feat: add proactive cross-chat issue review

.claude/skills/run-postgres-integration-tests/SKILL.md

@@ -0,0 +1,41 @@
+---
+name: run-postgres-integration-tests
+description: Run AgentsView PostgreSQL integration and backend-parity tests against a dedicated disposable local database. Use for pgtest failures, PostgreSQL storage changes, or release gates that require TEST_PG_URL.
+---
+
+# Run PostgreSQL integration tests
+
+1. Read `docs/agents/testing.md`, `docs/agents/storage.md`, and
+   `docs/agents/build.md`.
+2. Never use production, shared, or persistent archive databases. The tests
+   drop and recreate test schemas.
+3. Create a unique cluster below `$env:TEMP`, bind it only to `127.0.0.1`, and
+   use a free non-default port. Initialize it with PostgreSQL 17 `initdb`, UTF-8,
+   locale `C`, user `postgres`, and local trust authentication.
+4. Start with `pg_ctl -w`, create a dedicated `agentsview_test` database, and
+   verify the database and server version with `psql`.
+5. Set `TEST_PG_URL` only in the test process. Set `CGO_ENABLED=1` and verify
+   the compiler target is `x86_64-w64-mingw32` before running Go.
+6. Run the smallest gate first:
+
+   ```powershell
+   go test -tags 'fts5,pgtest' ./internal/postgres/... -run '^TestIssueReviewRowsConditionallyLoadsResultTail$' -v -count=1
+   ```
+
+7. Run the full canonical gate only after the focused test passes:
+
+   ```powershell
+   go test -tags 'fts5,pgtest' ./internal/postgres/... -json -count=1
+   ```
+
+   For large output, retain only failed test events and nearby output in the
+   conversation. Keep the unfiltered JSON outside Git if exact diagnosis is
+   needed.
+8. Do not repeat a failure unchanged. Identify the failing test, then rerun
+   that test with `-run '^ExactTestName$'` before another full suite.
+9. Stop the exact scratch server with `pg_ctl -w stop -t 360`. A full suite can
+   leave a large checkpoint that legitimately exceeds 30 seconds; while the
+   log shows checkpoint progress, wait instead of killing the process. Remove
+   the cluster only after resolving the absolute path and proving it is a child
+   named `agentsview-pgtest-*` below `$env:TEMP`. Preserve logs on failure until
+   the cause is recorded.

.claude/skills/run-postgres-integration-tests/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "PostgreSQL Integration Tests"
+  short_description: "Run isolated PostgreSQL parity tests safely"
+  default_prompt: "Use $run-postgres-integration-tests to run the PostgreSQL integration suite against a disposable local database."

docs/insights.md

@@ -16,6 +16,72 @@ live under the **More** dropdown as of 0.21.0, which leaves
 
 ![Insights page](/assets/generated/screenshots/insights.png)
 
+## Proactive Issue Review
+
+The top of the Insights page continuously ranks recurring problems and
+automation opportunities across the selected chats. It is deterministic and
+server-backed; generating an AI insight is not required. The review detects:
+
+- failed commands, edits, builds, tests, migrations, Git/GitHub operations,
+  missing files or dependencies, permissions, network errors, timeouts, and
+  tool crashes;
+- a successful retry after a failed identical call, persistent repeated waits
+  or polling, and the
+  same substantial workflow repeated across chats or projects;
+- slow non-wait tools, exact-normalized user requests repeated across chats,
+  explicit user corrections, and assistant-reported blockers;
+- allowlisted Codex response, tool-router, hook, session, and PowerShell
+  snapshot failures when local telemetry is available.
+
+When an `exec` wrapper calls only one nested tool, Issue Review attributes the
+finding and duration to that nested tool. Mixed-tool wrappers remain attributed
+to `exec` because the outer result cannot identify one responsible tool.
+
+The global date, project, machine, agent, termination, automation, and
+one-shot filters apply first. The panel adds exact chat, folder, category, tool,
+outcome, severity, confidence, status, suggested-action, and minimum-occurrence
+filters. Each finding keeps at most five redacted evidence excerpts and links
+to the exact message ordinal when one exists. Results are returned in pages of
+100; **Load more findings** continues through the full filtered result set.
+
+You can name and save up to 50 complete Issue Review filter sets. Selecting a
+saved view restores its filters and refreshes the results. Saving the same name
+updates that view, and deleting it removes only the preset. Saved views stay in
+the current browser profile; they are not stored in the archive or synced
+between devices.
+
+The panel refreshes when its filters or the global scope changes, after a
+debounced data-sync event, every hour while open, and on manual retry.
+Background refreshes use the one-hour analysis cache so frequent sync events
+cannot trigger repeated full-archive scans; **Refresh now** bypasses the cache.
+If a refresh fails, the last successful result remains visible with a warning.
+Acknowledgement and suppression decisions are loaded separately on every
+request, so they take effect immediately without invalidating or mutating the
+cached detector result. Acknowledgements reopen when a finding appears on a
+later date. Suppression can last 1, 7, or 30 days, or remain permanent; hidden
+findings stay available through the review-state filter.
+
+Durations are measured only when start and completion events can be paired.
+For slow tools, occurrences count calls at least 30 seconds long, while p95 is
+calculated from every measured sample for that tool. Coverage is measured
+samples divided by all scoped calls for the tool. The “excess” duration is a
+triage proxy above 30 seconds, not a claim that all of that time was wasted.
+Wait/sleep tools and negative or malformed durations are excluded.
+
+Large tool results retain bounded context from both the beginning and end, so a
+stable compiler, test, or command error near the tail remains classifiable.
+
+The optional Codex supplement reads `~/.codex/logs_2.sqlite` in read-only mode
+only for scoped session IDs and exact tool-call IDs. This supplement is
+available only with the local SQLite store; PostgreSQL and DuckDB use timing
+events already mirrored into their own stores. Telemetry and chat/tool
+excerpts are redacted before the API returns them. The panel shows a non-blocking
+warning when local telemetry is missing or unavailable; chat and tool-result
+analysis remains active.
+
+See [Proactive Issue Review handover](issue-review-handover.md) for the detector
+architecture, validation matrix, deployment gates, and follow-up roadmap.
+
 ## Insight Types
 
 There are three generation modes, selected from the dropdown at

docs/internal/session-format-sources.md

@@ -195,7 +195,11 @@ Grok section and remove the explicit registry exception in the coverage test.
   trees, so this evidence does not establish IDE, desktop, or `codex exec`
   activity-hint coverage. Locally observed Codex app builds can write the same
   schema, but that is observational evidence rather than a public
-  compatibility guarantee. Agentsview derives the hint path as
+  compatibility guarantee. Reverified 2026-08-08 against local desktop
+  rollouts: `event_msg` records with `payload.type="agent_message"`,
+  `phase="commentary"`, and `message` are ordered assistant progress updates;
+  Agentsview preserves them as assistant messages with
+  `source_subtype="commentary"`. Agentsview derives the hint path as
   `<configured-sessions-root>/../history.jsonl`; a custom sessions root
   without that sibling, or `HistoryPersistence::None`, degrades to ordinary
   watcher behavior, degraded-coverage polling when applicable, and the daily