Configuration Surface

Runtime behavior is declared by a flow manifest, a verification policy, an execution mode, and persistence settings. The manifest and policy are JSON; the CLI does not infer missing governance from ambient defaults.

Flow manifest

A v1 manifest owns the identifiers and boundaries that must survive planning and replay:

Surface	Required declarations
Identity	flow_id, tenant_id, flow_state
Determinism	determinism_level, replay mode and acceptability, entropy budget, optional non-determinism intent
Replay	minimum claim overlap and maximum contradiction delta
Dataset	identifier, tenant, version, hash, state, storage URI, and deprecation permission
Topology	agents and agent:dependency edges
Authority	retrieval contracts and verification gates

Unknown top-level keys are rejected. The loader fixes spec_version to v1 and requires an explicit determinism level; default is not accepted. Dataset tenant and flow tenant must agree, agents must be unique, and dependencies must form a reachable directed acyclic graph.

The checked-in minimal example is packages/bijux-canon-runtime/examples/boring/flow.json.

Verification policy

Live execution and replay require a policy JSON document. It declares the verification level, failure mode, randomness tolerance, arbitration rule and quorum, required evidence, maximum rule cost, rules, and the rule identifiers that cause failure or escalation. The policy fingerprint is persisted with the trace and compared during replay.

Execution modes

Mode	Behavior	Persistence
plan	resolve and validate the execution plan without running it	none
dry-run	exercise the dry-run executor and produce a trace	DuckDB required
live	invoke runtime integrations, verify, and finalize	DuckDB and policy required
observe	verify a supplied observed run	Python API; store, policy, and observed run required
unsafe	execute through the live path with an explicit warning event	Python API requires store and policy

The supported CLI surface is:

bijux-canon-runtime run MANIFEST --policy POLICY --db-path RUNS.duckdb
bijux-canon-runtime replay MANIFEST --policy POLICY \
  --run-id RUN_ID --tenant-id TENANT --db-path RUNS.duckdb
bijux-canon-runtime inspect run RUN_ID --tenant-id TENANT --db-path RUNS.duckdb

All three accept --json; run and replay accept --strict-determinism. Planning, dry-run, diff, failure explanation, and database validation are also implemented as CLI commands but hidden from top-level help. unsafe-run is present in the parser but cannot currently supply the verification policy that runtime validation requires; do not depend on that CLI path.

Python configuration and environment

ExecutionConfig additionally accepts artifact and execution stores, an ExecutionBudget, observers, an observed run, parent and child flow IDs, a resume run ID, verification and non-determinism policies, and strictness. Executable modes use a write store; resumed execution also needs a read store.

Set BIJUX_CANON_RUNTIME_STRICT=1 to force strict determinism. The legacy AGENTIC_FLOWS_STRICT=1 name is also recognized. Strict environment mode rejects dry-run and unsafe execution. AGENTIC_FLOWS_DB_PATH configures the DuckDB file checked by HTTP readiness; it does not configure CLI commands.