[polewskm]

Replaces the unused budgets.run with a per-review cost ceiling, runCeiling(parts) = 8 + 4 × parts ($12 for one part, $32 for six), overridable with CODE_REVIEW_MAX_RUN_USD. review() tracks the cost of every call so far and skips each reviewer or rule-check call queued after the total reaches the ceiling. Those reviewers show skipped: run budget in the check run and skipped in the summary comment. Synthesis still runs on what was found, and the summary warns, also as a ::warning:: in the log. Unit tests stub the agent's costs. The README documents the variable.

Decisions:

• CODE_REVIEW_MAX_RUN_USD must be a plain positive decimal (12, 2.5). Any other value, including 0, a negative number or 1e3, warns and keeps the default.
• The ceiling is checked when each call starts. Calls already running finish, so the total can go over the ceiling by up to the per-call caps of the calls in flight.
• A review the ceiling cut short never approves: an approve verdict becomes comment.
• A reviewer with some calls run and some skipped stays ran, with N of M calls skipped: run budget.
• review() and selectContext() take an optional agent function, which defaults to runAgent, so tests can stub the calls.

Closes #50

[ncode-review]

🤖 NCode Review at c6424e0: 7 new findings on the changed lines.

[ncode-review]

🤖 Review complete

Verdict: 💬 Comment — This PR replaces the unused budgets.run with a per-run cost ceiling that scales with diff parts and can be overridden with CODE_REVIEW_MAX_RUN_USD. Once spend reaches the ceiling, reviewer and rule-check calls that haven't started are skipped. Synthesis still runs, the summary gets a warning, and the verdict is downgraded from approve. The design holds together: it checks spend before each queued call and adds an injectable agent seam for tests, and the tests cover the main paths well. The remaining findings are small hygiene and style points.

8 reviewers run — 4 findings (run details)

Reviewer	Model	Findings	Top
synthesis	opus-5-5	4	🟡
architecture	sonnet-5-5	0	🟢
completeness	sonnet-5-5	0	🟢
conventions	sonnet-5-5	0	🟢
correctness	sonnet-5-5	0	🟢
hygiene	haiku-5-5	1	🟡
intent	haiku-5-5	0	🟢
rule-check	sonnet-5-5	2	🟡
tests	haiku-5-5	1	🟡

Skipped: api-compat, data-model, infrastructure, performance, security

Findings

• 🟡 F1 test/run-ceiling.test.ts:19: Tests don't pin the decimal edge cases of the override regex
• 🟡 F2 src/review.ts:285: review() gains a fifth positional parameter
• 🟡 F3 src/config.ts:63: Unnamed literal 40 when truncating the echoed override
• 🟡 F4 test/run-ceiling.test.ts:230: Test file has no trailing newline

_{Trial: findings are posted as review threads and in the check run. $0.52 on Amazon Bedrock, 60s, at 773338f. (how to interact)}

[polewskm]

/review

[ncode-review]

🤖 NCode Review at 686620e: 4 new findings on the changed lines.
1 more was already posted on this pull request.

[polewskm]

Autopilot: answering #61 (comment)

• F1 (context/selection spend) and F2 (spend equal to the ceiling): fixed in d5bf163, see Enforce a per-run cost ceiling as a runaway guard #61 (comment) and Enforce a per-run cost ceiling as a runaway guard #61 (comment).
• F3 (blank Model cell): fixed in 773338f, see Enforce a per-run cost ceiling as a runaway guard #61 (comment).
• F4 (is Enforce a per-run cost ceiling as a runaway guard #50 the right issue): a repeat, answered in round 1: Enforce a per-run cost ceiling as a runaway guard #61 (comment) (Enforce a per-run cost ceiling as a runaway guard #50 is this PR's issue and the description has Closes #50).
• F5 (the 40 in value.slice(0, 40)): declined as a nit, see Enforce a per-run cost ceiling as a runaway guard #61 (comment).

[polewskm]

Autopilot: Round 2 answered. No blocker or should-fix is left open; the remaining threads are nits, answered above. Recommend merging once you've resolved them.

[polewskm]

/review

[ncode-review]

🤖 NCode Review at 773338f: 2 new findings on the changed lines.
2 more were already posted on this pull request.

[polewskm]

Autopilot: merged origin/main (79c2447), which brought the opt-in verification stage from #72 (#41).

• src/config.ts: main still had budgets.run: 4; this PR's removal of it stands, next to main's new budgets.verification and maxTurns.verification.
• src/review.ts: the ceiling gates only reviewer and rule-check calls, as before. Verification, when CODE_REVIEW_VERIFICATION=on, runs after synthesis even when the ceiling was reached. It gets the same injectable agent as the other stages. The ceiling's warning and its approve-to-comment rule apply after verification. The README's ceiling paragraph now says verification still runs.
• test/run-ceiling.test.ts: the test charters gained main's new verificationPrompt (f8877ba).

If verification should also stop at the ceiling, that's a follow-up.