feat(claw-server): searchable audit via a task summary

[Dani Akash (DaniAkash)]

What

Lets users search their audit history by what a task was about, and records that same summary in product analytics. On name_session an agent declares a short, PII-scrubbed summary alongside the existing task category; the server stores it for full-text audit search, surfaces it in the audit UI, and attaches it to the task-declared analytics event.

flowchart LR
  A["name_session(name, category, summary)"] --> S["scrub structural PII"]
  S --> T["tasks.task_summary (SQLite)"]
  S --> F["FTS5 task_search index"]
  S --> AN["agent_session_task_declared (analytics)"]
  Q["audit search bar (?search=)"] --> L["list sessions"]
  T --> L
  F --> L
  L --> R["audit rows + detail header"]

Loading

How

• Capture. name_session gains an optional summary argument (last write wins, so a later call refines it). It is scrubbed server-side for structural PII (emails, URLs, file paths, bare domains, long digit runs) before it is stored, indexed, recorded on the audit dispatch, or sent to analytics.
• Storage and search. A new tasks.task_summary column holds the canonical value; a companion FTS5 task_search table holds the searchable index. Both are written in one transaction. The summary is written out of band, so it is excluded from the task recompute upsert and survives event replay; both stores are purged together on retention deletes. The existing ?search= path now also matches summary content through the FTS5 index (porter/unicode61 stemming for invoice vs invoices), alongside the current title / agent / site match.
• Analytics. The scrubbed summary is attached to the agent_session_task_declared event next to the category. It is the analytics catalog's one free-text property: optional (a category-only declaration still sends) and length-bounded at the boundary.
• Display. The summary rides on the SessionSummary contract and renders as a two-line snippet under the target in the audit table and as a lead line in the task detail header. The detail-page tab strip now scrolls horizontally (scrollbar hidden) for sessions that touch many browser tabs, and the search placeholder hints that summaries are searchable.

PII handling

Two layers. The tool description and agent prompt instruct the model to keep the summary free of names, emails, URLs, paths, and account numbers. Independently, the server scrubs structural PII from the summary once and uses that scrubbed value everywhere it is persisted or transmitted: the canonical column, the search index, the sessions API, the audit dispatch arguments, and the analytics event. The scrub is best-effort and structural, not a full classifier.

Testing

• Rust: full crate suite (lib + integration targets) green; clippy -D warnings clean; new tests cover the scrub, the dispatch-args substitution, recompute-survival plus retention purge, the FTS search path, and the optional free-text analytics property.
• Contract: schema and fixture tests green; the generated client is byte-identical to the spec.
• Extension: audit list and task-detail screen tests green, including the summary render; typecheck and lint clean.

[greptile-apps]

Greptile Summary

The PR adds locally stored, scrubbed task summaries and makes them available through audit search and session results.

• Adds summary capture and structural scrubbing to name_session.
• Stores canonical summaries alongside an FTS5 search index.
• Extends the session API contract and generated clients with taskSummary.
• Displays summary snippets in the audit table and expands the search hint.

Confidence Score: 3/5

The PR is not yet safe to merge because malformed name_session summary values can still place unsanitized sensitive text in the audit timeline.

The previous string-summary leak is fixed, but a non-string summary skips the scrubber and is serialized unchanged into persisted dispatch arguments.

Files Needing Attention: packages/browseros-agent/apps/claw-server-rust/src/api/mcp/service.rs

Security Review

The string-summary leak reported previously is fixed, but non-string summary values still bypass scrubbing and can persist sensitive nested text in audit dispatch arguments.

Important Files Changed

Filename	Overview
packages/browseros-agent/apps/claw-server-rust/src/api/mcp/service.rs	Adds summary capture and fixes the prior raw-string leak, but non-string summary values still bypass sanitization before audit persistence.
packages/browseros-agent/apps/claw-server-rust/src/db/audit_log.rs	Adds transactional summary/index writes, FTS-backed filtering, recompute preservation, and retention cleanup.
packages/browseros-agent/apps/claw-server-rust/src/db/migration.rs	Adds the nullable canonical summary column and standalone FTS5 index.
packages/browseros-agent/apps/claw-server-rust/src/api/http/sessions.rs	Projects stored summaries into both historical and live session API responses.
packages/browseros-agent/contracts/claw-api/schemas/sessions.yaml	Adds the optional taskSummary field consistently with generated Rust and TypeScript models.
packages/browseros-agent/apps/claw-app/screens/audit/audit.columns.tsx	Renders an optional two-line task-summary snippet beneath the task name.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A["name_session arguments"] --> B{"summary is string?"}
  B -->|Yes| C["scrub_summary"]
  C --> D["tasks.task_summary"]
  C --> E["FTS5 task_search"]
  C --> F["scrubbed audit args"]
  B -->|No| G["raw_args clone"]
  G --> H["audit dispatch args_json"]
  D --> I["SessionSummary"]
  E --> J["Audit search"]
  I --> K["Audit summary snippet"]

Loading

Prompt To Fix All With AI

### Issue 1
packages/browseros-agent/apps/claw-server-rust/src/api/mcp/service.rs:197-200
**Non-string summaries bypass scrubbing**

When `name_session` receives a valid string `name` and an object or array as `summary`, `Value::as_str` returns `None`, so this branch records the original arguments unchanged. Sensitive text nested in that value is therefore persisted verbatim in the audit detail timeline instead of being scrubbed or rejected.

**How this was verified:** The unvalidated non-string value follows the `None => raw_args.clone()` branch into the dispatch recorder's `args_json`.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (2): Last reviewed commit: "fix(claw-server): scrub the summary in t..." | Re-trigger Greptile}

[github-actions]

✅ Tests passed: 2550/2553

Ran 15 of 15 suites (0 not affected by this change).

Suite	Passed	Failed	Skipped
✅ server-agent	265/265	0	0
✅ server-api	170/170	0	0
✅ server-tools	254/254	0	0
✅ server-browser	10/10	0	0
✅ server-integration	10/10	0	0
✅ server-lib	141/141	0	0
✅ server-root	38/41	0	3
✅ agent	327/327	0	0
✅ claw-app	394/394	0	0
✅ claw-onboard	86/86	0	0
✅ build	44/44	0	0
✅ release	65/65	0	0
✅ claw-server-rust	637/637	0	0
✅ claw-server-rust-quality	passed	0	0
✅ claw-mcp	109/109	0	0

passed = ran successfully but emits no JUnit counts (a lint/format gate).

View workflow run

[Dani Akash (DaniAkash)]

Greptile (@greptileai)