Scope and Non-Goals

bijux-canon-agent owns governed coordination among bounded roles. It defines workflow shape, transition authority, shard handling, convergence, termination, result assembly, and trace evidence while keeping model providers and final run acceptance outside its authority.

flowchart LR
    input["task goal + context"]
    definition["roles + transitions + stop policy"]
    execution["bounded calls + shard merge"]
    decision["validation + convergence + termination"]
    trace["PipelineResult + RunTrace"]
    runtime["flow acceptance"]

    input --> definition --> execution --> decision --> trace --> runtime

In scope

• strict agent input, output, error, call, execution-plan, and runtime contracts;
• canonical pipeline definitions, role registries, lifecycle phases, transitions, stop conditions, and interruption handling;
• bounded file-reader, summarizer, critique, validator, stage-runner, planner, judge, and verifier roles;
• input preparation, context identity, cache keys, sharding, stage execution, result merge, final validation, revisions, warnings, and action plans;
• convergence strategies, snapshots, oscillation detection, decision windows, termination classification, and epistemic disposition;
• versioned traces, replayability metadata, ordering/completeness validation, result reconstruction, structured logging, counters, and telemetry;
• Python composition, CLI file/batch execution and replay, and the fixed offline HTTP v1 pipeline.

Non-goals

Not owned here	Owning boundary
Source normalization and chunk identity	bijux-canon-ingest
Vector backend selection and retrieval replay	bijux-canon-index
Claim support, evidence meaning, and reasoning verification	bijux-canon-reason
Final tenant authority, persistent governed-flow acceptance, and flow replay	bijux-canon-runtime
Guaranteeing model correctness, calibration, or provider determinism	model/provider evaluation and contract
Process isolation, distributed scheduling, secrets, network, or tenant policy	hosting system
Repository release and maintenance automation	maintenance tooling and handbook

Distinct outcome dimensions

Role verdict, convergence, termination, and execution success answer different questions. A veto can occur in a technically successful role call. Convergence can stop on stable but incorrect output. Resource exhaustion can terminate without convergence. A valid trace preserves these distinctions rather than compressing them into one success flag.

Provider boundary

Provider adapters supply model behavior under recorded metadata. Zero temperature is required for the package's replayable designation, but it cannot freeze provider weights, infrastructure, or hidden policies. Live tests demonstrate connectivity and metadata capture, not historical reproducibility or answer correctness.

Scope test

A change belongs here when it changes who acts, in which valid order, under which stop and convergence rules, or what trace is required to reconstruct the workflow. If it changes the meaning of a claim or the final authority over the whole run, it belongs below or above agent.

See the capability map and known limitations for the implemented and external boundaries.