agentleFS
DeL-TaiseiOzakiDeL-TaiseiOzaki/claude-code-orchestra/.claude/skills/feature/SKILL.md

feature

Skill199 starsChanged 2 months agoMITView on GitHub

Unified feature planning & implementation skill — replaces the old /add-feature and /start-feature skills (both trigger phrases still apply here). MODE=existing (formerly /add-feature): add a feature to an established codebase with Codex-first collaboration — Codex is consulted in every phase for scope analysis, architecture design, implementation planning, and validation. MODE=greenfield (formerly /start-feature): start a large or new feature that requires external research — Agent Teams (Researcher + Architect) do parallel research & design. Both modes share Phase 3 complexity routing (SIMPLE: Codex direct, MODERATE: Codex + /team-execute --review-only, COMPLEX: /team-execute).

73 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use bug, bugfix and 9 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • paulocsbSkill

    bug

    Investigation playbook for fixing a bug. Walks through 5 phases: reproduce → isolate → root cause → propose fix → validate. Forces concrete reasoning with code reads, log inspection, and git blame instead of pattern-matching guesses. Use for any non-trivial defect.

    Also in 8 projects that use this

  • team-attentionSkill

    bugfix

    Root cause based one-shot bug fix. debugger diagnosis → requirements.md generation → /execute. /bugfix "error description" Full investigation pipeline: debugger + gap-analyzer + standard verify. QA suggestion after successful fix.

    Also in 7 projects that use this

  • codebycoreySkill

    adopt

    Adopt an existing plan from conversation context into delta-spec format. Use when importing a plan, adopting planning that already happened, capturing plan mode output, or bringing in prior design work.

    Also in 4 projects that use this

  • Google CloudSkill

    fix

    Plan a remediation strategy after a test failure.

    Also in 10 projects that use this

  • marylinSkill

    recover

    Recover work from a crashed or interrupted session. Use when Claude crashed mid-task, after an unexpected exit, when you see 'stale sessions detected', or when you lost context and need to figure out what was happening.

    Also in 3 projects that use this

  • aegiswizardSkill

    context

    Also in 4 projects that use this

The file

Show its rules and text

We split this skill into the rules it gives an agent

25 rules. 0 are common, 9 are distinctive, and 16 are very specific to this project.

  • 0

    Common

    Most of the 67 edited versions we read keep these.

  • 9

    Distinctive

    This author's own choices. Worth a look.

  • 16

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If the signals are mixed or unclear, ask via AskUserQuestion — do NOT guess:

    line 52
  2. Very specific:

    Before anything else, if PROGRESS.md exists at the repository root, read it.

    line 79
  3. Distinctive:

    Ask the user to clarify:

    line 106
  4. Very specific:

    Every Codex consultation in this skill goes through the shared wrapper instead of a raw codex exec call, so a crashed CLI is never silently …

    line 157
  5. Very specific:

    Read the answer from the JSON output's response_file path.

    line 173
  6. Very specific:

    Write the typed input JSON to .claude/logs/design-input-{slug}.json (slug from Step 0-b, so two features cannot collide on one input file):

    line 198
  7. Distinctive:

    Run the dry-run, review the preview, then apply:

    line 221
  8. Very specific:

    Verify "ok": true and "result": "applied".

    line 231
  9. Very specific:

    Verify "ok": true and "progress_tracker_preserved": true in the output.

    line 281
  10. Very specific:

    All teammates spawned in MODE=greenfield write their work log to .claude/logs/agent-teams/{team-name}/{teammate}.md per the shared format: …

    line 289
  11. Distinctive:

    Use Codex's complexity classification to determine the implementation route in Phase 3.

    line 324
  12. Very specific:

    If Codex returns NEEDS_REVISION, update the plan and re-validate before proceeding.

    line 433
  13. Distinctive:

    If Codex needs more info → another researcher subagent round

    line 601
  14. Distinctive:

    Create the task list using TodoWrite:

    line 640
  15. Distinctive:

    Do not implement until the user approves the plan.

    line 720
  16. Distinctive:

    Every route below hands implementation to an agent that reports on its own work.

    line 731
  17. Very specific:

    Read the JSON: overall is pass / fail / no_gates.

    line 742
  18. Very specific:

    Read deletions, placeholders, weakened_tests, and out_of_scope_files.

    line 765
  19. Very specific:

    Use .claude/skills/_shared/gather_diff.py --base {ref} when you want the full patch to read: scope_empty: true with exit 2 means the …

    line 775
  20. Very specific:

    Read the implementation summary from the JSON output's response_file — as input to the verification, never as its result.

    line 815
  21. Very specific:

    When /team-execute returns, run Completion Verification above yourself.

    line 850
  22. Very specific:

    If the .claude/STATE.md block was not written at plan time (existing mode), write it now per the common protocol.

    line 856
  23. Distinctive:

    Codex-first (existing mode): Every phase consults Codex.

    line 884
  24. Distinctive:

    Complexity routing: Do not over-engineer simple features.

    line 887
  25. Very specific:

    Quality gates: Every route ends with Completion Verification — verify.sh (gate failure or no gate at all is exit 2) plus …

    line 888

“Common” is measured: kept by at least 6 in 10 of the 67 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Feature

One entry point for feature work, two modes:

  • MODE=existing (old /add-feature path): the feature goes into an established codebase whose conventions are already known. No external research needed → Codex-direct scope → design → plan.
  • MODE=greenfield (old /start-feature path): a large or new feature that needs external research and parallel design → Agent Teams (Researcher + Architect) with bidirectional communication.

Both modes converge on a shared Phase 3: user approval + complexity-routed implementation.

Preflight: ensure codex CLI is current (see codex-system skill).

/feature <feature description>
    | MODE determination (AskUserQuestion when ambiguous)
    ├─ MODE=existing   : Phase 1E SCOPE  -> Phase 2E DESIGN (Codex direct)
    └─ MODE=greenfield : Phase 1G UNDERSTAND -> Phase 2G RESEARCH & DESIGN (Agent Teams)
    | Phase 3 (shared): PLAN, APPROVE & IMPLEMENT
    SIMPLE   (1-3 files, <50 LOC) -> Codex danger-full-access direct
    MODERATE (3-5 files)          -> Codex danger-full-access + /team-execute --review-only
    COMPLEX  (5+ files)           -> /team-execute (implement + review)

Mode Determination (always first)

Decide the MODE before anything else. Signals:

SignalMODE=existingMODE=greenfield
Codebase stateEstablished, conventions knownNew area, or conventions absent
External research neededNo — Codex reasons about existing patterns directlyYes — libraries/tools/reference architectures must be researched
Feature sizeLocalized additionLarge, multi-module, or project kickoff
Design sourceCodex (read-only consults)Agent Teams (Researcher + Architect)

If the signals are mixed or unclear, ask via AskUserQuestion — do NOT guess:

question: "Which feature mode applies?"
multiSelect: false
options:
  - label: "existing"
    description: "Add to an established codebase; conventions known; no external research (old /add-feature)."
  - label: "greenfield"
    description: "Large/new feature; needs external research and parallel design via Agent Teams (old /start-feature)."

When NOT to Use

  • Bug diagnosis where root cause is unclear → /troubleshoot
  • Executing an already-approved implementation plan → /team-execute
  • Feasibility unknown / go-no-go decision needed first → /spike
  • Truly trivial changes (single function, <10 LOC) → edit directly, skip this skill

Full skill routing: AGENTS.md section "Routing Policy".


Common Protocols (both modes)

Step 0: Read PROGRESS.md (always first)

Before anything else, if PROGRESS.md exists at the repository root, read it. It is the rolling summary of the latest 5 checkpoints (maintained by /checkpointing) and carries the most recent session context, in-progress work, and the "将来のアクション" (next actions) from prior sessions. Use it to ground the new feature in what already happened and to avoid re-deciding settled questions. If it is absent (fresh repo), skip this step.

Step 0-b: Resolve the Workspace

Resolve this feature's paths once. The title becomes file and directory names, so give it a short English descriptor of the feature — not the user's raw wording, which the Language Protocol keeps out of paths:

python3 .claude/skills/_shared/workspace.py \
  --skill feature --title "<short English title>" --create

The JSON on stdout carries slug, team_name, and paths (brief, codebase_scan, research, state_input, team_dir). From here on, every {slug} / {team-name} / output path in this skill — and the /team-execute handoff in Route C — MUST come from this JSON verbatim, never be re-derived by hand: two independently hand-derived slugs are exactly how cross-phase artifacts drift out of sync.

Requirements Gathering

Ask the user to clarify:

  1. Purpose / feature description: What should the feature do? What do you want to achieve?
  2. Expected behavior: How should it work from the user's perspective?
  3. Scope boundaries: What to include / exclude?
  4. Technical preferences / constraints: Specific libraries, patterns, or constraints?
  5. Success criteria: How do you determine the feature is complete?
  6. Final design (greenfield): What form should the result take?

Opus Subagent Codebase Scan

Main orchestrator context is precious — large-scale codebase scanning is always delegated to general-purpose-opus (Opus, 1M context):

Task tool:
  subagent_type: "general-purpose-opus"
  prompt: |
    Analyze this codebase for feature: {feature description}

    Tasks (MODE=existing — affected-area scan):
    1. Identify the areas relevant to this feature:
       - Which modules/files will be affected?
       - What are the existing patterns in those areas?
       - What interfaces/contracts exist that the feature must conform to?
    2. Analyze existing conventions:
       - Code patterns (naming, structure, error handling)
       - Test patterns (test location, fixture usage, assertion style)
       - Import and dependency patterns
    3. Map dependencies:
       - What does the affected code depend on?
       - What depends on the affected code? (downstream consumers)
       - Are there shared utilities or base classes to leverage?

    Tasks (MODE=greenfield — comprehensive scan):
    - Directory structure and organization
    - Key modules and their responsibilities
    - Existing patterns and conventions
    - Dependencies and tech stack
    - Test structure

    Use Glob, Grep, and Read tools to investigate thoroughly.

    Save analysis to the `codebase_scan` path from Step 0-b (`.claude/docs/research/feature-{slug}-codebase.md`).
    Return concise summary (5-7 key findings).

Claude may supplement the subagent's analysis with targeted Glob/Grep/Read on specific files.

Codex Consult Protocol

Every Codex consultation in this skill goes through the shared wrapper instead of a raw codex exec call, so a crashed CLI is never silently mistaken for an empty answer. Write the prompt body to a file under the workspace (.claude/logs/codex/prompt-{label}.md, beside where the wrapper writes the response), then invoke. The wrapper creates its log directory only after it reads the prompt file, so create the directory first — otherwise the heredoc write fails in a fresh clone and the consult reads nothing (or a stale prompt from a previous label):

mkdir -p .claude/logs/codex
# write the prompt body to .claude/logs/codex/prompt-{label}.md, then:
python3 .claude/skills/_shared/codex_consult.py \
  --prompt-file .claude/logs/codex/prompt-{label}.md --label {label} --sandbox read-only

Read the answer from the JSON output's response_file path. Exit codes: 0 the call succeeded — read response_file; 1 bad args; 2 codex CLI not on PATH; 3 codex exited non-zero or timed out — inspect error and stderr_file before retrying or escalating.

Sandbox: read-only for every analysis/design/validation consult below; Phase 3 implementation (Route A/B) uses --sandbox danger-full-access instead — called out again at that call site.

Prompts below show only the prompt body (Objective / Context / Constraints / Output format) — that is the file content for --prompt-file. MODE=existing consultations are MANDATORY — do not skip them. The most important input to every Codex prompt is the existing codebase patterns from the Opus subagent scan — always include them.

DESIGN.md Update

In both modes, record the feature's architecture decisions in .claude/docs/DESIGN.md (the macro 要件定義書) before presenting the plan — never by editing the file directly. DESIGN.md is user-owned and, in MODE=greenfield, has two writers (the Architect in Phase 2G and the lead in Phase 3); a hand edit loses the atomic replace and the concurrent-modification guard, and one writer silently overwrites the other. Every write goes through the shared writer, exactly as design-tracker does.

Write the typed input JSON to .claude/logs/design-input-{slug}.json (slug from Step 0-b, so two features cannot collide on one input file):

{
  "decisions": [
    {"decision": "{design decision}", "rationale": "{why}", "alternatives": "{what was rejected}"}
  ],
  "tech_choices": [
    {"area": "{area}", "technology": "{library or tool}", "rationale": "{why}", "alternatives": "{rejected}"}
  ],
  "section_updates": [
    {"heading": "## アーキテクチャ (Architecture)", "content": "- {integration point}: {how the feature connects}"}
  ]
}

Table rows go through their typed key — decisions, requirements, nfr, tech_choices, agent_roles — which places the row in the right table and escapes | in every cell. Hand-writing a table row as section_updates content is refused. Use section_updates only for prose sections (Architecture overview, Constraints, TODO / Open Questions).

Run the dry-run, review the preview, then apply:

python3 .claude/skills/_shared/update_design.py \
  --input .claude/logs/design-input-{slug}.json
# Review the preview file path in the JSON output, then:
python3 .claude/skills/_shared/update_design.py \
  --input .claude/logs/design-input-{slug}.json --apply --require-change

Verify "ok": true and "result": "applied". --require-change makes a no-op result (every row a duplicate, or an empty payload) exit 2, so this step can never report "recorded" for a run that wrote nothing. Exit 1 is a bad input schema; exit 2 DESIGN.md is missing or structurally invalid (run /init) or the run was a no-op; exit 3 DESIGN.md changed under you (concurrent modification) or the write failed — re-read DESIGN.md, drop what the other writer already recorded, and re-run the dry-run before applying again.

Ordering in MODE=greenfield: the Architect teammate writes its design decisions during Phase 2G through this same script; the lead writes only after both teammates have finished (Phase 3 Step 3), and records only what the Architect did not.

Shared State Update

Append feature context to .claude/STATE.md for cross-session persistence, following .claude/rules/agent-state.md. Use the shared writer script for a deterministic, atomic update; never edit root AGENTS.md.

Gather these fields from the planning phases:

  • Context: Goal (1-2 sentences), Key files (new/modified), Dependencies, Complexity (SIMPLE / MODERATE / COMPLEX)
  • Architecture: Key decisions from Codex / Architect
  • Library Constraints (greenfield) or Codex Validation (existing)
  • Integration Points and Decisions with rationale

Write the input JSON to the state_input path from Step 0-b (.claude/logs/state-input-{slug}.json):

{
  "title": "{feature name}",
  "sections": [
    {"heading": "Context", "content": "- Goal: ...\n- Key files: ...\n- Dependencies: ...\n- Complexity: MODERATE"},
    {"heading": "Architecture", "content": "- {decisions}"},
    {"heading": "Decisions", "content": "- {Decision 1}: {rationale}"}
  ]
}

Run dry-run, review the preview, then apply:

python3 .claude/skills/_shared/append_state_block.py \
  --type feature --input .claude/logs/state-input-{slug}.json
# Review the preview file path in the JSON output, then:
python3 .claude/skills/_shared/append_state_block.py \
  --type feature --input .claude/logs/state-input-{slug}.json --apply

Verify "ok": true and "progress_tracker_preserved": true in the output. Exit code 2 means the state structure is invalid; stop before writing.

Timing: MODE=greenfield writes this at plan time (Phase 3); MODE=existing may defer it to post-implementation. Either way it is written exactly once per feature.

Work Logs (Agent Teams roles)

All teammates spawned in MODE=greenfield write their work log to .claude/logs/agent-teams/{team-name}/{teammate}.md per the shared format: .claude/skills/_shared/work-log-format.md.


MODE=existing — Phase 1E: SCOPE (Opus Subagent + Codex + Claude Lead)

Understand the feature's scope and impact on the existing codebase: run the Opus subagent scan (common protocol, existing task list) and consult Codex for scope and impact analysis, while Claude clarifies requirements with the user (common protocol).

Codex Scope & Impact Analysis (MANDATORY)

Via the Codex consult protocol:

Objective: Analyze the scope and impact of adding this feature to the existing codebase.
Context:
- Feature: {feature description}
- Affected modules: {from Opus subagent analysis}
- Existing patterns: {from Opus subagent analysis}
- Dependencies: {from Opus subagent analysis}
Constraints:
- Assess how many files need to change and estimate LOC
- Classify complexity: SIMPLE (1-3 files, <50 LOC), MODERATE (3-5 files), COMPLEX (5+ files)
- Identify integration points where the feature connects to existing code
- Flag risks: breaking changes, performance concerns, test coverage gaps
Output format:
## Scope Assessment
## Complexity Classification (SIMPLE / MODERATE / COMPLEX)
## Integration Points
## Affected Files (with change type: new / modify)
## Risks and Concerns
## Recommended Approach

Use Codex's complexity classification to determine the implementation route in Phase 3.

Create Feature Brief

Combine user requirements + codebase analysis + Codex scope assessment into a Feature Brief following the MODE=existing template in references/brief-templates.md, and write it to the brief path from Step 0-b (.claude/docs/research/feature-{slug}-brief.md).

The brief is the primary cross-phase artifact: it feeds all three Phase 2E Codex prompts, Phase 3, the Route A implementation prompt, and the /team-execute handoff. Interpolating it from conversation context is how a half-filled brief reaches three Codex prompts undetected — every downstream step reads the file.

Validate it before leaving this phase:

python3 .claude/skills/_shared/validate_doc.py \
  --contract feature-brief --file .claude/docs/research/feature-{slug}-brief.md

Exit 0 every required section is present (the MODE=existing / MODE=greenfield variant is auto-detected from the headings); 1 bad args or the file is unreadable — most often it was never written; 2 a required section is missing, listed in sections_missing. Do not continue to Phase 2E on a non-zero exit.

The ### Complexity Classification (from Codex) section is where the decided classification is recorded once. Phase 3's presentation and route selection read it from this file rather than re-typing it, so a MODERATE assessment cannot be presented and then routed as SIMPLE. Codex decides the classification; only its propagation is mechanical.


MODE=existing — Phase 2E: DESIGN (Codex Architecture + Plan + Validation)

Codex designs the architecture, creates an implementation plan, and validates completeness. All three consultations are MANDATORY.

Unlike MODE=greenfield which uses Agent Teams (Researcher + Architect) for design, MODE=existing uses Codex directly because the patterns and conventions are already established.

Step 1: Codex Architecture Design (MANDATORY)

Objective: Design the architecture for adding this feature to the existing codebase.
Context:
- Feature Brief: contents of .claude/docs/research/feature-{slug}-brief.md (from Phase 1E)
- Existing patterns: {conventions from codebase scan}
- Integration points: {from Codex scope analysis}
Constraints:
- Follow existing codebase conventions exactly (naming, structure, patterns)
- Minimize changes to existing code (prefer extension over modification)
- Maintain backward compatibility
- Design for testability
Output format:
## Architecture Design
## Module Structure (new files and modifications)
## Interface Design (function signatures, class APIs)
## Data Flow
## Error Handling Strategy
## Test Strategy

Step 2: Codex Implementation Plan (MANDATORY)

Objective: Create a step-by-step implementation plan for this feature.
Context:
- Feature Brief: contents of .claude/docs/research/feature-{slug}-brief.md (from Phase 1E)
- Architecture Design: {from Step 1}
- Complexity: {SIMPLE / MODERATE / COMPLEX}
Constraints:
- Order steps by dependency (what must be built first)
- Each step should be independently testable
- Include test writing as explicit steps (TDD where possible)
- Keep individual steps small and focused
Output format:
## Implementation Steps (ordered by dependency)
## File Changes (per step: file path, change type, description)
## Test Plan (per step: what to test)
## Dependencies Between Steps
## Estimated Effort per Step

Step 3: Codex Validation (MANDATORY)

Objective: Validate this implementation plan for completeness, correctness, and risk.
Context:
- Feature Brief: contents of .claude/docs/research/feature-{slug}-brief.md
- Architecture Design: {from Step 1}
- Implementation Plan: {from Step 2}
- Existing codebase patterns: {from Phase 1E}
Constraints:
- Check for missing edge cases or error handling
- Verify the plan maintains backward compatibility
- Ensure test coverage is adequate
- Identify potential integration issues
- Check that the plan follows existing conventions
Output format:
## Validation Result (PASS / NEEDS_REVISION)
## Missing Coverage
## Backward Compatibility Check
## Convention Compliance
## Integration Risks
## Additional Test Cases Recommended
## Revised Steps (if NEEDS_REVISION)

If Codex returns NEEDS_REVISION, update the plan and re-validate before proceeding.

Then update DESIGN.md (common protocol) and continue to Phase 3.


MODE=greenfield — Phase 1G: UNDERSTAND (Opus Subagent + Claude Lead)

Analyze the codebase with the Opus subagent scan (common protocol, greenfield task list) while Claude gathers requirements from the user (common protocol).

Create Project Brief

Combine codebase understanding + requirements into a Project Brief following the MODE=greenfield template in references/brief-templates.md, and write it to the brief path from Step 0-b (.claude/docs/research/feature-{slug}-brief.md).

Validate it before spawning the team — a teammate that starts from a truncated brief researches the wrong thing, and nothing downstream would notice:

python3 .claude/skills/_shared/validate_doc.py \
  --contract feature-brief --file .claude/docs/research/feature-{slug}-brief.md

Exit 0 every required section is present (variant auto-detected); 1 bad args or unreadable/never written; 2 a required section is missing, listed in sections_missing.

Phase 2G teammates receive the brief path as shared context and read the file, so lead and teammates work from the same bytes.


MODE=greenfield — Phase 2G: RESEARCH & DESIGN (Agent Teams — Parallel)

Launch Researcher and Architect in parallel via Agent Teams with bidirectional communication.

Key difference from subagents: Teammates can communicate with each other. Researcher's findings change Architect's design, and Architect's requests trigger new research.

Team Setup

Create an agent team for project planning: {feature}

Spawn two teammates:

1. **Researcher** — Uses WebSearch/WebFetch for external research (Opus 1M context)
   Prompt: "You are the Researcher for project: {feature}.

   Your job: Research external information needed for this project.

   Project Brief: read .claude/docs/research/feature-{slug}-brief.md

   Tasks:
   1. Research libraries and tools: usage patterns, constraints, best practices
   2. Find latest documentation and API specifications
   3. Identify common pitfalls and anti-patterns
   4. Look for similar implementations and reference architectures

   How to research:
   - Use WebSearch for comprehensive research:
     WebSearch: '{topic} best practices constraints recommendations'
   - Use WebFetch for targeted documentation lookup

   Save all findings to the `research` path from Step 0-b (.claude/docs/research/{slug}.md).
   Save library docs to .claude/docs/libraries/{library}.md

   Communicate with Architect teammate:
   - Share findings that affect design decisions
   - Respond to Architect's research requests
   - Flag constraints that limit implementation options

   IMPORTANT — Work Log:
   When ALL your tasks are complete, write your work log to
   .claude/logs/agent-teams/{team-name}/researcher.md per the shared format:
   .claude/skills/_shared/work-log-format.md
   Role-specific sections (between Tasks Completed and Communication):
   ## Sources Consulted
   - {URL or source}: {what was found}
   ## Key Findings
   - {finding}: {relevance to project}
   "

2. **Architect** — Uses Codex CLI for design and planning
   Prompt: "You are the Architect for project: {feature}.

   Your job: Use Codex CLI to design the architecture and create implementation plan.

   Project Brief: read .claude/docs/research/feature-{slug}-brief.md

   Tasks:
   1. Design architecture (modules, interfaces, data flow)
   2. Select patterns (considering existing codebase conventions)
   3. Create step-by-step implementation plan with dependencies
   4. Identify risks and mitigation strategies

   How to consult Codex:
   Write the question to .claude/logs/codex/prompt-<topic>.md, then:
   python3 .claude/skills/_shared/codex_consult.py --prompt-file .claude/logs/codex/prompt-<topic>.md --label <topic> --sandbox read-only
   Read the answer from the JSON output's response_file.

   Record architecture decisions in .claude/docs/DESIGN.md through the shared
   writer — never by editing the file. The lead writes the same document in
   Phase 3, so a direct edit is a lost update:
   write the typed JSON to .claude/logs/design-input-{slug}-architect.json
   (keys: decisions / tech_choices / agent_roles / section_updates — table rows
   only through their typed key), then:
   python3 .claude/skills/_shared/update_design.py --input .claude/logs/design-input-{slug}-architect.json
   # review the preview path in the JSON output, then:
   python3 .claude/skills/_shared/update_design.py --input .claude/logs/design-input-{slug}-architect.json --apply --require-change
   Verify "ok": true and "result": "applied". Exit 2 = invalid structure or a
   no-op; exit 3 = DESIGN.md changed concurrently — re-read it and redo the
   dry-run before applying. Report an exit 2 or 3 in your work log.

   Communicate with Researcher teammate:
   - Request specific library/tool research
   - Share design constraints that need validation
   - Adjust design based on Researcher's findings

   IMPORTANT — Work Log:
   When ALL your tasks are complete, write your work log to
   .claude/logs/agent-teams/{team-name}/architect.md per the shared format:
   .claude/skills/_shared/work-log-format.md
   Role-specific sections (between Tasks Completed and Communication):
   ## Design Decisions
   - {decision}: {rationale}
   ## Codex Consultations
   - {question asked to Codex}: {key insight from response}
   "

Wait for both teammates to complete their tasks.

Verify the Team Run (before Phase 3)

Both teammates were told to write a work log; a teammate that died mid-task, or wrote a log missing Issues Encountered, is otherwise indistinguishable from success — and Phase 3 would then synthesize from an incomplete run:

python3 .claude/skills/_shared/validate_doc.py \
  --contract work-log --dir .claude/logs/agent-teams/{team-name}/ --expect-files 2

Gate on files_failed == 0. Exit 0 both logs exist and satisfy the contract; 1 bad args or the team directory does not exist; 2 a required section is missing (see results[].sections_missing) or the directory does not hold exactly 2 logs — --expect-files 2 is what makes "no teammate wrote a log" distinguishable from "all logs valid". Do not proceed on a non-zero exit: find out what the missing teammate did or did not do first.

Why Bidirectional Communication Matters

Example interaction flow:

Researcher: "httpx has a connection pool limit of 100 by default"
    → Architect: "Need to add connection pool config to design"
    → Architect: "Also research: does httpx support HTTP/2 multiplexing?"
    → Researcher: "Yes, via httpx[http2]. Requires h2 dependency."
    → Architect: "Updated design to use HTTP/2 for the API client module"

Without Agent Teams (old subagent approach), this would require:

  1. Researcher subagent finishes → returns summary
  2. Claude reads summary → creates new Codex subagent prompt
  3. Codex subagent finishes → returns summary
  4. If Codex needs more info → another researcher subagent round

Agent Teams collapses this into a single parallel session with real-time interaction.


Phase 3 (shared): PLAN, APPROVE & IMPLEMENT

Both modes converge here: synthesize, get user approval, then route implementation by complexity.

Step 1: Synthesize Results

  • MODE=existing: the Feature Brief at .claude/docs/research/feature-{slug}-brief.md (including its recorded Complexity Classification) + Codex architecture / plan / validation outputs.
  • MODE=greenfield: read .claude/docs/research/{slug}.md (Researcher findings), .claude/docs/libraries/{library}.md (library docs), .claude/docs/DESIGN.md (Architect decisions).

In MODE=greenfield, validate each library doc the Researcher's work log claims it wrote, before its constraints are built into the plan:

python3 .claude/skills/_shared/validate_doc.py \
  --contract lib-doc --file .claude/docs/libraries/{library}.md

Exit 0 the doc has its sections and its > **Last Updated**: / > **Version Checked**: metadata; 1 the file is missing — the Researcher claimed a doc it never wrote; 2 a required section or metadata line is missing (sections_missing / metadata_missing). Validate one file per doc: running --dir .claude/docs/libraries/ also checks every pre-existing doc in the repository, which is a different question from "did this feature's research produce usable docs". A Researcher that recorded external library constraints in prose only, with no doc at all, is a finding — say so rather than silently proceeding.

Step 2: Create Task List

Create the task list using TodoWrite:

{
    "content": "Implement {specific task}",
    "activeForm": "Implementing {specific task}",
    "status": "pending"
}

Task breakdown should follow references/task-patterns.md.

Step 3: Update DESIGN.md and Shared State

Per the common protocols above (DESIGN.md Update / Shared State Update).

Gate Phase 3 on the artifacts this phase actually consumes, before presenting the plan. MODE=existing:

python3 .claude/skills/_shared/workspace.py --skill feature --slug {slug} --verify

MODE=greenfield consumes the Researcher's research artifact as well, and that key is not required by default — so name it explicitly, otherwise the gate verifies the scan and stays blind to the file Step 1 just read:

python3 .claude/skills/_shared/workspace.py --skill feature --slug {slug} \
  --verify --require research

Exit 0 means every required artifact (brief, codebase_scan, plus each --require key) exists and is non-empty; exit 1 is bad args or an unknown --require key; exit 2 means one is missing or effectively empty — read verify.missing in the JSON. Do not present a plan built on a brief or a scan that was never written.

Step 4: Present to User (approval gate)

## Feature Plan: {feature}

### Mode
{existing / greenfield} — {1-line rationale}

### Codebase / Scope Analysis
{Key findings from Phase 1 — 3-5 bullet points}

### Research Findings (greenfield: Researcher)
{Key findings — 3-5 bullet points; library constraints and recommendations}

### Complexity
- Classification: {SIMPLE / MODERATE / COMPLEX}
- Implementation route: {Codex direct / Codex + review / team-execute}

### Architecture Design (Codex / Architect)
{Architecture overview}
{Key design decisions with rationale}

### Implementation Plan ({N} steps) — Codex Validated: {PASS} (existing mode)
1. {Step 1}: {description}
2. {Step 2}: {description}
...

### File Changes Summary
| File | Change Type | Description |
|------|------------|-------------|
| {file} | {new/modify} | {what changes} |

### Test Plan
- {Test 1}: {what it verifies}

### Risks and Mitigations
- {Risk}: {mitigation}

---
Shall we proceed with this plan?

Do not implement until the user approves the plan.

Step 5: Complexity Routing

Greenfield features usually classify as COMPLEX (Route C); existing-mode features use the classification recorded in the brief's ### Complexity Classification section — read it from .claude/docs/research/feature-{slug}-brief.md rather than re-deciding it here, so the route matches what the user approved.

Completion Verification (MANDATORY on every route)

Every route below hands implementation to an agent that reports on its own work. A self-report is never completion evidence (AGENTS.md Guardrails), so all three routes end with the same two executable checks — the only difference between the routes is who wrote the code, not how much verification it gets.

1. Quality gates:

bash .claude/skills/_shared/verify.sh

Read the JSON: overall is pass / fail / no_gates. Exit 0 means overall: pass. Exit 2 means a gate failed, or no gate could run at all — inspect log_file and tools; no_gates is a failure by default because an implementation must not be declarable done with zero checks executed. If the project genuinely has no configured gates, re-run with --allow-no-gates, verify manually with the project's own commands, and say so in the report. Exit 1 bad arguments, 3 the log could not be written.

2. Diff evidence (the other half of the Guardrails):

python3 .claude/skills/_shared/verify_delegation.py \
  --base {ref the delegated run started from} \
  --expect-files {file the plan said would change} \
  --forbid-outside {directory the plan scoped the change to} \
  --label route-{a|b|c}

--expect-files and --forbid-outside each take a repo-relative path and are repeatable: name the files the approved plan named, and the directories it scoped the change to. --base defaults to HEAD, which is what a Codex run that left the tree dirty needs.

Read deletions, placeholders, weakened_tests, and out_of_scope_files. verdict is always needs-review: the script collects evidence and never accepts a delegated run on your behalf, so read the reported hunks and decide. Exit 0 nothing actionable and no violated expectation — deletions alone land here, reported but not actionable on their own, and exit 0 is still not an accept, so read the diff; 1 bad args or a --base that does not resolve; 2 an actionable finding (placeholders, weakened_tests) or a violated expectation (a missing expected file, an out-of-scope file, an empty scope); 3 git failed or the diff could not be written.

Use .claude/skills/_shared/gather_diff.py --base {ref} when you want the full patch to read: scope_empty: true with exit 2 means the delegated run changed nothing at all — a failed implementation, not a clean one.

Route A: SIMPLE (1-3 files, <50 LOC) — Codex Direct

Codex implements directly. Write the prompt body below to .claude/logs/codex/prompt-route-a-implement.md, then invoke with write access:

python3 .claude/skills/_shared/codex_consult.py \
  --prompt-file .claude/logs/codex/prompt-route-a-implement.md \
  --label route-a-implement --sandbox danger-full-access
Objective: Implement this feature following the approved plan.
Context:
- Feature Brief: contents of .claude/docs/research/feature-{slug}-brief.md
- Architecture Design: {from Phase 2}
- Implementation Plan: {from Phase 2}
- Existing conventions: {from Phase 1 codebase scan}
Constraints:
- Follow the implementation plan steps exactly
- Follow existing codebase conventions (naming, structure, patterns)
- Write tests for all new functionality
- Keep changes minimal and focused
Relevant files:
- {list of files to create/modify}
Acceptance checks:
- All new tests pass
- Existing tests still pass
- Code follows existing conventions
Output format:
## Changes Made
## Tests Written
## Validation Results
## Remaining Risks

Read the implementation summary from the JSON output's response_file — as input to the verification, never as its result.

Then run Completion Verification above (both checks).

Route B: MODERATE (3-5 files) — Codex + Review

  1. Implement with Codex (same prompt and --sandbox danger-full-access as Route A, with more files)
  2. Run Completion Verification above — both checks, same exit-code reading. Route B changes more files than Route A, so it gets no weaker a gate: record verify.sh's overall and the verify_delegation.py payload before moving on
  3. Hand off to /team-execute --review-only for parallel review (security, quality, test coverage), passing the slug from Step 0-b
After Codex implementation and Completion Verification:
/team-execute --review-only   <- Parallel review from multiple perspectives

Route C: COMPLEX (5+ files) — Team Execute

/team-execute   <- Phase 1: parallel implementation, Phase 2: parallel review

Hand /team-execute the brief path (.claude/docs/research/feature-{slug}-brief.md), the Architecture Design and Implementation Plan from Phase 2, and the slug resolved in Step 0-b — /team-execute reuses the slug verbatim, so its research and design artifacts resolve to the same files this skill wrote. Its work logs land in .claude/logs/agent-teams/team-execute-{slug}/, deliberately separate from this skill's feature-{slug}/ team directory: the slug is shared, the team directory is per-skill. Do not look for Phase 2G logs under the team-execute directory.

When /team-execute returns, run Completion Verification above yourself. Its own review phase is a teammate's report on teammates' work; the gates and the diff evidence are what close the route.

Post-Implementation

If the .claude/STATE.md block was not written at plan time (existing mode), write it now per the common protocol.


Output Files

Paths below are resolved once by workspace.py in Step 0-b (see paths in its JSON) rather than hardcoded here.

FileAuthorPurpose
.claude/docs/research/feature-{slug}-brief.mdLeadFeature / Project Brief — validated with --contract feature-brief
.claude/docs/research/feature-{slug}-codebase.mdOpus SubagentCodebase scan
.claude/docs/research/{slug}.md (greenfield)ResearcherExternal research findings
.claude/docs/libraries/{lib}.md (greenfield)ResearcherLibrary documentation
.claude/docs/DESIGN.md (updated)Lead / Architect (Codex-informed)Architecture decisions
.claude/STATE.md (updated)LeadCross-session feature context
.claude/logs/agent-teams/feature-{slug}/*.md (greenfield)Researcher / ArchitectWork logs — validated with --contract work-log --expect-files 2
Task list (internal)LeadImplementation tracking
Implementation filesCodex / Agent TeamsThe feature itself
Test filesCodex / Agent TeamsTests for the feature

Tips

  • Mode first: The biggest failure mode is picking the wrong path. When ambiguous, always AskUserQuestion — never guess
  • Codex-first (existing mode): Every phase consults Codex. Codex excels at understanding how new code fits into existing patterns and identifying integration risks; early scope classification picks the right implementation route from the start; validation catches missing edge cases and convention violations before implementation begins
  • Existing patterns: The most important input to Codex is the existing codebase patterns from the Opus subagent scan — include them in every Codex prompt
  • Agent Teams (greenfield mode): Bidirectional communication lets Researcher (Opus) and Architect (Codex) influence each other in real time
  • Complexity routing: Do not over-engineer simple features. 1-3 file changes should use Codex direct implementation, not Agent Teams
  • Quality gates: Every route ends with Completion Verification — verify.sh (gate failure or no gate at all is exit 2) plus verify_delegation.py diff evidence. A Codex or teammate summary is input to that check, never a substitute for it
  • Artifacts, not transcripts: the brief, the scan, the research file and the work logs are files with contracts. If a phase cannot point at a validated file, the phase did not happen
  • Ctrl+T: Toggle task list display
  • Shift+Up/Down: Navigate between teammates (when using Agent Teams)

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.