nfemmanuel/iranti

What's new

M7: Attendant response file-change autowrite

Fixes the root cause of file-change recall not surfacing: project/{id}/file/{name} entity facts were never written, so demand-driven recall had nothing to retrieve.

The Attendant now runs an automatic scan pass at every post-response attend call. It:

• Scans the agent's response text for file path mentions (Windows and Unix paths)
• Infers the most likely action (edited / created / read) from the ±150-char context window around each path
• Writes one durable fact per distinct file-scoped entity (project/{id}/file/{basename}) via the existing librarianWrite + attendant_autowrite pipeline
• Resets compliance pressure counters if any facts were written
• Surfaces the outcome as responseFileCapture on the attend result

No host changes required — collection is fully Attendant-side.

New public API

| Symbol | Kind | Description |
|---|---|---|
| detectFileAction(ctx) | pure function | Infers edited / created / read from a context window |
| ResponseFileCaptureResult | interface | Outcome of the post-response file scan (batchId, filesDetected, factsWritten, entities, skipped, durationMs) |
| AttendResult.responseFileCapture | field | Present on post-response attend results; absent on pre-response / mid-turn |

Test coverage

12 new M7 tests (test:m7-response-file-capture). All existing regressions clean (M2, B6, B7, B8, file-change-recall, artifact-capture, user-operating-rules).

Version bump

0.3.31 → 0.3.32 across npm, TypeScript SDK, and Python client.