implement

implement - Create a branch and implement from an Issue/instruction

Based on reading an Issue or a direct instruction, this performs everything from branch creation through implementation planning to code changes.

Input

• If an Issue number (#N or a bare number): fetch the content with gh issue view <N> and organize the requirements
• If a text instruction: treat it as-is as the requirements
• If no argument: confirm what to implement

Action classification and policy reference

This skill does not hardcode individual approval gates into prose; instead it follows the central autonomy policy (the SSOT for the 3 classes, gate vocabulary, and policy.yaml hierarchy defined by the policy skill). It classifies its own actions into the following classes, resolves the effective gate, and then executes:

Action in this skill	Class	policy reference	Zero-config default gate
Implementation on the branch (file edit/add/safe delete)	reversible-local	--action=branch-edit	proceed (present the plan and continue + audit trail)
Changes involving migration / data deletion / CI or release config changes	irreversible	--class=irreversible	ask (explicit approval before starting)

Resolve the effective gate with policy-read.mjs in the sibling policy skill (for Claude Code user-scope this is ~/.claude/skills/policy/policy-read.mjs, for dogfooding <repo>/.claude/skills/policy/policy-read.mjs, and for Codex/Gemini .agents/skills/policy/policy-read.mjs):

node <policy skill のディレクトリ>/policy-read.mjs --action=branch-edit --repo-root="$PWD"
# => .resolved.gate（既定 proceed）
node <policy skill のディレクトリ>/policy-read.mjs --class=irreversible --repo-root="$PWD"
# => .resolved.gate（既定 ask）

The gate vocabulary has only 3 values:

• proceed: execute without waiting for approval, and leave the plan/changes in the report as an audit trail
• batch-confirm: confirm once, all together, before starting
• ask: obtain explicit approval (Approval Gate) for each action

Does not break even if policy is absent: policy-read.mjs is fail-safe by design — unreadable or invalid values always fall back to the stricter side (ask). In environments where the policy skill itself is not deployed and policy-read.mjs cannot be called, apply the zero-config default gates from the table above directly (reversible-local=proceed / irreversible=ask).

Consistency under drive: drive is delegated autonomous execution by the user, and the reversible-local default of proceed (skip plan approval and continue) is consistent with this delegation. Standard branch implementation proceeds with proceed, and only changes judged irreversible require approval with gate=ask.

Procedure

1. Create a branch

1. Check the current state with git status and git branch --show-current
2. Determine a branch name in <type>/<slug> format from the requirements
3. Create the branch with git checkout -b <branch-name>

If already on a feature branch, confirm whether to continue working on that branch.

2. Implementation plan and action classification

1. Investigate the codebase
   • Identify relevant files
   • Understand existing implementation patterns
   • Check the scope of impact
2. Present the implementation plan:
   • Files to change and their content
   • Scope of impact
3. Classify the change into a class using the table in "Action classification and policy reference," and resolve the effective policy gate:
   • Only ordinary implementation on the branch → reversible-local (default proceed): present the plan and continue (do not wait for approval; the plan is kept as an audit trail)
   • Includes migration / data deletion / CI or release config changes → irreversible (default ask): obtain explicit approval before proceeding to implementation

3. Implementation

Execute code changes according to the gate resolved by policy (proceed starts without waiting for approval, batch-confirm confirms all at once before starting, ask starts after approval). Report progress as each file's changes are completed.

4. Verification (verify)

After implementation is complete, refer to .agents/skills/verify/SKILL.md and run the combined verification (build + type + test + lint) with the verify engine. verify auto-discovers verification commands through a discovery chain (AGENTS.md's 「検証」 section → package.json scripts → task runner → language heuristics) and runs them serially, each with its source attributed.

If errors occur, fix them on the spot and run verify again.

5. Completion report

実装完了:
  ブランチ: <branch-name>
  変更ファイル:
    A path/to/new-file
    M path/to/modified-file

Notes

• Do not read or stage .env files
• If the gh CLI is not authenticated, display an error message and abort
• Classify actions into the 3 classes and resolve policy before executing. Do not hardcode individual approval gates into prose
• For changes judged irreversible (migration / data deletion / CI or release config changes), always obtain explicit approval under gate=ask before starting