credence-governor-core 0.9.0

Harness- and domain-neutral Bayesian tool-call governance — a pure wire consumer of the Credence engine (credence-skin). Run credence-governor-daemon to serve it on :8787.

credence-governor-core

The harness- and domain-neutral Bayesian tool-call governance core for credence-governor — a pure wire consumer of the Credence engine (credence-skin). It carries no probabilistic code: all inference runs server-side in the engine, reached over the protocol-versioned JSON-RPC skin wire; this package only extracts features from a neutral event and drives proceed / block / ask decisions.

What's in it

• schema.py — the harness-neutral contract: AgentToolEvent, Session, Message. Every adapter translates its native events into these shapes.
• features/ + safety.py — the single source of truth for feature & taint extraction (waste loops, repetition, taint-flow, exfil chains). Adapters carry none of this; they send a Session and the daemon extracts.
• session.py (BrainSession) — boots the warm beliefs and drives the engine's structure_bma / structure_observe / structure_decide / routing_* verbs.
• daemon.py — the HTTP surface: POST /decide (synchronous, for subprocess-hook adapters), POST /sensor + GET /signals (SSE, for in-process plugins), /report, /ready. stdlib http.server; one lock serialises engine calls.
• data/ — "coding-ness" as data: the bdsl/ feature schema + utility, the brain/*.counts.json warm beliefs, and profiles/. Swappable.

Install & run

Zero-config — with docker or podman on PATH, the daemon auto-runs the published engine image (ghcr.io/gfrmin/credence-skin:latest); no env vars needed:

pip install credence-governor-core   # pulls the engine's wire client (credence-skin-client)
credence-governor-daemon             # zero-config: auto-runs the engine via docker/podman

With uv — no separate install step. The package is credence-governor-core but the console script is credence-governor-daemon, so the --from is required (uvx otherwise looks for a PyPI package matching the command name):

# one-off / ephemeral — re-resolves each run (cached after the first):
uvx --from credence-governor-core credence-governor-daemon

# persistent — installs `credence-governor-daemon` onto PATH (~/.local/bin):
uv tool install credence-governor-core
credence-governor-daemon

Prefer uv tool install for a long-lived service, and required if you use an adapter's opt-in autostart — it spawns the bare credence-governor-daemon, which must be on PATH (an ephemeral uvx install is not). uvx is ideal for a quick "spin it up to test". Either way uv supplies only the Python host; the engine still comes from docker/podman (or the overrides below).

Overrides (resolution order: CREDENCE_SKIN_COMMAND → CREDENCE_ENGINE_DIR → zero-config):

# prod: pin the engine image by digest (reproducible) — highest precedence
CREDENCE_SKIN_COMMAND="docker run --rm -i ghcr.io/gfrmin/credence-skin@sha256:<digest>" \
  credence-governor-daemon
# dev engine: a local checkout of the Credence engine
CREDENCE_ENGINE_DIR=/path/to/credence credence-governor-daemon

The daemon listens on http://127.0.0.1:8787 by default (CREDENCE_GOVERNOR_HOST / CREDENCE_GOVERNOR_PORT). Every adapter is a fail-open silent no-op until this daemon answers on :8787 — verify with curl -s localhost:8787/ready.

POST /decide {tool_name, input, session:{cwd, project_root, messages}} → {action, features, event_id}.

Raw-event capture (opt-in)

The observation log stores only the extracted features of each decision, so when the extractor gains a feature, past traffic cannot be re-derived. Set CREDENCE_GOVERNOR_CAPTURE=1 (or a path) to additionally capture the raw (tool_name, input, session) of each governed call to ~/.credence-governor/raw_events.jsonl — a re-extractable corpus a later feature-engineering pass folds in as benign negatives (training.capture_corpus.load_capture). Off by default (raw sessions carry file contents + commands). String fields over CREDENCE_GOVERNOR_CAPTURE_MAXLEN (default 1 MB — a runaway-size guard, not a privacy cap; truncated records are excluded from the fold since they would re-extract to different features) are cut. Non-causal: capture never feeds a belief or a decision.

Tests

python -m pytest tests -q   # parity tests; no engine needed (pure extractor/manifest code)

License

AGPL-3.0-or-later.