implement - SKILL.md Agent Skill

name: implement description: >- Orchestrate implementation of a spec, reads the spec, finds dependencies, plans architecture, validates plan with the current codebase, synthesize implementation plan, write code, does validation of the implementation with checkpoint asking user validation, does on round of refactoring for self-validation TRIGGER when: user says "implement spec", "implement" and passes spec file
argument-hint: [--verbose] [--iterations N] allowed-tools: Read, Glob, Bash, Write, Edit, AskUserQuestion, Skill, TaskCreate, TaskUpdate, mcp__plugin_solid-coder_pipeline__get_output_path user-invocable: true

Accepts a feature spec file and coordinates sub-skills to architect, validate, synthesize, implement, and review the feature — all in one command.

OUTPUT_ROOT: ~/.solid-coder/{project-slug}/implement-{spec-number}-<YYYYMMDDhhmmss> (computed via MCP tool)

SPEC_FILE: $ARGUMENTS[0] — path to a markdown spec file with YAML frontmatter (must contain at least a number field)
--verbose: (optional) Enable per-phase timing and implement-log.json output. Default: off.
--iterations N: (optional) Number of post-implementation refactor iterations. Default: 1. Use 0 to skip safety review.

0.1 Parse $ARGUMENTS — extract the spec file path and flags. If no spec file provided, fail with: "Input must be a spec file with frontmatter. Use /build-spec to create one."
- Set VERBOSE = true if --verbose is present, false otherwise
- Set ITERATIONS = N if --iterations N is present, 1 otherwise
0.2 Verify the file exists and ends in .md. If not, fail with the same message as 0.1.
0.3 Call mcp__plugin_solid-coder_specs__parse_spec with the spec file path. On error, stop and report the message.
0.4 Generate timestamp: run date -u +%Y%m%d-%H%M%S
0.5 Call mcp__plugin_solid-coder_pipeline__get_output_path with {operation: "implement", spec_number: "{number}"} → store result as OUTPUT_ROOT
0.6 Copy the spec file to {OUTPUT_ROOT}/spec.md
0.7 Check if a resources/ directory exists as a sibling of the spec file. If yes, copy it to {OUTPUT_ROOT}/resources/ (preserving all files — screenshots, schemas, etc.)
0.8 If VERBOSE: initialize implement-log.json at {OUTPUT_ROOT}/implement-log.json with empty phases array

1.1 Prepare a Task call:
- subagent_type: solid-coder:plan-agent
- prompt:
```
spec: {OUTPUT_ROOT}/spec.md
output: {OUTPUT_ROOT}/arch.json
```
1.2 Launch Task
1.3 If VERBOSE: capture timestamps and update implement-log.json with { "phase": "plan", "status": "success" }
1.4 If the Task failed → report: "Phase 1 (plan) failed: {error}. Run /plan manually to debug." STOP.

2.1 Prepare a Task call:
- subagent_type: solid-coder:validate-plan-agent
- prompt:
```
arch-json: {OUTPUT_ROOT}/arch.json
output: {OUTPUT_ROOT}/validation.json
```
2.2 Launch Task
2.3 If VERBOSE: capture timestamps and update implement-log.json with { "phase": "validate-plan", "status": "success" }
2.4 If the Task failed → report: "Phase 2 (validate-plan) failed: {error}. Run /validate-plan manually to debug." STOP.

3.1 Prepare a Task call:

prompt:

arch-json: {OUTPUT_ROOT}/arch.json
validation-json: {OUTPUT_ROOT}/validation.json
output: {OUTPUT_ROOT}/implementation-plan.json

3.2 Launch Task
3.3 If VERBOSE: capture timestamps and update implement-log.json with { "phase": "synthesize-implementation", "status": "success" }
3.4 If the Task failed → report: "Phase 3 (synthesize-implementation) failed: {error}. Run /synthesize-implementation manually to debug." STOP.
3.5 Read {OUTPUT_ROOT}/implementation-plan.json → extract matched_tags[] as comma-separated string. Store as MATCHED_TAGS. If empty or missing, set MATCHED_TAGS to empty string.

4.1 Prepare a Task call:

prompt:

mode: implement
plan: {OUTPUT_ROOT}/implementation-plan/
matched-tags: {MATCHED_TAGS}

4.2 Launch Task
4.3 If VERBOSE: capture timestamps and update implement-log.json with { "phase": "code", "status": "success" }
4.4 If the Task failed → report which plan items completed, which failed, and which remain. Do NOT rollback completed items. STOP.

SKIP THIS STEP MOVE TO PHASE 6

5.1 Use skill solid-coder:validate-implementation with: {OUTPUT_ROOT}
5.2 If result status == "skipped" or "approved" → proceed to Phase 6
5.3 If result status == "has_fixes":
- 5.3.1 Prepare a Task call:
  - subagent_type: solid-coder:code-agent
  - prompt:
```
mode: implement
plan: {fixes_path}
```
- 5.3.2 Launch Task
- 5.3.3 After fixes applied, ask user: "Fixes applied. Would you like to re-validate, continue to refactor, or stop?"
  - Re-validate → return to 5.1 (max 2 re-validates — after that, offer only continue or stop)
  - Continue → proceed to Phase 6 with remaining issues noted
  - Stop → STOP. Report what was implemented and what design issues remain.
5.4 If VERBOSE: capture timestamps and update implement-log.json with { "phase": "validate-implementation", "status": "success" }

6.0 If ITERATIONS == 0: skip this phase entirely — go to Phase 7.
6.1 Stage all files created/modified by Phase 4 and 5: run git add -A to capture all changes
6.2 Invoke skill solid-coder:refactor with: changes --iterations {ITERATIONS} (if VERBOSE: append --verbose)
6.3 If VERBOSE: capture timestamps and update implement-log.json with { "phase": "refactor", "status": "success" }
6.4 If refactor found violations that persist → report remaining violations to the user. This is informational — the implementation is complete.
6.5 If refactor found all code compliant → continue.

7.2 Print summary:

/implement complete
Spec:    {spec file}
Run:     {OUTPUT_ROOT}
Phases:  plan ✓ | codebase ✓ | synthesize ✓ | code ✓ | validate ✓ | review ✓

(Mark failed/skipped phases accordingly)

The orchestrator NEVER reads phase JSON outputs — it only passes paths forward between phases.
The orchestrator does NOT loop back to previous phases. The synthesizer reconciles conflicts.
The orchestrator runs phases 1-4 in sequence. Phase 6 (safety review) runs only when --iterations N is set with N > 0.
Do NOT build the project.
Do NOT deviate from the instructions. Follow them thoroughly.
ALL Task calls must run in foreground (never run_in_background: true) — background agents hit permission prompts silently and stall.