skills/dev-implement/SKILL.md
(Implementation Orchestrator) Execute tasks.md by dispatching subagents per milestone — parallel where possible, sequential where dependent. Review per milestone, escalate blockers.
npx skillsauth add dvduongth/skills dev-implementInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
| Field | Content |
|-------|---------|
| Role | Implementation Orchestrator |
| Goal | Execute confirmed tasks.md by dispatching subagents per task, organized by milestone. Produce working code that satisfies all mapped ACs. |
| Use when | plans/<feature_id>/tasks.md exists AND user has confirmed task breakdown (FIXED GATE at STATE-6 passed) |
| Constraints | Do NOT modify specs, design docs, or task definitions. If a design gap is found during implementation, log it as DESIGN_GAP in progress.md and continue if non-blocking. Do NOT skip milestones — execute in order (M0→M1→M2...). |
| Anti-patterns | Implementing all tasks in a single agent context (will overflow); skipping review gates; modifying tasks.md during execution; implementing without reading relevant ACs |
| Quality standard | Every task must pass Orchestrator review (95/100 threshold). Every milestone must pass integration review. All mapped ACs must be covered. |
| Output format | Vietnamese status updates; English code comments and commit messages |
Required reads (before execution):
plans/<feature_id>/tasks.md — the task breakdown to executespecs/<feature_id>/acceptance-criteria.md — AC mapping for each taskspecs/<feature_id>/requirements.md — requirement contextspecs/<feature_id>/use-cases.md — use case contextdocs/design-docs/<feature_id>/design-doc.md — design decisionsOptional reads:
docs/runs/<feature_id>/*_analyze/analysis-report.md — architecture analysisdocs/dev/<feature_id>/INDEX.md — diagram indexdocs/runs/<feature_id>/pipeline-state.json — verify STATE-6python tools/scaffold_run.py <feature_id> implement
Creates run folder + input.md. Note the path.plans/<feature_id>/tasks.md exists. If not → error: "No tasks.md found. Run /dev_tasks first."specs/<feature_id>/acceptance-criteria.md exists. If not → error: "No specs found. Run /dev_specs first."--milestone flag → filter to that milestone only.--task flag → execute that single task only (useful for retries).depends_on list, mapped ACs.progress.md:
# Implementation Progress — <feature_id>
## Overall: 0/N tasks complete (0%)
### M0 <name> ⏳ (0/K)
- [ ] TASK-001: <description> — QUEUED
- [ ] TASK-002: <description> — QUEUED
...
For each milestone in order (M0 → M1 → M2 → ...):
From the dependency DAG for this milestone:
For each READY task, build a context packet:
1. Task description (from tasks.md)
2. Mapped ACs (filtered from acceptance-criteria.md)
3. Related requirements (filtered from requirements.md — match REQ IDs in ACs)
4. Related design doc sections (match keywords from task description)
5. Existing code context (if task modifies existing files — read them first)
Dispatch strategy:
isolation: "worktree" if available)Each subagent prompt:
You are implementing a single task for feature <feature_id>.
## Task
<task description from tasks.md>
## Acceptance Criteria to satisfy
<filtered ACs>
## Requirements context
<filtered requirements>
## Design context
<relevant design doc sections>
## Existing code
<file contents if modifying existing files>
## Instructions
1. Use /orchestrator skill to implement this task
2. Follow all project conventions (CLAUDE.md)
3. Ensure all mapped ACs are testable
4. Run gitnexus_impact before modifying any existing symbol
For each completed task:
progress.md with result| Result | Action |
|--------|--------|
| Review score ≥ 95/100 | PASS — mark task DONE |
| Review score < 95/100, retries < 3 | Re-dispatch with reviewer feedback |
| Review score < 95/100, retries ≥ 3 | Mark ESCALATED, log in progress.md, continue other tasks |
| Task discovers design gap | Log DESIGN_GAP in progress.md; if non-blocking continue, if blocking pause milestone |
| Subagent crash/timeout | Mark FAILED, log error, continue other tasks |
After all tasks in milestone complete (or escalated):
milestone-M{N}-report.md:
# Milestone M{N} Report — <feature_id>
## Tasks: K/K complete
| Task | Status | Review Score | Files Changed |
|------|--------|-------------|---------------|
| TASK-001 | DONE | 97/100 | 3 files |
## AC Coverage
| AC | Status | Verified by |
|----|--------|------------|
| AC-001 | PASS | TASK-001 |
## Issues
- (none) or list of DESIGN_GAPs, escalated tasks
After all milestones complete:
Write implementation-summary.md:
# Implementation Summary — <feature_id>
## Result: N/N tasks complete, M/M ACs covered
## Files Changed
| File | Action | Task |
|------|--------|------|
| path/to/file.gd | Created | TASK-001 |
## Milestones
| Milestone | Tasks | Status |
|-----------|-------|--------|
| M0 | 3/3 | ✅ |
## Blockers / Escalations
- (list any unresolved items)
## DESIGN_GAPs Found
- (list any design gaps discovered during implementation)
Update progress.md with final status.
Print summary:
✅ Implementation complete for <feature_id>
Tasks: N/N complete
ACs: M/M covered
Files changed: X (Y new, Z modified)
Blockers: K unresolved
Output: docs/runs/<feature_id>/<ts>_implement/
If any tasks are ESCALATED or DESIGN_GAPs are blocking:
⚠️ Implementation completed with issues:
- N tasks escalated (need human review)
- M design gaps found (may need spec updates)
Review progress.md for details.
development
Hiểu sâu bất kỳ codebase nào đã được GitNexus index — architecture, execution flows, symbol relationships, blast radius. Dùng khi hỏi về codebase architecture, symbol context, impact analysis, hoặc index status.
tools
Search GIF providers with CLI/TUI, download results, and extract stills/sheets.
documentation
Fetch GitHub issues, spawn sub-agents to implement fixes and open PRs, then monitor and address PR review comments. Usage: /gh-issues [owner/repo] [--label bug] [--limit 5] [--milestone v1.0] [--assignee @me] [--fork user/repo] [--watch] [--interval 5] [--reviews-only] [--cron] [--dry-run] [--model glm-5] [--notify-channel -1002381931352]
tools
Gemini CLI for one-shot Q&A, summaries, and generation.