fix-harness-after-run

name	fix_harness_after_run
description	Repair harness-side crash causes after run-time triage while preserving target behavior and repro semantics.
compatibility	opencode
metadata	{"stage":"fix-harness","owner":"tianheng"}

What this skill does

Apply targeted harness/build-glue fixes for crashes classified as harness bugs in the post-run fix-harness stage.

When to use this skill

Use this skill only in fix-harness when crash triage/analysis indicates harness_bug.

Required inputs

• crash_info.md
• crash_analysis.md indicating harness-side fault
• crash_triage.json
• crash artifact metadata from coordinator context

Required outputs

• Minimal harness/build-glue patch under fuzz/

Workflow

1. Read crash diagnostics and identify harness misuse/precondition violations.
2. Apply a minimal patch in fuzz/ files.
3. Keep crash-path semantics and target behavior intact.

Constraints

• Focus on fuzz/ harness/build glue files.
• No unrelated refactor.
• Avoid upstream project source edits unless strictly required.
• Must produce textual code changes; pure no-op is invalid.
• Do not bypass acceptance by tampering with fuzz/repo_understanding.json.
• Use explicit actions when diagnostics include concrete paths: Read and fix <path>[:line].
• Prefer public/stable APIs; replace internal/private symbols first.
• If no public alternative exists, add api_surface_exception with non-empty reason and evidence.

Command policy

• Allowed: read-only commands only.
• Forbidden: build/execute commands.

Acceptance checklist

• Patch addresses harness-side root cause directly.
• Changes are minimal and evidence-driven.
• No semantic bypass and no doc-only/no-op patch.

Done contract

• Write one key modified path as the sole text of ./done (a relative path string like fuzz/foo.c; do not copy file contents).