Context Harness Debugger

Context Harness Debugger

Use this skill to diagnose agent failures as harness, context, tool, or evaluation problems before blaming the model.

Failure Triage

1. Capture the symptom: wrong output, wrong tool, loop, stale assumption, security issue, high cost, or brittle self-repair.
2. Map it to a likely harness cause:
   • Missing or vague instructions
   • No skill or wrong skill trigger
   • Too much static context causing context rot
   • Missing tool, oversized tool list, or bad tool schema
   • No deterministic guardrail/hook
   • Weak sandbox or authority boundary
   • No observability trace for the trajectory
   • No eval case covering the failure
3. Decide the repair surface:
   • Static project rule: small, universal behavior
   • Skill: reusable procedural workflow
   • Reference file: large conditional knowledge
   • Script/hook: deterministic enforcement
   • Tool/MCP change: external capability or schema issue
   • Eval: failure should be caught next time
4. Prefer reducing active context over adding more instruction text.
5. After the fix, add a regression prompt or test case that would have caught the failure.

References

• Read references/day1_vibe_coding_intro.txt for harness anatomy, context engineering, the 80% problem, and the factory model.
• Read references/day3_agent_skills.txt for context rot, progressive disclosure, skill composition, and context debt.
• Read references/day4_security_evaluation.txt for observability, trajectory quality, intent drift, and circuit breakers.

Anti-Patterns

• Do not fix every failure by adding "ALWAYS" rules.
• Do not stuff large documents into static prompts.
• Do not let the LLM context window become the database or message bus.
• Do not accept a correct final answer if the tool trajectory was unsafe for an action-allowed workflow.

Evaluation Prompts

• Positive: "My agent keeps looping and calling the wrong tool." Expected: diagnose context, tool list, instructions, guardrails, and eval gaps.
• Positive edge: "The answer is right but it used unsafe tool calls." Expected: trajectory/harness repair, not model praise.
• Negative: "The app has a runtime bug in this stack trace." Expected: normal debugging unless the failure is agent-harness behavior.