Claude finished @Yeraze's task —— View job

PR Review: chore(agents): default bug-investigator to Opus + add grounding contract

Tasks

• Read changed file
• Review changes and provide feedback

Review

This is a clean, focused improvement to the agent configuration. The changes are well-motivated and well-executed.

Model upgrade (line 5)

Bumping to model: opus is appropriate given the agent's job — it needs strong reasoning to avoid hallucinating file paths and line numbers. The PR description notes meshtastic-expert is already on Opus, so this brings meshmonitor-bug-investigator in line with that precedent.

Grounding rules (lines 14–45)

The 5 rules are specific, actionable, and directly address the failure modes described in the PR summary:

1. Cite only what you opened — closes the "memory/guessing" loophole cleanly.
2. Verify before you report — the ls + grep -n verification step is concrete and easy for the model to follow.
3. Prove absence — requiring the actual empty grep -rn output to substantiate "X doesn't exist" is the right fix for the most common failure mode.
4. Grep first, never bulk-read — good constraint; the ±30-line guidance is specific enough to be actionable.
5. Label confidence — CONFIRMED/HYPOTHESIS tagging makes uncertainty visible to downstream engineers.

Evidence section in output format (lines 275–278)

The inline comment in the output template reinforces the grounding rules at the exact point where the agent would produce citations. This is a nice touch — it makes the rules hard to overlook even if the agent "skims" the system prompt.

Minor observations (no blocking issues)

• The existing HYPOTHESIS / EVIDENCE / ROOT CAUSE block format at lines 128–135 predates the grounding rules and uses the HYPOTHESIS label. The new grounding rule 5 introduces CONFIRMED / HYPOTHESIS tagging on individual Evidence items rather than top-level sections. These aren't in conflict, but a future cleanup could unify the two patterns for clarity.
• Rule 4 says "read only the targeted ±30 lines" but the tools frontmatter (line 4) gives the agent the read tool with no explicit constraint. The rule is advisory text; whether the model reliably respects the 30-line bound depends on model behavior. Low-risk for now, worth monitoring.
• No test coverage for agent config files is expected or needed here; this is purely a prompt/config change.

Overall: Straightforward, well-scoped hardening. No blocking concerns. ✓