[cmgrote]

Summary

/lens recheck always called the model with verify_max_tokens (8,000). Reasoning tokens count against that limit, so a hard finding can use all of it before the model writes a verdict. The recheck then reports only "no verdict came back", and the author can't give it more room. Rounds already have an opt-in for this (/lens verify-budget); rechecks didn't.

This happened on #4731: two rechecks of one finding, both with evidence, each ended at out=8000 reasoning=8000 and returned no verdict.

Changes

• /lens recheck verify-budget F-… [reason] makes the recheck use verify_max_tokens_opt_in (24,000), the same budget /lens verify-budget gives a round. The word counts as the opt-in only right after recheck; anywhere later it is part of the reason.
• _recheck_call now reports when the call stopped at its output limit, using the existing _out_of_output_budget check.
• When a recheck runs out of budget:
  • the header says re-check: ran out of output budget;
  • the verdict comment names the exact retry command and the raised budget;
  • each kept finding says it was "not re-judged", which is different from "kept".
• If the recheck already used the raised budget, the comment says so and doesn't offer the opt-in again.
• The usage text, the event.py docstring and the .github/lens/config.toml comment are updated.

Testing

• 8 new tests in test_lens.py:
  • parsing the opt-in;
  • a cut-off recheck naming the opt-in;
  • an opt-in recheck sending max_tokens=24000;
  • an opt-in cut-off not offering the opt-in again;
  • a recheck that fails for another reason not mentioning the budget;
  • the CLI passing the opt-in through.
• Red/green: against main's lens code, 7 of the 8 fail. The one that passes covers the unchanged "no verdict for another reason" case.
• The full lens suite passes locally, apart from the tag-signing tests that only run in CI. Pre-commit and the scripts pyright baseline pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AbKqGbjx7uUk6aM6jzSxNF

[cmgrote]

/lens

[atlan-app-fleet]

lens summary · updated after round 2 (re-review · only commits since c48dd16)

✅ Ready to merge — every finding is resolved

Open findings: 🔴 0 critical · 🟠 0 high · 🟡 0 medium · ⚪ 0 low (nit)

lens reads this PR as — A one-off /lens recheck has only the standard 8,000-token verify budget. When reasoning consumes that output allowance before a verdict, the finding remains unresolved without a clear retry path, even though review rounds already support an opt-in raised budget. How: The PR accepts verify-budget only immediately after recheck, passes that choice through the CLI, and selects the same configured standard or opt-in budget used by round verification. It also reuses the existing output-budget detection to distinguish a cutoff from other no-verdict outcomes and explains that findings were not re-judged, with retry guidance only when the raised budget was not already used.

Approach check — sound.

✔️ Resolved (1)

id	severity	where	finding	found	resolved	how
F-e3d515	🟡 medium	.github/scripts/lens/review.py:1200	Output cutoffs with partial tool calls aren't reported	round 1	round 2	its code was removed or rewritten

🕘 Round history (2)

round	run	commits	new	resolved	blocking open	cost	log
1	first review	377cd57..c48dd16	1	0	0	$0.021	run
2	re-review · only commits since c48dd16	c48dd16..c0b10bb	0	1	0	$0.016	run

_{round 2 · $0.037 of $1.00 · 24 model calls · 0 failed requests · 79% prompt cache hits · gpt-6-luna}

[cmgrote]

/lens

[atlan-app-fleet]

✅ lens · round 2 · re-review · only commits since c48dd16 — ready to merge — every finding resolved · $0.04

Open findings: 🔴 0 critical · 🟠 0 high · 🟡 0 medium · ⚪ 0 low (nit)
Resolved this round: 1

Approach check — ✅ sound

• Problem: A one-off /lens recheck has only the standard 8,000-token verify budget. When reasoning consumes that output allowance before a verdict, the finding remains unresolved without a clear retry path, even though review rounds already support an opt-in raised budget.
• How the PR solves it: The PR accepts verify-budget only immediately after recheck, passes that choice through the CLI, and selects the same configured standard or opt-in budget used by round verification. It also reuses the existing output-budget detection to distinguish a cutoff from other no-verdict outcomes and explains that findings were not re-judged, with retry guidance only when the raised budget was not already used.

_{$0.037 of $1.00 · 24 model calls · 0 failed requests}

Full summary · Run log