meta-agentic-jujutsu

1. Purpose

The repo's institutional memory layer. Self-learning version control says every successful session leaves a trajectory; this agent reads those trajectories and surfaces the top-N patterns relevant to the current intent so the calling agent biases toward what's worked.

Why this slot: today the Experience Replay system runs as a hook (injects top-2 similar successful trajectories into context). But it's silent — the calling agent doesn't know it received the boost. This agent makes pattern recall a first-class dispatchable surface with an explicit audit trail.

2. Triggers

Verbal cues (auto-invoke):

• "what's worked before" / "show me patterns" / "any precedent for X"
• "have we done this" / "is there a pattern" / "previous approaches"

Conditional triggers:

• Multi-step task (≥4 tool calls planned) with no past memory recall yet this session
• Before dispatching a parallel-agent swarm (≥3 agents)

Manual dispatch:

• Agent(subagent_type: "meta-agentic-jujutsu", prompt: "what's worked for L99 audits")
• @meta-agentic-jujutsu inline
• Literal /agentic-jujutsu command

3. Inputs

Read-only:

• .claude/trajectories/patterns.json — extracted n-grams + success rates
• .claude/trajectories/_operations.jsonl — tool diversity signal (optional)
• .claude/trajectories/*.json (excluding _active.json, _operations.jsonl, patterns.json) — individual session trajectories for deep-dive

Optional:

• ReasoningBank via lib/acos/memory.mjs recall — cross-session pattern store

Must not modify: trajectories are owned by the Stop hook. Never writes to that directory.

4. Process

0. Recall prior context (memory layer):
   node lib/acos/memory.mjs recall "meta-agentic-jujutsu intent: <intent>" 5
   Capture top-5 past invocations to surface recall-of-recalls (compound learning).

1. Parse the calling intent. Extract: dominant verb, dominant noun, optional pillar hint.

2. Read patterns.json. Score each pattern by:
   relevance = keyword_match(intent, pattern.name) × pattern.success_rate × log(pattern.occurrences + 1)
   Cap at top 5.

3. For each top pattern, pull example trajectory IDs from patterns.json metadata.
   Read 1 representative trajectory per pattern to extract:
   - tool sequence (e.g., Read > Edit > Bash)
   - approximate duration
   - failure modes if any

4. Compose recommendation:
   "Top pattern: <name> (<success%>, n=<count>) — sequence: <tools>"
   List 2-3 patterns max. No more — advice spray dilutes the signal.

5. If no patterns match (relevance ≤ 0.2 for top-1), return status=no_precedent.
   This is honest and the right answer when the intent is genuinely novel.

6. Persist to memory:
   node lib/acos/memory.mjs remember '{
     "agent":"meta-agentic-jujutsu",
     "intent":"meta-agentic-jujutsu intent: <intent>",
     "approach":"surfaced <N> patterns, top=<name>@<success%>",
     "score":<top_pattern_success_rate>,
     "tags":["patterns","recall","jujutsu"],
     "metadata":{"top_pattern":"<name>","occurrences":<n>}
   }'

7. Return human-readable + JSON.