| name | agent-self-driving |
| description | Use only when the user explicitly invokes $agent-self-driving, or an active self-driving controller already owns the task, for long-task automation that must keep lifecycle owners, agents, review/repair, evidence, and closure moving until completion or a true user decision. Do not use merely because work is large or for routine debate, review, lanes, or integration. |
Agent Self-Driving
Opt-in automation controller. The main thread remains moderator and owns final
claims.
CONTROL -> DELEGATE -> VERIFY -> CONTINUE.
Boundary
Activate only by explicit $agent-self-driving; otherwise this skill must
already be the active controller. Size alone never activates it.
The active instance is the sole controller. For software delivery, read and use
dev-flow as the parent lifecycle router. When it selects
a leaf, read that leaf SKILL.md, let it own the action, and return evidence to
this same controller. Never restart, nest, or route back into self-driving.
Self-driving owns only its orchestration overlay: controller status, bounded
next action, private blackboards, task packets/raw outputs, convergence/findings
state, and evidence pointers. Parent/leaf workflows own Requirements, Spec,
Eval, Plan, implementation, review, lanes, integration, docs truth, and
lifecycle handoff. agent-runtime owns participant/model profiles, capability,
authorization, and worker/external session truth; link those records, never copy
them into controller state.
Read workflow-integration.md when another
workflow owns task state. Read artifact-layout.md
only when continuity, risk, auditability, agents, or handoff requires persistence.
Intensity and Shape
0 main agent only
1 one focused pass
2 multi-agent audit
3 bounded review -> repair -> recheck
4 full delivery lifecycle
Start at the lowest sufficient level. Read
lifecycle-intensity.md when intensity is
unclear or Level 3-4 applies.
Runtime Gate
A host-native worker already exposed by the current runtime, with no participant
or model choice and no external boundary, uses agent-runtime for worker
lifecycle—including resume—but does not require a new model profile.
Before named/model-selectable participants, capability or authorization checks,
external/paid/account/data-leaving use, or worker/session continuity, read and
run agent-runtime. Runtime—not this controller—owns
consent and session decisions. For direct delegation outside agent-review or
agent-lanes, read task-packets.md.
Converge and Continue
Start from source, treat agent output as claims, and normalize by scope,
evidence, severity, and action. Read
output-normalization.md for heterogeneous
outputs or conflicting findings; read
review-convergence.md for repeated review,
repair, rebuttal, or recheck.
Accepted blocker/major findings close only through evidence-backed repair or
rejection. A true pause condition keeps the finding and run open; it is not a
completion path. Persist only when the selected intensity needs recovery. If
status is in_progress, perform the next bounded action.
Pause only for a true user decision, unauthorized boundary, missing access, or
unavailable verification with no safe substitute. Claim complete only from
current owner evidence.
Bounded describes scope, ownership, and evidence—not elapsed time. Long-running
workers remain in_progress across short polling timeouts. Apply
agent-runtime worker lifecycle; do not invent a wall-clock cutoff or interrupt
merely to make the controller advance.
