Issue

Implement GitHub Issue

Overview

Own one ready GitHub issue from evidence gathering through a reviewable pull request. Remove routine reminders while retaining readiness, TDD, verification, and human publication gates.

Required workflow

Read references/workflow.md completely before acting. That file is the shared workflow body, symlinked to the same definition the Codex $issue skill uses, so both runtimes stay in step. This SKILL.md carries only the Claude Code specifics.

Treat the repository's AGENTS.md, current code/tests/Git state, and committed decisions as higher authority than this skill.

Extract exactly one issue number or URL from the invocation. If none is present, ask one concise question and stop.

Delegation

Use the issue-reader agent for read-only exploration and independent review. Dispatch it with the Agent tool and let its definition supply the model; it is pinned to haiku per the global model-routing rule for read-only investigation. Do not override that to a premium model for a routine scan.

Send parallel issue-reader dispatches in a single message so they run concurrently. Cap discovery at two agents, per the workflow's economy rule.

If issue-reader is unavailable, do not silently substitute a premium read-only agent. Perform the read in the main session, state the limitation once, and continue.

Delegate implementation only when independent write scopes genuinely justify the coordination overhead. Verification stays in the main session; the one exception is the mandatory independent adversarial review, which must run in a context that did not implement the change.

Enforcement limitation, state it rather than assume it away: Codex pins issue-reader to a read-only sandbox. Claude Code has no per-agent sandbox, so issue-reader is constrained by its tool list and its instructions, plus the global bash-guard PreToolUse hook. Treat read-only as a contract the agent is instructed to honor, not one the runtime enforces.

Claude Code process skills

The workflow calls for the repository-required brainstorming, TDD, and debugging workflows. In Claude Code those are:

• superpowers:brainstorming before design work on anything creative or new;
• superpowers:test-driven-development before any production code;
• superpowers:systematic-debugging before proposing any fix for unexpected behavior;
• superpowers:verification-before-completion before claiming the slice is done.

Invoke them through the Skill tool. They are gates, not suggestions.

Hard boundaries

• One issue, one branch, one PR.
• Do not implement blocked, unready, ambiguous, or dependency-gated work.
• Do not merge, force-push, push feature work to main, expose secrets, or widen external authority.
• Do not stage private brain content or unrelated user changes.
• Stop before Git publication until {{USER}} says green hand.
• In this workflow, green hand authorizes scoped commit, feature-branch push, ready PR creation, and the exhaustive issue comment as one sequence. It authorizes nothing beyond that sequence.
• Add Closes #N; let Sage's merge close the issue. Do not close it while the PR is open.

Completion output

Lead with the outcome. Give the branch, commit, PR, issue comment, CI state, external cleanup state, and the single next human action.