impl

name	impl
description	Implementation orchestration from a plan to execute approved scope predictably without losing constraints. Decomposes steps into parallelizable work items, spawns domain-appropriate agents, and verifies completion against BDD success criteria. Triggers: "cwf:impl", "implement this plan"

Impl

Execute an approved plan predictably while preserving constraints, dependencies, and BDD validation.

Quick Start

cwf:impl                    # Auto-detect most recent plan.md
cwf:impl <path/to/plan.md>  # Explicit plan path
cwf:impl --skip-branch      # Skip Phase 0.5 branch gate (stay on current branch)
cwf:impl --skip-clarify     # Skip Phase 1.0 clarify pre-condition check

---

Phase 0: Update Live State

Use the live-state helper for scalar fields, then initialize list fields in the resolved live-state file:

bash {CWF_PLUGIN_DIR}/scripts/cwf-live-state.sh set . \
  phase="impl" \
  task="{plan goal summary}"
live_state_file=$(bash {CWF_PLUGIN_DIR}/scripts/cwf-live-state.sh resolve)

In {live_state_file}, set live.key_files to the plan's key files list and initialize live.decision_journal as an empty list:

live:
  phase: "impl"
  task: "{plan goal summary}"
  decision_journal: []

Decision Journal

Throughout Phases 2-4, append significant decisions to live.decision_journal in the resolved live-state file (bash {CWF_PLUGIN_DIR}/scripts/cwf-live-state.sh resolve). Record decisions when:

• A design choice is made between alternatives
• A deviation from the plan occurs
• A trade-off is resolved

Format each entry as a timestamped string:

decision_journal:
  - "Phase 2: chose per-pattern commits — changes are cross-cutting"
  - "Phase 3b: agent-2 deviated from plan — added error handling for edge case"

The compact recovery hook reads this field to restore decision context after auto-compaction. Keep entries concise (one line each, max ~80 chars).

Phase 0.5: Branch Gate

Apply the Branch Gate. Never proceed on a base branch.

---

Phase 1: Load Plan

Find and parse the plan that drives implementation.

1.0 Clarify Completion Check

Apply the Clarify Completion Gate. Hard-blocks unless --skip-clarify is provided.

1.1 Plan Discovery

If an explicit path was provided, use it. Otherwise:

1. Scan .cwf/projects/*/plan.md using Glob.
2. Rank candidates with this priority (highest first):
   • Session pin from live state: if plan.md exists inside live.dir from the resolved live-state file, prefer it.
   • Plan metadata timestamp: newest parseable ISO-8601 from plan metadata/frontmatter (updated_at, created_at, generated_at, date).
   • Filesystem timestamp: newest plan.md modification time.
   • Directory name order: use only as a final deterministic tiebreaker after timestamp signals.
3. Select the top-ranked candidate only when it is unambiguous.

Ambiguity behavior (fail-closed):

• If two or more candidates tie on all available ranking signals, stop and ask the user for an explicit path.
• If metadata timestamps are missing/invalid for all tied candidates and file mtimes are identical, stop and ask the user for an explicit path.
• Never silently pick a plan based only on lexicographic directory naming when chronology is ambiguous.

If no plan is found, report to the user and stop:

No plan.md found in .cwf/projects/. Run cwf:plan first.

1.1b Phase Handoff Discovery

Check for a phase-handoff.md in the same directory as the discovered plan.md:

1. Derive the directory from the plan path: dirname {plan.md path}
2. Check if phase-handoff.md exists in that directory
3. If found, read and parse all sections:
   • Context Files to Read: Additional files the agent must read before implementation
   • Design Decision Summary: Key choices with rationale — treat as binding context
   • Protocols to Follow: Behavioral rules — treat as binding constraints alongside plan.md
   • Do NOT: Prohibitions — append to plan's "Don't Touch" list as behavioral constraints
   • Implementation Hints: Practical guidance — use during Phase 3 execution
   • Success Criteria: Additional BDD criteria — merge with plan's criteria for Phase 4 verification
4. If phase-handoff.md specifies "Context Files to Read", read each listed file before proceeding

If phase-handoff.md is not found, proceed normally with plan.md only. Phase handoff is optional.

1.2 Section Extraction

Read the plan and extract these sections (all optional except Steps):

Section	Key Content
Context	Background, rationale
Goal	Desired outcome
Steps	Ordered implementation steps (REQUIRED)
Files to Create/Modify	File paths with actions
Success Criteria — Behavioral (BDD)	Given/When/Then scenarios
Success Criteria — Qualitative	Non-functional quality attributes
Don't Touch	Files explicitly out of scope
Deferred Actions	Items not for this session

If the Steps section is missing or empty, report to the user and stop.

1.3 User Confirmation

Present a summary and confirm before proceeding:

## Plan Loaded

**Source**: {path to plan.md}
**Goal**: {goal from plan}
**Steps**: {count} steps
**Files**: {count} files to create/modify
**BDD Criteria**: {count} scenarios
**Don't Touch**: {list or "none specified"}
**Phase Handoff**: {path to phase-handoff.md, or "not found (proceeding with plan only)"}
**Phase Protocols**: {count} protocols to follow
**Phase Do NOT**: {count} prohibitions

Proceed with implementation?

Use AskUserQuestion with options: "Proceed", "Review plan first", "Cancel".

1.4 Lessons Baseline

Resolve lessons.md in the same directory as the selected plan.md.

1. If lessons.md does not exist, create it immediately using the structure from plan-protocol.md.
2. Treat lesson accumulation as continuous behavior:
   • user clarifications discovered during implementation
   • plan-vs-execution gaps
   • runtime/tooling constraints discovered mid-flight
3. Keep lesson entries in the user's language.

---

Phase 2: Analyze & Decompose

Break the plan into parallelizable work items and size the agent team.

2.1 Domain Detection

Read {SKILL_DIR}/references/agent-prompts.md Section "Domain Signal Table".

For each step, identify the primary domain by matching:

• File patterns mentioned in the step
• File patterns from the "Files to Create/Modify" table
• Keywords and descriptions in the step text

2.2 Dependency Analysis

Read {SKILL_DIR}/references/agent-prompts.md Section "Dependency Detection Heuristics".

Analyze step-to-step dependencies:

• File overlap: Steps touching the same file must be sequential
• Output references: Step N uses output from Step M → sequential
• Ordering keywords: "after", "then", "once X is done" → sequential
• Independent: No shared files, no output dependencies → parallel-safe

2.3 Work Item Grouping

Group steps into work items:

1. Place sequential-dependent steps in the same work item (preserve order)
2. Group steps with the same domain into one work item when possible
3. Keep parallel-safe items in separate work items
4. Each work item gets: assigned steps, file list, relevant BDD criteria

2.4 Adaptive Team Sizing

Read {SKILL_DIR}/../../references/agent-patterns.md for general agent principles.

Work Items	Agents	Strategy
1 (≤3 files)	0	Direct execution — Phase 3a
2-3	2	Group related items per agent
4-6	3	Balance parallelism vs coordination
7+	4 (hard cap)	Beyond 4, overhead exceeds gains

2.4.1 Web Debug Classification (Conditional)

Before finalizing work items, detect browser-runtime debugging scope from step text and file signals (for example: console, DOM, CDP, DevTools, click/focus/navigation regressions, viewport or mobile regressions, HTML/CSS/SPA runtime files).

If detected:

• mark the work item as web-debug
• require Web Debug Loop Protocol execution during Phase 3 (before/after evidence)
• resolve session_dir from live.dir in the resolved live-state file, then reserve {session_dir}/debug/{timestamp}-{slug}/

2.5 Cross-Cutting Assessment

Before finalizing decomposition, assess whether changes are cross-cutting:

1. Cross-cutting indicator: A single concept (protocol, pattern, interface) projected across 3+ files
2. If cross-cutting: Commit boundary = change pattern. Add a Commit Group column to group files by the concept they implement, regardless of work item assignment.
3. If modular: Commit boundary = work item (default). No Commit Group needed.

2.6 Present Decomposition

Show the decomposition before executing:

## Decomposition

### Work Items

| # | Steps | Domain | Files | Parallel | Commit Group |
|---|-------|--------|-------|----------|--------------|
| 1 | 1, 2 | {domain} | {file list} | yes | {group or "—"} |
| 2 | 3 | {domain} | {file list} | yes | {group or "—"} |
| 3 | 4, 5 | {domain} | {file list} | after #1 | {group or "—"} |

### Execution Strategy
- **Mode**: {Direct (3a) | Agent Team (3b)}
- **Agents**: {count}
- **Batches**: {count} (sequential batches of parallel items)
- **Commit Strategy**: {per work item | per change pattern}

### Don't Touch (enforced)
- {files from plan's Don't Touch section}

If Agent Team mode: use AskUserQuestion with "Execute", "Adjust grouping", "Cancel". If Direct mode: proceed without confirmation.

---

Phase 3a: Direct Execution

For simple plans (1 work item, ≤3 files).

1. Execute each step in the work item sequentially

2. Use Write, Edit, Bash, and other allowed tools directly

3. Follow the plan's step descriptions precisely

4. If the item is web-debug, run Web Debug Loop Protocol and persist before/after evidence under {session_dir}/debug/{timestamp}-{slug}/

5. Respect the "Don't Touch" list — never modify those files

6. If phase-handoff.md was loaded, apply its Implementation Hints during execution and respect its "Do NOT" constraints

7. Append newly discovered learnings to lessons.md before commit (user language).

8. Commit Gate: Apply the Commit Gate. Stage specific files and commit with Conventional Commit format.

9. After commit, proceed to Phase 4

---

Phase 3b: Agent Team Execution

For complex plans requiring parallel work.

3b.1 Prompt Construction

For each work item, build an agent prompt using the template from {SKILL_DIR}/references/agent-prompts.md Section "Implementation Agent Prompt Template". The prompt includes:

• The specific steps assigned to this agent
• The file list (create/modify) for these steps
• Relevant BDD criteria that this agent's work should satisfy
• The Don't Touch list (full list — every agent must respect it)
• Context from the plan (Goal, relevant Context sections)
• Phase handoff protocols, hints, and "Do NOT" constraints (if phase-handoff.md was loaded)
• For web-debug items, explicit instruction to run Web Debug Loop Protocol and persist evidence paths in the final report

3b.2 Parallel Launch

Launch agents for parallel-safe work items in a single message with multiple Task tool calls:

# All parallel items in one message
Task tool (item 1):
  subagent_type: general-purpose
  mode: bypassPermissions
  prompt: |
    {constructed prompt from 3b.1}

Task tool (item 2):
  subagent_type: general-purpose
  mode: bypassPermissions
  prompt: |
    {constructed prompt from 3b.1}

3b.3 Sequential Batches

If some work items depend on others:

1. Launch all parallel-safe items in batch 1
2. Wait for batch 1 to complete
3. Launch dependent items in batch 2 (include batch 1 results as context)
4. Repeat until all batches complete

Between batches, briefly verify that prior batch outputs exist and are valid before launching the next batch.

3b.3.5 Batch Commit

After each batch completes (before launching the next batch), apply the Commit Gate.

File-to-work-item mapping: The orchestrator maintains a map derived from Phase 2 decomposition ({work_item_id → [file_paths]}). Each agent's prompt (from 3b.1) MUST include the list of files it is expected to modify. After batch completion, use this map to attribute files to work items. If an agent modifies unexpected files, include them in that work item's commit with a note.

Commit granularity: Determined by Phase 2.6 cross-cutting assessment:

• Modular changes: One commit per work item
• Cross-cutting changes: One commit per change pattern (Commit Group)

Report after all commits in the batch:

Batch {N} committed:
- {short-hash} type(scope): summary

3b.3.6 Lesson-Driven Commits

Apply the Lesson-Driven Commits protocol. Lesson-driven changes get separate commits with "Driven by lesson:" reference.

3b.3.7 Lessons Checkpoint per Batch

After each batch (and before launching the next), update lessons.md with any new findings:

• corrected assumptions from user feedback
• design/implementation gaps discovered during execution
• constraints/tooling behavior that changed execution strategy

Keep entries concise and in the user's language.

3b.4 Result Collection

For each completed agent, append a structured result record (session artifact):

• path: {session_dir}/impl-result-provenance.md
• shape: one row per agent output (including failures)

For each row, capture:

1. Record which steps were completed
2. Note any files created or modified
3. Capture any issues or deviations reported by the agent
4. Capture provenance metadata (required):
   • source (REAL_EXECUTION | CACHED | FALLBACK)
   • tool (for example claude-task, codex, gemini)
   • agent_id (or slot identifier)
   • started_at_utc, finished_at_utc, duration_ms
5. If an agent reports failure, record the failure with the same provenance fields and continue with other agents

If any provenance field is unavailable, persist — explicitly (never omit key columns).

---

Phase 4: Verify & Suggest Review

Check implementation against the plan's success criteria.

4.1 BDD Criteria Checklist

For each BDD scenario from the plan (and from phase-handoff.md Success Criteria, if present):

1. Determine which work item(s) should have addressed it
2. Check if the relevant files were created/modified
3. Mark each criterion:
   • Covered: Implementation addresses this scenario
   • Uncovered: No implementation found for this scenario
   • Partial: Some aspects addressed, others missing

4.2 Qualitative Assessment

For each qualitative criterion, briefly note how the implementation addresses it (or flag if it doesn't).

4.3 Completion Summary

Present the verification results:

## Implementation Complete

### BDD Criteria Coverage

| # | Scenario | Status | Notes |
|---|----------|--------|-------|
| 1 | {Given/When/Then summary} | Covered | {brief note} |
| 2 | {Given/When/Then summary} | Uncovered | {what's missing} |

### Qualitative Assessment

| Criterion | Assessment |
|-----------|------------|
| {criterion} | {how it's addressed} |

### Files Modified

| File | Action | Agent |
|------|--------|-------|
| {path} | Created | {agent # or "direct"} |

### Issues

- {any deviations, failures, or concerns}

4.4 Uncovered Criteria

If any BDD criteria are uncovered:

1. List them explicitly
2. Ask the user whether to:
   • Attempt to address them now (additional implementation pass)
   • Defer them (add to Deferred Actions)
   • Ignore them (they may be covered by later stages)

4.5 Session Completeness Check

Run the session completeness check:

bash {CWF_PLUGIN_DIR}/scripts/check-session.sh --impl

If any FAIL items are reported, fix them before suggesting review. This ensures all required artifacts (plan.md, lessons.md, next-session.md) exist and cwf-state.yaml is properly updated.

4.6 Suggest Review

After presenting the summary:

Implementation complete. To review the code quality and verify against
the plan's success criteria, run:
  cwf:review --mode code

---

Rules

1. Plan is the contract: Implement what the plan says. Do not add features, refactor unrelated code, or reinterpret the goal
2. Don't Touch means don't touch: Never modify files in the Don't Touch list, even for "improvements"
3. Parallel when safe, sequential when needed: Only parallelize items with no file overlap or output dependencies
4. No idle agents: Size the team to the work. One agent with real work beats three agents where two wait
5. Fail forward: If one agent fails, collect its error and continue with others. Report all failures in Phase 4
6. Verify against criteria: Every BDD scenario must be checked, not assumed
7. Preserve existing patterns: Follow codebase conventions visible in existing files. Do not introduce new patterns without plan justification
8. Markdown discipline: All code fences must have a language specifier
9. Phase handoff protocols are binding: When phase-handoff.md is present, its "Protocols to Follow" and "Do NOT" sections carry the same weight as plan constraints. They are not suggestions.
10. Phase handoff is optional: Impl works with plan.md alone. Phase handoff enriches but is never required.
11. Never implement on a base branch: Phase 0.5 is a hard gate. The only bypass is --skip-branch.
12. Commit boundary = change pattern: When changes are cross-cutting (Phase 2.6), commit by concept/pattern. When modular, commit per work item. See Commit Gate.
13. Stage specific files, never git add -A: Only stage files known to belong to the current work item. Prevent accidental inclusion of unrelated changes.
14. Lesson-driven changes get separate commits: If a lesson triggers a code change, commit it independently. See Lesson-Driven Commits.
15. Incremental lessons are mandatory: Update lessons.md as implementation progresses, not only at the end.
16. Structural Triage Contract: For each triage item referencing analysis documents, enforce the triage record contract in agent-patterns.md: include source_ref, source_recommendation, triage_action, runtime_caller_check, deletion_premortem, and decision_id. If source_recommendation is missing, stop and complete triage context first. If triage_action diverges from source_recommendation, ask the user before proceeding. For deletion actions, do not continue until runtime-caller check and pre-mortem are both recorded.
17. Language override is mandatory: implementation/code artifacts stay in English, while lessons.md stays in the user's language.
18. Web-debug evidence is mandatory: Work items classified as web-debug must include before/after browser evidence and a reproducible path using Web Debug Loop Protocol.

References

• references/impl-gates.md — Branch, Clarify, and Commit gate definitions
• references/agent-prompts.md — Agent prompt template, domain signals, dependency heuristics
• agent-patterns.md — Shared agent orchestration, Web Research, and Web Debug protocols