Report the last finished attempt in telemetry

.agent/handoff/243.md

@@ -0,0 +1,68 @@
+# Handoff note: issue #243
+
+- issue: 243
+- started_at: 2026-10-14T13:48:04.076479Z
+- reason: token_soft_threshold
+- last_work_commit: 24d85f3236730f040cec81b2646e6aba687da873
+
+## Summary
+
+Implemented the last-finished attempt telemetry end to end on branch
+`agent/issue-243` (work uncommitted; sealing will commit it).
+
+- `simple_coding_agent/telemetry.py`: `finish_attempt(summary=None)` normalizes
+  and stores `last_finished` (`attempt_id`, `issue_number`, `issue_title`,
+  `outcome`, `finished_at`, `budget_used`/`turns` as `measured` figures,
+  `pull_request: {number, url} | null`); legacy summaries seed title
+  `"unknown"` and `pull_request: null`; malformed summaries log once and keep
+  the previous view. `snapshot()` now serves `last_finished`. Added
+  `load_last_finished(data_dir)` built from the latest completed attempt
+  archive record.
+- `simple_coding_agent/observability.py`: added `read_latest_completed_attempt`
+  scanning `data_dir/logs/<issue>/<started>/attempt.json` (skips
+  unreadable/incomplete records; latest by `(completed_at, started_at)`).
+- `simple_coding_agent/lifecycle.py`: new `data_dir` constructor param with
+  startup seeding (`_seed_telemetry_last_finished`, never raises);
+  `_telemetry_finish(summary)` passes summaries through;
+  `_archive_pull_request` writes the PR into the archive at publication
+  (hooked into `_publish_bounded`, `_publish_terminal`, recovery retry);
+  `_commit_finalization` writes issue number/title into the outcome record,
+  backfills the PR safety net, and finishes telemetry with the full summary;
+  `_write_outcome` now returns `completed_at`.
+- `tests/fakes.py`: `FakePublisher`/`PublishedAttempt` accept `pull_request`.
+- `simple_coding_agent/__main__.py`: passes `data_dir` to `AgentLifecycle`.
+- New `tests/test_telemetry_last_finished.py` (15 tests): reporter shape,
+  legacy/unknown handling, unavailable figures, malformed-summary behavior,
+  finalisation wiring + archive PR, restart seeding, null with no history,
+  latest-wins seeding, alongside-active reporting.
+- `code-review` skill invoked this run: Standards (0 hard violations,
+  4 smells, worst = duplicated PR normalizer across telemetry/lifecycle) and
+  Spec (0 missing, 0 creep, 3 wrong-looking, worst = shared
+  `_build_failed_logged` flag) reports aggregated in the run transcript.
+
+## Remaining work
+
+1. Re-run `tests/test_telemetry_last_finished.py` (was 14/15 green; fixed the
+   malformed-summary log-count expectation from 2 to 1 to match the
+   log-once-per-episode pattern; re-run was blocked by the budget hook).
+2. Decide on review findings: extract one shared PR normalizer (strongest
+   Standards finding); consider separate warning flags for snapshot vs finish
+   failures, timestamp parsing instead of string ordering in the archive scan,
+   and narrowing the `except TypeError: finish()` shim.
+3. Run `ruff check`, full `pytest` on both gate Pythons (3.12 and 3.13).
+4. Small docs touch: `docs/agents/telemetry-snapshot.md` `last_finished`
+   bullet (measured basis, legacy `"unknown"` title) and
+   `examples/telemetry/snapshot-idle-with-last-finished.json` (`budget_used`
+   basis `estimated` -> `measured` to match the issue spec).
+5. Commit the work on `agent/issue-243`.
+
+## Evidence
+
+- `tests/test_telemetry_last_finished.py` (new, untracked): 14/15 passing;
+  the single failure was the test's own log-count expectation, corrected to 1
+  (mirrors `test_malformed_seed_keeps_the_previous_view_and_logs_once`).
+- Touched files (uncommitted): `simple_coding_agent/__main__.py`,
+  `simple_coding_agent/lifecycle.py`, `simple_coding_agent/observability.py`,
+  `simple_coding_agent/telemetry.py`, `tests/fakes.py`, plus the new test file.
+- Gate not yet run: ruff + full pytest suites still pending.
+- Code-review gate satisfied: `code-review` invoked with both axes reported.

docs/agents/telemetry-snapshot.md

@@ -58,7 +58,10 @@ synthetic examples under `examples/telemetry/` are the conformance target.
   `finished_at`, `budget_used`, `turns`, `pull_request: {number, url} |
   null`) is held in memory, updated at finalisation, and seeded at startup
   from disk, so it survives restarts. It is shown alongside `current_attempt`
-  too.
+  too. Its `budget_used` and `turns` figures use the `measured` basis (or
+  `unavailable` when no token budget was reported); archives written before
+  the title and pull-request fields existed seed an `"unknown"` title and a
+  null pull request.
 - `recent_events` is the last up to 50 redacted log events, newest last, each
   `{timestamp, level, event, issue_number, phase, text}` with `text` truncated
   to 300 characters. High-volume `limits_checked` and `tool_call` events are

examples/telemetry/snapshot-idle-with-last-finished.json

@@ -38,7 +38,7 @@
     "finished_at": "2026-09-29T15:40:00Z",
     "budget_used": {
       "value": 812345,
-      "basis": "estimated"
+      "basis": "measured"
     },
     "turns": {
       "value": 23,

simple_coding_agent/__main__.py

@@ -144,6 +144,7 @@ def _attempt_sink(issue_number: int) -> AttemptArchive | None:
         control_store=ControlStore(config.data_dir),
         repository=config.target_repo,
         telemetry=telemetry,
+        data_dir=config.data_dir,
     )
     interruption_handler = AttemptInterruptionHandler(
         event_log=lambda event, detail="", level="INFO", issue_number=None: logger.emit(