Retrospective

Retrospective

Turn execution experience into institutional knowledge. This skill orchestrates a fixed Phase 0 through Phase 5 workflow that gathers evidence, generates insights, diagnoses root causes, decides actions, scores atomicity, and persists learnings. The long-form rubrics live verbatim in references/; this file is the orchestration contract.

This skill replaces the former retrospective agent (.claude/agents/retrospective.md). It is callable from Skill("retrospective") and from /retro fill <date> (Issue #2079). The Stop hook that once invoked it automatically was deleted in #3349, so a retrospective is written when someone asks for one.

Triggers

When to Use

The output artifact is a Markdown file. The Learning Extraction Template in references/learning-template.md defines the exact structure. Save to .agents/retrospective/YYYY-MM-DD-[scope].md. When filling an auto-retro skeleton, write to the existing YYYY-MM-DD-auto-retro.md file produced by the Stop hook.

Inputs

Treat the session log as the system of record for what happened. Git history and GitHub activity are derived evidence that corroborate or extend it. When a source is unavailable, degrade gracefully: produce the artifact from the evidence you have and mark the missing sections, never substitute invented data.

Process

The workflow is six sequential phases. Phase 0 gathers facts. Phases 1 and 2 interpret them. Phase 3 decides actions. Phase 4 extracts and scores learnings. Phase 5 persists them. Each phase links to the rubric and template it uses. Run them in order; do not interpret before you observe.

Phase 0: Data Gathering

Gather facts before interpretation. Observation precedes diagnosis.

• Run the 4-Step Debrief (Observe, Respond, Analyze, Apply): see frameworks.md, 4-Step Debrief.
• Build the Execution Trace chronology: see frameworks.md, Execution Trace Analysis.
• Run Outcome Classification (Mad, Sad, Glad): see frameworks.md, Outcome Classification.

Evidence sources: the most recent session log under .agents/sessions/, git log for the period, and optional GitHub activity through the github skill. Do not use raw gh.

Phase 1: Generate Insights

Make meaning from data. Look past symptoms to find causes.

• Five Whys is mandatory for every failure: see frameworks.md, Five Whys.
• Fishbone Analysis for complex failures with multiple contributing factors: see frameworks.md, Fishbone Analysis.
• Force Field Analysis when a pattern recurs despite knowing better: see frameworks.md, Force Field Analysis.
• Patterns and Shifts for multi-session trends: see frameworks.md, Patterns and Shifts.
• Learning Matrix for quick categorization when short on time: see frameworks.md, Learning Matrix.

Phase 2: Diagnosis

Prioritize findings for action. Diagnostic priority order: critical error patterns, success analysis, near misses, efficiency opportunities, skill gaps, traceability health. The full priority order, traceability metrics, and diagnosis template live in diagnosis-and-actions.md, Diagnosis.

For each root cause that Five Whys surfaces, store a root-cause pattern for future prevention: see diagnosis-and-actions.md, Root Cause Pattern Management.

If the work touched diagnosis or action classification, stop here for the diagnose this failure trigger; otherwise continue to Phase 3.

Phase 3: Decide What to Do

Move from insights to action.

• Action Classification (Keep, Drop, Add, Modify): see diagnosis-and-actions.md, Action Classification.
• SMART Validation of every proposed learning before storage: see diagnosis-and-actions.md, SMART Validation.
• Dependency Ordering of the resulting actions: see diagnosis-and-actions.md, Dependency Ordering.

Phase 4: Learning Extraction

Transform insights into stored knowledge. Score every learning 0 to 100 percent for atomicity and reject vague statements. The scoring rubric, quality thresholds, worked examples, and evidence-based tagging live in diagnosis-and-actions.md, Atomicity Scoring.

Assemble the artifact using the byte-exact Learning Extraction Template. Save to .agents/retrospective/YYYY-MM-DD-[scope].md. When filling an auto-retro skeleton, overwrite the placeholder sections in the existing YYYY-MM-DD-auto-retro.md and remove the UNFILLED banner.

Phase 5: Persist and Close

Persist learnings to memory and evaluate the retrospective itself.

• Persist learnings with atomicity at or above 70 percent to Serena memory (ADR-037). Search for existing patterns before creating new entries to avoid duplicates: see diagnosis-and-actions.md, Memory Protocol.
• Close with +/Delta, ROTI, and Helped, Hindered, Hypothesis: see frameworks.md, Closing Activities.
• Route any P0 or P1 delta item to a GitHub issue through the github skill; store P2 and P3 items in backlog memory.

Success Criteria

Before the retrospective is complete, confirm:

• One Markdown file exists at .agents/retrospective/YYYY-MM-DD-[scope].md (or the existing auto-retro skeleton was filled and its UNFILLED banner removed).
• The artifact structure matches the Learning Extraction Template byte-for-byte, with placeholders filled.
• Every extracted learning carries an atomicity score and an evidence reference.
• Learnings at or above 70 percent atomicity are persisted to Serena memory, or the memory write failure is noted in the artifact.

Boundaries

• This skill reads evidence and writes one artifact plus memory entries. It does not open PRs itself; it routes delta items to the github skill.
• Memory and GitHub are integration points. A failed memory call degrades to a documented fallback (write the artifact, note the memory write failed), never a silent context loss.
• Keep entry points thin. The Stop hook and /retro fill parse inputs and call this skill; they do not re-implement the workflow.

Scripts

References

• frameworks.md: Phase 0, 1, and closing activity rubrics (4-Step Debrief, Execution Trace, Outcome Classification, Five Whys, Fishbone, Force Field, Patterns and Shifts, Learning Matrix, +/Delta, ROTI, Helped/Hindered/Hypothesis).
• diagnosis-and-actions.md: Phase 2, 3, 4, and 5 rubrics (diagnosis priority and traceability, root-cause patterns, action classification, SMART validation, atomicity scoring, evidence-based tagging, memory protocol).
• learning-template.md: the byte-exact Learning Extraction Template that the output artifact must match.
• .claude/agents/retrospective.md: the source agent body these references were lifted from (canonical source for the rubrics; retired once the skill ships).