Releases
0.8.0Latest
macOS
Windows
Agents can read other apps’ local stores through context sources, use reviewed add-ons, and research one question across many archived conversations. Stats adds You, Optimize, and Outcomes pages, and Do it for me hands a finding to the person’s own agent, which changes a file only after the person applies the change. Accounts opens on an overview of limits and resets, each account page leads with its plan and limits, and API keys, now including OpenRouter, Amazon Bedrock, and Vertex AI, spend only after explicit consent. Skills & plugins adds a skill or MCP server that one agent has to the others, and a page an agent publishes opens beside its conversation. Terminal prompts can reach Review, the phone, and Slack together, and Changes reads like a code review and sends line comments to the agent that made the edits. Memories shows the conversations that changed and read each memory. Hansard’s terminals and the agents in them keep running when the app quits, crashes, or updates, and the built-in browser blocks ads and trackers, keeps several profiles, and serves up to four agents at once. Skills & plugins → Services switches each developer tool’s profile, team, or context from Hansard. The Library keeps useful work on its own, and a person’s Personal Sync devices now update to one release.
Added
- Context sources. Agents can read other apps’ local stores without
Hansard writing to them: the messages, contacts, and calls that Cued or
Nexus collect, and any SQLite file.
hansard context discoverfinds known apps,hansard context adddrafts a mapping for review, and Settings → Imports → Context sources orhansard context enableapproves it and chooses which agents may read each source. Agents read throughlist_context_sources,search_context, andget_context_itemwithin daily budgets, and every read enters a read log thathansard context readsprints. Results are redacted, one-time codes are masked, and text other people wrote is marked untrusted and pauses delegation and messaging tools until the next prompt. Nothing from a source leaves the computer. - Add-ons. Skills & plugins → Add-ons lists a curated catalog:
installed add-ons first, then a card for each add-on by what it does, with
Install and the space it takes on disk, and the add-ons still under
review in a closed Coming later group. Each add-on’s page says What
it can do, its Tools, the Agents that reach it, and its
Information in plain words. Context7, Memory, and Sequential Thinking
install from npm: each entry pins an exact version and the integrity of
every package it needs, and the installer checks every download and runs
no install script. DeepWiki and Microsoft Learn are hosted add-ons that
download nothing: a Hansard relay that the plugin bridge starts reaches the
server at the address the catalog pins, and an install confirms that
address. Installed add-ons reach every agent as servers of Hansard’s plugin
bridge, and a bridged item’s page names each agent with Hansard’s MCP entry
and links to the add-on’s page.
hansard addonslists, shows, installs, updates, and removes add-ons. An install or update asks in the person’s own terminal, outside agent sessions, or in a dialog on the page, and names what the add-on can do and its risk. A newer catalog only proposes updates. Nothing downloads until a person installs an add-on. - One skill for every agent. A Skills & plugins item page lists every
agent Hansard detects with how it has the item, and offers Add to each
agent that lacks it, such as Add to Codex, and Add to all, or says
in one line why an agent can’t take it. A dialog confirms each add with the
files, entries, command or address, what the item can do, and fields for
the values Hansard leaves out. A server’s copy carries its whole entry, with
environment values and headers shown masked, and the integration sync keeps
each copy current with its source. Remove takes out only what Hansard
added. Goose, Mistral Vibe, Windsurf, and Warp also take copies, and the
Augment CLI, Cline CLI, Kimi Code, and Kiro take remote servers, each in the
fields its own documentation names.
hansard bridge list --refreshfinds items installed since the last scan, andhansard bridge check <item>reports where each agent has one. - Published pages beside the conversation. A page, document, or picture
an agent saves to the Library, and a Claude Code Artifact publish, shows as
a card in its conversation. The card, or the agent’s link to the item,
opens it in a pane beside the conversation that shows each new version as
the agent revises it. A
save_to_librarysave by path publishes the page with a link that opens it in Hansard, and saving the same file again adds a version to the same item. Every session is told at its start to publish a page, report, or document it makes for the person this way, so Claude Code publishes them as Codex does. The phone shows the same cards. - Agent pane. Subagent runs, managed agents, and rows of the Agents view open in a pane beside the conversation, with Back, Open in full, their own transcript pages, and a composer while the agent runs. In Running’s projects layout they open in an Agent tab of the right pane. Words for a subagent go to the conversation that started it, wrapped so it passes them on with its harness’s own tool, and the pane offers to resume that parent when it is not running. The phone opens a subagent’s conversation from its call.
- Research across conversations.
research_past_sessionsandhansard history --research "<question>"plan up to 8 searches, read candidate conversations, keep only findings whose quotes appear in the messages they cite, and answer with a timeline and reversals. The readers and the summary run as background model calls through the signed-in Claude Code or Codex CLI, within a time and call budget. The deep-recall skill uses it for every integration target. On five synthetic questions it found 20 of 22 planted events against 19 for the earlier flow, while the calling agent read 17,835 characters instead of 206,914. - Questions to recall tools.
search_past_sessions,get_past_session,search_memories,get_memory, andboard_threadtake aquestionand answer from what they read, with citations.hansard history --questionandhansard show <id> --questiondo the same from a terminal. Each answer is one tool-less call through the signed-in Claude Code or Codex CLI, withclaude-sonnet-5orgpt-6-luna. Hosted calls and calls inside a taint window drop the question. - Stats You and Optimize. You compares the last 28 days with the 28 before them: approval waits, tasks that succeeded by work mode, redone agent work, cost per finished task, and limit hits, with last week’s recurring findings. Optimize ranks fixes by the share of usage they cover, such as repeated approvals, idle plugins, long conversations, warm account switches, short Claude Code cache lifetimes, and cheaper models for a kind of work, each with its evidence and Dismiss.
- Do it for me and Ask why. On an Optimize finding, Do it for me starts the person’s own agent with a brief, shows its progress and the proposed change as a diff, and writes the change only when the person applies it. Follow-up questions can propose new changes. Ask why asks the agent about a finding, a failing tool on Reliability, a broken integration, or a Memories review item, and can suggest a better wording for a proposed lesson.
- Outcome labels. Off by default. Check how conversations end in
Settings → Classification labels each finished conversation’s outcome,
whether the person corrected the agent, and whether the agent claimed work
that the evidence contradicts, each with the messages that show it. Labels
appear on Outcomes by label, week, harness, model, and project, and in
a conversation’s Details, and
search_past_sessionsandhansard historyfilter on them.hansard outcomesshows status with its cost, runs a pass, and backfills older conversations after it shows the estimated calls and cost. A judge that its rules cannot decide makes a background model call, and the labels are not yet checked against hand labels. - Accounts overview. Accounts opens on a limits board and a reset timeline, and every account and account group has its own page. Cards show each account’s email and workspace, the agent that uses it, a plan’s model pools, a line for each extra limit, and Use reset when a limit is reached and a usable reset remains. A limit a provider’s plans never have leaves its line empty instead of Not reported. An account page leads with its plan and its worth as figures, then Limits on one line each in one panel, Resets, and one history chart for the limit chosen in the section head. One Check now in the header checks the open account. The workspace ID is explained in its row, sign-in history and the account’s sessions are full-width rows, and each sign-in row offers the one sign-in that fixes it, with the rest in More. Attention notices span the content column with their fix at the right edge, and a capture from another device is named by that device. Each account records its plan changes, lists them newest first, and notifies once when a usage check finds another plan.
- Plans without a 5-hour limit. A Claude plan whose usage answer has a weekly limit and no session limit, and a Codex plan whose limit is only weekly, read No 5-hour limit on that line in Accounts and Routing, through direct reads, router quotas, snapshots, and published usage, and keep no stale 5-hour window. Routing rules place such an account by its weekly reset, after the accounts that have a 5-hour limit.
- Account groups.
hansard account groupand the group pages in Accounts give API key accounts shared allow and deny rules for organizations, projects, portfolios, and folders, and deny rules win over every allow rule. A rule change applies at each conversation’s next request. Subscription accounts can join a group, but routing does not apply group rules to them yet. Groups stay on this device, outside Personal Sync and Teams. Routing refuses a project that a turned-off Use only key holds and names the project, key, and group. - API key credits and reservations. A key can link to a subscription
account with a person-entered monthly credit and renewal day, and Accounts
shows this device’s routed spend in the cycle and the estimated amount
left. Reserve for explicit requests keeps a key out of automatic
routing until a model prefix, account header, matching project rule, or
hansard account run <key-account> -- <command>names it. A key’s page leads with its spending against each cap, the coding tool that uses it, and its group. Its name and key share one row, Use only this key and Reserve for explicit requests are switch rows, the account group is a Used by row, and Remove API key sits in More. A key’s row in the Accounts rail says what stops its spending, such as Key needed, Spending off, Cap reached, or On hold, or else its spend against its nearest cap. A reached cap tells the client when it resets. - Cloud API keys. OpenRouter, Amazon Bedrock, and Vertex AI express-mode
keys join OpenAI, Anthropic, Google Gemini, and xAI in Accounts → Add
account → API key account and
hansard account api-key save. Hansard checks a key without generating output. Bedrock needs--regionand offers its Mantle model catalog, and Vertex AI needs comma-separated Gemini IDs in--models. OpenRouter caps use its saved token prices, and a Bedrock key with a cap stays blocked because its model list gives no rates. AWS profiles and Google service accounts stay cloud connections. - Developer services. Skills & plugins → Services shows the account
each developer tool is signed in as: Modal and Hugging Face, and Vercel,
Netlify, Supabase, Railway, Fly.io, AWS, Google Cloud, Azure, Cloudflare,
Docker, and Kubernetes once their tool is installed. Cloudflare is read
through
wrangler whoamiand keeps its email and account names, never an account ID, permission, or token. Each row also lists every other account, team, organization, profile, subscription, or context the tool knows. A shared dropdown switches Modal profiles, Vercel teams, the Google Cloud account and configuration, the Azure subscription, the Docker context, and the Kubernetes context by running the tool’s own local switch, only when the person picks one in Hansard’s window. Hansard never switches an account on its own, and an AWS profile or a Hugging Face token, whose switch asks or lives in the shell, is named with its command instead. The tab draws at once from each service’s last answer, kept instate/dev-services.json, and asks an aged service again in the background, where a read had waited up to 1.5 seconds for the tools. A tool installed only inside a project, such as Modal in a project’s.venv, is found too. Sign in runs the tool’s own sign-in in a terminal inside Hansard, and Hansard never reads its tokens. AWS is checked only on Check, since its lookup can start a credential helper. Add as MCP server lists services such as Linear, Sentry, Notion, Atlassian, Stripe, Neon, and Replicate with the address of each one’s official remote MCP server to copy. Settings → Projects keeps one Developer services row that opens the tab. With Modal signed in, the + menu of a pane in Running offers Modal shell. - Timeline. View options → Timeline draws a conversation’s trace on one wall-clock axis, with long pauses cut to breaks and waits for the person and approvals marked. A Worked for row opens the Timeline at its turn, and a trackpad pinch zooms.
- Line comments in Changes. Comment on diff lines in Changes or a file’s diff view, and send the pending comments as one message, with file and line anchors and the hunk text, to the conversation that made the edits. Changes also marks files as viewed until their content changes.
- Changes as a review. Changes reads like a code review’s diff view.
Its View options menu shows or hides the changed-files list
(
Primary+Shift+Y) and sets its position, groups files by folder, sets test, build, and generated files apart, collapses or expands every file, and turns on side by side, changed-word marks, and hidden whitespace changes. Each file’s header has a Viewed check and a More menu with Attach as context, Copy, Show whole file, Mark as viewed, Open file, Open in any code editor installed on this computer, and Show in Finder. Unmodified lines fold into rows that open in place. Running’s right pane Changes list uses the same View options and More menu. A diffed file’s lines are counted once per file version, and a fold that cannot open says the file is not text. - Pull requests and GitHub. A pull request card opens failed checks and
unresolved review threads on request, reads the tail of a failed check’s
log, re-runs failed jobs, and sends review threads to the agent as line
comments. Beside Create PR, Details pushes a conversation checkout’s
commits without force, publishes a branch, opens draft pull requests, and
opens the repository, branch, pull request, and checks on GitHub.
Settings → Projects signs the GitHub CLI in through its web sign-in
without reading its token, and a shared dropdown in its GitHub row
switches between the github.com accounts the GitHub CLI holds with
gh auth switch, only when the person picks one in Hansard’s window. The GitHub status answers from its last reading while a fresh one runs. - Approvals everywhere. With Terminal approvals on your phone on in Settings → Remote access, a Claude Code permission prompt, plan, or question in a terminal also shows in Review and on the paired phone, and the first answer from any of them counts. Codex and Devin CLI prompts, and those of Claude Code subagents in the background, wait for the phone while the Mac is locked or unused, and return to the terminal within a second of the person using the Mac. OpenCode prompts are offered to the phone through OpenCode’s own API. An open conversation shows its waiting request as a live card, a plan shows whole on its Review card, and a terminal question takes its options by number key. Under the Slack owner prompts switch, waiting terminal prompts and plugin bridge approvals also reach the bot owner’s direct message as cards, which never offer Always.
- Phone relay (preview). When Tailscale does not answer, the iPhone app
reaches the Mac through Hansard’s phone relay at
relay.hansard.dev. The Mac and the phone encrypt every request end to end with keys the relay never holds, so the relay forwards frames it cannot read and keeps no content. The Mac keeps one outbound connection to the relay only while a phone paired through it or a connection link waits. Only connect over Tailscale in Settings → Remote access → Advanced turns the relay off for new and existing pairings. - Pause more agents. Running pauses and resumes Amp, Augment CLI, Kiro CLI,
Qoder, Goose, Crush, OpenHands, Hermes, OpenClaw, Junie, and Kilo Code
through their archived conversations, and offers Continue in another
terminal for them. Archived Grok Build conversations offer their resume
command. Running lists the Muse Code CLI and pauses and resumes it with
muse resume <id>, and an Antigravity CLI conversation that ran on remote agents resumes withagy --remote-agent --conversation <id>.cancel_agentlets an agent close the visible agents it started withstart_agentsthrough Pause, never one the person started or typed in. - Managed agents.
hansard agents --watchshows projects, the selected agent’s output, and its managed agents in three terminal columns, and--from <session>opens one archived conversation. Details lists a conversation’s managed agents as dense rows with when each ended. - Run threads. A
start_agentsor coordinator batch of two or more agents gets a project board thread for each run. A thread where two or more agents post has a Run tab with claimed files, commits, test counts, blockers, off-task flags, and refused messages on one timeline.hansard boards run <thread-id>prints the same summary, andhansard coordinator metricsand therun_metricstool measure each run. - Off-task language flags. Review and a conversation’s Details show advisory flags for evaluation awareness, hidden evaluation material, rule evasion, concealment, and person impersonation in board posts, peer messages, and tool calls. The rules run locally, make no model calls, never block delivery, and can miss or misclassify text. Details shows the section only for a run with a flag.
- Collaboration settings. A Collaboration group in Settings holds Agent messaging, Boards, AI assistants, Remote access, and Slack, and Handoff & helpers becomes Handoff. Agent messaging sets who agents can reach, message limits, attachments, and muted sessions, and Boards shows follows, channel subscriptions, harness boards, team sharing, and board sync. A live session can be muted from its direct messages page. Teams → Members shows each teammate’s messaging restrictions.
- Library exemplars and assets. Stars mark selected item versions as
exemplars with short reasons, and
/learnsaves a playbook with its final exemplar. A collected page keeps the files it names in the same version, a multi-file page previews in its sandbox, andhansard library exportor Download all files writes a whole version. Copy ID appears on every Library card, row, and item menu. Waiting playbook proposals show at the top of Library → Playbooks with Approve and Dismiss. - More collected work. The Library collects plans that Claude Code, Codex,
and Qwen Code propose, drawings from claude.ai’s visualize tool,
sites Codex publishes with its Sites connector, canvases Cursor and Grok Bot
agents write, Claude Design pages with their scripts and pictures, files
Claude Code sends with SendUserFile, Antigravity walkthrough screenshots,
and pages a shell heredoc writes after
mkdir,cd, or option steps. - New integrations. Goose and Mistral Vibe become full integration targets with Hansard’s MCP entry and recall skills. Qwen Code gains its MCP entry, skills, and session hooks, Crush its MCP entry, and GitHub Copilot CLI the recall, continue-from, deep-recall, library, and learn skills with a session start hook. The deep-recall skill also installs for Antigravity, Devin, Grok Build, OpenCode, and Factory Droid. Devin, Factory Droid, and Qwen Code add the archived evidence a first prompt names.
- ChatGPT texts. On macOS, the texts a person sends a ChatGPT dot from
Messages import as ChatGPT texts, grouped with OpenAI and kept apart from
ChatGPT conversations and exports. Enter the dot’s number in
Settings → Imports → Messages assistants → ChatGPT dot number, or set
imports.chatgptMessagesHandle, and then runhansard import --sources chatgpt-messages --since all. Hansard opens no Messages store for ChatGPT texts until the number is set. The texts list under the archive’s ChatGPT dot room, beside the dot’s tasks and threads that the browser extension captures, and every dot room is pinned first. Messages assistants’ reactions show as badges with who reacted, a thread reply quotes the message it answers and jumps to it, and a vCard draws as a contact card, in the viewer and on the phone. - Web chat folders. A chat service’s folder in Sessions shows an Activity calendar of the last 26 weeks, the topics its newest titles share, and what the service remembers and made. Its label leads with the service’s mark and the person’s own picture where Settings holds one. A web chat or a Messages assistant’s chat continues in a local agent without a checkout, in a folder Hansard keeps for each service, or in a listed folder chosen with Change.
- Try Hansard. Guided setup adds a phase of short steps that run on the
person’s own archive: try
/recall, find a remembered conversation, let a second agent find the first one’s work, ask a running agent a question, continue from a past conversation, star a script agents reuse, let two agents talk, and fix one thing from the week. Settings → About lists every step. Feature guides cover the Library, Running, Review, Projects, and Context sources, and a guide opened on one computer counts as seen on the person’s other Personal Sync devices. - Slack bot setup. Settings → Slack → Interactive bot creates the Slack
app from Hansard’s manifest, connects pasted tokens, and claims the bot from
a direct message code, then lists the channels it was invited to with
switches. Settings shows when the Slack app needs an update and can update
its manifest with a pasted configuration token.
hansard notify slack bot connect,claim, andchannelsrun the same steps. - GitHub checkouts as projects. The watcher finds GitHub repositories cloned beside known checkouts and makes them projects before their first conversation. With the GitHub CLI signed in, API key and account group pickers offer repositories and organizations not yet in the catalog.
- Readable export options. Shared pages show edits as diffs by default,
and can add a running totals panel and harness context. The Share dialog
has a switch for each.
hansard export <session-id> --format htmltakes--totalsand--harness-context, and with--tools,--no-diffsshows edit inputs instead of diffs. Usage over time adds bar charts of each prompt’s tokens and estimated cost, with a table, to a web page. A step that ran on a faster service tier is priced at that tier’s rate in the page, trajectories, and Trace. Theatifandeatifformats write ATIF-v1.8, and the Share dialog no longer offers Copy link, whose local link opened only in the person’s own viewer. - Share with team. The Share dialog shares one conversation with a Team
whose rules include it while the member’s folders leave it out, after a
confirmation of what the Team receives and its redaction.
hansard team share <session-id>does the same from a terminal, with--teamand--dry-run, andhansard team unsharestops uploading its later changes. Sharing refuses inside an agent session. - Terminals that outlive the app. Hansard’s built-in terminals and the agents in them run in a terminal host process, so quitting, a crash, an update, or a viewer restart ends none of them. The reopened app shows each terminal’s screen with what it wrote meanwhile and types into it again. Agent runs and Deep Recall answers that the desktop app started also keep running after it quits, and the watcher pauses idle agents in those terminals by the usual rules whether or not the app is open. Quit Hansard? now lists only classification jobs.
- Scratch folders. Folders a tool makes one per run fold into families. The Sessions Folder filter shows one Scratch folders row with a row for each name stem, and Sessions, project pages, and Stats gather a device’s scratch folders into one group. Projects suggests each family once with the project its evidence names, and adds a family to its project on its own when scripts of exactly one project’s checkout create its folders, saying Added because with the file and Undo.
- Model, speed, and effort picker. The composer, New conversation, the coordinator workbench, and the phone pick the model, speed, and effort in one chip. Its popover lists each model with a cost mark, the speed and effort rows for the chosen model, and the list price per million tokens at the chosen speed, and the chip shows a bolt for Fast and two for Ultrafast. Claude Code’s fast mode is offered on Claude Fable 5.2 and Mythos 5.2, and its model list keeps the release’s own rows that a cached account option displaced.
- Doctor repairs.
hansard doctor --fix, or one question per repair in a terminal, resumes a paused search index, reinstalls missing integration files and hooks, restarts the watcher from this copy, and records this copy’s Node.js in login startup. - Codex memory v2. Memory capture reads Codex memory v1 and v2 from the
default or configured home and every saved Codex profile. Run
hansard memory backupto refresh without a conversation reimport. - Windows login startup. The Windows app starts the bundled watcher at
login without a console window, restarts it after failures with a backoff,
and removes its logon task when the app is uninstalled. Notifications post
without a Hansard window open on Windows and through
notify-sendon Linux. - Built-in browser preview. The macOS desktop app on macOS 14 and later
includes a browser whose sites agents can use after the person grants a
lease, answered in the browser or on the phone. It docks in the right pane
of the view on screen, a Sessions conversation’s Browser pane or a
Browser tab of Running’s right pane, and opens in a window of its own
only after Open in a window. View → Built-in Browser
(Option-Command-B), the conversation header’s Browser toggle, and the
command palette open it. On first open it offers to import the sign-ins of
the person’s main browser. Import sign-ins takes three steps: the
browser and profile, the sites with the signed-in ones checked, the most
recently used first, and email, banking, and sign-in services in their own
section, then a short review that names the services that may ask to sign
in again. Advertising and tracking domains stay unchecked. Sign-ins come
from installed Chromium browsers and Firefox into an encrypted profile
store, each import is listed in Settings → Coding agents → Agent
features, and imported sign-ins clear on request. Block ads and
trackers, on by default in Browser options, blocks the domains of a
bundled EasyList and EasyPrivacy list, never sign-in pages, and
Allow ads on this site exempts one site. Browser options also create,
rename, and delete profiles, each with its own sites, tabs, and sign-ins.
Up to four agents can hold leases at once, each on its own sites and tabs,
named in the agent bar with a link to its conversation, and Stop all
and Take over all end every lease.
browser_accesscan name a profile. Let agents use the built-in browser, off by default in Browser options and Settings → Coding agents, lists thebrowser_*tools for new agent sessions, about 1,260 tokens of tool definitions. The tools return outlines and visible text, never cookies, secret field values, or a profile export. The tab spinner and loading bar pause with the page’s other repeating indicators while the page is idle or hidden.
Improved
- Lean recall replies. Recall results merge a conversation’s copies from
other devices, drop weak matches, and weigh the caller’s repository higher.
search_memoriesmerges copies of one memory and pages 5 at a time, andget_past_sessionkeeps the person’s and agent’s words whole while folding runs of tool steps. On a synthetic run thread of 176 replies, catching up withboard_updatestook 11 calls and 73,053 characters instead of 45 and 232,145.board_contextlists threads within about 750 tokens, and session start lines carry less board text.hansard show <id> --json --msg-limitreplies stay within 48 KiB, withtranscript_nextfor the next page, and a long message’s exact text reads in pieces through--msg-offset,--msg-limit 1, and the new--content-offset. Research keeps time for reading transcript evidence, keeps the evidence that matched, and says when its coverage was incomplete.code_historymerges a memory’s copies from other devices, and the recall tools’ descriptions say thatsincefilters by when a conversation started. A board read or post stays in the caller’s working repository, even when the harness changed folders after its launch. - Stats pages. The sidebar lists the Stats pages in groups, with a page menu in the title when the sidebar is folded, and page controls sit under the title and in the heads of the cards they change. Work splits into Work and Outcomes, and Limits & accounts, Speed, and Context & cache hold subscription value, agent runtime, and cache rewrites. Library and Skills & plugins pages show what the Library keeps, reuses, and never uses, skill runs by period, agent, and project, installed skills that never ran with their listing cost, and MCP servers with failures, bridged calls, approvals, and call times. Overview’s lead figures show their change against as many days before the range. Session time reads in hours past two days, and Stats Changes leaves out work-mode and success comparisons when the earlier period has too few classified sessions. Ranked lists and tables share one table style.
- Stats at large archives. At about 20,000 synthetic conversations, an unchanged Stats page polls with a 304 instead of fetching a 12 MB payload, and a quiet minute went from 48.6 MB and up to 96 rebuilds to none. The Stats engine sizes its heap by the rows it reads and no longer exits on a shape switch. Three-month bodies fell from 11.9 to 7.9 MB. Stats Changes now answers at that size instead of failing, and a held comparison answers it in 57 to 78 ms instead of 2.9 to 3.3 seconds. A busy refresh answers from the saved result.
- Subscription value. On the benchmark fixture the value computed in about 3 seconds instead of 29, and later changes cut the fit from 1.9 seconds to 0.17. A restarted viewer answers at once with the last saved value, marked with its age, while a newer one computes.
- Accounts polling. An open account page patches in place, and with a 32-day weekly chart it holds about 1,400 elements instead of 9,650. Two minutes of polls on a synthetic home with eight saved logins took 34 ms of server time instead of 268. A refresh of ten long transcripts into the subscription ledger takes about 90 ms instead of 1.1 seconds.
- Plan names and usage. Plans show by their current names: Claude’s Max
and Ultra, ChatGPT’s Pro 100, Pro 200, Pro 500, and Business, Grok’s
Premium, Super, and Ultra under The X Subscription, Cursor’s X Subscription
memberships, Muse Code’s own plan names, and Google AI tiers for router
Antigravity accounts. A Google AI Ultra plan’s price can be chosen per
account with
hansard account value price. Cursor’s Cursor Models and Other Models pools show as separate windows, the Devin CLI quota comes from its status line, and every plan-limit account shows unreported 5-hour and weekly rows as Not reported or Not read yet. Accounts also read the Augment CLI’s balance with its included allowance and cycle end, the Kiro CLI’s sign-in, and Windsurf’s cached daily and weekly percentages, without opening credential files. A Muse Code check that carries no usage keeps the account’s last readings, marked stale, instead of showing no limits until the next window’s first use, and a Muse plan saved under its tier name no longer reads as a plan change to itself. A Muse Code usage check never runs its launcher’s updater, a timeout names the step that stalled, and a credential failure says whether the file is not valid JSON, too large, not a regular file, or in an unsupported format. - Model prices and fast tiers. Stats, the composer, and exports price
Claude Fable 5.2 and Mythos 5.2, released November 9, 2026, GPT-6.1 Terra,
Muse Wonder 1.0 and Muse Ember 1.0, Kimi K3.1 Pro and K3.1 Pro Highspeed,
MiniMax-M3.1-Flash, Mistral Large 4, LongCat-2.5, and SWE-2 and SWE-2 Flash
at their makers’ published rates, and Grok Build’s
grok-4.7-buildat Grok 4.7’s. The model catalog adds 127 identities and marks 97 as retired. Tokens a harness recorded on a fast tier are priced at that tier’s published rate: Claude fast mode, OpenAI Fast and Ultrafast as Codex records them per turn, and Cursor’s fast Grok models. Stats lists what each tier cost, and the effort and speed card’s tooltips name what fast mode cost. Newer releases of a model family keep their own shade. - Fewer manual commands. Sign-in, routing, the router, pairing, and the
CLI now repair what they can instead of naming a command. Accounts of a
stopped router offer Start router, which installs a missing router, a
router worker that never became ready or stops answering is replaced, a
router whose port another program took moves to a free port, and router
account writes retry once.
hansard updateandhansard statusoffer to set up a computer never set up,hansard psoffers to pause agents idle for over an hour, andhansard coordinator statuslists open decisions and offers to reopen agents a restart stopped. - Router restarts. A released router worker finishes the responses its API key and Muse gateways are streaming and holds their ports for the next worker, so an app installation or update no longer cuts a conversation or refuses a connection. Gateway session bindings survive worker restarts, so a warm conversation stays on its key. New conversations start away from accounts near their limit.
- Router down. While agents route through Hansard and its router does not answer, the title band, Accounts, a refused request’s error row, and Running say so with Start router and Stop routing, which returns each routed agent to its own sign-in. A notification follows once the router stays down for two minutes, never for a stop the person made. A router start is recorded before it runs, and the watcher adopts a router that answers with no worker, so routing comes back on its own.
- Accounts → Routing. Automatic account switching lists each agent installed on this computer on one line, with how many of its accounts are ready, the account it uses next, and one action, Enable or Add account. The former Codex app account reads Codex sign-in, which supplies the workspace, plugins, and connected apps, and the Codex row shows it only when it needs attention. Requests that a key’s Use only rule refuses read as one sentence. Each order lists its plans and limits on the Accounts limit line, the rule Five-hour accounts reads Accounts with a 5-hour limit, and an account placed by the saved order says why in plain words. A usage check redraws only that account’s lines.
- Sign-in safety. A router account’s rejected sign-in answers with a retryable 503 instead of a 401, so Codex no longer spends its sign-in renewal on it. Hansard never renews the default Codex folder or a folder the person registered, an expired Codex access token reads as a wait for Codex to renew it, and a router copy holding a rejected token stays off until the saved login has a new token. The sign-in window shows who signed in and closes on its own.
- Sign-in from Running. ACP agents that refuse a run for want of a login, including the Cursor CLI, Augment CLI, GitHub Copilot CLI, Kimi Code, Kiro CLI, Cline CLI, and Mistral Vibe, end it as sign-in required, and Running offers Sign in with the agent’s own login command. Antigravity sign-in explains its one shared Google sign-in and offers Sign in through the router where routing is on.
- Cloud connections. Changing, switching, or removing a cloud connection no longer asks to turn it off first, and a subscription login runs beside a cloud connection that holds its default profile, with that process’s settings given only to it.
- Plugin bridge approvals. A bridged server’s tool groups are set from one dropdown that says what agents could do without asking, for every agent or for one. Ask every time asks before each call and never remembers an answer, and the other asking mode reads Ask once per conversation. Computer control in Settings → Coding agents turns tools that control the computer on or off for every agent and for each one.
- Skills & plugins sidebar. Under Search, the sidebar offers Show
(All, Needs attention, New this week, Changed this week, and
Not used in 30 days), Agents, each harness with its mark, and
Sources, such as agent folders, the shared
.agentsfolder, plugins, projects, apps, and caches, each row with the items it would show. The tabs alone choose the type, and the Harness, Type, and Scope dropdowns are gone. A row shows the description most of its copies give, with an N descriptions badge for the others. A connector known under several names is one row that keeps each name’s tool access. Moving from Services to another tab no longer passes through All first. - Running. Conversation tabs use one style with a status glyph and keep only running and recent conversations, with History, Resume, and New conversation at the strip’s end. A project whose lead agent sat idle past the hour reads Idle for over an hour. Each agent’s memory counts its process tree’s physical footprint on macOS, with resident memory beside it, and its CPU counts only live processes, so a busy child that exits no longer makes a working agent read as idle. An idle Codex conversation can pause automatically, since its own app server, sandbox helper, and native binary no longer count as work, and a conversation’s managed agents leave them out. A project card counts several agents that need the person, as 4 need you beside 10 running, and a rail card that needs the person shows its amber state words and count without a ring. Activity overview is offered in every grouping, takes its rail row’s name, and its meta line counts the agents that need the person, work, and wait idle. Its memory past what is installed reads as compressed or swapped instead of a share over 100%, and a poll redraws only the checkouts that changed.
- Agent graph and Lanes. Running’s agent graph fits the page down to 70% scale, then zooms and pans, including with a trackpad pinch. Pointing at a node lights only its links, and Graph and Lanes take no text selection, so a pointer drag no longer washes the graph in the selection color. Graph and Lanes pages run up to 1,440 px wide and end with room under their last row. A live agent’s lane draws its state when no turn is open in the window, so a lane that needs the person is never empty, and a count of subagents with no row opens the conversation that started them.
- Running view options. View options is one menu of Group by
(Project, Folder, Environment, Status, Date, or None), Sort by with its
direction, Show, Agents, and Display, with Reset to
defaults while an option differs. Show → Subagents and side
conversations, off at first, nests each conversation’s newest children
under it. New section and Move to section keep the person’s own
sections of projects. The right pane adds a Context tab that draws the
conversation’s context window. In By folder, programs outside every
checkout fold into one Other row with their count and memory. Rows keep
their place when
ps,lsof, orosascripttimes out on a loaded machine, with the last complete reading marked stale. Every menu opened from the sidebar, View options and section menus included, stays inside the sidebar and opens toward its start, as the Sessions filter does. - Delegation. A session can start any number of helpers, and a project’s
Helper budget counts only helpers that are waiting or running.
spawn_agentsays that no window opens, and an agent whose terminal app refuses Hansard is told so withspawn_agentfor background work instead of opening a Terminal window for each agent.coordinator start --resume,start_agents,hansard show,get_past_session, and the viewer’s links take the id a harness printed as well as the archived id. - Conversation reading. Each run of steps between two agent messages folds under its own Worked row with its time, and a turn with tool calls ends on one finish line whose control opens or closes every fold. An opened fold shows every step, and read steps name each file they read. The header keeps one metadata line, View options opens as a menu, and Details opens only when the reader opens it. Consecutive API retries draw as one row that counts the attempts. Show and Hide N tool calls open and close every step at once, a code-mode script draws inside the card of the call it ran, and find in a transcript streams its count and current match from the rows in view, in slices of a frame.
- Agent messages. Transcript cards name each agent message’s recipient: board posts link their thread, conversations open in the Agent pane, and the person reads To you. Delivery shows only recorded states. A board post from an unnamed session names its harness, and a delivered message previews by its own words. The Coordinator’s Needs you lists approvals before questions, oldest first, and each wait turns amber and then red as in Review. Board posts, agent messages, and Codex subagent messages draw as one quiet card that names the other party with the time and a recorded delivery state, folds long words, and shows Reply and Copy on hover. A Direct messages session that never named itself is named by its harness.
- Details pane. A conversation’s Details list the memories it wrote, loaded, or read, the agents it messaged, heard from, or started, the other folders it used, and the folders it was given: a Codex sandbox’s writable folders and the folders added to a Claude app session. The Session section sits under Overview, with each row’s role in its tooltip beside the pane. Memories gains the same Details pane beside a memory: its account, file, saved versions, and front matter, and the conversations that changed and read it, newest first and a page at a time, each opening at the message that did it. They come from the experience index, so a memory’s older conversations appear once the index has derived them. On a phone the pane spans the screen.
- Person’s name. Settings → General shows the person’s name with the device and time it came from and whether Personal Sync carries it, and changes it through Edit. Conversations received from the person’s own devices show the current name, and a new name reaches each Team without waiting on Team storage.
- Phone. The served phone script fell from 335.7 KiB to 147.1 KiB, a repeat launch takes two round trips instead of five, and the last three conversations reopen at once. Listing 40 agents across 2,000 folders takes 13 to 16 ms instead of about 2 seconds. The iPhone app’s request cards show the agent, conversation, project, risk, and a subagent’s brief at every text size, and Devices sends a test approval. An agent the Mac no longer lists reads Ended, and its controls leave.
- Boards. Boards order by their latest post or their project’s latest conversation, harness boards follow the person’s harness order, and a temporary folder’s board stays out of the list until it holds a post. Following lists one row for each thread with its update count. The outline rail draws every tick on the first paint, and a board list holding a 120,000-post board builds in 0.3 ms instead of 33.6 ms.
- Graph. A catch-all group, such as web chats outside any project, splits into the communities among its members, and unlinked conversations gather into small month clusters at the layout’s edge. A new conversation joins the cluster its links, or its nearest conversations by meaning, point to, instead of landing as a lone cluster beyond the view, and the Groups arrangement draws each group as a named region in its color. Visible views and Projects read from a read-only reader while the Graph worker indexes, and the worker prepares views only after indexing has been quiet for a minute. On a copy of a large archive while agents wrote, opening the six largest clusters’ members took 2.7 seconds instead of 8.4 to 10.2, a warm open of All Conversations took 0.50 seconds instead of 1.74 at the median, and the worker used 17 seconds of CPU instead of 54. The Graph sidebar has one Reset beside Filters that returns every control, group, and pin to its default, Arrange and the forces join Display, and the fields read Show, Link by, and Node size.
- Teams. A lone Teams card’s facts read as one line of labels and values, the overview’s header says how teammates see the person, each teammate sits under Conversations in the sidebar, and a shared conversation’s harness is named in its tooltip.
- Faster first loads. After a viewer restart, saved project lists, unchanged Memories lists, and recent telemetry results answer at once. The Coordinator page draws from the view its last poll kept, and an unchanged poll answers 304 instead of 1.1 MB. The first peer listing no longer holds the viewer for 2 seconds, pull request discovery starts beside Git status, and Workspace line counts read only changed tracked paths. After a restart, the first Settings and setup answers, a conversation’s Managed agents, and a checkout’s pull request status answer from what the last viewer kept, with its age, while fresh reads run. Skills & plugins → Services and the GitHub status in Settings → Projects answer from their last readings while fresh ones run. A project page starts its agents, catalog, and coordinator board reads with the page, so Waiting on you and Working no longer wait for a process scan.
- Lighter imports. In archive format 2, a Claude Code transcript of 4 MiB or more imports as a stream, in time that follows what was appended, and watcher imports of Codex rollouts from 256 KiB continue from settled stream checkpoints. Small appends reuse raw hashes and view pages, and scheduled imports yield after 100 changed items, 256 MiB of source, or 15 seconds. A watcher-owned child recovers an interrupted archive write in small steps: on 30,000 synthetic sessions it peaked at 167 MB instead of the 1,122 MB of the session list rebuild it replaces. View pages of views over 1 MiB compress at gzip level 4 instead of 6, which cuts a re-read’s CPU by about 16% for 4% to 17% more page bytes. Sources whose readers parse their own store, and the Claude app, import within the shared work budget in batches, and a scheduled Claude Code or Codex re-read reads later batches from a queue its first batch writes instead of listing every source again. A due import heartbeat stops yielding after five minutes to sources whose files keep changing, so a constantly rewritten Claude Code import no longer holds back Muse Code and later sources.
- Not in a project. Folders of conversations outside any project read as, for example, “Codex, not in a project” in Sessions, Projects, Stats, Graph, boards, and help, instead of (Uncategorized).
- Personal Sync with a full outbox. Archive writes continue while the upload outbox has no room, and the captures they owe upload once it has. Remote cleanup’s history caches yield to uploads past 60% of the outbox limit, a capacity failure retries at once after a restart, and status says that imports continue while backups wait.
- Personal Sync storage budget. Each backup has a storage budget: twice
the current backups of every device, at least 16 GiB, or the size set with
hansard config set sync.remotePrune.budgetGiB <GiB>. Status,hansard doctor storage, and Settings → Personal Sync show the backup’s size, its growth over the last month, and its share of the budget. Report-only remote cleanup plans in resumable steps, so waiting uploads no longer hold its report, and it names reclaimable bytes by kind. - Web chat capture health. Settings → Imports → Browser extension and
hansard browser statusshow the last capture, each extension’s last report and version, each paired browser with its extension ID and last contact, and each waiting provider’s reason. Sites turned on with no open tab and nothing waiting stay out of the list. While capture is stale, the title band says since when no web chats were saved, with the reason. - Web chat capture imports. Captures import smallest first. A capture that stopped the import worker twice is held as too large instead of stopping later imports, and the worker yields to archive maintenance after an interrupted write. Archiving a Claude Code on the web capture reads only the providers it needs, where it had read the whole archive for about a minute and held back every other capture.
- Watcher background work. A ready source holds the archive refresh back for at most 10 minutes, watcher ticks count only time the computer is awake, source watches drop events for files whose status did not change, and Antigravity, Devin, and Codex SQLite stores are read through one reader process per store. Prompt and session start hooks answer within 2.5 seconds of their start, leaving a slow lookup or notice for the next prompt, and sweep their per-conversation records at most once an hour.
- Library collection. A file that both the scratch source and a
conversation’s write collected is one item, and a published page’s
template beside it is not kept as an item of its own. Card previews start
at a document’s own words instead of its YAML front matter. An item page’s
header holds Star, Keep, Download, and More actions, and
save_to_librarylists the Library’s kinds and answers a file type given as a kind with that list. - Settings and pages. Every view uses one four-level heading scale, Settings descriptions fit one line with detail in each row’s info, and row hover fills stay clear of their text. Settings rows and wells share one style. Each assistant in Settings → AI assistants has its own section with a row for each thing it may do and Remove. Remote access reads as This computer, Allowed devices, Shared projects, Your other devices, and Advanced. Each agent in Coding agents is one row with Add, Repair, and Remove, and About folds its feature guides after the Try steps. Coding agents holds the built-in browser, playbooks, and plan usage in one Agent features section. Settings → Projects shows three combined projects and folds the rest, and keeps the GitHub repositories row to one line. A project page’s environment opens Edit environment, with each folder’s branch on this device, Worktree on the start folder, and variable names that say whether this device sets them. The shared dropdown and the command palette rank their search by fuzzy match, and the Sessions and Memories Folder filters pick from a searchable list of folders, whose next page reads Load 100 more folders. Every item inside a section, such as a row, card, list, or chart title, uses one item scale below its section heading: titles at 13 px and 600 weight, and meta lines at 12 px, muted. Empty states and the coordinator’s welcome take the shared empty-state title instead of a heading’s size. Every page column ends with the same padding at every width, Accounts included.
- Product logos. Developer services, remote MCP servers, connectors, API key types, browsers, web chat folders, and published Library items show their service’s logo in its brand colors, from Simple Icons, Lobe Icons, and the Chromium source with their notices. Storage services whose terms permit no logo show their name beside a generic glyph.
- Composer and launches. A one-row chip says who chooses the permissions of OpenCode, Devin, Amp, and the ACP agents. A launch through a turned-off account offers to turn it on in the same confirmation, and a helper that no account can serve is refused at once with each account’s reason, the soonest reset, and the harnesses that can start instead.
- A conversation is never titled by a delivered Hansard agent message, even when a launch appended session-tools text after it or the harness titled the conversation by its label.
- Review’s Update Hansard item proposes an update only for a device that runs an earlier release, waits a day after a device answers that it is current, and ends once a check finds the device no longer needs it.
- On Windows, a saved login’s sign-in, a harness sign-in, and the Running
monitor run as the terminal’s own process, so
cmd.exeno longer reads their quotes as backslashes or expands%NAME%in their values. Workbench agents launch directly in Windows terminals, and leaving Running closes the monitor it started. - Phones, the Mac’s live run, and
follow_remote_runno longer draw Claude Code Read, MCP, ToolSearch, Skill, WebSearch, and WebFetch calls as failed. - A long conversation keeps its subagent runs on each transcript page instead of dropping the whole list over the response cap.
- A model routed through a router account alias reads as the model in every view.
- Graph’s Groups arrangement no longer stops on a link between two catch-all clusters, which had left every conversation where the earlier arrangement put it. Meaning links appear as soon as the first conversation vectors arrive.
- An
agentcommand is told apart by the file it runs, so the Cursor CLI and Grok Build, whose installers both linkagent, launch and list correctly. - A refresh never pulls back a reader who scrolled up, a transcript that follows its end glides there instead of jumping, and the line in view stays still while code above it draws. The phone conversation follows new output only while the reader is at its end.
- A damaged saved Running list is replaced after a fresh scan instead of blocking later writes.
- A router no longer reads as stopped after the computer’s time zone changes, which had skipped usage checks and blocked app installation, and a stop ends a router whose start was recorded in another time zone.
- Requests through the API key gateway reach Anthropic as the Claude Code version that sent them, with its own headers, instead of as Claude Code 2.1.280, which Anthropic refuses for Claude Fable 5.2. The router’s claimed version follows the newest installed Claude Code, lowered to the oldest one running.
- Remote control stays up when a network change renames the Mac, since the host lock names this installation as well as the host name.
- Codex’s own trust entries for scratch folders no longer block integration
edits of
~/.codex/config.toml. - A session hook an earlier Hansard wrote reads as Needs an update instead of Missing, while the sync rewrites it.
- A held approval prompt returns to its terminal when an installation replaces the app or the agent session ends, and requests whose hook or session is gone leave Review, the phone, Slack, and the Mac’s notification.
- A failed or stopped turn keeps one Failed line across live refreshes, and retries Claude Code recorded before a failed request count as one failure.
- The subscription pool keeps serving while the account group store cannot be read, with no API key in routing.
- Stats Agents no longer counts subagent runs from spawn calls twice, and Stats Changes counts human turns as the prompts a person typed.
- Terminal completion setup sources a file, so a slow shell start no longer cuts the setup line at 1,024 bytes. OpenCode plugin and Codex terminal sockets stay within the macOS path limit.
- An offload restore reached through the macOS home’s
.noindexfolder is no longer refused as another archive. - The Library keeps pages an agent published with the Artifact tool from any folder, including a project’s hidden working folder.
- Exported, traced, and shared steps are priced at the rate in effect at their own time instead of the newest rate.
- The phone composer clears the keyboard’s echo of sent words, so a committed autocorrection no longer leaves a delivered message in the field.
- Claude Code and Codex, which show a tool result’s structured content in place of its text, now show every Hansard tool’s whole answer, including recall notices, code history, work context, and the built-in browser’s outline and page text.
- A Codex conversation recorded in full-auto mode resumes with
-s workspace-write -a on-request, since Codex 0.167 refuses--full-auto. A Codex terminal session’s app server ends with its TUI instead of finishing a turn on its own with no terminal. - Agent output pasted into a person’s message reads as the person’s text instead of drawing tool calls inside the bubble.
hansard doctorno longer reports every line of Cursor’s per-projectworker.logfiles as a malformed record.- A full search index build can replace a live index that holds a received conversation by its title alone, while its events are not yet derived.
- A peer notice that a hook stopped before writing is announced at the next tool call instead of being lost.
- Settings → Imports → Maintenance hides recovery rows for agents that are not on this computer.
- The menu bar’s Memory module keeps agents with similar footprints in their order instead of swapping rows between reads.
- A Codex run that Hansard starts keeps recording its events and its end. The processes Hansard owns keep their standard output blocking, so a Codex app server that one of them starts keeps writing when its reader falls behind, and a run whose connection stops delivering notifications ends at its rollout’s turn end.
- A project’s coordinator section shows a loading notice after 1.5 seconds, keeps Retry when a load fails, and keeps its cached agents and decisions with an out-of-date notice when a refresh fails. A project known on this computer whose folder is missing says how to set one, and Project folders opens its Environments section. An answer for a project the person left no longer replaces the open project’s section.
- A repository under a folder that macOS protects, such as Documents, keeps its project board on a loaded computer: a permission check that fails leaves reads on the last completed check.
- Dependency updates close advisories in rustls, follow-redirects, and zip.
All upgrade notes- Update every Hansard copy that shares an archive, including the desktop app and npm installation, and update each Personal Sync device. The coordinator workbench store moves to writer and reader version 2, so earlier copies leave it unopened instead of resuming conversations without their account placement. Router API accounts use writer version 6 and require reader version 5, which records spending consent, or reader version 6 once the store holds an OpenRouter, Bedrock, or Vertex AI key, so earlier copies cannot read an API account store this release saved. The plugin bridge’s switch store moves to writer and reader version 2, so an earlier copy keeps the switches it last loaded and, once restarted, serves no bridged tools until it is updated. Helper records with account placement or startup retries require reader version 5. Library collection state requires reader version 5, so an earlier copy stops collecting into the Library. The add-on record moves to version 2 with reader floor 1 for hosted add-ons, which a copy that reads only version 1 leaves out and never rewrites. Integration state uses writer version 5 with reader floor 2, and earlier copies read it without rewriting it. The Library catalog uses writer version 5, playbooks with exemplar links require reader version 2, and playbook proposals with an exemplar require dreaming record reader version 3. Streamed import state moves to version 2 and starts again on its own. A project merge that adds a folder family syncs as project merge schema 2, which earlier copies carry without applying. These changes need no complete archive reimport.
- A person’s Personal Sync devices now update to one release. When another
device in the same Personal Sync profile reported a newer release in the
last three days, and the update feed or npm already lists that release for
this platform, the desktop app opens an Update required sheet.
Update and restart runs the in-app updater, and while tasks run,
Update when they finish waits for them without stopping any. On an npm
installation,
hansard status,hansard doctor, and the interactive mode print the npm command that installs the release. The watcher keeps archiving meanwhile. A release not yet published for this platform, such as a Windows preview that lags behind macOS, a prerelease, and a source checkout show a notice instead. Only copies from this release on take part, so the first sheet can appear when a later release ships. - Personal Sync now uploads Library items at once, in a stream of their own, instead of waiting until every device can read them. A device on 0.7.1 receives that stream and lists its items under Other devices. Devices on 0.5.6 to 0.7.0 refuse that one stream, mark a cloud account as Needs update, and keep receiving everything else, though each receive reports a failure and their remote cleanup report is refused while the stream is listed. From this release, a receiver leaves a stream it cannot read, says how many items from which device need a newer Hansard, and receives them after it updates. Sync formats that change how conversations are stored still turn on only once every device reads them, but they hold no upload: conversations upload in the earlier form until then.
- The person’s name in Settings → General becomes one Personal Sync setting that every device in a profile holds. The first pass after the update offers each device’s current name, and the device that runs it first sets the name the others take, so a device may change to another device’s name. An edit made after that wins on every device, and archive maintenance relabels this device’s archived conversations with a changed name. Device names stay on each device. Earlier copies carry the setting without applying it and keep their own name until they update.
- Release preparation resets the 150 parser entries to the new baseline
0.8.0.0, replacing their0.7.1.*versions, including that of the newchatgpt-messages-historysource. Refresh after updates therefore re-reads every enabled source once, including sources whose importer did not change. Provider files still present are imported again, files older than the import window are read by one low-priority import run, and conversations whose provider files are gone are parsed again from Hansard’s raw copies without keeping fewer messages. The work runs in background batches on AC power while imports are idle. Settings → Imports → Maintenance andhansard statusshow its progress. No complete archive reimport is required. - Importer changes that the refresh brings to earlier history:
- Claude Code and the Claude SDK classify the tools-used session record as hidden instead of counting it as an unknown record, and know ten record and content types that Claude Code 2.1.275 through 2.1.319 write: a refusal without a fallback model and stopped agents keep a context row, and bridge sessions join the Remote Control evidence. Headless Claude SDK subagent runs import as separate conversations linked to their parent’s Agent calls.
- Codex CLI, desktop, and SDK imports keep the files a code-mode patch adds,
and a credential-free part of an MCP result’s structured content that
its text leaves out. They also know the records that Codex through
rust-v0.167.0 writes, so
additional_toolsno longer counts as an unknown record in every session from Codex 0.165.0 on. A Codex attached thread reads as a fork of the thread it came from, image generation draws as a call row and a result row with its image, and the session summary keeps the folders a turn’s sandbox may write. Claude app sessions keep the folders the person added to them. - Antigravity app, CLI, and IDE imports link subagent runs to their spawning
calls with their role and final status, and the Antigravity CLI’s failed
terminal steps no longer count as unknown records. Muse Code imports link
subagent runs and keep provider-started reminder children apart from
delegated runs. Muse Code 1.4.14 journals fold approvals into runtime
approvals, show display prompts, steers, and readable reasoning, and keep
the archived history when Muse prunes a journal at a context checkpoint.
A Muse Code conversation is titled by its first request instead of the
handle Muse gives it, such as
lake-pisces, which stays as an alias. - Grok Build permission, retry, hook, and subagent records, Devin CLI compaction and context-only records, and Kimi CLI approval, hook, retry, status, and compaction records gain a line of content and a context kind, so they no longer appear as blank System rows. Grok Build’s automatic compactions read as started, finished with token counts, failed, or cancelled, and a failed MCP server reads as an MCP row.
- pi and iFlow compactions, Crush finish parts, and Qwen Code tool telemetry become runtime evidence, with context sizes, durations, statuses, and error classes. Crush 0.100.0 subagent reports become task notification rows that keep their output, and the host context and notes it sends ahead of a prompt read as a prompt context row. Factory Droid 0.23x compaction notices count the skills and images carried over, and Prime 0.9.5 and 0.9.6 compaction rows keep their harness state digest.
- Sessions that failed to import now import: Mistral Vibe 3 Unified stores at store format minors 9 and 10, MiMo Code 0.1.16, and pi 1.5.0 sessions whose usage records a provider-reported total. A moved Mistral Vibe session keeps its new working directory.
- Text that was not the person’s no longer reads as theirs: Qwen Code
0.26.2’s memory catalog snapshot is hidden, and Cline SDK messages that
its runtime wrote become system reminder, compaction, and provider error
rows. MiniMax Code 0.7 system errors become provider error rows instead of
model text, and its
run_codechild calls link to their parent. Cline SDK tool activity that the provider ran reads as call and result rows, generated media as attachments, and a forked session as a branch of its source. MiniMax Code sessions read from the pi history gain therun_codechild calls their display rows name, with each saved image on its tool result. Qwen Code 0.27.0’s title sidecar names a session whose transcript lost its custom title or first prompt. - A GitHub Copilot CLI 1.0.106 session whose
events.jsonlrolled over imports every segment instead of only the newest, and a gptme 0.35.0 conversation first imported after a manual compaction keeps its earlier history fromlossless.jsonl. - New Warp conversations keep their billed cost after Warp 0.2026.10.14
retired its earlier cost field, and the fields Warp 0.2026.10.07 through
0.2026.10.21 add, such as web fetch charges and a command that did not
start, no longer count as unknown. Advisor calls that Auggie 0.37.0
writes draw as Augment tool calls. Kimi Code 2.3.0 through 2.5.0, DeepSeek
Harness dsh-v0.2.1, Prime v0.10.1, and Devin Desktop records no longer
count as unknown. Kimi Code 2.5’s workspace catalog supplies a working
directory, and the Devin CLI’s sessions are read from
DEVIN_SESSIONS_DIRwhen it is set, which CLI 3000.11.92 supports. - Claude.ai browser captures and exports keep a long history of inputs in
pages, and capture continues from a verified checkpoint. Raw rebuild
reads the original inputs page by page. Failed captures resume through
their retries or
hansard browser retry. Histories already marked as unavailable for raw replay keep that status and their archived content. Devin browser captures keep messages that hold only an attachment. - Instinct, Poke, and ChatGPT texts fold each Messages reaction, tapback, and sticker onto the message it names, record the message a thread reply answers, and read a vCard attachment as a contact card. The refresh counts a folded reaction as a kept row, so these conversations are parsed again even though they list fewer rows.
- If the watcher or Refresh after updates is off, run
hansard rebuild --stale-parsersto re-read available native history with current importers. Usehansard rebuild --from-raw --since allfor supported preserved captures and for the Devin CLI, OpenCode, and MiMo Code SQLite snapshots, which automatic raw replay skips. - The inferred title version moves from 14 to 16. A conversation whose title
came from a delivered Hansard agent message is titled again from its
transcript. A request that only points elsewhere, such as one to read the
message above or to treat it as the goal, is titled by the words it
names, else by the next request, and a provider title that is only such a
pointer titles nothing. Refresh after updates retitles stored
conversations in the background. With it off, run
hansard rebuild --since all. Derived file and search index versions are otherwise unchanged. - Cached conversation lists refresh on their own for the new
Not in a project folder names and for ChatGPT texts, which now list
under their ChatGPT dot room. Persisted Stats refresh on their own for the
new model prices and fast-tier pricing. The experience index’s rule
version moves to 3, which records the memories a conversation read: the
watcher’s pass derives conversations from the last 30 days again, and
hansard experience rebuildcovers older ones. - Playbook tools are now on by default. New local agent sessions list
save_playbook,propose_playbook, andlearn, which add about 2,000 tokens of tool definitions. A configuration written before this default records no choice and takes it. An agent asked for a playbook saves it in the same conversation, and an agent may offer one playbook on its own after substantial work, which waits for approval. Turn off Let agents save playbooks in Library → Playbooks or Settings → Coding agents, or runhansard mcp tools disable playbooks. Hosted assistants and sessions using--no-memoriesdo not list these tools. - Approvals that the plugin bridge remembered before this release expire once, so each bridged tool that asks prompts again. The bridge now reads a tool’s own MCP annotations to decide whether it only reads, and a tool group’s mode can differ for each agent. Computer control in Settings → Coding agents starts on for every agent. Turned off for an agent, it withholds every tool that controls the computer from that agent.
- API key accounts now spend only after explicit consent. Keys saved before
this release stay off and serve no routed request until spending is
allowed. Turn on Allow API key spending on the key’s page in Accounts,
or run
hansard account api-key save --id <account-id> --allow-spend true. Replacing a key turns spending off until it is allowed again. The key’s page replaces the former Manage dialog, and turning off Allow API key spending replaces Pause account. Caps and project rules still apply. Update every routing device before adding an OpenRouter, Bedrock, or Vertex AI key: earlier copies carry its synced record without applying it. No reimport is required. - The managed router now pins CLIProxyAPI 8.0.63, which fixes the two
denial-of-service paths present in 7.3.20 and adds Codex refresh recovery,
Claude Fable 429 classification, and Muse Code header fixes. An app update
or restart keeps the installed router binary, whether 7.3.20 or 8.0.47.
Run
hansard account proxy install --allow-major-upgradeto install the pinned release. The flag is needed only when moving from version 7. A running router keeps serving until it has been quiet for 65 minutes, and then its worker starts the installed release. For an immediate upgrade, runhansard account proxy stop, the install, and thenhansard account proxy start, which ends a router the stop left answering, where earlier builds asked for a computer restart. Stopping interrupts active requests, which need a retry. The upgrade preserves keys and saved sign-ins. Version 8.0.63 keeps Codex refresh claims and their receipts in.codex-refresh/under the router’s auth directory: backups and restores must keep that folder with the credentials, and a router that reports an uncertain refresh needs a new sign-in or the credentials of the copy that refreshed successfully. - Running roster agents that stop on a usage or rate limit now continue once
their account can serve them again. An agent pinned to a router account
waits for that account’s own reset plus two to five minutes and is never
moved. Receipts and the roster say when each one continues.
hansard config set orchestrate.reviveAfterLimit falseturns this off. Revival does not reopen agents that a computer restart stopped. - Claude Code and Codex agents that
start_agentsstarted now pause once they finish their turn and sit idle, like other conversations, unless the person typed in them. A session can start any number of helpers, andorchestrate.delegation.maxChildrenPerParentis retired: it is ignored on load andhansard config setrefuses it. A project’s Helper budget counts only helpers that are waiting to start or running. - In Agent messaging → All projects, agents can now message sessions in other projects on their own, not only find them, when the device knows the sender’s project. Same project keeps their messages in their project unless the person asks them to contact another one.
- On macOS, a new Hansard home stores its contents in a sibling
.noindexfolder,~/.hansard.noindexby default, with~/.hansardkept as a compatibility link, so Spotlight leaves archive contents out of its index. Existing homes stay where they are unless migration is chosen in Settings → Imports → Maintenance or withhansard config set archive.spotlightExclusion migrate. The installed app’s watcher then moves the home once it can pause archive services safely. Filesystem backups must include the physical folder, not only its link.hansard doctorreports the layout. Windows and Linux keep their layout. No rebuild or reimport is needed. - Restart agent sessions to load
research_past_sessions, the context source tools, and the playbook tools. Where Hansard’s managed local MCP server is installed, integration sync turns off Hansard’s discovered ChatGPT app in local Codex and its claude.ai connector in Claude Code, and removing the integration restores switches that still hold Hansard’s values. Unlessintegrations.autoInstallisfalse, the integration pass installs Hansard’s integration for supported agents it detects for the first time, which now include Goose, Qwen Code, and Crush. It refreshes skills and commands whose installed copy is still exactly what Hansard wrote, and keeps copies the person edited. A Mistral Vibe install that had only the turn-end hook gains the MCP entry, skills, and tool hook on the next sync. The sync also writes Hansard’s MCP entry for the Cline CLI into~/.cline/data/settings/cline_mcp_settings.jsonand for Kimi Code into~/.kimi-code/mcp.json, the files each one reads, instead of files they do not read. - A notification for a subscription plan change starts on, under its own
switch in Settings → Notifications. A Personal Sync backup that goes
over its storage budget, or would pass it within 30 days at its last
month’s growth, raises one notification under Import or sync needs you.
Hansard deletes nothing to meet a budget. While a device follows the remote
host, the Mac stays awake on AC power. Turn off Keep this Mac awake in
Settings → Remote access → Advanced, or run
hansard config set remoteAccess.keepAwake false. - The Library reads archived conversations again once, in resumable watcher
passes, for its new collection rules, and gives stored items whose files
can still be read a version that holds them. It now also keeps some found
items on its own: pages and components, files an agent ran or read again
after writing them, files a later conversation used, and files that a
conversation’s last reply names. Each such item says
Kept automatically, and Personal Sync uploads it with the other kept
items. Return to Found in the item’s More actions, or
hansard library unkeep <id>, undoes this, and the rules never keep that item again. A file that the scratch source and a conversation’s write both collected becomes one item once. This needs no rebuild. - Team uploads and receives now leave out every Library file. Earlier releases uploaded kept Library items’ records and files to the storage of a Team with no match rule, where teammates’ Hansard stored them without listing them. Copies already uploaded stay in that storage.
- Extension
0.8.0.0retries each conversation on its own schedule, so another account’s conversations, unfinished replies, and failing conversations no longer hold up the rest. A ChatGPT or Claude conversation that fails, times out, or is refused fails alone and is set aside after three tries instead of pausing the provider, and large conversations get 90 seconds to load. It follows newer Gemini, Z.ai, Devin, Kimi, and NotebookLM pages, lists Devin sessions with their attachments and recordings, captures a ChatGPT dot’s visible tasks and threads into its room, and reports its browser, health, and scan counts to Hansard. A page observer that an extension reload left in an open tab stops instead of throwing on each change. After it updates, Rescan all history fetches the Devin attachments and recordings that earlier captures left out. It needs no new permission and ships separately from npm and desktop updates, with its own store review. With an earlier extension, Settings explains a quiet archive from Hansard’s side alone. - The Windows x64 preview moves from 0.7.1 and stays unsigned. Download the installer with Settings → About → Check now or from hansard.dev, quit Hansard, and run it. The installer registers login startup for an archive that is already set up.
- Context sources. Agents can read other apps’ local stores without
Hansard writing to them: the messages, contacts, and calls that Cued or
Nexus collect, and any SQLite file.
0.7.1
macOS
Windows
The Library keeps reusable work beside the conversations that made it. Growing conversations import and sync with fewer repeated reads, and messages to working agents keep their delivery state. Accounts explains sign-in problems, Graph names its groups, and the phone uses the desktop transcript.
Added
- Library. Search scripts, components, pages, documents, images, and
snippets by project and tags. Each item keeps its source conversation,
versions, and a sandboxed preview. Found groups collected work by task
and supports keeping or dismissing several items together. Agents use
search_library,get_library_item, andsave_to_librarythrough the installed library skill.hansard libraryprovides the command interface. - Add reusable work. Save to library appears on conversation code
blocks and tool output. The Library’s Add button and drag-and-drop accept
files and folders.
hansard library add <path>links large folders and font collections as References instead of copying them. Automatic linking applies above 16 MB or 200 files, or to collections that are mostly fonts.--referenceselects linking explicitly. References include file search and font previews, and tags remain editable withhansard library tag. - Collected work. The Library collects preserved Claude artifacts, ChatGPT Canvas versions, generated files and images, agent plans, and complete scratch-file writes from archived tool calls. Historical scratch copies wait in Found for 90 days unless kept. Stored Devin browser reports and experimental Gemini Canvas documents join this collection. Private links without captured files remain unavailable. Items that may contain secrets are held without their content.
- Library on other devices. Personal Sync carries kept items and their versions. Other devices lists received items read-only. A Reference syncs its manifest, not its linked files or local search index. Found items remain local until kept.
- Sign-in recovery. Accounts offers guarded sign-in within the window
and Use a code instead for Codex, Grok Build, and Muse Code. The equivalent command
is
hansard account login <label-or-id> --device-code. Accounts also shows local sign-in failures and recovery history, alongside each connection’s current state and reason. Native sign-in options follow each agent’s supported methods. Cursor CLI adds Sign in on another device with Copy link, also available withhansard account login <label-or-id> --browser-link. Devin adds Paste a token, also available withhansard account login <label-or-id> --manual-token. Hansard forwards the token to Devin’s native prompt without retaining it. Both methods keep the saved account’s identity check. - Poke imports. On macOS, Settings → Imports → Messages assistants
imports Poke conversations from Messages after its contact number is set.
Poke uses its own source and number setting beside Instinct. The CLI reads
it with
hansard import --sources poke --since all. - Remote agent work. Agents can list paired computers, start a run,
follow its steps, and stop it through the remote-device tools.
hansard remote devices,hansard remote follow, andhansard remote stopexpose the same work in the CLI. Approvals and questions return to the originating computer. Device rows show agent readiness and offer an update when the other computer runs an older release or build, with the person’s approval. - Running and conversation controls. Running adds a Messages pane
with queued words and delivery receipts.
hansard psincludes subagents and paused conversations, withpause,resume, andevictcontrols.hansard show <id> --agentslists a conversation’s runs, andhansard rename <id> <title>sets a name from the terminal. Show idle agents is one remembered switch shared by Running’s Graph and Lanes overviews. It starts off, while agents with recent work or a working subagent remain visible. - Stats comparisons. Compare up to four projects, including their weekly
activity ranks. The Agents tab compares kinds of agents and work passed
between agents. Runtime evidence separates approvals, waits, failures, and
retries by harness. The Team picker and
hansard stats --team <team>scope Stats to a joined Team. - Graph groups. Large conversation views open as named clusters that expand into their members. All Conversations covers the archive, search highlights matching nodes and clusters in place, and keyboard navigation reaches groups and conversations. Temporary folders can form one group. Cluster names in Settings → Imports → Search uses Apple’s on-device model where available, or an optional 1.28 GB model download elsewhere. Keyword names remain available without a model. Naming stays on the computer.
- Saved Muse Code accounts.
hansard account add <label> --provider musecreates a separate file-based login. Sign in withhansard account login <label>and launch it withhansard account run <label> -- <arguments>. Hansard verifies the native account identity and reads its allowance before retaining the login. Muse accounts do not join native account pools or change Meta routing support. - Account preferences on other devices. Personal Sync carries explicit
account names, credit-use choices, chat-account links, and dismissed link
suggestions through verified account memberships. API account names, caps,
models, enabled state, and portable project rules also follow the account.
Native credentials, sign-in folders, and connection enablement stay local.
hansard account api-key listlists the API accounts on this device. - Edit local files. Local text and Markdown files up to 1 MiB offer Edit in a conversation’s Files pane and Running’s right pane. Save checks the opened revision and reports external edits. Drafts and undo history survive pane changes for the current page’s lifetime. Received files, remote workspaces, and files outside the project stay read-only.
- Reviewed playbooks. Library → Playbooks keeps procedural preferences
with steps, source quotations, linked examples, and earlier versions.
After the playbooks tool group is enabled,
save_playbookrecords an explicitly requested preference.propose_playbooksends corrections to Memories → Proposed for review. Only a person’s approval adds the proposed procedure to the Library. Kept playbooks use Personal Sync, while automatic prompt or tool-hook suggestions remain unavailable. - Personal Sync category status. Settings and guided setup show counts
for conversations, account records, and kept Library items, with memory
and settings records counted separately. Show sync activity lists pending
changes and last uploads. Show received items lists each device’s copied
counts.
hansard personal-sync status --jsonreturns the same categories. Missing indexes show unknown counts, and an offline device’s received counts do not establish its current upload state. The storage review also names the local conversation count, destination, and sync direction. - Recent project conversations. Running’s History menu opens older conversations beyond the main list’s one-hour window, with paging and Resume for ended conversations. It also lists detached project processes, their ports, and possible conversation origins. Ambiguous origins remain unknown. Stop requires confirmation and rechecks the process identity. Stale idle agents are labeled Stale, and newer finished work can lead the project card.
- Work taxonomy choices. Choose ONET 31.0 or 30.2 in Classification.
TaskNet’s October 2026 task inventory is available through an explicit
local CSV import with
hansard classify taxonomy import-tasknet <file.csv> --accept-terms, after reviewing the source’s terms for permitted personal noncommercial use. Hansard bundles no TaskNet data and does not download it automatically. The import verifies the pinned source file. Select it withhansard classify taxonomy select tasknet-2026-10-27after import.hansard classify taxonomy listshows the choices. Stats keeps TaskNet occupations and tasks separate from ONET’s activity hierarchy.
Improved
- Claude Code subagent imports. An existing subagent update no longer rewrites its parent’s transcript, views, or raw inventory. With a synthetic parent holding 1,270 subagents, measured writes per change fell from 16–20 MB to 0.84–0.90 MB, and CPU time from 4.4–4.7 seconds to 0.33–0.37 seconds. Later SQLite checkpoint writes are outside that import window.
- Codex background imports. Coalescing nearby changes and reusing an import worker reduced a six-append synthetic burst from six children to one. Median import CPU fell from 3.16 seconds to 0.97 seconds, and measured JavaScript file reads from 106.27 MB to 26.24 MB. Median peak child memory rose from 151.67 MiB to 177.76 MiB in that workload.
- Personal Sync reads. Verified transcript prefixes and earlier uploaded raw objects let later uploads reuse unchanged chunks. Three synthetic transcript uploads read 3.15 MB instead of 18.88 MB, with identical destination writes. A separate replay of eight copied Claude Code and Codex sources reduced Personal Sync reads from 4.130 GB to 1.345 GB when several imports occurred between uploads. These are logical file reads, not measurements of physical disk traffic.
- Growing-conversation reads. After checkpoint and raw-hash reuse, verified log buffers, saved messages, and lazy attachment lookup cut further repeated reads. A matched replay of eight copied Claude Code and Codex sources, appending 25.69 MB, reduced logical reads from 2.440 GB to 1.621 GB, a 33.57% reduction. Measured CPU rose from 43.82 to 47.27 seconds. These trials establish a read reduction, not a CPU improvement or an installed-app result. A later matched comparison reused decoded batches and verified stream states, reducing reads from 1.620 GB to 1.409 GB. CPU changed from 47.299 to 46.871 seconds, with overlapping trial ranges. That result is consistent with unchanged CPU, and the added cache retains more memory. The two comparisons do not establish a combined speedup.
- Search storage. Removing redundant two-character prefix postings made an index of 3,010,096 stored documents 10.05% smaller: 3.494 GB became 3.143 GB. Stored search text and embedding signatures stay valid, and ranked results matched across 576 checked query executions. Short queries in unspaced scripts retain prefix matching.
- Responsive status and navigation. Watcher permission checks refresh outside the request thread. An isolated synthetic replay reduced median permission-poll latency from 61.44 ms to 0.75 ms. This measures that check, not a complete viewer request. Sessions and Running can open from saved answers while fresh reads run, and unchanged polls retain existing rows. Composer discovery also reuses file listings and refreshes expired catalogs without blocking their readers.
- Background indexing. Stats schedules refreshes from changed facts. Graph and experience indexing resume from verified transcript positions, and completed similarity rankings survive unchanged passes. Library lists, received-history checks, Team receives, and source-presence checks use bounded indexes or saved change positions.
- Phone conversations. The web phone and iPhone conversation share the desktop transcript renderer, with thinking and compaction details behind expandable rows. Composers remain available while an agent works or the connection recovers. Messages keep their delivery state, and supported live channels receive them without replacing the running terminal session. Other channels deliver queued words when the turn ends or resume the conversation when needed. New conversations offer each agent’s model, effort, permissions, and image attachments. Phone links can cover selected projects, a portfolio, or all allowed projects. Project and conversation responses use smaller pages, and conversation cursors keep their position as messages arrive. Large conversations without prepared pages show a preparation state while the host builds them in the background. On iPhone, active agents appear after the first project page, and later project pages preserve the current navigation.
- Year in review. Each card centers on one figure from the selected year, with daily activity, work rhythms, model arrivals, and the year’s agent and model shares. Charts follow the theme, use readable labels on a phone, and retain their harness and model colors.
- Conversation reading. Long conversations receive a stored outline for their first visible window. Agents and Trace return to the transcript position they replaced. Image previews reserve their own aspect ratio, and terminal panes preserve their screen and scrollback when panes change. Continued and forked conversation chips mark their snapshot boundaries, show the launch time, and identify source conversations that continued later. Final replies stay outside collapsed turns when recorded provider boundaries divide the work, including turns started by agent notes. Existing metadata supplies these boundaries without an archive rebuild.
- Messages assistant numbers. Instinct and Poke settings format numbers as they are entered. The Country selector starts from the Mac’s region and accepts a different country for each assistant. Changing that selector changes the display without changing an already saved number.
- Portfolio suggestions. Settings → Portfolios reads maintained folder identities instead of waiting for project Stats. Unchanged suggestions reuse their saved answer. An unavailable index or failed request shows a reason and Try again. The next archive write prepares a missing index without a reimport.
- Concurrent Settings saves preserve changes made by another window or process. Maintenance requests queue behind running work instead of being lost.
- Failed and interrupted runs retain their reasons in Running. Signed-out runs offer the harness’s own sign-in, and account checks cannot restore a stale failure after a newer sign-in succeeds.
- The router preserves queued usage through restarts, waits for startup readiness, and reports held models as retryable limits. Standby requires evidence that no paid usage can bill beyond the account’s limit.
- A background import without a watcher updates search when it completes. A new archive’s first Sessions list refreshes as soon as conversations arrive, and long conversations find their latest request past large tool outputs.
- Hosted assistant links recover after sleep and keep an open device socket online. Phone conversations remain on screen through temporary host failures, and a reconnect replaces an obsolete event wait. Slow or interrupted phone reads retry with bounded delays. Leaving a conversation cancels its retries, hidden pages start no retry request, and denied reads do not retry.
- Import failures retain retry backoff across watcher restarts. Augment reads only its own editor-store value, Devin Desktop retries a store that changed during its read, and one signed-out Claude cloud profile does not hide a healthy profile’s import.
- Embedded terminals answer foreground, background, and cursor color queries before their renderer attaches, so startup does not wait for the pane to open. Restored screens retain their colors, and changing the theme clears obsolete program color overrides.
- iPhone development and release builds use one application identity. Updates from the same signing team replace that app and retain its pairing records without uninstalling it.
- Peer requests to a live Codex conversation use its native queue. If that queue cannot accept the request, the message stays in Hansard’s inbox. Delivery no longer starts a competing app-server resume for that live thread.
- Running’s folder-history index tolerates malformed catalog entries without blocking archive writes or index creation.
All upgrade notes- Update every Hansard copy that shares an archive, including the desktop app and npm installation, and update each Personal Sync device. Indexed view pages, subject assignments, storage-copy receipts, and helpers with inherited permissions have newer reader requirements. Saved-login records now require account reader version 6, and launch preferences require routing reader version 2, so earlier copies cannot rewrite Muse Code account choices. API accounts awaiting local keys or folder mappings require reader version 3. The derived memory catalog requires reader version 3. New playbook proposals require record reader version 2, preventing earlier copies from approving them as instruction lines. New classification records use schema version 4 and schema registry version 2. Earlier copies preserve these newer records instead of overwriting them. Library and experience stores use writer version 4 with reader floor 1, preserving earlier reads while blocking older writes. Library sharing and sharded Personal Sync inventories wait until every participating reader supports them. These changes need no complete archive reimport.
- Refresh after updates now includes enabled sources whose older
conversations fall outside the normal import window. The watcher queues
one low-priority import run for those sources, replays supported manual
exports from preserved raw copies, and retries imports left failed by
earlier releases. Background refresh runs on AC power while imports are
idle. Settings → Imports → Maintenance and
hansard statusshow its progress. Release preparation resets the 147 existing parser entries to baseline0.7.1.0, replacing their0.7.0.*versions. The new Dot task source,codex-dot-task-browser, uses0.7.1.1. The first refresh can therefore re-read sources without a behavior change listed below. The suffix reset does not reverse earlier importer fixes or require a complete archive reimport. - If the watcher or Refresh after updates is off, run
hansard rebuild --since allto refresh derived files. Runhansard rebuild --stale-parsersto re-read available native history with current importers. Usehansard rebuild --from-raw --since allfor supported preserved captures and SQLite source cohorts, including Devin CLI, OpenCode, and MiMo Code, which automatic raw replay skips. A derived rebuild alone cannot recover fields an earlier importer omitted. Unsupported export replay, missing raw evidence, and a replay that would lose messages remain unchanged. Reimport the original export when its replay is unsupported. Rebuild everything, also available ashansard rebuild --everything, requests a full background refresh and is not required for an ordinary update. - Conversation-name changes re-read Claude Code, its SDK and desktop/cloud surfaces, Codex CLI and SDK, Cursor CLI, Copilot, Factory Droid, Grok Build, Qwen Code, Kimi Code, Mistral Vibe, CodeBuddy, MiMo Code, and the other importers that retain explicit names. They also cover Augment, Devin CLI, OpenCode, Kilo Code, Crush, Cline CLI, Continue, Posit Assistant, gptme, OpenHands, DeepSeek Harness, Windsurf, pi, Prime Agent, Goose, Zed, Hermes, OpenClaw, Junie, Qoder, Warp child agents, and Roomote. Perplexity and DeepSeek browser captures retain recorded renames. Codex app adaptive titles remain provider titles. Names set in Hansard continue to outrank all of them.
- Claude Code, Codex, Cursor, Copilot cloud, Factory, Grok Build, iFlow, Qoder, and Antigravity imports distinguish more harness records from human messages. Stats can therefore report fewer human turns after refresh. Antigravity also counts retries only after the first attempt, records file writes as edits, and retains screenshots and recordings from brain folders. Claude Code imports store changed subagent runs separately from their parent. The watcher re-reads these sources automatically. The manual native and raw commands above provide the same importer updates where their evidence exists.
- NotebookLM browser imports retain open Studio notes, and experimental Gemini
browser imports retain Canvas documents present in captured history.
Existing supported captures can be replayed with
hansard rebuild --from-raw --source-type notebooklm-browser --since allandhansard rebuild --from-raw --source-type gemini-web-browser --since all. Earlier captures without those bytes need a new capture. The browser extension ships separately, and a source collector is not a claim of Chrome Web Store availability. Gemini Canvas live capture remains experimental. Experimental Dot imports also retain task references, including reported parents and visibility. Refresh retained references withhansard rebuild --from-raw --source-type chatgpt-dots-browser --since all. A reference does not contain a delegated task’s transcript. This refresh does not fetch missing tasks or infer absent visibility. - View-page manifests move to version 4. The refresh converts compatible page
manifests, builds conversation outlines, and converts existing Claude Code
subagent summaries from archived evidence.
hansard rebuild --since allperforms these steps without reading provider stores. Search moves from index version 22 to 26 through schema changes and a copy of stored search rows. Its document signatures and embeddings remain valid. Stats, Graph, subject assignments, and Library indexes refresh their derived data without a complete reimport or manual index deletion. The memory catalog rebuilds its derived counts automatically. Playbook store updates need no reimport. - Resume command placement now accompanies automatic pause unless
agents.resumeCommandHookwas explicitly disabled. Hansard places the command in supported cmux or tmux terminals, or installs a shell hook in~/.zshrcor~/.bashrc. The separate setup and Settings switch is gone.hansard config set agents.resumeCommandHook falsedisables placement and removes the hook. Pause idle agents remains available. - Notifications now have their own Settings → Notifications page.
Approvals, questions, failed runs, lost Claude or ChatGPT sign-ins, and
import or sync problems notify by default. Finished runs, coordinator
follow-ups, and available updates start off. Destinations include macOS,
cmux, and a paired phone, with Focus, quiet hours, and burst controls.
hansard notify statusshows the choices.hansard config set notifications.enabled falseturns them off. Slack keeps its separate configuration. - Saved logins belonging to different accounts no longer share a sign-in
folder. The next usage check repairs shared folders, preserves earlier
settings and history, and names any account that needs a new sign-in.
Check Accounts → Connections or
hansard account statusafter updating. Signing in again also restores a paused router copy when its account matches. - Personal Sync → Cloud accounts chooses one device to capture each
shared cloud account. Other devices receive that history through Personal
Sync.
hansard personal-sync cloud-accountsshows the selected device, andhansard personal-sync cloud-accounts owner <account-id> --device <device-id|this>changes it. The former Instinct-only settings become Messages assistants, with separate Instinct and Poke number fields. Existing Instinct settings remain. Both fields now validate complete phone numbers. Invalid saved numbers need correction before capture resumes. CLI configuration still requires a leading+and country code. - Privacy & Sharing → Hide secrets now shows redaction rules directly.
Enabled redaction also masks high-entropy Bearer credentials. Run
hansard redact reapplyto apply current rules to earlier local conversations whose originals the reveal cache can reconstruct. This does not redact preserved provider files or remove copies already shared elsewhere. - Restart agent sessions to load refreshed integrations and the Library and
remote-device tools.
hansard account login,hansard personal-sync auth, andhansard browser pairnow require a person’s interactive terminal, outside an agent session. The watcher and remote host can move to a newer installed Hansard build automatically, and service recovery respects a recorded stop and pauses during desktop installation. - Personal Sync now carries API account metadata. A received account needs Accounts → Add key on this device before it can route requests. Folder-restricted accounts also need local folder mappings. Keys, folder paths, spend totals, and pending spend reservations remain local. A synced cap limits each device’s observed spend, not combined spending across devices. These account changes need no archive rebuild or reimport.
- Extension
0.7.1.1adds optional access tohttps://codex-cloud-backend.chatgpt.com/*for experimental Dot task transcripts. It ships separately from npm and desktop updates. After installing that extension release, enable its ChatGPT switch, grant the additional host, and use Rescan all history. Tasks arrive as separate Codex cloud sessions with their recorded parents and Dot room roots. Replaying old room references cannot fetch those transcripts. Store review and publication remain separate, and live-provider verification of Dot task capture remains pending. - Settings → Classification → Work taxonomy now defaults to ONET 31.0
for new work labels. ONET 30.2 remains available with
hansard classify taxonomy select onet-30.2. Existing labels keep their recorded taxonomy and remain current. Selecting a taxonomy starts no backfill and does not enable classification. Stats groups work labels by taxonomy and release and refreshes its derived data without a reimport. Optional relabeling useshansard classify taxonomy backfill --since 30d --limit 200after work performed labels are enabled. It uses the configured backend, requires a finite window, and allows at most 1,000 jobs with one worker. The task run also refreshes horizon and difficulty labels and respects retry and usage limits. A derived rebuild alone does not relabel work. - Playbook write tools start off.
hansard mcp tools enable playbooksenablessave_playbookandpropose_playbookfor new local agent sessions. Existing playbooks remain searchable and readable while writes are off. Hosted assistants and sessions using--no-memoriesdo not expose either write tool. - The Windows x64 preview moves from 0.7.0 and stays unsigned. Download the installer with Settings → About → Check now or from hansard.dev, quit Hansard, and run it.
- Library. Search scripts, components, pages, documents, images, and
snippets by project and tags. Each item keeps its source conversation,
versions, and a sandboxed preview. Found groups collected work by task
and supports keeping or dismissing several items together. Agents use
0.7.0
macOS
Windows
Running becomes a workspace for each project’s agents: conversations open as tabs beside their terminals, changes, and files, Graph and Lanes show what every running agent is doing, and idle agents pause on their own and resume where they ran. A Review view gathers what waits on the person, and a conversation gains a step timeline, Trace and Agents views, and a Details pane that opens pull requests. Agents can ask their person a question and read the history of a file or a task, projects gain environments and plugins that Hansard keeps, and a phone connects through a one-use link. The router adds API key accounts with spend caps and keeps each conversation on one account, and after an update the watcher brings the archive to the new release’s files on its own.
Added
- Work in Running by project. The rail holds one card per project with its
lead agent’s line, state, branch, folder, and ports, the selected project’s
conversations open as tabs, and a right pane holds up to three panes of
Changes, Files, terminals, and the ports agents listen on. The
Activity overview lists memory and CPU by checkout, Graph draws
running agents with their subagents, peers, tools, and MCP servers, and
Lanes shows each agent’s last 30 minutes of work.
hansard ps --by-folderlists memory by checkout. - Resume a conversation in a terminal inside Hansard with Resume in terminal, continue it in cmux, Terminal, iTerm2, or Ghostty, show up to four agents side by side in the Terminals display, and bring an older conversation back with Open in Running.
- Pause idle agents and resume them where they ran, from Running, the menu bar, and Settings → Handoff & helpers → Pause and Resume. A message to a paused agent waits with a receipt that says so, or resumes the agent while Resume for requests is on.
- Open a dashboard from a menu bar item in the macOS desktop app: a popover of Agents, Memory, Usage, Accounts, and Imports that marks what needs attention.
- Decide what waits in one Review view in the dock: teammates’ requests, held messages, agent questions and approvals, plugin access, coordinator decisions, Slack approvals, and proposed lessons, with All, High risk, and Overdue filters and Undo.
- Read a conversation as a Trace, with turns, model responses, tool calls, and subagent runs on one active-time axis, or as Agents, with its subagents, the agents it started, and its peers, from View options → Transcript view.
- Open a conversation’s Details beside the column, or as a drawer where it
does not fit: its context by category with its growth across compactions,
its MCP servers under Connectors, its model, effort, and running
process, and its changes beside the checkout’s uncommitted ones.
Create PR commits the checked files from a private index, pushes them,
and opens the pull request with
gh. - Manage projects on the Projects page and with
hansard projects. Every GitHub folder becomes a project in its owner’s organization, portfolios group projects, and environments name the folders, account, and plugins a conversation starts with, in a new git worktree for each conversation when turned on. Plugins added from a marketplace, a git repository, or a folder are kept by Hansard and turned on per environment, and Skills & plugins charts each skill’s and server’s uses. - Let agents ask their person a question with
ask_userand wait for the answer, given in a dialog from the title band, a notification, Review, a paired phone, or the Slack bot’s direct message. Agents also attach files to messages, and read a file’s or symbol’s history withcode_historyand a task’s arc withwork_context, whichhansard why <file>andhansard work-contextprint. The watcher keeps the index behind them without model requests, andhansard experience statusreports it. - Add OpenAI, Anthropic, Gemini, and xAI API keys as router accounts with
hansard account api-key saveor Add account → API key account: daily and monthly spend caps in USD, rules that limit a key to organizations (in Accounts), projects, portfolios, or folders, and Pause account. A key without rules serves only when the subscription pool cannot.hansard account proxy rotate-keyreplaces Hansard’s router client keys and retires the old ones after a grace period, 24 hours by default. - Use a usage reset from Accounts or Usage. Use reset redeems a Codex
reset after one confirmation, and for a Claude login whose Claude Code
offers it, opens Claude Code’s own
/limit-reset, whichhansard account limit-resetalso opens. - Turn on Send Claude Code activity to Hansard in Settings → Handoff & helpers to read Claude Code’s metadata-only telemetry into Running errors, compaction sizes, and a Stats Claude Code activity section. Hansard keeps 14 days of it on the computer and never syncs it. Stats adds a Runtime evidence card of approvals, waits for the person, retries, and failures, and Why was the prompt cache rewritten?, which ranks cache rewrites by cause and cost.
- See whether a conversation’s prompt cache is warm, and what its next message would process again, on a chip in the composer, and in the time’s tooltip on idle Running rows.
- Import Grok Bot desktop conversations from the app’s local copies, which
keep at most 200 records each, and its full history through the opt-in
Grok Bot cloud history source (
grok-bot-cloud) on macOS. - Rename a conversation from its Actions menu or its Sessions row’s menu. The name outranks every other title, and imports and rebuilds keep it. A long conversation’s Sessions row shows its latest request on a Now: line.
- Summarize conversations after they end with the switch of that name in
Settings → Classification, off by default, or with
hansard summaries run. Continuations and recall read the summaries. - Find and switch agents’ integrations, shared MCP servers, apps, plugins, skills, and services from one Integrations picker, and add, repair, or remove what Hansard installed in each coding agent from Settings.
- Connect a phone from Settings → Remote access → Connect a phone.
Hansard prints the
tailscale servecommand to run, then a one-use link and QR code, which expire after 10 minutes, open an installable phone page of the chosen projects’ agents, conversations, and questions. An opt-in switch offers Claude Code approvals on the phone for up to 15 seconds before the terminal prompt takes over. Guided setup beside Connect a phone walks the phone through Tailscale, HTTPS for the tailnet, the private address, and the code to scan, and ends with adding the page to the Home Screen. - Move to a new computer: Hansard notices an identity copied from another
computer, holds uploads, routing, and remote hosting until the person
chooses Keep both or Replace with the old computer’s name, and
hansard device adoptgives the new computer its own identity and Team member keys. - Let a Team’s agents message each other over a direct connection, which a
manager allows for the Team and each member turns on with
hansard team sharing <team> --agent-messages on. It needs the macOS app on each computer and a relay that the Team runs itself and names withhansard team network <team> --relay <https-url>; Hansard does not ship a relay. Teammates see each other’s live sessions in shared projects, and held messages wait under From other people in Direct messages. - Share a project’s board with a Team that allows agent messages, from
Board sharing in Boards or with
hansard boards sharing, off by default for each project. Boards travel through Team storage and need no relay. - Offload local raw and attachment objects older than 30 days with
hansard storage offloadonce two backup destinations hold verified copies, and bring them back withhansard storage restore. Both preview until--yes, andstorage.offload.enabledturns offload on. - Take prompts from Slack: the bot’s owner can start runs without an approval click, send tasks by direct message, and message live agents on the Mac, and Channel access sets who may ask in each channel, whether their runs need approval, and run limits. A run someone other than the owner asks for runs in a home and sandbox of its own, and messages are cards edited in place.
- List the record types each source holds, and any an importer does not
recognize, with
hansard doctor records.
Improved
- Open the Graph faster. With 20,000 synthetic conversations, the first
All graph draws in about a second instead of about 20 seconds, and a
large view opens arranged on the next visit, since each view’s settled
places are kept. Graph keeps its answers across restarts, shows a read-only graph
on battery power instead of Graph unavailable, rebuilds its renderer
after a lost WebGL context, and in the macOS app ranks through a native
service. A Work tab draws conversations with the files, commits, test
failures, decisions, and memories they produced and their outcomes
(Node kinds adds tasks), and
hansard graph compactpacks the Graph index’s references and reclaims free pages without rebuilding it, with twice the index size plus 2 GiB of free disk. - Open long conversations of large tool outputs faster. View pages break at 1 MiB as well as at 300 messages, so a synthetic conversation of 600 messages with 200 outputs of 250 KB opens in 10 ms instead of 85, at a peak of 212 MB instead of 764. A view the viewer cannot hold is derived in a child sized for it, so a 188 MB transcript opens with the viewer at 58 MB instead of running it out of heap, and a search result opens at its message and stays there while rows settle.
- Build the session tree in a worker thread, so the longest pause of the tree route on 21,393 changing synthetic sessions fell from 719 ms to 16. The first frame takes the stored theme, and Sessions shows the rows of the last visit before the archive answers.
- Read a turn’s tool calls as one step timeline, with dead ends, parallel calls, and subagent branches. The Worked for row counts active time without approval waits and leads with the turn’s outcome, a conversation pins its latest task list at the top, and a recap card waits at the point the reader left.
- Open every menu and picker in one shared dropdown with the same look, submenus, and keyboard behavior: the Actions menu with its Continue conversation, Open in, and Copy submenus, the composer’s pickers, Deep Recall, account and Teams menus, and row and right-click menus. Disabled rows keep focus and name their reason.
- Pause the viewer’s motion while the window is hidden or unfocused, or after 60 seconds without input: animations, spinners, shimmer, Running clocks and pulses, countdown displays, and setup demos stop, and resume on focus or input.
- Pick a model, effort, and speed for every harness the composer starts. Each harness lists its own models or says who chooses and why, Speed offers Claude Code’s fast mode, the Cursor CLI’s fast model, and Codex’s tiers where they exist, a routed harness lists what its router accounts and API keys serve, and an Account row starts a conversation on one saved sign-in. Composer images reach Claude Code, Codex, and ACP agents as images.
- Read Boards like a chat app: projects unfold into channels, threads list newest first with new posts divided from earlier ones, the header names threads, participants, and the last post beside a Follow this board switch, and mentions open the shared picker.
- Set up Hansard on a page that fills the window, with one settings card per step beside a live demo. New in Hansard opens as a short sequence that shows each highlight and runs the chosen steps only at Finish, with the reason and Run again beside a step that fails. The Web chats guide walks through adding, pinning, and pairing the browser extension and turning on its sites, then waits for the first saved chat. On a phone-wide sheet, the close button keeps room of its own with a 44 px touch target, and a guide’s name and step count sit on their own line above the buttons.
- Set up bucket storage step by step. Under More storage options, Cloudflare R2, Amazon S3, and Other S3-compatible open a walkthrough that links each page to open and the provider’s own guide, draws that page with the site’s labels, and holds the fields its step produces. An R2 address that ends in a bucket fills Bucket as well, and guided setup’s Personal Sync guide uses the same walkthrough. Google Drive and OneDrive walk through registering the OAuth app that Sign in with this app uses.
- Find a setting by its configuration key in Settings search, pause background work in Settings → Imports → Search, see what each sharing feature sends under Privacy & Sharing, and open setup again with Run setup.
- Run search by meaning in less memory: the text server holds about 110 MiB when idle and peaks at 292 MiB during a pass instead of 647, images embed in a server of their own that stops after its pass, and the model stays loaded between passes.
- Release what idle processes hold: the viewer’s Graph worker stops after ten idle minutes once its index is complete, the Stats engine, the watcher, and the MCP host return freed heap, and recall calls from every agent session share one warm query child that never waits for a worker slot.
- Import growing conversations with less work: the watcher’s targeted imports run in one warm worker, parser checkpoints cover transcripts up to about 1 GB, a complete Claude Code scan after a restart ends in one batch, and transcripts are read a line at a time, so a 185 MB transcript holds 365 MB of heap instead of 735.
- Receive Team changes with less waiting: a first receive runs 32 downloads at once instead of 16, so 2,000 synthetic shared conversations arrive in about 27 seconds instead of 34, an upload on a fixed interval runs as soon as the archive changed, and a growing conversation’s Markdown uploads at most every 15 minutes, and again once its transcript rests. Compress shared files stores text files as gzip when that saves at least a tenth.
- Keep a receiver’s versions of a growing transcript as clones that share their unchanged blocks, so a 36 MB transcript that grew 8 times takes 36.4 MB of disk instead of 310.6 MB.
- On a Mac, make PDFs from the Share dialog without a print panel, with one 18 mm margin on every page, and save every format of any conversation through files the viewer writes.
- Group every captured chat service as a web chat, with the chat account’s
avatar and a
<service> (Uncategorized)folder for chats outside a project, link a chat account to an Accounts account with Link chat account…, and show Dots as its own app beside ChatGPT. - File a Codex cloud task that the browser extension’s source build captures under the repository its recorded GitHub pull requests name, when they name exactly one. An environment label alone assigns no repository.
- Refresh model pricing and the model catalog while keeping earlier entries.
- Import a Codex rollout that Codex compressed after a week without writes with its title, surface, and subagent links, and count a rollout that is gone as a removed source instead of a read failure that held back the Codex source.
- Serve a conversation whose transcript is over the viewer’s limit from its stored Markdown, so a 440 MB live transcript no longer runs the viewer out of heap, and count Personal Sync setup’s conversations without reading the whole catalog, which ran a viewer with 21,000 conversations out of heap.
- Keep one oversized transcript from ending
hansard rebuildout of heap. The command leaves that conversation as it is, counts it under Left for a larger heap, and prints theNODE_OPTIONS=--max-old-space-sizevalue that refreshes it. It also measures the heap it has left correctly on Node.js 24, which the desktop app carries. - Replace the desktop app under agent sessions’ recall servers, which move to
the new app’s MCP host, instead of waiting for them. Restart to update?
names the work a restart would stop and can wait for it, and
hansard app processeslists what holds the app. - Finish a large Personal Sync first receive on Windows: storage copies
inside the archive no longer count toward the 10,000 tracked files,
requests that fail with
ECONNABORTEDare retried, and a receive reports progress while it replays history, so the watcher no longer stops it. - Report a Personal Sync capture that automatic recovery is finishing as in
progress instead of an error.
hansard personal-sync statuslists it under Captures recovering automatically, Devices in Settings → Personal Sync counts it among the changes waiting to upload, and the conversation’s next capture waits behind it instead of failing the upload. An interrupted capture of a conversation whose source grew past the 64 MiB reread limit is withdrawn at its first recovery instead of after a 15-minute wait, and the conversation uploads again from the archive’s copy. - Re-parse conversations received from other devices without aborting on every pass. Each re-parse process holds one conversation’s metadata at a time, so one that sets thousands aside for a full cache no longer runs out of heap between them. A process that exits between conversations raises the heap of the one it was preparing, and the rest go on in a new process. A cache held at its size limit is no longer measured again at every receive check.
- Keep a saved login’s new sign-in only when it belongs to that account, leave the earlier sign-in otherwise, and show a refused sign-in with Sign in again and Add as a new account.
- Refuse a relative
HANSARD_HOME, which created an archive in the working directory, and read a leading~or~/in it as the home folder. - Read a Devin CLI node whose metadata holds JSON null instead of failing the whole Devin CLI import, and link Antigravity subagent runs to their child conversations again.
- Treat an encrypted TRAE database without its key as a setup condition
instead of an import error on every import.
hansard doctorlists it once under source setup, the TRAE row in Settings → Imports names both ways to unlock it, and memory backup skips it. Hansard tries it again when the database changes orTRAE_DB_KEY,HANSARD_TRAE_DATABASES, orHANSARD_TRAE_SQLCIPHERis set, and a key that SQLCipher refuses stays an import error. - Scan for TRAE Agent trajectories only in
trajectories/at the top of each discovery root and git checkout, instead of five levels into every project. A trajectory in a checkout’s subfolder needsHANSARD_TRAE_AGENT_ROOTSorHANSARD_TRAE_AGENT_FILES. A scan that reaches its limit no longer fails the rest of the TRAE import, and the watcher retries only the scan after 30 minutes, then at doubling intervals up to six hours. - Stop asking for attention for a failed import source once a later import of it completes.
- Count only time the computer is awake toward the watcher’s time limits: an import’s 60-second stall limit and its 10- and 30-minute run limits, the 10-minute remote sync stall check, and the limits on the watcher’s indexing, Stats, classification, maintenance, embedding, and summary work. Work that a sleep interrupts continues after the computer wakes instead of being stopped and reported as stalled.
- Keep a phone’s open agent conversation from failing with This agent is no longer available in this project between process scans.
- Cut an embedding input longer than the server’s 2,048-token batch to fit instead of failing its conversation, and ask Ollama not to think when labeling, so a thinking model returns its label before the request times out.
- Keep saving a Claude conversation whose browser captures reached their
limit through repeated settings observations. Hansard retires the
redundant observations and keeps the clocks later updates need, captures
that failed retry on their own after the update, and
hansard browser retryretries them at once. - Show Perplexity Computer replies, user messages, and thinking from the browser extension’s source build in the conversation instead of hiding them as provider context. Refuse a Claude Code on the web capture whose last event page says more events follow, so it waits as a failed capture until Rescan all history in the extension reads it in full, instead of archiving the conversation incomplete.
- Keep a turned-on site’s browser access when another site that shares it is turned off in the extension’s source build, such as ChatGPT’s when Codex cloud is turned off.
All upgrade notes- This release is published for the macOS desktop app on Apple silicon, the
Windows x64 preview, and the npm package. Mac apps from 0.6.1 through 0.6.5
update in place with Update and restart, and npm installations upgrade
with
npm install -g hansard@latest. npm 0.6.3, 0.6.4, and 0.6.5 were not published, so npm installations move from 0.6.2 to this release, and the 0.6.5, 0.6.4, and 0.6.3 notes below apply to them as well. - The Windows preview moves from 0.6.4 and stays unsigned. Download the installer with Settings → About → Check now or from hansard.dev, quit Hansard, and run it. No Windows installer was published for 0.6.5, so the 0.6.5 notes below apply to a Windows installation as well.
- The npm package now needs Node.js 22.12 or later, or 20.19 or later on
Node.js 20, the first releases that load ES modules with
require(). Node.js 21 is not supported. On an older Node.js,hansardstops with a message that names the installed version. The desktop app carries its own Node.js. - First update every copy of Hansard installed on a computer: the desktop app and an npm installation alike. As in 0.6.5, archive format 2 conversions, raw compaction, and the Personal Sync formats that need newer readers wait while an earlier copy is installed or running.
- After an update, the watcher now brings archived conversations to the files
the new release writes. In background batches on AC power, while no import
runs, it does the work of
hansard rebuild --since all, converts conversations to archive format 2 and compacts raw source copies once every installed copy and Personal Sync device reads them, and parses a conversation whose provider file is gone again from Hansard’s own raw copy, never keeping fewer messages. This covers thehansard rebuild --since allandhansard rebuild --compact-rawsteps of the 0.6.5 notes and brings this release’s runtime summaries, cache rewrite causes, task lists, and titles to existing conversations without a reimport.hansard statusshows its progress as Archive refresh. Once a compacted conversation uploads, its Personal Sync backup requires 0.6.0 or later on every device for good. - Turn the refresh off with Refresh after updates in
Settings → Imports → Maintenance or
hansard config set archive.autoRefresh false, and then runhansard rebuild --since allonce. An npm installation that does not run the watcher runs that command too. - The search index derives every conversation’s documents again, and search by meaning embeds every conversation again, since both now carry this release’s event and index versions. Both run in the background and can take hours on a large archive, and search keeps answering meanwhile. Search by meaning pauses on battery power, and the watcher installs its new llama.cpp runtime, about 12 MB on macOS and up to 20 MB on Windows and Linux, on its next pass and keeps the downloaded models.
- Most importers changed. Claude Code, the Claude SDK, Codex, Cursor, Devin,
Grok Build, Antigravity, Kimi, MiMo Code, OpenCode, Aider, and many others
now keep runtime evidence (approvals, waits for the person, failures,
retries, compactions, credits, and context usage) and the records they do
not recognize. The Codex importer reads a rollout that Codex compressed as
.jsonl.zstwith its title and subagent links, so a conversation an earlier release archived from one imports again once. The Claude Code importer reads names set with/renameorclaude -n. The watcher reads the conversations in its import window (30 days by default) again with the new importers, and the archive refresh parses again those whose provider files are gone. - Refresh older history with
hansard import --sources claude,claude-sdk,codex-cli,codex-desktop,codex-sdk --since all, adding other sources as needed. Each release starts a new parser baseline, sohansard rebuild --stale-parsersand Re-read with current importers would re-read every local source; prefer that command. Devin CLI, OpenCode, and MiMo Code histories kept as SQLite snapshots stay out of the automatic refresh, andhansard rebuild --from-rawrebuilds them. - The Codex cloud and Perplexity Computer browser importers changed. The
archive refresh parses earlier captures again from Hansard’s raw copies, so
Codex cloud tasks take their pull requests’ repository and Perplexity
Computer replies show. With the refresh off, run
hansard rebuild --from-raw --source-type codex-cloud-browser --since alland the same command with--source-type perplexity-computer-browser. - Two behaviors start on by default. Hansard pauses an agent that waits at its
prompt in a terminal 65 minutes after its last model request, once its
prompt cache is cold, and after 15 idle minutes under critical macOS memory
pressure. The conversation stays saved, and Resume in Running brings it
back where it ran. Turn off Pause idle agents in
Settings → Handoff & helpers → Pause and Resume or run
hansard config set agents.autoPause false. Resuming or messaging a Claude Code or Codex conversation in the viewer also stops its copy running in another terminal app first: a copy at its prompt stops at once, and the viewer asks before stopping one that works.hansard config set orchestrate.takeOverElsewhere falsekeeps both running. - Conversation labeling and automatic labeling are now off unless turned on,
so an installation whose configuration never set them stops labeling new
conversations. Turn them on in Settings → Classification or with
hansard classify enableandhansard classify auto enable. Labeling requests now leave out common secret patterns even when archive redaction is off. - The watcher and the shared MCP host move to the new code on their own.
Restart agent sessions to load the
ask_user,peer_file,code_history, andwork_contexttools. Every tool call now answers within 120 seconds, or within the longest waitask_session,ask_user, andwait_agentsallow plus 30 seconds, and ends with atool_timeouterror past it. - Commands that act for the person now refuse to run inside an agent session,
and most of them also need an interactive terminal. They include
hansard coordinator startandsend,hansard peers answer,hansard team accept,hansard remote host pair,hansard personal-sync devices --confirm,hansard device pair, and the review commands ofhansard memory dream.hansard coordinator startreads a brief only from--briefor--brief-file, andhansard coordinator sendits words only from--textor--text-file. In the desktop app, answers, approvals, and grants need the app’s own window, and a browser tab needs the code thathansard webprints. - Routing now keeps each conversation on the account that first served it in
every mode, and moves it once when that account reaches a limit or fails.
Run
hansard account proxy client configure claudeagain and restart Claude Code to give a routed setup from an earlier release the one-hour prompt cache for main conversations. - Update every member of a Team. Team people records now carry revisions and removed device keys, which members on 0.6.5 and earlier cannot read, so those members keep the details they last verified. A manager can turn on Compress shared files on a Team’s Rules page, which cannot be turned off later; a member on an earlier release then refuses the Team’s rules and stops syncing it until it updates, so turn it on once the Rules row counts every device as ready.
- To let the Slack bot take direct messages, update the Slack app from the
manifest that
hansard notify slack bot manifest --direct-messages onprints and reinstall it, since an app installed before has no direct message scopes.
- Work in Running by project. The rail holds one card per project with its
lead agent’s line, state, branch, folder, and ports, the selected project’s
conversations open as tabs, and a right pane holds up to three panes of
Changes, Files, terminals, and the ports agents listen on. The
Activity overview lists memory and CPU by checkout, Graph draws
running agents with their subagents, peers, tools, and MCP servers, and
Lanes shows each agent’s last 30 minutes of work.
0.6.5
macOS
A project’s page now starts, resumes, and messages visible agents in any harness and keeps the project’s roster, decisions, follow-ups, and watches, which agents also reach through four coordinator tools. Teams get their own view in the dock, with a join preview, rules each member accepts, and a folder scope for each member, and receive another member’s changes within seconds. Archive format 2 keeps conversation Markdown compressed, Personal Sync sends compressed file pieces and shows a new device’s newest conversations first, and the CLI, the viewer, and the desktop app start with less work.
Added
- Coordinate a project’s agents from its page (Start agent and the
Needs you, Agents, Follow-ups, and Activity groups), or with
hansard coordinator: start or resume a visible conversation in any harnesshansard psrecognizes, in cmux, Terminal, iTerm, Ghostty, or the Hansard app, deliver exact words to one agent or several, and keep the project’s roster, decisions and approvals, follow-ups, workstreams, and watches. Needs you lists agents waiting at a prompt and decisions to answer,hansard coordinator statusand therostertool show each agent’s tokens and estimated cost, and a usage pace watch notifies when spending crosses a rate. The watcher follows each agent’s state, held messages, commits, board posts, and due follow-ups without any model request, and agents use the same records through thestart_agents,deliver_message,roster, andledgertools. - Open Teams in the dock to see each Team as a mailbox, with its members,
rules and their signed history, newest conversations, and the
conversations it receives from this computer under From you, and to
create or join a Team on guided pages.
hansard team previewshows a Team’s rules, members, storage, and how many local conversations its rules match before joining. Each member keeps the rules they accepted: a change that shares less applies at once, and one that shares more waits until the member accepts it withhansard team rulesandhansard team accept.hansard team scopelimits a member’s uploads to chosen folders, andhansard team sharingrecords per-Team memory and agent-message choices, both off by default. - Set a profile picture in Settings or the Teams view; Team members see each other’s pictures and names from signed people records.
- Share a conversation from a redesigned Share dialog with a live preview
of each format, a Privacy section, and Save to Downloads beside
Save to…, which opens the macOS save panel in the desktop app. Shared
web pages carry harness and model marks, grouped tool steps, a prompt
index, and a light and dark switch, print as PDF with page margins, and
credit Hansard unless Hansard credit is turned off. Privacy counts
the values redaction hid and flags possible secrets the copy still holds,
and the copy names
paths from the conversation’s folder, with the home folder as
~. - Open the command palette with Command-K (Ctrl-K on Windows and Linux) to find conversations, open views, and run the page’s actions; a button above the session list opens it too, and an Appearance switch hides that button.
- Retire an account you no longer use with
hansard account retire <account>, which takes it out of routing and usage checks while keeping its usage history, and bring it back withhansard account restore <account>. Accounts lists retired accounts apart. - Run an interactive Slack bot from Settings → Slack → Interactive bot or
hansard notify slack bot run: a private Socket Mode app with Broadcaster, Tag-to-run, Digest, and Persistent watcher roles for each channel, where the owner approves every launch and continuation. Native Claude Code and Codex runs need macOS, and Socket Mode needs Node.js 22 or later. - Share a Codex installation’s app connectors with every agent through the plugin bridge when Codex is signed in with ChatGPT. Connector tools that are not read-only ask before each call.
- Let agents on a person’s paired devices list, message, and ask each other over remote access, with an Agent messages switch, on by default, when adding a device. Board posts, follows, and read marks reach a person’s other devices through Personal Sync.
- Import Claude Code workflow subagent transcripts as subagent runs of their parent conversation.
- See a first look at Stats for Claude Code and Codex while first-run setup imports them, read from their transcripts without archive writes. Setup now asks which coding agents get Hansard before the import, and, in the macOS app, asks for each macOS permission the chosen sources need, one at a time.
- Set up Personal Sync on its own page with a storage picker of provider logos that finds Google Drive, Dropbox, OneDrive, and iCloud Drive folders synced on this computer, and see the backup as one status card with a Needs attention list and its devices.
- Open links that lead outside the viewer in the default browser from the desktop app, instead of losing them or navigating the app’s window away.
Improved
- Start commands and the viewer with less work: each
hansardcommand loads only its own modules, compiled modules are kept in the archive’s compile cache, a restarted viewer serves its first page from a cached build, the Share dialog, Year in review, Accounts, and the project page load when first opened, and the Codex and Claude Code transport checks run off the startup path. A viewer served to other devices compresses its page and assets. - Child processes Hansard starts, such as import and index workers, provider servers, and the Cursor model listing, stop when the process that started them exits, even after a crash or a forced quit.
- A new device’s first receive of a backup with at least 256 conversations lists and opens the newest 64 first, before the rest arrives: on a synthetic backup of 2,000 conversations, the first one lists after about 8 seconds. A Team first receive downloads 16 files at once instead of 4, as Personal Sync already did, and a Team member’s first receive of 2,000 synthetic shared conversations takes about 42 seconds.
- Keep less on disk: conversation Markdown under archive format 2, view pages
at a higher compression level, parser checkpoints, and bundled model data
are stored compressed, and
hansard rebuild --compact-rawcompresses existing raw source copies. Personal Sync releases delivered upload payloads, keeps its outbox caches within bounds, and repacks sparse outbox database pages above 60% of its limit, at most daily. - Send Team changes sooner and with less data: a member receives within seconds of another member’s upload, a grown transcript of 5 MiB or more on S3-compatible storage travels as its appended bytes, uploads leave conversation views out, ready uploads run at most every 2 minutes while the archive keeps changing, the Teams view reads only each Team’s count and newest conversations, and Team status no longer lists and downloads the same objects on every request.
- Build the conversation list faster: a 6,400-session tree builds in about 140 ms instead of 610, filtered and paged lists build two to four times faster, a list reload after an import reuses every row it did not change, and a conversation opened again shows its kept copy at once while it refreshes.
- Spend fewer tokens on agent tools: tool schemas leave out default values and repeated descriptions, recall replies send one local timestamp per message, board replies name each board and thread once, and session hooks print shorter instructions.
- Redesign Running around agent rows with folder, harness, terminal, and a state dot, and an agent page with its actions, facts, and latest messages.
- Open Stats Changes, Tokens & spend, Tools & harnesses, and Projects on two headline figures with the rest behind a disclosure, and Overview on two lead figures and four supporting ones, with the rest behind Show more numbers, name hours on a 12-hour clock with AM and PM, count sessions without a recorded model under their harness as Model not recorded by that harness, such as Model not recorded by Devin, and draw Which models work most on their own? as a bubble plot of tool calls against output words per human turn. Year in review adds cards for the assistant, model, projects, and hours of a year, drawn in each harness’s own colors and mark.
- Price GLM, MiniMax, and Kimi models at their makers’ published rates, resolve Factory custom model ids to the model they name, and resolve Devin version names and Ask Devin modes as Cognition models.
- Show a harness’s asked questions as cards with every option and the chosen or typed answer, proposed plans as plan cards, API errors as error rows, a compaction as one divider, and web search results as links. Runs started in Hansard draw live Codex web searches, subagents, plans, and compactions as they happen, and an edit’s one-line success reply folds under its diff.
- Give the viewer one design language: borderless pill-shaped buttons, tabs, and segments with a dark pill marking the selection, neutral grays, reading text at 14px, provider and model colors drawn from each maker’s family hue, and a scene in each empty view. Dock names are off by default, Command-0 opens the tenth dock view, and Command-Shift-K opens Deep Recall.
- Name Graph communities from their distinctive title words, draw GitHub project nodes as the GitHub mark, and keep a large graph responsive during an import.
- Keep watcher imports going when a source fails: a source that reports no progress for 60 seconds is stopped and retried with backoff, other sources continue, absent sources are skipped in automatic imports, and the import status names the import that holds the lock.
- Stop with a clear fix on Node.js older than 20, and list the agent files
setup changes with
hansard init --dry-run.HANSARD_HARNESS_MODEL_MENUS=0keeps model menus from starting agent CLIs or reading agent stores.
- Count a Claude response’s tokens once at its final usage when its content blocks repeat a partial output count.
- Repair a received copy whose manifest misses a node by rolling it back to its newest complete snapshot and receiving again, and leave a received copy whose manifest cannot be read out of every list instead of failing the viewer.
- Retry a transcript page that fails behind a busy viewer after 3, 10, and 30 seconds, so a conversation no longer shows only part of its messages, and keep a reader’s text selection while a running conversation updates.
- Keep the viewer signed in when a request is refused for one feature, and clear an Accounts notice after a successful check.
- Report a router credential whose provider rejected its token as needing sign-in, recognize Claude Code and Codex login refusals, and report an agent signed in only when it holds a credential its CLI can use. Temporary provider homes no longer receive single-use refresh tokens.
- Name a Codex process’s own thread as its conversation in Running, find a long Codex turn’s start so it does not read as idle, and stop prompt words such as “server” from marking a conversation as a server.
- Withdraw a Personal Sync capture whose source is gone for good, so one lost source no longer fails every upload pass.
- Stop showing Dec 31, 1969 as the last check of an account list that was never captured.
All upgrade notes- This release is published for the macOS desktop app on Apple silicon. Mac apps from 0.6.1 through 0.6.4 update in place with Update and restart. The npm package and the Windows x64 preview skip this release and move to 0.7.0, whose notes list the earlier notes that apply to them.
- First update every copy of Hansard installed on a computer: the desktop app
and an npm installation alike. Archive format 2 now keeps each local
conversation’s Markdown gzip-compressed, which copies from 0.6.4 and earlier
cannot read, so while such a copy is installed or running, imports keep
writing archive format 1 for every file, as they do for copies that cannot
read format 2.
hansard config set archive.format 2names the copies that hold it back. Personal Sync’s release of delivered upload payloads and its deferred Markdown captures also wait until every installed copy supports them, and an earlier copy lists a large first receive that this release started without its conversations until the receive completes. - The watcher and the shared MCP host move to the new code on their own.
Restart agent sessions to load the coordinator tools (
start_agents,deliver_message,roster, andledger) and the shorter descriptions and replies of the recall, messaging, and board tools. - Then run
hansard rebuild --since allonce. It stores the Markdown of every local conversation as*.conversation.md.gz, at about 35 to 40 percent of its plain size; without it, conversations are compressed as they are next written or read. No reimport is required. Commands that name the file (hansard show --path,--open, and--json,hansard history, andhansard status) name a plain copy under~/.hansard/cache/conversation-markdown/, which cache cleanup removes after 30 days unused, so a script that readsconversation.mdfrom the archive directly should take its path from those commands. Exports, Team uploads, and Personal Sync uploads still carry plainconversation.mdfiles. - View pages, which archive format 2 keeps for long conversations, are now stored at gzip level 6 and stay readable by earlier releases. Pages written before are rewritten when their conversation’s view is next written.
- Compress the raw source copies already in the archive with the explicit
hansard rebuild --compact-rawpass, which never runs on its own. Preview it withhansard rebuild --compact-raw --dry-run --json, which lists the eligible files and their sizes without estimating the savings, then runhansard rebuild --compact-raw. It stores each eligible file losslessly in verified 1 MiB chunks without reparsing, needs free space for both forms while it converts, and leaves the replaced objects until archive maintenance collects them after their grace period. It waits while a copy of Hansard or a Personal Sync device cannot read chunked raw storage, and once a compacted conversation uploads, the backup requiresraw-chunk-storage-v1, which version 0.6.0 and later read, on every device for good. Later imports keep the storage modearchive.rawStoragesets. - The Claude Code and Claude SDK importers changed: workflow subagent
transcripts join their parent conversation as subagent runs, and a response
whose rows repeat a partial output count counts its final usage. The
watcher reads the conversations in its import window (30 days by default)
again with the new importers, and the first import of each active
conversation parses it in full once to write the new compressed parser
checkpoint. Refresh older history with
hansard import --sources claude,claude-sdk --since all. Each release starts a new parser baseline, sohansard rebuild --stale-parsersand Re-read with current importers would re-read every local source; prefer that command. Saved Stats payloads are computed again on their own. - Conversations received through a Team now stay out of Sessions, Stats, and
the graph. Turn on Show team conversations in Sessions, Stats, and Graph
in Settings → General, or run
hansard config set privacy.teamConversations shown, to keep them there. - Update every member of a Team. A member on this release lists the change marks other members’ uploads leave and receives as soon as one moves; uploads from a member on 0.6.4 or earlier leave no mark, so they arrive at the Team’s receive interval. Team uploads now leave out conversation views, which every receiver derives from the transcript.
- Personal Sync adds the Compressed file pieces format. Once every
confirmed device reads it, each piece of a file of 1 MiB or more, and a
text file of 64 KiB or more, uploads deflate-compressed when that saves
bytes. Update every device;
hansard personal-sync deviceslists which ones read it. Until then files upload as they are and nothing waits. - Personal Sync now releases the local copy of an upload payload once every
configured destination holds its exact files, after a 15-minute grace, and
leaves a conversation’s Markdown out of a capture until a destination needs
it. Archive maintenance does both; preview the cleanup with
hansard prune sync-outbox --dry-runand apply it at once withhansard prune sync-outbox --yes. After either has run, a copy from 0.6.4 or earlier on that computer cannot open the upload queue, so its Personal Sync captures and uploads stop until it updates. - Remote cleanup stays report-only by default. Its report no longer fails when
rebuilding this computer’s release log would bring the outbox database
within a fifth of its 512 MiB limit; it replays the backup’s history
instead. Preview removals with
hansard personal-sync prune-remoteand apply them with--yes. - On macOS, the updated app moves the watcher’s login startup from its launch agent to the Hansard app’s own login item on its own, so System Settings → General → Login Items lists it as Hansard, and macOS may show a notice that Hansard added a background item. The watcher and its workers then read Documents, Desktop, Downloads, cloud storage folders, other apps’ data, and external and network drives only where macOS allows the Hansard app, so macOS never asks for them in the background. History in a folder the app may not read waits until the Hansard app is allowed under System Settings → Privacy & Security.
- Plugin bridge tools set to ask before each call ask once more, since approvals recorded under the earlier risk rule expire.
- Coordinate a project’s agents from its page (Start agent and the
Needs you, Agents, Follow-ups, and Activity groups), or with
0.6.4
macOS
Windows
Teams receive every conversation another member shares. A Team upload no longer names the raw provider snapshot behind a conversation, which no Team carries, and a Team receive accepts conversations that an earlier upload sent with that link.
Improved
hansard app installpauses the watcher, the account router, and the remote host that run from the installed app while it replaces the app, and starts them again from the new app.--fromalso installs a builtHansard.appbundle.
- Receive a teammate’s conversations in full. A Team receive refused the metadata of any conversation that named a raw provider snapshot, such as Claude Code subagent runs and Codex conversations, so their transcripts arrived but the conversations never appeared. Each member’s next upload sends its Team conversations again without the link.
All upgrade notes- This release is published for the macOS desktop app on Apple silicon, the
Windows x64 preview, and the npm package. Mac apps from 0.6.1 through
0.6.3 update in place with Update and restart, and npm installations
upgrade with
npm install -g hansard@latest. npm 0.6.3 was not published, so npm installations move from 0.6.2 to this release, and the 0.6.3 notes below apply to them as well. - The Windows preview stays unsigned. Download the installer with Settings → About → Check now or from hansard.dev, quit Hansard, and run it. A Windows installation from 0.5.6 skips 0.6.0 through 0.6.3, so their upgrade notes below apply to it as well.
- Update every member of a Team. A member on 0.6.3 or earlier receives only the conversations of members on this release.
0.6.3
macOS
Windows
Hansard lists the agents running on the computer and stops idle ones from the new Running view or
hansard ps, and the Share dialog saves a conversation as a web page, PDF, Markdown, or an ATIF or EATIF trajectory. Archive format 2 becomes the default, growing conversations import only their new messages, and long transcripts keep only the rows near the viewport. Imports add Mistral Vibe’s Unified Harness, Gemini Apps HTML activity, Qwen Chat exports, and Cursor cloud run streams, and the browser extension’s source build captures many more chat and agent products.Added
-
See every agent harness running on the computer in the new Running view below Sessions, or with
hansard ps: each agent’s conversation, folder, terminal, state (working, waiting for an answer, or idle and for how long), and the CPU and memory of every process it started, with agents waiting for an answer first. Stop idle agents andhansard ps stop --idlestop agents idle for over an hour after a confirmation, and a stop leaves the conversation ready for the harness’s own resume command. An agent started in a Hansard terminal opens that live terminal, and the Terminal display andhansard ps --watchshow the same list as a full-screen monitor. -
Share a conversation from the redesigned Share dialog as a Web page, a PDF, Markdown, or an ATIF or EATIF trajectory, with tool calls, reasoning, and subagent runs chosen under Include and the text editable before saving.
hansard export <session-id> --format atif|eatif|html|markdownwrites the same files. ATIF is the Agent Trajectory Interchange Format (v1.7) that Harbor reads, and EATIF adds Hansard’s provenance, token accounting, file changes, and harness records underextra.eatif. The desktop app saves them to the Downloads folder and opens the print panel for Save PDF. -
Send a tool’s requests to one account until it reaches a limit, such as an account with a 5-hour limit and no weekly limit, with the Always try first routing rule. The tool’s Rules list in routing setup offers it with an account picker,
hansard account proxy always-first <client> <auth-index> on|offsets or removes it, and a Try first override still leads. -
Import Mistral Vibe’s Unified Harness history, the default session store of Vibe 2.26.0, with recorded message clocks, tool failures, cancelled approvals, inline images, and session attachments. Hansard verifies each snapshot’s hashes, chunks, and recovery journal before it accepts it.
-
Import Gemini Apps activity that Google Takeout exports as English HTML, Qwen Chat single-chat JSON downloads with
hansard import qwen-chat <chat-export.json> --username <account-label>, and Gemini Notebook / NotebookLM Takeout exports from Settings → Imports → Import export…. Session details of a notebook conversation open Gemini Notebook details with its dates, sharing state, and resources. -
Capture more products with an experimental source build of the browser extension: Claude Code on the web sessions, which join the conversations
hansard import claude-cloudreads; Codex cloud tasks; the workspace a Claude.ai chat links, archived as its own Claude Cowork cloud conversation with the files its recorded calls name; Claude Design project chats and files; Perplexity Computer tasks; and the visible chat of open Gemini Notebook / NotebookLM notebooks. -
Capture Grok chat, Meta AI, DeepSeek chat, Kimi chat, Qwen Chat, Mistral chat (Le Chat), Poe, and Z.ai chat history with the same experimental source build. Each collector reads saved history through the signed-in tab and keeps the raw responses for replay.
-
Rebuild Claude.ai browser captures and account exports, Claude Code cloud conversations, Perplexity captures, Devin cloud and Ask captures, and Mistral Vibe sessions from the archive’s raw copies with
hansard rebuild --from-raw, without reading the provider. A Claude.ai conversation shaped by captures and exports replays from the ordered inputs its import recorded; one archived before that record refuses replay and keeps its archive. -
Archive the one-to-one macOS Messages conversation with an Instinct number, set under Instinct number in Settings → Imports → Agents or with
hansard config set imports.instinctHandle. Capture stays off until a number is set, and only that conversation’s rows and attachments enter the archive, never the Messages database. Capture from the Messages store on a Mac has not been verified yet. -
Share a Team’s conversations by folder: a Folder rule, or
hansard team create --path-prefix ~/work/api, shares conversations started in that folder or inside it on each member’s computer. A folder in the home folder is saved as~/..., so the policy never names an account. -
Upload a Team’s changes as they’re ready: As changes are ready, or
--upload-interval ready, uploads about 15 seconds after the archive stops changing and at least every 2 minutes during a long conversation. Team uploads and update checks also offer every minute and every 5 minutes.
Improved
- The Team rule picker names each repository and folder by its folder name, with its path beneath.
- Keep only the transcript rows near the viewport in the page, so long conversations scroll and update with less work. Leaving a conversation of more than 4,000 transcript elements no longer stalls the next one while its rows are removed, and transcript find, the outline, and folding still reach every loaded message.
- Import growing conversations faster: the watcher derives only the messages
after the part an earlier import cached, builds the event record and view
pages from there, and leaves files that already hold the bytes it would
write untouched. An import that finds a conversation of 2,000 turns with 24
finished subagent transcripts unchanged writes 0.31 MB with 4 file syncs
instead of 22.3 MB and 30, and a parent whose subagent grows renders its
conversation.mdonce per interval instead of at every import. - Start imports as soon as a source is ready. The watcher runs its 15-minute
maintenance beside its ticks, runs import and capture workers at background
priority, holds back a changed file for three times its last import’s
duration, at most a minute, and applies changed source selections within 30
seconds without a restart.
hansard doctorreports the slowest steps of recent watcher ticks. - Refresh Stats after an import by folding again only the parts of the payload that changed conversations touch. On a synthetic archive of 2,345 conversations, a refresh after one to three changes takes 107 to 121 ms instead of 161 to 177 ms.
- Answer full-text searches by ranking matches before joining and excerpting only the page a search returns, and keep the search index near the size of a fresh build by merging its segments after refreshes replace documents. A refresh that derives every conversation again writes its plans in batches instead of holding them all in memory.
- Give agents leaner recall:
get_past_sessioncuts each tool message to 1,500 characters, markedtruncated, unless the read covers one message,search_memorieshits leave out empty fields and the repeated excerpt, andget_memoryopens a window centered on a hit’sline.board_contextover thousands of unread posts answers in milliseconds instead of seconds. - With search by meaning on, keep the embedding model loaded for ten minutes after an agent’s recall query and reuse the vectors of recent queries, unless the system is short of memory; on an Apple silicon Mac with screenshot search on, the loaded model holds about 440 MB.
- Stop agent hooks at 3.5 seconds, before the 5 seconds agents allow: a
session-start hook then returns the session token without the thread
digest, and a hook that fails returns an empty result and records the
failure.
hansard doctorlists hooks that failed in the last seven days under agent hooks. - Poll less: the project coordinator page answers its poll from its last view while nothing it read changed, instead of rebuilding the view every 2.5 seconds, and reads a run’s journal only when the run changed. The sidebar keeps conversation run states while nothing changes, and the composer’s file search runs in its own process, so it never waits behind an agent CLI inventory.
- Read Cursor cloud agents’ retained run streams, with thinking, tool calls, and their JSON arguments and results in a separate ordered block, and check enabled Cursor cloud agents every minute instead of every six hours.
- Download the public images and code charts that Perplexity captures reference, keeping original bytes, a labeled preview, or the reason a file is unavailable, within 100 MiB per file and 500 MiB per conversation. A host that asks for a pause holds back only its own downloads, and the receiver takes turns between providers so one provider’s queue does not hold up another’s.
- Keep the text that completed Write calls in Claude Cowork cloud tasks recorded, as ordered revisions of each file, and link Artifact calls to the latest completed write of their files.
- Upload only the appended bytes of a growing file of 1 MiB or more in Personal Sync once every confirmed device reads the Appended file segments format; the pieces fold into a full chunk at the next chunk boundary or once they number 32.
- Read newer harness stores: Hermes schemas through 31 and
HERMES_DATA_DIR_SUFFIX, OpenClaw schema 24, Kimi Code 2.1.1, pi 1.1.0, Qwen Code 0.25.0, MiMo Code 0.1.15, CodeBuddy 2.163.0, and GitHub Copilot request token counts and response times in VS Code. The historical Gemini CLI importer also finds the macOS Seatbelt store under~/.cache/.gemini/tmp. - The browser extension’s source build keeps ChatGPT generated files, reads
Muse (Personal Agent) side chats, goals, task runs, approvals, and the
original
MEMORY.md, downloads Devin attachments that redirect, slows ChatGPT collection after repeated rate limits, and captures the tab whose conversation changed, with current conversations taking turns with older history.
- Keep the screenshots and files a tool result returns in its tool card, including cards inside a folded tool group, show a message’s attachments once, and pair tool results with their calls again when earlier messages load across a page boundary.
- Pair a tool result that names its call only with that call, so a result whose call is missing no longer shows as the output of another call.
- Show Time unknown in conversation lists and search results for a conversation without a usable time, instead of a 1970 date.
- Include activity without a known date in all-time Stats totals and favorites, mark the Tokens tile when it counts estimates, show model spend as unavailable when a model’s usage has no prices, and caption all-time charts All time.
- Label Claude Code Remote Control sessions as Claude Code Remote Control and keep them in their Claude Code projects, or in Claude Code (uncategorized) without a recorded workspace. Claude Code cloud imports skip the sessions of other Anthropic products, such as Cowork tasks, workflow agents, and warm workers.
- Group recovered Cursor app conversations under Cursor 1, Cursor 2, or Cursor 3 without a reimport, and stop showing the app’s recovery warnings for Cursor CLI imports.
- Import the complete records of a Codex rollout before an unfinished last line, instead of failing the conversation while Codex is still writing it.
- Give a Claude Code or Claude SDK fork its own identity with a link to its parent, keep copied parent history out of search and Stats, keep child links when the session store moves, and keep a truncated source from replacing even a short archived conversation. Workflow scripts are kept as raw evidence and never run.
- Archive the bytes of Claude Science text document attachments, which earlier imports kept without content, and restore the tool usage and recorded results of Antigravity conversations that only its database holds.
- Restore ChatGPT project membership from captured project identities, identify ChatGPT dots replies from captured room membership, and keep failed Perplexity workflow results marked as failed.
- Follow every Jules list page, including empty continuation pages, and keep a truncated activity list from publishing a session. Grok Build imports read a stable snapshot of each session, so an incomplete or concurrent write leaves the archive as it was.
- Keep each repeated Mistral Vibe prompt as its own message, and stop a Muse (Personal Agent) export import without changing the archive when a workspace file is missing or does not match its declared size.
- Report Warp SQLite read failures and unreadable iFlow folders in import diagnostics, and retry a GitHub Copilot mutation log or a CodeBuddy session once its last record is complete.
- Give the watcher’s search index process a larger heap after a refresh runs out of memory, as imports do, instead of failing at the same size on every retry.
- Save the files the viewer offers to the Downloads folder in the desktop app, whose macOS web view canceled every download before.
- Keep outline jumps working while transcript rows are held outside the page and after a live refresh that changes no rows.
All upgrade notes- This release is published for the macOS desktop app on Apple silicon and
the npm package. Mac apps from 0.6.1 and 0.6.2 update in place with
Update and restart. The npm package skipped 0.6.2, so
npm install -g hansard@latestmoves an npm installation from 0.6.1 straight to this release, and the 0.6.2 upgrade notes below apply to it as well. The watcher and the shared MCP host move to the new code on their own, and the managed router after ten minutes without requests. Restart agent sessions to load the recall tools’ new descriptions and thelineoption ofget_memory. - The Windows x64 preview moves from 0.5.6 to this release and stays
unsigned. Download the installer with Settings → About → Check now or
from hansard.dev, quit Hansard, and run it. Then restart an enabled watcher
once with
hansard watcher restart. A Windows installation skips 0.6.0, 0.6.1, and 0.6.2, so their upgrade notes below apply to it as well. - Update every copy of Hansard installed on a computer. A watcher that runs
an earlier copy never rewrites a conversation this release archived or the
search index it built, so new messages in those conversations wait until
this release imports them;
hansard statusandhansard doctorname such a watcher. - No reimport is required. Run
hansard rebuild --since allonce to bring the stored events, tool usage, and outcome totals of earlier conversations to this release: a tool result that names its call now pairs only with that call, and Mistral Vibe Unified Harness edits count as edits in Stats. The watcher reads the conversations in its import window (30 days by default) again with the new importers, and refreshes the search documents of every conversation once because this release derives events differently; search keeps working meanwhile. Saved Stats payloads and the Cursor app generation of recovered conversations are computed again on their own. Each release starts a new parser baseline, sohansard rebuild --stale-parsersand Re-read with current importers would re-read every local source; prefer the commands below. - Importers for Claude Code, the Claude SDK, Claude Science, Codex, Grok
Build, the Gemini CLI, Mistral Vibe, Kimi Code, Roo Code, pi, Qwen Code,
Devin Desktop, Antigravity, Hermes, OpenClaw, MiMo Code, CodeBuddy, iFlow,
GitHub Copilot in VS Code, and Warp changed. Refresh history older than the
import window, including Claude Code fork identities and workflow scripts,
Codex sessions that stopped at an unfinished last record, Antigravity tool
usage, and Claude Science text attachments, with
hansard import --sources claude,claude-sdk,claude-science,codex-cli,codex-desktop,codex-sdk,grok-build,gemini-cli,vibe,kimi-code,roo-code,roo-code-cli,pi,qwen-code,devin-desktop,antigravity-cli,antigravity,antigravity-ide,hermes,openclaw,mimo-code,codebuddy,iflow,copilot-ide,warp --since all. - Refresh cloud histories with
hansard import claude-cloud --since all, which moves Remote Control sessions without a recorded workspace to Claude Code (uncategorized),hansard import claude-cowork-cloud --since all, which adds the files completed Write calls recorded,hansard import cursor-cloud --since all, which reads the run streams Cursor still retains, andhansard import jules --since all. - Browser captures and exports rebuild from the archive’s own raw copies
without a browser:
hansard rebuild --from-raw --provider chatgptrestores ChatGPT project membership and dots reply roles,hansard rebuild --from-raw --provider perplexityre-reads Perplexity captures,hansard rebuild --from-raw --source-type muse-agent-exportkeeps the full tool output of Muse (Personal Agent) side chats, andhansard rebuild --from-raw --source-type claude-web-exportkeeps the Design chats of Claude.ai exports. Choose Rescan all history in the browser extension to download the images of earlier Perplexity captures. - Archive format 2 is now the default, and a configuration in which an
earlier release saved format 1, its default then, reads as format 2. New
imports keep the events of a conversation above 1 MiB as an event record
and the view of a long conversation as compressed pages, and update a
growing transcript in place. Archived conversations keep their files until
hansard rebuild --format 2 --since allconverts them. Imports still write format 1 while a copy of Hansard from 0.6.0 or earlier is installed or running, since those copies read neither form, andhansard config set archive.format 2names the copies that hold it back.hansard config set archive.format 1keeps format 1. Before moving back to 0.6.0 or earlier, runhansard rebuild --format 1 --since allandhansard config set archive.format 1. - Personal Sync adds the Appended file segments format. Once every
confirmed device reads it, a file of 1 MiB or more that only grew uploads
just the bytes appended since its last upload. Update every device;
hansard personal-sync deviceslists which ones read it. Until then such a file uploads its whole last chunk again and nothing waits. - The watcher keeps a cache of what it derived from each conversation active
in the last six hours, up to 1 GiB in all under the Hansard cache folder,
and skips it when the volume would keep less than 2 GiB above the storage
floor.
hansard config set archive.derivationCache falseturns it off. - The managed router’s verified CLIProxyAPI release is now 7.3.20, which
halves the time the router adds before forwarding a 1 MB request. An
installed router keeps its version until Check for update in the
Local router menu in Accounts installs the new one, or until
hansard account proxy stop,hansard account proxy install, andhansard account proxy startdo. - The browser collectors this release adds come from a source build of the extension, which the Chrome Web Store release does not include yet. Update the extension and Hansard together, since this release’s receiver imports their captures.
- A Team that uses a Folder rule or As changes are ready needs every member on this release. Earlier copies refuse that Team’s policy and stop syncing it until they update. A Mac app that installed the first 0.6.3 build keeps it until the next release; download the app again from hansard.dev for these Team options.
-
0.6.2
macOS
Hansard searches conversations, memories, and screenshots by meaning with an optional local embedding model, imports Claude Science and Claude Cowork cloud history, and estimates what each subscription plan is worth. Personal Sync adds Google Drive, OneDrive, Dropbox, and storage copies, and this release fixes router sign-ins and the viewer’s checks of cross-site requests.
Added
- Search conversations and memories by meaning with a local embedding model.
hansard embeddings enable, or Search by meaning in Settings → Imports → Search, downloads EmbeddingGemma 2 and a llama.cpp runtime, about 330 MB, and embeds the archive on this computer. The sidebar search,hansard history,hansard memory search, and thesearch_past_sessionsandsearch_memoriestools then merge keyword and meaning matches, and--mode keywordor--mode meaningruns one alone. - Search the screenshots that conversations attach with
hansard embeddings enable --imagesor Screenshot search, which downloads about 880 MB in all and keeps about 540 MB; a match names the message that holds the image. - With embeddings on, open the closest conversations from
Related conversations in a conversation’s Actions menu or with
hansard embeddings similar, follow Similar subject and Similar screenshots links in Graph, and rank the subjects of the selected range in the Stats Projects tab. - List memories from different providers that cover the same subject with
hansard memory duplicates.hansard classify neighborsmeasures how often the labels of the most similar conversations match the model’s, andclassify.neighbors.enabledanswers a request from them only when every label agrees at least 90% of the time. - Import Claude Science conversations, tool results, attachments, child
tasks, and artifact revisions from its local and organization stores. The
watcher follows the stores,
hansard import --source claude-science --since allimports earlier history, andhansard memory backupkeeps Science memories. - Import tagged Claude Cowork cloud tasks across saved Claude Code sign-ins
with
hansard import claude-cowork-cloud --since all, and refresh them every six hours with Claude Cowork cloud in Settings → Imports → Cloud agents. - Capture ChatGPT Work Cloud tasks and dots ChatGPT rooms with an
experimental source build of the browser extension, shown as
ChatGPT Work Cloud and ChatGPT dots. ChatGPT browser captures
rebuild from the archive without a browser through
hansard rebuild --from-raw --provider chatgpt. - Estimate what each Claude and Codex plan is worth at API prices for the
recorded workload with
hansard account valueand the Subscription value section of the Stats Tokens & spend tab, from limit samples, archived transcripts, and the managed router’s new per-request records.hansard account value price <plan> <usd>prices a plan that lists none, such as a Team seat. - Compare models in the Stats Analysis tab, which places each model or reasoning effort by judged success, or by the Artificial Analysis Intelligence Index with a free API key, against its median cost, active time, and output per conversation, with a Pareto line.
- Change routing for a while with temporary router overrides: use only one
account, skip one, try one first, or bypass the router until a set time or
a usage reset. Temporary overrides under Routing in Accounts and
hansard account proxy overridestart and end them, and the title band shows one while it lasts. - Back up Personal Sync to Google Drive, OneDrive, or Dropbox with OAuth app
credentials the user provides, from Settings → Personal Sync or with
hansard personal-sync auth,create, andconnect. - Keep up to 16 additional Storage copies of a Personal Sync backup or a
Team from Settings,
hansard personal-sync storage, orhansard team storage. Scheduled sync updates each copy at most every 30 minutes, and a fresh installation can recover from a completed copy while the primary storage is unavailable. - Browse a remote host’s project files and working-tree changes, read-only, in the workspace pane of an attached remote conversation; both computers need this release.
- Capture Codex subagent assignments, follow-ups, and results through
capture_subagent_messages, which the managed Codex integration asks each agent to call in its current conversation. The viewer labels them agent-reported copy.
Improved
- Narrow recall:
hansard historyandsearch_past_sessionstake a message role (--role user), an end date on the local clock (--until), several hits per conversation (--hits-per-session), and marks for drafts pasted again (--group-revisions), andget_past_sessionreads a role-filtered window with long messages cut short. - With embeddings on,
board_contextalso matches threads by meaning,list_peersmarks same-project sessions whose recent prompts overlap the caller’s, and Codex and Claude Code session hooks list up to three earlier conversations close to a new session’s first prompt;orchestrate.messaging.relatedConversationsturns that list off. - Show Codex desktop sessions recorded by ChatGPT Work Local as
ChatGPT Work Local in history, search, and Stats, without a reimport.
Source filter options count the conversations they keep, every harness
works as a
--providerfilter, and web chat account folders, including Muse (Personal Agent), carry their service name. - Show up to 100,000 conversations in Graph through the new Nodes display setting, 500 by default instead of 400. Layouts of 10,000 conversations settle in about a quarter of the time, and indexing keeps moving during imports and retries a locked session index or a failed start.
- Install Hansard’s skills as
hansard-<name>, such as/hansard-recall, when another skill or command already uses the name, and install the rest when one agent’s configuration file cannot be read. Integrations change only Hansard’s own entries in agent configuration files and keep their comments, formatting, symlinks, and read-only files. - Give each account one name: a rename in Accounts or with
hansard account renamealso names its saved logins and router sign-ins. A saved Codex login and the router copy made from it show as one connection, and signing in to that login renews the router copy too. - Replace the transcript’s Latest messages bar with a floating button above the composer, and glide Sessions rows to their new places, with new conversations dropping in, when the list updates on its own.
- Answer Pricing, the Stats Changes tab, the one-shot Stats worker, and the composer’s model choices from the session store as committed, instead of rebuilding the session index for tens of seconds when a live import leaves one row behind its file. The Changes tab paints its last comparison while classification results or account names change, and navigation reads each received Personal Sync copy once, not once per conversation.
- Open the Codex composer’s plugin list from its installed plugins instead of reading the whole marketplace catalog, about 13 MB, on the viewer’s thread.
- Transfer Muse (Personal Agent) automatic exports from a source build of the browser extension in resumable chunks, up to the 256 MiB archive limit instead of 20 MiB; update the extension and Hansard together.
- In the browser extension’s next Chrome Web Store release, the Claude.ai history scan also lists archived and unified conversations, and its settings show each delayed retry with its reason and countdown while other conversations keep capturing.
- Refuse every state-changing viewer request that a browser sends from a page other than the viewer’s own, before any route runs. Browsers send the viewer’s session cookie from pages that other servers on 127.0.0.1 serve, and several routes, including agent turns and plugin approvals, checked only that cookie; command-line clients are unaffected.
- Make router sign-ins in Accounts reliable: overlapping sign-ins each finish with their own account, a sign-in to an account or workspace other than the selected one is removed instead of added as a duplicate, a canceled sign-in stays with its own request, a renewed router account leaves one copy holding its refresh token, and checks and renewals of one profile run one at a time.
- Show Codex sign-in ended on Accounts when OpenAI ends the Codex
sign-in on this device, with the last known account, the
codex loginstep, and Check again, and count the connection in the title band’s attention notice, instead of reporting usage as unavailable. - Run the account commands of Node-based agent CLIs with Hansard’s own
Node.js runtime, including the desktop app’s, so they work without a
separate
nodeon the PATH. - Keep the screenshots and image previews that completed MCP tool calls return in Codex conversations, and label a JPEG preview that Codex declares as PNG by its bytes.
- Classify a batch that a model refuses on usage-policy grounds one conversation at a time, so only the declined conversation waits a week instead of every automatic pass halting; Settings reports how many the model declined.
- Keep received Personal Sync conversations listed while a receive replaces their snapshot, including a conversation that one snapshot holds under several paths.
- Keep storage accounting’s tracked files valid after a macOS restart, which gives APFS volumes new device numbers, and retire the records of replaced archive files and moved received cache folders, which stopped storage reconciliation.
- Keep the desktop app’s guided setup going when one source’s project discovery fails, and stop searching for projects from the filesystem root when the app starts there.
- Accept Muse (Personal Agent) workspace exports whose file names contain colons; import a rejected export again with the same account label.
- Truncate a long value in key-value lists, such as a Graph project label, instead of squeezing its term to one letter per line.
All upgrade notes- This release is published for the macOS desktop app on Apple silicon; the
npm package skipped it, and the Windows preview stays on 0.5.6. Mac apps
from 0.6.1 update in place with Update and restart. The watcher
and the shared MCP host move to the new code on their own, and the managed
router after ten minutes without requests. Restart agent sessions to load
the recall tools’ new options, and resume Codex sessions to load
capture_subagent_messages. - No reimport and no
hansard rebuild --since allis required. The Codex, Claude.ai export and browser capture, ChatGPT browser capture, and Muse (Personal Agent) export importers changed. The watcher reads Codex conversations from its import window (30 days by default) again with the new importer; to add MCP tool screenshots to older Codex desktop conversations, runhansard rebuild --from-raw --source-type codex-desktop-history, which replays the archive’s own raw copies. Each release starts a new parser baseline, sohansard rebuild --stale-parsersand Re-read with current importers would re-read every local source; neither is needed. - Graph builds its index again once after the update, and its status line counts the conversations the first pass has read. Navigation and Stats facts, list payloads, and saved Stats payloads are computed again on their own.
- Stats spend estimates change with the refreshed prices. Claude Sonnet 5 is priced at $2 input and $10 output per million tokens throughout, without the September 1, 2026 increase, since Anthropic kept its introductory price as the standard price. The LiteLLM snapshot prices Claude Sonnet 5.5 cache reads at $0.10 per million tokens and adds Claude Haiku 5.5.
- Search conversations and memories by meaning with a local embedding model.
0.6.1
macOS
Long conversations open and jump to search results faster, Personal Sync uploads and receives less, and agents share each other’s skills, MCP servers, and plugins through Hansard.
Added
-
Share each agent’s MCP servers, plugins, and skills with every other agent that has Hansard’s MCP entry, through the shared MCP host. Tools that run code, control the computer, or change cloud services ask before each use until they are set to Allow, and Skills & plugins, each conversation’s Actions menu, and
hansard bridgechoose what is shared.HANSARD_PLUGIN_BRIDGE=0in the watcher’s environment turns sharing off. -
Report reclaimable and outdated storage with
hansard doctor storage, each finding with its size and the command that repairs it, and list the Personal Sync space by kind withhansard doctor storage --remote.hansard doctoralso compares session counts across metadata files, the session store, and the search index, names the command that repairs a difference, and summarizes slow and abandoned viewer requests. -
Search with quoted phrases,
-exclusions, and word prefixes, with identifiers such assqlite-vecmatched as one term, in the app,hansard history, and agent recall. Conversation results show highlighted excerpts and open at the matching message with find started on the match. -
Route Grok Build, Antigravity CLI, and Muse Code through Automatic account switching, set up account pools for new Cursor CLI and Devin CLI sessions under Accounts for new CLI sessions, show each account’s credit balance on its card, and choose for each Codex account whether routing spends its credits once included usage runs out (
hansard account credits). -
Save Amazon Bedrock, Vertex AI, and OpenRouter connections under Routing → Cloud model connections, grant each to the harnesses its provider supports, such as Claude Code, Codex, OpenCode, and Zed, and activate it separately in each harness; Turn off restores that harness’s earlier settings without changing the others, and
hansard account inferenceoffers the same controls. Provider API billing applies to a selected connection. -
Import Muse (Personal Agent) agent data exports with
hansard import muse-agent <zip-or-folder> --username <account-label>: message text, times, and channels form one account timeline, workspace files stay apart from the conversation text, and every exported file is kept in raw storage. -
Restore conversations that only the Personal Sync backup still holds with
hansard personal-sync restore-sessions, which lists them and with--yesrestores them byte for byte, never over an existing file;--sessiontakes one id or a comma-separated list, and a conversation removed on purpose is restored only when--sessionnames it. -
Run
hansardwith no command in a terminal to open an interactive mode that lists the matching subcommands, flags, and values while a command is typed, and complete the same with Tab in bash, zsh, and fish throughhansard completion.hansard initoffers the mode as thecli.interactivesetting, andhansard config set cli.interactive falseturns it off; scripts, pipes, and agent shells still get the command list. -
Choose the permission mode of a remote run in the composer, from the modes the host offers the selected agent.
-
Install the browser extension from the Chrome Web Store and pair it from the guided setup’s new Browser extension page or from Settings → Imports. Hansard keeps a capture from a provider it cannot import yet until an update can.
-
Opt in to a more compact archive with
hansard config set archive.format 2: events above 1 MiB are kept as references into the transcript, views of long conversations as compressed pages, and growing transcripts are extended in place.hansard rebuild --format 2 --since allconverts archived conversations and--format 1converts them back; imports write the format only while every copy of Hansard that may open the archive reads it.
Improved
-
Speed up first-run and full-history imports again after the 0.6.0 slowdown. A session’s own files share one storage reservation and one folder sync, files Hansard can rebuild from the transcript skip durability syncs, and each write takes far fewer SQLite transactions. A parent’s re-import parses only the subagent transcripts that changed, import children reuse compiled code on Node 22.8 and later, the import ledger appends only what changed, and transcript checkpoints save only new messages.
-
Capture shared provider stores less often: the Codex state database again only when a thread’s recorded facts change, and the Cursor global store through a recent capture that holds a conversation’s rows. Daily maintenance moves Cursor conversations to the newest capture once older captures take at least 256 MiB, so those can be removed. Files of 1 MiB or more are cloned instead of copied on volumes with copy-on-write clones (APFS, Btrfs, XFS).
-
Answer Stats and Pricing from one long-lived Stats process per Hansard home, which reads again only the conversations that changed and computes again only when an input or the day moves;
HANSARD_STATS_ENGINE=0computes each request in a one-shot process instead. The same archive now gives the same Stats in every locale and row order. -
Update the search index in place instead of rebuilding it: a pass replaces only the passages of changed conversations, and the index takes about half the space. A rebuild continues past a conversation whose stored events cannot be read and names it, and index passes back off while the storage limits refuse them.
-
Give agents more from recall: session search takes the
hansard historyfilters and reports when the index was built,get_past_sessionreads a window around a hit, memory search ranks project and global memories in one list with live paths and line ranges, and a failed tool call returns a stable error code and whether a retry can succeed. Peer messages, board posts, andask_sessionanswers arrive whole up to 2,000 characters,list_peerspages with each peer’s last reported status, prompts announce unread messages only when new ones arrive, and helper tools return previews with windowed full results. -
Treat skills as capabilities, not memories: memory backup no longer archives skill folders, memory search, the Memories view,
hansard memory list, and Stats leave out the ones earlier backups kept, and skill revisions store each unchanged file once. -
Keep long-running processes healthy: login startup, agent hooks, and MCP entries start Node.js through Homebrew’s stable link, so a Node.js upgrade no longer breaks them; process locks free themselves however their holder exits; the worker queue frees places that finished, exited, or silent jobs held; and the watcher log rotates at 4 MB.
-
Prepare for mixed versions on one computer: from this release on, a copy of Hansard that shares the archive with a newer one reads the stores and conversation files the newer version wrote without rewriting them, works out its own reading of a conversation in memory, and leaves the search index, memory catalog, and classification store to the newer version. Every copy records itself under
state/runtimes, so features that need every copy wait for each one. -
Use less disk: the reveal cache keeps only the originals of messages redaction changed (
hansard prune reveal-cacheconverts earlier copies at once), classification revisions are stored compressed (hansard classify compact --applyconverts earlier ones), and maintenance removes index files and tables only earlier versions used. -
Serve the viewer faster: a password-protected
hansard web --hostviewer compresses conversation and Stats responses for other devices, navigation validates archived files in the background only after they change or every ten minutes, large conversations are trimmed to the response limit without serializing them repeatedly, and each response reports its phases in aServer-Timingheader. -
Open long conversations faster. A search result or graph link opens the conversation directly at its message, newer messages load below it, and Latest messages returns to the end. Scrolling a long conversation in the desktop app stalls far less, since inserting and folding rows no longer makes the browser engine recheck the styles of the whole transcript. Conversation requests reuse configuration and validated page indexes, and a lookup reuses each received Personal Sync copy while it stays unchanged and opens one snapshot per conversation. An attachment resolves through that snapshot instead of failing while a newer copy arrives, and a conversation without stored pages opens with a full first window.
-
Answer the Behavior tab’s repeat requests with 304 again after local writes change its data, report an import status without an error when no import has run, keep Markdown previews of large sources within 1 MiB, and have
hansard doctorname a viewer older than the views it reads. -
Tidy the Routing summary and the account switching dialog. The managed router keeps its output in rotating logs, and
hansard account proxy statusreports a failed start and the latest routing refresh error with their times. Enabled router accounts appear in the same quota history as native sign-ins, including Antigravity’s separate model quotas, and Accounts, Usage history, and Routing share verified account identities. -
Open the composer’s working folder list and file search faster: new conversation targets list each folder once per session snapshot and check a missing folder once, and file and command queries check only the requested folder while git lists a checkout in the background.
-
Keep remote access available: the tunnel retries with growing waits and reports remote access offline after six failed attempts, and the watcher starts a remote host again after a restart, logout, or crash.
-
Stamp a device or person rename onto archived conversations in short batches so imports keep running, and show a rename still owed in Settings and
hansard status.hansard rebuild,hansard redact reapply,hansard delete, andhansard classify cleanupwait for a running import, andhansard redact reapplyrewrites each conversation in place. -
Lay out Graph’s connected clusters as packed islands, with unconnected conversations in an outer ring.
-
Keep the green ring on finished and done conversations in the sidebar after they are viewed, and narrow the open dock to smaller icons without names by turning off Dock names in Settings → Appearance.
-
Speed up Team sync: uploads read only the conversation files that changed since the last pass and list the Team’s storage once a day, receives check again only the files whose copy changed, each Team follows its policy’s sync intervals, and
hansard team statusand Settings list the devices the last receive found. -
Clean up with
hansard personal-sync repair, which retires received snapshots past retention, removes versions no kept snapshot uses, trims the chunk cache, and links upload payloads to the archive objects they copy;hansard personal-sync migrate, which repairs state that earlier versions, departed devices, and interrupted work left behind; andhansard personal-sync prune-remote, which removes backup objects only this computer uploaded, after a grace period and through a trash copy. Each previews first and applies with--yes. -
Upload less: a conversation uploads without its view file, and without events when it has messages; later uploads delete files that earlier versions uploaded and raw source objects that no conversation names; and
hansard personal-sync sweepqueues settled conversations that still carry such files. Payloads that copy archive objects are frozen as links, the outbox deletes obsolete records and compacts itself, and profile state is written only when it changes. -
Leave
conversation.mdout of uploads by default once the backup’s devices are confirmed and each reads such conversations, since receiving devices render it from the transcript;sync.sessionPayload fullkeeps it, andindexalso leaves out the transcript of a conversation every receiving device re-parses.archive.rawStorage chunkedkeeps snapshots of shared SQLite stores above 1 MiB as 1 MiB chunks, so a change uploads only the chunks it touched. -
Upload only the provider file versions a capture names, not the versions in between that the archive itself discards. Check the heads of streams with nothing to publish through one listing, reading a head again at most every 15 minutes, read the profile manifest again only after six hours, and reuse an unchanged large file without reading its chunk manifest.
-
Re-parse received Codex, Claude Code, Claude SDK, Grok Build, and Cursor app agent transcripts on each receiving device with its own importer, in a memory-limited background process.
hansard personal-sync interpret --allruns it for every received conversation, andhansard personal-sync paritychecks the same path against this computer’s own conversations. Receives skip view files that senders uploaded, and checks slow to every 15 seconds after a quiet minute and to once a minute after ten. -
Answer an idle receive check with one listing and no local writes, wait at least 30 seconds after a receive that applied changes before the next check, and speed up a first receive by skipping the chunk cache, hashing each file once, and downloading more files at a time.
-
Report which sync formats each device reads and confirm the backup’s devices:
hansard personal-sync deviceslists, for each format a later release may turn on, the devices and copies that read it, lack it, or have not reported, and names a device on a release Personal Sync no longer supports.hansard personal-sync devices --confirm, or Confirm under Device progress in Settings → Personal Sync, records the backup’s devices, and a format turns on only once every confirmed device reads it. -
Back up a running conversation at most every two minutes by default; Back up running conversations every in Settings → Personal Sync and
hansard personal-sync timing --checkpoint-secondsset another interval, andhansard personal-sync timingprints this computer’s upload timing. -
Warn in
hansard personal-sync statusandhansard doctor storageonce a stream’s file count or file list, or the upload outbox database, reaches half of the limit that stops the backup, and again from four fifths. -
Recover interrupted captures on their own during imports and archive maintenance, withdraw the capture of an archive write that fails, and keep an unfinished capture from stopping its source’s import.
hansard personal-sync recover-capture --allrecovers every stopped capture at once, and Settings → Personal Sync shows each device’s problem under its name with Recover, Use this version, or Update. -
Report conversations the backup still holds after they left the archive, and remove them from the backup only once
hansard personal-sync repair --only withdrawals --yesconfirms them. -
Report how much backup space cleanup could reclaim: an idle upload pass plans at most once a day which objects only this computer’s earlier uploads still hold, and
hansard personal-sync status,hansard doctor storage, and Settings → Personal Sync show the result.hansard personal-sync prune-remoteuses the same plan. -
Leave files that have not changed out of each upload revision once every installed copy of Hansard on the computer supports it;
revision.elisioninhansard personal-sync status --jsonreports the gate, and receiving devices need no new format.
- Keep imports from removing the last archived copy of a conversation as a
duplicate. A copy of a Cursor conversation from another of its stores, a
web export conversation imported under another account, or the Claude SDK
copy of a session started from Hansard goes only when conversations
archived on this computer, never copies received from another device, hold
every one of its prompts, replies, tool calls, tool output, thinking, and
attachments, with text compared exactly. A copy that holds anything no
other copy has stays, and
hansard doctorlists the conversations kept this way. An archived Claude Code copy whose file now holds only utility commands goes only when it was read from that same file, and a legacy Claude.ai memory session stays until its text is recorded as a memory. - Answer a malformed viewer request target, such as
//?x, with 400 instead of stopping the viewer. - Reconnect a Codex account whose credential its own Codex sign-in holds by renewing that sign-in, and open the account’s Codex sign-in if renewal fails, instead of creating a second routing sign-in. A revoked or expired routing credential now shows as such instead of an older usage failure, and a repeated sign-in callback no longer sends its code again.
- Play the transitions of the Import exports page, which loaded the shared components without their motion tokens.
- Keep Back, Forward, and the expand button in the macOS title band while the session browser is collapsed.
- Build Stats for archives whose all-time Stats payload passes 16 MB, which failed near 12,700 conversations.
- Show Stats at once from its saved counts after an agent changes a memory file or a setting or account changes, which left the page loading until a full count finished. The watcher now refreshes the saved counts at least every ten minutes while agents write without pause, and the Stats process answers again seconds after it finds a conversation written moments before its index, instead of passing requests to a slower one-off process for a minute or more.
- Keep large Personal Sync uploads from failing with
chunk source changedwhen another link to the file is added or removed while it uploads. - Finish indexing a Personal Sync first receive of several hundred conversations, which wrote them in one transaction past the session store’s growth limit and rolled back on every pass. Conversations received earlier stay visible while a large receive indexes in batches, an interrupted receive resumes without duplicates, and an earlier copy of Hansard that reads the archive meanwhile rebuilds the list in one step.
- Rebuild the session list of an archive with thousands of conversations, which rolled back past the store’s transaction growth limit.
- Drop a queued Personal Sync capture of a conversation that left the archive, which failed every upload pass.
- Keep the watcher running when the disk fills: a log line it cannot write is dropped instead of stopping the watcher and its MCP host until the next login.
- Relabel existing conversations when classification runs with a prompt override or a narrowed classifier list, and retry the results that failed.
- Keep account dialogs closable while a sign-in request runs, and close a canceled sign-in at once; a late answer to a dismissed sign-in no longer closes or marks an error in the dialog opened after it.
- Keep the Accounts page’s scroll position, focus, and open menus while its account list or routing queue updates.
- List Hansard’s recall-memories skill once in Codex and Grok Build, which also read the shared skills folder that held OpenCode’s copy.
- Count a peer message delivered whole into a session’s prompt as read, so coordination hooks stop announcing it as unread.
- End the processes that interactive usage probes start, remove the temporary homes they leave, and keep the Grok CLI from updating itself into a probe’s home.
All upgrade notes- This release is published for the macOS desktop app and the npm package;
the Windows preview stays on its earlier release. Mac apps from 0.6.0
update in place with Update and restart, and
npm install -g hansard@latestupdates an npm installation. Then restart an enabled watcher once withhansard watcher restartor from Importer version in Settings → Imports. From this release on, the watcher restarts itself when the code it loaded changes on disk, when a Node.js upgrade removes the binary it runs from, and once a day, each at a moment when no import, sync, or other pass runs;watcher.restart.onCodeChangeandwatcher.restart.dailyturn either off. The MCP host and the remote host move to new code once no call or run is active, the watcher restarts the managed router on new code after ten minutes without requests, andhansard statusandhansard doctorlist processes that still run earlier code. - Update every copy of Hansard installed on a computer. Until an earlier copy
is updated, Importer version in Settings → Imports moves a watcher
that runs from it, and its login startup, to the app’s copy. While an
earlier copy imports or runs the watcher, it rebuilds the search index in
its own format, and Personal Sync raw object deletes, archive format 2, and
chunked raw storage wait until every installed copy of Hansard supports
them. A copy from 0.6.0 that still runs imports deletes archived duplicates
without this release’s check, so update it or move its watcher to the
app’s copy before running
hansard personal-sync restore-sessions. - No reimport and no
hansard rebuild --since allis required: no importer changed what it archives, and archived events, views, Markdown, tool usage, and titles stay current. Each release starts a new parser baseline, sohansard rebuild --stale-parsersand Re-read with current importers would re-read every local source; neither is needed for this release. Saved Stats payloads are computed again on their own. - The first watcher pass after the update rebuilds the search index once, in
a layout about half its size. The rebuild needs free space above the
storage floor for twice the current index, and at least 512 MiB; search
answers from the previous index until it finishes. Agent sessions and other
processes that started before the update then match exact titles only until
they restart;
hansard watcher restartrestarts the shared MCP host that agent sessions search through. - Personal Sync uploads leave out view files and the events of conversations
with messages, which receiving devices build from the transcript, and each
conversation’s next upload deletes those files from the backup.
hansard personal-sync sweepqueues settled conversations whose backup still lists them, 200 a run. Upload passes also delete raw source objects that no backed-up conversation names. Receivers on 0.6.0 and earlier apply these deletes and keep working, but their knowledge graph loses the links it read from sent events until they update. Each receiving device now re-parses received conversations with its own importer;hansard config set sync.receive.interpret falseturns that off. - Confirm the backup’s devices under Device progress in
Settings → Personal Sync or with
hansard personal-sync devices --confirm. Once every confirmed device reads it, uploads leaveconversation.mdout by default, since receiving devices render it from the transcript;hansard config set sync.sessionPayload fullkeeps uploading it. Without a confirmation, an upload pass confirms the devices on its own 14 days after uploads began waiting on the list, if every device seen in the last 30 days reports the formats it reads. - Personal Sync backs up a running conversation at most every two minutes by
default, instead of every 30 seconds. Back up running conversations
every in Settings → Personal Sync and
hansard personal-sync timing --checkpoint-secondschange the interval, and a device that stored the earlier default moves to the new one. - Personal Sync no longer supports 0.5.5 and earlier, which cannot read what
later releases upload.
hansard personal-sync devices,hansard status,hansard doctor, and Settings → Personal Sync name such a device as unsupported; update it. - Check reclaimable storage with
hansard doctor storage, which names the command for each finding.hansard personal-sync repairtrims received copies and the upload outbox,hansard personal-sync migraterepairs state that earlier versions, departed devices, and interrupted work left behind, andhansard prune reveal-cachereplaces the whole unredacted copies earlier versions kept; the first two preview until given--yes, and the prune previews with--dry-run.hansard personal-sync prune-remoteremoves backup objects that only this computer uploaded; apply it only after every device’s first upload has finished. Upload passes now plan that cleanup at most once a day and report the reclaimable space without removing anything. The next release turns removal on by default;hansard config set sync.remotePrune.enabled falsekeeps it report-only, andtrueturns it on now. hansard doctor storagereports conversations that left the archive, throughhansard deleteor because an import replaced their copy, while the Personal Sync backup still holds them, possibly as their only copy. They leave the backup only oncehansard personal-sync repair --only withdrawals --yesconfirms them, so check the list its preview prints first; a conversationhansard deleteremoves is confirmed at once and leaves the backup a day later.hansard personal-sync restore-sessionslists the conversations the backup holds and the archive no longer does, and with--yesrestores them.- Team uploads remove structural local-path fields and wholly absolute path values from metadata, and transcript metadata and event strings shorten only paths the same record also holds as separate values; a path found only inside a title, summary, or other text can remain, and conversation Markdown clears its working-directory fields but keeps its title and body. The next upload replaces bytes shared earlier when this changes them, and a member on an earlier release keeps sharing paths until it updates. The watcher runs each Team’s passes at its signed policy’s intervals. Team sync removes no files from the Team’s storage.
- Routing and account pools skip a Codex account whose included usage is
used up while it would bill its credits;
hansard account credits <account> onor the account’s Connections dialog allows one. The managed router installs only the CLIProxyAPI release Hansard lists as verified. After a restart, logout, or crash, the watcher starts a remote host again unlesshansard remote host stopor Stop work was the last command. - The watcher writes a live conversation’s
conversation.mdat most every two minutes and again once it settles, and commands that read the file bring it current first; a program that opens it by path reads a complete rendering at most two minutes behind. The reveal cache keeps only the originals of messages redaction changed, and each conversation’s next import replaces the whole copy earlier versions kept, after which an earlier copy of Hansard finds no unredacted text for it. Memory search, the Memories view, and Stats leave out skill folders that earlier backups archived. - The watcher’s integration sync, also run by
hansard integrations sync, gives Codex entries a 31-minute tool timeout forask_sessionandwait_agentsand moves OpenCode’s recall-memories skill into OpenCode’s own skills folder. Restart agent sessions to load the new MCP server instructions. - The browser extension’s next Chrome Web Store release sends nothing to a receiver that cannot prove it holds the browser’s pairing, which this release can, and asks for a Hansard update instead; a browser paired by an earlier release may be asked to pair again.
- The shared MCP host shares each agent’s MCP servers, plugins, and skills
with the other agents by default. Restart agent sessions to see the shared
tools, or set
HANSARD_PLUGIN_BRIDGE=0in the watcher’s environment to turn sharing off.
-
0.6.0
macOS
Added
- Open a workspace pane beside any conversation with its changes, files, and interactive terminals. It previews PDFs, tables, Markdown, and documents, highlights code, draws side-by-side diffs, opens files in their default app, lists the commands the agent ran, and follows file paths and links in the conversation. Resume in terminal types a conversation’s resume command at a workspace terminal’s prompt.
- Import Claude Code cloud and Codex cloud sessions as their own providers. Codex cloud tasks import full transcripts through the saved ChatGPT sign-in, and both imports read every registered profile. Settings → Imports → Cloud agents turns on cloud histories for Claude Code, Codex, the GitHub Copilot coding agent, Amp, Augment, Devin, Jules, Cursor, Factory, OpenHands, Warp, Kilo Code, and Qoder, which the watcher then refreshes every six hours.
- Support the Cursor CLI as a current agent: import its chat and ACP stores,
launch and resume it over
cursor-agent acp, sign in its accounts with isolated profiles, read its usage pools and on-demand spending, and let helpers start it in Agent, Plan, or Ask mode. Cursor sessions are listed as Cursor CLI, Cursor cloud, and Cursor 1, 2, and 3. - Import gptme histories, Posit Assistant conversation trees, NotebookLM Takeout exports, and Ask Devin conversations with their child sessions.
- Open the alternate branches of Claude.ai and ChatGPT conversations from their archived captures, and keep each chat’s model, settings, project, and provider state in one conversation record.
- Show everything a provider recorded about a session, including Devin status, billing, and pull requests, in a record dialog opened from Session details.
- Forecast when each plan limit runs out, list the sessions spending it, and alert once before a provider runs out.
- Answer Claude questions and plan reviews, reply to MCP elicitations, and pick each harness’s own approval choices from one request card in the conversation. Grok Build, Devin, OpenCode, and the Cursor CLI now run over the Agent Client Protocol, so their prompts, questions, and stops reach Hansard.
- Add model-aware Codex speed controls to the composer and settings, carried through launches and resumed runs.
- Add a short first-visit guide to each main page, project coordinators, Imports, and Personal Sync, with rows that sign in or turn a feature on in place, replayable from Settings → About → Guides.
- Announce each release once with New in Hansard, a dialog with the
release’s highlights and the setup steps that apply to this computer, a note
after the first interactive command, and
hansard whats-new. - Add
hansard why <file>:<line>to find the archived agent edit that wrote a line. - Render TeX formulas in conversation text as MathML.
Improved
- Resume turns that quitting, updating, or a crash cut short when Hansard opens again (Resume after a restart turns it off). The desktop quit review now also covers macOS menu, Dock, scripted, and logout quits, orchestrator runs, and terminals running a command; an idle shell prompt does not block quitting. Agent runs and the desktop app start without the launching terminal’s session variables.
- Fold steered messages into a running Claude turn, offer Stop in place of Send while a turn runs (Escape stops it), keep runs streaming across view switches, and paste rich text into composers as Markdown so links keep their addresses.
- Rework remote control (alpha) around Settings → Remote access: one switch allows access, devices list the agents installed and signed in on each, direct connections run over Tailscale by default, a self-hosted relay adds one-use links and a browser conversation page, and a grant reaches every project folder or one portfolio.
- Serve every agent session’s recall tools from one shared MCP host run by the watcher, and keep prompt hooks and MCP startup independent of peer discovery.
- Keep older Hansard copies from rewriting conversations a newer importer archived, and alert when another device’s newer Hansard holds back conversations until this copy updates.
- Offer every installed harness unless it is turned off, list installed
harnesses first in Settings, hide unavailable harnesses from conversation
controls, and import agents installed after setup, as
hansard initdoes. - Run imports, archive maintenance, and diagnostics from Settings, which also shows older imports, the watcher’s copy, and the command line tool. Setup offers detected recoveries on its import page and explains incomplete storage estimates.
- Pair the browser extension and manage its captures from Settings → Imports. The extension gains Hansard icons, a toolbar badge for failures, per-check provider status, and automatic retries of failed capture imports.
- Gather resume, session details, and other conversation actions into one Actions menu led by Resume in terminal, lead the header with the most-used model, show cache reads and writes in the Tokens row, and show a time recorded without a time zone as its recorded clock.
- Show web chat account portfolios on the Projects page and in scoped Stats, and count Claude.ai conversations under their selected model.
- Surface relevant board threads and unread reminders to agents, show board participation, and bring new posts and messages in place.
- Animate the viewer through one shared motion library: panels, sheets, lists, the workspace pane, composer attachments and requests, boards, and charts.
- Recover legacy Antigravity conversations through the running app’s language server, recover Cursor models and token counts from the account’s usage history, and read Devin cloud models from the recorded Devin version.
- Preserve more of what each provider records, with model settings kept apart from reply evidence, for Claude desktop and cloud, Copilot, OpenCode, Factory, Grok Build, Goose, pi, Qwen, MiMo, Qoder, Trae, and Antigravity.
- Bound archive, index, sync, and cache writes against configurable storage
allowances (
storage.limits), keep incomplete raw snapshots from blocking cleanup, explain cleanup blockers in plain language, and recover the archive after an interrupted update withhansard update --list-recoveriesandhansard update --recover. - Offer in-app desktop updates only on Developer ID signed installs and check the signature before downloading.
- Speed up large archives: skip idle heartbeat imports of unchanged sources, re-archive only the Claude subagent runs whose files changed, refresh a dirty search index at most once a minute, load boards from a saved index, write only changed session rows, and build transcript rows within a frame budget.
- Add Sol 6.1, Sonnet 5.5, and Cursor Composer model identities and pricing.
- Harden the Windows desktop app and CLI: resume conversations in workspace
terminals with
cmd.exesyntax, copy and paste with Windows Terminal keys, minimize to the taskbar while work keeps running, open archived Codex threads in Codex Desktop, find the Cursor CLI, the Codex CLI bundled with Codex Desktop, whisper.cpp, and the running Antigravity language server, and read Claude desktop data from the Microsoft Store package. - Replace in-use binaries when upgrading the Windows app, flush and replace read-only archive files, identify tracked storage by volume and file ID, and move the shared MCP host to another port when Windows refuses the recorded one.
- Match native Claude peer deliveries, messages received during a turn, and archived message reads inside Codex tool batches.
- Keep title band alerts visible after their entrance animation, and let an alert’s hover fill reach both edges.
- Stop repeating session briefings on follow-up turns and keep conversation identity stable across follow-up runs.
- Keep Warp imports and memory backup running past inaccessible or macOS-withheld folders.
All upgrade notes- Update the desktop app and every CLI installation on a computer together,
then restart an enabled watcher with
hansard watcher restart. Mac apps from 0.5.6 update in place with Update and restart. 0.6.0 upgrades the local sync outbox the first time it writes, after which a 0.5.6 copy on the same computer cannot write to it. Older copies also no longer rewrite conversations a newer importer archived;hansard statusandhansard doctorreport held conversations and a watcher running a different copy. - Importers for Claude Code, Claude desktop and Cowork, Cursor, Devin
Desktop, Copilot, Factory, Grok Build, OpenCode, MiMo Code, Antigravity,
Goose, pi, Qwen Code, Trae, and Qoder changed. Refresh earlier sessions and
import the new Cursor CLI, Posit Assistant, and gptme sources with
hansard import --sources claude,claude-code-desktop,claude-cowork,cursor-cli,cursor,devin-desktop,copilot-cli,copilot-ide,factory,grok-build,opencode,mimo-code,antigravity-cli,antigravity,antigravity-ide,goose,pi,qwen-code,trae,qoder,posit-assistant,gptme --since all, then recover Cursor models and token counts withhansard import cursor --recover. - Refresh cloud histories with
hansard import claude-cloud --since all,hansard import codex-cloud --since all,hansard import devin-cloud --since all, andhansard import factory-cloud --since all. The Claude and Codex imports move earlier hosted copies under the new Claude Code cloud and Codex cloud providers; Remote Control sessions stay with Claude Code. Import ChatGPT and Claude.ai exports again to add conversation records and alternate branches, and choose Rescan all history in the rebuilt browser extension for ChatGPT, Claude.ai, Perplexity, and Devin captures. - No
hansard rebuild --since allis required: viewer payloads, model and session-kind rules, Stats, and navigation refresh on read or in the background. Each release starts a new parser baseline, sohansard rebuild --stale-parsersand Re-read with current importers re-read every local source; prefer the targeted commands above. - Recall tools now run through one MCP host the watcher starts. Existing
agent sessions keep working and switch when restarted; run
hansard integrations syncso Codex connects to the host, then restart agents. - Direct remote connections now listen on the Tailscale address only.
Restart a running host after upgrading, and run
hansard remote host direct on --network anyon a host that other devices reach over the local network. New grants and pairing codes reach every project unless--projectsor--portfoliolimits them; existing grants keep their project lists.HANSARD_REMOTE_CONTROL=0turns remote control off. - The integration target
cursor-ideis nowcursor, and the earlier names still work. Resume, fork, and copy actions moved into the conversation’s Actions menu, and the remote control dock action into Settings → Remote access. Newly installed harnesses are offered unless turned off throughorchestrate.disabledHandoffProvidersandorchestrate.delegation.disabledProviders. Grok Build and OpenCode run over the Agent Client Protocol;hansard config set orchestrate.acpTransport execrestores the earlier launcher. Writes keep 2 GiB free by default (storage.limits.minFreeBytes).sqlite3andunzipare no longer required. - The Windows desktop app remains an unsigned preview with manual installer
updates. Before moving an older archive to Windows through Personal Sync,
run
hansard update --dry-run --since all, thenhansard update --since all.
0.5.6
macOS
Windows
Added
- Update the desktop app in place. It checks for releases on launch and hourly, and Update and restart downloads and verifies the signed update, then relaunches with rollback. Active runs and imports defer the restart.
- Control agents on another computer (alpha). A macOS or Linux host accepts
work over SSH or a direct encrypted connection paired with a one-time code;
the composer’s device menu selects the host, provider, and project folder,
and sends, live output, approvals, and interruptions stay in the
conversation.
hansard remoteand the Remote control dock action manage hosts, access grants, and pairing. - Add Boards for persistent discussions: General and topic boards, a
board for each project, portfolio, and detected agent harness, threads with
Follow and in-place search, direct messages with live sessions, and
hansard boards. Agents post project findings on their project board. - Replace the desktop first-run page with a guided multi-page setup that selects history sources, offers to raise agent retention limits, reviews disk space, imports with live progress, and configures agents, the account router, classification, Personal Sync, and Slack. Deferred steps stay at the top of Settings.
- Review running agent work, Deep Recall, and classification before the desktop app quits, with Keep Running in Background, Cancel, or Stop Work and Quit.
- Move primary navigation into an app dock that collapses into a floating island, and add Back and Forward (Command-[ and Command-]), Command-1 through 9 for dock views, and keycap shortcut hints revealed while Command is held.
- Add sidebar themes under Appearance: solid, gradient, and custom themes that color the dock, title band, alerts bar, and optionally the session list, plus animated pixel animals for empty states.
- Attach dropped or pasted images in the conversation composer, including files dropped in the desktop app, and insert dropped file paths at the caret.
- Mark conversations with activity since they were last viewed, and show each conversation’s run state and unread activity in one dot beside its title.
- Add range-scoped Stats headline tiles with marked favorites, leader timelines for models, projects, harnesses, and kinds of work, and Session rhythm charts by local hour, hour of the week, and when each day starts and ends.
- Count mid-turn interruptions for Claude Code, Codex, Grok Build, and OpenCode and compare models by them in Work & autonomy, and show time savings, effective horizons, expertise, and decision ownership from classified work.
- Import Devin Desktop ACP message stores with full tool inputs, capture Devin cloud sessions from signed-in tabs in the browser extension, and keep one archive session per Devin cloud or CLI session across Desktop, web, API, and CLI copies. Devin cloud is listed as its own harness.
- Install managed integrations for Factory Droid, Augment CLI, Cursor IDE,
Cline CLI, GitHub Copilot CLI, Amp, Kimi Code, Kiro, Zed, and Claude
Desktop, and report status and identity from conversations opened directly
in supported agents through
hansard session. - Let Source and Model filters hold several values, and order model and harness filters by capability, use, name, or most recent use.
Improved
- Rebuild Stats charts as ranked tables and bars in place of scatter plots and donuts, give every provider, model, project, and device one color per range, label series with brand and project marks, and open every folded tail in place instead of capping it silently.
- Fold the Behavior page into a Stats Changes tab that compares the
selected range with the equally long period before it;
hansard behavior --startsets the first day of a comparison. - Rework Reliability around ranked error rates, a project error table, API error kinds, and the spend of sessions that hit API errors; rank models by spend in Tokens & spend; and rank Work & autonomy models on judged success with cost per success.
- Label each conversation with one conversation request and one task request, rate ten difficulty facets from 1 to 10, default to the light claude-sonnet-5 and gpt-6-luna models, limit automatic labeling to the last 30 days, run classifier CLIs without tools, MCP servers, project files, or hooks, and add a Jev backend.
- List Accounts sections and accounts in the session browser, rename accounts in place, gather connections that need attention at the top of the page, chart each usage limit’s history, and read Antigravity, Grok Build, Cursor, and Codex workspace plan details.
- Redesign Routing as a summary per tool with one ordered account list, read rule usage about every two minutes with rate-limit backoff, include router quotas in Capture, fold router sign-in into Add account, restart the managed router from the watcher after a restart or crash, and warn that Claude Code routing turns off claude.ai connectors.
- Run every project coordinator from one Settings → Coordinator configuration, start it with the project’s first message, show its whole conversation with helpers as threads, persist its tasks and rules, show its plans, and require verification before a task completes. Orchestrator runs no longer stop at a concurrency cap or elapsed-time deadline.
- Separate passive peer notices from work requests, bound inbox previews
with paged
hansard peers read, list live Codex threads beside Claude Code sessions, and expose coordinator identity through peer and session tools. - Receive Personal Sync changes incrementally through verified inventory
checkpoints, show progress per device, receive device profiles
concurrently, add an upload-only
hansard personal-sync publisherpreset, and keep failed captures from blocking other sync work. - Show ten conversations per folder, hang subagent fan-outs off one guide line, list each project folder with the device that holds it, fold Claude Code scratch folders into the uncategorized folder, and keep rows in place as activity changes.
- Show harness and GitHub marks in place of typed provider and repository names, and use shared dropdowns, styled tooltips, date pickers, and show-more disclosures across the viewer, with focus rings only after keyboard navigation.
- Cluster the Graph by connection communities, connect conversations whose agents read the same files or fetched the same pages, keep placed nodes still through refreshes, and add Reset beside Fit.
- Draw recovered memory snapshots as numbered file lines linked to their source conversation, read memories in the centered conversation column, and keep scratch checkouts and skill folders out of the memory archive.
- Describe what each skill, plugin, and server is, who uses it, and where it comes from, including skills Warp and Amp load from other agents’ folders, and name Codex plugins by their manifest display names.
- Separate generated launch instructions from user prompts, name delegated conversations from their assignments, accept large Codex tool results, keep Perplexity research steps, cited sources, and media, and classify Perplexity chats as web chats. The watcher keeps a browser receiver running for paired browsers.
- Pin the conversation composer to the reading column, float its status and the full-access confirmation above it, and stop repeating archived turns in the live run overlay.
- Share one resource budget across background workers, bound regenerable caches, serve memory and Stats reads from indexed catalogs, load optional viewer features when their views open, checkpoint SQLite stores without blocking readers, and release idle Codex servers after fifteen minutes.
- Add a Windows x64 installer with the Node.js runtime and native dependencies bundled. Installation uses the current user’s account without elevation.
- Add a Windows download to the website and manual installer updates through Settings → About → Check now. This Windows preview is unsigned: Windows may show an unrecognized-app warning or block installation under Smart App Control or an organization’s security policy.
- Align navigation, branding, sidebar controls, alerts, and themed window controls in the Windows title bar while preserving resizing and Snap Layouts.
- Keep background setup, import, and supervisor subprocesses from opening transient console windows on Windows.
The published macOS 0.5.6 installer and its signed native update remain available. Windows updates use a downloaded installer; finish active work and quit Hansard before running it. Automatic CLI installation continues to require a trusted signed release.
- Render conversations that hold a redacted secret instead of failing to display them.
All upgrade notes- Quit the desktop app before replacing it with the 0.5.6 installer, then
reopen it. CLI installations should restart an enabled watcher with
hansard watcher restart, then restart the viewer and any agent MCP connections. - 0.5.6 is the first desktop release with in-app updates. Later macOS releases
install from the app with Update and restart; Mac apps from earlier
releases need this one manual upgrade.
hansard app installnow downloads the signed app from hansard.dev and verifies it before installing. - Importers for Codex, Claude Code, Claude SDK, Devin, Grok Build, OpenCode,
and Antigravity changed. Refresh earlier sessions with
hansard import --sources codex-cli,codex-desktop,codex-sdk,claude,claude-sdk,devin-cli,devin-desktop,grok-build,opencode,antigravity-cli,antigravity-ide,antigravity --since all. - Conversations archived before this release gain the revised tool usage
fields, including interruptions and the prompt times Stats places by local
hour, and revised inferred titles after one
hansard rebuild --since all. No reimport is required. - npm installations upgrading from 0.5.4 should also follow the 0.5.5 upgrade notes below.
- Managed agent integrations now include memory search, status reporting,
session identity, peer messages, boards, and helper launches by default.
The watcher applies them on its next integration pass;
hansard integrations syncapplies them immediately. Reload agents afterward.--no-memoriesand explicit off settings are kept. - Codex routing keeps Codex’s own ChatGPT sign-in for plugins and connected
apps. Refresh an existing Codex routing configuration with
hansard account proxy client configure codex, then restart Codex. - Classification labels each conversation with one conversation request and
one task request, so labels from earlier releases are out of date. The
watcher relabels conversations active in the last 30 days;
hansard classify run --profile all --limit allrelabels older ones. - Rebuild and reload the preview browser extension to capture Devin cloud sessions, then choose Rescan all history once to reparse earlier Perplexity captures.
- The Behavior page is now the Changes tab in Stats. Project coordinators
share one harness, model, effort, and helper budget under
Settings → Coordinator. The Chart color setting and
orchestrate.maxActiveRunsare retired; other saved settings are unchanged.
0.5.5
macOS
Added
- Import local history from Crush, Goose, Continue, Roo Code, Kilo Code, Amp, Zed, DeepSeek Harness, Hermes, OpenClaw, OpenHands, Junie, Warp, Kiro, Augment, Qoder, TRAE, and MiniMax Code, including provider instruction files and memories where the agent keeps them. Explicit imports cover Amp, Kilo Code, OpenHands, Warp, Roomote, MiniMax Code, Qoder, Augment, and Kiro cloud or export records.
- Capture ChatGPT, Claude, Gemini, and Perplexity conversations from
signed-in Chromium tabs through a preview browser extension, a durable local
receiver, and
hansard browser serve,pair,status, andrevoke. - Share a reviewed text copy of a conversation as standalone HTML or Markdown. Tools, reasoning, context, and attachments are omitted.
- Add Stats → Year in review, a full-screen recap of one year with one assistant as downloadable 9:16 cards.
- Filter conversations by device, person, edited file, and commit, and filter Memories by harness, type, scope, device, update time, and folder.
- Count the lines each conversation’s file edits add and remove and show them on project overview rows, run rows, and coordinator workstreams.
- Recover memory history from archived file operations, including files deleted before backups began, with recovered versions and gaps labeled.
- Present the project coordinator as a chat with a composer, setup card, and settings popover.
- Add Copy resume command, Fork, and Continue from actions to the conversation header, and Dictation settings for the transcription engine and language.
Improved
- Rebuild Stats around range-aware charts: Tools & harnesses, Reliability, Projects, and Work outcome tabs; weekday rhythm by local day, spend per prompt, top models, busiest day, model takeover, working styles, and a combined memory activity chart.
- Redesign Settings as grouped cards with searchable rows, keycap shortcut editing, and rebuilt Imports, Projects, Portfolios, Privacy & Sharing, Personal Sync, Teams, Slack, Classification, and Handoff & helpers pages.
- Rebuild the conversation view with a compact header, hover message metadata, one-line thinking previews, finished turns folded under a Worked for row, a magnifying outline rail, overlay find bars, and full tool inputs, exit codes, and durations in expanded tool rows.
- Rebuild the continuation composer around a context strip, a permission label, and one model and effort chip that shows the settings that will apply. The coordinator and Deep Recall use the same chip.
- Show Accounts as compact rows with usage meters, unify router and native account identities, connect new coding accounts through one router sign-in, and keep router selections through cooldowns.
- Represent cloud accounts as separate Personal Sync owners, receive updates promptly by default, and retry blocked or interrupted transfers.
- Pair agent messages with verified inbox receipts and link them across archived conversations. Keep stale peer leases from hiding live sessions.
- Redesign Skills & plugins as a grouped list and show kind icons and a key on Graph nodes.
- Move viewer navigation validation off the request path, bound native transcript loading, and cancel superseded project statistics requests.
- Build every main-pane control from shared component primitives.
- Read LevelDB histories through isolated database snapshots.
- Refresh pricing and model catalogs while retaining historical entries.
- Illustrate the documentation website with product screenshots and diagrams.
All upgrade notes- Quit the desktop app before replacing it with the 0.5.5 installer, then
reopen it. CLI installations should restart an enabled watcher with
hansard watcher restart, then restart the viewer and any agent MCP connections so new recall tools, Claude model discovery, and peer delivery load. - Conversations archived before this release gain per-conversation line
counts and the revised tool usage fields after one
hansard rebuild --since all. No reimport is required. - The watcher archives newly supported local agents going forward. Import
their earlier history with
hansard import --sources crush,goose,continue,roo-code,roo-code-cli,kilo-code,kilo-code-legacy,amp,zed,deepseek-harness,hermes,openclaw,openhands,junie,warp,kiro,augment,qoder,trae,minimax-code --since all. Cloud and export sources remain explicit imports; seehansard help import. - Memory backup reconstructs earlier file versions from archived tool
operations. The watcher runs this with its next memory backup;
hansard memory backupruns it immediately. - Settings sections are regrouped under Personal, Archive, Agents, and Sharing, and Harnesses is renamed Coding agents. Saved settings are unchanged.
0.5.4
macOS
Build 2
Build 1
Added
- Import ChatGPT, Claude.ai, and Gemini Takeout exports directly from the app, including ZIPs, split exports, and extracted folders. The bundled backend provides progress, cancellation, account selection, and recovery after navigating away without requiring a separately installed CLI.
- Add guided two-step desktop setup using shared settings controls, visible defaults, and an export-import link after setup.
- Make Accounts the primary destination for provider connections, captured usage, reported subscription details, reset allowances, and capture status. Support saved native profiles, isolated browser profiles, and private usage API connections, with verified identity and duplicate-workspace protection.
- Add guided local request routing for Codex and Claude Code through a verified managed CLIProxyAPI runtime, with explicit client configuration and restore controls. Native account pools support ordered and quota-aware selection.
- Add Graph v2 Projects, Conversations, and Local views with evidence-backed relationships, background indexing, bounded layout workers, search, filtering, and connection inspection.
- Add persistent project coordinators, independent helpers, durable result delivery, and peer coordination across connected sessions.
- Configure ordered account-routing rules, preview account selection, reuse saved native logins, and continue routed conversations after usage limits.
Improved
- Upload ready Personal Sync snapshots promptly, capture the final snapshot after conversations become idle, and remove obsolete local outbox payloads while preserving pending uploads and recovery data.
- Preserve provider-reported account activity, optional subscription fields, availability, and reset eligibility without treating local session totals as account-wide usage.
- Improve account actions, composer picker sizing, model defaults, and dictation placement in the desktop webview.
- Resolve installed harness executables before terminal wrappers and allow Codex continuation outside Git repositories.
- Exclude received sessions from missing-source warnings, retain visible import errors alongside progress, and let classification skip sessions awaiting a parser refresh.
- Refresh pricing and model catalogs while retaining historical entries.
- Unify sidebar controls and typography, add shared motion behavior, and support multiple Graph filters. Open conversations at the latest message.
- Improve large and synced conversation reads with targeted archive access, consistent browser requests, and bounded replica validation.
- Simplify transcript context cards, show Claude subagents in expandable cards, and preserve original image links when a preview cannot decode.
- Reorganize website documentation around features with separate app and CLI instructions and verified signed download links.
All upgrade notes- Quit the desktop app before replacing it with the 0.5.4 installer, then
reopen it. The archive, settings, saved accounts, and credentials remain
outside the app bundle. CLI installations should restart an enabled watcher
with
hansard watcher restartand reopen the viewer after upgrading. - Graph v2 builds its derived index in the background on the first Graph visit. Its new relationships and the expanded account usage fields do not require a provider reimport. The next usage capture fills supported account fields.
- Automatic archiving, plan usage tracking, and secret redaction are selected by default in fresh desktop setup. Existing preferences are preserved.
- The standalone macOS download is available for Apple silicon. Intel packaging and the previously recorded Windows compatibility work remain separate follow-ups.
- Earlier 0.5.4 desktop builds can be replaced with the refreshed installer.
Refresh Claude subagent titles, Grok Build search queries, and Codex browser
request parsing with
hansard import --sources claude,codex-cli,codex-desktop,grok-build --since all.
0.5.3
macOS
Release
Preview 1
Added
-
Add standalone macOS packaging for Apple silicon and Intel, with a bundled Node.js runtime, native SQLite support, and on-device dictation helper. Public downloads depend on signing, notarization, and download hosting.
-
Import Muse Code native session journals, including linked subagent runs and named skill usage, and restore missing originals from preserved raw files.
-
Explore archived relationships through the knowledge graph viewer.
-
The transcript View menu adds Thinking, Tool activity, and Subagents switches beside the context-card filters, plus a Messages only preset that leaves just the user and assistant prose. The switches are display only.
-
List the harness’s full catalog under
/,$, and@in the composer. Claude Code discovery now covers skills synced from claude.ai, subagents, plugin commands and MCP servers, every MCP configuration scope, claude.ai connectors, and the bundled skills and built-in commands reported by the last turn; Codex discovery covers system and plugin skills from disk and MCP server status from the app server; Grok discovery covers agents and MCP servers. Discovery CLIs resolve through the known install locations, and a catalog the server reports as still pending is retried instead of cached. -
Use each harness’s own model and mode vocabulary in the composer. Claude Code offers its alias menu with older generations under More models and its permission modes (Manual, Accept edits, Plan, Auto, Bypass permissions); Codex offers the models and reasoning levels from its models cache with descriptions and its Read only, Auto, and Full access presets; Grok offers the models its CLI lists. Handoffs and new conversations share the same pickers. Claude Code turns can set the effort level (low, medium, high, xhigh, max) through the CLI’s
--effortflag.orchestrate.claudePermissionModeandorchestrate.grokPermissionModeacceptauto. -
Dictate into the composer. The microphone button records while held (or toggles, per its menu), the menu picks the input device, and punctuated text lands at the caret. The page records 16 kHz WAV and the viewer transcribes it with the best engine on the machine: the on-device macOS 26 speech recognizer through the bundled
hansard-speechhelper (hansard dictation setupbuilds it for the CLI), whisper.cpp with a local model (hansard dictation setup --whisper), OpenAI transcription withOPENAI_API_KEY, or a local command fromdictation.command. The browser’s own speech recognition is the fallback.hansard dictation statusshows the engines;dictation.backendanddictation.localepin an engine and a language. -
The plus menu and the inline pickers open flush with the composer bubble’s left edge and share its corner radius.
-
Offer every harness with a live adapter as a handoff target by default (
orchestrate.handoffProvidersnow defaults to Codex, Claude Code, Antigravity, OpenCode, and Grok). -
Open a project overview from the Conversations rail. Selecting a project name expands its folder and shows a work board for the project: header with working directories and providers, then Pinned, Waiting on you, Working, Resolved, and Recent conversations as collapsible groups with counts. Group membership comes only from orchestrator run state and from state a session reported about itself; archived history alone stays under Recent. The folder glyph only expands or collapses rows, the project label in a conversation header returns to the overview, and Settings → General → When opening a project can open the last viewed conversation in the project instead (
web.projectOpenBehavior). -
Add
report_statusandsession_contextto the collaboration-enabled Hansard MCP entry. A live Codex or Claude Code session can report working, blocked, needs_review, or done with a note and next step; the project overview shows the report marked as reported by the session, below any observed approval request or terminal run outcome. Conversations Hansard starts receive a short instruction naming the tools when messaging or helper launches are enabled. -
Fork an archived Codex or Claude Code conversation into a new native session from the composer’s Continue in… menu. The fork runs the harness’s own fork command (
codex exec fork,claude --resume --fork-session), keeps the full history, and links both conversations through lineage as “Forked into” and “Forked from”. The web route isPOST /api/orchestrate/fork. -
Find harness CLIs outside
PATH. Continuation preflight and the exec adapters resolve binaries from known install locations, including the Codex CLI inside the ChatGPT desktop bundle, the Claude Code launcher under~/.local/bin, and Grok under~/.grok/bin, so the desktop app offers the same harnesses as a login shell. Harnesses that cannot start stay in the menu as disabled rows with the reason. -
Rework the continuation composer: the insert control sits in the toolbar, popovers open above the bubble instead of overlapping it, the chosen harness tints the composer border and send action, and slash, dollar, and at completions answer from prefetched per-harness catalogs on every keystroke.
-
Track whether archived sessions still have their original provider files. The watcher records
state/source-presence.jsonevery 15 minutes and logs a warning naming the repair command when archived sessions lose their originals, for example after a reinstall recreates~/.claude.hansard statusshows the same count, andhansard doctornow runs the unavailable-sources check without--deep. -
Keep an archived conversation when its append-only JSONL source shrinks to less than half the archived message count (for example
claude -rstarting a fresh file under an old session id after~/.claudewas recreated). The shorter transcript is not imported; the event is recorded instate/source-regressions.jsonwith the archived raw copy’s path, the watcher logs a warning, andhansard statusandhansard doctorreport it. SetHANSARD_ACCEPT_SOURCE_REGRESSIONS=1on an import to accept the shorter transcript deliberately.
-
Fix Windows archive manifest flushes, refreshes of read-only raw projections, and folder-backed personal sync publication.
-
Commands typed in Claude Code bash mode render as a command card with their output instead of raw
<bash-input>and<bash-stdout>markup. -
Setup source inventory and the import storage preflight count Cursor’s workspace and global SQLite stores, so a Cursor install whose history exists only in those stores is offered for the initial import and measured.
-
Cursor workspace chats are fingerprinted by parser revision and conversation content, so parser changes and grown conversations reach the archive while store rewrites alone do not trigger a re-import.
-
Foreground imports of sources that continue across several bounded slices print one scan line, keep cumulative imported and skipped counts, and report whole-source session totals. Cursor slices no longer recount sessions an earlier slice already handled.
-
Make the stats section tabs work on the Projects page. The tab strip was meant to be hidden there with every section shown, but its layout rule kept it visible with no effect, so switching tabs changed nothing. The Stats page and a project’s stats now share the strip, and either view keeps the selected tab in its URL.
-
Classify Claude Code transcripts whose first prompt line exceeds a megabyte, for example a desktop session opened with pasted screenshots. The importer read only the first megabyte when deciding whether a
~/.claude/projectsfile was a conversation, saw nothing but the queue records ahead of that prompt, and left the session out of every import. The classifier now streams the leading lines, and the classification cache version is bumped so cached verdicts are recomputed on the next scan; no reimport is needed. -
Escalate an import worker’s heap after it dies of V8 heap exhaustion (768, 1536, 3072, then 4096 MB, remembered in
state/import-heap.json) instead of retrying at the same size on every tick, and log the sizes rather than the garbage-collector trace. -
Let
hansard repair claude-code-backupsfall back to the per-session.raw/copy, verified against the recorded digest, when a raw record’s content-addressed archive object is missing.
-
Keep the npm README focused on installation, first use, and upgrade guidance. Detailed maintainer references remain in the source repository’s
docs/tree. -
Ship only the compiled graph viewer. Graphology, Sigma, and the browser build tools remain development dependencies, with bundled licenses included in the package notices.
-
Refresh model identities and pricing while retaining historical rates.
-
Earlier pages of a long transcript load automatically. Ordinary sessions stream in completely; a very large session loads ahead of the scroll position. The former Load earlier messages button remains only as a retry after a failed page request.
-
Subagent runs render as pills that share one wrapped row. A pill opens the run’s transcript and its arrow expands the prompt and result in place. Runs that started before the loaded page now appear with the page that holds their start time instead of collecting at the top.
-
Dates and times now share one display standard across the viewer and the CLI: a medium date with a short time (
Sep 7, 2026, 9:10 AM) and a medium date (Sep 7, 2026), both in the local time zone and locale. Relative times keep their4h agoform and fall back to the standard date after a week. Raw ISO days no longer appear in the pricing footnote, the model lifecycle chart, behavior trend buckets, or the import summary, and the Memories history list uses the standard datetime. -
Default new Personal Sync spaces to the versioned immutable-revision protocol. Compatible Hansard releases exchange complete revisions by wire capability rather than package-version equality. Upload and receive schedules now run independently of provider imports while the desktop app or watcher is active; existing
archive-filesspaces retain their protocol. -
Capture OpenCode and MiMo Code SQLite state coherently, then publish only an audited history projection for conversion and raw replay. Credential, account, and share-secret tables are excluded, session share URLs are cleared, and raw manifests record the privacy filter. Legacy OpenCode CLI, Desktop, and Web selectors forward to the canonical source without creating separate roots or parser identities.
-
Preserve OpenCode and MiMo Code ordered current todo snapshots and OpenCode global JSON session diffs with bounded summaries, full-list digests, exact source accounting, and offline replay. OpenCode JSON
session/shareandsession_sharerecords remain outside raw publication and replay, and early embedded share secrets are removed before conversion and raw publication. -
Render web links in conversations, memories, and recall answers in a dedicated link color without a standing underline, led by the linked site’s favicon and a hover title with the full URL. The page requests icons only from the viewer’s own
/api/faviconroute; the viewer process fetches each icon from the linked site over HTTPS, refuses private and reserved addresses, caches results undercache/link-favicons/for seven days, and never uses a third-party favicon service. A letter badge stands in while an icon loads or when a site has none. Setweb.linkFaviconstofalsewithhansard config set web.linkFavicons falseto keep the viewer from contacting linked sites. Restarthansard webor the desktop app to pick up the new rendering. -
Make the session tree folder glyph the disclosure control, replacing the chevron: it crossfades between its closed and open forms as a folder opens or closes. Attachment cards redraw their document lines on hover or keyboard focus. Both use the shared motion tokens and follow reduced-motion preferences, and the attachment glyphs now come from the vendored icon set.
All upgrade notes-
Claude Code and Claude SDK histories (
cli-history,claude-sdk-history) include corrected timestamp handling. Records without their own clock, such as file history snapshots, no longer take the import time, which had moved the session start after its end. Refresh existing archives with:Terminal window hansard import --sources claude,claude-sdk --since all -
Cline task and SDK histories now use enforced source-audited cohorts. Refresh existing Cline archives and their search entries with:
Terminal window hansard import --sources cline,cline-cli --since allhansard index rebuild -
OpenCode now uses one canonical import selector and physical store owner with four enforced JSON cohorts and five enforced SQLite cohorts. MiMo Code remains an independent enforced SQLite source rather than inheriting an OpenCode storage contract. Refresh both archives and their search entries with:
Terminal window hansard import --source opencode --since allhansard import --source mimo-code --since allhansard index rebuild -
Cursor workspace chats (
cursor-workspace-sqlite) include corrected legacy timestamps. Legacy chat tabs without per-message timestamps are dated from the tab’s recorded send time instead of the workspace store’s modification time, which moved whenever Cursor reopened the workspace. Refresh existing archives with:Terminal window hansard import --source cursor --since all
-
0.5.2
Added
- Add resumable in-view conversations, handoffs, new sessions, steering, approvals, context resources, and persistent transports for supported agent harnesses.
- Add local and cloud import coverage for Prime Agent, Claude Cowork, Copilot in VS Code and the Copilot coding agent, Cursor cloud, Factory cloud, Codex cloud, Devin cloud, Jules, Qwen Code, iFlow, Kimi, Mistral Vibe, MiMo Code, CodeBuddy, Cline SDK sessions, and recoverable Windsurf histories.
- Add a first-class memory archive with snapshots, search, viewer history, project grouping, agent retrieval integrations, and memory statistics.
- Add subscription-usage history, task and work-frontier analysis, model/lab lifecycle views, reliability statistics, and model effort and service-tier breakdowns.
- Add structured file and commit facets to archived tool usage and expose them
through
hansard history --fileandhansard history --commit.
- Keep conversation navigation responsive during imports by using bounded projections, targeted detail reads, persisted Stats payloads, background reconciliation, bounded viewer caches, and one reusable Stats generation for presentation-range switches during archive writes.
- Keep the Pricing reference canonical by excluding unused provider-route aliases from the release catalog and disclosing archive-observed identities that still need mapping separately with their recorded usage.
- Preserve deeper provider lifecycle, tool, token, model, timestamp, raw-source, and attachment evidence while keeping machine probes, SDK batch jobs, and subagents out of human-conversation metrics where appropriate.
- Add transactional raw recovery, stale-parser refresh, source-retention diagnostics, and safer full-update behavior for provider histories that have aged out.
- Add desktop deep links for cold and warm opens, app-first Slack session links, adaptive splash themes, and a public documentation site.
- Treat archived content as untrusted across recall and viewer surfaces, enforce archive path containment, restrict link schemes, and strengthen viewer response headers.
- Remove install-time side effects, ship an audited shrinkwrap, verify exact package dependency closure, and gate publishing through the reviewed trusted workflow and package tripwire.
- Exclude raw provider backups from Teams, require explicit sharing consent, remove legacy raw team objects, scan public files for identifying data, and use visibly synthetic fixtures throughout release checks.
All upgrade notes- Version 0.5.2 advances the archive, view, tool-usage, search-index, behavior,
SDK-session, and parser contracts. Existing archives remain readable, but a
complete local refresh is required to backfill the corrected importer metadata
and normalized events. After upgrading, run
hansard doctor, preview withhansard update --dry-run --since all, then runhansard update --yes --since allafter confirming the original provider histories or usable raw backups are available. - Restart a running watcher after a normal package-only upgrade with
hansard watcher restart. The full update command handles watcher restart as part of its maintenance flow.