Laya Orchestrator

laya is the bundle's System-1 decision engine: typed answers (choice, score, noul) in one forward pass, offline, free. This skill teaches WHEN to interrogate it and how to treat the answers. Contract modes (off|shadow|assist) and the resident daemon are in .devin/research/LAYA_MASTER_ARCHITECTURE.md.

When to interrogate

Muscle calls (run inside extensions/laya-tools/ via its .venv):

.venv/Scripts/python filebool.py --glob "src/**/*.py" \
    --question "contains authentication data"
.venv/Scripts/python laya_cli.py predict --state '{"body": "..."}' \
    --questions-file q.json          # ad-hoc advisory only

When NOT to interrogate

• The fact is directly checkable with read/grep/exec → verify, don't ask (Rule 12). Laya answers what to read first, never what the fact is.
• Fewer than ~10 files: reading them is cheaper and certain.
• Binary/unreadable files: they abstain anyway.

Output doctrine

• abstain = "check it yourself", never silently dropped from the result map.
• yes/no = reading-order priority, not a conclusion. Never cite a Laya verdict to the user as evidence without tool verification.
• confidence is entropy-based, not calibrated accuracy; thresholds come only from the profile's own calibration.

Fencing (Camada 10)

laya_cli.py predict with ad-hoc questions is advisory: its output MUST NOT feed hooks, policies, or gating scripts. If a question recurs, promote it to a frozen -vN profile in decision_contract.

Layer 7: LAYA-GATE notices

The compact-gate hook injects a one-line LAYA-GATE directive when pressure and trajectory warrant it (prune | fold | handoff | clear). Execute it inside the current turn, never mid tool-batch:

• prune → run extensions/laya-compactor/compact.py on the session transcript.
• fold → offload >50k-token artifacts to files (context-folding).
• handoff/clear → invoke handoff skill, then clear.

Cost table