Astra Ticket

Input: a GitHub issue URL or owner/repo#number.

• Before task work, verify the active orchestrator is GPT-6 Astra
  (gpt-6-astra) using authoritative runtime/session metadata.
• Stop if different or unverified; defaults and user assertions are not proof.
• Except QA, use gpt-5.6-luna at max for every subagent.
  Verify model/effort support before spawning; stop if unavailable.
  Never substitute another model.
• Set model and effort explicitly on every spawn; use fresh context when
  required. Pass applicable workflow overrides, ticket, workspace, and artifacts.
• Reuse passing checks at the same commit; after changes, rerun affected checks unless a concrete finding requires broader validation.
• Before each review and final handoff, inspect human/bot feedback and CI; address actionable findings, report unfinished checks after five minutes, and claim ready only with actionable threads resolved and required checks passing on the final SHA.
• For low-risk changes whose correctness is quickly verifiable with deterministic checks—such as copy, translations, docs, formatting, metadata, or simple config—mark internal Luna/Astra reviews (steps 6–8) skipped and default step 5 QA to skipped unless requested; retain applicable checks, affected-flow retests, final-head CI, and available GitHub feedback handling.

1. Read the issue/comments and relevant artifacts in repo TMP/ or tmp/,
   $TMPDIR, and /tmp; check stale context against the ticket and code.
   If Grain is connected, reuse the linked workspace. New work defaults to Development Artifacts/<org>/<repo>; explicit destinations win. Clarify ambiguous names.
   Verify organization, folder, and audience; return the location with the link. Rename the workspace for the PR, retaining its ID and shared folder. Without Grain, continue locally.
   When Grain is connected, keep all development artifacts in one canonical task workspace in that folder, retain local copies, and pass its workspace ID to every subagent.
2. As Astra, read simple-plan and use its planning steps to investigate and
   produce a concise plan. Preserve the ticket's intent, constraints, and scope.
   Use create-ticket's intent guidance; update the linked brief with the approach and tradeoffs before coding.
   Reuse Socrates' verdict while its premise and evidence hold. Otherwise dispatch fresh Luna Max Socrates with the ticket, plan, and repository evidence. Resolve material findings with the user; if existing behavior meets the outcome, finish with evidence and guidance.
   Autonomously derive detailed specs: files, changes, dependencies, edge cases,
   acceptance criteria, and checks. Save the plan/specs under task-specific tmp/.
   This workflow authorizes proceeding without simple-plan's routine approval pause once premise findings are resolved.
3. Spawn Luna Max implementers with the ticket, plan, specs, and artifacts;
   sequence dependent tasks, wait for completion, and inspect work/check results.
4. Spawn a Luna Max agent to use prepare-pr and open or update the PR.
   Pass explicit code/branch-preparation and draft-PR overrides to prepare-pr; after QA (or an explicit skip) and internal reviews finish, mark ready, then check any triggered automated reviews and final-head CI and address actionable feedback before handoff.
5. Before reviews, ask asynchronously whether to run end-to-end PR QA; state
   it will run after 60 seconds without a reply. Yes starts QA; no skips it.
   Use a timed, interruptible wait; do not treat a pending question as silence
   before 60 seconds elapse. If timed input is unavailable, wait for an answer.
   For QA, spawn GPT-5.6 Sol (gpt-5.6-sol) at reasoning effort medium
   using pr-test-automation; verify model support without substitutions.
   Delegate QA bug fixes and pushes to Luna Max, then have Sol rerun affected
   flows before reviews. Report blocked QA honestly; never call it a pass.
   Capture QA screenshots; when Grain is connected, save and verify media/reports there instead of release assets; otherwise follow pr-test-automation's durable-publication behavior.
   Carry the QA handoff into the final brief, separating future improvements from current blockers.
   Return the same verified evidence link in the final handoff; explicitly report unavailable capture or publication rather than claiming upload.
6. Run up to three fresh Luna Max reviews sequentially; stop after a clean review. Use
   the review skill on the current PR. After each review, delegate actionable
   fixes, checks, and pushes to Luna Max before starting the next review.
   For all reviews, use COMMENT when authenticated as the PR author.
7. Review the resulting PR yourself as Astra, using the same review skill.
   Delegate any fixes to Luna Max, then verify those fixes yourself.
8. If review fixes invalidate completed QA, Sol retests affected flows, then
   one fresh Luna Max agent reviews the fixes covered by that rerun.
   Report remaining findings; do not restart the review loops.
9. Return the PR URL, QA/check results with tested commits, and open findings. Do not merge.
   After confirming an authorized or existing merge, mark the Grain task complete; move its workspace to the configured completed-work destination when supported. Preserve IDs, shares, and the shared folder; verify location/links and report move failures.
   After QA (or a skip), reviews, and final-head CI, extend the same brief from the published PR with before/after behavior and verified results. Retain intent, sources, and decisions; verify content and reciprocal links, then open Grain last.

Preserve required tracker publication, local paths, and evidence contracts alongside Grain.
Use the bundled sibling ../review/SKILL.md and its CRITERIA.md, resolving paths from this skill's directory.
Read referenced criteria; report missing skills. These workflow overrides take precedence.