agents: Instructions file for Claude Code

Global Claude Code Instructions

All work follows rules/core-principles.md.

Workflow

Begin file-modifying sessions with /workflow-init; skip for read-only investigation.

Steps use prefix WF-<TYPE>-N; WF-CODE-N = standard implementation. WF-META-N = planning-only (meta issues, no worktree).

After each skill completes, run: node "$AGENTS_MAIN_ROOT/bin/workflow/next-step" --session $CLAUDE_CODE_SESSION_ID. Follow ACTION / NEXT_SKILL / NEXT_HINT output — invoke means run the named skill immediately without pausing to ask the user (reserve confirmation for CONFIRM_* gates, the USER_VERIFIED dialog that /get-release-approval issues, and in-skill AskUserQuestion), done means proceed to session close, blocked/abort means see NEXT_HINT for recovery, paused means next-step is deliberately quiet — take no workflow action and continue the user's out-of-workflow work, resuming with echo "<<WORKFLOW_NEXT_STEP_RESUME: {reason}>>" when REASON=next-step-paused or echo "<<WORKFLOW_ENFORCE_WORKFLOW_ON: {reason}>>" when REASON=workflow-off-quiet. Run node "$AGENTS_MAIN_ROOT/bin/workflow/next-step" --list for the full step plan. Emergency partial reset: echo "<<WORKFLOW_RESET_FROM_{step}: {reason}>>" (reason mandatory; marks prior steps complete, resets target step and after to pending).

Notes

• Skill procedures settle a step and read the next action in one call: next-step --advance --step <step> --complete|--skipped|--pending [--skip-reason <reason>] --next.
  • record-skip-judgment / set-deliverable-type / set-workflow-type support the same --advance [--next] pair for their own step.
  • Pass --next only when the call will consume the returned ACTION block immediately; otherwise omit it and proceed to the skill's own next documented step.
  • --next emits no ACTION= line when the settled step is not the session's current step (ADVANCE_SCOPE=not-current-step) — treat a missing ACTION= as "proceed to your own next step," not an error.
• Do not use --permission-mode plan. Always use default mode for implementation tasks.
• Mark AskUserQuestion metadata.source: jev-auto-skip-askuser only on a choice you raise whose answer the conversation already settles; detail: docs/architecture/jev.md.
• Workflow state reset is main-conversation only — emit <<WORKFLOW_RESET_FROM_{step}: {reason}>> only when holistic context justifies it.
• Not auto-injected — Read each before the work it governs:
  • rules/github-issues.md — any issue or PR operation.
  • rules/architecture-principles.md — designing or changing a mechanism, guard, gate, or check.
  • rules/branch.md — naming a branch or recording branching.
  • rules/worktree.md — creating, entering, or removing a worktree.
  • rules/ops.md — any destructive or system-state-changing operation.
  • rules/test.md — writing, reviewing, or running tests.
  • rules/coding.md — writing or editing code.
  • rules/docs.md — writing or updating documentation.
  • rules/mid-workflow-findings.md — bugs, follow-ups, or next-task findings discovered mid-workflow.
• For trivial edits that temporarily suspend workflow enforcement, see rules/workflow-off.md (when it applies) and /enforce-workflow-off (how).
• When working inside the agents repository itself, also consult docs/agents-repo-dev.md.
• When you encounter an issue, concern, or unexpected outcome that core-principles + workflow don't resolve, report it: rules/supervisor-reporting.md says when, /supervisor-report does it.
• Asked where a plan is: run node "$AGENTS_MAIN_ROOT/bin/plan-link" --step <step> and write its output as-is in the response.
• write-code is not a tracked next-step step — invoke it manually once write_tests/review_tests complete, before /run-tests.
• For keeping the Stop guard quiet during long-running work — automatic while a delegated step is in flight, NEXT_STEP_PAUSE otherwise — see rules/stop-guard-exemptions.md.