feat(agents): pin the four inherit toolkit reviewers to opus (−10.6% pr-reviewer spend)

[DJRHails]

Operator decision materialized as a PR — merge or close; it is deliberately not self-merged because it changes which model runs four review agents.

The gantry pr-reviewer cost anatomy (gantry#942, N=109 runs / 18,815 requests, 2026-08-25) measured the §1 review battery at 28.6% of $/review. Within it, $847 of $1,142 bills at fable rates ($10/$50 per MTok) through these four agents' model: inherit, while their siblings code-reviewer and code-simplifier have pinned model: opus ($5/$25 — half price) all along.

This pins the remaining four:

• pr-test-analyzer
• silent-failure-hunter
• comment-analyzer
• type-design-analyzer

Projected saving: 10.6% of pr-reviewer spend ($4.2k/mo at the current ~$40k/mo run rate).

The quality case for parity rather than degradation: code-reviewer — the battery's highest-spend, highest-yield finder — has run on opus the whole time these numbers were gathered, and Opus 5 is the exact model every review already falls back to fleet-wide whenever fable is unavailable (whole review windows have run on it, verdicts accepted). Still a model change on finding-producing agents, so it ships only on your merge. Trivially reversible (four one-line frontmatter edits).

Note these files are the vendored copies consumed everywhere the dotfiles bake (gantry worker images pick it up via the djrhails-dev rebuildProbe on the next image freshen), so the pin applies to interactive sessions too, not only pr-reviewer.

via gantry cost-opt worker

[DJRHails]

Review Summary

Direct single-pass review (4 changed lines of agent frontmatter — under the multi-agent thresholds).

Findings

No P1–P3 findings. The diff does exactly what the description says: pins model: opus on the four model: inherit toolkit reviewers, matching code-reviewer and code-simplifier.

#	Severity	Finding	Resolution
1	P4	The subagents README's re-sync instruction ("copy the updated *.md files") would silently revert all six opus pins on the next upstream refresh, re-introducing the fable-rate billing this PR eliminates.	Fixed: README now records the local modification next to the re-sync step (8f65558).
2	P4	silent-failure-hunter.md's description line is not strict-YAML (unquoted colon-space inside a plain scalar). Pre-existing, untouched by this PR, and Claude Code's lenient frontmatter parser loads it fine — noted only in case a strict-YAML tool ever consumes these files.	Noted, no change.

Verified

• Diff is exactly 4 one-line model: inherit → model: opus edits; all six vendored agents now carry exactly one model: opus line (valid frontmatter value).
• The "siblings pinned all along" claim holds: at the vendoring commit (7857935), code-reviewer/code-simplifier were already opus while the four changed agents were inherit.
• No stale references: nothing else in the repo documents these agents' model, and the review-pr SKILL.md carries no model text to rot.
• gantry#942 (the cited cost anatomy) exists and is open. The fable-vs-opus spend split and projected saving are the issue's own measurements — not re-derived here.
• Quality pipeline: no CI workflows in this repo; the pre-commit hook suite (gitleaks, TruffleHog, glassine checks, symlink/behaviour suites) ran green on the fix commit.

Commit

8f65558 — docs(subagents): guard the opus pins against a plain re-vendor

The model-pin decision itself is untouched and remains yours to merge or close, as the PR intends.

Verdict: approve

via gantry