retrospective-learning

You orchestrate a retrospective audit. Subagents do the scanning (read-only). You categorize, present findings indexed, and apply only what the user accepts per item.

$ARGUMENTS is targeted feedback — any free-form hint ("the tool-use felt off today", "audit my skills budget") gets highest priority but never narrows scope. The audit is always a full sweep.

Pre-flight

1. Announce: Starting retrospective.
2. Resolve state paths once and cache them:

   ATOMIC_STATE="${HOME}/.atomic"
   RUNS_DIR="${ATOMIC_STATE}/retro-runs"
   LEARNINGS="${ATOMIC_STATE}/retro-learnings.md"
   SCRATCH="tmp/$(date +%Y-%m-%d)-retro"
   mkdir -p "$RUNS_DIR" "$SCRATCH"
   RUN_ID="$(date +%Y-%m-%d-%H%M%S)"
   

3. Resolve the current Claude project session dir (used by history scan):

   PROJECT_SLUG=$(pwd | sed 's|/|-|g')
   SESSIONS_DIR="${HOME}/.claude/projects/${PROJECT_SLUG}"
   

   If $SESSIONS_DIR does not exist, the history scan will degrade to current-only — note that in the run summary, do not abort.
4. Read $LEARNINGS if it exists. It carries: acceptance rates per category, modify-signal patterns, deprioritized finding types. Apply as soft weights during Phase 4 categorization. If absent, proceed — it will be created at the end of the run.

Step 1 — Pick scope

Prompt via AskUserQuestion:

Question: What scope should this retrospective cover?
Options:
  - Historical + current conversation (recommended) — last 5 .jsonl sessions + prior /retrospective-learning audits + this conversation
  - Current conversation only — skip history scan, no prior-audit cross-check

Store the answer as $SCOPE.

Step 2 — Dispatch background agents in parallel

Single message, two or three Agent tool calls. Briefs live in $SCRATCH/.

2a. Discovery — atomic-investigator (always)

Write $SCRATCH/discovery-brief.md:

# Discovery brief

Catalog every config-shaped file at BOTH installed (~/.claude) and project levels. Return a `file:line — purpose` table grouped by kind and level.

Look for:

- Global: ~/.claude/CLAUDE.md, ~/.claude/CLAUDE.local.md, ~/.claude/commands/, ~/.claude/agents/, ~/.claude/skills/, ~/.claude/output-styles/, ~/.claude/rules/, ~/.claude/settings.json, ~/.claude/settings.local.json, ~/.atomic/config.toml, ~/.atomic/profile.md
- Project: ./CLAUDE.md, ./CLAUDE.local.md, .claude/commands/, .claude/agents/, .claude/skills/, .claude/settings.json, .claude/settings.local.json, docs/wiki/index.md, .claude/project/followups/INDEX.md, .claude/project/followups/*.md
- Memory: ~/.claude/projects/${PROJECT_SLUG}/memory/MEMORY.md and topic files

For each artifact: path, one-line purpose, char count, line count.
Do NOT propose changes. Inventory only.

Respond in atomic style. Drop filler, pleasantries, hedging. Fragments OK. Technical terms exact. Inventory table only — no preamble, no echo of this brief.