fix(review-runs): census the prefixes Step 2 prices, not just tend-*

[tend-agent]

Step 1 of review-runs enumerates the fleet with a hard-coded single prefix, while Step 2 documents the opposite — that non-tend-* workflows using the tend action are in scope and their prefixes come from the repo's running-tend skill. The two steps therefore disagree about what the fleet is, and Step 1 wins: a workflow outside the tend- prefix is never classified for duration, never near-timeout-checked, and never reaches Step 3's log analysis. Nothing in the output marks the omission.

Measured on this repo

Over a 24 h window (SINCE=2026-08-06T08:42:07Z), review-reviewers — which runs the tend composite action but is not named tend-*, and which tend's own running-tend overlay already lists as an extra prefix for Step 2 — has 21 completed runs that Step 1 never sees:

$ gh api repos/max-sixty/tend/actions/workflows --jq '.workflows[] | select(.name | startswith("tend-")) | .name' | grep -c '^review-reviewers$'
0
$ gh api "repos/max-sixty/tend/actions/workflows/250009605/runs?created=>=$SINCE&status=completed&per_page=1" --jq '.total_count'
21

Change

Replace the inline startswith("tend-") with a PREFIXES array defaulting to ("tend-"), matched as an anchored alternation. Step 2's sentence now points at the same list rather than describing a parallel one, so a single repo-level source drives both steps.

Behaviour is unchanged for an adopter with no extra prefixes — the default array reproduces the old filter exactly. Running the edited Step 1 block verbatim:

# as written, default PREFIXES=("tend-")
tend-review 30, tend-mention 30, tend-notifications 26, tend-ci-fix 9, tend-triage 6, tend-nightly 1

# with tend's running-tend prefix list, PREFIXES=("tend-" "review-")
tend-review 30, tend-mention 30, tend-notifications 26, review-reviewers 20, tend-ci-fix 9, tend-triage 6, tend-nightly 1

The 30s in that output are #886's separate bug (the unpaginated endpoint capping at a page), still live on main — this change doesn't address it and doesn't depend on it.

Scope and conflict note

This is part 2 of #888. Part 1 of that issue — token-report.sh capping the same window at --limit 100 — is already fixed by #887, so nothing here touches that script.

#886 edits the same two gh api lines in this block to add --paginate. The changes are independent in intent but overlap textually, so whichever lands second needs a trivial rebase; the two edits compose (a PREFIXES-driven filter on a paginated fetch).

Not included: an explicit prefix-list line in tend's own .claude/skills/running-tend/SKILL.md. Its "Usage analysis" section already names review- as the extra prefix, so this repo's Step 1 resolves correctly today; adding a dedicated line there is a separate overlay concern.

Gate assessment

• Structural. The filter is fixed in the recipe text, so it excludes identically on every run. No decision point.
• Evidence: High. Exclusion reproduced directly against the API, and the fix verified by executing the edited block verbatim at both prefix lists.
• Change type: targeted fix — one code block plus one cross-reference sentence, no new sections.

---

Refs #888

Follow-up commits

3d4fcda — review on this PR found the same disagreement one paragraph later: the near-timeout instruction resolved a workflow's timeout-minutes by globbing .github/workflows/tend-*.yaml, which doesn't match the very workflow the widened census now admits. It now reads the workflow's own file. f823494 reworded the PREFIXES comment from rationale into an instruction.

[tend-agent]

Step 1 now admits non-tend-* workflows, but the near-timeout check two paragraphs down still resolves the cap by globbing a tend-* filename — the sentence beginning "To determine the timeout cap for a workflow, read timeout-minutes from the workflow YAML file". On this repo that glob doesn't match review-reviewers.yaml, the exact workflow this change exists to admit, so such a run now enters the census and gets duration-classified but the reader following the recipe finds no file to read a cap from. Same Step-1-disagrees-with-the-next-step shape, on the same axis, one paragraph later. Dropping the glob to the directory would close it — happy to push that here rather than leave it for a follow-up.

Otherwise the mechanism checks out: PREFIX_RE builds ^(tend-) and ^(tend-|review-) correctly, and the two prefix lists return the census the description claims against the live API.

[tend-agent]

Both findings from the previous review are addressed. 3d4fcda drops the tend-* glob from the near-timeout paragraph, and the instruction now resolves against the workflow's own file — on this repo that reaches .github/workflows/review-reviewers.yaml, the file the earlier glob missed. No workflow here sets timeout-minutes, so the 360-minute-default sentence following it still holds. f823494 applied the inline suggestion verbatim; thread resolved.

Nothing further from me — self-authored, so no approval to give.