Repository navigation
Conversation
Every refusal / rollback / blocked row exec.sh writes (25 log_build call sites) now carries " · guard=<id>" right after its detail text, inside the detail cell: column readers (status, build-log-read, script-owners, done-drift-check, row-groups, doctor steps:, progress-report, obligation-check, ci-gates, metrics-ledger classify) are unchanged. metrics-trend.sh adds a "blocked by guard:" count line from docs/BUILD-LOG.md, rows from an older exec.sh counted as unnamed, so the harvested metrics file reports it. Same verdicts, exit codes and refusal text; no new process per exec (process budget 132 starts on master and branch). Fixture tests/wave2/test-buildlog-guard-name.sh: master 10 passed 5 failed, branch 15 passed 0 failed. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
|
Review: MERGE-AFTER-FIX. Value: The mechanism is sound and minimal:
The largest bucket misses the goal.
Also, Regression risk: Low. I checked every BUILD-LOG reader myself, not from the PR body:
No verdict, exit code or on-screen text changed. The code is bash 3.2-safe and works in both layouts. Correctness: Fixtures:
The input is real exec runs plus captured rows. Minor: Context cost: 65568 → 65568 (measured). Time cost: 0 process starts per exec. Exact fixes required:
Generated by Claude Code |
…ch-exit is not a guard Independent review of #813 (MERGE-AFTER-FIX): - mxbuild-gate covered four different stops. Now: mxbuild-missing (toolchain absent), pre-existing-errors (model already broken), mxbuild-crash (gate could not run), and mxbuild-gate only for errors the script introduced (the real catches). - patch-exit is called out as not a guard; metrics-trend.sh counts it on a separate "script failed" line instead of "blocked by guard". - metrics-trend.sh matches " · guard=<id>" only, so a bare guard= quoted in detail text is not counted. Fixture: real mxbuild-missing refusal run, per-site id checks for the split, a derived row with a bare guard= in detail text, and a patch-exit row (5 of 22 fail on the pre-fix head, 22/22 pass). Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
|
Review fixes pushed in f41b78c:
Fixture results:
Needs a re-review. Generated by Claude Code |
|
Review: MERGE-AFTER-FIX — the split landed at the right lines, but the two
The remaining gap is the
A
Context cost: 65,568 → 65,568 (measured).
Generated by Claude Code |
…, not mxbuild-gate (#813 review) The --patch rollback branch only runs when the gate was skipped, crashed or unverified, so its rows were miscounted as catches. Tag by GATE_STATE; fixture D checks both rows statically and runs a --patch crash end to end. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
|
Review: MERGE — re-review of dc784df: the
Context cost: 65,568 → 65,568 (measured). Generated by Claude Code |
Direct lane
There is no issue. This comes from the over-engineering audit of 2026-11-20, which found that in the one field metrics file we have (#804), 13 of 136 execs were "blocked or refused", and no row said which guard stopped them. Until rows name the guard, nobody can measure a guard's catch rate against its false-refusal rate, and so nobody can decide which guards to keep.
What changed and why (one paragraph):
project-bin/exec.sh'slog_buildtakes an optional third argument, the guard id.· guard=<id>right after its detail text. That covers all 25 call sites with status 🚫 / ✗ refused / ❌ / ↩.plan-rule,stop-row-scan,row-group-gate,lock-studio-pro,commit-first,uncommitted-model,marketplace-gate,mxcli-check,mxbuild-gate,patch-exit,hash-verify.project-bin/metrics-trend.shadds one line under "outcomes":blocked by guard: mxbuild-gate 15, plan-rule 7, …, unnamed N. Rows from an older exec.sh count asunnamed.bin/harvest-learnings.shalready pastes this summary into the inbox metrics file, so the next field harvest reports per-guard counts.Readers checked. The token appears only in cell 6, before
steps:, and no guard id containsapplied,PARTIAL,UNVERIFIED,dirty model,cr: CR-orsteps:. Readers that parse rows:status.shandscript-owners.shread cells 3–5.build-log-read.sh([bug] Stage readout in PROJECT.md goes stale after the last gate pass, and late asks skip change-intake #767) andprogress-report.pyread the date or cells 1–3, plus thecr: CR-pattern.done-drift-check.shandrow-groups.shread applied rows only.doctor.shparsessteps:to the end of the cell.obligation-check.shandci-gates.shmatch only the date.metrics-ledger.shreads the result cell only.ready.shandtoken-burn.shdon't read the BUILD-LOG.Impact
project-bin/exec.sh: one parameter expansion inlog_buildplus a trailing argument on 25 calls.project-bin/metrics-trend.shnow also readsdocs/BUILD-LOG.md(written by exec.sh'slog_buildon every exec). Projects pick both up viasync-project.sh --upgrade-bin. Any open PR that adds alog_build "🚫 …"call must pass a guard id, or the new static check fails.${3:+ · guard=$3}starts no process.test-exec-process-budget.shcounts 132 starts on both master and the branch.Time cost: none per exec · 0 extra process starts (132 → 132) · can force rework: n · offset: n/a
Field evidence: the audit's count from #804. The row shape was checked against the captured
tests/wave2/fixtures/verify-446/BUILD-LOG.captured.md, which the new fixture appends verbatim. The next harvest is the first field run of the new token.Fixture:
tests/wave2/test-buildlog-guard-name.sh [exec.sh]mxcli check, uncommitted model). Each row carriesguard=<id>; the result cell, exit 1 and the 5-cell shape are unchanged.blocked by guard: mxcli-check 1, uncommitted-model 1, unnamed 2.--sincefilters, and a missing BUILD-LOG is reported rather than skipped.Existing fixtures that target the changed files, all green on the branch:
check-scripts,check-portability, the leak guard andcheck-pr-disciplineare clean.CHANGELOG.mdline appended, crediting the audit and the field metrics (contrib/inbox: learnings from a requirements-driven full build #804)Not done, by design: the Studio Pro lock and exec-lock refusals write no BUILD-LOG row today. Adding rows would change behaviour, so they stay uncounted.
Review: pending independent review
Generated by Claude Code