Expose timeout-minutes as an input on the two agent workflows

[d-morrison]

Closes #879.

Problem

A caller cannot set timeout-minutes on a job that uses uses:, so the agent job's limit is unreachable from the consumer repo. Both workflows hard-coded 60:

• claude.yml:439
• claude-code-review.yml:533

That default is reasonable. It is only a problem for a consumer with a stricter policy, which had no way to express one.

Change

An optional timeout-minutes input on both, defaulting to 60 so no existing caller changes behavior.

This extends an established pattern rather than inventing one — r-cmd-check.yml, altdoc-multiversion-docs.yml and check-code-similarity.yml already accept the same input, and the description follows r-cmd-check.yml's "hang ceiling, not a budget" framing.

Scoped to the agent job in each workflow (claude and claude-review — the two that carried the 60). The short helper jobs keep their fixed limits:

Workflow	Job	Before	After
claude.yml	mention-filter	5	5
claude.yml	claude	60	${{ inputs.timeout-minutes }}
claude-code-review.yml	preempt-previous / gather-context	5 / 10	unchanged
claude-code-review.yml	claude-review	60	${{ inputs.timeout-minutes }}
claude-code-review.yml	post-review / require-*	15 / 10 / 10	unchanged

A consumer capping the long agent run has no reason to shorten a five-minute mention filter, and parameterizing those would invite someone to set a value that starves them.

Motivating case

Morrison-Lab/qbt#28 requires every job at or under 50 minutes. qbt#58 brought everything that repo controls under the cap — and these two were the only jobs left above it, with no caller-side way to reach them. So the policy held everywhere except the two longest-running jobs, which is the opposite of useful.

Docs

Updated alongside the code so the input lists don't drift:

• README rows for both workflows
• ## Inputs tables in website/reference/claude.qmd and claude-code-review.qmd
• changelog fragment per changelog.d/README.md (.added category)

Verification

• Both workflows parse; the input resolves onto the intended job in each, helper jobs untouched (checked by parsing, not grepping)
• check-non-standard-chars clean across 363 files
• check-new-line-breaks and check_list_item_splices clean

One note on that last pair: the first run warned my changelog fragment was untracked and therefore not examined. I staged it and re-ran so it was actually checked — otherwise the "clean" result would have covered every file except the one I added.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lx1hhHu58mXq9DfJsAxkHn

---

Generated by Claude Code

[d-morrison]

All seven findings addressed in 8067210. Every one was correct.

The finding I'd argue is the most valuable is #5, the README "Job timeouts" prose — it isn't in CLAUDE.md's three-site rule at all. It's a sentence naming three workflows that expose their timeout as an input, and this PR makes it five. Nothing links the input to that paragraph, so it would have gone stale silently and the next person would have read an enumeration that quietly excluded the two workflows they were looking at.

Site	Fix
website/workflows.qmd:42-43	Key-inputs cells extended (the table that drifts independently of README's)
website/reference/claude.qmd, claude-code-review.qmd	commented timeout-minutes: 50 in each Example block, matching r-cmd-check.qmd:172
examples/claude.yml, claude-code-review.yml	same line in the commented with: block, matching examples/r-cmd-check.yml:42
README.md "Job timeouts"	enumeration now lists all five

On method, since this is the second time I've claimed a sweep was complete when it wasn't. I didn't work from the list above — I grepped for extra-secret-names, the previous last input, which returns every file enumerating these workflows' inputs. That makes the sweep complete by construction rather than by inspection, and it's what I should have done before the first push instead of updating the three surfaces I happened to think of.

The PR description claimed "docs updated alongside the code so the input lists don't drift." That was the claim, and it was wrong in four places. Corrected, and the description now reflects what actually shipped.

Verified: all four YAML files parse; check-non-standard-chars clean across 363 files; check-new-line-breaks and check_list_item_splices clean with the changes staged.

---

Generated by Claude Code

[d-morrison]

Took the nit rather than deferring it, in 28aef98.

Normally an optional finding wouldn't start a push. This one is different because it's a regression I introduced, not a pre-existing nit — checked rather than assumed:

origin/main   README.md:762  76 chars
8067210       README.md:762  96 chars   <- my reflow merged "additionally expose" onto it
28aef98       block max      74 chars

The previous commit reworded the enumeration from three workflows to five and, in doing so, made that line worse than the one this branch started from — inside a commit whose entire purpose was documentation hygiene. The "don't churn on optional findings" rule is there to protect against noise on pre-existing nits; it shouldn't be used to lock in something I broke two commits ago.

Rewrapped at a clause boundary, no wording changed, block max now 74 characters. check-non-standard-chars and check-new-line-breaks both clean.

---

Generated by Claude Code

[github-actions]

Claude finished review — View run

---

Review: PR #881 — Expose timeout-minutes as an input on claude.yml and claude-code-review.yml

What changed since the last review round

Per the computed range (8067210e..89e84dc7), the only change is a pure re-wrap of the "Job timeouts" prose in README.md:

-additionally expose their timeouts as a `workflow_call` input, which is the pattern to follow if
-a consumer ever needs to raise one.
+additionally expose their timeouts as a `workflow_call` input,
+which is the pattern to follow if a consumer ever needs to raise one.

This is exactly the non-blocking semantic-line-break nit flagged in the prior round (README.md:762 being ~96 chars). The rendered Markdown text is identical before and after — only the source wrap point moved — so per this review's own instructions ("don't flag cosmetic source-only formatting that renders identically"), there is nothing further to raise here.

No other files changed in this round. The two workflow files (claude.yml, claude-code-review.yml) and the four other doc/example sites (website/workflows.qmd, both website/reference/*.qmd Example blocks, both examples/*.yml stubs) are unchanged from commit 8067210, which the prior round already confirmed fully addressed the doc-sync gaps from round 1.

Confirmation of prior verdict

The second review round (posted 2026-09-15T02:04:06Z, #881 (comment) — run 34919636257) confirmed all four CLAUDE.md doc-sync gaps from round 1 were closed and found no remaining blocking issues, only the now-resolved line-length nit. That verdict was effectively "Ready for merge" (the residual nit was explicitly non-blocking), and this round's sole change is that exact nit being addressed. Nothing in this round warrants revisiting that conclusion.

I re-verified independently:

• README.md's "Job timeouts" enumeration correctly lists all five workflows (altdoc-multiversion-docs.yml, r-cmd-check.yml, check-code-similarity.yml, claude.yml, claude-code-review.yml).
• website/workflows.qmd, both reference .qmd Example blocks, and both examples/*.yml stubs all carry timeout-minutes per the diff file (.claude-review-pr.diff).
• The workflow YAML changes (claude.yml, claude-code-review.yml) are unchanged from the previously-approved commit — no new functional review is needed.

Verdict

Ready for merge

Structured Review Data (JSON)

{
  "schema_version": "1.1",
  "reviewer": "claude",
  "commit_sha": "89e84dc7c0842d17fb2a8fdab63465772c75c4cf",
  "verdict": "CLEAN",
  "findings": [],
  "detailed_assessment": "The only change since the prior round is a cosmetic re-wrap of README.md's Job timeouts prose to satisfy the semantic-line-break guideline; rendered Markdown is unchanged and no other file was touched.",
  "holistic_assessment": "All four doc-sync sites CLAUDE.md requires for a new workflow_call input were already verified complete in the prior round, and the workflow YAML wiring for timeout-minutes on claude.yml and claude-code-review.yml remains unchanged and correct."
}

Reviewed commit: 28aef98

[github-actions]

💰 Cost: $0.6926 (review) — run

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.

_{2026-09-15 02:19 UTC}