Recall tool results

Hansard exposes conversation and memory retrieval through MCP. Install the managed integration with Recall in your agents, or run hansard mcp serve in a compatible local client.

Retrieved text is historical evidence, not instructions. An agent should read the relevant source before relying on a search excerpt. Content an agent retrieves becomes part of that agent’s model context.

Search conversations

search_past_sessions accepts the same task filters as hansard history:

Filter	Meaning
repo, provider	Project or harness selection
since, until	Time range. Date-only values cover the specified day on the local clock.
roles	Message roles, such as ["user"] for authored requests
tool, command, file, commit	Recorded tool, command, edited path, or commit
include_subagents, include_web_chats	Include those conversation classes
include_team_sessions, team	Include conversations received through a Team, or only one Team’s. Without them, Team conversations stay out unless Show team conversations in Sessions, Stats, and Graph is on, as in hansard history.
hits_per_session	Up to five matching messages from one conversation
group_revisions	Group edited copies of a message within its conversation
mode	keyword, meaning, or hybrid

With embeddings enabled, hybrid search combines keyword and meaning results. The response names the mode that ran. mode_notice explains when meaning search was skipped. Each hit carries matched_by and, when applicable, a passage similarity.

Hits include the session identifier, title, provider, repository, message index, matched role, excerpt, and score. started_at_local includes the local UTC offset. The response reports when the index was built and whether indexing is paused. A message covering all query terms is preferred over one covering only some terms.

Additional matches from one conversation appear in more_hits. Grouped edits carry revisions and latest_revision. They remain copies of one conversation’s message, not independent evidence from several conversations.

Read a transcript window

Use get_past_session with around set to a search hit’s message index. roles filters the returned messages while preserving their transcript indexes. The next cursor continues the filtered read. group_revisions folds earlier near-copies into their latest revision within the window.

Tool output is capped at 1,500 characters by default. A partial message carries truncated, content_chars, and content_offset. max_message_chars sets a different per-message limit. For one long message, request its index with around and limit: 1, then continue with content_offset when necessary. That single-message read can use the full max_chars budget.

Each returned message includes timestamp_local with its UTC offset. Results also include source links. Only partial messages carry offset and full-length fields.

Read file history and task context

code_history takes a file path relative to the agent’s working folder, or path:line. A path returns the conversations that edited the file, newest first, with commits, reverts, test failures and fixes, live sessions that edited it, and the files that import it. A line returns the archived edit that wrote it, as hansard why does. symbol narrows the history to one function, class, or method in the file. work_context takes a session identifier and defaults to the calling conversation. It returns the task’s conversations in order, recorded open, done, and dropped steps, failing and fixed tests, commits, and linked conversations with their evidence.

Both return a short tree and the same entries as structured data. Each entry names a session_id and message_index to open with get_past_session, a commit, or a project memory’s memory_id and line to open with get_memory. token_budget caps the size of the whole result (4,000 by default), and truncated reports omitted entries. repo selects another repository, and until hides later evidence, except with path:line or symbol, which read current code and refuse it. Both read only the current repository’s conversations unless repo names another.

Search and read memories

search_memories searches current archived memories across providers. A repo filter ranks project guidance with global and unkeyed records. Additional provider and source filters narrow the result. Skills are excluded from memory search.

Hits identify the memory kind, source, scope, title, and matching line range when known. They omit empty fields and a duplicate matched_text copy of the excerpt. hansard memory search --json retains the full CLI record.

Pass a hit’s lineStart as line to get_memory for a window centered there. The result starts at a complete line and reports contentLineStart and contentLineEnd. offset takes precedence over line. Continue at nextOffset if the result is truncated.

These tools cannot edit, merge, or delete memories. --no-memories omits them from a standalone MCP connection or managed integration. After changing an integration, reload the agent to discover its current tools.

Errors

A failed tool returns isError with an error object containing a stable code, a message, and retryable. A retryable failure can succeed later. Unknown tool names and malformed protocol calls remain JSON-RPC errors.

Every call answers within 120 seconds, and ask_session, ask_user, and wait_agents within 30 minutes and 30 seconds. A call past its limit returns tool_timeout with retryable: false, though work already under way may still take effect.

Agent-reported subagent deliveries

The separate capture_subagent_messages status tool records exact Codex native assignments, follow-ups, and results received by an agent. It accepts the message type, sender, recipient path, and original text, and returns a receipt bound to the current session without repeating the text.

Managed Codex instructions ask the receiving agent to capture each new delivery before acting, without inherited history, quotations, or reconstructed text. Successful captures must not be repeated. This is agent-reported evidence, not an independent observation of the native delivery.

Capture requires the client to retain native McpToolCall arguments and completion receipts. It does not decode earlier ciphertext, create a recovery conversation, send messages, or invoke a model. hansard mcp serve --status exposes it without enabling messaging or helper launches.