Subagent-Driven Development

Execute plan by dispatching fresh subagent per task, with three-stage review after each: spec compliance → code quality → specialist lens (STRIDE/OWASP/ATAM/TEMPORAL/CPT).

Core principle: Fresh subagent per task + evidence-gated review = high quality, fast iteration.

Continuous execution: Do not pause between tasks. Execute all tasks without stopping unless: BLOCKED status unresolvable, genuine ambiguity, or all tasks complete.

Announce: "Using subagent-driven-development to execute the plan."

Pre-Check

# Confirm isolated workspace exists
git worktree list
# If not — run using-git-worktrees first

Read plan from docs/superskills/plans/ once. Extract all tasks.

Create task tracker:

• Claude Code: TodoWrite tool
• Other harnesses: maintain - [ ] checklist inline in this session
• Fallback: echo "[ ] Task N: <name>" >> /tmp/task-tracker.txt

Per-Task Loop

For each task:
  1. Dispatch implementer subagent (full task text + context)
  2. Answer questions if asked BEFORE implementation
  3. Implementer: implements, tests, commits, self-reviews
  4. Stage A: Spec compliance review (spec reviewer subagent)
     → issues found? implementer fixes → re-review → repeat until ✅
  5. Stage B: Code quality review (quality reviewer subagent)
     → issues found? implementer fixes → re-review → repeat until ✅
  6. Stage C: specialist-review (context-aware lens)
     → CRITICAL? fix immediately. HIGH? fix before next task.
     → Bug confirmed? trigger pattern-globalize before proceeding.
  7. verification-before-completion gate — evidence anchor required
  8. Mark task complete in TodoWrite

Do NOT start Stage B before Stage A is ✅. Do NOT mark task complete before verification-before-completion.

Subagent Context Templates

Implementer

TASK: [Full task text from plan]
PLAN: [Feature goal + architecture from plan header]
CODEBASE: [Relevant file structure]
CONSTRAINTS: Follow tdd-verified. No implementation before failing test.
             No completion claim without verification-before-completion.
STATUS: Report one of: DONE | DONE_WITH_CONCERNS | NEEDS_CONTEXT | BLOCKED

Spec Reviewer

TASK: Verify implementation matches spec exactly.
SPEC: [Relevant spec sections]
BASE_SHA: [before task]
HEAD_SHA: [after task]
Check: all requirements implemented (no under-build), nothing extra (no over-build).
Report: ✅ COMPLIANT or ❌ ISSUES: [list]

Quality Reviewer

TASK: Review code quality only (not spec compliance — already verified).
BASE_SHA: [before task]
HEAD_SHA: [after task]
Check: naming, structure, duplication, error handling, test quality.
Report: ✅ APPROVED or Issues by severity: CRITICAL | HIGH | MEDIUM | LOW

Stage C — Specialist Review Integration

After quality review ✅, run specialist-review on the task diff.

Signals to check:

• Auth/session/token imports → STRIDE
• User input + web surface → OWASP
• Architecture decision in diff → ATAM
• Async/await/worker patterns → TEMPORAL
• Write chain ≥ 3 components → CPT

Pass specialist lens context to the review subagent. See specialist-review for full signal taxonomy.

Pattern-Globalize Trigger

If Stage A, B, or C reveals a confirmed bug (not just style issue):

After fix is committed:
→ trigger pattern-globalize immediately
→ do NOT mark task complete until pattern-globalize has run

This is mandatory. One confirmed bug = check globally.

Handling Implementer Status

Never ignore BLOCKED. Something must change before retry.

Model Selection

After All Tasks

1. Dispatch final code review subagent (full implementation scope)
2. Invoke adr-commit — ADR Gate + PATTERN-DEBT lifecycle + merge options

Do NOT invoke finishing-a-development-branch. Use adr-commit.

Red Flags — Never

• Start on main/master without explicit user consent
• Skip spec compliance review
• Start code quality review before spec ✅
• Skip specialist-review Stage C
• Skip pattern-globalize after confirmed bug
• Mark task complete without verification-before-completion
• Dispatch parallel implementation subagents (conflicts)
• Give subagent your session context (construct exactly what they need)
• Accept "close enough" on spec compliance

Gotchas — Observed Failure Points

<!-- Populated by knowledge-compound skill after cycles where this skill underperformed --> <!-- Format: [YYYY-MM-DD] What failed | Root cause | What to do instead --> <!-- DO NOT pre-populate with speculation — real observations only --> <!-- Example: [2026-06-01] Agent skipped risk scoring for "trivial" CSS task that touched auth config | Triage gate not read carefully | Triage gate now checks file content, not just task description -->

Evidence Boundary

Subagent reports are not T1 evidence for the parent agent. Treat them as T3 until the parent verifies raw output, diff, or runtime behavior directly. Use shared/subagent-evidence.md and verification-before-completion claim grammar before marking tasks complete.