local-automation-runtime-operate

atlas-tools-generated: source=skills/local-automation-runtime-operate/SKILL.md manifest=atlas-tools.v1 checksum=sha256:02b81fed648c74023c5d3e4635a885f0160383c7a22d881da2ecd5b74dd0121d

atlas-tools-generated-end

name: local-automation-runtime-operate description: "Operate the local automation runtime: queue issues, run bounded or unattended cycles, long shift supervision, review/validate/repair/finalize PRs, and summarize outcomes. Use when running issue-to-PR automation after setup."

Local Automation Runtime Operate

Purpose

Run the local GitHub issue-to-PR lane with explicit bounds, review gates, validation, repair, finalization, and summaries. Use repos.txt to add target repos; use another runtime only for isolation, secrets separation, or intentional concurrent lanes. Run these commands from an installed runtime directory. Do not operate from templates/local-automation-runtime/ inside the tools repository; that path is source material and direct execution will create runtime state in the template tree.

Operator Roles

• Codex operator owns runtime host, secrets, worker execution, validation, repairs, finalization, and merges.
• Gemini operator may own planning, issue quality, semantic review notes, and readiness labels.
• Only one operator may run mutating flags at a time: --apply, --publish, --merge, or --close-issues.

Operating Flow

1. Refresh repo/project state:
   • ./atlas-agent-reconcile --repos-file repos.txt --apply
   • ./atlas-agent-project-reconcile --projects-file projects.txt --apply
2. Preview queue eligibility:
   • ./atlas-agent-plan-queue --plan ../path/to/plan.md --repo OWNER/REPO --dry-run
3. Audit manual/dependency-gated items before queue approval:
   • Use project-queue-audit for items with Manual, review-before-dispatch, dependency, or blocker state.
   • Do not approve dispatch if issue body runtime fields and Project fields disagree.
4. Queue approved work:
   • ./atlas-agent-plan-queue --plan ../path/to/plan.md --repo OWNER/REPO --apply --queue
   • Add --publish only when eligible queued issues should immediately open draft PRs.
5. Run a bounded cycle:
   • ./run_e2e_chain.sh --cycles 1
   • Add mutating flags only after preview looks correct.
6. Run unattended cycles when the lane is stable:
   • ./atlas-agent-unattended --cycles 3 --dispatch-max-per-repo 2 --review-apply --post-cycle-summary
7. Run a long bounded shift when the lane is stable and durable resume state is needed:
   • ./atlas-agent-shift --cycles 12 --max-minutes 480 --sleep-seconds 300 --publish --apply --review-apply --post-cycle-summary
   • Use the shift wrapper for wall-clock limits, a supervisor lock, heartbeat/status JSON, and an exit handoff.
8. Review blockers:
   • ./atlas-agent-review --summary review.json
   • ./atlas-agent-semantic-review OWNER/REPO#PR --apply
9. Validate and repair:
   • ./atlas-agent-local-validate OWNER/REPO#PR --apply
   • ./atlas-agent-deployed-validate OWNER/REPO#PR --apply
   • ./atlas-agent-pr-repair OWNER/REPO#PR
10. Finalize only when review approves:

• ./atlas-agent-finalize --required-checks-file required-checks.json --merge --close-issues

Guardrails

• Keep cycles bounded with --cycles, --dispatch-max-per-repo, and per-stage repair/review/validation max and concurrency flags.
• For multi-hour operation, use atlas-agent-shift instead of treating a Codex chat or subagent as the durable supervisor.
• Treat Open dependencies: and Manual gates remaining: as the runtime dispatch contract; Project fields are advisory unless they match the issue body.
• Keep unattended dispatch one-point only. Larger points:N issues must be decomposed or explicitly handled outside unattended dispatch.
• Treat no-check PRs as blocked unless local validation and required-check policy explicitly allow them.
• Do not repair failed workflows until atlas-agent-review --apply has classified the failure.
• For human-action, secret/config, infra/env, or dependency-blocked workflow classes, stop worker repair and hand the issue to the responsible operator.

Subagents

• Delegate bounded side tasks only: inspect one PR, classify one log set, draft one repair plan, or review one handoff/status file.
• Do not delegate the runtime control loop. atlas-agent-shift owns the long-running loop, lock, heartbeat, and handoff.
• Require subagents to return changed files, commands run, validation results, blockers, and the next safe command.
• If a subagent works across more than one runtime cycle, require a handoff note before using its result.

Output

Summarize:

• queued issues
• PRs opened
• review labels/routes
• validation results
• repairs attempted
• merged/closed items
• remaining blockers and owner actions