Delivery

Carry an implementation outcome from premise to an honestly reported state. Loading this skill is guidance only: nothing is executed, dispatched, reviewed or published until a real tool call or process does it and its receipt is collected.

Scope

• Use for outcomes that change source, configuration or external state.
• Read-only questions, explanations and lookups do not launch workers.
• Keep effort proportional. A one-file fix is one task; there are no fixed task counts, file caps or coverage targets.

Roles

• Default: Codex coordinates and independently verifies; an actual Claude execution implements. Only explicit owner or project instructions override it.
• The default assigns roles; it does not start a dispatcher. A role is fulfilled only when a real controller or CLI invocation runs it.
• Never hardcode accounts, profiles, models or paths. Inherit configured CLI defaults and report the model actually observed.
• An unavailable provider or reviewer stays a named gap, for example independent review: not performed (verifier provider unavailable). Same-provider self-review is never reported as cross-provider review.
• An existing configured controller owns dispatch and STOP. Without one, a direct CLI invocation is a fallback: it preserves identity (provider, profile, session) and every gate, and is not reported as managed delivery.

Workflow

1. Ground. Read project instruction files and verify the request's premise against source. Record the base identity: exact SHA, toolchain versions and dirty state. For an unfamiliar repository run /repo-context first.
2. Contract. Write one live outcome contract (task contract). Amend it in place when scope changes; do not keep parallel copies.
3. Decompose. Tasks under about 30 minutes, each with dependencies, a write scope and capability routing. Disjoint write scopes may run in parallel in separate worktrees; coupled work runs sequentially.
4. Predeclare. Name the checks and the expected changed-path manifest before execution starts. Map requirements to discriminating evidence using acceptance by task; unavailable required evidence means incomplete acceptance, not a successful zero.
5. Execute and collect. Gather actual terminal receipts (command, exit status, log location) with /verified-handoff. A pending, killed or timed-out run is not a pass. See Provider lifecycle for when a provider run counts as complete.
6. Freeze and reconcile. Freeze the result identity: the result commit SHA, or a digest of the exact dirty bytes when uncommitted. Worktree diff, task result, report and PR must agree with that result and the manifest. Explain every difference or treat it as a defect.
7. Independent review. The reviewer sees requirements, source, diff and raw artifacts before any producer conclusions. Findings and the verdict are bound to the input revision, base identity and frozen result identity. A blocker states the violated requirement and a reproduction; low-impact suggestions are labelled as such.
8. Remediate. Preserve failure evidence; do not overwrite logs. Remediation, a new head or input drift invalidates the bound verdict: freeze the new result and review it again. After repeated failure through the same mechanism, change strategy instead of retrying.
9. Publish. Required local checks pass and an independent verdict without unresolved blockers covers the exact result before push or PR; otherwise hold. Merge requires the configured CI and approval at the exact head SHA, verified ancestry and existing authorization. Report implemented, reviewed, merged and deployed as distinct states.
10. Continue. Proceed through remaining authorized slices while attended. Do not create watchers, schedules or operational authority that was not granted.

Provider lifecycle

• A child process exiting 0 is not completion. Completion also requires the provider's own terminal result to report success (not an error, permission denial or interruption) and no unresolved background task or later lifecycle event.
• Before cleanup or removing anything a run depends on, collect every predeclared check's actual receipt and both complete logs: the child process output and the provider event log.
• Record child exit code or signal separately from the admission outcome (whether the result is accepted as complete).
• An interrupted or incomplete run keeps its recorded outcome. A later recovery is a new boundary with its own receipts; it never relabels the earlier history.

Recovery

• Preserve task ID, provider, profile, session, worktree, SHA and evidence.
• Incomplete provider termination (timeout, crash, quota, kill) is never completion.
• No silent retry, profile switch or billing switch.
• Clean up only identified processes this delivery owns, and only after their checks are terminal.
• Revalidate drift (head SHA, base, dirty state) before resuming.
• Every blocker names the gate, its owner and the next executable step.

Project adapters

Repository-specific policy (base branch, release, issue tracking, required checks) lives in that project's own instructions or adapter skill. Adapters add policy for their repository; they never become universal exceptions to this procedure.

These procedures are guidance, not an enforcement engine; compliance is shown by the receipts, not assumed.

Resources

• references/task-contract.md — outcome contract, task card, receipt, review and final report fields.
• references/acceptance-by-task.md — proportional behavioral acceptance profiles.
• references/repository-measurements.md — optional measurement and task-grounding limits.