feat(gh-body-audit): dispatch the cold reader on Sonnet

[ultimatile]

Summary

gh-body-audit step 3 dispatched its cold-reader subagent without naming a model, so the subagent ran on whatever model the session used. The cold reader only lists, from the body alone, the tokens a reader cannot reach and the sentences that point at a text instead of stating it; step 4 then judges every reported hit in main context. Step 3 now names model: "sonnet" in the Agent invocation.

Closes #231

Changes

• skills/gh-body-audit/SKILL.md: step 3 invokes Agent with subagent_type: "general-purpose" and model: "sonnet".
• .claude-plugin/marketplace.json: metadata.version is 2026.10.5.

Verification

• skills/gh-body-audit/test-body-math-scan.sh passes. It covers the math scan only; no automated test reads the skill's prose.
• A subagent dispatched with subagent_type: "general-purpose" and model: "sonnet" ran as claude-sonnet-5-5.
• The step 3 prompt was run once on Sonnet and once on Opus against the bodies of two issues, gh-body-audit: the cold reader runs on whatever model the session uses #231 and gh-post's behavior is described in six files and owned by none #222. On gh-body-audit: the cold reader runs on whatever model the session uses #231 the Opus run reported 13 hits and the Sonnet run reported every one of them. On gh-post's behavior is described in six files and owned by none #222 the Opus run named four spans the Sonnet run did not. Two name things that exist where a reader of the repository would look for them: the ## Math heading of skills/gh-body-conventions/SKILL.md and the published package mdformat. The other two sit inside sentences gh-post's behavior is described in six files and owned by none #222 quotes verbatim from skill files. Step 4 judges a reported token by whether its referent is there, so none of the four calls for an edit to the body.

Notes

• sonnet is an alias and moves with model releases. gh-body-audit: the cold reader runs on whatever model the session uses #231 compared models by re-auditing thirteen bodies and found that Claude Sonnet 5 flagged three hits Claude Haiku 4.5 missed; the two-body check above ran on claude-sonnet-5-5 and does not reproduce that comparison.
• done-check and todo-check also dispatch subagents without naming a model. This change leaves them as they are: no measurement covers their prompts.

[Copilot]

Copilot review overview

🟢 Approval recommended

The narrow changes match the requested behavior and versioning requirements, with no identified blocking issues.

Review effort: Balanced
Findings: None

What changed in this PR

Updates gh-body-audit to run its cold-reader subagent on Sonnet rather than inheriting the session model, addressing #231.

Changes:

• Explicitly selects model: "sonnet" without changing the audit prompt or subsequent evaluation.
• Bumps the plugin version to 2026.10.5.

File	Description
skills/​gh-body-audit/​SKILL.md	Selects Sonnet for the cold-reader dispatch.
.claude-plugin/​marketplace.json	Updates the plugin version.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.