ci: switch wheels-bot triage from Sonnet to Opus (#2589)

* ci: switch wheels-bot triage from Sonnet to Opus

Triage's job is now harder than pure classification — it reads the
issue body, optionally greps the codebase, and self-rates confidence
that downstream Opus stages (propose-fix, research, write-docs) auto-
fire on. The auto-downgrade rules force `medium` for sensitive
surfaces regardless of model, so the gate's guard-rails stay intact;
upgrading the model just gives the bot a better chance of resolving
genuine uncertainty (vs. hedging to `medium` when a deeper read of
the code would have produced a clean `high`).

Concrete example from #2582 today: triage rated `medium` because the
reporter listed `/env` as working even though it routes through the
same `consoleExec` function as the failing commands. An Opus pass on
that issue would have either read `consoleExec` to resolve the
inconsistency (likely `high`) or confirmed it as a genuine scope-
initialization nuance (justified `medium` with sharper rationale).

Tools and turn budget unchanged — same read-only allowlist (`gh`,
read-only `git`, `Read`/`Grep`/`Glob`), same 200-turn ceiling. Cost
delta is roughly 5× per triage; volume is modest (~3-4 issues/day).

Also synced the model column in CLAUDE.md's bot stages table.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs: address Reviewer A/B consensus findings (round 1)

- docs/contributing/wheels-bot.md:260 — replace stale "only stage that
  runs Opus on a non-coding task" claim with advisor+triage framing
  (A's main finding; B confirmed)
- docs/contributing/wheels-bot.md:224 — clarify "analytical Sonnet
  stages above" by naming Reviewer A and B explicitly (B's missed-issues
  finding)

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>

Author: 3 people

.github/workflows/bot-triage.yml

@@ -71,6 +71,6 @@ jobs:
           prompt: |
             /triage-issue ${{ github.event.issue.number || inputs.issue_number }}
           claude_args: |
-            --model claude-sonnet-4-6
+            --model claude-opus-4-7
             --max-turns 200
             --allowedTools "Bash(gh:*),Bash(git status),Bash(git log:*),Bash(git diff:*),Bash(git show:*),Bash(git grep:*),Read,Grep,Glob"

CLAUDE.md

@@ -1057,7 +1057,7 @@ Workflow orchestration (multi-step planning, feature development) is not a frame
 
 | Stage | Trigger | Model | Output |
 |---|---|---|---|
-| Triage | issue opened/reopened | Sonnet | Comment classifying as `bug` / `framework-design` / `other` (+ confidence on `bug` path). |
+| Triage | issue opened/reopened | Opus | Comment classifying as `bug` / `framework-design` / `other` (+ confidence on `bug` path). Reads code with the allowlisted tools to resolve uncertainty before rating. |
 | Research | bot triage emits `framework-design` marker | Opus | Comment comparing Rails / Laravel / Django / Phoenix / Spring Boot / +1 and recommending a Wheels-idiomatic path (+ confidence). |
 | Propose Fix | bot triage emits `triage-confidence:high` OR research emits `research-confidence:high` (or `workflow_dispatch`) | Opus | TDD-mandatory draft PR on branch `fix/bot-<issue>-<slug>`. Spec-then-implementation, both required by `bot-tdd-gate.yml`. |
 | Reviewer A | PR opened / synchronized / ready_for_review | Sonnet | Single PR review with line comments, verdict, and `wheels-bot:review-a:<pr>:<sha>` marker. |

docs/contributing/wheels-bot.md

@@ -221,7 +221,8 @@ loop on the updated PR state.
 
 This is a *coding* stage — Opus model, 60-minute timeout, broad
 allowlist with the test runner (mirrors propose-fix's setup). Different
-from the analytical Sonnet stages above.
+from Reviewer A and Reviewer B, which remain Sonnet-based analytical
+stages.
 
 The bot:
 
@@ -257,11 +258,11 @@ round 11). The advisor reads the full A↔B exchange, the disputed
 code, and the canonical references (`CLAUDE.md`, `.ai/wheels/`),
 then issues a tie-breaking verdict.
 
-**This is the only stage that runs Opus on a non-coding task.** The
-reasoning depth is justified because the advisor's verdict overrides
-the analytical reviewers' deadlock — it must be right. Address-
-review uses Opus because it modifies code; advisor uses Opus because
-it adjudicates code-related disputes Sonnet couldn't resolve.
+**The advisor and triage are the two stages that run Opus on
+non-coding tasks.** Triage justifies the cost by reading code to
+resolve uncertainty before rating confidence; advisor justifies it
+because its verdict overrides a deadlocked A↔B exchange — it must be
+right. Address-review and propose-fix run Opus for code edits.
 
 The advisor: