implement

name	implement
description	Executes tasks from the current track's plan using TDD workflow (red-green-refactor). Implements one task at a time, commits after each, runs three-stage review at phase boundaries, and tracks progress in plan.md. Use when the user asks to implement the next task, start coding, continue a plan, run test-driven development, or says 'start implementing'.

Implement Track

Implement tasks from the active track's plan following the TDD workflow.

Red Flags - STOP if you're:

• Implementing without an approved spec and plan
• Skipping TDD cycle when workflow.md has TDD enabled
• Marking a task [x] without fresh verification evidence
• Batching multiple tasks into a single commit
• Proceeding past a phase boundary without running the three-stage review
• Writing production code before a failing test (when TDD is strict)
• Assuming a test passes without actually running it

Verify before you mark complete. One task, one commit.

Constraints

Draft skills are designed for single-agent, single-track execution. Do not run multiple Draft commands concurrently on the same track.

---

Step 1: Load Context

1. Find active track from draft/tracks.md (look for [~] In Progress or first [ ] track)
2. Read the track's spec.md for requirements
3. Read the track's plan.md for task list
4. Read draft/workflow.md for TDD and commit preferences
5. Read draft/tech-stack.md for technical context
6. Read draft/guardrails.md (if exists) for hard guardrails and learned conventions
7. Check for architecture context:
   • Track-level: draft/tracks/<id>/architecture.md
   • Project-level: draft/.ai-context.md (or legacy draft/architecture.md)
   • If either exists → Enable architecture mode (Story, Execution State, Skeletons)
   • If neither exists → Standard TDD workflow
8. Load production invariants (if draft/.ai-context.md exists):
   • Read the ## INVARIANTS section (and ## CONCURRENCY if present)
   • Identify which invariants reference files this task will modify (same file or same module)
   • Keep matching invariants as active constraints for this task — these govern code generation, not just review
   • If invariants reference lock ordering, fail-closed behavior, or data integrity rules: these are non-negotiable during implementation
9. Load graph context (if draft/graph/schema.yaml exists):
   • Read draft/graph/hotspots.jsonl — check if any files this task will modify appear as hotspots
   • If modifying a hotspot file (high fanIn), warn: "This task modifies {file} (fanIn={N}). Changes here affect many downstream files. Consider running a graph impact query."
   • Read draft/graph/modules/<module>.jsonl for the module(s) being modified — gives file-level dependency context
   • See core/shared/graph-query.md for on-demand query subroutines (callers, impact)
10. Update the track's entry in draft/tracks.md from [ ] to [~] In Progress

If no active track found:

• Tell user: "No active track found. Run /draft:new-track to create one."

Architecture Mode Activation:

• Automatically enabled when .ai-context.md or architecture.md exists (file-based, no flag needed)
• Track-level architecture.md created by /draft:decompose
• Project-level .ai-context.md created by /draft:init

Step 1.5: Readiness Gate (Fresh Start Only)

Skip if: Any task in plan.md is already [x] — the track is in progress, this check has already passed.

Run once, before the first task of a new track:

AC Coverage Check

For each acceptance criterion in spec.md:

• Verify at least one task in plan.md references or addresses it
• If an AC has no corresponding task, flag it: "⚠️ AC: '[criterion]' has no task in plan.md"

Sync Check (if .ai-context.md exists)

Compare the synced_to_commit values in the YAML frontmatter of spec.md and plan.md.

• Skip if either file has no YAML frontmatter or no synced_to_commit field (quick-mode tracks omit it).
• If they differ: "⚠️ Spec and plan were synced to different commits — verify they are still aligned."

Result

Issues found: List them, then ask:

Readiness issues found (see above). Proceed anyway or update first? [proceed/update]

• proceed → add a ## Notes entry in plan.md listing the issues, then continue to Step 2
• update → stop here and let the user refine spec or plan before re-running

No issues: Print Readiness check passed. and continue to Step 2.

Step 1.7: Testing Strategy Loading

Before starting TDD cycle for the first task:

1. Check for testing strategy:
   • Track-level: draft/tracks/<id>/testing-strategy.md
   • Project-level: draft/testing-strategy.md or draft/testing-strategy-latest.md
2. If found: load coverage targets, test boundaries, and strategy into TDD context
3. If not found and TDD is enabled: suggest "Run /draft:testing-strategy to define test approach"

Bug Track Test Guardrail

If track type is bugfix (from metadata.json):

BEFORE writing any test file:
  ASK: "This is a bug fix track. Want me to write tests as part of the fix? [Y/n]"
  If declined: skip TDD cycle, note in plan.md: "Tests: developer-handled"

Step 2: Find Next Task

Scan plan.md for the first uncompleted task:

• [ ] = Pending (pick this one)
• [~] = In Progress (resume this one)
• [x] = Completed (skip)
• [!] = Blocked (skip - requires manual intervention)

IMPORTANT: If blocked task found, notify user:

• "Task [task description] is marked [!] Blocked"
• Show the blocked task details and recovery message
• "Resolve the blockage manually before continuing implementation"
• Do NOT attempt to implement blocked tasks

If resuming [~] task, check for partial work.

Step 2.5: Write Story (Architecture Mode Only)

Activation: Only runs when .ai-context.md or architecture.md exists (track-level or project-level).

When the next task involves creating or substantially modifying a code file:

1. Check if file already has a Story comment - If yes, skip this step
2. Skip for trivial tasks - Config files, type definitions, simple one-liners
3. Write a natural-language algorithm description as a comment block at the top of the target file

Story Format

// Story: [Module/File Name]
//
// Input:  [what this module/function receives]
// Process:
//   1. [first algorithmic step]
//   2. [second algorithmic step]
//   3. [third algorithmic step]
// Output: [what this module/function produces]
//
// Dependencies: [what this module relies on]
// Side effects: [any mutations, I/O, or external calls]

Adapt comment syntax to the language (# for Python, /* */ for CSS, etc.).

CHECKPOINT (MANDATORY)

STOP. Present the Story to the developer for review.

• Developer may refine, modify, or rewrite the Story
• Do NOT proceed to execution state or implementation until Story is approved
• Developer can say "skip" to bypass this checkpoint for the current task

See core/agents/architect.md for story writing guidelines.

---

Step 3: Execute Task

Step 3.0: Design Before Code (Architecture Mode Only)

Activation: Only runs when .ai-context.md or architecture.md exists (track-level or project-level). Skip for trivial tasks - Config updates, type-only changes, single-function tasks where the design is obvious.

3.0a. Execution State Design

Study the control flow for the task and propose intermediate state variables:

1. Read the Story (from Step 2.5) to understand the Input -> Output path
2. Study similar patterns in the existing codebase
3. Check .ai-context.md Data Lifecycle — Align execution state with documented state machines (valid states/transitions), storage topology (which tier data targets), and data transformation chain (shape changes at boundaries)
4. Check .ai-context.md Critical Paths — Identify where this task sits in documented write/read/async paths. Note consistency boundaries and failure recovery expectations.
5. Propose execution state: input state, intermediate state, output state, error state

Present in this format:

EXECUTION STATE: [Task/Module Name]
─────────────────────────────────────────────────────────
Input State:
  - variableName: Type — purpose

Intermediate State:
  - variableName: Type — purpose

Output State:
  - variableName: Type — purpose

Error State:
  - variableName: Type — purpose

CHECKPOINT (MANDATORY): Present execution state to developer. Wait for approval. Developer may add, remove, or modify state variables. Developer can say "skip" to bypass.

3.0b. Function Skeleton Generation

Generate function/method stubs based on the approved execution state:

1. Create stubs with complete signatures (all parameters, return types)
2. Include a one-line docstring describing purpose and when it's called
3. No implementation bodies — use // TODO, pass, unimplemented!(), etc.
4. Order functions to match control flow sequence
5. Follow naming conventions from tech-stack.md

CHECKPOINT (MANDATORY): Present skeletons to developer. Wait for approval. Developer may rename functions, change signatures, add/remove methods. Developer can say "skip" to bypass.

See core/agents/architect.md for execution state and skeleton guidelines.

---

Step 3.0c: Production Robustness Patterns (REQUIRED)

Applies to all code generation — architecture mode or not. These patterns are generation directives, not a post-hoc checklist. Apply them while writing code, not after.

When your implementation hits any of these triggers, use the corresponding pattern. Do not write code that violates these and plan to "fix it later."

Atomicity

Trigger	Required Pattern
Multi-step state mutation (DB + memory, multiple records)	Wrap in transaction or try/finally with rollback on failure
File write	Write to temp file + atomic rename to target path. Never write directly to the target.
DB write paired with in-memory state update	DB-first: persist to DB, update memory only on DB success. Never update memory optimistically.
Resource acquisition (locks, file handles, connections, capital)	Release in finally / defer / RAII — never rely on happy-path-only cleanup

Isolation

Trigger	Required Pattern
Method mutates shared/instance state	Acquire the class's or module's existing lock before mutation
Lifecycle operations (start/stop/reset/reconnect)	Use a dedicated lifecycle lock, separate from data locks
Returning internal state to callers	Return a deep copy or frozen snapshot — never a mutable reference to internal state
Acquiring a second lock while holding one	Follow documented lock ordering. If no ordering exists, do not nest locks — restructure to acquire sequentially.
DB I/O while holding a state lock	Move DB I/O outside the lock scope. Lock only the in-memory mutation, not the I/O.

Durability

Trigger	Required Pattern
Critical state that must survive crashes	Ensure state is recoverable from DB/disk alone — no reliance on in-memory-only state for recovery
Async DB write (fire-and-forget)	Await the write. Check return value or propagate exceptions. No fire-and-forget on data persistence.
Event log / audit trail / fill history	Use append-only pattern where specified by architecture

Defensive Boundaries

Trigger	Required Pattern
External numeric data used in arithmetic	Guard with isFinite() / isnan() / equivalent before any calculation
External API/webhook response consumed	Validate expected fields exist and have correct types before accessing nested properties
SQL query with dynamic values	Parameterized queries only — zero string interpolation for values
Dynamic column names, table names, or identifiers in SQL	Validate against an explicit allowlist — never pass user-controlled strings as identifiers

Idempotency

Trigger	Required Pattern
Operation that may be retried (network calls, queue consumers, webhook handlers)	Use a dedup key (UUID, request ID, fill ID) — check-before-write or upsert
State transition (status changes, lifecycle events)	Validate the transition is legal from the current state. Reject terminal→terminal transitions.
Alert / notification emission	Dedup on (alert_type, entity_id, time_window) to prevent re-firing on retries

Fail-Closed

Trigger	Required Pattern
Error path or exception handler that determines access/action	Default to the safe/restrictive/deny state — never default to permissive on error
Missing data, null, or undefined where a decision depends on it	Treat as deny/reject/skip — not as allow/proceed
Config or feature flag missing/unparseable	Use the restrictive default — system runs in safe mode, not open mode

Resilience

Trigger	Required Pattern
Any retry logic	Exponential backoff with jitter — never fixed-interval or immediate retries. Prevents retry storms.
Cache population under high concurrency	Cache stampede prevention: use probabilistic early expiration or request coalescing to prevent thundering herd
External dependency call (HTTP, RPC, DB to external service)	Circuit breaker pattern: track failure rate, open circuit on threshold, allow periodic probes to recover
Non-critical dependency failure	Graceful degradation: return cached/default/partial result rather than failing the entire request

Enforcement: These patterns override convenience. If following a pattern makes the code more verbose, that's correct — the verbosity is the safety. If a pattern is genuinely N/A for the current task (e.g., no DB in a pure utility function), skip it — only apply relevant patterns.

If project invariants were loaded in Step 1: Cross-reference them here. Project-specific invariants (lock ordering, concurrency model, consistency boundaries) take precedence over these general patterns when they conflict.

---

Step 3.1: Implement (TDD Workflow)

For each task, follow this workflow based on workflow.md. If skeletons were generated in Step 3.0b, fill them in using the TDD cycle below.

Characterization Testing (Refactoring Existing Code Without Tests)

When refactoring code that lacks tests, write characterization tests first to capture current behavior as a baseline. Identify seams (interfaces for test doubles, swappable imports), record actual outputs for representative inputs, then proceed with the TDD cycle for new behavior.

If TDD Enabled:

Iron Law: No production code without a failing test first.

3a. RED - Write Failing Test

1. Create/update test file as specified in task
2. Write test that captures the requirement
3. RUN test - VERIFY it FAILS (not syntax error, actual assertion failure)
4. Show test output with failure
5. Announce: "Test failing as expected: [failure message]"

Test Quality Checklist (REQUIRED for every test):

• No shared mutable state between test cases — each test sets up its own state
• Assertion density: every test must have at least one meaningful assertion (not just assertTrue(true))
• No logic in tests: no conditionals, loops, or try/catch in test code — tests should be trivially readable
• DAMP over DRY: prefer descriptive and meaningful test names and setup over deduplication
• Test behavior, not implementation: verify observable outcomes, not internal method calls
• One behavior per test: each test should verify exactly one logical behavior
• Reference: Google SWE Book Ch. 12, Google Testing Blog "Test Behavior, Not Implementation"

Property-Based Testing Checkpoint: After writing example-based tests, consider property-based tests for pure functions (algebraic properties, round-trip serialization, sort invariants). Not mandatory — skip if properties are not obvious.

3b. GREEN - Implement Minimum Code

1. Write MINIMUM code to make test pass (no extras)
2. RUN test - VERIFY it PASSES
3. Show test output with pass
4. Announce: "Test passing: [evidence]"

Observability Prompts (consider during implementation): Structured logging at decision points, metrics for latency-sensitive ops, tracing at service boundaries, error classification (transient vs permanent). Use engineering judgment — not mandatory for every task.

Contract Testing Checkpoint (Service Boundaries Only): For new API endpoints or service-to-service interfaces, suggest consumer-driven contract tests. Skip for purely internal modules.

3c. REFACTOR - Clean with Tests Green

1. Review code for improvements
2. Refactor while keeping tests green
3. RUN all related tests after each change
4. Show final test output
5. Announce: "Refactoring complete, all tests passing: [evidence]"

Red Flags - STOP and restart the cycle if:

• About to write code before test exists
• Test passes immediately (testing wrong thing)
• Thinking "just this once" or "too simple to test"
• Running tests mentally instead of actually executing

If TDD Not Enabled:

3a. Implement

1. Implement the task as specified
2. Test manually or run existing tests
3. Announce: "Implementation complete"

Implementation Chunk Limit (Architecture Mode Only)

Activation: Only when .ai-context.md or architecture.md exists (track-level or project-level).

If the implementation diff for a task exceeds ~200 lines:

1. STOP after ~200 lines of implementation
2. Present the chunk for developer review
3. CHECKPOINT (MANDATORY): Wait for developer approval of the chunk
4. Commit the approved chunk: feat(<track_id>): <task description> (chunk N)
5. Continue with the next chunk
6. Repeat until the task is fully implemented

This prevents large, unreviewable code drops. Each chunk should be a coherent, reviewable unit.

---

Step 4: Update Progress & Commit

Iron Law: Every completed task gets its own commit. No batching. No skipping.

After completing each task:

0. Quick robustness scan (30-second check before committing):

   • Scan the code you just wrote against the Step 3.0c triggers
   • If any trigger is present but the pattern wasn't applied: fix it now
   • This is a rapid pattern-match, not a full review — you should have applied these during generation, this catches anything missed

1. Commit FIRST (REQUIRED - non-negotiable):

   • Stage only files changed by this task (never git add .)
   • git add <specific files>
   • Verify staged changes exist before committing: git diff --cached --quiet. If nothing staged, skip the commit step.
   • git commit -m "type(<track_id>): task description" (Conventional Commits — see core/shared/vcs-commands.md)
   • If a Jira ticket is linked in spec.md, reference it in the commit body: Refs: <JIRA_ID>.
   • Get commit SHA: git rev-parse --short HEAD
   • Do NOT proceed to the next task without committing
   • Do NOT batch multiple tasks into one commit

2. Update plan.md:

   • Change [ ] to [x] for the completed task
   • Add the commit SHA next to the task: [x] Task description (abc1234)

3. Update metadata.json:

   • Increment tasks.completed
   • Update updated timestamp

4. Verify state updates (CRITICAL):

   • Read back plan.md - confirm task marked [x] with SHA
   • Read back metadata.json - confirm tasks.completed incremented
   • If EITHER verification fails:
     • Mark task as [!] Blocked in plan.md
     • Add recovery message: "State update failed after commit . Recovery: manually edit plan.md line X to mark [x], update metadata.json tasks.completed to Y"
     • HALT - require manual intervention before continuing

5. If .ai-context.md or architecture.md exists for the track:

   • Update module status markers ([ ] → [~] when first task in module starts, [~] → [x] when all tasks complete)
   • Fill in Story placeholders with the approved story from Step 2.5
   • If updating project-level draft/.ai-context.md: also update YAML frontmatter git.commit and git.commit_message to current HEAD. Update draft/architecture.md with structural changes, then run the Condensation Subroutine (defined in core/shared/condensation.md) to regenerate draft/.ai-context.md.

Verification Gate (REQUIRED)

Iron Law: No completion claims without fresh verification evidence.

Before marking ANY task/phase/track complete:

1. IDENTIFY: What command proves this claim? (test, build, lint)
2. RUN: Execute the FULL command (fresh, complete run)
3. READ: Full output, check exit code
4. VERIFY: Does output confirm the claim?
   • If NO: Keep task as [~], state actual status
   • If YES: Show evidence, then mark [x]

Red Flags - STOP if you're thinking:

• "Should pass", "probably works"
• Satisfaction before running verification
• About to mark [x] without fresh evidence from this session
• "I already tested earlier"
• "This is a simple change, no need to verify"

---

Step 5: Phase Boundary Check

When all tasks in a phase are [x]:

1. Announce: "Phase N complete. Running three-stage review."

Three-Stage Review (REQUIRED)

Stage 1: Automated Validation

• Fast static checks: architecture conformance, dead code, circular dependencies, performance anti-patterns. Review for common security anti-patterns (OWASP top 10). For automated checks, use language-specific tools (e.g., npm audit for JS, bandit for Python, cargo audit for Rust).
• If critical issues found: List them, return to implementation

Stage 2: Spec Compliance (only if Stage 1 passes)

• Load track's spec.md
• Verify all requirements for this phase are implemented
• Check acceptance criteria coverage
• If gaps found: List them, return to implementation

Stage 3: Code Quality (only if Stage 2 passes)

• Verify code follows project patterns (tech-stack.md)
• Check error handling is appropriate
• Verify tests cover real logic
• Classify issues: Critical (must fix) > Important (should fix) > Minor (note)

See core/agents/reviewer.md for detailed review process.

Quick Review Alternative

At phase boundaries, offer the lightweight alternative:

"Phase {N} complete. Review options:
  1. Full three-stage review (recommended) — spec compliance + security + quality
  2. /draft:quick-review — lightweight 4-dimension check (faster)
  Choose [1/2, default: 1]:"

If quick-review chosen, invoke /draft:quick-review with the phase's changed files.

2. Run verification steps from plan (tests, builds)
3. Present review findings to user
4. If review passes (no Critical issues):
   • Update phase status in plan
   • Update metadata.json phases.completed
   • Refresh blast-radius memory (see "Impact Memory" subsection below)
   • Proceed to next phase
5. If Critical/Important issues found:
   • Document issues in plan.md
   • Fix before proceeding (don't skip)

Impact Memory (blast-radius snapshot)

After a phase passes review, refresh metadata.json.impact so future tracks can detect overlap with this work.

1. Compute touched files: From plan.md, find the first commit SHA recorded for this track (earliest [x] line with (<sha>)). Run:

   git diff --name-only <first_sha>^..HEAD
   

   That is the files_touched list. Derive modules_touched as the unique top-level path segments (e.g. auth/login.go → auth).

2. Compute downstream blast radius (graph-aware, optional): If draft/graph/schema.yaml exists, for each file in files_touched query:

   graph --repo . --out draft/graph --query --file <path> --mode impact
   

   Aggregate across all files: downstream_files = total unique downstream files (deduped), downstream_modules = union of affected_modules, max_depth = max across queries, by_category = sum of each query's by_category. If the graph is absent, leave these fields as zeros / empty arrays — the snapshot still records the directly-touched files.

3. Write metadata.json with the populated impact block and computed_at set to the current timestamp.

This snapshot is consumed by /draft:new-track to surface overlap warnings when a new track touches the same modules as a recently completed track.

Step 6: Track Completion

When all phases complete:

1. Run review (if enabled):

   • Read draft/workflow.md review configuration
   • Check if auto-review enabled:

     ## Review Settings
     - [x] Auto-review at track completion
     

   • If enabled, run /draft:review track <track_id>
   • Check review results:
     • If block-on-failure enabled AND critical issues found → HALT, require fixes
     • Otherwise, document warnings and continue

2. Update plan.md status to [x] Completed

3. Update metadata.json status to "completed"

4. Update draft/tracks.md:

   • Move from Active to Completed section
   • Add completion date

5. Verify completion state consistency (CRITICAL):

   • Read back plan.md - confirm status [x] Completed
   • Read back metadata.json - confirm status "completed"
   • Read back draft/tracks.md - confirm track in Completed section with completion date
   • If ANY file shows inconsistent state:
     • ERROR: "Track completion partially failed"
     • Report: "plan.md: , metadata.json: , tracks.md:
       "
     • Provide recovery: "Manually complete updates: [list specific edits needed]"
     • Do NOT announce completion until all three files verified consistent

6. Announce: "Track <track_id> completed!

Summary:

• Phases: N/N
• Tasks: M/M
• Duration: [if tracked]

[If review ran:] Review: PASS | PASS WITH NOTES | FAIL Report: draft/tracks/<track_id>/review-report-latest.md

All acceptance criteria from spec.md should be verified.

Next: Run /draft:status to see project overview."

Error Handling

If blocked:

• Mark task as [!] Blocked
• Add reason in plan.md
• REQUIRED: Follow systematic debugging process (see core/agents/debugger.md)
  1. Investigate - Read errors, reproduce, trace (NO fixes yet)
  2. Analyze - Find similar working code, list differences
  3. Hypothesize - Single hypothesis, smallest test
  4. Implement - Regression test first, then fix
• Do NOT attempt random fixes
• Document root cause when found

Recommended: Instead of inline debugging, invoke /draft:debug skill for a structured session:

"Task blocked: {description}. Run /draft:debug for structured investigation? [Y/n]"

The debug skill provides: Reproduce → Isolate → Diagnose → Fix methodology with debug report output.

If test fails unexpectedly:

• Don't mark complete
• Follow systematic debugging process above
• Announce failure details with root cause analysis
• Show evidence when resolved

If unsure about implementation:

• Ask clarifying questions
• Reference spec.md for requirements
• Don't proceed with assumptions

Tech Debt Log

During implementation, track technical debt decisions in the track's plan.md:

When you encounter a shortcut, workaround, or known-imperfect solution during implementation:

1. Add an entry to the ## Tech Debt section at the bottom of plan.md
2. Use this format:

## Tech Debt

| ID | Location | Description | Severity | Payback Trigger |
|----|----------|-------------|----------|-----------------|
| TD-1 | `src/api/handler.ts:45` | Hardcoded timeout instead of config | Low | When adding config system |
| TD-2 | `src/auth/session.ts:12` | In-memory session store | Medium | Before horizontal scaling |

Severity levels:

• Low — Cosmetic or minor maintainability issue
• Medium — Will cause problems at scale or in specific scenarios
• High — Actively impeding development or risking production issues

Payback Trigger — The condition or event that should trigger debt repayment (e.g., "before launch", "when adding feature X", "before scaling past N users").

Only log genuine debt — intentional shortcuts with known consequences. Not everything imperfect is debt.

---

Progress Reporting

After each task, report:

Task: [description]
Status: Complete
Phase Progress: N/M tasks
Overall: X% complete

---

Cross-Skill Dispatch

At Track Completion (Step 6)

After announcing track completion, suggest relevant follow-ups based on context:

If track modifies production code:

"Track complete! Consider:
  → /draft:deploy-checklist — Pre-deployment verification"

If track added new APIs/services/components:

  → /draft:documentation — Update documentation for new components"

If implementation contains TODO/FIXME/HACK comments:

  → /draft:tech-debt — Catalog any new technical debt introduced"

If new patterns or dependencies not in tech-stack.md:

  → /draft:adr — Document this design decision"

Jira Sync at Completion

If Jira ticket linked, sync via core/shared/jira-sync.md:

• Post comment: "[draft] implementation-complete: All {n} tasks done. Ready for review."

Bug Track with rca.md

If implementing a bug track and draft/tracks/<id>/rca.md exists:

• Load rca.md as context for the implementation
• Reference root cause, blast radius, and prevention items during fix
• After fix: update rca.md "Proposed Fix" section with actual fix details