SuperAPI AI Directory

long-task-control

Design & creation

View repository

Updated 10-16 07:56

About this skill

Long Task Control

Preserve verified work while correcting the course. Constrain acceptance and ownership; leave execution tactics free. This workflow needs no other skill.

Target

Reconstruct the outcome, scope and acceptance criteria from the original request and later authorized changes. Newer authorized intent wins; expose unresolved conflicts. Compare the work against this target at each major phase boundary.

Observe

Inspect artifacts, commands, checks and delegate state directly. Separate liveness, progress toward the target and verified completion. Silence, elapsed time, an active process or a dirty diff alone establishes neither progress nor failure. Useful diagnosis and necessary preparation count as progress before acceptance checks finish.

Wait for a meaningful checkpoint while work advances or a known dependency is pending. A missed checkpoint prompts targeted diagnosis, not automatic replacement. Repeated failures without changed conditions, a new hypothesis or diagnostic information call for a new route. Preserve accepted work; stop the unproductive attempts, not the overall task.

Challenge

Default to separate, clean-context, read-only reviewers at major phase boundaries. Challenge Design (soundness, ROI, gold-plating) and Drift (intent, progress, evidence, stale work). Reviewers form their findings independently before reconciliation. Complete contract-changing review before dependent writing; non-blocking review may run concurrently.

Give one owner the findings: deduplicate, check the evidence, resolve conflicts and classify actionable defects, boundary notes, overdesign and false positives. Explain material findings with concrete evidence. Separate review and repair ownership. Re-review material repairs. At closure, run both lenses twice over the entire cumulative delivery; preserve any stronger requested review cadence. Report unavailable independence or incomplete required rounds as review gaps.

Correct

Disposition material findings: keep, prune, repair, replan or escalate. Break remaining work into independently verifiable outcomes. Preserve useful context and artifacts while retiring stale work and its descendants.

For delegated work or handoff, read Delegation.

Verify

Map the scoped acceptance criteria to real artifacts, actions and observable outcomes. Identify the delivered version and relevant configuration; confirm the instance and prerequisites before driving it. Use canonical repository checks and public user paths, including relevant side effects and required failure or rollback cases. For documents, check the actual output against its sources. Existing project verification instructions may supply recipes; their absence does not prevent direct verification.

Retain the actions, actual results, artifact identity and evidence locations. Separate review findings from observed behavior. Diagnose environment or driver failure separately from product failure; preserve the intended expectation. Repair within scope, then repeat the affected checks and disturbed integration paths. Reuse evidence that still applies. Clean only this run's resources and keep its proof readable.

Continue and close

Continue with the next action that advances the target. Claim task completion only after necessary delegated results and required reviews are consumed, material findings are resolved, and applicable acceptance checks pass. Report failed, blocked and unexecuted paths explicitly.

At a durable boundary, update the existing run record with surprises, changed decisions and reasons, unresolved judgments, deferred work, evidence pointers and the next action. A control pass can end with a corrected next action while the overall task remains pending.