LaPis AGENTS.md

Protocols

1. Code & Doc Retrieval — Use memory-code / memory-doc

LaPis is the canonical Pi memory stack. For any code/doc lookup, prefer LaPis tools first.

The memory-layer extension enforces structured retrieval over raw file reads:

• Code → memory-code tool
  • Modes: callers, callees, blast-radius, dead-code, complexity, deps, outline, churn, hotspots, cycles, importance, coupling, extractable, hierarchy, signal-chains, layer-violations, index-repo, reindex-repo
• Docs → memory-doc tool
  • Modes: search, outline, backlinks, broken-links, glossary, tutorial-path, code-examples, orphans, coverage, stale-pages, duplicates, index-docs, reindex-docs

If a repo isn't indexed yet, the tool will tell you exactly how to index it.

Enforcement rules:

• read on a code file in an indexed repo without offset/limit → BLOCKED. Use memory-code outline first.
• read on a code file with offset/limit → ALLOWED (editing targeted lines).
• bash grep/rg for browsing or scanning source code in an indexed repo → BLOCKED. Use memory-code instead.
• bash grep/rg for targeted single-symbol lookup (e.g., grep -rn "rankObservations" src/) → ALLOWED when faster than memory-code for exact-name searches.
• bash find for file discovery in an indexed repo → BLOCKED. Use memory-code search instead.
• After calling memory-code outline on a file, subsequent reads are allowed.
• Prefer memory-code search for semantic queries (e.g., "how does context injection work"). Prefer bash grep for exact symbol queries (e.g., "where is rankObservations defined").
• If a lookup seems to need raw source browsing, stop and use memory-code / memory-doc instead.

2. Persistent Memory — Automatic

Memory is handled automatically by the memory-layer extension. It:

• Injects context at session start (decisions, preferences, recent memories)
• Re-injects context after compaction — /compact no longer destroys memory awareness
• Auto-detects decisions — pattern-matches assistant messages for decisions, bugfixes, discoveries, and auto-saves them
• Periodic progress checkpoints — every 10 turns, saves an auto-checkpoint with files edited
• Persistent memory reminder — every 5th LLM call, injects a lightweight reminder if no memory tool was used recently
• Git-triggered trust sync — after git pull/checkout/merge/rebase, auto-syncs code trust scores
• Auto-saves session summaries on shutdown with topics, files, and turn count
• Auto-recovers incomplete sessions
• Detects stale indexes and warns when code indexes are out of date
• Hard-blocks reads of code files in indexed repos — must use memory-code outline first; partial reads (offset/limit) allowed for editing
• Auto-dreams at turn 50 of each active session — runs the Dream Cycle to clean stale memories (not just old). Targets superseded, zero-recall auto-saved, stale corrections, and replaced configs. Age alone is NOT a signal.