Capability Map

bijux-canon-agent provides role-based workflow control whose lifecycle, decisions, convergence, termination, and trace can be inspected independently of the final text. Provider access is an adapter capability, not the source of orchestration authority.

flowchart LR
    input["task goal + context"]
    definition["pipeline definition"]
    lifecycle["owned transitions"]
    roles["bounded role calls"]
    merge["shard merge + validation"]
    converge["convergence + termination"]
    output["PipelineResult + RunTrace"]

    input --> definition --> lifecycle --> roles --> merge --> converge --> output

Workflow capabilities

Capability	Owning area	Produced evidence
Strict role contracts	contracts/	immutable inputs, outputs, errors, plans, and runtime records
Canonical pipeline definition	pipeline/definition.py and canonical pipeline	role eligibility, lifecycle phases, transitions, terminal states
Lifecycle control	pipeline/control/	validated transition, stop reason, causal order
Role registry	pipeline/agent_registry.py	explicit role-to-implementation resolution
Bounded role execution	agents/	role output or typed error with call metadata
Input sharding and merge	pipeline/execution/ and pipeline/results/	shard status, merged result, warnings, revisions, action plan
Goal-aware final validation	result finalization	success/failure state that cannot be cached as success when invalid
Cache selection	preparation and runtime support	context-derived key and explicit cache-hit state

Decision and evidence capabilities

Capability	Owning area	Produced evidence
Convergence evaluation	pipeline/convergence/	strategy, window, snapshots, hash, verdict history, reason
Termination classification	pipeline/termination.py	completion, convergence, failure, abort, or exhaustion
Epistemic disposition	pipeline epistemic models	pass/veto and epistemic status separate from execution success
Versioned trace	traces/	header metadata, ordered entries, run fingerprint, replay fields
Trace validation	pipeline/trace_validation/	ordering, completeness, epistemic, and replayability findings
Trace reconstruction	replay support	terminal result projection and documented summary parity
Structured observability	observability/ and execution telemetry	stages, shards, duration, counters, contextual logs
Provider adapters	llm/	provider/model identity, prompt/model hashes, usage, result or error

Role capabilities

The package includes bounded roles for file reading, summarization, critique, validation, stage execution, planning, judging, and verification. A role supplies local behavior within the active lifecycle phase. It cannot override the controller, invent a terminal transition, or silently become runtime authority.

Capability status by authority

Status	Capabilities	Required qualification
package-owned orchestration	role contracts, pipeline definition, lifecycle, sharding/merge, convergence, termination and trace validation	deterministic contract/lifecycle evidence under the selected configuration
included local role behavior	file reading, extractive/simple processing, critique, validation, planning, judging and verification implementations	input format, role eligibility, parser/resource and result-contract evidence
provider-dependent	configured remote model calls and provider usage metadata	credentials, client/provider/model identity, network policy, failure handling and live evidence
interface-specific	complete Python composition, file/directory CLI, fixed offline HTTP pipeline	callers use only the capabilities promised by that interface
host-governed	authentication, tenant isolation, sandboxing, egress, durable storage and confidential-data policy	controls outside the package process

An implementation can exist without being reachable through every public surface. In particular, source-level provider adapters do not make the fixed offline HTTP operation provider-configurable, and a Python-composed workflow does not prove that runtime's package-root live loader can invoke it.

Interpret workflow outcomes

Outcome	Evidence that must agree	What it does not mean
converged completion	convergence strategy/window/history, admitted final result, terminal state and complete trace	factual correctness or runtime acceptance
completed without convergence	stopping rule, last candidate, explicit non-convergence and trace	convergence because execution ended normally
partial result	every successful/failed shard, merge lineage, usable artifacts and partial disposition	whole-input success
veto or validation refusal	source role, issues/findings, target artifact, action plan and disposition	provider or infrastructure failure
provider/role failure	call identity, typed error, retries/fallbacks and controller response	empty role content or skipped history
interrupted/aborted	last causal event, lifecycle state, incomplete work and recovery boundary	finalized replayable completion
fatal orchestration failure	terminal reason, partial evidence and non-success interface result	ordinary role-level rejection

PipelineResult, RunTrace, interface status/exit behavior and stored artifacts must report the same outcome. A trace-first reconstruction should reach the same terminal disposition as the summary file.

Public boundary

Python exposes the complete composition model. The CLI runs files or immediate directory entries and publishes result/trace files. HTTP v1 deliberately runs one fixed offline simple/extractive pipeline. Source-level model adapters do not expand that versioned HTTP contract.

Convergence proves only that a configured condition was observed. Trace replay proves retained fields are coherent. Neither establishes model correctness or recreates a remote provider's historical serving environment.

See Invariants for lifecycle and trace laws and Known limitations for model, convergence, replay, credential, and hosting boundaries.