Improve Harness

Analyze capability-gap failure patterns recorded in episodic memory and generate structured improvement reports. Supports creating OpenSpec proposal stubs from high-priority findings.

Arguments

$ARGUMENTS - Optional flags:

• --time-window <days> (default: 30)
• --create-proposal (create an OpenSpec proposal stub from the top finding)
• --output <path> (write report to file; default: stdout)
• --candidate-work-output <path> (explicit candidate sidecar; file reports default to adjacent improve-harness-candidate-work.json)

How It Works

1. Queries the coordinator episodic memory for entries with capability_gap:* tags
2. Groups findings by capability_gap value
3. Ranks by (frequency x severity_weight) where severity weights: critical=4, high=3, medium=2, low=1
4. Generates a markdown report with summary stats, ranked findings table, and recommendations
5. Writes schema-valid candidate-work stubs when a file report or explicit candidate destination is requested; stdout-only runs remain write-free
6. Optionally creates an OpenSpec proposal stub from the top finding

Data Sources

The skill consumes capability-gap signals from four emitters via the shared D4 tag schema:

Source	Tag	How it gets there
Agent self-report	source:self-reported	Agent calls remember MCP tool during failure
Coordinator audit-triage	source:coordinator-emitted	LLM classifier over audit batches
Session-log	source:session-log	Agent fills ### Capability Gaps Observed at phase boundary
Transcript mining	source:transcript-mined	/collect-transcripts deep-analysis pass

The skill also scans openspec/changes/**/session-log.md for ### Capability Gaps Observed sections to catch gaps not yet mirrored to memory.

Deduplication is keyed on (capability_gap, affected_skill, session_id). When the same gap appears from multiple sources, all sources are preserved — cross-source agreement is the strongest signal.

Prerequisites

• Python 3.11+
• Coordinator running at COORDINATOR_URL (default: http://localhost:8000)
• Falls back gracefully with a warning if the coordinator is unreachable

Steps

1. Analyze Failure Patterns

python3 <agent-skills-dir>/improve-harness/scripts/analyze_failures.py \
  --time-window ${TIME_WINDOW:-30} \
  --json

2. Generate Report

python3 <agent-skills-dir>/improve-harness/scripts/generate_report.py \
  --time-window ${TIME_WINDOW:-30} \
  ${CREATE_PROPOSAL:+--create-proposal} \
  ${OUTPUT:+--output "$OUTPUT"}

3. Review and Act

• Review the ranked findings table
• For high-priority gaps, use --create-proposal to generate an OpenSpec proposal stub
• Refine the proposal with human guidance and submit via /plan-feature