recall: look up before you act

Never invent history or guess the rationale behind architecture and rules. Always retrieve physical evidence before acting or explaining past design decisions.

Four-level deterministic retrieval protocol

When investigating past decisions, incident context, or rule rationale, execute this four-level retrieval chain:

Level 1: dev_env SSOT check (single source of truth in dev_env)
   ↓
Level 2: Global memory search (basic-memory across all projects)
   ↓
Level 3: Git commit archeology (git log -p / git log --grep)
   ↓
Level 4: Physical runtime verification (worker.log, live probes, exit codes)

Level 1: dev_env SSOT check

• dev_env is the only editable source for root rules (workspace-iac/etc/rules/subagents.md) and common skills (skills/common/*).
• Always inspect the dev_env repository directly to verify current authoritative definitions rather than relying on projected worktree snapshots.

Level 2: Global memory search

• Run search_notes with search_all_projects: true.
• Query exact keywords (e.g. audit, swarm, decisions, sessions, handover).
• Read full notes with read_note when relevant contracts or decisions match. Do not rely on short search excerpts.

Level 3: Git commit archeology

• When the question asks "why is this designed this way" or "when did this change":
  • Run git log -p -n 3 -- <target-file> to read the exact diff and commit message.
  • Run git log --grep="<keyword>" --oneline to find the introducing PR and measured failure evidence.
  • Read the exact problem description. Do not synthesize or assume rationale.

Level 4: Physical runtime verification

• When a tool, worker, or command fails:
  • Check live daemon logs: tail -n 50 ~/.local/state/subagent-worker/worker.log.
  • Check the LLM roles and the provider: subagent-worker --check-models.
  • Check physical git status and branch locks: git status -s, git worktree list.
  • Never classify a transient network or rate-limit failure as an architectural conflict.

How to use results

1. Empty result: Continue silently. Report "no memory found" only when the owner asks.
2. Cite briefly with anchors: Cite the conclusion, date, commit SHA, note id, or exact line numbers (file#Lxx-Lyy). A statement without an anchor equals an invention.
3. Code wins over memory: Memory records a past decision. The live code is the present fact. On conflict, follow the code, cite both, and report the drift.
4. Anchor small-model screening: When a light model screens notes or a knowledge base, its output must carry anchors (file#Lxx-Lyy or note permalink).