AGENT SKILL

implement-tasks

Use to implement a feature from its tasks.json with test-driven development — writes a failing test first, makes it pass, refactors, gates, and commits per task. Triggers on "/sdlc:implement-tasks {slug}", "implement {slug}", "build {slug}", "TDD {slug}", "code up the tasks for {slug}", "імплементуй {slug}", "реалізуй фічу {slug}", "напиши код за задачами {slug}". Reads docs/features/{slug}/tasks/tasks.json + the upstream artifacts, detects the repo's test/lint/vet commands stack-agnostically, builds a dependency DAG, and runs one of three modes — sequential single-agent TDD, an agent team (TeamCreate), or a dynamic Workflow — chosen from settings + DAG shape with graceful fallback. Output: committed, tested code + updated tasks/tracker.md. Hard-refuses if docs/features/{slug}/tasks/tasks.json is missing (the legacy docs/features/{slug}/tasks.json is accepted as a fallback) → run `/sdlc:break-tasks {slug}` first.

License 许可
MIT
最近更新
2026-11-12
Stars
★ 15
Forks
13

01 前言 Front Matter

name
implement-tasks
model
inherit
stage
17
agents
[ "test-author", "implementer", "refactor", "reviewer" ]
effort
medium
triggers
[ "/sdlc:implement-tasks" ]
description
Use to implement a feature from its tasks.json with test-driven development — writes a failing test first, makes it pass, refactors, gates, and commits per task. Triggers on "/sdlc:implement-tasks {slug}", "implement {slug}", "build {slug}", "TDD {slug}", "code up the tasks for {slug}", "імплементуй {slug}", "реалізуй фічу {slug}", "напиши код за задачами {slug}". Reads docs/features/{slug}/tasks/tasks.json + the upstream artifacts, detects the repo's test/lint/vet commands stack-agnostically, builds a dependency DAG, and runs one of three modes — sequential single-agent TDD, an agent team (TeamCreate), or a dynamic Workflow — chosen from settings + DAG shape with graceful fallback. Output: committed, tested code + updated tasks/tracker.md. Hard-refuses if docs/features/{slug}/tasks/tasks.json is missing (the legacy docs/features/{slug}/tasks.json is accepted as a fallback) → run `/sdlc:break-tasks {slug}` first.

02 资源文件 8

03 SKILL.md 正文

Skill: implement-tasks (SDLC stage 17 — Code)

The implementation engine — the keystone that closes the cycle. It turns tasks.json into committed, tested code through a strict TDD cycle per task — SELECT → RED → GREEN → REFACTOR → GATE → COMMIT — and orchestrates that cycle in one of three modes (sequential / agent-team / dynamic-workflow) picked by an unambiguous decision tree. Everything is stack-agnostic: the test, lint, and vet commands are detected, never hard-coded.

This file is the spine. Each step delegates to a file in references/.

Owner

Tech Lead drives; the engine runs the cycle. Four subagents ship with the plugin: test-author (sdlc:test-author, RED), implementer (sdlc:implementer, GREEN), refactor (sdlc:refactor, REFACTOR — isolated, tidies while green), reviewer (sdlc:reviewer, read-only review). GATE + COMMIT stay on the engine, which owns the settings and the require_integration policy.

Inputs

  • <slug> — feature slug.
  • Gate (hard refuse): docs/features/<slug>/tasks/tasks.json (legacy fallback: docs/features/<slug>/tasks.json, used only when the first is absent). Both missing → «run /sdlc:break-tasks <slug> first». Greenfield skeleton: <slug> = _scaffold (written by map-architecture).
  • Read for context (the agents read these directly, not via paraphrase): PRD.md (AC), data-model.md + the staged migrations under docs/features/<slug>/migrations/ (a layer: migration task promotes these into the live migrations tree — see ./references/inputs.md), contracts/openapi.yaml, test-plan.md, sad.md, Accepted adr/.
  • Settings: .claude/sdlc.local.md (auto-created with documented defaults if absent — normally by write-prd at the backbone start; implement-tasks creates it too if you jump straight here) → ./references/settings.md.

Protocol

  1. Preconditions. Resolve the plan path — docs/features/<slug>/tasks/tasks.json, else the legacy docs/features/<slug>/tasks.json (warn that it's legacy) — verify it parses; load the upstream artifacts list. Detail → ./references/inputs.md.
  2. Settings. Read .claude/sdlc.local.md; if absent, auto-create it with the documented defaults (frontmatter + the «What each key does» body, self-documenting) and patch .gitignore (.claude/*.local.md, .worktrees/) — the same template write-prd writes. → ./references/settings.md.
  3. Detect commands. Run the stack-agnostic cascade (settings override → Makefile → package scripts → language manifests → Docker probe for the integration tier) to resolve unit / integration / lint / vet commands. Print what was detected. → ./references/command-detection.md.
  4. Build the DAG. Parse tasks.json, validate deps is acyclic, topologically sort into phases (Kahn). Compute task_count, longest_chain, parallel_width. Mark serialization lanes (layer: migration; tasks with overlapping files_hint). layer: ui is not auto-serialized — UI tasks parallelize unless their files_hint overlaps.
  5. Pick the mode. Run the decision tree (below; full form → ./references/decision-tree.md). Apply the guards.
  6. Generate the run-plan. Sequential → an ordered task list. Team → a shared TaskList with the full task text in each body. Workflow → a generated Workflow script (DAG → Kahn phases → fan-out pipeline). → ./references/team-exec.md / ./references/workflow-exec.md.
  7. Banner. Print the active mode and the settings that drove it: mode=<…> tdd=<…> isolation=<…> parallel=<n> integration=<…>. The user sees exactly how the engine will behave before it acts.
  8. Execute in the chosen mode. Every task runs the TDD cycle → ./references/tdd-loop.md. A layer: migration task first promotes its staged migration(s) (docs/features/<slug>/migrations/<NN>_*) into the live migrations tree — assigning the real sequence number / timestamp per the repo's convention, in ordinal order — then applies + reverts them; detail → ./references/inputs.md.
  9. Per-task gate + commit. RED runs in sdlc:test-author, GREEN in sdlc:implementer, REFACTOR in the isolated sdlc:refactor agent. After GREEN+REFACTOR the engine runs the gate — unit + (integration if available) + lint + vet must be clean — then commits task-scoped with trailers SDLC-Task: <id> and SDLC-AC: <id> (one per satisfied AC). GATE + COMMIT are the engine's, not any agent's. Update tasks/tracker.md → done.
  10. Summary + hand off. Report covered AC, commits made (with SDLC-Task trailers), any task dropped/blocked, and the per-task gate results. Then emit the stage-handoff block per ../_shared/handoff.md — What I did (covered AC, commits with SDLC-Task trailers, gate results) + Review (the committed diff + tasks/tracker.md) + Run next (/clear, then /sdlc:review-feature <slug> — a clean-context pass over the whole diff), then /sdlc:ship-feature <slug>. In team mode the reviewer may also run per-task, but the authoritative independent review of the whole change lives in the review-feature skill — implement-tasks does not self-certify.

Decision tree (compact)

parallel_eligible := isolation==worktree AND max_parallel>1 AND parallel_width>=2
                     AND (size in {M,L,XL} OR task_count>=4)

if team_mode AND parallel_eligible:                          → AGENT TEAM (TeamCreate) over the DAG
elif workflow_mode=="auto" AND parallel_eligible AND Workflow-available: → DYNAMIC WORKFLOW
else:                                                        → SEQUENTIAL single-agent TDD (topo order)

Guards (apply before dispatch): team_mode but not eligible → warn + downgrade to the next mode. max_parallel>1 with isolation: inplace → clamp parallel to 1 (no two agents edit one tree). workflow_mode: off → never generate a Workflow. tdd: false → skip RED (warn loudly — you lose the safety net). require_integration: always but Docker absent → BLOCK before dispatch; auto → run unit-only and mark integration NON-red; never → skip the integration tier. Full table → ./references/decision-tree.md. Graceful degrade: if Workflow/TeamCreate is unavailable at runtime, fall through to sequential.

TDD cycle (per task)

SELECT → RED → GREEN → REFACTOR → GATE → COMMIT — RED in sdlc:test-author, GREEN in sdlc:implementer, REFACTOR in the isolated sdlc:refactor, GATE + COMMIT on the engine. The RED step is load-bearing: write the test first, run it, and classify the first run — GOOD red (assertion fails / unimplemented) vs BAD red (the test itself won't compile → fix the test) vs false-pass (green immediately → the test is too weak, strengthen it) vs NON-red (skipped because Docker is absent → governed by require_integration, counts as neither red nor green). Quote the failing line before writing any production code. Escalation on persistent red → ./references/escalation.md: more-capable model → retry → split the task → if the test encodes a wrong AC, ask a human (never weaken the test) → rollback to the last green. stop_on_red decides halt vs drop-and-continue (dependents auto-block).

Definition of Done

  • Every task in docs/features/<slug>/tasks/tasks.json is either committed (test-first, gate-clean, SDLC-Task/SDLC-AC trailers) or explicitly reported as dropped/blocked with the reason.
  • Unit gate green; integration green where available (or NON-red recorded with the policy reason); lint + vet clean per the detected commands.
  • The active mode + settings were printed in the banner before execution.
  • tasks/tracker.md reflects final status; the summary reports the gate results and hands off to review-feature (the independent review gate) — implement-tasks does not self-certify the whole change.

Anti-patterns

  • Code before the test. RED first, always (unless tdd: false, which warns).
  • Weakening a test to make it pass. If the AC is wrong, ask a human and fix the AC; never edit the test to be less strict.
  • Skipping the RED classification. A false-pass that looks green hides a useless test.
  • Parallel agents editing one working tree. Parallelism requires worktree isolation — the guard clamps it.
  • Committing with a red or skipped gate and calling it done. A NON-red integration tier must be labelled, not hidden.
  • Spawning a team for <4 tasks — coordination overhead exceeds the gain; the eligibility check forbids it.
  • Claiming integration passed when Docker was absent. Report NON-red honestly.

References & template

inputs.md · settings.md · command-detection.md · decision-tree.md · tdd-loop.md · team-exec.md · workflow-exec.md · escalation.md — all in ./references/.

04 同仓库更多 Skills

genkovich
genkovich/agentic-engineering-course-public

release

Run the whole local release pipeline in order, stopping at each human gate. Triggers on "run the release", "prepare a release", "/release", "зроби реліз", "підготуй реліз". Drives the four stages back to back — bump-version → curate-changelog → release-notes → check-docs-drift — pausing after each so the human reviews the diff before the next stage. Prepares everything in the working tree; never tags, commits, or pushes. The tag push (which fires CI) stays the human's action.

★ 15
genkovich
genkovich/agentic-engineering-course-public

release-notes

Draft partner-facing release notes from the same commits as the changelog. Triggers on "draft release notes", "release notes for the next version", "/release-notes", "нотатки релізу", "release notes партнеру". Same input as curate-changelog, different audience and tone: a short, benefit-first narrative for a user or partner — "what's new for you" — not a categorized engineer's list. Keeps only changes a user actually notices; drops internal refactors and tooling. Prints the notes to the chat to seed a GitHub Release description or a partner announcement. Writes no file.

★ 15
genkovich
genkovich/agentic-engineering-course-public

curate-changelog

Curate the [Unreleased] section of docs/CHANGELOG.md from git history. Triggers on "curate the changelog", "update the changelog", "/curate-changelog", "онови changelog", "курування журналу". Reads git log since the last tag, FILTERS out what a reader never feels (test/ci/chore/merge), GROUPS commits about one feature into one bullet, and REWRITES technical subjects into plain language under the six Keep a Changelog categories (Added/Changed/Deprecated/ Removed/Fixed/Security). The result must have FEWER lines than the raw log — that is the proof of curation, not a dump. Leaves the edit in the working tree; never commits or pushes.

★ 15
genkovich
genkovich/agentic-engineering-course-public

codify-rule

Turn a repeated mistake into a durable, path-scoped rule in .claude/rules/. Triggers on "codify this rule", "make a rule so this doesn't happen again", "/codify-rule", "закодифікуй правило", "петля помилка-правило". Use only when the same mistake showed up TWICE (twice = a pattern, once = noise). Names the wrong shape and the right shape with one short example of each, scopes the rule to the paths that matter, and leaves the edit in the working tree so the next agent on this repo inherits the fix. Edits only .claude/rules/; never commits or pushes.

★ 15
genkovich
genkovich/agentic-engineering-course-public

check-docs-drift

Find and fix where docs/api.md fell behind the code. Triggers on "check for doc drift", "is the API doc up to date", "/check-docs-drift", "перевір дрейф документації", "docs drift". Runs the deterministic scripts/check-docs-drift.py (parses the public NoteBook methods with ast, compares to what docs/api.md documents — NO LLM finds the gap), reports the drift, and PROPOSES the doc edit that closes it. The gap is found by parsing; the agent writes the missing documentation in the doc's existing style. Edits only docs/; never commits or pushes.

★ 15
genkovich
genkovich/agentic-engineering-course-public

bump-version

Decide and apply the next semver for a release. Triggers on "bump the version", "what version is next", "/bump-version", "простав версію", "яка наступна версія". Reads the next number from the deterministic scripts/next-version.sh (NEVER guesses it), explains in plain language WHICH part moved and why (e.g. "MINOR, because a feat landed: filter notes by tag"), edits the version line in pyproject.toml, and proposes the matching `git tag vX.Y.Z`. The number is decided by a rule; the agent explains and applies — the human confirms the tag. Never pushes a tag itself.

★ 15