# Implement Feature > Implement a whole feature as one PR through ticket waves on an integration branch. Use when implementing a spec and its tickets, or an explicit ticket set, end-to-end. - Skill: `sleeplessv/implement-feature` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add sleeplessv/implement-feature` - Raw SKILL.md: https://api.skillmd.com/api/skills/sleeplessv/implement-feature/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Integrations & APIs - Author: sleeplessv (https://skillmd.com/u/sleeplessv) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/sleeplessv/implement-feature --- # Implement feature Deliver the selected tickets as one reviewed feature PR into the default branch. Use `orchestrator-mode` for this run and `implement-ticket` for each ticket. Delegate git, tracker, file, and test work; keep bulk evidence in handoff artifacts. ## Scope and contracts - The **work-set** is the selected tickets. An explicit list controls implementation and review scope; a larger spec provides context, not additional completion obligations. - The **integration branch** collects verified ticket work. Keep its name on resume. Follow environment branch conventions, using `spec--` or a descriptive slug. Recognize prior `feat/spec-*`, `feat/prd-*`, `feat/ticket-*`, and `feat/issue-*` names. - A **wave** contains tickets whose blockers are satisfied in the current integration tip. Dispatch within available capacity, allocating separate external resources or serializing checks that share ports, databases, or output paths. Each ticket has an isolated worktree, including single-ticket waves. A setup worker can create it manually in a writable location when native isolation is unavailable. If isolation cannot be established, report the blocker. - Pin each wave's immutable `base_sha`; return and verify it. The integration branch name is destination metadata, not a substitute for the SHA. Preserve the **original baseline** SHA and failures across resumes; the wave base advances only after verification. - The coordinator owns ticket eligibility and lifecycle through its tracker worker. Implementing workers do not claim tickets, comment, relabel issues, or create per-ticket PRs. Follow implement-ticket's [Ticket lifecycle comments](../implement-ticket/SKILL.md#ticket-lifecycle-comments) for every attempt. Reuse the user's comment authorization without asking again. Existing authorization and configured permissions govern external actions. - Pass absolute artifact paths, relevant decisions, and compact return summaries. Reuse loaded skills and honor explicit prose and testing guidance. One worker owns each ticket; reserve capacity for leaf work and review children instead of spawning helpers inside it. ## Workflow ### 0. Resolve the work-set and prepare the tree Dispatch setup using [Work-set resolution](references/reference.md#work-set-resolution). It resolves spec-only, tickets-only, or combined inputs, checks every selected ticket's eligibility, snapshots bodies and criteria, and builds the dependency graph. The setup return contains a compact work-set/index and proposed branch or resume state. Announce the work-set before branch creation, then continue setup with [Tree preparation](references/reference.md#tree-preparation). A fresh one-ticket work-set uses one solo `implement-ticket` dispatch. A resumed integration branch keeps its feature workflow even if one ticket remains. Zero tickets without an existing branch means nothing to implement; a completed existing integration branch still needs its gates. Surface dependency cycles, mismatched parent specs, unresolved outside blockers, ambiguous ownership, or required overlapping dirty inputs. Preserve unrelated dirty work through isolation. Infer a short slug from the selected scope when no spec is present. Done when the work-set is announced, dependencies resolved, integration checkout and SHAs known, and the original baseline, ticket snapshots, and durable run record are available by path. ### 1. Implement in waves 1. Select the ready tickets and pin `base_sha` to the verified integration tip. A setup worker records each assigned worktree and task-created branch before implementation. Have the tracker worker claim eligible tickets and post their start comments before dispatch. 2. Dispatch one worker per ticket within capacity. Supply its snapshot, handoff and run-record paths, `base_branch`, `base_sha`, original baseline, resource allocation, `open_pr: false`, and any evidence-backed `resume_branch`. Tell it to read implement-ticket's [Orchestrated dispatch contract](../implement-ticket/references/orchestrated.md). Include the start-comment URL, or the recorded posting failure. On each worker's return or interruption, have the tracker worker post its stop comment before integration or redispatch. 3. A dedicated integration worker follows [Wave integration](references/reference.md#wave-integration). It checks the result SHAs, merges successes, tests the merged tree, records decisions, and reports actual merge state. Any conflict resolution or subsequent fix goes through a separate verification worker using [Post-resolution tests](references/reference.md#post-resolution-tests). Repeat fix then verification until green or a concrete blocker. Release dependants and clean up only after the tested integration tip is preserved. An integrator that authored no code can supply the independent wave verification; the final gates still check the whole feature. 4. A failed ticket keeps its WIP. Record its dependants as blocked, continue independent tickets, and report the failure. Accept `already_satisfied` only with evidence for every criterion at the pinned base. This marks behavior resolved, not WIP safe to delete. Report what merged and what is next once per wave, including unresolved preservation or cleanup failures. Done when every selected ticket is satisfied or recorded as failed or blocked downstream. Every started attempt must also have a stop-comment URL or an explicitly reported posting failure. ### 2. Integration gates Use fresh workers to verify the fixed integration SHA and review its diff from the recorded PR base. Verify runs configured lint/type checks, tests, and the applicable artifact or runtime check. Review checks standards and the selected tickets' criteria, using the full spec only as context. See [Integration review](references/reference.md#integration-review). Schedule according to total capacity. On a four-slot runtime, finish Verify before starting a `code-review` worker that needs two children, or flatten its review axes into root-owned workers. Use `code-review` when available; otherwise dispatch equivalent independent standards and criteria reviews. No verifier or reviewer mutates the checkout while another reads its fixed SHA. Send actionable in-scope findings to one fix worker, then rerun the affected gates on the new SHA. Unselected requirements and adjacent problems remain reported, not implemented. Repeated failures require diagnosis or another approach; stop only at a concrete blocker. For a partial work-set, run Verify on preserved integration work and report the incomplete state without opening a PR. Done when both gates pass on the preserved integration head and every selected ticket is satisfied. ### 3. Verification plan and feature PR One worker authors and executes the [Verification plan](references/reference.md#verification-plan), then creates or updates the [Feature PR](references/reference.md#feature-pr). Pass the verified SHA, selected-scope evidence, run record, and verification report paths. A newly discovered defect returns to the fix and gate loop. Refresh tracker coverage before claiming the entire spec is complete. Open one ready-for-review PR with explicit base and head, Summary, Test plan, Verification plan, and evidence-backed closing lines. Never merge it. Report its URL and any unexecuted checks. If publication is blocked, preserve the completed branch and draft body and report that limitation. ## Failure and resume Keep original baseline evidence, decisions, ownership records, and WIP across invocations. An interrupted merge is not proof the tree equals the last pushed tip. Return actual HEAD, merge state, integrated tickets, dirty paths, and preservation status. Push only verified integration states, or explicitly authorized WIP to a separate recovery branch. Resume from that record; never relabel a current integration regression as a pre-existing baseline failure. Before pausing, cancelling, or handing off the feature, stop active ticket workers and post each started attempt's stop comment. Reconcile missing comments from interrupted runs before resuming.