| name | routing-cybernetic-workflows |
| description | Use when a formed task needs a decision between ordinary direct work and a durable commitment-controlled run. |
Routing Cybernetic Workflows
Purpose
Choose the smallest control entry that preserves the formed task. This skill
routes only; it does not create files or execute target work.
Use $framing-cybernetic-intent first for pre-task input that is still dissatisfaction,
risk sense, symptoms, method preference, or process distrust rather than a
formed task.
For that branch, emit the literal response-only handoff:
$framing-cybernetic-intent <pre-task input>
Entries
ordinary_direct_work
commitment_controlled_run
ordinary_direct_work: normal Codex work is enough to preserve the request.
commitment_controlled_run: execution needs a durable human-approved
commitment covering meaning, blocking obligations, completion evidence,
non-goals, or live/destructive authority.
Do not choose a commitment-controlled run merely because solution design,
implementation strategy, file selection, commands, debugging, parallelism, or
subagent scheduling are uncertain. Those are agent-owned execution choices.
A material fidelity/cost tradeoff also does not choose the control entry by
itself. Framing may state the user value and truthful approximation boundary.
Before either route implements a materially visible tradeoff, perform a
fact-bound outcome-fidelity check: compare a highest-fidelity reference with
truthful lower-cost paths, reject misleading claims, and ask the user only if
non-dominated paths materially differ in visible semantics. In ordinary direct
work this is an agent-owned current-turn check; in a controlled run it happens
during commitment formation before approval.
The v1 bounded_runtime and controlled_run entries are legacy compatibility
paths for existing run directories only. Never start a new v1 run.
Decision Test
Use commitment_controlled_run when at least one of these must remain stable
across a long or risky execution:
- the user's objective or blocking result states;
- observable completion conditions and rejected substitutes;
- evidence claims that must bind one exact candidate;
- explicit non-goals;
- concrete live, external, destructive, or forbidden action boundaries.
Otherwise use ordinary_direct_work.
Re-route an ordinary task before completion or a consequential action when new
facts reveal durable obligations, explicit non-goals, a same-subject evidence
requirement, or authority semantics. Discovery is agent-owned; do not wait for
the user merely because the route must become more controlled.
For ordinary direct work with a material Value / Fidelity Frame, do not skip the
outcome-fidelity check merely because no durable commitment is needed. Keep the
comparison ephemeral, choose the best truthful path from current facts, and
execute directly. Do not turn candidate paths, cost estimates, or a plan into a
new approval object.
Gates
Every commitment-controlled run uses:
- fixed pre-commitment counterexample review before human approval;
- explicit human approval of one commitment hash;
- final independent counterexample review against one evidence subject;
- deterministic final verification;
- fresh action grants only for authority entries declared in the commitment.
Schema validation is structural. Counterexample review is semantic.
Response
Return:
Control entry decision.
Why.
Commitment scope: the meanings, obligations, evidence, non-goals, or
authority that require durable control.
Recommended next step.
Rejected control: process details deliberately left agent-owned.
For an ordinary-direct material fidelity tradeoff, Recommended next step must
say that the agent will first perform the current-turn outcome-fidelity check,
then implement and verify the selected truthful path. It still emits no handoff
prompt.
For a commitment-controlled run, the handoff must be:
$forming-cybernetic-commitments <formed task>. 允许使用 subagent 执行 counterexample review。
Do not create the commitment in the routing turn.