Skip to content

run_retrospective

Perform a structured post-engagement retrospective analysis.

Read-only: Yes (optionally writes to disk)

Description

Produces five structured outputs:

  1. Inference rule suggestions — patterns the graph shows that existing rules missed
  2. Skill gap analysis — skills unused vs. techniques attempted without skills
  3. Context-improvement recommendations — where context, logging, validation, and coverage should improve
  4. Attack path report — client-deliverable markdown (timeline, findings, recommendations)
  5. Heuristic RLVR traces — state→action→outcome triplets with explicit confidence and trace quality

Parameters

Parameter Type Default Description
write_to_disk boolean false Save all outputs to files
output_dir string "./retrospective/" Directory for output files

Returns

Field Type Description
summary string High-level summary
inference_suggestions array Suggested new inference rules with evidence
skill_gaps SkillGapReport Unused skills, missing skills, failed techniques
context_improvements ContextImprovementReport Frontier observations, context gaps, OPSEC observations, logging quality
training_traces_count number Number of RLVR training traces
trace_quality TraceQualityReport Quality assessment of training traces
report_preview string First 500 chars of the markdown report
output_dir string Output directory (if write_to_disk: true)

Output Files (when write_to_disk: true)

Written to <output_dir>/<engagement_id>/:

File Content
report.md Client-deliverable attack path report
inference-suggestions.json Suggested inference rules
skill-gaps.json Skill gap analysis
context-improvements.json Context improvement recommendations
training-traces.json RLVR training traces
trace-quality.json Trace quality assessment
summary.txt Summary text

Usage Notes

  • Run at the end of an engagement or after significant progress
  • The attack path report (report.md) is designed for client delivery
  • Inference suggestions can be applied to future engagements via suggest_inference_rule
  • Training traces can be used for model fine-tuning (RLVR)
  • Also available as a CLI: npm run retrospective