Qwen Code is a separate product. | fixture-tested; 8 open gates. 2026-10-08 native trial (blocked): Arc opened chat.qwen.ai and displayed Log in. Public client 0.3.12 establishes the single-chat JSON serializer, but no authenticated history or native export was acquired. The implemented export route has synthetic acceptance only. Bulk and archived schemas remain unverified. This records an access gate, not provider-side impossibility. |
| Qwen Code | qwen-code |
local-store | Qwen Code historical chats, checkpoints, archives, and subagents from the runtime tmp/projects stores. Active JSONL branches, tools, native failures, actual patches, embedded images and token splits are preserved. Assistant-record requested models remain distinct from response-model ledger totals. External attachment references stay unavailable. Offline raw rebuild requires the parser 0.5.6.1 ledger binding. Other producing interfaces remain unverified. Current 0.25.0 state and managed records remain typed evidence; managed transaction validation and resource reconstruction are unimplemented. | fixture-tested; 9 open gates. 2026-09-29 native trial (partial): Official CLI 0.24.7 with a local synthetic endpoint: initial/two resumed turns retained twenty messages, three human prompts, seven tool calls, controlled edit/shell errors, a skipped-image notice and a byte-verified image. Prefix, repeat, source-byte and full offline raw rebuild checks passed. All 164 observed paths across 45 chat/ledger rows audited. Other interfaces and full viewer acceptance remain open. 2026-10-09 native trial (blocked): No local data under ~/.qwen and no configured runtime override. Upstream CLI 0.25.0 adds fifteen record subtypes. Parser 0.6.2.1 preserves current state and managed records with synthetic import, repeat and raw replay coverage. Installing and creating a conversation, or supplying a store, would permit a native trial. Managed transaction validation and resource reconstruction remain limits. |
Amazon
Section titled “Amazon”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Kiro local history | kiro |
local-store | Unified sessions, earlier CLI and IDE stores, Crew transcripts, tools, and recorded usage. Unified, V2, Crew, and export readers retain credit units separately from tokens and preserve steering or compaction evidence where recorded. Token and credit counters depend on the engine. Credits sum only within the same unit and do not establish model-specific charges. Steering document counts are not token counts. Cloud sessions require an explicit export. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): Read-only discovery found settings residue and no retained local conversations. Native field, repeat/resume, replay, automatic capture, and consumer trials are blocked by absent history. |
| Kiro session exports | kiro-export |
manual-export | Native Kiro ZIP, JSON, or unified-directory exports preserve messages, tools, reasoning, attachments, and recorded accounting. Cloud enumeration is unavailable. Export ZIPs can omit snapshots and tool-output files; missing bytes remain explicit. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No native Kiro export was supplied. Export import remains fixture-tested; current native field fidelity and round-trip acceptance are unverified. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Amp Cloud | amp-cloud |
cloud | The signed-in Amp CLI lists and exports hosted threads, including orbs, with the same identity as full JSON exports. Exported threads can omit internal subagents and compaction details. Live hosted acceptance is still pending. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No verified native Amp executable or authenticated cloud export is available. The discovered terminal integration shim was not invoked as a provider CLI. |
| Amp exports | amp-export |
manual-export | Full Amp JSON exports from the CLI or web app overlap safely with local and cloud thread imports. Public exports can omit internal subagents and compaction details. A displayed thread is not evidence of a complete export. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No native Amp export was supplied. Export normalization and repeat import remain fixture-tested without current native fidelity evidence. |
| Amp local history | amp |
local-store | Earlier CLI and extension JSON threads, tool calls, usage, and file-change backups. Current server-held threads use the explicit amp-cloud import. Historical local JSON threads. Current CLI, app, and orb threads require explicit cloud export; missing message clocks and ledger cache counts remain explicit gaps. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No retained local Amp thread candidates were found. The available command is a terminal integration shim, not a verified native CLI. Native round-trip and new-turn capture remain blocked. |
Anthropic
Section titled “Anthropic”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Claude Code CLI | claude |
local-store | Interactive Claude Code JSONL transcripts from ~/.claude/projects, including desktop-launched Claude Code sessions when stored there. Hansard-launched turns retain bounded context summaries in the device-local live evidence archive, with native category labels and named token counts. Bounded native CLI and Desktop-linked trials retain forks, child identity, and workflow files. Installed acceptance passes a 22-session linked sample. Large transactional raw rewrites, complete normalized field fidelity, rendered viewer, and full native supervisor remain unverified. | fixture-tested; 4 open gates. 2026-10-09 native trial (partial): Native CLI versions 2.1.275 through 2.1.295: 57 selected parents and 67 sampled children pass isolated import, byte-identical repeat, raw hashes, search, viewer payload, Stats, and source removal. This includes nineteen Desktop-linked parents. Full offline raw rewrite passes for 68 sessions; 56 larger sessions pass dry reparse and normalized message counts only. Copied growth preserves identity and repeat bytes. Raw-only field accounting covers every selected file. Complete normalized fidelity, rendered viewer, native supervisor, and installed acceptance remain open. 2026-10-09 native trial (partial): Installed b69492f2 passes the nineteen-parent Desktop-linked CLI sample and its three children using bundled Node 24.20.0. All 22 sessions pass import, identical repeat, raw hashes, full offline raw rewrite with transcript and metadata parity, search, viewer payload, Stats, source removal, copied growth, and repeat after growth. All 1384 installed manifest files match. This bounded installed result retains the large-rewrite, normalized-value, rendered-viewer, and settled-supervisor limits. |
| Claude Code cloud | claude-cloud |
cloud, browser | Read-only hosted Claude Code events preserve messages, tools, usage, and Git context. The saved-profile and browser routes share the archive writer, exclude recorded non-Code product tags and known chat-linked workspaces, and rebuild from retained full and incremental responses. Native hosted-session acquisition remains unverified. Synthetic Chromium extension-to-archive capture and offline replay pass. An untagged workspace read before its chat is archived can still appear as Code. Current remote files and oversized sessions require separate acceptance. Remote Control bridge evidence establishes local execution only. | fixture-tested; 6 open gates. 2026-10-08 native trial (blocked): The initial valid saved profile returned zero sessions, another had expired credentials, and two lacked credentials. Installed browser capture later archived six Remote Control bridge sessions, which do not establish hosted execution. The signed-in native Code page requires connecting GitHub before repository selection and offers no repository. Clicking Send once with the approved synthetic prompt reopened that picker and left the prompt unsent. No session was created or GitHub access granted. Profile sign-in and repository selection remain pending. 2026-10-08 native trial (partial): Seventeen native detail and event-page responses from six Remote Control bridge sessions passed 328-path field accounting. Installed 8a143c3b retained all 9780931 response bytes exactly. A representative 24-message isolated import passed 209 direct source-value comparisons. This cohort validates the shared parser for bridge execution only. Hosted non-bridge execution and browser transport fields remain unverified. |
| Claude Code Remote Control (local runtime) | claude, claude-cloud |
local-store, cloud, browser | A local Claude Code runtime controlled remotely. Local transcripts and retained bridge events share identity only when the recorded bridge ID establishes the link. Installed browser acquisition archived six bridge sessions. Isolated repeat, raw replay, integrity, search, viewer payload, and Stats checks passed for those native captures. Remote access does not establish hosted execution. Local JSONL and retained cloud events can have different coverage. Hosted Claude Code and Cowork remain separate surfaces. | fixture-tested; 3 open gates. 2026-09-29 native trial (partial): Six bridge sessions passed repeated import, full-event refresh, raw-hash checks, and native viewer review. This provides no hosted-execution acceptance. 2026-10-08 native trial (partial): Installed build 731cc154 acquired six native bridge sessions through the browser extension. Their 3709 events produced 2910 normalized messages. Repeat imports, archive-only replay, raw integrity, search, viewer payload counts, and Stats passed in an isolated archive using the installed runtime. All six production receiver jobs reached archived. Native resumed attachments and cross-route identity remain unverified. These sessions provide no hosted-execution acceptance. 2026-10-08 native trial (partial): Installed build 10cf96ee matches twelve inspected runtime files to committed source. Its bundled Node runtime imported a representative 24-message native bridge capture into an isolated archive. Repeat import, raw replay, raw integrity, search, viewer payload, and Stats passed. This repeat supplements the earlier six-session production trial and does not establish hosted execution. 2026-10-08 native trial (partial): Installed build 079919d2 matches twelve inspected runtime files to committed source. Its bundled Node runtime imported the representative 24-message native bridge capture into a fresh isolated archive. Repeat, raw replay, raw integrity, search, viewer payload, and Stats passed. The bridge remained classified as local Remote Control. Hosted execution and rendered acceptance remain separate gates. 2026-10-08 native trial (partial): Installed 079919d2 rendered the local bridge classification and all 24 messages without a transcript-load error in Chromium and WebKit. Session details still used the hosted source label. Source repair f7495c47 names bridge history Claude Code Remote Control and preserves the hosted label. Installed rendering of that repair remains pending. 2026-10-08 native trial (partial): Six native bridge sessions passed bounded accounting for 328 field paths across seventeen detail and event-page responses containing 3709 events. Fifty-nine declarations describe normalized fields, and 269 describe raw preservation. No unreviewed fields or drift occurred. Read-only installed 8a143c3b inspection matched every response byte in the archive. A fresh isolated 24-message import passed 209 direct value comparisons, including nine tool calls. Resumed attachments and cross-route identity remain separate gates. 2026-10-09 native trial (partial): Installed ca2af9bd passed 209 native source-value comparisons for a 48-event, 24-message Remote Control sample. Chromium and WebKit rendered the correct Remote Control Source label, all 24 messages, and 15 tool cards. Both failed the project button label because the archived automatic scope still names Claude Code cloud. Installed scope correction and rendered recheck remain pending. 2026-10-09 native trial (partial): Installed 738e0ece matches all 1350 inspected runtime and package files. The bundled CLI rebuilt one native Remote Control session from retained raw inputs with no errors. Parser 0.6.2.1 corrected the automatic scope to claude-code/uncategorized. All 24 messages, nine tool calls, nine tool results, identity, empty workspace, repository, and six raw digests remain intact. Only the per-message parser version changed in transcript bytes. Restoring that field in memory reproduces the original complete transcript digest. Chromium and WebKit passed the corrected project button, Remote Control Source label, all 24 reachable messages, all 15 expandable tool cards and results, the recorded Error badge, and horizontal overflow checks. No inspected label misclassifies the session as hosted cloud execution. |
| Claude Code Desktop | claude-code-desktop |
local-store | Claude Code sessions launched from the Claude app when local metadata is available. Metadata includes bounded shared-plan utilization history and tool-family identity hashes. Records with this evidence remain importable without an audit or initial message. Replay preserves sanitized plan samples without account identifiers or tool definitions. Audit transcript when available, with local Claude app metadata used to fill gaps. App 2.31226.0 supplies metadata without direct audit transcripts; its linked CLI history is verified separately. Current model settings remain separate from reply and usage attribution. | fixture-tested; 5 open gates. 2026-10-09 native trial (partial): Claude app 2.31226.0 supplies nineteen local metadata records with no audit transcript or initial message. The direct metadata route discovers nineteen records and imports zero sessions. The CLI join imports all nineteen linked parents and three children, with repeat, raw hashes, full offline raw rewrite, search, viewer payload, Stats, copied growth, and source removal passing. The direct audit route remains blocked by absent transcript data. Its 8111 observed records retain 10061 raw-only field paths. Installed and rendered acceptance remain open. 2026-10-09 native trial (partial): Installed b69492f2 passes the CLI join for all nineteen Desktop metadata records and three children. The 22 joined sessions pass import, repeat, raw hashes, full raw rewrite, search, viewer payload, Stats, source removal, and copied growth. The direct Desktop audit route remains without a native transcript. This result does not promote that direct route or rendered browser acceptance. |
| Claude Duet (partial) | claude-duet |
browser, cloud | The browser extension’s Claude Duet collector reads each organization’s channels on claude.ai with the browser’s sign-in; an opt-in cloud reader reads the same channels with a saved Claude Code sign-in. Channels are kept by structure (task statuses, tasks turned on, thread-started notices), never by name, and Claude Code projects are skipped. Each channel’s main thread and each task become linked conversations with status, needs-input, completion, decision cards with their answers, inferred Link and Shop Pay purchases, Make outputs with inline bytes, and the steps of each task’s bound cloud sessions, with retained raw slices and offline replay. The record format comes from the claude.ai web client and is unverified against a native capture. On the inspected Mac the Claude Code sign-ins reached no Duet channel, so the browser route is required there. Tracks have no confirmed field and fall back to the channel; the u-prefixed task ids in Duet’s addresses match no known record. Purchases are inferred. Outputs without inline bytes keep metadata only, and signed addresses are never followed. | fixture-tested; 5 open gates. |
| Claude Cowork (cloud runtime) (partial) | claude-cowork-cloud, claude-cowork |
cloud, desktop-cache, browser | Saved profiles enumerate explicitly tagged Cowork tasks. The shared browser route reads workspaces explicitly linked by Claude.ai chats, keeping separate transcripts. Successful Write calls preserve recorded file bytes. Browser captures separately download current files named by supported tools or artifact references and retain observed file versions in raw evidence. Saved-profile enumeration requires Cowork tags and does not download remote files. Browser files use recorded paths, with 64 plans per workspace, 8 MiB per file, and 64 MiB per capture. Missing files, unread paths, and incomplete event pages remain explicit. The retained cohort has a 256 MiB bound. Periodic account reconciliation requires source enablement. Native linked-workspace and remote-file acquisition remain unverified. | fixture-tested; 3 open gates. 2026-10-06 native trial (partial): Nineteen tagged tasks were enumerated. Eighteen conversations preserved 276 messages, verified raw snapshots, tools, models, and usage. Repeated import added no duplicates. Offline raw rebuild preserved every transcript byte. Search, source filters, viewer payloads, and Stats passed. One task had metadata only. A separate resumed Design task reached a succeeded scheduled run with twelve messages in an isolated supervisor runtime. Remote file downloads remain unverified. 2026-10-08 native trial (partial): After authorized source enablement, the installed scheduled import reached succeeded: twenty candidates, nineteen imported, one skipped, zero errors, and no remaining candidates. Earlier import-lock waits did not consume the source failure budget. Separately, native Design events passed recorded-file hashes, repeat, offline replay, search, viewer payload, and Stats in an isolated archive. Remote-file acquisition remains unverified. 2026-10-08 native trial (partial): Installed build 079919d2 parsed retained native Design events and wrote twelve messages into a fresh isolated archive using bundled Node. Both recorded-write files retained exact source digests at 272 and 842 bytes. Repeat, raw replay, raw integrity, search, viewer payload, and Stats passed. Current remote-file acquisition and linked-workspace acquisition remain unverified. 2026-10-08 native trial (partial): Installed 079919d2 passed 172 bounded source-value comparisons across 74 events and twelve messages. Chromium and WebKit rendered twelve messages and four expandable tool calls/results. Both recorded-write attachment controls were missing, although direct resource reads returned exact 272-byte and 842-byte files. Source repair 8a143c3b awaits installed rendering. Complete field accounting remains open. 2026-10-08 native trial (partial): The event-stream field contract accounts for 711 generalized paths across 184 native records: 49 normalized declarations and 662 raw-preservation declarations. No unreviewed fields, type drift, discriminator drift, or missing required paths occurred. Installed 8a143c3b retained exact raw bytes for all 74 settled events and 110 transport records. This bounded cohort does not close account-detail, linked-workspace, or current-file fidelity. 2026-10-08 native trial (partial): A read-only viewer with runtime files verified against installed 8a143c3b passed twenty headless Chromium and WebKit assertions for both recorded-write attachment controls. The controls become visible after opening the turn fold, tool group, and edit callout. Both clicked downloads match their complete native Write inputs at 272 and 842 bytes, with 36 independent size, digest, response, and visibility comparisons. Current remote files, rendered exports, and the desktop application window remain separate boundaries. |
| Claude Cowork (local audits and desktop cache) | claude-cowork |
local-store | Claude Cowork legacy local audits and tagged remote sessions cached by Claude Desktop. Audit results retain estimated cost increments, bounded model usage and context limits, and rate-limit notifications. Empty metadata entries are excluded. Repeated cache imports retain earlier turns and raw snapshots; partial parser refreshes remain explicit. Legacy audit transcripts, or cached remote transcripts after the conversation is opened or revisited in Claude Desktop. Current model settings are separate from reply and usage attribution; missing reply models remain unknown. The September 2026 unified chat/task rollout does not establish additional capture coverage; keep recorded source and execution identity. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): Claude app 2.31226.0 has no usable local session metadata, audit files, or remote mappings under ~/Library/Application Support/Claude/local-agent-mode-sessions. The cache inventory finds 69 relevant entries: 68 watch SSE records and one event stream. None is Cowork-tagged, and the cache reader discovers zero sessions. Native import, replay, growth, and consumer trials remain blocked by absent local session data. Provider creation and sign-in were outside the trial. |
| Claude SDK jobs | claude-sdk |
local-store | Programmatic SDK and agent batch runs under ~/.claude/projects; stats split delegated runs from one-shot batch jobs. Hansard-launched turns retain bounded cumulative model usage and plan snapshots as device-local evidence without adding duplicate charges. Full JSONL transcript and usage; delegated runs count in stats while one-shot jobs stay in the SDK jobs split. | fixture-tested; 3 open gates. 2026-09-29 native trial (partial): Three retained SDK sessions passed repeat-import and raw-hash checks; full metadata and viewer acceptance remain open. 2026-10-09 native trial (partial): All seven retained SDK jobs pass isolated import, byte-identical repeat, exact raw hashes, full offline raw rewrite with transcript and metadata parity, search, viewer payload, SDK-inclusive Stats, and source removal. Copied native growth adds two messages under the existing identity and repeats identically. Field accounting covers 1505 records and 2213 raw-only paths, including the saved tool-result sidecar. Complete normalized fidelity, rendered viewer, native supervisor, and installed acceptance remain open. 2026-10-09 native trial (partial): Installed b69492f2 and bundled Node 24.20.0 pass all seven retained SDK jobs: import, byte-identical repeat, exact raw hashes, full offline raw rewrite with transcript and metadata parity, search, viewer payload, SDK-inclusive Stats, source removal, copied growth, and repeat after growth. All 1384 installed manifest files match. Full normalized values, rendered browser acceptance, and settled native supervisor capture remain open. |
| Claude.ai (consumer and organization chat) | claude-web |
manual-export, browser | Claude.ai account exports and browser capture preserve conversation trees, attachments, account identity, and provider metadata; export memories become memory records. ZIP, JSON, folder, and combined-input exports replay after digest and identity checks. New browser and mixed histories bind the ordered capture and export inputs needed for replay, including retained files, models, and metadata. Models and token usage remain unknown when absent. Browser captures depend on retained account data and complete conversation loading. A five-message native mixed-history replay passed after explicit account alignment. All 873 nonblank sessions in the downloaded export pass repeat and offline replay. Automatic cross-route identity, broader provider history, and complete metadata fidelity remain open. Older browser histories without verified input bindings refuse replay. A new import binds only when its inputs reproduce the retained history; otherwise the richer archive stays with an explicit reason. Legacy conversation-style memory replay remains deferred. The unified chat/task rollout does not establish additional capture coverage. | fixture-tested; 4 open gates. 2026-10-08 native trial (partial): Opening an existing signed-in synthetic conversation refreshed the installed collector automatically. Four visible messages and one provider summary survived isolated repeat import. After the replay repair, all five messages rebuilt offline on committed source. A second replay and a fresh import matched byte-for-byte. Search, viewer payload, and Stats passed. The sample had no attachments. New-turn and account-wide acceptance remain open. 2026-10-08 native trial (partial): One retained native capture accounted for 66 generalized field paths under parser 0.6.2.1. Forty-four paths map to normalized fields and twenty-two remain raw. A separate fidelity audit passed 105 assertions against source values and the preserved archive. The sample has four text messages and one provider summary. It provides no file, tool, project, linked-workspace, alternate-branch, or complete account-history acceptance. Organization fields lack a separate saved-account receipt. 2026-10-08 native trial (blocked): The bounded native inventory found no shared Claude export directory at the configured archive location and no matching candidates in the top-level Downloads scan. This does not establish absence elsewhere. No new export was requested. Synthetic export replay, repeated rebuilds, digest failures, staged-manifest checks, and path confinement passed. Native export acceptance remains open. 2026-10-08 native trial (partial): Fresh import on 90b00ac8 under parser 0.6.2.2 retained a verified input sequence for the same five-message native capture. Repeat import, raw replay, recovery after an empty transcript, search, viewer payload, and Stats passed. All 105 source-value assertions passed. The older unbound native archive refused replay without transcript or metadata mutation. This sample provides no native export or mixed-history acceptance. 2026-10-08 native trial (partial): Installed build 079919d2 imported the retained five-message native capture into a fresh isolated archive using bundled Node. The verified input sequence, repeat import, raw replay, raw integrity, search, viewer payload, and Stats passed. This retained-input check does not establish new-turn, export, mixed-history, or attachment acceptance. 2026-10-08 native trial (partial): The authorized native account export request returned Export started. The request selected all conversations and included Users and Projects. Its email download remains pending, so native export and mixed-history replay are not yet verified. 2026-10-08 native trial (partial): After confirmed 8a143c3b activation, a fresh baseline archived four visible messages and one provider summary. One authorized resumed text turn produced the exact requested reply and reached installed capture automatically under the same identity. The seven-message result passed installed isolated repeat, stale-prefix retention, raw integrity, offline replay, search, viewer payload, and Stats. No manual import established the production continuation. New-session and attachment acceptance remain open. 2026-10-08 native trial (partial): The native export downloaded a version 1.0 manifest with seven ZIP links for light metadata, projects, memories, feedback, Design chats, frames, and conversations. The manifest contains no conversation payloads. Its links expire after twenty-four hours and may work only once. The ZIP downloads remain pending, so native export and mixed-history replay are still unverified. 2026-10-09 native trial (partial): Installed 738e0ece imported one native export conversation followed by its native browser capture. Five normalized messages passed byte-identical repeat, offline replay after intake removal, empty-transcript recovery, raw integrity, search, viewer payload, and Stats. The export user account and browser organization differ, so the isolated trial used explicit account alignment. Automatic cross-route identity remains open. The complete export holds 891 ordinary conversations, one Design chat, and 7315 ordinary source messages. Its first import produced 873 sessions and 19 skips without importer errors. All 873 transcripts, with 22521 normalized messages, pass byte-identical repeat and full offline replay. Sixteen empty conversations and three blank-only conversations explain every skip. All 7280 ordinary source-message identities, 4646 plain-text comparisons, 4343 call arguments, 4335 result identities, 37 normalized memory records, and eight preserved-file digests pass. Search, representative viewer payload, and full Stats pass. The installed Design projection gap remains separately recorded. The isolated archive was removed. |
| Claude unified chat and tasks (partial) | claude-web, claude-cowork-cloud |
cloud, browser, manual-export | Browser collection enumerates active and archived v2 chat lists and reads explicitly linked workspace event pages in resumable steps. It validates organization and session relationships, archives each workspace separately, and preserves the chat link through replay. Current files named by supported tools or artifact references download separately from recorded Write bytes, with retained earlier versions. Native acquisition of linked workspaces and current files remains unverified after confirmed extension activation. Workspace files have recorded-path and size limits. Missing files and incomplete event pages retain their reasons while the complete chat remains importable. Events beyond the capture envelope remain incomplete. Full scheduled-run history and account-wide completeness remain unverified. Product search excludes older Cowork tasks. | fixture-tested; 4 open gates. 2026-10-08 native trial (partial): Four retained native chat-list rows explicitly link upgraded chats to workspace session IDs. An existing native chat passed repeat, offline replay, search, viewer payload, and Stats. Linked-workspace admission, separate identity, recorded file writes, organization refusals, raw rebuild metadata, and resumable event acquisition pass synthetic tests. Shared extension activation is confirmed. Native linked-workspace acquisition remains unverified, with browser control and receiver throughput impaired by host overload. 2026-10-08 native trial (blocked): The existing linked-workspace page displays a failed response, an exhausted usage limit, and an unavailable local computer. No additional prompt was sent. This establishes an access boundary for the inspected trial, not successful linked-workspace or current-file acquisition. The installed receiver is at 10cf96ee, while the latest extension reload remains unconfirmed. |
| Claude Science (partial) | claude-science |
local-store | Local and organization-scoped SQLite stores preserve native messages, tools, structured/server outputs, background completion, captured output files, images, artifact revisions, child summaries, generated context, and recorded usage. Live tool capture passed repeat, offline replay, search, canonical tool counts, viewer payload, and Stats checks. Credential tables are excluded. Account and project memory backup preserves permitted native records and revisions independently of frame links. Native Science 0.1.56 trials cover local tools, files, a bundled PubMed call, compute inventory, memory records, and installed rendering. Remote compute jobs and routines remain unverified: the inspected account has no compute provider, and the client routine interface is disabled. Other connectors, versions, and platforms need separate acceptance. Nested host SDK methods remain enclosing code unless the provider records distinct events. Branch archives, system prompts, and reproducibility snapshots remain raw evidence. Missing response models remain unknown. | fixture-tested; 3 open gates. 2026-10-07 native trial (partial): Eleven native tool calls and results, failed exit status, background completion, three artifact revisions, captured PNG, and 22039-byte full output passed source capture and offline replay. 2026-10-07 native trial (partial): A resumed PubMed call and compute inventory reached installed Hansard automatically. All thirteen sessions replayed offline, 660 envelopes passed field accounting, and three native memory records were captured independently. Compute providers and an enabled routine interface were unavailable. 2026-10-08 native trial (partial): Installed Science 0.1.56 supplied thirteen nonempty sessions and two empty frames. Parser 0.6.2.1 retained 844 normalized messages, 87 matching attachment digests, and 114 matching raw digests. Repeat added no sessions; all thirteen offline replays preserved transcript bytes. Search, viewer payloads, and Stats with subagents passed. The full field audit accounted for 660 envelopes and 312 paths without drift. The store had three frame-linked memories, no routine schedules, no compute usage, and no native text-document blocks. Public documentation lists 0.1.60. |
| Claude Design (partial) | claude-cowork-cloud, claude-design, claude-web |
cloud, browser, manual-export | Design conversation and tool events import through Cowork tasks. Successful Write calls retain exact source bytes and artifact references. A standalone Design page starts bounded project enumeration across signed-in organizations. Persisted queries, pending projects, and page cursors survive restarts and tab navigation. The collector reads available chats, current files, and exposed revisions through read-only Connect methods. Separate project/chat identities retain observed file versions and replay from verified raw inputs. One native standalone project passed automatic installed acquisition, exact file comparison, repeat, offline replay, field accounting, and consumer checks. Synthetic pagination, restart, integrity, and preservation checks pass. The accessible organization’s one-project native inventory matches the extension scan. A second organization fails an access check. Account export Design chats now preserve nested text, thinking, tool calls and outputs, usage, attachment descriptors, project identity, and distinct wrapper and content IDs. The native candidate and installed runtime pass repeat, offline recovery, integrity, and consumer checks. Installed rendering passes in Chromium and WebKit. Account-wide project enumeration remains incomplete because a second organization fails a sign-in or browser check. Complete project metadata, direct edits, other export formats, and metadata-only closed chats remain unverified. The inspected native client disables its former closed-chat reader. A native thumbnail was refused because its inventory declares 3266 bytes while the response contains 9034 bytes. Standalone ListVersions returned HTTP 501, while the earlier Cowork version list was empty. Scans stop at 200 pages and 2000 projects per organization, 500 projects per page, and 1000 chats per project. File bounds are 8 MiB per file, 64 MiB per project group, and 500 retained attachments. Provider revisions, observed file versions, recorded writes, and rendered exports remain distinct. Standalone Design closes December 14, 2026. Chats and comments do not migrate to Claude artifacts. Design export chats remain under the Claude.ai export source. Their identities do not automatically merge with browser Design identities. The native account-export sample does not verify nonempty thinking rendering. | fixture-tested; 4 open gates. 2026-10-06 native trial (partial): A private Design task and its resumed final reply reached a succeeded scheduled import with twelve normalized messages and one artifact reference. Raw replay, search, viewer payload, and Stats checks passed. Artifact bytes and complete revisions remain unverified. 2026-10-08 native trial (partial): Retained native events yielded twelve messages and two recorded-write files of 272 and 842 bytes, with exact hashes. Repeat, offline replay, search, viewer payload, and Stats passed. A separate HTML export contained 107869 bytes. An authorized standalone synthetic project then produced a blue rectangle in the native UI. Its Connect API response and installed automatic acquisition remain unverified. The earlier Cowork version list was empty. 2026-10-08 native trial (partial): The approved standalone project displays its completed assistant reply and blue rectangle. Its native Download action returned a 799-byte HTML file, retained privately with a digest for independent comparison. Twelve inspected runtime files in installed build 10cf96ee match committed source and include the standalone receiver. Extension acquisition and captured-file comparison await confirmed activation of the rebuilt extension. The download alone does not establish automatic capture. 2026-10-08 native trial (blocked): After all collectors were enabled, the signed-in native project refreshed and rendered its completed design. A stable extension database snapshot recorded C.claudeDesign is not a function. The shared distribution includes that helper and injects it first, indicating an older loaded worker. Native Connect acquisition awaits a full extension reload. The free Project archive download is a 25947-byte ZIP with three members, and its 799-byte HTML matches the independent file download. Official migration documentation says standalone closes December 14, 2026 and chats and comments do not migrate. 2026-10-08 native trial (partial): Installed build 079919d2 passed isolated native-event acceptance for the twelve-message Cowork Design sample and its two recorded-write files of 272 and 842 bytes. Repeat, offline replay, raw integrity, search, viewer payload, and Stats passed. This evidence does not establish standalone Design acquisition, current remote files, project enumeration, or rendered attachment controls. 2026-10-08 native trial (partial): Installed 079919d2 passed 172 bounded source-value comparisons for the Cowork Design sample. Both browser engines rendered twelve messages and four tool calls/results. Two missing file controls remain an installed rendering failure until source repair 8a143c3b is rechecked. Direct archived resource reads returned exact file bytes. Standalone Design and native current-file acquisition remain separate gates. 2026-10-08 native trial (partial): The Cowork Design event stream passed bounded accounting of 711 generalized field paths. Installed 8a143c3b preserved the complete 202079-byte verification envelope, including all 74 settled events and 110 transport records. The separate normalization audit passed 172 value comparisons. Standalone Design acquisition remains outside this cohort. 2026-10-08 native trial (partial): After the confirmed 8a143c3b reload, automatic standalone Connect acquisition archived five normalized messages and two files of 799 and 69150 bytes. Both files match the independent native ZIP, and the HTML matches a separate native download. Twenty-four source-value comparisons and a 124-path field audit passed. Installed repeat, raw integrity, offline replay without receiver intake, search, viewer payload, and Stats passed. The thumbnail failed a size check, and ListVersions returned HTTP 501. Account-wide inventory, direct edits, full project metadata, and closed chats remain separate gates. 2026-10-08 native trial (partial): A read-only viewer with runtime files verified against installed 8a143c3b passed twenty headless Chromium and WebKit assertions for both recorded-write attachment controls. The controls become visible after opening the turn fold, tool group, and edit callout. Both clicked downloads match their complete native Write inputs at 272 and 842 bytes, with 36 independent size, digest, response, and visibility comparisons. Current remote files, rendered exports, and the desktop application window remain separate boundaries. 2026-10-09 native trial (partial): After the confirmed ca2af9bd reload, the native Design home and extension scan each listed one project in the accessible organization. The next organization returned a sign-in or browser-check error and did not appear in the native account menu. Two fresh automatic captures archived five messages under the same identity. Installed acceptance passed 24 source-value comparisons, both independently checked file hashes, repeat, raw integrity, offline replay after intake removal, search, viewer payload, and Stats. The thumbnail diagnostic records 3266 bytes expected and 9034 received. ListVersions remains HTTP 501. Account-wide completion remains an access gate. 2026-10-09 native trial (partial): One official Design chat export exposed two omitted tool calls and outputs, thinking, and a skill attachment descriptor. Candidate parser 0.6.2.2 against base eff41c41 passed 46 native value and recovery checks and all 50 bounded field declarations. Two wrappers produce four normalized messages with both identity namespaces and clocks intact. Repeat, offline recovery after intake removal, raw integrity, search, viewer payload, and Stats passed with bundled Node 24.20.0 and signed dependencies. The 200-session base/candidate/candidate/base benchmark passed every enforced budget at 21.2, 20.9, 21.6, and 20.0 sessions per second. This measures the shared synthetic write path, not Design acquisition. Installed acceptance remains pending. File bytes, revisions, and automatic browser/export identity remain separate boundaries. 2026-10-09 native trial (partial): Installed 313e215f modified passed the same 46 native Design value and recovery checks and all 50 bounded field declarations. All 1364 runtime files match the build manifest, and the five Design runtime files match the measured candidate. The four-message transcript remains byte-identical after import, repeat, and offline recovery. Both calls and results, thinking, usage, the attachment descriptor, wrapper UUIDs, content IDs, and clocks remain intact. Raw integrity, search, viewer payload, and Stats pass. Chromium 153 and WebKit 26.6 each pass six requested rendering checks: four reachable messages, both expandable tool cards and exact outputs, the intact skill descriptor, distinct wrapper/content IDs, one-conversation Stats, and no horizontal overflow. The seventh assertion incorrectly expected visible thinking. Independent ZIP inspection finds only an empty thinking text field, with no signature or summary. Both engines preserve that empty payload and correctly draw no disclosure. The corrected expectation passes against recorded payload and DOM evidence without a new browser run or runtime change. The original failed assertion remains in private evidence. This sample does not verify nonempty thinking rendering. Source session files and raw bytes remain unchanged, no screenshots were taken, and the viewer stopped before isolated archive cleanup. Exact 313e215f and the modified build snapshot each passed the complete project check. The modified snapshot excludes three unrelated derivation-cache commits. Browser/export identity, current file bytes, provider revisions, and account-wide completeness remain separate boundaries. |
Augment
Section titled “Augment”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Augment Cosmos cloud | augment-cloud |
cloud | Cosmos API pagination imports conversations, tools, measured tokens, credits, compaction, and child relationships. Cloud pages must agree with the reported count. Local synchronized caches can stop at 200 messages and must not replace fuller API captures. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No configured Augment cloud session or credential is available. Authenticated Cosmos listing, tools, child histories, and field fidelity remain unverified. |
| Augment local history | augment |
local-store | Auggie JSON sessions, IDE LevelDB exchanges, locally synced Cosmos sessions, tool arguments, edits, and recorded usage. CLI and IDE exchanges, tools, recorded models and tokens. Cosmos local sync can truncate history; credits and child cost summaries remain separate from tokens. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): Read-only discovery found settings residue without retained CLI, IDE, or Cosmos conversation candidates. Native round-trip and resumed capture remain blocked. |
ByteDance
Section titled “ByteDance”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| TRAE | trae |
local-store | Agent trajectories and desktop history retain recorded agent tokens, context and compression counters, and child-run context evidence. Encrypted stores require a supplied key and SQLCipher. Desktop context and compression counters are not billed usage. No desktop credits are inferred. Model selections do not establish request attribution, and remote histories remain limited to local caches. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): An encrypted 831488-byte desktop database is present. No SQLCipher key or unlocked snapshot is configured. Native content and field acceptance remain blocked; no readable Agent candidates were found. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Crush | crush |
local-store | Crush project-local .crush/crush.db sessions, including reasoning, tools, attachments, and child sessions. SQLite messages, reasoning, tools, embedded attachments, and child sessions normalize. File history, channel, and Prism savings stay raw. Token counters describe latest context and cost can include children. Built-in raw replay is deferred. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no project .crush/crush.db, ~/.crush/crush.db, ~/.local/share/crush/crush.db, or ~/Library/Application Support/crush/crush.db, and no configured alternative. Audited v0.98.0. Installation plus separately authorized activity could create a trial store. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Cline | cline |
local-store | Historical Cline task folders from editor globalStorage, including UI chronology, matching API blocks, and task metadata. Current SDK stores use the cline-cli route. Historical task UI messages define chronology, including an empty UI history. Matching API blocks enrich that chronology. Checkpoint repositories are excluded. Current 4.x sessions use the SDK route registered as cline-cli. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no task stores in the checked editor, profile, JetBrains, or remote-server roots, ~/.cline/tasks, or ~/.cline/data/sessions. Audited v4.1.23. A current installation writes SDK sessions; retained task files are needed for the historical task route. |
| Cline CLI | cline-cli |
local-store | Cline SDK sessions from ~/.cline/data/sessions, including CLI, desktop, and IDE surfaces plus subagent transcripts. Version-1 SDK manifests and message artifacts preserve tools, usage, cost, and linked child transcripts. Compaction is structural evidence outside visible chronology. Shared indexes, queues, and team state outside the bundle are excluded. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no ~/.cline/data/sessions, ~/.cline/tasks, ~/.cline/data/tasks, or Cline Application Support session roots, and no environment override. Audited cli-v3.0.70 and extension v4.1.23 SDK writers. Installation plus separately authorized activity could create a trial store. |
Cognition
Section titled “Cognition”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Ask Devin | devin-ask-browser |
browser | Browser collection separates Ask query IDs from agent session IDs and retains questions, answers, repository and branch context, citations, and query-to-agent links. Native repository, branch, query-to-agent links, repeat import, search, viewer, and Stats pass for one bounded query conversation. Installed 8a143c3b raw replay preserves the retained 21-message native transcript. Fresh Ask receipts at 22:54 UTC were skipped as older snapshots, and their contents remain unverified. Broader history coverage, per-question clocks, and measured token usage remain incomplete. | fixture-tested; 1 open gates. 2026-10-08 native trial (blocked): The browser reached the Devin sign-in page. Native repository/branch comparison and query-to-agent round trips could not run. No paid prompt or provider write was made. 2026-10-08 native trial (partial): Signed-in installed 731cc154 trial: one Ask query conversation and its Construct Devin Prompt step produced 21 messages, five paired tools, two citations, and links to the single launched agent. Repository and branch match the native UI and rendered Devin details. Automatic receiver delivery, raw hash verification, repeat byte equality, search, viewer navigation, and Stats pass. Production jobs remain queued. Offline replay returns no-raw-parser-entrypoint. Source replay repair rebuilt both retained native captures with unchanged transcripts. At 19:13 UTC seven trial jobs were archived, one stale capture was skipped, and six later jobs were queued. Installed repair acceptance remains pending. 2026-10-08 native trial (partial): Installed 10cf96ee matches six committed receiver files. Bundled-runtime replay preserves the 21 Ask and six agent messages exactly. Repeat import, raw integrity, search, viewer payloads, and Stats pass in an isolated archive. By 20:15 UTC fourteen production trial jobs archived and one stale capture skipped, with none queued. Both complete production transcripts are indexed and searchable. The bounded native repository, branch, and linked-agent comparisons remain valid. Broader history coverage remains unverified. 2026-10-08 native trial (partial): Installed 079919d2 matches six receiver files byte for byte. A bounded production read verifies 21 Ask messages, six agent messages, and both original attachment sizes and hashes. The bundled runtime passes isolated repeat and offline replay after temporary receiver-intake deletion, preserving exact transcripts and both file hashes. Raw integrity, search, viewer payloads, and Stats pass with two conversations, three user messages, six tool calls, and no measured tokens. This uses retained native evidence without a new provider request. Fresh extension capture remains held for reload confirmation. 2026-10-08 native trial (partial): Installed 8a143c3b matches six receiver files and uses bundled Node 24.20.0. After the confirmed extension reload, two automatic agent captures at 22:53 and 22:54 UTC archived six messages and the exact 2,097,152-byte binary and 2,289-byte MP4 hashes. The fresh agent capture and retained 21-message Ask capture pass isolated repeat, offline replay after temporary receiver-intake deletion, raw integrity, search, viewer payloads, and Stats. Stats retains two conversations, three user messages, six tool calls, and no measured tokens. Fresh Ask captures were skipped as older-cloud-snapshot, with their contents unavailable for comparison. No new provider prompt ran. Rendered attachment verification remains pending. |
| Devin CLI | devin-cli |
local-store | Local sessions.db active-chain history with recorded generation models, usage, reasoning, tool timing, rules, and linked subagents. Native repeat, cross-route identity, search, viewer, and Stats checks passed for one retained session. Stored node context, compaction ancestry, step timing, and ACUs or credits use the shared runtime contract. One-session native coverage does not verify a new or resumed automatic capture. Current captures can retain database references without the snapshot identity needed for raw replay. Hidden branches and uncommon records remain partial. Runtime evidence extensions have synthetic regression coverage; native emission across installed versions remains unverified. | fixture-tested; 4 open gates. 2026-09-29 native trial (partial): One retained session passed repeat-import and raw-hash checks; full metadata and viewer acceptance remain open. 2026-10-08 native trial (partial): One retained session normalized 24 messages; repeat imported zero with identical transcript bytes. Its Desktop mirror merged into the same identity. Search and source-tree viewer passed, and Stats counted the merged archive. Raw replay failed with SOURCE_COHORT_SOURCE_FORMAT_REIMPORT_REQUIRED because the capture lacked snapshot identity. |
| Devin cloud | devin-cloud |
cloud, browser, desktop-cache | API chat and session metadata combine with browser event capture and Desktop mirrors. A bounded native browser trial automatically acquired binary and MP4 originals with verified sizes and hashes. Repeat import and offline replay preserved both files. API snapshots retain ACUs, Teams on-demand USD excluding children, and status transitions at their observation time. Installed 8a143c3b and its reloaded extension pass fresh agent capture and replay with both original binary and MP4 hashes. Both production conversations remain complete. Forced native continuation, recording events, and API listing remain unaccepted. Redirect-hop details and the earlier collector revision were not retained. Rendered attachment verification and remote copies remain untested. Runtime evidence extensions have synthetic regression coverage; native emission across installed versions remains unverified. | fixture-tested; 3 open gates. 2026-10-08 native trial (blocked): The API credential probe returned auth_unavailable and the browser required sign-in. Nine Desktop cloud mirrors are local-cache evidence only. Native redirected attachment and recording acceptance remains pending. 2026-10-08 native trial (partial): Installed 731cc154 received the approved synthetic session automatically in running and completed states. Its terminal event page normalizes 18 events into six messages with shell evidence and an Ask launch link. Both native files, a 2 MiB binary and a one-second MP4, failed after three attempts with invalid byte range. Isolated repeat bytes, raw integrity, search, rendered viewer, and Stats pass. Production jobs remain queued. API listing returns auth_unavailable; offline replay returns no-raw-parser-entrypoint. Source replay repair rebuilt both retained native captures with unchanged transcripts. At 19:13 UTC seven trial jobs were archived, one stale capture was skipped, and six later jobs were queued. Installed repair acceptance remains pending. 2026-10-08 native trial (partial): Installed 10cf96ee matches six committed receiver files. Bundled-runtime replay preserves the 21 Ask and six agent messages exactly. Repeat import, raw integrity, search, viewer payloads, and Stats pass in an isolated archive. By 20:15 UTC fourteen production trial jobs archived and one stale capture skipped, with none queued. Both complete production transcripts are indexed and searchable. Both native files remain unavailable in the retained capture. Download repair acceptance awaits a fresh capture after extension reload. The installed bounded API listing again returned auth_unavailable at 20:15 UTC. 2026-10-08 native trial (blocked): At 20:32 UTC the existing native page still displayed both synthetic files and no repository changes, but reported that it could not refresh and was showing its last loaded version. The latest extension reload remained unconfirmed. No post-reload download trial or new provider prompt ran. Verified attachment bytes, native continuation, and recording-event acceptance remain open. Earlier installed transcript and replay passes remain valid. 2026-10-08 native trial (blocked): After confirmation of extension f006c25b, the existing native session loaded signed in without the earlier refresh error. Both files were visible and the Changes panel showed no repository edits. No new prompt ran. The 20:42:54 UTC session-linked job query showed no fresh capture. Later terminal receipts establish that pending jobs without archive session linkage were omitted. The latest archived capture then visible still recorded both invalid-range failures, with six messages and valid raw hashes. Git confirms f006c25b predates download repair a2f1b624; this version observation does not identify the collector that later acquired verified files. 2026-10-08 native trial (partial): Automatic recapture acquired at 20:41:27 UTC archived at 20:44:34 UTC on receiver 10cf96ee. The binary has 2,097,152 bytes and the MP4 2,289 bytes; both SHA-256 hashes match the native originals. Installed isolated repeat and offline replay preserve all 21 Ask and six agent messages and both exact file hashes after temporary receiver-intake deletion. Raw integrity, search, viewer payloads, and Stats pass. No new provider prompt ran. Forced native continuation, recording_stopped events, rendered attachment verification, redirect-hop details, and exact collector revision remain unverified. API listing remains auth_unavailable. 2026-10-08 native trial (partial): Installed 079919d2 matches six receiver files byte for byte. A bounded production read verifies 21 Ask messages, six agent messages, and both original attachment sizes and hashes. The bundled runtime passes isolated repeat and offline replay after temporary receiver-intake deletion, preserving exact transcripts and both file hashes. Raw integrity, search, viewer payloads, and Stats pass with two conversations, three user messages, six tool calls, and no measured tokens. This uses retained native evidence without a new provider request. Fresh extension capture remains held for reload confirmation. 2026-10-08 native trial (partial): Installed 8a143c3b matches six receiver files and uses bundled Node 24.20.0. After the confirmed extension reload, two automatic agent captures at 22:53 and 22:54 UTC archived six messages and the exact 2,097,152-byte binary and 2,289-byte MP4 hashes. The fresh agent capture and retained 21-message Ask capture pass isolated repeat, offline replay after temporary receiver-intake deletion, raw integrity, search, viewer payloads, and Stats. Stats retains two conversations, three user messages, six tool calls, and no measured tokens. Fresh Ask captures were skipped as older-cloud-snapshot, with their contents unavailable for comparison. No new provider prompt ran. Rendered attachment verification remains pending. |
| Devin Desktop | devin-desktop |
local-store | ACP schema-6 SQLite message stores and legacy event logs preserve ordered transcript, thinking, tools, plans, modes, usage, and workspace context. Selected model settings stay separate from session-version attribution. ACP metadata retains hooks, terminal exits, subagents, three-section context snapshots, cache writes, request USD, and recorded credits. Ten supported stores were found among 44 database files. Exact generation models were absent from native message records. Adaptive/Fusion internals and new-turn capture remain unverified. Current message-store raw replay has no parser entry point. Runtime evidence extensions have synthetic regression coverage; native emission across installed versions remains unverified. | fixture-tested; 4 open gates. 2026-09-29 native trial (partial): Ten retained sessions passed repeat-import and raw-hash checks. Credential-database references contain no copied bytes; model provenance remains under review. 2026-10-08 native trial (partial): Ten native stores imported and the repeat found zero changed candidates, preserving transcript bytes. The installed app is 3.10.48, with editor runtime 1.126.0. Twenty-seven cloud-mirror raw records passed hash checks; twelve provider files retained their hashes. Selected model config and session-version provenance were separated. Search, source-tree viewer, and all-time Stats passed. Message-store raw replay remains unsupported. |
| Windsurf | windsurf |
local-store, local-service, manual-export | Legacy Cascade transcript hooks, explicit local-service recovery, trajectory Markdown exports, brain artifacts and opaque protobuf preservation; exact trajectory IDs reconcile routes. Stored ACP metadata and decoded step protobuf metadata retain runtime evidence, context occupancy, credits, and token dimensions. Only one opaque native identity is currently accessible. Readable recovery requires a compatible service or export. Current Devin releases removed Cascade, while older guides remain online. Transcript hooks retain at most 100 files. No complete account-history or current automatic acceptance is claimed. Runtime evidence extensions have synthetic regression coverage; native emission across installed versions remains unverified. | fixture-tested; 6 open gates. 2026-09-29 native trial (partial): One raw-backup session passed repeat-import and hash checks but contains zero normalized messages. Full conversation import remains unverified. 2026-10-08 native trial (partial): One opaque Cascade file matches the sole Cascade metadata-cache ID; no extra metadata-only identity, transcript-hook store or readable brain history was found. Raw hash and repeat import pass with zero messages. No compatible language server is running. Installed Devin 3.10.48 postdates documented Cascade removal. App startup remains unapproved, so the earlier one-conversation recovery was not reproduced. |
| Poke (partial) | poke |
local-db | Opt-in macOS Messages import for one configured exact handle, through the same scoped reader and importer as Instinct. A single read-only query captures only one-to-one iMessage, RCS, and SMS rows and selected attachments. Stable timeline identity, typed text, observed revisions, reactions, replies, raw preservation and offline replay pass synthetic tests. CLI configuration and the Settings number field share the same setting. Native and installed capture remain blocked on macOS Messages access and the approved reader deployment. Poke publishes no number every person texts, and its number depends on the channel a person picks. Recipes, integrations, models, measured usage and billing are absent from this route. Poke’s WhatsApp and Telegram chats are not read. | fixture-tested; 7 open gates. |
Continue
Section titled “Continue”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Continue | continue |
local-store | Continue shared JSON sessions, including tool results, context, reasoning, attachments, and recorded usage. Shared IDE and CLI histories do not establish the originating interface, exact message clocks, or absent model IDs. Full session and index bytes preserve unmapped editor state, prompt logs, and UI fields. Offline raw replay has no parser entrypoint. Final Continue 2.0.0-vscode was reviewed without a current native cohort. | fixture-tested; 5 open gates. 2026-10-09 native trial (blocked): No local data at ~/.continue/sessions or the inspected VS Code-family continue.continue extension roots. The Continue home exists, but no session store or cn executable was found. No inspected JetBrains installation exists; its documented history uses the same shared Continue store. Final 2.0.0-vscode source was reviewed. An existing session file or separately authorized native session is required. Installing an extension alone does not provide history. |
DeepSeek
Section titled “DeepSeek”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| DeepSeek chat (partial) | deepseek-chat |
browser | Experimental browser capture inventories saved and pinned chats through account-bound time cursors. Stable cache-free history preserves selected ancestry, alternate messages, text, reasoning, file metadata and paired search/open tools. Permitted raw evidence supports offline replay. Native account and history fidelity remain unverified. Original files, image and voice resources, some fragment types, legacy history and independent resources remain incomplete. Official current and legacy export import is unimplemented. DeepSeek Harness and the developer API are separate products. | fixture-tested; 8 open gates. 2026-10-08 native trial (blocked): Read-only Arc inspection displayed the sign-in form and client 2.5.0. Public assets identify client commit 4955e097. No authenticated history or current/legacy export was acquired. This access gate does not establish provider-side impossibility. |
| DeepSeek Harness | deepseek-harness |
local-store | Versioned JSONL and Zstandard logs from local web, headless and agent deployments, including tools, edits, attachments and measured usage. Verified LOG_ONLY approval, hook, compaction, and retry records project into the runtime timeline. Current rc.2 and alpha sources retain format v4. Local logs preserve model routes, tools, usage, attachments, and children; custom roots require HANSARD_DSH_SESSION_ROOTS. Raw replay is deferred. Earlier native claims are historical; current installed and automatic acceptance have no local data. | fixture-tested; 7 open gates. 2026-10-09 native trial (blocked): No local data. No dsh binary or ~/.dsh/sessions store exists, and documented overrides are unset. Installation alone is insufficient. Existing logs or separately authorized conversation creation are required. Earlier 0.1.5-rc.3 and 0.1.7-rc.2 native claims remain documented separately and were not repeated. |
Factory
Section titled “Factory”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Factory cloud sessions | factory-cloud |
cloud | Factory cloud imports retained messages, tool output, usage, and session metadata and preserves richer local session history. The cloud normalizer currently imports tool results without call arguments and preserves full response bodies. Public Sessions API documentation omits message response fields; authenticated tool-argument availability remains unverified. Current session settings remain separate from recorded message models. Full refreshes detect unchanged-clock message changes and retain archived history on incomplete, stale or regressed captures. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): The cloud UI shows Log in/Sign up and no Sessions API key is configured. Current docs omit message response fields, so native tool-argument availability remains unverified rather than proven absent. |
| Factory Droid | factory |
local-store | Factory droid JSONL transcripts and token-usage sidecars from ~/.factory/sessions. Local transcripts retain recorded reply model IDs and reasoning effort independently of current settings. Model-only context is preserved without counting human prompts. Hosted resume, full metadata coverage and cloud access remain unverified. | fixture-tested; 4 open gates. 2026-09-29 native trial (partial): Official Droid 0.229.0 created a synthetic conversation against a local endpoint. One archive retained three records, one human prompt, reply model and effort, declared usage and two hash-verified raw files; repeated import was unchanged. Native resume required hosted authentication and remains unverified. 2026-09-29 native trial (partial): Direct content-addressed and CLI raw replay retain native messages, settings and usage. A second native run retained three tool calls, arguments and results including a controlled error; repeated import and raw-hash checks passed. The bounded metadata contract accounts for 73 observed paths and rejects field, type, role and record-kind drift. Structured diffs, attachments, provider errors, workers, hosted resume and viewer verification remain open. 2026-09-29 native trial (partial): A third official Droid run retained five calls and results, a controlled edit failure, correct file-change counts and a byte-verified tool-result image. Repeat import, raw hashes and direct/forced raw replay passed. The bounded contract accounts for 96 paths across 36 wrapped records, including decoded JSON-string edit results. User uploads, native documents, workers, provider failures, hosted resume and viewer acceptance remain open. 2026-10-08 native trial (blocked): Read-only discovery found cache/log residue without retained Droid sessions. Current native round-trip and resume are blocked. Earlier dated native trials remain historical evidence. |
GitHub
Section titled “GitHub”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| GitHub Copilot CLI | copilot-cli |
local-store | Copilot CLI events.jsonl session state and workspace metadata from ~/.copilot/session-state, with per-request token and premium-request usage from ~/.copilot/session-store.db. Recorded response models, native message lineage, explicit zero counters, and shell exit codes remain distinct. Enriched raw replay needs the preserved usage-ledger binding; older archives require reimport. CLI 1.0.94 and SDK 1.0.18 drift review leaves stable bounded reads, binary-asset retrieval, newer events, complete child identity, and hosted billing as explicit limits. | fixture-tested; 8 open gates. 2026-09-29 native trial (partial): Official CLI 1.0.89 with a local synthetic endpoint: one session resumed from 6 to 18 normalized messages, two human prompts, view success and controlled bash failure, five verified raw records, and unchanged repeat import. All 130 observed event paths audited. Full offline raw rebuild retained transcript, manifest, metadata and identity; a subsequent live import remained unchanged. Hosted billing and other surfaces remain unverified. 2026-10-09 native trial (blocked): No local data at ~/.copilot/session-state or ~/.copilot/session-store.db. No standalone copilot executable was found. Positron has an embedded CLI shim, but its copied session-store.db has schema version 3 and zero sessions or turns. CLI 1.0.94 and SDK 1.0.18 were reviewed upstream. Installation alone does not provide history. The earlier CLI 1.0.89 partial trial remains valid only for its recorded scope. |
| GitHub Copilot coding agent | copilot-agent-cloud |
cloud | GitHub Copilot coding-agent sessions import task transcripts, repository and pull-request context, and recorded billing metadata. Transcript retrieval uses an undocumented endpoint. Premium requests and AI credits are not inference tokens. New task-event streams and artifacts need separate acquisition. Offline cloud replay is unsupported. Non-JSON HTTP 401/403 responses retain authentication/access classification. The native session-list trial stopped at HTTP 403 without returning a session page. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): Read-only listing with existing credentials reached GraphQL with HTTP 200, but /agents/sessions returned HTTP 403, text/plain, 42 bytes. No session page was returned. The response does not establish absent account history or plan eligibility. Installation does not resolve account/endpoint access. VS Code 1.141.0 cloud-backend code was reviewed. Native import, repeat, growth, consumers, and installed acceptance remain blocked. |
| GitHub Copilot in VS Code | copilot-ide |
local-store | Copilot Chat sessions from VS Code-family chatSessions stores, including the append-log format used since VS Code 1.109. Version-3 documents and mutation logs retain requests, responses, tools, edits, and available usage. VS Code 1.141 request token fields and response clocks are supported with stable bounded reads. Composer selections remain separate from historical models. Extension transcripts, Agent Host stores, and JetBrains legacy, relocated, or Nitrite histories have no verified reader. A compatible JetBrains CLI event store uses the separate CLI route. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data in the checked ~/Library/Application Support/<editor>/User/workspaceStorage/*/chatSessions, globalStorage chat roots, profiles, extension transcripts, or agentSessionData. VS Code 1.97.2 and Positron 1.130.0 are installed. Copied Code interactive.sessions values were empty. No JetBrains installation, ~/.config/github-copilot chat roots, or .idea/copilot/chatSessions was found in the bounded project walk. Retained history is required. Installing JetBrains alone does not implement its unsupported readers. Upstream review covered VS Code 1.141.0 and JetBrains plugin 1.18.0-261-linux-arm64. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Antigravity 2.0 | antigravity |
local-store, local-service | Readable transcripts, lifecycle notices, task artifacts, decoded database steps, and partial trajectory summaries from ~/.gemini/antigravity. Opaque conversation protobufs require a compatible running recovery service. The inspected app store is empty and supplies no native legacy corpus. Database tools and log replay pass fixtures; full metadata, artifacts, automatic capture and installed acceptance remain open. | fixture-tested; 6 open gates. 2026-10-08 native trial (blocked): Installed app 2.17.0 has no retained conversation rows or legacy conversation protobuf corpus in the inspected roots. The summary database is empty. No compatible recovery service is running. Provider startup was not authorized. This records bounded local absence, not successful encrypted-body recovery or account-wide absence. |
| Antigravity CLI | antigravity-cli |
local-store | CLI persistent logs, generated context, task artifacts, database-only messages and tools, and recorded generation models and usage. Native retained history, repeat import, log replay, search, viewer payloads and Stats have bounded acceptance. Logs take precedence over database messages, so a result missing from a log can remain raw-only. Database and language-server replay are deferred. Full artifacts, new/resumed automatic capture, rendering and installed-runtime acceptance remain open. | fixture-tested; 6 open gates. 2026-09-29 native trial (partial): Three retained sessions passed repeat-import and raw-hash checks; other Antigravity storage formats and full acceptance remain open. 2026-09-29 native trial (partial): CLI 1.2.13 created and resumed a synthetic conversation from 3 to 11 messages with one file tool call/result. Corrected accounting matches native CLI totals; model selections stay separate from response models. Stable identity, repeat import, seven raw hashes and viewer reload passed. Additional Sonnet and Gemini trials confirm the response/reasoning split. Positive native cache writes, full metadata and attachment coverage remain open. 2026-10-08 native trial (partial): CLI 1.2.16 inspection found six retained conversations. The preferred route preserves 114 messages, 41 calls and 39 results; a database-only trial preserves 112 messages, 41 calls and 40 results. Repeat imports, raw hashes, all six log replays, search, viewer payloads and Stats pass. One completed database result remains raw-only when a log takes precedence. Native automatic capture, rendered UI, database replay and full artifact coverage remain open. |
| Antigravity IDE | antigravity-ide |
local-store | Antigravity IDE transcripts, lifecycle notices, task artifacts, and conversation-database usage from ~/.gemini/antigravity-ide; stale migration copies are skipped. No IDE application or store was available for native acceptance. Synthetic fixtures cover database tools, stale migration copies and log replay. Database/recovery replay, full artifacts and automatic/installed acceptance remain open. | fixture-tested; 6 open gates. 2026-10-08 native trial (blocked): The inspected native IDE home and application are absent. Registered-profile discovery found zero candidates. Synthetic database-event, migration-identity, and offline-log-replay tests pass, but no native IDE conversation was available. |
| Gemini | gemini-web |
manual-export, browser | Gemini browser capture and offline replay plus JSON and English HTML Takeout activity imports. Available export media enters attachment storage. Experimental Canvas parsing keeps document and code entries when the selected history candidate contains them. Native JSON preserves 109 media files. Native English HTML preserves all 80 prompts, available response text, and 52 linked uploads, but omits references to 57 generated files retained in the raw ZIP. Models, usage, complete history, browser/export equivalence, and export raw replay remain unverified or unavailable. | fixture-tested; 5 open gates. 2026-10-08 native trial (partial): Native JSON Takeout: 82 activity records, 80 prompts, 37 conversations, 151 messages, and 109 checksum-matched attachments. Installed revision 731cc154 passes repeat and fidelity checks. Native English HTML exposed caption-link and pre-clock subtitle framing defects. Parser 0.6.2.2 preserves all 80 exact prompts, available response text, and second-precision clocks in 37 conversations and 111 messages. All 52 linked upload hashes match, repeat skips all 37, and the shared raw ZIP hash matches. The HTML omits references to 57 generated files present in that ZIP. Installed revision 079919d2 and bundled Node 24.20.0 repeat every HTML check in a fresh temporary archive at 21:32 UTC. |
| Gemini CLI | gemini-cli |
local-store | Historical official saved JSON/JSONL chats, prompt logs, manual/tool checkpoints and linked children under ~/.gemini/tmp and ~/.cache/.gemini/tmp, with per-root project registries and explicit copy overrides. Historical import only. No executable installation, probing or live integration. Shadow Git stores, shell history, arbitrary exports and inaccessible container stores are excluded. Pruned provider history remains unavailable. | fixture-tested; 4 open gates. 2026-10-09 native trial (blocked): No local data at ~/.gemini/tmp or ~/.cache/.gemini/tmp, their project registries, or shadow-history roots. Historical source audit covers v0.63.0, including Seatbelt storage and collision filenames. Native validation requires an existing historical store. Gemini executable installation and probing are excluded. |
| Gemini Notebook (formerly NotebookLM) (partial) | notebooklm |
manual-export, browser | Takeout ZIP/folder import through CLI and app picker, plus extension capture of visible settled notebook chat and expanded Thoughts. Open Studio notes can enter the archive and Library through experimental browser capture. Native chat continuation reaches the production archive and installed viewer. Both routes support offline raw replay. Browser history is visible-only and requires an English account control. Open Studio note capture remains unverified on a live notebook. Unopened notes, audio overviews, mind maps, original media, tool telemetry, and citation-source mapping are not captured. Takeout resources remain raw-only. Message clocks, models, usage, and account-wide completeness remain unknown. Exact-prefix retention rejects shorter or rewritten histories. Browser and export identities remain separate. | fixture-tested; 5 open gates. 2026-09-29 native trial (partial): One native Takeout prompt/reply pair, five hash-verified members, unchanged repeat import and raw rebuild after removal of the import copy. Browser viewer displayed the transcript, separate provider and unknown time. Metadata field accounting covers the bounded notebook/text-source cohort; multi-turn and full metadata display remain open. 2026-10-08 native trial (partial): Native Takeout: one pair, five raw members, unchanged repeat, identical replay, search, and expandable viewer metadata. Three records cover 15 field paths. Installed revision 731cc154 repeats those import checks; six Google modules in 10cf96ee retain identical bytes. Automatic extension intake grows from one to three pairs, with the approved prompt exactly once. Installed isolated import retains the earlier pair and five Thoughts, one source ID, and seven citation markers; repeat, raw replay, search, and Stats pass. Stats counts one chat, six messages, three user messages, and estimated tokens without dated activity or model attribution. Native reload retains all three pairs. The production six-message transcript matches the isolated transcript byte for byte; both retained raw objects and projections pass hash checks. On installed 10cf96ee, the latest repeated capture completes at 20:33 UTC. All 26 receiver jobs are archived under one identity, none queued or failed; the six-message transcript and retained raw hashes still pass. The installed macOS viewer renders all three pairs and expands Thoughts and capture metadata. Source commit e126a694 fixes unknown-time search labels and passes 81 focused viewer tests; installed acceptance remains pending. Takeout continuation and artifact/media gates remain open. 2026-10-08 native trial (partial): Installed revision 8a143c3b passes native continuation, repeat import, exact content and retained Thoughts, raw hashes, offline replay, and search under bundled Node 24.20.0. Nine relevant bundled files match the commit. After the confirmed extension reload, a fresh three-pair capture archives at 22:49 UTC. All 54 NotebookLM jobs are archived under one identity. The six-message transcript remains byte-identical, and both raw objects and projections pass hash checks. Takeout download and rendered search, Stats, and app-picker acceptance remain pending. 2026-10-09 native trial (partial): Clean 8a143c3b snapshot rendering in Chromium and WebKit passes 18 assertions, fails four Stats assertions, and leaves two app-picker assertions unrun. Search shows Time unknown without an epoch date or tooltip. Ordered messages, Thoughts, expandable capture metadata, and coverage disclosures pass. Unknown-time Stats headlines and favorite provider are missing. Source 0b33e2de48d1 corrects all-time totals and estimate labels with 45 passing focused tests. Nine inspected Google files match installed ca2af9bd; Google importer bytes remain unchanged. The Stats correction is not installed. The client ABBA is inconclusive under unstable load, with greater-than-tenfold variation between bases. Installed rendering, stable timing, and export download remain pending. 2026-10-09 native trial (partial): After the confirmed ca2af9bd extension reload, two fresh NotebookLM captures finish archival by 00:33:52 UTC. At 00:35 UTC, all 72 recorded jobs are archived, with none queued or failed. The signed-in native page retains all three ordered pairs. The production identity and six-message transcript remain unchanged, and both current raw objects and projections pass SHA-256. No new prompt was submitted. The Takeout download still reports ERR_BLOCKED_BY_CLIENT in Arc. Stats follow-up 65e9e6039857 passes six headline and seven existing rendering tests. A stable serialized client comparison against ced1274f shows no material regression, with all-time headline medians improving from 11.56 ms to 0.84 ms. The clean full check passes at 2b68a810d921, including the follow-up. Installed rendering remains pending. 2026-10-09 native trial (partial): Five inspected Google files in installed f6c2b2d0 match the commit. Its Stats functions process the retained native payload under bundled Node 24.20.0 and generate one conversation, six messages, three user messages, and 244 estimated tokens. Generated markup labels all tokens estimated and identifies NotebookLM as favorite provider. Model, calendar, and spend tiles remain unavailable, and the payload is unchanged. This is installed function and generated-markup verification, not browser rendering. The clean shared full check passes at 415fdd26e13a, including f6c2b2d0. Fresh extension trials await reload confirmation. Installed rendering and the export-dependent gates remain open. 2026-10-09 native trial (partial): Installed f6c2b2d0 passes twelve rendered assertions in sequential Chromium 153 and WebKit 26.6 trials. All 1339 tracked runtime files match the commit. The isolated cohort shows one conversation, six messages, three user messages, and 244 tokens labelled All tokens estimated. NotebookLM is the favorite provider; model, dated activity, and spend remain unavailable. Search and header show Time unknown without an epoch date or search date tooltip. Each engine imports one native Takeout conversation into a fresh temporary archive, then skips one on repeat, with numeric counts and Complete summaries. Different explicit account labels exactly explain the earlier identity difference. Four screenshots pass visual review. A separate fallback-month caption defect is corrected by e97ea1502a5d with eight passing rendering tests; installed caption acceptance remains pending. The shared full check at f4dabac95eaf covers installed 738e0ece but precedes that correction. Export-dependent and broader fidelity gates remain open. 2026-10-09 native trial (partial): Installed 313e215f modified, including e97ea150 with three unrelated commits reverted, passes four caption assertions per engine in Chromium 153 and WebKit 26.6, eight total. All 1364 runtime manifest entries match the bundle, and its full check passes. Retained native captures replay into an isolated archive. Both Models captions read The models you leaned on, All time. The cohort retains one conversation, six messages, three user messages, and 244 estimated tokens. NotebookLM remains the favorite provider; model, dated activity, and spend headlines stay unavailable. Finite and custom labels pass. Include Chat changes the headline while preserving empty secondary heatmap hosts and Models ranking. Eight screenshots pass visual review and are deleted with hashes retained. The caption gate is closed. Takeout access and the five export-dependent or broader fidelity gates remain open. |
| Jules | jules |
cloud | Jules API sessions and activities preserve prompts, replies, plans, tool activity, generated patches, repository context, and status. The API does not expose inference model or token usage. Only an explicit recorded model hint can supply attribution. Unordered and empty continuation pages are followed within limits. Truncated activity lists cannot publish a successful checkpoint. Native cloud validation and offline raw replay remain open. | fixture-tested; 4 open gates. 2026-10-09 native trial (blocked): No local data exists at candidate Jules stores; official history is cloud-based. Native validation requires authorized API access to existing Jules history, and CLI installation alone cannot provide it. Rolling v1alpha documentation was audited on this date. Synthetic pagination and incomplete-publication regressions pass. Offline raw replay remains unsupported. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| gptme CLI / local store | gptme |
local-store | Main-branch JSONL, recorded reply model/effort, reasoning, usage and JSON tools; exact branch/view/recovery files plus a declared chat-settings projection. Originating interface is not recorded. Config credentials/MCP/prompt settings are excluded. Naive clocks remain timezone-unknown. Alternate histories, attachment references and markdown/XML tool syntax are raw-only. Child linkage, hosted billing and other hosts remain unverified. Log-root moves create a separate path-based identity and retain the original archive. | fixture-tested; 7 open gates. 2026-09-29 native trial (partial): Official gptme 0.34.0 against a local synthetic endpoint: five initial and ten resumed records, a real file read, native undo/view/checkpoint captures, unchanged repeated imports and raw replay. 223 independent value/hash/round-trip checks passed; all 111 observed field paths accounted for. Hosted billing and other hosts remain unverified. 2026-10-09 native trial (blocked): No local data: ~/.local/share/gptme/logs is absent and ~/Library/Application Support/gptme/logs is empty. Upstream v0.34.0 serialization remains compatible. A synthetic regression verifies path-based identity after moving the root and retention after removal. Installing and creating a conversation, or supplying a store, would permit a native trial. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| iFlow | iflow |
local-store | iFlow CLI transcripts from ~/.iflow/projects, including tool calls and compaction records. The audited 0.5.19 journal preserves transcripts, tools, compaction, and raw replay. Interactive logs omit measured usage. Denied directories and rejected sources report errors. Hosted service shutdown is announced; current native acceptance requires retained local history. | fixture-tested; 7 open gates. 2026-10-09 native trial (blocked): No local data. No iFlow binary or ~/.iflow/projects store exists, and documented overrides are unset. Latest 0.5.19 matches the earlier artifact. Installation does not restore history. The announced April 17, 2026 service shutdown means historical logs are the available validation route. |
JetBrains
Section titled “JetBrains”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Junie | junie |
local-store | Local journals and IDE task chains preserve child agents, tools, request usage, estimated cost, context-window reports, mode, and authorization availability. The public 3659.7 release verifies runtime field schemas. Authorization availability does not establish tool approval, and repeated cost snapshots do not add request charges. Older CLI arguments can be lost after compaction. Installed runtime acceptance remains unverified. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): Read-only discovery found instance/log residue but no retained CLI or IDE conversations. Remote mode shares a running CLI session; no native session is available for acceptance. |
Kilo Code
Section titled “Kilo Code”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Kilo Code | kilo-code |
local-store | Shared CLI, VS Code, and JetBrains SQLite sessions, including tools, reasoning, attachments, and usage. Shared CLI and current IDE SQLite history. Interface attribution is absent; recorded costs remain estimates when their provenance is ambiguous. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): Current Kilo data contains metadata/log residue without a history database. Read-only discovery found no conversation candidates; native round-trip and resume remain blocked. |
| Kilo Code Cloud | kilo-code-cloud |
cloud | Kilo synchronized session exports and legacy transcript blobs provide retained conversations; targeted Cloud Agent requests also import final results. Cloud Agent result lookups require session and message IDs and return status/final text only, without the original prompt, tools, model, or tokens. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): The cloud UI redirects to sign-in and no API key is configured. Synthetic synchronized-export tools, arguments, outputs, raw preservation, and repeat import pass; authenticated acquisition remains blocked. |
| Kilo Code legacy tasks | kilo-code-legacy |
local-store | Earlier IDE and CLI JSON task histories and legacy task indexes. Earlier JSON tasks, including tools and human prompts. Missing model IDs and legacy clocks remain explicit gaps. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No retained legacy Kilo IDE or CLI task history was found in inspected roots. Native legacy field fidelity and migration identity remain unverified. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Meta AI consumer assistant (partial) | meta-ai |
browser | Experimental profile-bound browser capture follows separate pinned and unpinned inventories and backward message pagination. Durable snapshots preserve branches, structured responses, recorded reasoning and GraphQL envelope metadata. Every page is verified again before publication. Native account history and automatic capture remain unverified. Original files and media, separate device and messaging histories, account instructions, memory and official export import remain incomplete. Tool details, citations and nested responses remain raw. Muse Personal Agent and Muse Code are separate products. | fixture-tested; 9 open gates. 2026-10-08 native trial (blocked): Read-only Arc inspection displayed signed-out Meta AI. Public rolling client assets supplied persisted identity, history, message and branch contracts. No authenticated GraphQL request, terms acceptance, prompt or export request occurred. The sign-in gate does not establish provider-side impossibility. |
| Muse Code | muse-code |
local-store | Muse Code journals retain linked nested runs, recorded request usage, response ids, completion timestamps, and finish reasons. Preserved model-call records exclude inputs. Muse Code 1.4.14 tool approvals become runtime approvals, and a journal whose start Muse pruned at a context checkpoint is joined to the history already archived. Response-creation events do not establish request start or first-token latency. Native journals were observed from 1.0.3 and 1.4.14 only, and no checkpoint-pruned journal was observed. External hooks remain unverified. Muse Personal Agent is a separate product. | fixture-tested; 4 open gates. 2026-11-09 native trial (partial): Four Muse Code 1.4.14-R7309.1 journals were read without writing to the provider store and imported into a scratch archive: a live session with 101 messages, 47 tool calls, and 5 approvals, its two reminder subagents, and a session that ended during startup with no prompt. No journal with a checkpoint-pruned prefix was available, so the join of archived history rests on synthetic fixtures. 2026-10-09 native trial (blocked): No local data at the documented XDG muse/sessions store or configured override. The muse command is absent from PATH, which is not an exhaustive installation inventory. Stable 1.4.4-R5419.1 and current SDK/cookbook source contracts were audited. Installation plus an existing or authorized new journal would enable a native trial. Earlier 1.0.3 native evidence does not verify 1.4.4. |
| Muse (Personal Agent) (partial) | muse-agent |
manual-export, browser | Explicit exports preserve turns and raw members. Structured native evidence adds side chats, tool runtime, and exact-session approval joins. Usage footers retain context and response estimates without asserting billed spend. Model-call inputs and approval payloads are excluded. Installed Muse 7.0 verifies field names, not populated structured acquisition. Signed-in capture remains unverified and unpublished. Goals, activity, schedules, task runs, unjoined approvals, and unknown lineage remain raw-only. Complete history, nonempty resources, model attribution, billed usage, and native memory capture remain unverified. | fixture-tested; 6 open gates. 2026-10-03 native trial (partial): Muse 6.0 on macOS: native data export preserved nineteen turns and two byte-verified raw members. Repeat import, raw rebuild, bounded manifest field audit, transcript disclosure and Stats passed. Nonempty resources, background work, side chats and account completeness remain unverified. 2026-10-06 native trial (partial): A native web download appended twenty-eight turns and two workspace files to the earlier timeline. Exact message fields, member bytes, repeat-import identity, source-free raw rebuild and manifest field accounting passed. Chromium fixtures verified automatic page-change capture and restart delivery of a 21 MiB export. Signed-in automatic acquisition remains unverified. 2026-10-08 native trial (partial): Muse 6.0 build 1082791187 read-only Mac inventory found no complete local transcript store. The bundled cache retains forty messages for three days. Signed-in web inspection exposed newer messages, goals, activity and memory. The side-chat list and Library directory were empty. Memory View source and Download include editor text outside the file. No Muse receiver job or new archived turn was observed. The installed parser lacks the subsequent workspace validation fix. Separate feature acquisition and live automatic acceptance remain open. 2026-10-08 native trial (partial): Read-only web inspection found empty active and archived side-chat, approval and completed-goal inventories. Two recurring tasks were visible; one exposed four successful scheduled runs. A tracked goal exposed six activity entries and a cron association, while earlier one-off task links did not open details. Association state, goal state and run status are separate facts. Structured acquisition and history completeness remain unverified. No goal, task, permission or approval was changed. 2026-10-08 native trial (partial): The installed 0.6.2 receiver now includes workspace validation. One authorized synthetic side chat and its resumed turn preserved the native identity and exact short replies. The main chat remained connected, but the earlier collector reported no ready history connection and no Muse job arrived. A newer shared extension build includes connection discovery and structured acquisition changes. Its reload and native capture remain pending. No additional prompt, approval or task mutation was sent. 2026-10-08 native trial (blocked): Read-only extension IndexedDB confirmed the rebuilt collector and the diagnostic history connection not ready. The probe reached an authenticated main-chat page and an RPC context, but no Muse receiver job arrived. Empty currentPage is also produced by Sync now and does not establish a missing tab. Browser control and native accessibility timed out under severe host contention. The structured receiver update is committed with twenty-five focused tests passing, but installed and live acceptance remain pending. 2026-10-08 native trial (partial): Installed bundles 10cf96ee and 079919d2 include structured parser 0.6.2.2. Seven focused synthetic trials passed under each bundled runtime, including side-chat resume and independent raw replay. All six Muse, dispatch and version files match 079919d2. An installed two-session synthetic archive preserves resume identity and repeat skipping, and indexed search finds the resumed marker only in its side chat. The required isolated project check passed. Native UI inspection still shows Muse Connected. A 21:28 UTC production read found no Muse jobs and the unchanged nineteen-message timeline. Rendered checks, the confirmed extension reload, fresh connection diagnostic and live acquisition remain pending. 2026-10-08 native trial (partial): Installed synthetic consumer checks pass for source filtering, main and side-chat identities, resumed text and estimated-usage Stats. Installed 8a143c3b parser 0.6.2.3 passes eight focused tests, including complete multiline tool-detail output, separate operation times, repeat capture and raw replay. Four serial base/candidate/candidate/base throughput trials measured 8.8, 16.4, 10.7 and 17.2 sessions per second. Both revisions missed the absolute floor once, recording an environment limit; every non-throughput gate passed. The confirmed extension reload reached a ready connection but rejected its secure WebSocket metadata. The collector correction passes thirteen focused tests and awaits deployment. A 21:44 UTC production read found zero Muse jobs among 3632 receiver jobs and the unchanged nineteen-message timeline without either approved marker. Multiline rendering and live capture remain unverified. 2026-10-08 native trial (partial): A verified 8a143c3b installed-runtime snapshot passes fourteen rendered assertions across Chromium and WebKit for the existing nineteen-message timeline. Response disclosure, search and matching provider-filtered Stats pass. Session details labels estimated usage, while the Stats Tokens tile omits that label. Neither approved capture marker is present. The native System files inventory exposes eight Markdown files. The Soul editor download includes outside-file explanation text and exceeds its listed byte count, so it cannot establish original-byte fidelity. Synthetic multiline rendering and automatic capture remain pending. 2026-10-08 native trial (partial): The installed 8a143c3b synthetic rendering trial passes fourteen of sixteen assertions across Chromium and WebKit. Separate identities, resumed markers, exact 32-line tool output, collapse/reopen, search and one-second duration pass. Stats fails in both browsers: estimated tokens lack a label, and the Models table shows zero spend for unpriced usage. Search displays the account ID instead of its name. Corrected Stats rendering and live automatic capture remain pending. 2026-10-09 native trial (partial): After the confirmed ca2af9bd reload, a native capture checkpoint holds a complete 28135-byte export with forty-one turns, nine members and seven workspace files. The installed parser passes 410 native-value comparisons and all member digests. Both approved replies occur once in the export and the same native side chat. Eight structured responses include two active sessions, two schedules, one hundred task runs and one tracked goal. Acquisition stops at activity.list; a structural guard correction passes fifteen focused tests but awaits deployment. Export main-channel labels include side-chat turns and cannot establish conversation boundaries. Empty voice inventory fields remain raw. Terminal receiver publication, native side-chat identity, memory bytes and replay remain open. 2026-10-09 native trial (partial): Installed f6c2b2d0 synthetic rendering passes all twelve assertions across Chromium and WebKit. The same two conversations show 238 tokens with All tokens estimated, an em dash for unpriced model spend, the Unpriced token usage tooltip and estimated Session details. Both browsers preserve all 32 tool-output lines after collapse and reopen, fetch details once and display a one-second duration. The estimated-usage and unpriced-spend gate is closed. This does not establish native automatic capture. Installed 738e0ece retains unchanged Muse receiver modules; its extension reload and terminal native capture remain pending. 2026-10-16 native trial (blocked): Muse 7.0 build 1086317620 desktop inventory copied six SQLite cohorts before opening them; all passed integrity checks. Name-authority records contain session identifiers and naming operations, but no messages. URL-cache and analytics queues are empty, and other stores hold routing, resource statistics and preferences. No transcript IndexedDB store was found; the optional bundled cache still retains forty settled messages for three days. Official documentation identifies the cloud VM as the system of record. Desktop acquisition requires an unimplemented authenticated remote route. Credential inspection, credential reuse and private server calls await authorization. The native app reports failed sign-in. All 26 existing Muse parser, binding and replay tests pass in isolated homes, but desktop history, three-route identity, watcher capture and installed acceptance remain unverified. No app rebuild occurred. |
MiniMax
Section titled “MiniMax”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| MiniMax Code cloud | minimax-code-cloud |
cloud | Explicit MiniMax Code cloud session imports paginate retained messages and preserve full response bodies with stable identity. Supplied display tool calls, arguments, results, thinking, and usage are normalized. Returned message usage and context snapshots retain reported costs and the verified archived-state snapshot. Account-wide enumeration is not implemented. A session ID is required, and nonadvancing or partial pagination must not replace an archive. The official reader projects a smaller text schema; native availability of richer tool records remains unverified. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): The Code web app shows Sign in; no cloud token or selected native session is available. Supplied display tools, arguments, results, thinking, usage, and raw bodies pass synthetic tests only. |
| MiniMax Code local history | minimax-code |
local-store | Canonical transcripts, verified compaction snapshots, older ledgers and SQLite history, tools, and durable request usage. Durable per-request tokens and reported costs, context snapshots, and local lifecycle metadata retain their source provenance. CLI and desktop share a session identity. Canonical history and verified compaction snapshots take priority over mirrored stores; durable usage supersedes message counters. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): Read-only discovery found no retained MiniMax Code local history in inspected roots. Native CLI/desktop fields, compaction, resume, and consumer acceptance remain blocked. |
Mistral
Section titled “Mistral”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Mistral Vibe for work (formerly Le Chat) (partial) | mistral-chat |
browser | Experimental browser capture binds merged Chat and Work history to user, organization and workspace. Private, archived, shared and pinned inventories include project chats. Durable pages preserve message versions, canvas and context-linked research. Complete snapshots recheck every page before publication. Native account and history fidelity remain unverified. Original bytes, separate workflow streams, archived schedule inventories, project resources, Libraries, Knowledge Base and Skills remain incomplete. Official export import is unimplemented. Vibe CLI and remote coding products are separate. | fixture-tested; 9 open gates. 2026-10-08 native trial (blocked): Read-only Arc inspection displayed Vibe Terms of Service and sign-in controls. No terms were accepted, authenticated history read or export requested. Public rolling client assets supplied query and message contracts. This access gate does not establish provider-side impossibility. |
| Mistral Vibe CLI (partial) | vibe |
local-store | Legacy v1/v2 and Vibe 2.26.0 default Unified Harness v1 minor-8 history, tools and failures, cancelled approvals, reasoning, session tokens, configuration, lineage, inline and session-owned images, verified raw replay, and retained-history protection. Native CLI and ACP trials pass import, resume, repeat, field audit, image hashes, usage, search, consumer APIs, and offline replay. Parser 0.6.2.3 preserves repeated physical records, and production watchers capture both engines automatically. Stored Unified approvals, tools, hook notices, and compaction checkpoints project into runtime evidence. Saved context snapshots and legacy window thresholds retain their provenance. Legacy clocks retain session order. Local model configuration does not establish response models or per-model usage. Legacy costs use current rates; Unified billed cost is absent. Arbitrary external files remain unavailable. Native branches, compaction, delegation, voice, nonempty recovery journals, engine migration, and hosted billing remain open. Other Unified store minors fail visibly. | fixture-tested; 2 open gates. 2026-10-08 native trial (blocked): The initial discovery found no default native store. Later authorized temporary-store execution superseded this access boundary; see the subsequent trial. 2026-10-08 native trial (partial): Official Vibe 2.26.0 ran legacy and Unified CLI/ACP trials against a synthetic loopback endpoint without paid calls. Native import, resume, repeat zero writes, exact offline replay, field accounting, seven tool outcomes including two errors and one cancelled approval, two images per engine, configuration-only model scope, usage, search, viewer APIs, and Stats APIs pass. Parser 0.6.2.3 preserves all 28 legacy and 27 Unified normalized messages. Production source watchers and supervisor workers automatically captured a new two-message continuation for each engine without changing earlier messages. Both sessions have 260 recorded tokens. CLI, ACP, continuation, and replay retain each native identity. The initial installed 10cf96ee trial passed the Unified round trip but merged one repeated legacy prompt. Subsequent trials below close rendered and installed acceptance after correction 234cc6dfc855. Native alternate history and engine migration remain open. 2026-10-08 native trial (partial): Headless Chromium and WebKit pass both engines with exact rendered message counts, two available images, two failed tools, expanded cancelled-approval details, seven tools, 260 tokens, configuration-only models, and correct clock provenance. No page or HTTP errors occurred. Canonical events v16 and tool usage v14 correct Unified edit categories and rebuild outcome totals from retained transcripts. Native APIs and both browsers confirm three edit calls in the Stats total and category rows for each engine. Both Unified search_replace rows say Edited, and the failed edit retains its Error badge. 2026-10-08 native trial (partial): The installed 079919d2 and 8a143c3b bundles and their bundled Node pass both native round trips in fresh temporary archives. Legacy preserves all 28 messages and three raw components. Unified preserves all 27 messages and eleven raw components. Both retain resume prefixes and native identity, repeat with zero writes, replay exactly with provider roots unavailable, preserve both image hashes, record seven tool outcomes and 260 tokens, return exact-session search hits, and report three edit calls. Eleven relevant bundled files match each commit byte for byte. Native alternate history, engine migration, and the documented resource and billing limits remain open. |
Moonshot
Section titled “Moonshot”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Kimi CLI | kimi-cli |
local-store | Kimi CLI wire journals, context history, and subagent runs from ~/.kimi/sessions. Wire approvals, hook records, retries, and each context update retain native timestamps and identities. Historical context and wire journals preserve reasoning, tools, usage, and child sessions. Main-session model identity can be absent. Latest 1.52.0 redirects to the Kimi Code installer. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no ~/.kimi/sessions, ~/.config/kimi/sessions, or ~/.local/share/kimi/sessions, and no KIMI_SHARE_DIR override. Audited 1.52.0 with unchanged wire types. Its installer redirects to Kimi Code; retained CLI history or a historical executable is needed. |
| Kimi Code | kimi-code |
local-store | Kimi Code wire journals from ~/.kimi-code/sessions, including per-agent runs and linked child sessions. Stored approvals, retries, compactions, and context measurements or estimates use the runtime evidence contract. Wire journals preserve reasoning, tools, directional usage, and linked agent runs. Scheduling, mirrored child lifecycle, and file-history references stay raw-only. Dedicated durable hook invocation records remain unverified. External checkpoint bytes are not acquired. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no ~/.kimi-code/sessions and no KIMI_CODE_HOME override. Audited 2.1.1 protocol 1.5. Twenty-four added durable record types now have explicit raw-only accounting. Installation plus separately authorized activity could create a trial store. |
| Kimi consumer chat and agent tasks (partial) | kimi-chat |
browser | Experimental browser capture enumerates global, pinned and project chats. It follows message and external tool pages with stable snapshots, normalizes selected ancestry, text, reasoning, tool arguments and results, and preserves raw evidence for offline replay. Native access requires sign-in. Original files, generated artifact revisions, project resources, independent tasks and hidden branches remain uncollected or unverified. Export availability and import remain pending. The current model selector is not historical model evidence. Kimi Code and Kimi CLI are separate products. | fixture-tested; 8 open gates. 2026-10-08 native trial (blocked): Arc displayed Log in to sync chat history. The interface exposed project, plugin, scheduled and artifact modes. No authenticated history or export was acquired. This access gate does not establish provider-side impossibility. |
Nous Research
Section titled “Nous Research”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Hermes | hermes |
local-store | SQLite transcripts, compression continuations, tools, main and auxiliary model usage, recorded billing provenance, and stored tool outcome metadata. Schemas 1 through 32 retain actual, estimated, included, and unknown billing states. Stored terminal results expose approval decisions only when structured outcome fields exist. API retry/error hooks and context composition are not persisted in the state database. Raw replay and native acceptance remain unverified. | fixture-tested; 7 open gates. 2026-10-09 native trial (blocked): No local data. Hermes is installed, but its copied schema-30 ~/.hermes/state.db has zero sessions and messages. No fallback logs or named profiles exist. Installing again would not supply history. Existing history or separately authorized conversation creation is required. |
OpenAI
Section titled “OpenAI”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| ChatGPT | chatgpt |
manual-export, browser | ChatGPT account exports and browser capture preserve conversation branches, files, recorded models, account identity, and provider metadata. Workspace plan and export permissions affect availability. Current settings cannot establish historical settings or measured usage. Native project attribution and offline replay pass in source and installed isolated trials. Installed Chromium and WebKit verify project attribution, eight messages, and four working image controls. GPT identity, voice fields, and collection recovery remain pending; Codex cloud is separate. | fixture-tested; 7 open gates. 2026-10-08 native trial (partial): The installed extension reported 63 collection tasks, zero waiting deliveries, and HTTP 429 or unfinished-generation errors before the update. Native project membership is visible, but the matching export record has no project metadata. Updated-extension recovery and project, GPT, and voice field comparisons remain pending. 2026-10-08 native trial (partial): A read-only inventory of 3,038 archived records and a separate 151-capture raw scan found no nonempty values at the inspected project, GPT identity, or voice metadata paths. Raw hashes and identities passed. The 32 MiB raw scan stopped at its byte limit. These bounded samples do not establish account-wide absence. Native GPT and Voice conversation samples remain pending. 2026-10-08 native trial (partial): Updated-extension activation is confirmed by new receiver jobs and later tab-targeting fields in IndexedDB. Native sidebar and project-list responses verify a captured snorlax project identity. The correction passes native import, repeat import, and offline replay in source and installed 079919d2 trials, preserving all eight messages and four hash-verified images totaling 2,935,830 bytes. Interleaved timing is inconclusive under host load. Rendered project attribution, GPT identity, and Voice samples remain pending. 2026-10-09 native trial (partial): Installed ca2af9bd renders the recorded project identity consistently in Chromium and WebKit. All eight messages remain accessible, including the paired tool call and result. Four image controls decode and open their lightboxes. No project name was captured, so the native identifier remains the displayed value. This bounded project trial does not establish GPT identity, Voice fields, or complete account history. |
| Codex CLI | codex-cli |
local-store | Retained CLI rollouts and their recorded child sessions from Codex state and JSONL history. Installed b69492f2 copied-file repeat, raw replay, source-removal preservation, search, bounded viewer payloads, and Stats passed. Full supervisor capture through a settled native turn remains unverified. Ephemeral runs, complete paginated histories, account-wide resources, and external attachment bytes are not established by local rollout support. Three large CLI samples retain explicit scratch or heap limits. | fixture-tested; 3 open gates. 2026-10-09 native trial (partial): 25 of 27 selected CLI/child rollouts passed import, byte-identical repeat, verified raw hashes, offline replay, source removal, search, viewer payloads, and Stats evaluation. Eight copied native sessions passed synthetic continuation and repeat. A 134,964,918-byte rollout passed import/repeat/raw but replay reached the scratch safety threshold. A 160,313,502-byte rollout exhausted a 640 MiB JS heap during derived serialization. Shared field accounting covers all 34 selected rollouts, 109 SQLite columns, and 81 title-index records. A 90-second native watcher callback trial captured two CLI updates into an isolated archive; the supervisor queue and settled final turn remain unverified. 2026-10-09 native trial (partial): Installed b69492f2: 24 of the same 27 CLI/child rollouts passed import, byte-identical repeat, verified raw hashes, offline replay, source removal, search, bounded viewer payloads, and Stats. Seven native copies passed synthetic continuation and repeat. One 160,313,502-byte source exhausted the 640 MiB JS heap. Two sources of 134,964,918 and 112,448,249 bytes passed import/repeat/raw but replay reached the 750 MB scratch threshold. All 1,384 installed runtime files matched the manifest before and after, including 21 Codex-named modules. Bundled Node v24.20.0 ran with isolated archives and archive.derivationCache=false. All 19 installed inspector and robustness tests passed. Full supervisor capture and large-session acceptance remain partial. |
| Codex cloud | codex-cloud |
cloud, browser | Saved ChatGPT sign-in reads retained task turns and events. The Codex cloud browser collector reads current and archived WHAM tasks, selected-turn events, raw alternate attempts, and task diffs. The ChatGPT switch separately captures visible Dot task threads through durable turn and item pages, preserving native parent and room-root identities. CLI fallback preserves task status and the unified diff. Native browser task acquisition remains unverified. Browser captures use workspace-scoped identities and remain separate from older CLI/API records without matching account evidence. WHAM collection is bounded to 500 turns per task. Dot task collection requires the optional cloud thread host and stays within 10,000 pages and 32 MiB. File downloads are not exposed by these routes. CLI fallback lacks full conversation history. | fixture-tested; 6 open gates. 2026-09-29 native trial (blocked): Seven authenticated profiles returned empty active and archived task lists; one required sign-in. No native task was available for an archive trial. 2026-10-08 native trial (blocked): Saved-sign-in reads returned zero current tasks and zero archived tasks, one page each without continuation or truncation. Native Codex cloud navigation entered the unified ChatGPT cloud composer. No native WHAM task was available for the new browser collector. 2026-10-08 native trial (blocked): The signed-in unified cloud composer has no published environment. The authorized repository-preserving native trial still needs a disposable repository-backed environment. No task has been submitted. Existing empty task lists do not verify browser fidelity. |
| Codex Desktop | codex-desktop |
local-store | Retained Desktop and IDE-origin rollouts under the shared vscode source, with independent product attribution. Installed b69492f2 copied-file repeat, raw replay, source-removal preservation, search, bounded viewer payloads, and Stats passed. Full supervisor capture through a settled native turn remains unverified. Ephemeral runs, complete paginated histories, account-wide resources, and external attachment bytes are not established by local rollout support. Project and section tables remain raw without a complete normalized organization model. | fixture-tested; 2 open gates. 2026-10-09 native trial (partial): Four selected native rollouts passed import, byte-identical repeat, verified raw hashes, offline replay, source-removal preservation, search, bounded viewer payloads, and Stats. Two copied sessions passed synthetic continuation and repeat. Shared field accounting covers 178,778 rollout records, 109 SQLite columns, and 81 title-index records. No eligible Desktop activity was captured during the bounded passive watcher trial. Full supervisor capture remained unverified; installed acceptance was outside this source-tree trial. 2026-10-09 native trial (partial): Installed b69492f2: all four rollouts from the same cohort passed import, byte-identical repeat, verified raw hashes, offline replay, source removal, search, bounded viewer payloads, and Stats. Two native copies passed synthetic continuation and repeat. All 1,384 installed runtime files matched the manifest before and after, including 21 Codex-named modules. Bundled Node v24.20.0 ran with isolated archives and archive.derivationCache=false. All 19 installed inspector and robustness tests passed, including partial-tail completion for this product. Full supervisor capture and cross-route identity remain unverified. |
| Codex SDK jobs | codex-sdk |
local-store | Retained exec and SDK-style rollouts; delegated runs and one-shot jobs retain separate Stats classification. Installed b69492f2 copied-file repeat, raw replay, source-removal preservation, search, bounded viewer payloads, and Stats passed. Full supervisor capture through a settled native turn remains unverified. Ephemeral runs, complete paginated histories, account-wide resources, and external attachment bytes are not established by local rollout support. Exec evidence does not establish all Python app-server SDK routes. | fixture-tested; 2 open gates. 2026-10-09 native trial (partial): Three retained exec rollouts passed import, byte-identical repeat, verified raw hashes, offline replay, source-removal preservation, search, bounded viewer payloads, Stats, synthetic continuation, and repeat-after-growth. Shared field accounting covers 178,778 rollout records, 109 SQLite columns, and 81 title-index records. No eligible exec activity was captured during the bounded passive watcher trial. Python app-server SDK attribution and full supervisor capture remained unverified; installed acceptance was outside this source-tree trial. 2026-10-09 native trial (partial): Installed b69492f2: all three exec rollouts from the same cohort passed import, byte-identical repeat, verified raw hashes, offline replay, source removal, search, bounded viewer payloads, Stats, synthetic continuation, and repeat-after-growth. All 1,384 installed runtime files matched the manifest before and after, including 21 Codex-named modules. Bundled Node v24.20.0 ran with isolated archives and archive.derivationCache=false. All 19 installed inspector and robustness tests passed, including partial-tail completion for this product. Python app-server SDK attribution, full supervisor capture, and cross-route identity remain unverified. |
| ChatGPT Work Local (partial) | codex-desktop |
local-store | The existing Codex rollout importer captures Work Local and its recorded originator assigns a separate ChatGPT Work Local harness in history, search, navigation, and Stats. Native retained history preserves messages, tools, usage, model metadata, and screenshot bytes. Received native history verifies child attribution. Installed Chromium and WebKit verify the product label, harness version, all 54 messages, and eight working image controls. Canonical events v17 preserve native call identities and produce identical first and full render pairings: nineteen cards, two results without recorded calls, and zero duplicate result references. Installed viewer acceptance passes. Native automatic resume and generated-file acceptance remain pending. Local execution does not establish Work Cloud orchestration or dots coverage. | fixture-tested; 2 open gates. 2026-10-06 native trial (partial): ChatGPT desktop 26.930.51102 retained-history import preserved product identity, messages, tool records, usage, and model metadata. Repeat import added no duplicate. 2026-10-07 native trial (partial): Native retained-history import preserved eight screenshot images with matching byte hashes, repeat-import identity, and offline replay. Received native child and parent raw records verified originators and parent attribution. Generated-file acceptance remains pending. 2026-10-08 native trial (blocked): The existing codex-desktop source is enabled. Computer Use refused control of the Work Local application. A user-assisted two-turn generated-file trial remains pending. Installed 0.6.2 rendering verifies ChatGPT Work Local, harness version 0.158.0-alpha.2.1, and all 54 messages. Eight archived images return successfully and decode, but grouped tool results hide their controls. The source correction passes 58 focused tests. Installed image controls and automatic resume remain open. 2026-10-08 native trial (partial): Installed 079919d2 passes Chromium and WebKit checks for the Work Local label, all 54 messages, and eight image controls inside tool cards. Every control opens the correct decoded image by double-click and Enter. Reviewed screenshots were removed after retaining hashes. The separate first-render page-boundary duplication is corrected in 0d0614d4a0cd and awaits installed verification. Native automatic resume and generated-file acceptance remain pending. 2026-10-09 native trial (partial): Installed ca2af9bd passes all eight image controls in Chromium and WebKit. First render still shows 21 tool cards, four without call references, and one duplicate result reference. A full re-render gives the expected 19, two, and zero. A synthetic regression reproduces the unloaded-call result displacing later interior groups. Source correction 3f949245 prevents identity-bearing results from using type or adjacency fallback. Eleven paging tests, eleven harness/version tests, and the targeted server-window assertion pass. Private screenshots were deleted after review and hash verification. Installed first-render acceptance and the benchmark remain pending. 2026-10-09 native trial (partial): Installed 738e0ece renders identical first-page and full-render pairings in Chromium and WebKit: eighteen tool cards, one without a call reference, and zero duplicate results. All 54 messages and eight image controls remain available. These v16 counts do not establish native pairing fidelity. Original-provider and retained-raw identity comparisons expose a separate derived-event defect: an unmatched nested completion consumes a wrapper call before its exact output arrives. Canonical events v17 make recorded call identities binding. Three focused regressions and 23 native identity and preservation checks pass. A serial 200-session import comparison passes all enforced budgets in base, candidate, candidate, base order. Installed parent verification remains pending; the paging pass alone does not close native pairing fidelity. 2026-10-09 native trial (partial): Installed 313e215f plus17 serves canonical events v17 after a bounded raw rebuild. Chromium and WebKit each render 54 messages, nineteen tool cards, two results without recorded calls, and eight decoded image controls. First and full renders have identical pairings and zero duplicate result references. Independent review of fifteen native call records and 23 transcript calls confirms both unmatched completion identities. All 48 native identity, installed-module, and preservation checks pass. The original eighteen-card and one-unpaired-result target omitted the second unmatched completion. The failed receipt remains unchanged; a separate review accepts all eight requirements in each engine against the corrected native expectation. No runtime change or new browser run was needed. Installed viewer acceptance is closed. Native automatic resume and generated-file acceptance remain open. |
| ChatGPT Work Cloud (partial) | chatgpt-browser |
browser | The ChatGPT collector scans active and archived Work task lists, preserves full conversation trees, and assigns a separate Work Cloud harness. Generated sandbox-file links use the native conversation download resolver and existing attachment queue. Automatic capture retains the first native turn, its tpp origin, and the exact 29-byte generated file. Installed repeat import and offline replay preserve the conversation, attachment identity, and original filename. The bounded native field audit covers seven messages and the acquired asset. Installed Chromium and WebKit render the product labels, seven messages, and a working file control that downloads the exact original bytes. Command presentation summaries and Work turn identifiers remain raw-only. Command arguments and results, measured usage, resumed file capture, cloud execution provenance, and account completeness remain unverified. Older captures cannot replay enrichment whose supporting responses were not retained. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): The signed-in native Work history filter displayed No chats. The unified cloud composer required an environment selection. No Work Cloud conversation was available for capture or output-file acceptance. An authorized user-created trial remains necessary. 2026-10-08 native trial (partial): An authorized synthetic task completed its first native turn and produced the exact 29-byte requested file. The installed archive retains all seven messages and native conversation_origin tpp. The original capture retained its sandbox link without an asset. Native cached download evidence establishes the new resolver. Source asset planning and installed isolated import, repeat import, and offline replay preserve the exact bytes. Automatic download, the unsent resumed turn, cloud execution provenance, and rendered consumers remain open. 2026-10-08 native trial (partial): Installed 079919d2 isolated import, repeat import, and offline replay preserve seven messages and the exact 29-byte generated file. The first-turn field audit accounts for 149 native paths without drift: 73 normalized and 76 raw-only. Bundled normalization passes 123 independent comparisons against the native tree and retained archive. One command presentation summary is preserved raw, but the capture contains no command arguments, command results, or measured usage. Automatic download, continuation, cloud execution provenance, and rendered consumers remain open. 2026-10-08 native trial (partial): Installed 079919d2 renders the Work Cloud header, details, Source filter, seven messages, and first-turn marker in Chromium and WebKit. No page, console, HTTP, or network errors or horizontal overflow occurred. The generated file is absent from the production capture. Automatic download, continuation, and independent cloud execution provenance remain unverified. 2026-10-09 native trial (blocked): After the confirmed ca2af9bd reload, fifteen recent ChatGPT deliveries archive with parser 0.6.2.1. The existing synthetic task remains listed in the signed-in native interface, but its conversation cannot load after one UI retry. The retained baseline still has seven messages and no acquired file. This observation does not establish the cause of the native load failure. Automatic file acquisition remains open, and the approved continuation remains unsent. The temporary recovery tab was closed. 2026-10-09 native trial (partial): Automatic acquisition after the ca2af9bd reload retains the first-turn file in the existing seven-message conversation. Installed verification confirms the exact 29 bytes, source filename, original availability, resource and message identities, and all seven attachment fields. Isolated repeat import and offline replay without intake preserve the transcript and file digest. Field accounting covers 138 paths: 70 normalized and 68 raw-only, with no unreviewed paths or drift. The separate native page-load failure remains recorded. The approved continuation awaits user-assisted submission. Resumed file capture, generated-file rendering, and independent cloud execution provenance remain open. 2026-10-09 native trial (partial): Installed ca2af9bd renders the automatically acquired Work Cloud baseline in Chromium and WebKit. Both engines show the product label, seven messages, and the named file control. Activating the control downloads the exact 29 bytes with the original filename and digest. The direct attachment response returns 200. No horizontal overflow or page error occurs. The approved continuation still awaits user-assisted submission. Resumed capture and independent cloud execution provenance remain open. |
| Dots (partial) | chatgpt-dots-browser |
browser | The source-build ChatGPT collector discovers dot profiles and captures paginated ChatGPT direct-message rooms with task links, stable room identity, raw evidence, and source filters. Current-room capture resolves primary selection and thread routes. Room-file attachments use the native file resolver and verified archive bytes. Native automatic acquisition retains 23 room records. Read-only comparisons preserve the earlier nineteen identities and timestamps and verify four additions. Installed ca2af9bd raw replay corrects sender roles while preserving every record identity, timestamp, and raw digest. The isolated native baseline passes installed data consumers and Chromium and WebKit rendering with thirteen unique records: ten assistant, two user, and one provider-status. Its six task links remain in the payload without viewer controls. The latest capture retains seven task references without destination URLs or titles. Controlled continuation and file downloads remain pending. Avatar image bytes, widget internals, other messaging channels, saved notes, voice, and delegated task transcripts lack normalized coverage. Older captures require verified raw replay before continued capture when their archived roles differ. | fixture-tested; 5 open gates. 2026-10-08 native trial (partial): The signed-in native interface exposed existing direct messages plus Activity and Outputs. The empty Dots archive is a capture gap. The installed extension remains on its earlier build, so updated collection, source-field comparison, and automatic follow-up acceptance remain pending. 2026-10-08 native trial (partial): The enabled collector retains HTTP 422 after nine native attempts, with zero detail pages. Its active diagnostics identify the room-message endpoint. Successful cached native requests preserve compound profile IDs and use room-message limits of 20 or 32. The collector now requests 32 instead of 100 for room messages, retaining 100 for task threads. All 19 focused tests pass. Native acquisition after the correction, attachments, continuation, and installed consumers remain open. 2026-10-08 native trial (partial): A native capture acquired at 21:35 UTC retains 13 room records, three complete pages, six task links, and verified raw hashes. The messaging API labels both participants as users. Parser 0.6.2.1 uses room membership to identify the dot. Eight focused tests pass. Isolated source replay upgrades an installed-baseline archive to ten assistant messages, two user messages, and one provider-status record, preserving identity, repeat import, and 51 native value comparisons. The field contract accounts for 200 observed paths: 20 normalized and 180 raw-only, without type drift. Two embedded widget messages remain raw-only. Installed corrected roles, rendered consumers, automatic continuation, and files remain pending. 2026-10-08 native trial (partial): A later automatic capture retains 19 records, including six additions, and preserves the original thirteen messages. Corrected source normalization yields fourteen assistant messages, four user messages, and one provider-status record. Eighty-eight native value comparisons pass. The expanded field audit accounts for 227 paths: 20 normalized and 207 raw-only, with no drift or unreviewed fields. A native widget has no supported file identity. Avatar image bytes and widget internals remain outside normalized coverage. The production archive still uses the earlier role parser. Installed role replay and rendered acceptance remain pending. No synthetic continuation was sent by this trial. 2026-10-09 native trial (partial): Installed ca2af9bd replays the 13-record baseline as ten assistant messages, two user messages, and one provider-status record. All 39 identity and metadata comparisons pass. Repeat import retains identities and roles. Installed navigation, source filtering, Stats attribution, viewer payload roles, and indexed identity search pass. Bounded production raw replay updates one 19-record conversation to parser 0.6.2.1, with fourteen assistant messages, four user messages, and one provider-status record. Record identities, timestamps, and raw hashes remain unchanged. Added sender and source-role metadata match the retained source. These checks do not inspect message text. Rendered acceptance, controlled continuation, and file capture remain open. 2026-10-09 native trial (partial): Installed ca2af9bd passes Chromium and WebKit rendering for the Dots label, source filter, and thirteen unique records with ten assistant, two user, and one provider-status role. Six task links remain in recorded metadata without a rendered control. The widget appears as one unavailable, unlinked card rather than a captured file. Controlled continuation, identifiable file downloads, task-link rendering, and the other Dots surfaces remain open. 2026-10-09 native trial (partial): A retained ordinary capture received on 2026-10-09 contains 23 unique records: seventeen assistant messages, five user messages, and one provider-status record. The receiver archived it with parser 0.6.2.1. All earlier nineteen identities and timestamps remain. Read-only inspection verifies the raw hash, 115 native values, and 69 installed normalization comparisons on 313e215f modified. Field accounting remains 227 paths: twenty normalized and 207 raw-only, with no drift or unreviewed fields. Seven task references lack destination URLs and titles, and no available file asset appears. No provider trial was started. Controlled continuation, file downloads, and updated-extension acceptance remain unverified. |
| ChatGPT texts (partial) | chatgpt-messages |
local-db | Opt-in macOS Messages import of the texts a person sends their ChatGPT dot, for one configured exact handle, through the same scoped reader and importer as Instinct and Poke. A single read-only query captures only one-to-one iMessage, RCS, and SMS rows and selected attachments. Stable timeline identity, typed text, observed revisions, reactions, replies, raw preservation and offline replay pass synthetic tests. CLI configuration and the Settings number field share the same setting. Native and installed capture remain blocked on macOS Messages access and the approved reader deployment. OpenAI publishes no number every person texts: each dot has its own number, sent as a contact card after its owner verifies a phone number. The dot’s tasks, tools, models, measured usage and billing are absent from this route. Whether the Dots capture also records texted messages is unverified. | fixture-tested; 6 open gates. |
OpenClaw Foundation
Section titled “OpenClaw Foundation”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| OpenClaw | openclaw |
local-store | Per-agent transcripts, tools, fork lineage, request billing provenance, exact approval joins, run/tool durations, and estimated context composition from stored reports. Agent schemas 23 through 31 and legacy JSONL are supported. Runtime evidence needs the optional shared state database. Character categories are estimates; latest-run billing snapshots are not cumulative totals. Raw replay remains deferred. Native and installed acceptance remain unverified. | fixture-tested; 7 open gates. 2026-10-09 native trial (blocked): No local data. No binary or store exists at ~/.openclaw, ~/.openclaw-<profile>, or ~/.clawdbot; documented environment overrides are unset. Installation alone is insufficient. Existing Gateway history or separately authorized conversation creation is required. |
OpenCode
Section titled “OpenCode”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| OpenCode | opencode |
local-store | Canonical OpenCode JSON and SQLite history from its core data root, independent of the CLI, desktop, or web surface that used it. Canonical JSON and SQLite transcripts with reasoning, tools, session totals, enforced source-shape accounting, and coherent snapshot replay. Model-switch selections remain separate from historical response models and effort. Coexisting transcript tables are selected per session; competing populated stores under one native ID are rejected. | fixture-tested; 4 open gates. 2026-09-29 native trial (partial): CLI 1.18.31 against a local endpoint: legacy-table session grew from four to six archived messages on resume with stable identity, two recorded models and unchanged repeated import. A separate v2 prompt and unused model-switch session imported alongside it without false model attribution. Two raw references resolve to a verified history-only snapshot; v2 generation, full metadata and viewer verification remain open. 2026-09-29 native trial (partial): Shared-snapshot replay reproduces both native sessions, eight messages, and serialized metadata including native identity. The shared identity constructor also passes synthetic MiMo nested-actor replay tests; no native MiMo acceptance is implied. Native CLI cohort rebuild passed; returning the resulting archive to ordinary live import hit the documented full-refresh requirement. 2026-10-09 native trial (partial): Retained producer 1.18.31 SQLite cohort: 3 sessions and 14 messages pass isolated and installed import, byte-identical repeat, offline replay after excluding only wrapper stamps, search, viewer payload and Stats checks. All 245 observed paths have dispositions and 270 SQLite cells match independently. Copy-only growth preserves identity and earlier content; removed stores preserve archived bytes. Native v2 records and automatic capture remain unverified. |
OpenHands
Section titled “OpenHands”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| OpenHands | openhands |
local-store | CLI, SDK, Agent Canvas, and agent-server events plus legacy application sessions, tools, file edits, branches, child sessions, and measured usage. SDK event folders and legacy application events preserve tools, edits, and recorded usage. Legacy pickle state remains raw-only; arbitrary SDK directories need an explicit root. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No retained default SDK, Agent Canvas, or legacy history was found. No arbitrary SDK persistence directory was supplied. Native field, repeat/resume, and consumer gates remain open. |
| OpenHands Cloud | openhands-cloud |
cloud | OpenHands cloud API imports persistent conversation events and combined usage snapshots, including paused runtimes. Usage snapshots lack historical model attribution and events lack an active-branch pointer. Runtime files, cloud memories, and ZIP downloads are not fetched. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No OpenHands API key is configured. Native cloud event pagination, child histories, usage, and field fidelity remain unverified. |
Perplexity
Section titled “Perplexity”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Perplexity (partial) | perplexity |
browser | Search threads retain prompts, answers, research tools, citations, metadata, and bounded media, shopping, and code-chart image downloads. Explicit Search workflows preserve recorded code, outputs, answer identity, and unified image ownership. Four native conversations verified 38 originals, seven previews, repeat import, offline replay, search, viewer, and Stats. Native evidence is a bounded retained cohort, not complete account history. Both live discovery routes archive the native Search chart under one identity. Installed ca2af9bd captures its continuation with 13 records, three Python calls, one recorded failure, and four independently hashed PNGs. Installed 738e0ece preserves the corrected failed status through fresh import, repeat import, and offline replay. Earlier source records, raw hashes, field accounting, search, viewer payload, and one-conversation Stats pass. Chromium and WebKit each pass all 16 installed rendering assertions. Only the failed call shows a Failed badge. All tools expand and collapse, four PNG lightboxes work, and both Stats ranges count one conversation. The scoped Search gates are closed. Alternate branches, interactive visualization rendering, and original video remain incomplete. | fixture-tested; 0 open gates. 2026-10-08 native trial (partial): Four native conversations preserve 44 messages, 38 original images and seven previews. Repeat, raw hashes, resumed shopping downloads, archive-only replay, search, viewer and Stats pass. All 139 retained captures pass 550-path field accounting and raw hash checks. Fifty payloads pass current/list identity. Six code steps expose no chart images. Shared build 731cc154 activation and nine matching runtime/collector files are verified. Existing-thread navigation produced no new Perplexity job during observation. Native chart, automatic continuation and installed acceptance remain open. 2026-10-08 native trial (partial): A new Search chart trial reached the receiver automatically at 19:51 UTC and archived on installed build 10cf96ee at 20:13 UTC. Its three messages pass bundled-runtime repeat import, raw hashes, offline replay, search, viewer payloads and one-conversation Stats. The earlier response omitted all chart descriptors, so bytes remain absent. Corrected native block parameters await extension activation and recapture. Safari renders the prompt, plan and answer, but its Stats page remained at Loading current scope. Five fresh captures pass 391-path field accounting. Native continuation and live cross-route identity remain open. 2026-10-08 native trial (partial): The corrected read-only thread request exposes explicit SEARCH mode with workflow_root and unified_assets. Independent downloads verify two PNGs of 37,252 and 18,641 bytes. Parser 0.6.2.2 preserves Search identity, recorded code and output, and earlier answer identity under synthetic tests. The complete direct response passes 255-path field accounting and 34 native value comparisons. Native acquisition and extension capture field accounting remain pending. Installed 079919d2 independently passes the earlier three-message repeat, raw hash, offline replay, search, viewer payload, and Stats checks. The combined deployment must include the later Search workflow correction. 2026-10-08 native trial (partial): The 22:47 UTC reload of 8a143c3b produced current-tab and list captures with parser 0.6.2.2. Both resolve the same account, conversation, six messages, and two chart resources. All 343 observed paths pass field accounting without type/value drift or required omissions. Thirty-four native value comparisons pass under the installed runtime. A provider-wide image cooldown still delays PNG acquisition and enriched archival. Host-scoped scheduling fixes pass 31 focused tests and await installation. Native continuation remains unverified. 2026-10-09 native trial (partial): Both reloaded discovery captures archived at 23:53 UTC on October 8 under the earlier Search identity. Installed ca2af9bd verifies their six messages and native PNG files of 37,252 and 18,641 bytes. Bundled-runtime enrichment, repeat import, raw hashes, exact offline replay, 322 archived field paths, 26 source-value checks, search, compact viewer payload, and one-conversation Stats pass. The serial 200-session import ABBA passed every enforced budget. Native continuation and enriched installed rendering remain open. 2026-10-09 native trial (partial): After the ca2af9bd reload, the native Search follow-up submitted once at 00:34 UTC archived automatically by 00:38 UTC. Its 13 records retain the five earlier source records exactly and recompute one aggregate notice. Three Python calls include one recorded failure and a successful retry. All four independently acquired PNGs survive repeat import and offline replay byte for byte. The installed runtime passes 323-path field accounting, 66 native value comparisons, search, viewer payload, and isolated one-conversation Stats. The shared full check passes at 2b68a810. Installed chart/tool/Stats rendering remains open. 2026-10-09 native trial (partial): Installed ca2af9bd renders all 13 records in Chromium and WebKit. All three Python inputs and results expand and collapse. Four PNGs load with matching hashes and working lightboxes. Default and All time Stats each count one conversation. Each engine passes 39 of 41 assertions. Both remaining failures identify one missing Failed badge despite preserved error text. Parser corrections and installed failed-status acceptance remain pending. 2026-10-09 native trial (partial): Installed 738e0ece matches all 1,349 inspected runtime files. Its bundled Node 24.20.0 and Search parser 0.6.2.3 reproduce all 13 verified records in a fresh isolated archive. Exactly one Python result has failed status. Repeat import, offline replay, both raw hashes, all four PNG byte sequences, search, compact viewer statuses, and one-conversation Stats pass. The covering full check passes at f4dabac95eaf. Failed-badge rendering remains pending. 2026-10-09 native trial (partial): Installed 738e0ece passes all 16 rendering assertions in each of Chromium 153.0.8010.12 and WebKit 26.6. All 13 records remain present. Only Python call 2 shows a Failed badge and expands to its recorded error. Successful calls have no Failed badge. All three inputs and results expand and collapse. Four PNGs match independently verified hashes and have working lightboxes. Default 3M and All time Stats each count one Perplexity conversation. No page errors, HTTP errors, non-GET requests, or horizontal overflow occur. The viewer stopped and its isolated home was released. This closes the scoped Search rendering gate without establishing complete account history or broader artifact and branch coverage. |
| Perplexity Computer (partial) | perplexity-computer |
browser | Experimental browser capture enumerates retained, archived and scheduled Computer tasks. Workflow text, tools, nested subagent identities, statuses, recorded step models, inline artifacts, and permitted download descriptors retain raw evidence and offline replay. Native access exposed only a transient welcome task whose retained-thread request returned THREAD_NOT_FOUND. After shared extension activation, an attempted CSV/PNG task on the signed-in Free plan opened Upgrade to Pro instead of a durable task. Native Computer history, field fidelity, artifact resolution and automatic capture remain unverified. Account-wide memory, Brain, Skills, connectors, automation definitions, independent subagent history and device-local execution are outside this route. | fixture-tested; 8 open gates. 2026-10-08 native trial (blocked): Signed-in web inspection exposed a transient welcome task whose retained-thread request returned THREAD_NOT_FOUND. Shared build 731cc154 activation and matching runtime/collector bytes are verified. Submitting a CSV/PNG task on the Free plan opened a blocking Upgrade to Pro dialog. No task response or artifacts appeared, and no purchase or follow-up submission occurred. Native history, fields, artifacts, continuation and installed round-trip acceptance remain blocked. Source, fixture, Chromium restart and automatic-continuation tests pass. |
Vercel
Section titled “Vercel”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| fx | fx |
local-store | Read-only history and automatic capture from FX_HOME sessions, including checksummed v2 logs, legacy snapshots, commit-watermark-fenced logs, and conversation logs. Three native-generated sessions against a local scripted endpoint pass import, repeat, raw rebuild, search, viewer payload, and Stats checks. Synthetic fixtures additionally cover tools, images, child links, interruption, compaction, and source regression holds. Native tool and image generation, historical producer trials, full viewer verification, and installed acceptance remain pending. Superseded turns and provisional running records remain raw. Missing required v2 blobs hold the session. Large sessions above the bounded snapshot limit require a later capacity extension. | fixture-tested; 4 open gates. 2026-11-14 native trial (partial): Native fx 0.0.17 generated three synthetic sessions against a local scripted endpoint. Import, repeat, exact offline rebuild, search, viewer payload, and Stats checks pass without changing provider files. All 61 observed paths across 102 records are classified: 30 normalized and 31 raw-only. No external inference or real credentials were used. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| pi | pi |
local-store | pi coding agent JSONL session trees from ~/.pi/agent/sessions, including branches and compactions. Only the active branch is normalized; abandoned branches remain raw. Response models, requested models and configuration remain separate. Calculated costs are estimates and zero counters remain explicit. Resume can omit effort changes; unmatched effort stays unknown, while omitted same-model changes are unobservable. Version-3 system, standalone usage, context-edit and summary records from 1.1.0 retain exact typed evidence without rewriting historical messages. | fixture-tested; 7 open gates. 2026-09-29 native trial (partial): Official CLI 0.73.1 with a local synthetic endpoint: initial/resumed conversation retained sixteen messages, two human prompts, five tool calls, controlled edit/shell errors and a byte-verified image. Initial prefix, repeat import, source bytes and full offline raw rebuild passed. All 53 observed field paths audited. Hosted usage, other session shapes and full viewer acceptance remain open. 2026-10-09 native trial (blocked): No local data under ~/.pi/agent/sessions and no configured agent/session override. Upstream v1.1.0 retains version 3 while adding system, usage and context-edit records. Parser 0.6.2.1 preserves these records with synthetic import, repeat and raw replay coverage. Installing and creating a conversation, or supplying a store, would permit a native trial. |
Prime Intellect
Section titled “Prime Intellect”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Prime Agent | prime-agent |
local-store | Prime Agent JSONL session trees from ~/.prime/agent, including legacy per-cwd stores, RLM child sessions, and session-local continual-harness state. Versioned JSONL transcript with model routing, reasoning effort, service tier, tools, token and cost usage, attachments, lifecycle state, linked RLM child sessions, and raw-preserved kernel and session-local continual-harness state. | fixture-tested; 4 open gates. 2026-09-29 native trial (partial): Three retained sessions passed repeat-import and raw-hash checks; full metadata, tool, attachment, and viewer acceptance remain open. 2026-10-09 native trial (partial): Three retained version-3 sessions contain 6 configuration notices. Isolated and installed import, repeat, raw replay, viewer payload, zero-human-turn Stats and source removal pass. All 32 observed paths have dispositions. Copy-only synthetic growth preserves identity and becomes searchable; native notices have no text-search hits. Native conversation, tool, attachment and child-record fidelity remain unverified. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Qoder Cloud Agents | qoder-cloud |
cloud | Qoder cloud tasks import retained messages, tools, output files, repository context, task relationships, and billing credits. Event-id deduplicated request spans, confirmations, compactions, and thread events retain request credits separately from cumulative session credits. Credits are separate from inference tokens. Reasoning and measured token counts are unavailable; attachment download failures and limits remain explicit. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No Qoder cloud access token is configured. Current public events, child threads, file downloads, and memory fidelity require authenticated native inspection. |
| Qoder local history | qoder |
local-store | Native transcripts, desktop SQLite history, subagents, recorded edit patches, and available request usage. CLI and Work transcript request credits remain separate from token counts and populate canonical usage. Native transcripts and desktop histories with tools and subagents. Hosted requests hide token counts; credits remain separate. Selected desktop models are not request attribution. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): Read-only discovery found no retained CLI, app, IDE, or Work history in inspected roots. Native round-trip, fields, resume, and consumers remain blocked. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Poe (partial) | poe |
browser | Experimental account-bound browser capture reuses the mounted Relay network for saved history, backward message pagination, forward canvases and separate human and bot member lists. Every response is checked again before publication. Raw replay preserves settled text and participant identity without guessed models. Native authenticated history and automatic capture remain unverified. Original resources, prior edit versions, separate device content, account memory, bot instructions and official export import remain incomplete. Canvas content, reactions, parameters and job metadata remain raw. Temporary and deleted history has no recovery claim. | fixture-tested; 10 open gates. 2026-10-08 native trial (blocked): Arc displayed Poe’s login page with Terms of Service acceptance on continuation. Public rolling client assets supplied persisted query and Relay network contracts. No authenticated query, sign-in, terms acceptance, prompt or export request occurred. This records an access gate, not provider-side impossibility. |
Roo Code
Section titled “Roo Code”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Roo Code CLI | roo-code-cli |
local-store | Historical Roo Code task files, tool calls, reasoning, attachments, child tasks, and recorded usage. Historical CLI tasks use the VS Code shim store and version-1 shared indexes. Deleted ephemeral sessions are unavailable and models can be absent. Built-in raw replay and cross-file transactional snapshots are deferred. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no ~/.vscode-mock/global-storage/tasks or ~/.roo/tasks. Audited CLI 0.1.17 in final v3.54.0 source. The package is private and its npm lookup returns 404. Retained history or an authorized historical executable is needed. |
| Roo Code IDE | roo-code |
local-store | Historical Roo Code task files, tool calls, reasoning, attachments, child tasks, and recorded usage. Historical IDE tasks preserve tools, reasoning, attachments, and recorded usage. Models can be absent and legacy clocks can be missing. Only version-1 shared indexes are accepted. Built-in raw replay and cross-file transactional snapshots are deferred. Native acceptance is blocked by no local data. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): No local data: no rooveterinaryinc.roo-cline/tasks in checked editor, profile, JetBrains, or remote-server roots, and no configured custom storage path. Audited final v3.54.0. The product documents its shutdown; retained history or an authorized historical installation is needed. |
Roomote
Section titled “Roomote”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Roomote cloud | roomote-cloud |
cloud | Roomote API imports complete visible conversation pages and linked tasks; a matching database import keeps its richer history. Cloud text is sanitized and omits archived or hidden conversations from listing. Known IDs can be requested explicitly; native hosted acceptance is pending. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No Roomote deployment origin or authorized cloud token is configured. Native hosted history and newly documented artifact/receipt routes remain unverified. |
| Roomote database | roomote-db |
database-copy | Roomote database snapshots import messages, tool arguments/results, reasoning, usage ledgers, child tasks, and scoped memories. Only history tables and projections are preserved. Cloud and database copies join by verified native identity, and partial projections cannot replace complete history. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No Roomote deployment origin or database connection is configured. Read-only native schema, tool, accounting, and artifact inspection cannot start. |
Tencent
Section titled “Tencent”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| CodeBuddy | codebuddy |
local-store | CodeBuddy Code project transcripts and subagent runs from ~/.codebuddy/projects. Supported message token counts, compaction summaries, and subagent parent links populate runtime evidence without unverified approval or hook formats. The current published bundle defines the reviewed shape. Stable bounded reads preserve transcript and metadata bytes, latest titles, executed models, and session-meta. Unknown records remain raw and counted. Active-branch reconstruction and external image/blob acquisition remain incomplete. Native acceptance has no local data. | fixture-tested; 7 open gates. 2026-10-09 native trial (blocked): No local data. No CodeBuddy binary or ~/.codebuddy/projects store exists, and documented environment overrides are unset. Installation alone is insufficient. Existing project journals or separately authorized conversation creation is required. |
Spear Street Technology
Section titled “Spear Street Technology”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Instinct (partial) | instinct |
local-db | Opt-in macOS Messages import for one configured exact handle. A single scoped read-only query captures only one-to-one rows and selected attachments. Stable timeline identity, typed text, observed revisions, reactions, replies, raw preservation and offline replay pass synthetic tests. CLI configuration and the app source input share the same setting. Native and installed capture remain blocked on macOS Messages access and the approved reader deployment. Historical edit plists and rich app payloads stay raw-only. Attachments are bounded and offloaded files remain unavailable. Models, internal tools, measured usage and billing are absent from this route. WhatsApp Desktop store presence is confirmed, but protected-store access prevented verifying an Instinct chat. WhatsApp is unimplemented. | fixture-tested; 6 open gates. 2026-10-09 native trial (blocked): Terminal access to Messages returns Operation not permitted. No native Messages database or attachment was opened or copied. Synthetic acceptance is separate. WhatsApp ChatStorage.sqlite exists, but a read-only open failed before any chat rows were read. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Warp cloud | warp-cloud |
cloud | Warp cloud pages provide normalized conversation steps, recorded models, parent/child links, and aggregate usage. Aggregate usage belongs to the parent when no child allocation is recorded. Unsupported third-party harness sessions can return HTTP 422. | fixture-tested; 4 open gates. 2026-10-08 native trial (blocked): No Warp cloud API key is configured. Native run pagination, tool trajectory, child accounting, and artifact fidelity remain unverified. |
| Warp local history | warp |
local-store | Local SQLite task history retains tools, reasoning, models, usage, and linked runs. Request ids join recorded timing, outcomes, incompleteness, and context segments with measured counts where present. The installed Warp 0.2026.09.16.08.27.02 protobuf descriptors verify request timing and context field layouts. Occupancy fractions do not establish token counts. Runtime decoding has synthetic coverage; live native-record and installed Hansard acceptance remain unverified. | fixture-tested; 4 open gates. 2026-10-08 native trial (partial): Warp 0.2026.09.16.08.27.02 is installed. macOS denies its container and database open with EPERM. The fixed native diagnostic preserves that denial; transcript fidelity, resume, and consumers remain blocked. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Cursor CLI | cursor-cli |
local-store | Cursor CLI interactive, headless, and ACP conversation stores from ~/.cursor/chats and ~/.cursor/acp-sessions, including saved Cursor account profiles. Stores preserve available models, reasoning, tool arguments/results, edits, and step clocks. Native existing-history, new/resumed headless and new ACP capture, repeat, direct parser replay, field checks, search, viewer, and Stats passed in isolated archives. Headless transcript/store identity passes both import orders. Store context counts are not stdout usage totals. Headless system-prompt model context is not response-model attribution. Unobserved field families, completed-turn ACP resume, interactive permissions, and built-in raw replay remain open. | fixture-tested; 3 open gates. 2026-10-08 native trial (partial): CLI 2026.09.28-64d2043 existing-history trial: one chat, 18 messages, five calls and five results. Repeat imported zero. Five raw hashes, 34 blob hashes, and direct parser replay passed. A private audit retained 200 observed field paths and independently compared tool arguments/results, models, thinking, clocks, durations, seven summary values, and identity. Search, Arc viewer, and Stats passed. The later authorized trial below adds new/resumed headless and ACP evidence. 2026-10-08 native trial (partial): Authorized CLI 2026.09.28-64d2043 trial: new/resumed headless and new ACP file-read prompts completed. The production source watcher and supervisor import worker captured final states in an isolated archive after 14.3, 5.8, and 8.1 seconds without manual import. Headless retained 16 messages and two calls, ACP ten messages and one call. Repeat imported zero and skipped three with unchanged transcripts. Six raw and 51 blob hashes, direct parser replay, 72 headless and 74 ACP field paths, independent field values, search, Arc rendering, and Stats passed. Headless transcript-first and store-first routes retained one complete identity. Later trials below record installed configuration and startup catch-up evidence. ACP resume, unobserved fields, and built-in replay remain open. 2026-10-08 native trial (partial): The supervisor configuration-refresh fix passes five new regression tests. A native trial enabled Cursor CLI only in a temporary configuration after watcher startup. Its normal 30-second check attached both roots and imported all three stores with no errors, preserving the headless and ACP transcripts exactly. No manual import or new prompt was needed. The later bundled trial records installed verification. 2026-10-08 native trial (partial): Installed bundle 10cf96ee restarted the supervisor with 28 watched roots. Its configuration watcher and supervisor match the tested runtime bytes. Bundled Node and modules then passed native configuration catch-up in a temporary archive: the ordinary 30-second poll attached both CLI roots and imported three stores with zero errors. Headless retained exactly 16 messages and two calls, ACP ten messages and one call. No new prompt, manual import, or real configuration change occurred. A separately authorized tool-free continuation completed at 20:30:41 UTC. An early 20:34 UTC read still found neither conversation. Later production evidence below records their successful capture on this bundle. 2026-10-08 native trial (partial): Startup fix aeaeba6ac9cc queues catch-up on the first successful source selection. Both added regressions fail on the old helper, and all seven watcher tests pass with the fix. With Cursor CLI enabled before isolated startup, the native trial attached both roots and imported three stores with zero errors. Headless reached 19 messages and two calls, preserving its earlier 16 messages exactly and the new continuation marker. ACP remained at ten messages and one call. No manual import, new prompt, or post-start configuration change occurred. Later bundled replay and field checks of the continuation are recorded below. Installed startup-fix acceptance is recorded below. 2026-10-08 native trial (partial): Installed 10cf96ee production capture passed at 20:34 UTC. Read-only inspection found one native identity per trial, exactly matching isolated transcripts: 19 headless messages and two calls, ten ACP messages and one call. The earlier headless prefix remains unchanged. Six raw objects and six projections pass size and SHA-256 checks. Four loaded modules match the bundle. The approved continuation arrived after 253.197 seconds without manual import. Bundled direct parser replay of its latest store, three raw hashes, 42 blob hashes, 72 JSON field paths, and independent tool, text, clock, summary, and identity checks pass. Arc blocks installed loopback rendering with ERR_BLOCKED_BY_CLIENT. Earlier isolated consumers remain verified. 2026-10-08 native trial (partial): Installed 079919d2 passed read-only production checks and bundled native startup catch-up at 21:32 UTC. Four loaded module stamps match the bundle, including aeaeba6ac9cc. Production retains one exact native identity per trial and six verified raw objects plus six projections. With Cursor CLI enabled before temporary startup, the initial catch-up attached both roots and imported three stores with zero errors. Headless retained exactly 19 messages and two calls, ACP ten messages and one call. No manual import, new prompt, or configuration change after startup occurred. Installed rendering is recorded below. 2026-10-08 native trial (partial): The read-only installed 079919d2 viewer passes all five required assertions in both Chromium and WebKit. Both headers identify Cursor CLI, with 19 headless and ten ACP messages. All three headless replies and the ACP reply render. Expanded file-read cards show the exact results and durations of 416 ms, 173 ms, and 285 ms. No page, console, HTTP, or network errors or horizontal overflow occurred. Four screenshots were reviewed and removed after retaining their checksums and the assertion receipt. Optional installed Stats was not checked because no per-session control is available. Earlier isolated Stats checks remain verified. |
| Cursor | cursor |
local-store | Cursor app workspace/global SQLite history and current agent transcripts. The versioned generation rule also classifies recovered app conversations as Cursor 1, Cursor 2, or Cursor 3. Twenty recovered records now receive an era when their dates permit it. The bounded native SQLite trial covers Cursor 1 history only. Modern transcripts, app-assisted transcript recovery, new/resumed automatic capture, and full field fidelity remain native gates. Usage recovery does not reconstruct chat text. Built-in SQLite raw replay is deferred. | fixture-tested; 6 open gates. 2026-10-08 native trial (partial): App 3.22.12: 67 workspace databases and one global database with 273 Composer records. The global snapshot has 267 Cursor 1 records, no Cursor 2 records, and six Cursor 3 records without conversation bodies, headers, or bubbles. Two bounded slices archived 96 Cursor 1 sessions, leaving 263 global candidates. Repeat imported zero with no errors. All 963 raw hashes passed. Five global samples matched direct snapshot parser replay. Search, Arc viewer, and Stats passed. No current project transcripts were present. App-assisted recovery awaits prompt approval. |
| Cursor cloud agents | cursor-cloud |
cloud | Read-only v0 conversation text and v1 agents, runs, repositories, usage, and resumable run streams with thinking, tool calls, arbitrary JSON arguments/results, and truncation flags. Enabled sources poll every minute. Run streams expire. Simplified events omit model identity and per-event clocks. SDK updates and attachment references remain raw without duplicate projection or binary download. Retained evidence and request limits can stop capture visibly. Built-in raw replay is deferred. Native API access, resume, and automatic capture remain unverified. | fixture-tested; 6 open gates. 2026-10-08 native trial (blocked): Credential discovery again returned auth_unavailable after the CLI trials. A Cursor Dashboard API key or appropriately scoped service-account key is required. Ordinary CLI sign-in is insufficient. No native list, conversation, or run-stream response was obtained. Fixtures cover stream values, resume, repeat, expiry, truncation, metadata retention, corruption, resource bounds, scheduling, indexed nonpath arguments/output, and Stats tool counts. The isolated cloud trial changed no source setting. |
| Grok Bot (partial) | grok-bot, grok-bot-cloud |
desktop-cache, cloud | Grok Bot 0.72.1 local replicas accumulate sanitized evidence with offline replay and filesystem capture. Local and cloud identities share one account scope; verified earlier duplicate copies retain their evidence in one conversation. An opt-in cloud reader traverses available history through read-only methods and joins local tails. Accounts reads cached weekly usage. Skills and plugins lists observed connector, Bot, and template metadata without callable controls. Native local snapshots and replay passed in an isolated archive. Local replicas retain at most 200 records. Cloud API and Keychain access remain unverified. Native messages, widgets, connector cards, uploads, image references, and reply links are covered. Structured tool/approval types have renderer and synthetic evidence only. Original attachment bytes, cloud skills, per-turn tokens, and native model attribution remain unavailable. | fixture-tested; 3 open gates. |
| Grok Build | grok-build |
local-store | Grok Build durable ACP updates, prompt usage ledgers, structured chat sidecars, and provider request failures from ~/.grok/sessions, including encoded nested temporary project folders. Stored permission, retry, hook, and subagent records use the shared runtime contract; recorded SessionStatus updates retain current context occupancy. Durable ACP updates, typed hook/subagent/permission/MCP lifecycle, prompt-level token/cost ledgers, source-shape telemetry, structured chat enrichment, and raw session sidecars. Current model settings remain separate from recorded model usage. Unhandled update types produce explicit warnings and raw-only records instead of silent omissions. Recorded format versions, zero counters and unambiguous repository remotes are retained; explicit repository overrides remain authoritative. Stable bounded snapshots protect parsed/raw consistency and archived update prefixes. Snapshot limits and mutable-sidecar regression gaps remain documented. Directories without updates.jsonl are excluded. Runtime evidence extensions have synthetic regression coverage; native emission across installed versions remains unverified. | fixture-tested; 5 open gates. 2026-09-29 native trial (partial): Ten retained sessions passed repeat-import and raw-hash checks; historical client versions and full field coverage remain unverified. 2026-09-29 native trial (blocked): Grok 1.0.44 rejected fresh synthetic prompts on two saved profiles because the usage balance was exhausted. Successful replies and resume remain unverified. 2026-09-29 native trial (partial): Two native failed turns retained provider error text, five messages and eleven hash-verified raw records each. Repeats were unchanged; this does not verify successful replies or resume. 2026-09-29 native trial (partial): All 95 observed paths in fourteen retained summary/signal records pass explicit field accounting. Parser 0.5.6.3 replay preserved 662 messages and 127 verified raw records; event and attachment field audits remain open. 2026-10-09 native trial (partial): Thirteen copied sessions and 1,385 messages passed isolated import, byte-identical repeat, offline raw replay, search, viewer payload, Stats, copied growth, movement and source-loss checks. All 166 raw references matched source hashes. The field inventory covered 1,062 paths; 97 summary/signal paths have explicit accounting. Two metadata-only directories lack updates.jsonl. Producer binary versions, live automatic capture and installed acceptance remain unverified. |
| Grok chat (web and mobile) (partial) | grok-chat |
browser | Experimental browser capture inventories saved and workspace chats, reads complete response trees and batched bodies, and follows external tool results with saved cursors. Selected ancestry, text, scalar or array reasoning, response models and paired tool output normalize. Native continuation, exact raw preservation, offline replay, search, Stats and rendered tool output pass in isolated archives. Installed ca2af9bd passes native replay. Installed 738e0ece repeats rendering successfully in Chromium and WebKit after shared tool-pairing changes. Fresh automatic archival passed on 8a143c3b. Extension-dependent trials remain pending until confirmed reload. Complete account inventory, native checkpoint restart, original media, workspace files, independent tasks, organization switching, Imagine and Companion histories remain incomplete. The official export request completed, but its downloaded schema and importer remain unavailable. Grok Build, Grok Bot and Grok on X are separate products. | fixture-tested; 5 open gates. 2026-10-08 native trial (blocked): Read-only Arc inspection displayed Sign in. An unauthenticated HTTP trial of the account route returned a Cloudflare browser-signature denial. No authenticated history or export was acquired. These observations do not establish provider-side impossibility. 2026-10-08 native trial (partial): A later signed-in Arc inspection reached Grok Data Controls and the Command Menu. The Command Menu displayed History is empty. Export Account Data is available, but no export request or prompt was submitted. Direct browser navigation to the account-read route returned ERR_BLOCKED_BY_CLIENT. The collector account route, native payloads and automatic capture remain unverified. 2026-10-08 native trial (blocked): A later signed-in Arc trial submitted a synthetic prompt and reached an age confirmation prompt before any response. The trial stopped at Year of birth and Continue. An extension-state inspection separately reported account binding and workspace-inventory progress without a recorded error. No native conversation payload or official export was acquired. Automatic capture and export import remain unverified. 2026-10-08 native trial (partial): After age confirmation cleared, two marker turns completed in one native conversation. Automatic captures with 14 and 16 primary messages reached the receiver. Installed 10cf96ee rejected both because the first human response named an unlisted root parent. The official export request returned an email confirmation. The downloaded archive remains unavailable. 2026-10-08 native trial (partial): The corrected receiver passes isolated native import, repeat, continuation, stale-snapshot protection and offline replay. It preserves 16 primary messages, eight tool calls, five tool results and both complete raw observations. Four identified X engagement counters may refresh without changing post content. Two terminal-initialization cards remain raw and counted as unknown. 2026-10-08 native trial (partial): Explicit field accounting classifies all 294 observed native paths: 93 normalized and 201 raw-only, with no unreviewed paths or type drift. Both markers reach search and the rendered viewer. Stats counts one Grok conversation. The search tool card displays all 14 output lines and expands and collapses. All 17 focused tests pass. Installed archival, the import benchmark and full project check remain pending. 2026-10-08 native trial (partial): Installed bundle 079919d2 and bundled Node pass both native envelopes with parser 0.6.2.1: initial import, repeats, continuation, stale protection, exact raw preservation, offline replay, search, viewer details and Stats. The result retains 16 primary messages, eight calls and five results. The receiver module matches the commit. This isolated result is separate from production capture. 2026-10-08 native trial (partial): An explicit retry at 21:43 UTC let the installed receiver publish five retained automatic captures and skip one stale snapshot. Read-only production checks verify parser 0.6.2.1, one stable conversation, 16 primary messages, eight calls, five results and both markers. Both raw object hashes match, five complete observations remain, and 48 native content, clock, model and reasoning comparisons pass. Production keyword search finds both markers after the index refresh at 21:58:54 UTC. The clean shared full check at caf220a44b96 includes the receiver correction and passes. Production rendering and fresh capture after the extension reload remain pending. 2026-10-08 native trial (partial): Installed bundle 8a143c3b and bundled Node 24.20.0 repeat isolated native acceptance. Four receiver and version files match the commit. After the confirmed extension reload, a fresh capture received at 22:55:47 UTC archives automatically at 22:57:42 UTC on its first attempt. Production checks retain one identity, 16 primary messages, eight calls, five results and both markers. All 48 native value comparisons and both raw hashes pass. Eight retained observations preserve the original trial captures and the fresh envelope, whose digest matches its receiver job. All 295 observed paths are classified: 93 normalized and 202 raw-only, without drift or unreviewed fields. Read-only Arc history displays one saved chat and an empty Projects page. Complete inventory, native checkpoint restart and installed rendering remain unverified. 2026-10-08 native trial (partial): Serial 200-session performance comparisons use base 918b09da9a04 and candidate 73f707923b05 with identical bundled Node 24.20.0 and dependencies. Grok initial and continuation ABBA rounds verify all final transcripts and show no candidate slowdown in that window. The first generic comparison retains one candidate throughput and scale-ratio failure. Both repeated candidates pass every enforced budget; the repeated base misses the absolute throughput floor, establishing an environment limit in that window. All generic projection, operation and memory checks pass. No speedup or default-size release performance is claimed. 2026-10-09 native trial (partial): Installed ca2af9bd and bundled Node 24.20.0 repeat native import, continuation, stale protection, exact raw retention, replay, search, viewer payload and Stats checks. A separate installed production viewer passes eight required assertions and five supplementary checks in each of Chromium and WebKit. Both assistant markers, eight Thinking blocks, three tool groups, all five results and the complete 14-line search output render correctly without horizontal overflow. All 1338 runtime source and entry-point files match the installed revision. Four inspected receiver and version files remain unchanged in f6c2b2d0. Complete inventory, native checkpoint restart, original resources and official export import remain open. 2026-10-09 native trial (partial): Installed 738e0ece passes the production rendering recheck after shared tool-pairing changes. Chromium and WebKit each pass eight required assertions and six supplementary checks. All 21 messages, both assistant markers, eight Thinking blocks, three tool groups, five recorded results and the complete 14-line search output render correctly without horizontal overflow. All 1349 installed source and entry-point files match the revision. The trial used the existing conversation without a provider request, rebuild or reimport. Extension-dependent trials remain pending until confirmed reload; the five existing coverage gates remain open. |
Xiaomi
Section titled “Xiaomi”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| MiMo Code | mimo-code |
local-store | MiMo Code SQLite history from its independent mimocode data root and enforced storage cohort. Actor registry snapshots retain parent links; persisted system and tool prefixes expose UTF-8 byte sizes without invented token counts. The distinct SQLite history projection preserves reasoning, tools, usage, actors, title authority, prompt configuration, and prefix snapshots through v0.1.15. Credential and runtime tables stay excluded. Native, automatic, and installed acceptance have no local data. | fixture-tested; 7 open gates. 2026-10-09 native trial (blocked): No local data. No MiMo binary or ~/.local/share/mimocode store exists. XDG, vendor, and Hansard MiMo overrides are unset. Installation alone is insufficient. Existing SQLite history or separately authorized conversation creation is required. |
Zed Industries
Section titled “Zed Industries”| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Zed | zed |
local-store | Native SQLite history retains tools, settings, and cumulative tokens. Allowlisted local Agent/Subagent telemetry adds attempts, cache usage, compaction, turn durations, and authorization, with raw replay across log truncation. Separate sidebar records identify external Claude, Codex, Gemini, and Cursor threads. Zed truncates telemetry at launch; unread events cannot be recovered. Batch offsets are not wall clocks. Native database and telemetry replay has synthetic coverage; installed capture remains unverified. External transcripts stay in provider stores. Historical text-thread replay remains unsupported. | fixture-tested; 5 open gates. 2026-10-09 native trial (blocked): No local data at ~/Library/Application Support/Zed/threads/threads.db or the checked db/0-stable, db/0-preview, db/0-nightly, and db/0-dev sidebar roots. No Zed app or executable was found. Preview, ~/.local/share/zed, and standard Flatpak roots were also checked. Zed 1.23.2 native database source was reviewed. A Zed installation with a retained native thread would permit a trial. 2026-10-09 native trial (blocked): No local data at ~/.config/zed/conversations or ~/Library/Application Support/Zed/conversations. Historical SavedContext 0.4.0 source and older upgrade paths were reviewed separately from Zed 1.23.2 native-agent persistence. A historical text-thread file is required. Installing current Zed does not establish availability of this historical format. |
| Product | Source selector or format | Route | Coverage and limitations | Verification |
|---|---|---|---|---|
| Posit Assistant | posit-assistant |
local-store | Version 3 trees normalize the selected branch, recorded models, main usage, and tools. Optional usage-events journals preserve keepalive and classifier usage separately. Nested subagent stores retain explicit parent links. Current model and effort settings remain separate from historical replies. Alternate branches, UI-only records, and unknown blocks remain raw-only. Auxiliary usage and nested links have synthetic regression coverage; broader native cohorts, older schemas, JSON/HTML exports, and other hosts remain unverified. | fixture-tested; 6 open gates. 2026-09-29 native trial (partial): Official TUI v0.1.0 build dev against a local synthetic endpoint: first and resumed imports grew from two to six LM records with stable identity, one tool call/result, four hash-verified raw files and unchanged repeated import. All 75 observed field paths audited; hosted billing and other hosts remain unverified. 2026-10-09 native trial (partial): Retained schema-3 cohort: 2 sessions and 8 messages pass isolated and installed import, byte-identical repeat, offline replay, search, viewer payload, Stats and source removal. All 75 captured and 16 excluded auxiliary paths have dispositions. Copy-only growth preserves identity and earlier content. Hosted models, branches, children, attachments and other hosts retain separate gates. |
Unsupported products and missing routes
Section titled “Unsupported products and missing routes”This is a running backlog. Priority 1 covers explicit product gaps and major hosted histories; priority 2 covers additional maintained clients; priority 3 covers narrower or unverified surfaces. Product identity remains unresolved where no authoritative match is available. A failed public-page request is not evidence that a product does not exist. A blocked native trial records the tested access boundary, not successful import or automatic capture.
| Company or project | Product | Candidate routes | Next evidence needed | Verification | Priority |
|---|---|---|---|---|---|
| Agent37 | QM (Agent37 hosted) | cloud, browser | Upstream identifies Agent37 as a third-party hosted QM service. Hosted account/instance identity, access, version, retained-history API, pagination and export availability remain unverified. Self-hosted database access does not establish hosted coverage. | research; 5 open gates. | 1 |
| Amazon | Amazon Bedrock Managed Agents powered by OpenAI | cloud, manual-export | The official product is in limited preview. AWS hosts the agent loop and model inference; AgentCore or self-hosted compute executes tools. AWS IAM, endpoints and retained state differ from OpenAI Agents API. Shared harness concepts do not establish identical source contracts. | research; 5 open gates. | 1 |
| Amazon | Kiro cloud sessions | cloud, browser | Local and supplied Kiro exports are implemented; account-wide hosted enumeration needs separate validation. | research; 5 open gates. 2026-10-08 native trial (blocked): Dedicated cloud acquisition is unsupported. No authorized cloud history or native export was available; local discovery found no retained conversation. | 1 |
| Amp | Amp Puck | cloud, browser, manual-export | Puck conversations are distinct from coding threads created by Puck. The documented puck:true search selects the former only. Verify personal-assistant history, delegated thread links, voice, Slack, and web/TUI/mobile identity independently of Amp thread export. | research; 5 open gates. 2026-10-08 native trial (blocked): Puck remains a separate unsupported surface. No authorized Puck history or native export was supplied; ordinary Amp thread support does not establish Puck coverage. | 1 |
| Anthropic | Claude for Excel | local-store, manual-export, cloud, telemetry-export | Excel history is documented in pivot.claude.ai IndexedDB, scoped by user, organization and app, retaining the latest 50 conversations. Cross-app identity, compaction, tools and attachment bytes require native inspection. OTEL omits assistant text and truncates content attributes. Enterprise Compliance API coverage needs separate acceptance. Excel sandbox Data enumeration returned EPERM. | research; 5 open gates. 2026-10-08 native trial (blocked): Excel sandbox Data enumeration returned EPERM. Official export-script inventory copied nothing and found no readable stores. Permission-denied Office roots are not empty-history evidence. No transcript sample, Compliance API export or automatic capture was verified. | 1 |
| Anthropic | Claude for Outlook | local-store, manual-export, cloud, telemetry-export | Outlook history is documented in pivot.claude.ai IndexedDB, scoped by user, organization and app, retaining the latest 50 conversations. Cross-app identity, compaction, tools and attachment bytes require native inspection. OTEL omits assistant text and truncates content attributes. Enterprise Compliance API coverage needs separate acceptance. The readable Outlook cache contained six unrelated add-in manifests. Inspected WebKit roots contained no transcript IndexedDB. | research; 5 open gates. 2026-10-08 native trial (blocked): The readable Outlook cache contained six unrelated add-in manifests. Inspected WebKit roots contained no transcript IndexedDB. Official export-script inventory copied nothing and found no readable stores. Permission-denied Office roots are not empty-history evidence. No transcript sample, Compliance API export or automatic capture was verified. | 1 |
| Anthropic | Claude for PowerPoint | local-store, manual-export, cloud, telemetry-export | PowerPoint history is documented in pivot.claude.ai IndexedDB, scoped by user, organization and app, retaining the latest 50 conversations. Cross-app identity, compaction, tools and attachment bytes require native inspection. OTEL omits assistant text and truncates content attributes. Enterprise Compliance API coverage needs separate acceptance. PowerPoint sandbox Data enumeration returned EPERM. | research; 5 open gates. 2026-10-08 native trial (blocked): PowerPoint sandbox Data enumeration returned EPERM. Official export-script inventory copied nothing and found no readable stores. Permission-denied Office roots are not empty-history evidence. No transcript sample, Compliance API export or automatic capture was verified. | 1 |
| Anthropic | Claude for Word | local-store, manual-export, cloud, telemetry-export | Word history is documented in pivot.claude.ai IndexedDB, scoped by user, organization and app, retaining the latest 50 conversations. Cross-app identity, compaction, tools and attachment bytes require native inspection. OTEL omits assistant text and truncates content attributes. Enterprise Compliance API coverage needs separate acceptance. Word sandbox Data enumeration returned EPERM. | research; 5 open gates. 2026-10-08 native trial (blocked): Word sandbox Data enumeration returned EPERM. Official export-script inventory copied nothing and found no readable stores. Permission-denied Office roots are not empty-history evidence. No transcript sample, Compliance API export or automatic capture was verified. | 1 |
| Anthropic | Claude in Chrome | cloud, browser, local-store | Cowork side-panel account history, classic extension history, and Code/Desktop-driven browser work require separate provenance. Browser actions, approvals, screenshots, downloads, scheduled tasks and recorded workflows remain unverified. Chrome profile access was denied on the inspected Mac; Arc is not a supported substitute. | research; 5 open gates. 2026-10-08 native trial (blocked): Chrome is installed, but profile enumeration returned EPERM. The readable Arc extension inventory contained no Claude manifest. No Chrome-origin session sample was identified. Extension installation, classic history and account-backed browser records remain uninspected. | 1 |
| Anthropic | Claude Security (cloud) | cloud, manual-export | The documented scan surface requires Enterprise access. Project/scan/finding CSV or Markdown exports and webhooks do not establish full scan transcripts, validator activity, approvals or patch-review lineage. Remediation creates a separate Code web session. The tested Max account did not expose scan history. | research; 5 open gates. 2026-10-08 native trial (blocked): Opening the documented /security route redirected to /new on the signed-in account. No project, scan, finding export or webhook history was reachable. No scan or remediation session was started. | 1 |
| Anthropic | Claude Security plugin for Claude Code | local-store, manual-export | The local plugin is distinct from hosted scanning. Timestamped reports contain findings, revision context and patches alongside Claude Code parent/validator transcripts. Generic Code capture does not prove report fidelity or lineage. Cached plugin source is not installation or usage evidence; no native report sample was found in the bounded inventory. | research; 5 open gates. 2026-10-08 native trial (blocked): The official marketplace contains plugin 0.12.0, but user settings have no enabled Security entry. A depth-three inventory of project, task, download and desktop roots found no CLAUDE-SECURITY report directories. Other locations and project-specific settings remain uninspected. No scan was started. | 1 |
| Anthropic | Claude Tag in Slack | cloud, manual-export | Slack messages and Anthropic session transcripts are different histories. Edited versions, retrieved context, tools, attachments, routines, memory and action attribution need dedicated acquisition. Owner network exports omit Git and MCP traffic. The inspected Slack member has no connected Claude account or discovered bot messages. | research; 5 open gates. 2026-10-08 native trial (blocked): Native Slack shows the installed Claude app, but Home requires account connection. The personal DM is empty and a bot-specific workspace search returned no results. The connected Slack API exposes a different workspace. No full Tag transcript or Owner Activity record was reachable. | 1 |
| ByteDance | TRAE Work web | cloud, browser, manual-export | The official site links this web application, but its static response is only an application shell. Cloud enumeration, retained history and overlap with decrypted local TRAE Work caches remain unverified. | research; 5 open gates. 2026-10-08 native trial (blocked): Hosted SOLO/Work history has no registered acquisition route. Local desktop history is encrypted, and no authorized hosted history was available for a separate trial. | 1 |
| CloudStack (Msty) | Msty Go | local-store, manual-export | Desktop task agents, Mission Control helpers, scheduled jobs, approvals, and Memory Bank require separate history accounting from Studio conversations. Mobile steering does not establish cloud execution. | research; 5 open gates. | 1 |
| GitHub | Claude on GitHub (cloud agent) | cloud, browser | GitHub-hosted Claude partner agent, separate from Copilot and the vendor-owned cloud product. Verify GitHub account, repository, agent identity, iteration/PR lineage, model selection, event history, billing and artifact retention. | research; 5 open gates. | 1 |
| GitHub | Codex on GitHub (cloud agent) | cloud, browser | GitHub-hosted Codex partner agent, separate from Copilot and the vendor-owned cloud product. Verify GitHub account, repository, agent identity, iteration/PR lineage, model selection, event history, billing and artifact retention. | research; 5 open gates. | 1 |
| GitHub | GitHub Copilot app (cloud sandbox) | cloud, browser | App sessions can execute in cloud sandboxes. Do not infer a local transcript or equate them with issue-assigned coding-agent logs; verify cloud enumeration, agent identity, worktrees, approvals, canvases and retained events. | research; 5 open gates. | 1 |
| GitHub | GitHub Copilot app (local) | local-store, manual-export | Desktop app built on Copilot CLI with local sandboxes, worktrees, chats, automations and canvases. Shared runtime does not prove shared session roots or complete app metadata coverage; native app attribution and artifact retention are unverified. | research; 5 open gates. | 1 |
| GitHub | GitHub Copilot Chat on github.com | browser, cloud, manual-export | Web chat is separate from the coding agent, CLI and VS Code history. | research; 5 open gates. | 1 |
| GitHub | GitHub Copilot SDK (cloud) | cloud, sdk-events | GitHub-hosted sessions use cloud-agent entitlements and optional repository context. Streaming must be enabled for message deltas, and the remote worker must start before the first prompt. Remote Control of a local runtime is a different surface. | research; 5 open gates. | 1 |
| GitHub | GitHub Copilot SDK (local) | local-store, sdk-events | SDK sessions use a local Copilot runtime. Origin, application identity, machine prompts, child runs and ephemeral versus persisted event stores need verification independently of interactive CLI history. | research; 5 open gates. | 1 |
| GitHub | GitHub Copilot SDK Remote Control | local-store, cloud, sdk-events | The SDK shares a locally executing session to GitHub web/mobile. A remote URL is transport evidence, not proof of GitHub-hosted compute. Preserve local origin and verified cross-route identity. | research; 5 open gates. | 1 |
| Firebase App Testing agent | cloud, manual-export | Preview Android test-generation and execution agent. YAML test definitions are not run history; preserve test/run/step identity, actions, assertions, screenshots, recordings, logs, versions and failures independently of Gemini in Firebase chat. | research; 5 open gates. | 1 | |
| Gemini Spark | cloud, browser, local-store | Persistent personal cloud agent using the Antigravity harness, with desktop integration also announced. Antigravity transcript support does not establish Spark account history, scheduled/background work, approvals or local-versus-cloud execution coverage. | research; 5 open gates. | 1 | |
| Google AI Studio | browser, manual-export | Saved prompts, app-building sessions and Gemini chat are different stores. Inspect native exported prompts and attachments. | research; 5 open gates. | 1 | |
| Manus | Manus | cloud, browser, manual-export | Inspect tasks/messages/files and ownership evidence; avoid treating task summaries as full transcripts. | research; 5 open gates. | 1 |
| Microsoft | Microsoft Copilot consumer chat | browser, manual-export | The updated experience separates personal and organizational histories. Current authenticated inventory, conversation, feature and resource contracts remain unavailable behind Sign into Copilot. Privacy Dashboard CSV exports exist, but native columns and identities remain unverified. Legacy endpoints and GitHub Copilot importers do not establish current consumer support. | research; 9 open gates. 2026-10-08 native trial (blocked): Read-only Arc inspection reached copilot.com and displayed Sign into Copilot. The landing page separates account histories and names Notebook, Cowork, Researcher and Guided Learning. Public unauthenticated scripts supplied no usable history contract. Official guidance confirms distinct personal-account CSV categories, but no authenticated history or export was acquired. This is an access gate, not provider-side impossibility. | 1 |
| MiniMax | MiniMax Agent | browser, manual-export | Do not conflate consumer Agent tasks with MiniMax Code local or cloud sessions. | research; 5 open gates. | 1 |
| Mistral | Mistral Vibe remote coding agents | cloud, browser, manual-export | Cloud sessions can start from CLI or Vibe chat and move between local and cloud execution. Preserve execution transitions, repository/diff context, attempts, and original identity without claiming local logs contain the complete hosted run. | research; 5 open gates. | 1 |
| OpenAI | ChatGPT for Word | local-store, cloud, manual-export | Officially announced Word add-in for proofreading, suggested edits and formatting checks. ChatGPT history coverage does not establish document-scoped conversation, tracked-change, attachment or cross-app identity preservation. | research; 5 open gates. | 1 |
| OpenAI | ChatGPT in Slack and Microsoft Teams | cloud, manual-export | Service-account surfaces have workspace/channel identity and approved connections distinct from participants’ private ChatGPT accounts. Codex cloud delegation from Slack runs as the requesting user. Preserve parent messages and task links without merging by display name. | research; 5 open gates. | 1 |
| OpenAI | ChatGPT Space agent interactions | cloud, browser, manual-export | Page-side conversations, inline requests, comments, agent replies and accepted document changes are separate evidence. Shared pages do not share private chats or memory. Preserve page/version/action links and agent identity. | research; 5 open gates. | 1 |
| OpenAI | ChatGPT Team Tasks | cloud, browser, manual-export | Scheduled and event-triggered runs use a team service account and workspace connections. Preserve team/task/run identity, configured-versus-recorded models, generated files and workspace-credit billing. Runs do not inherit the creator’s private history or become human activity. | research; 5 open gates. | 1 |
| OpenAI | ChatGPT Workspace Agents | cloud, browser, manual-export | Published agents can receive asynchronous ChatGPT API-channel triggers with separate access tokens. Trigger submission is not a history export. Verify agent versions, run enumeration, event retention and overlap with Team Tasks before any deduplication. | research; 5 open gates. | 1 |
| OpenAI | Codex Security CLI | local-store, manual-export | Default scan bundles and workbench.sqlite3 live under Codex Security plugin state; CODEX_SECURITY_STATE_DIR overrides that root. Native evidence must reconcile shared workbench identity with plugin scans and distinguish partial coverage from completed scans. | research; 5 open gates. | 1 |
| OpenAI | Codex Security Cloud | cloud, manual-export | Existing Codex cloud transcripts do not establish coverage of scans, monitoring configuration, findings, artifacts or generated patches and draft-PR links. | research; 5 open gates. | 1 |
| OpenAI | Codex Security local plugin and workbench | local-store, manual-export | Ordinary Codex transcripts may retain scan activity but do not preserve the complete sealed scan directory, workbench triage state or cross-scan finding identity. JSON, CSV and SARIF exports cover different fields. | research; 5 open gates. | 1 |
| OpenAI | OpenAI Agent Builder | cloud, browser, manual-export | Preview runs and node traces need historical preservation separately from workflow definitions. Official documentation schedules shutdown for 2026-11-30. Exporting Agents SDK code or recreating a Workspace Agent does not export past conversations or run history. | research; 5 open gates. | 1 |
| OpenAI | OpenAI Agents API | cloud, manual-export | Managed Codex sessions use API-project identity, independently of Codex SDK and ChatGPT cloud tasks. Streams do not replay missed events; saved items omit some intermediate events. Trace export requires organization enablement and may arrive later. Usage is best effort, lacks cache-write counts and is not a final bill; command-output truncation is not reported. | research; 5 open gates. | 1 |
| OpenAI | OpenAI Agents SDK applications | local-store, cloud, manual-export | Application-owned multi-agent workflows, sessions and traces are distinct from the Codex SDK CLI harness. Inspect the application’s persistence and tracing policy, handoffs, tool runs and attachments; no universal retained-history store is established. | research; 5 open gates. | 1 |
| Perplexity | Comet assistant | browser, desktop | Verify whether browser-side tasks are ordinary Perplexity threads or a separate store; shared branding is insufficient. | research; 5 open gates. | 1 |
| Replit | Replit Agent | browser, manual-export, cloud | Project source exports alone are not conversation exports; retain task history, checkpoints, tools and artifacts. | research; 5 open gates. | 1 |
| Replit | Replit Chat | cloud, browser, manual-export | General chat, research, connected context, and Routines can precede conversion into Projects. Preserve chat-to-project lineage and scheduled runs; Replit Agent project exports do not establish this coverage. | research; 5 open gates. | 1 |
| Vercel | v0 | cloud, browser, manual-export | Evaluate documented chats/messages API with pagination, versions, attachments and access boundaries. | research; 5 open gates. | 1 |
| Vercel | Vercel Agent | cloud, browser, manual-export | Dashboard and Slack investigations, approved actions, and code review need team/project identity, plans, approvals, tool evidence, deployments and pull-request lineage. v0 conversations are separate. | research; 5 open gates. | 1 |
| Y Combinator | QM (self-hosted) | database-copy, cloud, browser | Durable deployments use PostgreSQL; the default in-memory store disappears on restart. Preserve scoped session identity, append-only tape, harness changes, approvals, compaction, source roles, artifacts and request usage. Viewer history is an audience-filtered projection. Codex, Claude Code, Pi and OpenCode backends do not imply QM coverage. | research; 5 open gates. | 1 |
| Alibaba | Qwen Code DashScope Batch jobs | cloud, manual-export | The current official CLI exposes batch submit/manage workflows through the DashScope Batch API. Verify input/output records, job lifecycle, partial failures, request/account identity and usage without counting machine batch prompts as human conversations. | research; 5 open gates. | 2 |
| Alibaba | Qwen Code Desktop | desktop-cache, local-store | The official macOS, Windows and Linux app is separate from the consumer Qwen chat product. Verify configured homes, native client identity, workspaces, attachments and overlap with local CLI JSONL before claiming desktop coverage. | research; 5 open gates. | 2 |
| Alibaba | Qwen Code editor and ACP integrations | local-store, acp-events | Official VS Code, Zed and JetBrains interfaces require separate discovery and origin checks. A retained CLI-compatible session may overlap with qwen-code, but client-owned state and ACP streams are not proven by the CLI trial. | research; 5 open gates. | 2 |
| Alibaba | Qwen Code messaging channels | local-store, local-api | Official docs list Telegram, DingTalk, WeChat, Feishu, QQ, WeCom, GitHub, GitLab and DingTalk Workspace channels. Verify per-user/group/workspace session scope, platform message identity, attachments, webhook-triggered jobs and human-versus-machine prompts. A channel transports requests to Qwen Code; it does not establish a vendor-hosted runtime. | research; 5 open gates. | 2 |
| Alibaba | Qwen Code SDK applications | local-store, sdk-events | Official TypeScript, Python and Java SDKs are documented separately. Verify durable versus ephemeral transcripts, application and subagent identity, machine-prompt accounting, and overlap with the CLI store. | research; 5 open gates. | 2 |
| Alibaba | Qwen Code Web UI and daemon | local-store, local-api, manual-export | The browser UI controls a local or explicitly selected daemon; browser access is not hosted execution. A native 0.24.7 probe reached the end of workspace-qualified persisted paging and read JSON/JSONL exports. Both omitted embedded image bytes; exports also omitted per-message models and some usage. Export channel labels do not establish original client identity. Capability, workspace, cursor and archived-source boundaries need explicit handling. | research; 5 open gates. | 2 |
| Amazon | Amazon Q Business | cloud, manual-export | Business conversations have a different account/application API from coding agents. | research; 5 open gates. | 2 |
| Amazon | Amazon Q Developer CLI historical history | local-store | Check migration into the Kiro classic SQLite reader with actual historical artifacts; do not install a retired client for current integration. | research; 5 open gates. | 2 |
| Amazon | Amazon Q Developer IDE | ide | Kiro imports do not establish Q Developer IDE chat coverage. | research; 5 open gates. | 2 |
| Amazon | Kiro Mobile | cloud, manual-export | Early-access iOS client uses the same cloud sessions as Kiro Web. This is a separate access surface, not evidence of a new conversation identity or an implemented mobile export. | research; 5 open gates. 2026-10-08 native trial (blocked): Dedicated mobile acquisition is unsupported. No mobile history or native export was supplied; local Kiro stores cannot establish mobile coverage. | 2 |
| Amazon | Kiro Web | browser, cloud, manual-export | Vendor documentation exposes browser-based multi-repository agents. Verify whether native exports include all web sessions and whether the cloud identity is shared with Kiro local sessions. | research; 5 open gates. 2026-10-08 native trial (blocked): Dedicated web acquisition is unsupported. No authorized hosted session history was available; local Kiro stores cannot establish web coverage. | 2 |
| Anthropic | Claude Tag in Microsoft Teams (announced) | cloud | The official product page still offers the Teams waitlist. The inspected Mac has no Teams app or com.microsoft.teams2 store. No tenant history or export was reachable; neither Slack installation nor its format establishes Teams support. | research; 5 open gates. 2026-10-08 native trial (blocked): Neither the Teams application nor its sandbox container was present. Current official documentation still directs Teams users to a waitlist. Native tenant availability, history and capture remain unverified. | 2 |
| Apple | Xcode coding intelligence | ide | Separate native Xcode chats from externally hosted Codex or Claude runs; do not count a shared model twice. | research; 5 open gates. | 2 |
| ByteDance | Coze | cloud, manual-export | Bot conversations, workflows and development runs need separate identifiers and capture contracts. | research; 5 open gates. | 2 |
| ByteDance | Doubao | browser, manual-export | Consumer chats are separate from TRAE stores. Export, media and account boundaries need a native sample. | research; 5 open gates. | 2 |
| Chatbox | Chatbox | local-store, manual-export | Inspect native exported sessions, local attachment storage and provider metadata. | research; 5 open gates. | 2 |
| ChatHub | ChatHub | browser, local-store, manual-export | Multi-model conversations span web, extension, desktop and mobile clients. Preserve each response model, side-by-side grouping, files and image output; a link from Chatbox does not establish common ownership or shared history. | research; 5 open gates. | 2 |
| Cherry Studio | Cherry Studio | local-store, manual-export | Database backup may contain credentials; preserve only allowlisted conversation and attachment data. | research; 5 open gates. | 2 |
| Forge | Forge | local-store, cloud | Determine retained conversation storage; configuration and analytics files do not prove transcript availability. | research; 5 open gates. | 2 |
| Genspark | Genspark | browser, manual-export | Hosted agent tasks and generated artifacts need a retained-history capture contract. | research; 5 open gates. | 2 |
| Firebase Studio | browser, cloud, manual-export | Cloud workspace chat and prototyping sessions require account/workspace enumeration and task history. | research; 5 open gates. | 2 | |
| Flow | browser, cloud, manual-export | Creative project prompts, generated assets and revisions are a distinct artifact-oriented surface. Conversation import scope and export fidelity require review. | research; 5 open gates. | 2 | |
| Gemini Code Assist | ide, cloud | IDE chat history requires its own store contract; Antigravity and historical Gemini CLI are not substitutes. | research; 5 open gates. | 2 | |
| Gemini Enterprise and Agentspace | cloud, manual-export | Organization search and agent histories need tenant-specific APIs and retention evidence. | research; 5 open gates. | 2 | |
| Gemini in Firebase | cloud, manual-export | Console and Firebase CLI assistance require separate history and retention verification from Firebase Studio workspaces. | research; 5 open gates. | 2 | |
| Gemini in Workspace | cloud, manual-export | Workspace side-panel conversations may have different retention and export policies from Gemini Apps. | research; 5 open gates. | 2 | |
| Opal | browser, manual-export | Workflow definitions and execution/chat histories must be distinguished. | research; 5 open gates. | 2 | |
| Stitch | browser, manual-export | Design conversation and generated assets need a distinct task/project model. | research; 5 open gates. | 2 | |
| Workspace Studio | cloud, manual-export | Flow definitions, agent activity, and execution histories are separate from Gemini chat. Takeout lists flows and activity logs; their native schema and import scope require validation. | research; 5 open gates. | 2 | |
| gptme | gptme Android | cloud-api, manual-export | Establish whether the client retains history bytes or only connects to a server; no Android-native capture or backup has been validated. | research; 5 open gates. | 2 |
| gptme | gptme Cloud | cloud-api, browser, manual-export | No account-wide hosted history acquisition, authentication contract, instance pagination, artifact capture or cross-route identity has been validated. | research; 5 open gates. | 2 |
| gptme | gptme Desktop | local-store, cloud-api | Locate the bundled local store and distinguish local execution from cloud sign-in; existing compatible-store parsing does not establish native desktop acceptance. | research; 5 open gates. | 2 |
| gptme | gptme Web UI (self-hosted) | local-store, cloud-api | Verify server instance identity, branches, artifacts, panels, browser/computer previews and API pagination; a local-store trial does not prove web UI acceptance. | research; 5 open gates. | 2 |
| Granola | Granola Chat | desktop, cloud, manual-export | Private chat threads about meetings, folders and the full note collection are separate from meeting transcripts. The documented CSV export contains notes, summaries and meeting transcripts; it does not establish chat-history export coverage. | research; 5 open gates. | 2 |
| Hugging Face | HuggingChat | browser, manual-export | Hosted chat history is separate from local model inference and Hub datasets. | research; 5 open gates. | 2 |
| Jan | Jan | local-store, manual-export | Read local threads and attachments with measured usage and model identity when recorded. | research; 5 open gates. | 2 |
| JetBrains | JetBrains AI Assistant chat | ide | Junie task-chain history does not establish coverage for general AI Assistant chats. | research; 5 open gates. | 2 |
| LibreChat | LibreChat | cloud, database-copy, manual-export | Multi-user conversation and agent history needs account-scoped capture and branch identity. | research; 5 open gates. | 2 |
| LM Studio | LM Studio chat | local-store, manual-export | Chat application history is separate from OpenAI-compatible API requests; inspect native persistence. | research; 5 open gates. | 2 |
| Lovable | Lovable | browser, manual-export | Project chat, edits, files and integration state need an explicit conversation export. | research; 5 open gates. | 2 |
| Microsoft | Copilot Studio | cloud, manual-export | Dataverse conversation transcripts and agent runs need independent schema validation. | research; 5 open gates. | 2 |
| Microsoft | Microsoft 365 Copilot | cloud, manual-export | Work account interactions, app context and compliance exports need their own identity and source contract. | research; 5 open gates. | 2 |
| MiniMax | Talkie character chat | browser, manual-export | Consumer character conversations are separate from MiniMax Code and Agent. The public data-export page lists profile fields without promising chat history; inspect native exports, character/thread identity, memories, retries and media. | research; 5 open gates. | 2 |
| MiniMax | Talkie Claw | cloud, browser, manual-export | Proactive character companions span messaging channels and generated media. Verify conversation/channel identity, background messages, attachments, and diamond billing independently of ordinary Talkie chat and OpenClaw local logs. | research; 5 open gates. | 2 |
| Mintplex Labs | AnythingLLM Desktop Assistant | local-store, manual-export | Desktop Assistant threads are saved in the Assistant Chat workspace. Preserve screenshots, tools and originating interface when recorded; do not duplicate shared workspace threads or treat screenshots as underlying application data. | research; 5 open gates. | 2 |
| Mintplex Labs | AnythingLLM Meeting Assistant | local-store, manual-export | Meeting transcripts and generated summaries are distinct from agent follow-up actions. Ask Questions creates a workspace containing the transcript; preserve links, source revisions, approvals and tool results without treating meeting speech as chat prompts. | research; 5 open gates. | 2 |
| Mintplex Labs | AnythingLLM Mobile | local-store, manual-export | Android is a standalone harness even when inference uses desktop or cloud providers. JSON/Markdown exports, on-device threads, forks, scheduled jobs and optional instance sync need native schema and cross-route identity checks. | research; 5 open gates. | 2 |
| Mintplex Labs | AnythingLLM scheduled jobs | local-store, cloud, manual-export | Run records preserve configured prompt, thinking, tool arguments/results, files, clocks and recorded token counters. The UI shows the latest 50 runs. Continue in Thread copies prompt and final response, not the full run; retain lineage and avoid double-counting human activity. | research; 5 open gates. | 2 |
| Mintplex Labs | AnythingLLM workspace chat | local-store, cloud, manual-export | Desktop, self-hosted and cloud workspace chats expose CSV, JSON, Alpaca and OpenAI fine-tune exports with account permissions. Native database/API fidelity, agent tools, branches, attachments and account boundaries remain unverified. Training exports may omit conversation evidence. | research; 5 open gates. | 2 |
| Mistral | Mistral Vibe for code in IDEs | local-store, manual-export | Official pages distinguish VS Code, JetBrains, and Zed access. Verify host, ACP transport, retained store, and local/remote session identity; the Vibe CLI importer alone does not establish these surfaces. | research; 5 open gates. | 2 |
| Notion | Notion Agent | desktop, browser, cloud, manual-export | Personal agent chat, tool actions, model selection, page context and connected sources need a native history contract. Workspace document exports are not proof of conversation coverage. | research; 5 open gates. | 2 |
| Notion | Notion Custom Agents | desktop, browser, cloud | Triggered and scheduled agent runs, interactive chats, activity logs, approval history and configuration versions are separate from personal Notion Agent conversations. | research; 5 open gates. | 2 |
| oh-my-pi | oh-my-pi | local-store | Separate data root and diverged schema; the pi importer is not sufficient. | research; 5 open gates. | 2 |
| Open Interpreter | Interpreter Desktop | local-store, manual-export | The document workstation is separate from the terminal agent. Verify native conversation stores, Word/Excel/PDF edits, browser/computer tools, attachments and local versus hosted model evidence. | research; 5 open gates. | 2 |
| Open Interpreter | Open Interpreter terminal | local-store, manual-export | Inspect retained terminal conversation logs and computer/tool semantics by client generation. The current terminal agent and Interpreter Desktop workstation have separate source repositories; neither has an implemented importer. | research; 5 open gates. | 2 |
| Open WebUI | Open WebUI | cloud, database-copy, manual-export | Use account-scoped exports/API before database reads; preserve branches, tools and files. | research; 5 open gates. | 2 |
| OpenClaw Foundation | Remote OpenClaw Gateway | cloud, manual-export | The existing importer requires a locally accessible Gateway state directory; remote history API is a separate route. | research; 5 open gates. | 2 |
| Qoder | Qoder Agent SDK | browser, cloud, manual-export | Vendor documents an SDK separately from its CLI. Verify transcript persistence and machine-run provenance before claiming coverage through the CLI reader. | research; 5 open gates. 2026-10-08 native trial (blocked): Standalone SDK classification has no registered route. No SDK history was supplied, and inspected local Qoder roots contain no conversation candidates. | 2 |
| StackBlitz | Bolt | browser, manual-export | Capture project chat and revision/checkpoint history independently of exported source code. | research; 5 open gates. | 2 |
| Tencent | Yuanbao | browser, manual-export | CodeBuddy coding sessions do not cover the consumer assistant. | research; 5 open gates. | 2 |
| Vercel | eve agent framework | cloud, local-store, manual-export | History belongs to each deployed application and its access policy. Durable session IDs differ from executing workflow runs and channel continuation tokens. Replay must preserve stream versions, delta/final semantics, retry ambiguity, compaction and child tasks. | research; 5 open gates. | 2 |
| xAI | Grok in X | browser, manual-export | Verify whether X history and grok.com history share account and conversation identity before joining captures. | research; 5 open gates. | 2 |
| You.com | You.com | browser, manual-export | Search, research and custom-agent histories need native account export evidence. | research; 5 open gates. | 2 |
| CloudStack (Msty) | Msty Nexus model testing | local-store, manual-export | Nexus includes interactive model, preset, and route testing inside its gateway console. Verify whether tester conversations persist. Gateway request logs and current routing settings cannot stand in for consumer history or historical reply-model attribution. | research; 5 open gates. | 3 |
| CloudStack (Msty) | Msty Studio | local-store, browser, manual-export | Official exports contain conversation/split structure and either embedded attachment data in JSON or files in .mstyconv ZIPs. Branches, crew/persona replies, tools and missing older attachments require native schema validation. Go and Nexus are separate. | research; 5 open gates. | 3 |
| Cognition | DeepWiki chat | browser | Repository documentation Q&A may be separate from Ask Devin. Inspect persistent identity and export availability. | research; 5 open gates. | 3 |
| Dify | Dify | cloud, database-copy, manual-export | Chat messages, workflow runs and application telemetry require separate event identities. | research; 5 open gates. | 3 |
| Flowise | Flowise | cloud, database-copy, manual-export | Chat history and workflow traces require versioned contracts and credential-safe raw preservation. | research; 5 open gates. | 3 |
| Gemini in Android Studio | ide | Verify Android Studio chat persistence separately from Code Assist. | research; 5 open gates. | 3 | |
| n8n | n8n AI agents | cloud, database-copy, manual-export | Workflow execution records are not automatically human conversations; preserve machine-run classification. | research; 5 open gates. | 3 |
| Nous Research | Remote Hermes profiles | cloud, manual-export | Local profile SQLite coverage does not prove remote-only account enumeration. | research; 5 open gates. | 3 |
| OpenHands | Remote self-hosted OpenHands | cloud, manual-export | Local store and hosted Cloud adapters exist; verify arbitrary self-hosted API compatibility and auth scope. | research; 5 open gates. 2026-10-08 native trial (blocked): No mounted remote history or authorized remote runtime was supplied. The separate remote surface remains unsupported; default local discovery found no retained history. | 3 |
| Plandex | Plandex | cloud, manual-export | Inspect server-side plan/conversation exports and retained version history. | research; 5 open gates. | 3 |
| Posit | Databot (historical) | local-store, manual-export | Historical exploratory data-analysis agent superseded by Posit Assistant. Native legacy storage and migration continuity remain unverified. | research; 5 open gates. | 3 |
| Posit | Positron Assistant (historical) | local-store, manual-export | Superseded by Posit Assistant as of Positron 2026.07. Preserve historical stores independently; shared branding does not establish schema or identity compatibility. | research; 5 open gates. | 3 |
| SWE-agent | mini-SWE-agent | local-store | Versioned task trajectories need separate job classification and cost accounting. | research; 5 open gates. | 3 |
| SWE-agent | SWE-agent | local-store | Trajectory files are machine-prompted jobs, not human conversation activity. | research; 5 open gates. | 3 |
| Xiaomi | MiMo hosted chat or playground | browser, manual-export | A model/API brand is not proof of durable account history. Inspect the actual hosted history before defining a source. | research; 5 open gates. | 3 |
Format and drift coverage
Section titled “Format and drift coverage”The source-shape registry declares 76 enforced, 73 observation-only, 2 unsupported, and 1 inapplicable contracts. Enforced contracts reject unfamiliar structural cohorts. Observation-only contracts record shape evidence without establishing exhaustive field coverage. An unsupported shape contract can accompany a raw-preservation route.
Shape policy, raw replay support, and import availability are separate properties. A matching shape does not prove every metadata field is normalized or displayed. Tests associated with every implemented product are listed in the machine-readable catalog; a fixture pass does not establish native account access or acceptance.
Bounded JSON field accounting is declared for 53 formats, and selected protobuf schema accounting for 6 formats. 93 formats have no declared field audit. These audits cover only the scopes listed below; none establishes complete provider metadata fidelity.
| Source format | Parser version | Product entry | Shape policy | Field audit | Raw replay |
|---|---|---|---|---|---|
acp-runtime-history |
0.8.0.0 |
acp-runtime | observe | Not declared | supported |
aider-chat-history |
0.8.0.0 |
aider | enforce | Not declared | supported |
amp-cloud-history |
0.8.0.0 |
amp-cloud | enforce | Not declared | deferred |
amp-thread-export |
0.8.0.0 |
amp-export | enforce | Not declared | deferred |
amp-thread-history |
0.8.0.0 |
amp | enforce | Not declared | deferred |
antigravity-cli-brain |
0.8.0.0 |
antigravity-cli | observe | Not declared | deferred |
antigravity-cli-conversation-db |
0.8.0.0 |
antigravity-cli | observe | Selected protobuf schema | deferred |
antigravity-cli-language-server-steps |
0.8.0.0 |
antigravity-cli | observe | Selected protobuf schema | deferred |
antigravity-cli-transcript-log |
0.8.0.0 |
antigravity-cli | observe | Bounded JSON fields | supported |
antigravity-conversation-db |
0.8.0.0 |
antigravity | observe | Selected protobuf schema | deferred |
antigravity-history |
0.8.0.0 |
antigravity | observe | Not declared | deferred |
antigravity-ide-brain |
0.8.0.0 |
antigravity-ide | observe | Not declared | deferred |
antigravity-ide-conversation-db |
0.8.0.0 |
antigravity-ide | observe | Selected protobuf schema | deferred |
antigravity-ide-language-server-steps |
0.8.0.0 |
antigravity-ide | observe | Selected protobuf schema | deferred |
antigravity-ide-transcript-log |
0.8.0.0 |
antigravity-ide | observe | Bounded JSON fields | supported |
antigravity-language-server-steps |
0.8.0.0 |
antigravity | observe | Selected protobuf schema | deferred |
antigravity-protobuf |
0.8.0.0 |
antigravity | unsupported | Not declared | unsupported |
antigravity-summary-proto |
0.8.0.0 |
antigravity | observe | Not declared | deferred |
antigravity-trajectory-summary |
0.8.0.0 |
antigravity | observe | Not declared | deferred |
antigravity-transcript-log |
0.8.0.0 |
antigravity | observe | Bounded JSON fields | supported |
augment-cli-history |
0.8.0.0 |
augment | enforce | Not declared | deferred |
augment-cloud-history |
0.8.0.0 |
augment-cloud | enforce | Not declared | deferred |
augment-cosmos-history |
0.8.0.0 |
augment | enforce | Not declared | deferred |
augment-ide-history |
0.8.0.0 |
augment | enforce | Not declared | deferred |
chatgpt-browser |
0.8.0.0 |
chatgpt, chatgpt-work-cloud | observe | Bounded JSON fields | supported |
chatgpt-dots-browser |
0.8.0.0 |
chatgpt-dots | observe | Bounded JSON fields | supported |
chatgpt-export |
0.8.0.0 |
chatgpt | observe | Not declared | deferred |
chatgpt-messages-history |
0.8.0.0 |
chatgpt-messages | enforce | Not declared | supported |
claude-cloud-history |
0.8.0.0 |
claude-cloud, claude-remote-control | observe | Bounded JSON fields | supported |
claude-code-desktop-metadata |
0.8.0.0 |
claude-code-desktop | observe | Bounded JSON fields | supported |
claude-cowork-cloud-history |
0.8.0.0 |
claude-cowork-cloud, claude-unified, claude-design | observe | Bounded JSON fields | supported |
claude-design-browser |
0.8.0.0 |
claude-design | observe | Bounded JSON fields | supported |
claude-duet-history |
0.8.0.0 |
claude-duet | observe | Not declared | supported |
claude-science-history |
0.8.0.0 |
claude-science | enforce | Bounded JSON fields | supported |
claude-sdk-history |
0.8.0.1 |
claude-sdk | observe | Bounded JSON fields | supported |
claude-web-browser |
0.8.0.0 |
claude-web, claude-unified | observe | Bounded JSON fields | supported |
claude-web-export |
0.8.0.0 |
claude-web, claude-unified, claude-design | observe | Bounded JSON fields | supported |
claude-web-memory |
0.8.0.0 |
claude-web | observe | Not declared | deferred |
claude-workspace-desktop |
0.8.0.0 |
claude-cowork-cloud, claude-cowork | observe | Bounded JSON fields | supported |
cli-history |
0.8.0.1 |
claude, claude-remote-control | observe | Bounded JSON fields | supported |
cline-sdk-session-history |
0.8.0.0 |
cline-cli | enforce | Not declared | supported |
cline-task-history |
0.8.0.0 |
cline | enforce | Not declared | supported |
codebuddy-history |
0.8.0.0 |
codebuddy | observe | Bounded JSON fields | supported |
codex-cli-history |
0.8.0.0 |
codex-cli | enforce | Bounded JSON fields, Bounded JSON fields, Bounded JSON fields | supported |
codex-cloud-browser |
0.8.0.1 |
codex-cloud | observe | Not declared | supported |
codex-cloud-history |
0.8.0.0 |
codex-cloud | observe | Not declared | deferred |
codex-desktop-history |
0.8.0.0 |
codex-desktop, chatgpt-work-local | enforce | Bounded JSON fields, Bounded JSON fields, Bounded JSON fields | supported |
codex-dot-task-browser |
0.8.0.1 |
codex-cloud | observe | Not declared | supported |
codex-sdk-history |
0.8.0.0 |
codex-sdk | enforce | Bounded JSON fields, Bounded JSON fields, Bounded JSON fields | supported |
continue-history |
0.8.0.0 |
continue | enforce | Not declared | deferred |
copilot-agent-session |
0.8.0.0 |
copilot-agent-cloud | observe | Not declared | deferred |
copilot-cli-history |
0.8.0.0 |
copilot-cli | observe | Bounded JSON fields | supported |
copilot-ide-history |
0.8.0.0 |
copilot-ide | observe | Not declared | supported |
crush-history |
0.8.0.0 |
crush | enforce | Not declared | deferred |
cursor-agent-transcripts |
0.8.0.0 |
cursor | observe | Not declared | supported |
cursor-cli-history |
0.8.0.0 |
cursor-cli | observe | Not declared | supported |
cursor-cloud-history |
0.8.0.0 |
cursor-cloud | observe | Bounded JSON fields | deferred |
cursor-global-sqlite |
0.8.0.0 |
cursor | observe | Not declared | deferred |
cursor-raw-sqlite-salvage |
0.8.0.0 |
cursor | observe | Not declared | deferred |
cursor-workspace-sqlite |
0.8.0.0 |
cursor | observe | Not declared | deferred |
deepseek-chat-browser |
0.8.0.0 |
deepseek-chat | observe | Not declared | supported |
deepseek-harness-history |
0.8.0.0 |
deepseek-harness | enforce | Bounded JSON fields | deferred |
devin-ask-browser |
0.8.0.0 |
ask-devin | observe | Not declared | supported |
devin-cli-history |
0.8.0.0 |
devin-cli | observe | Not declared | supported |
devin-cloud-browser |
0.8.0.0 |
devin-cloud | observe | Bounded JSON fields | supported |
devin-cloud-history |
0.8.0.0 |
devin-cloud | observe | Not declared | deferred |
devin-desktop-acp-events |
0.8.0.0 |
devin-desktop | observe | Not declared | supported |
devin-desktop-acp-messages |
0.8.0.0 |
devin-desktop | observe | Not declared | deferred |
factory-cloud-history |
0.8.0.0 |
factory-cloud | observe | Not declared | deferred |
factory-droid-history |
0.8.0.0 |
factory | observe | Bounded JSON fields | supported |
fx-cli-history |
0.8.0.0 |
fx | observe | Bounded JSON fields | supported |
gemini-cli-history |
0.8.0.0 |
gemini-cli | enforce | Not declared | supported |
gemini-web-browser |
0.8.0.0 |
gemini-web | observe | Not declared | supported |
gemini-web-export |
0.8.0.0 |
gemini-web | observe | Bounded JSON fields | deferred |
goose-jsonl-history |
0.8.0.0 |
goose | enforce | Not declared | deferred |
goose-sqlite-history |
0.8.0.0 |
goose | enforce | Bounded JSON fields | deferred |
gptme-history |
0.8.0.0 |
gptme | enforce | Bounded JSON fields | supported |
grok-bot-replica |
0.8.0.0 |
grok-bot | enforce | Bounded JSON fields | supported |
grok-build-history |
0.8.0.0 |
grok-build | observe | Bounded JSON fields | supported |
grok-chat-browser |
0.8.0.0 |
grok-chat | observe | Bounded JSON fields | supported |
hermes-jsonl-history |
0.8.0.0 |
hermes | enforce | Not declared | deferred |
hermes-sqlite-history |
0.8.0.0 |
hermes | enforce | Bounded JSON fields | deferred |
iflow-cli-history |
0.8.0.0 |
iflow | enforce | Not declared | supported |
import |
0.8.0.0 |
Generic normalized input, not a provider contract. | inapplicable | Not declared | unsupported |
instinct-messages-history |
0.8.0.0 |
instinct | enforce | Bounded JSON fields | supported |
jules-cloud-history |
0.8.0.0 |
jules | observe | Not declared | deferred |
junie-cli-history |
0.8.0.0 |
junie | enforce | Not declared | deferred |
junie-ide-history |
0.8.0.0 |
junie | enforce | Not declared | deferred |
kilo-code-cloud-history |
0.8.0.0 |
kilo-code-cloud | enforce | Not declared | deferred |
kilo-code-sqlite-history |
0.8.0.0 |
kilo-code | enforce | Not declared | deferred |
kilo-code-task-history |
0.8.0.0 |
kilo-code-legacy | enforce | Not declared | deferred |
kimi-chat-browser |
0.8.0.0 |
kimi-chat | observe | Not declared | supported |
kimi-cli-history |
0.8.0.0 |
kimi-cli | enforce | Bounded JSON fields | supported |
kimi-code-history |
0.8.0.0 |
kimi-code | enforce | Bounded JSON fields | supported |
kiro-classic-history |
0.8.0.0 |
kiro | enforce | Not declared | deferred |
kiro-cli-history |
0.8.0.0 |
kiro | enforce | Not declared | deferred |
kiro-crew-history |
0.8.0.0 |
kiro | enforce | Not declared | deferred |
kiro-ide-history |
0.8.0.0 |
kiro | enforce | Not declared | deferred |
kiro-session-export |
0.8.0.0 |
kiro-export | enforce | Not declared | deferred |
kiro-unified-history |
0.8.0.0 |
kiro | enforce | Not declared | deferred |
meta-ai-browser |
0.8.0.0 |
meta-ai | observe | Not declared | supported |
mimo-code-history |
0.8.0.0 |
mimo-code | enforce | Bounded JSON fields | supported |
minimax-code-cloud-history |
0.8.0.0 |
minimax-code-cloud | enforce | Bounded JSON fields | deferred |
minimax-code-history |
0.8.0.0 |
minimax-code | enforce | Bounded JSON fields | deferred |
mistral-chat-browser |
0.8.0.0 |
le-chat | observe | Not declared | supported |
muse-agent-export |
0.8.0.0 |
muse-consumer | enforce | Bounded JSON fields, Bounded JSON fields | supported |
muse-code-history |
0.8.0.0 |
muse-code | enforce | Not declared | supported |
notebooklm-browser |
0.8.0.0 |
notebooklm | observe | Not declared | supported |
notebooklm-export |
0.8.0.0 |
notebooklm | observe | Bounded JSON fields | supported |
openclaw-jsonl-history |
0.8.0.0 |
openclaw | enforce | Not declared | deferred |
openclaw-sqlite-history |
0.8.0.0 |
openclaw | enforce | Bounded JSON fields | deferred |
opencode-history |
0.8.0.0 |
opencode | enforce | Not declared | supported |
opencode-sqlite-history |
0.8.0.0 |
opencode | enforce | Bounded JSON fields | supported |
openhands-cloud-history |
0.8.0.0 |
openhands-cloud | enforce | Not declared | deferred |
openhands-legacy-history |
0.8.0.0 |
openhands | enforce | Not declared | deferred |
openhands-sdk-history |
0.8.0.0 |
openhands | enforce | Not declared | deferred |
perplexity-browser |
0.8.0.0 |
perplexity | observe | Bounded JSON fields | supported |
perplexity-computer-browser |
0.8.0.0 |
perplexity-computer | observe | Bounded JSON fields | supported |
pi-cli-history |
0.8.0.0 |
pi | enforce | Bounded JSON fields, Bounded JSON fields | supported |
poe-browser |
0.8.0.0 |
poe | observe | Not declared | supported |
poke-messages-history |
0.8.0.0 |
poke | enforce | Not declared | supported |
posit-assistant-history |
0.8.0.0 |
posit-assistant | enforce | Bounded JSON fields | supported |
prime-agent-history |
0.8.0.0 |
prime-agent | enforce | Bounded JSON fields | supported |
qoder-app-history |
0.8.0.0 |
qoder | enforce | Not declared | deferred |
qoder-cli-history |
0.8.0.0 |
qoder | enforce | Bounded JSON fields | deferred |
qoder-cloud-history |
0.8.0.0 |
qoder-cloud | enforce | Bounded JSON fields | deferred |
qoder-ide-history |
0.8.0.0 |
qoder | enforce | Not declared | deferred |
qoder-work-history |
0.8.0.0 |
qoder | enforce | Bounded JSON fields | deferred |
qwen-chat-browser |
0.8.0.0 |
qwen-chat | observe | Not declared | supported |
qwen-chat-export |
0.8.0.0 |
qwen-chat | observe | Not declared | supported |
qwen-code-history |
0.8.0.0 |
qwen-code | enforce | Bounded JSON fields, Bounded JSON fields | supported |
roo-code-cli-history |
0.8.0.0 |
roo-code-cli | enforce | Not declared | deferred |
roo-code-task-history |
0.8.0.0 |
roo-code | enforce | Not declared | deferred |
roomote-cloud-history |
0.8.0.0 |
roomote-cloud | enforce | Not declared | deferred |
roomote-db-history |
0.8.0.0 |
roomote-db | enforce | Not declared | deferred |
trae-agent-history |
0.8.0.0 |
trae | enforce | Not declared | deferred |
trae-ide-history |
0.8.0.0 |
trae | enforce | Not declared | deferred |
trae-legacy-history |
0.8.0.0 |
trae | enforce | Not declared | deferred |
trae-work-history |
0.8.0.0 |
trae | enforce | Not declared | deferred |
vibe-cli-history |
0.8.0.0 |
vibe | enforce | Bounded JSON fields, Bounded JSON fields | supported |
warp-cli-history |
0.8.0.0 |
warp | enforce | Not declared | deferred |
warp-cloud-history |
0.8.0.0 |
warp-cloud | enforce | Not declared | deferred |
warp-desktop-history |
0.8.0.0 |
warp | enforce | Not declared | deferred |
web-chat-export |
0.8.0.0 |
Generic historical export compatibility envelope, not a distinct product. | observe | Not declared | deferred |
windsurf-cascade-brain |
0.8.0.0 |
windsurf | observe | Not declared | deferred |
windsurf-cascade-protobuf |
0.8.0.0 |
windsurf | unsupported | Not declared | unsupported |
windsurf-language-server-transcript |
0.8.0.0 |
windsurf | observe | Not declared | supported |
windsurf-trajectory-export |
0.8.0.0 |
windsurf | observe | Not declared | supported |
windsurf-transcript-jsonl |
0.8.0.0 |
windsurf | observe | Not declared | supported |
zai-chat-browser |
0.8.0.0 |
zai-chat | observe | Not declared | supported |
zed-agent-history |
0.8.0.0 |
zed | enforce | Bounded JSON fields | supported |
zed-text-history |
0.8.0.0 |
zed | enforce | Not declared | deferred |
Field-audit scopes
Section titled “Field-audit scopes”These versioned contracts detect changes in the selected fields or declarations. Runtime normalization, preserved bytes, and viewer behavior still require separate checks. A parser version change requires review of every associated contract, and a scope change regenerates this reference.
| Source formats | Audit | Declared scope |
|---|---|---|
antigravity-conversation-db, antigravity-cli-conversation-db, antigravity-ide-conversation-db, antigravity-language-server-steps, antigravity-cli-language-server-steps, antigravity-ide-language-server-steps |
Selected protobuf schema | Selected usage and generic-result protobuf descriptors audited in the installed official Antigravity CLI and desktop language server. This is a schema field contract, not complete record or billing acceptance. |
antigravity-transcript-log, antigravity-cli-transcript-log, antigravity-ide-transcript-log |
Bounded JSON fields | Persistent transcript log envelopes observed in six retained CLI conversations. Other record shapes, generated-message sidecars, database payloads, and live capture remain separate acceptance boundaries. |
chatgpt-dots-browser |
Bounded JSON fields | Native direct-message room, profile, membership, task links, and widget records observed on 2026-10-08, plus synthetic raw-message and file shapes. Thread parent and visibility fields follow the desktop client inspected on 2026-10-28; synthetic tests cover their normalized projection. Widget internals remain raw-only. Calls, notes, other channels, and delegated task transcripts remain unverified. |
chatgpt-browser |
Bounded JSON fields | One native ChatGPT Work first-turn browser capture with seven messages, a generated sandbox-file link, command presentation summaries, and the subsequently acquired original file. Command arguments, command results, measured usage, cloud execution provenance, resumed turns, consumer chat features, GPT identity and Voice need separate native evidence. |
claude-cloud-history |
Bounded JSON fields | Seventeen native detail and event-page responses from six Remote Control bridge sessions acquired through the installed browser receiver. The cohort contains 3709 events. Browser transport envelopes, hosted non-bridge sessions, resumed attachments, and cross-route identity need separate evidence. Raw-only fields are retained without claiming complete normalized display. |
claude-cowork-cloud-history |
Bounded JSON fields | One retained native Cowork Design cache stream. This accounts for settled event fields and separately preserved transport fields. It does not establish account detail, API pagination, linked workspaces, current remote files, or standalone Design. |
claude-workspace-desktop |
Bounded JSON fields | Cowork audit result cost, bounded model usage and context limits, and rate_limit_event overage metadata. USD costs remain provider estimates and cumulative snapshots are differenced once. Native capture and complete audit-record fidelity remain unverified. |
claude-design-browser |
Bounded JSON fields | One native standalone Design project and chat acquired through Connect JSON. The cohort includes two conversation messages, two tool calls and results, current-file inventory, two verified files, and an unavailable thumbnail. Other projects, closed chats, provider revisions, direct edits, and additional message shapes need separate evidence. |
claude-web-export |
Bounded JSON fields | One native Design chat from an official account export, with two message wrappers, two tool calls and outputs, thinking, one skill attachment descriptor, usage, author attribution, and file-change metadata. This is a bounded Design export field audit, not accounting for ordinary conversation exports, memories, artifacts, feedback, or other Design shapes. |
cli-history |
Bounded JSON fields | Linked Desktop session toolSurfaceSnapshot hashes and version metadata only. Transcript content, complete Desktop metadata, and live control responses are outside this bounded audit. |
claude-code-desktop-metadata |
Bounded JSON fields | Desktop tool identity hashes and version 1/2 plan utilization history, including metadata-only records without transcripts. Samples describe the shared plan, never session charges. Account identifiers and tool definitions are excluded. This does not establish native capture or complete Desktop field coverage. |
claude-science-history |
Bounded JSON fields | Claude Science 0.1.56 bundled examples and local native tool and reviewer conversations on macOS. Harness notices, attachment receipts, reviewer metadata and bounded child summaries supplement tool lifecycles, image sidecars, output spills, and artifact revisions. A resumed PubMed connector and compute-inventory trial includes native memory rows. Remote compute jobs and routines need separate native trials. |
claude-sdk-history |
Bounded JSON fields | Linked Desktop session toolSurfaceSnapshot hashes and version metadata only. Transcript content, complete Desktop metadata, and live control responses are outside this bounded audit. |
claude-web-browser |
Bounded JSON fields | One retained native Claude.ai browser conversation with four text messages and a provider summary. The sample covers account identity, branch ancestry, conversation settings and message state. Files, tools, projects, linked workspaces, alternate branches and other source shapes require separate acceptance. |
codebuddy-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
codex-cli-history, codex-desktop-history, codex-sdk-history |
Bounded JSON fields | Observed tagged JSONL fields across CLI, desktop, exec, and child sessions. Tool dictionaries use wildcard keys because keys can contain private paths and identifiers. A field declaration is not a normalized-value or native acceptance assertion. |
codex-cli-history, codex-desktop-history, codex-sdk-history |
Bounded JSON fields | Observed title-index records. The importer archives this sidecar only when it supplies the selected session title. This does not establish complete index-history capture. |
codex-cli-history, codex-desktop-history, codex-sdk-history |
Bounded JSON fields | SQLite schema and row accounting from a coherent copied native database. BLOB values are represented as base64 strings for this private audit. Empty tables retain declared columns without invented value types. Raw preservation does not establish normalized support. |
copilot-cli-history |
Bounded JSON fields | Copilot CLI 1.0.89 JSONL events from an offline local OpenAI-compatible endpoint: initial and resumed text turns, view tool, and controlled bash failure. Endpoint tokens are injected test counters, not measured hosted usage. SQLite ledger columns, workspace/checkpoint sidecars, attachments, subagents, compaction, Remote Control, SDK, desktop and cloud histories require separate field audits. |
cursor-cloud-history |
Bounded JSON fields | Synthetic v0 conversation and v1 stream capture envelope. This field audit checks only the fixture shape used to test catalog enforcement. It is not native account evidence or a complete provider field audit. |
deepseek-harness-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
devin-cloud-browser |
Bounded JSON fields | Synthetic browser capture with string and object attachment descriptors, one unavailable file, and one recording annotation response. This is regression field accounting, not a native field census. Nonempty annotation records and other session or event fields need separate audits. |
factory-droid-history |
Bounded JSON fields | Droid 0.229.0 native text, Execute, Read, Create, Edit, controlled command and edit failures, embedded tool-result image and session-settings records. User-uploaded documents, workers, provider failures and hosted resume require separate evidence. |
fx-cli-history |
Bounded JSON fields | Bounded synthetic session-manager v2 fixture, native 0.0.17 seed, and three native-generated sessions against a local scripted endpoint. This inventory does not establish complete upstream or legacy field coverage. |
gemini-web-export |
Bounded JSON fields | Native Gemini Apps JSON and English HTML activity. HTML caption links supply chat identity, pre-clock subtitles describe media, and prompt whitespace is retained. HTML has second precision and can omit generated-file references that JSON supplies. Models, usage, alternate replies, and account-wide completeness remain unverified. |
goose-sqlite-history |
Bounded JSON fields | Goose CLI 1.52.0 SQLite initial and resumed synthetic conversations through a local OpenAI-compatible endpoint, including write/edit, a failed edit, a nonzero shell result and an image result. Usage counters are injected, not measured billing. Legacy JSONL, desktop/API interfaces, hosted providers, recipes, schedules, subagents, compaction, user uploads and other tool/block kinds need separate audits. |
gptme-history |
Bounded JSON fields | Observed gptme main, branch, view, recovery, scoped configuration, and runtime sidecar shapes. Metadata-only sidecars and provider cost provenance have synthetic regression coverage; hosted billing and installed runtime acceptance remain unverified. |
grok-bot-replica |
Bounded JSON fields | Replica schema 1 fields observed in Grok Bot 0.72.1, plus synthetic tool-call fields. Arbitrary tool payloads, other renderer cards, and future fields remain subject to separate field review. Authentication fields, private forms, and URL queries are deliberately excluded before persistence. |
grok-build-history |
Bounded JSON fields | Metadata sidecars only. Update, event, chat, workflow, resource and attachment schemas require separate contracts. A pass verifies declared field accounting, not normalization or raw fidelity. |
grok-chat-browser |
Bounded JSON fields | Rolling Grok web capture envelopes, response trees, selected-message bodies and paired tool output. Explicit paths cover observed native captures and synthetic code/X search cases. Original resource bytes and independent account exports are outside this route. Empty collections do not establish provider absence. Raw-only data is not claimed as rendered. |
hermes-sqlite-history |
Bounded JSON fields | Hermes schema 32 session and model-ledger billing plus persisted message outcome fields. The state database has no dedicated API retry/error or approval ledger. Hook payloads, complete transcript fidelity, and native acceptance are outside this scoped audit. |
instinct-messages-history |
Bounded JSON fields | Synthetic macOS Messages selected-row envelope v1. Covers seven-table fixtures and Foundation-generated typed text. Native Messages schema and field fidelity remain blocked on macOS access. |
kimi-cli-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
kimi-code-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
mimo-code-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
minimax-code-cloud-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
minimax-code-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
muse-agent-export |
Bounded JSON fields | Muse (Personal Agent) native agent data exports from macOS 6.0 and the web app. The native cohort contains user and assistant turns labeled main-channel, empty turn resources, bundled workspace files and an empty voice inventory. Export channel labels do not establish native conversation boundaries. Nonempty turn resources or voice inventory, unresolved workspace files, other roles and channels require further native validation. |
muse-agent-export |
Bounded JSON fields | Scoped Muse Personal Agent approval metadata and usage-footer text. Field names and decisions were checked in the installed Muse 7.0 bundle. Populated native acquisition remains unverified. |
notebooklm-export |
Bounded JSON fields | NotebookLM consumer Takeout notebook and text-source JSON metadata from one native prompt/reply export. Chat HTML, generated artifacts, binary sources, and other source types require separate contracts. Metadata for notebooks without chats is raw-only. |
openclaw-sqlite-history |
Bounded JSON fields | Runtime evidence fields reviewed against OpenClaw 2026.10.5, agent schema 31 and shared state schema 29. This scoped audit does not claim complete transcript or native runtime coverage. |
opencode-sqlite-history |
Bounded JSON fields | OpenCode 1.18.31 legacy message/part rows inside a v2-capable SQLite store. All populated table fields are accounted for. The privacy projection excludes transport, project, migration, account, and credential tables. |
perplexity-browser |
Bounded JSON fields | Browser capture envelopes. Explicit field paths cover retained and freshly acquired native captures, current-tab discovery metadata, acquired attachments, synthetic chart resources, and Search workflow asset associations. Table column names are dynamic. Raw-only fields are preserved, not claimed as rendered. A field absent from this sample is not provider-unavailable. |
perplexity-computer-browser |
Bounded JSON fields | Synthetic Computer workflow envelopes only. No native Computer payload was available. The explicit paths cover text, tools, nested subagents, statuses, step models, and inline artifacts. Other fields require a new audit. |
pi-cli-history |
Bounded JSON fields | pi coding-agent 1.1.0 additions from pinned upstream source: usage, context_edit, system messages, summary usage and compaction system checkpoints. Synthetic schema evidence only. No native 1.1.0 store was available. |
pi-cli-history |
Bounded JSON fields | pi coding-agent 0.73.1 v3 tree JSONL: initial and resumed text turns through a local OpenAI-compatible endpoint, current edit arrays, a failed edit, a failed bash call and an image result. Injected tokens and configured-rate costs are synthetic. Forks, compaction, extensions, uploads, thinking content, other APIs, hosted use and older producer versions require additional cohorts. |
posit-assistant-history |
Bounded JSON fields | Native v3 tree, LM/UI sidecars, workspace metadata, and excluded workspace index and execution state. Optional usage-events rows retain allowlisted keepalive and classifier usage separately from conversation totals. Nested children use explicit parentTreeId links. Auxiliary usage and nested relationships have synthetic coverage; broader native cohorts remain unverified. |
prime-agent-history |
Bounded JSON fields | Retained version 3 configuration-only sessions, semantic edges, and kernel-state JSON. No native user, assistant, tool, attachment, or child record occurred. Binary kernel state is verified separately by byte hash. |
qoder-cloud-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
qoder-cli-history, qoder-work-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
qwen-code-history |
Bounded JSON fields | Qwen Code 0.24.7 tree JSONL and version-one usage ledger: one initial and two resumed turns through a local OpenAI-compatible endpoint, file creation/edit, controlled edit/shell failures, a capability-rejected image and an embedded image result. Injected token counters are synthetic. Hosted metering, user uploads, extensions, subagents, branches, compaction, other backends and older producing versions require additional cohorts. |
qwen-code-history |
Bounded JSON fields | Qwen Code 0.25.0 additions from pinned upstream source: fifteen record subtypes, including managedSession envelopes. Synthetic schema evidence only. Managed transactions remain typed preservation without execution, commit validation or billing projection. No native 0.25.0 store was available. |
vibe-cli-history |
Bounded JSON fields | Synthetic stored runtime evidence fields only. Native provider capture and complete source-field coverage remain unverified. |
vibe-cli-history |
Bounded JSON fields | Vibe CLI 2.26.0 legacy and Unified Harness v1 minor 8 initial, resumed, image, model-change, failed-tool, and cancelled-approval trials with a synthetic local endpoint. No hosted billing, native branching, compaction, voice, or delegation coverage is implied. |
zed-agent-history |
Bounded JSON fields | Scoped Zed Agent and Subagent telemetry metadata and external sidebar identities, verified against upstream 1d8a29e. Synthetic import, truncation, and replay tests do not establish installed-app capture. |
Acceptance and drift
Section titled “Acceptance and drift”Every route needs a versioned native sample and a repeatable import trial covering messages, tool arguments/results, edits, attachments, models, measured usage, repository/branch context, account/organization/project identity, lifecycle, billing, and parent/child relationships where the provider records them. A metadata field must be identified as normalized, preserved only in safe raw data, intentionally excluded with a reason, unavailable, or still unreviewed.
Trials must preserve complete earlier transcripts and raw records, update a resumed session in place, distinguish absent fields from explicit clears, avoid stale metadata overwriting newer evidence, and deduplicate only with verified account and conversation identity. Unknown record types and metadata paths need a visible drift report. Partial pagination, inaccessible attachments, unsupported formats, and unavailable clocks must never be reported as complete captures.
The coverage audit checks every registered source, parser format, and browser route against this inventory. Adding a source or format without a product entry fails the audit. Public vendor discovery is a separate network check: candidate links require review before becoming support claims.