trace-analysis

Trace Analysis

Overview

Trace analysis turns historical learnings into actionable signal for skill improvement. Uses scripts/trace-analyzer.sh to classify failure patterns from .agent-harness/learnings.jsonl.

Core principle: Data over intuition. Run the analyzer before deciding what to improve.

When to Use

• During retrospective (pair with retrospective skill)
• When reviewing .agent-harness/learnings.jsonl for trends
• When noticing recurring failure modes across sessions
• When planning which skills to improve next

How to Run

bash "${CLAUDE_PLUGIN_ROOT}/scripts/trace-analyzer.sh" [project_root]

Pattern Reference

The analyzer classifies learnings by keywords in key and insight fields:

Pattern	Meaning	Address With
loop	Re-editing same file without converging	loop-detection
drift	Wandering away from task goal	writing-plans with explicit task boundaries
oversight	Missing steps or requirements	verification-before-completion checklists
scope-creep	Adding work beyond what was asked	sprint-contract to lock scope
verification-gap	Claiming success without evidence	computational-sensors for deterministic verification
other	Unmatched failure mode	Investigate manually

Interpreting Output

• Top pattern = highest priority. Analyzer sorts by count.
• If loop appears, that's urgent. Doom loops waste the most time.
• Multiple patterns at similar counts suggest systemic issues. Planning skills (writing-plans + sprint-contract) may prevent several at once.

TDD Baseline (RED Phase)

Pressure scenario: agent conducting retrospective, learns failure patterns exist but ignores data, "improves" based on gut feeling instead of analyzer output.

Baseline failures: agent skips trace-analyzer.sh, rationalizes "I know the patterns", does not cross-reference with skill improvement actions.

Recommendations

Match analyzer recommendations to skill invocations:

1. loop-detection → agent-harness:loop-detection — ensure tracking is active
2. computational-sensors → agent-harness:computational-sensors — configure sensors
3. sprint-contract → agent-harness:sprint-contract — lock scope before next implementation
4. verification checklists → agent-harness:verification-before-completion
5. writing-plans → agent-harness:writing-plans — plan before next multi-step task

Common Mistakes

Mistake	Fix
Treating one learning as a pattern	Need 3+ occurrences to confirm a real pattern
Skipping analysis because learnings are sparse	Even sparse data shows direction
Acting on every pattern at once	Address top pattern first, then re-run

Rationalization Table

Excuse	Reality
"I know my failure modes"	Feelings are not data. Run the analyzer
"The patterns are obvious"	Obvious patterns would have been fixed already