05 Fix

Per sub-agent (one finding per invocation):

• thread_id, file, line from step 3.
• The finding summary and Copilot’s suggested fix (if any).
• The triage rationale from step 4 — the agent already decided this is a fix; this sub-agent only implements it.

Return contract

Where status ∈ complete | partial | blocked and summary is a one-line description of the change.

Procedure

1. Discover repo conventions for the area being edited — first. Before writing any code, read:
   • .github/instructions/*.md whose applyTo glob matches the file’s path,
   • .github/skills/,
   • AGENTS.md,
   • CONTRIBUTING.md,
   • neighbor-file patterns in the same directory and recent commits touching similar files.
2. Apply the fix in line with those conventions.
3. Return files_touched + a one-line summary + status.

Gotchas

• Max 5 concurrent fix sub-agents. The cap prevents fix-fanout chaos; the parent merges results and reconciles file conflicts between waves before step 6.
• Never invent a generic answer that contradicts repo practice. That’s the “elephant in school” anti-pattern — a Copilot suggestion in isolation looks right but breaks the project’s lint, format, spell-check, license-header, or framework conventions. The discovery step is mandatory, not optional.
• One finding per sub-agent. Fixes that need to touch the same file get serialized by the parent between waves — don’t merge multiple findings into one sub-agent invocation.
• Project policy beats Copilot suggestion. If discovery surfaces a documented convention (spell-check allowlist mechanism, lint suppression style, etc.) that contradicts the suggested fix, follow the convention and reflect that in the reply body drafted in step 8 (see 04-triage.md).
• Push back with written rationale if implementing the fix would over-engineer the design for a hypothetical edge case — flip the triage to decline and return status: blocked with the rationale so step 8 drafts the right reply.