| name | prp-deliver |
| description | This is an experimental skill. Never use it unless the user explicitly tells you to invoke /prp-deliver. |
| argument-hint | <issue|PRD|document|plan|description|reviewed PR> [--base <branch>] [review scopes] |
Deliver to a Reviewed PR
Own one outcome from source input to a published READY TO MERGE review and green CI. Act as the sub-orchestrator for this delivery: start and steer specialist agents, preserve their handoff artifacts, and do not implement or review the work in this context.
Input: $ARGUMENTS (if absent, use the conversation).
Ownership contract
- Continue autonomously through plan, implementation, PR, review, correction, re-review, and CI. A published report is observable progress, not a pause point.
- Run stages sequentially in the delivery owner's checkout. Only one implementation or correction agent may mutate it at a time.
- Use a fresh agent for planning, a fresh agent for implementation, and a fresh independent agent for every review. Resume the implementation agent for corrections while it remains available.
- Keep the plan path, implementation report, PR, latest review report, publication URL, and agent handles. Durable artifacts—not recalled summaries—cross context windows.
- Keep a verbatim caller-decisions record containing explicit scope, standing decisions, and finding dispositions. Update it when the caller resolves a blocker and include every applicable decision in each fresh-agent prompt.
- Carry the burden of proof. Prove completion with green validation, the live PR, the complete published review, its
READY TO MERGE verdict, and green required CI.
- Stop only when a product decision, missing prerequisite primitive, inaccessible dependency, permission boundary, or repeated no-progress failure cannot be resolved autonomously. Report the exact decision or access needed and a recommendation.
- Do not merge the PR. The caller owns the merge gate.
1. Resolve and plan
Accept an issue or tracker URL, PRD, document, existing .plan.md, free-form request, conversation context, or reviewed PR.
- Review-only request or contributor PR: use
/prp-review; this is not a delivery run.
- Existing plan: use it. If it is issue-derived and lacks a verified
Plan Publication, start a planning agent with /prp-plan publish <absolute-path> first.
- Issue whose plan was already published: let
/prp-implement resolve it from source metadata and require that agent to return the persisted absolute plan path.
- Existing PR with a published PRP review: resolve its plan and implementation report, then enter correction or review without repeating completed stages.
- Otherwise start a fresh agent with this prompt:
Invoke /prp-plan against <complete input>. Additional caller context: <relevant context and explicit base, if any>. Applicable caller decisions, verbatim: <decisions or "None">. Return the absolute plan path, source and publication URLs, and any concrete blocker.
Require an absolute plan path before review and, for issue-derived plans, a verified published-plan URL.
2. Implement and open the PR
Start a fresh agent in the same checkout with this prompt:
Invoke /prp-implement on <absolute plan path or source issue> with base <explicit base, if any>. Applicable caller decisions, verbatim: <decisions or "None">. Own implementation through validation, scoped commit, PR creation, linked PRD updates, and the implementation report. Return VALIDATION: GREEN, the resolved absolute plan path, absolute report path, and PR URL; otherwise return the concrete blocker.
Do not begin review without VALIDATION: GREEN, a live PR, and the implementation report.
3. Review independently
Start a fresh agent with this prompt:
Invoke /prp-review on <PR URL or number> with scopes <requested scopes, if any>. Applicable caller decisions and scope constraints, verbatim: <decisions or "None">. Publish the complete review to GitHub. Return the verdict, canonical review-report path, verified publication URL, and any review blocker.
Require the canonical $PRP_DIR/reviews/pr-{number}-review.md, its complete GitHub publication, and the verified publication URL.
4. Correct and re-review until ready
For NEEDS FIXES, REVIEW INCOMPLETE, or any OPEN finding, send the complete review report—not an abbreviated finding list—to the implementation agent:
Continue this delivery by invoking /prp-implement in review-correction mode for <PR>. Read <plan>, <implementation report>, and the complete review at <review report>. Applicable caller decisions and finding dispositions, verbatim: <decisions or "None">. Disposition every finding under the skill's fix-now, follow-up, and decline rules. Restore VALIDATION: GREEN and update the implementation report. Commit and push only when repository changes are required; for an evidence-only disposition, prove the PR head is unchanged.
Let the implementation agent disposition the report using its plan and code context. Prefer fixing valid, narrow, low-risk findings now; track only valuable distinct outcomes, and decline speculative or directionally wrong work without creating backlog noise.
If that agent is unavailable, start a fresh correction agent with the same complete artifact bundle. After every correction or disposition, start another fresh review agent and update the canonical publication. Repeat until READY TO MERGE with every finding terminal, or a genuine blocker survives without new evidence or progress.
5. Require green CI
After READY TO MERGE, wait for every required CI check. A pending check is not green. Return a PR-caused failure and its complete evidence to the implementation agent and tell it to invoke /prp-implement in CI-correction mode, then run a fresh review against the changed head. When the repository has no required CI, rerun its authoritative local validation instead and record that evidence.
6. Return the outcome
Return the implemented outcome, absolute plan and implementation-report paths, PR URL, READY TO MERGE verdict, latest review-report path, GitHub publication URL, validation summary, and CI results. These are the proof the caller verifies before accepting the workstream. Only then suggest meaningful remaining non-blocking follow-ups. If genuinely blocked, return the same artifact bundle plus the exact blocker and recommended next action.
