fix(trace-repair): stop rejecting a repair for how it labels itself

[drewstone]

The action budget rejected an intervention whose declared kind disagreed with its measured payload. The scaffold runs the action identically either way, so the rule scored an arm on how it described a repair rather than on the repair itself.

It also made the analyst contract unsatisfiable. The prompt fixes the declared kind to shell and recommends authoring a whole file with a heredoc, and the classifier reads any heredoc as edit, so an answer that followed the prompt was rejected before a container opened.

Measured on the milestone-2 repair corpus

arm	findings	rejected on the label alone
prime	10	7
bare-framing	1	1
oracle-fix control	43	0
inert-probe control	43	0

The controls could never trip the rule: the admit runner derives their declared kind from classifyActionPayload, the same classifier the check measures with. The gate therefore applied to the analyst arms alone, which is what makes it a comparison defect rather than a shared cost.

Re-graded on the same corpus, the answer that was rejected in 1 ms now reaches execution and returns a real outcome (outcome=measured).

Change

declared joins payload in BudgetMeasurement, and the mismatch stops being a BudgetViolation. Everything the budget still rejects is an executability bound: byte cap, one top-level statement, one authored file, no-op, submit-instead-of-repair.

Checks

• npx vitest run tests/trace-repair/ — 128 passed (11 files)
• npx tsc --noEmit — one pre-existing error in src/agent-profile.test.ts, present on clean main and unrelated to this change
• merges cleanly into origin/main (git merge-tree --write-tree)

[tangletools]

✅ Auto-approved drewstone PR — cd4d62ef

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-10T11:07:45Z}

[tangletools]

🟢 Value Audit — sound

Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	94.6s (2 bridge agents)
Total	94.6s

💰 Value — sound

Removes a budget gate that rejected an intervention whose self-declared kind disagreed with its measured payload — a property that never affected execution, so the gate taxed analyst arms (not controls) on labeling rather than on the repair.

• What it does: Deletes the payload-kind-mismatch BudgetViolation from checkInterventionBudget and stops treating a declared-vs-measured kind disagreement as a rejection. The declared kind is now recorded (new BudgetMeasurement.declared field) alongside the measured payload for reporting, but no longer gates admission. Every remaining rejection (bytes, statements, heredocs, no-op, submit) is an executabil
• Goals it achieves: (1) Stop rejecting a repair for a label that the scaffold ignores at run time — wrapActionForExec(action, cwd) in grade.ts:296/301/349/370 runs the text identically either way, and response.intervention.kind is read exactly once (grade.ts:141), only to feed this check. (2) Remove a comparison defect: the control arms derive their declared kind from classifyActionPayload (the same classifier
• Assessment: A clean, correct fix in the grain of the codebase. The file's own header states the budget bounds 'what the scaffold could have done' (one statement, one file, a byte cap); the removed gate was the sole rejection that was not an executability bound. The degenerate-strategies.ts catalog (src/trace-repair/degenerate-strategies.ts:76-83) never enumerated label-mismatch as a defended strategy, confi
• Better / existing approach: none — this is the right approach. Searched for an existing equivalent: checkInterventionBudget is the single admissibility site and this edits it in place; no parallel classifier or duplicate kind logic exists to consolidate. The control-arm classifier reuse (scripts/tb-repair-milestone1.ts:516 calling classifyActionPayload) is already shared. The only alternative would be to drop the `declar
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 2
• Bridge warning: opencode/kimi-for-coding/k2p7: opencode: opencode error

🎯 Usefulness — sound

Removes the one budget rejection that scored an arm on its self-label rather than its executable payload, fixing a comparison defect that only the analyst arms could trip; every remaining rejection is a true executability bound.

• Integration: Wired through the live grader path. checkInterventionBudget(action, response.intervention.kind, budget) is called at src/trace-repair/grade.ts:139-143, so the declared kind already flows from the analyst response into the check; the measurement flows into the rejection record at src/trace-repair/grade.ts:151 and src/trace-repair/funnel.ts:37. Removing the payload-kind-mismatch variant is type-
• Fit with existing patterns: Fits the grain and reconciles the code with its own stated design. The module's top doc (src/trace-repair/action-budget.ts:5-8) says the budget bounds what the scaffold can execute, and src/trace-repair/degenerate-strategies.ts:76-83 documents the buy-a-bigger-action defeat purely as a statements/heredocs/bytes bound. The label-mismatch rule was the sole non-executability rejection, so dropping
• Real-world viability: Holds on all paths. It is a pure-function relaxation: one rejection branch removed, one always-populated recorded field added (required declaredKind param, no null path). An action that mislabels itself now reaches the reproduction gate — correct, because the scaffold runs the text identically either way, which is the load-bearing fact the PR's measured corpus confirms (controls 0/43 rejected on
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

---

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260810T113349Z}

[tangletools]

✅ No Blockers — cd4d62ef

Review health 100/100 · Reviewer score 89/100 · Confidence 70/100 · 3 findings (3 low)

	opencode GLM 5.2	opencode DeepSeek v4 Pro	opencode DeepSeek v4 Flash	aggregate
Readiness	92	95	89	89
Confidence	70	70	70	70
Correctness	92	95	89	89
Security	92	95	89	89
Testing	92	95	89	89
Architecture	92	95	89	89

Reviewer score is advisory once the run is complete and the verdict has no blockers.

Full multi-shot audit completed 2/2 planned shots over 2 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 2 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 2 changed files. Global verifier still owns final merge decision.

🟡 LOW declared kind is recorded but never persisted for admitted rows — src/trace-repair/action-budget.ts

The doc comment says 'A reader who wants the mismatch counts declared against payload', and the commit message says the mismatch is 'measured so a report can count it'. But the measurement only survives into the record path for rejected rows: grade.ts:151 puts check.measurement into RepairRejection.measurement, and for the admissible branch grade.ts discards the check result entirely (produceGrade proceeds without storing check.measurement in any of the not-reproduced/did-not-execute/measured outcomes). So for every admitted row — the rows whose label-vs-payload mismatch the change exists to observe — declared and payload are absent from the recorded grade, and the mismatch count the change promises cannot actually be computed from artifacts. Either surface the measurement on the measured/

🟡 LOW Admission test lives under a 'rejects' describe block — tests/trace-repair/action-budget.test.ts

The describe block at line 76 is framed around rejection, but the renamed test at line 103 now asserts admission. This is pre-existing (the 'admits one command inside the cap' test at line 79 already broke the framing), so not a regression — but the block title now misleads a reader scanning for what is rejected vs admitted. Cosmetic; no behavior impact. Optional: split admits/re

🟡 LOW Rejection-path measurement does not pin declared — tests/trace-repair/action-budget.test.ts

The new test (line 103-111) pins measurement.declared on the admitted path, but the rejected-path test (lines 130-134) asserts only statements/bytes. Because measurement is built once before all reject branches (action-budget.ts:413-425), declared is present there too, so this is not a correctness risk — but pinning declared/payload on a rejected action would lock the invariant that reports can count the mismatch even for rejected repairs. Optional hardening, not a blocker.

---

_{tangletools · 2026-08-10T11:36:11Z · trace}

[tangletools]

✅ Approved — 3 non-blocking findings — cd4d62ef

Full multi-shot audit completed 2/2 planned shots over 2 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 2 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 2 changed files. Global verifier still owns final merge decision.

Full immutable report for this review: trace

Summary comment for this run: full summary

---

_{tangletools · 2026-08-10T11:36:11Z · immutable trace}