Issue To PR
Use only when the user explicitly invokes this lane by name or directly
commands this exact pipeline, or when user-invoked epic-to-prs delegates an
issue. Ordinary issue work does not trigger it.
Take the issue to a merge-ready PR with the smallest change that delivers its accepted scope. Use Pro for initial planning and final review; use the working agent's judgment to carry the work between them.
For an Astra or Fable coordinator, apply $delegated-implementation throughout
the work: delegate code, reproduction, tests, and repairs; personally review
every deliverable and changed code line. All skill authorship stays with the
parent. An assigned implementation worker keeps its bounded role and reports
to the originating coordinator under that contract.
Install
git clone git@github.com:aelaguiz/arch_skill.git
cd arch_skill
make install
When to use
- "issue-to-pr on ", singly or as a list.
- A direct command to plan an issue, consult Pro, implement, test, publish a PR, get Pro's final review, and stop at merge-ready.
- An issue handed over by user-invoked
epic-to-prs.
For investigation-only work, standalone planning or review, or work without a GitHub issue, use the requested workflow instead.
Delivery contract
- Freeze acceptance and non-goals from the issue and the user's direction. Reviewers and bots cannot expand or silently shrink that scope. Resolve routine scope interpretation from the contract; a change to what the user asked for belongs to the user.
- Write the implementation plan on disk before implementation and apply
$startup-pragmatism. Keep the same judgment during delivery: enough investigation and verification for the actual change, without invented approval gates or proof machinery. - The run starts authorized for accepted in-scope work. Work in a dedicated worktree under the target repo's AGENTS.md. Require self-documenting code with clear comments at boundaries and role seams, relevant tests, and required repo checks. The execution contract above determines who authors each deliverable and who runs verification.
- Use
$pr-authoringand$pr-review-followthrough. PR Agent and other bots are advisory; assess findings against the issue and code rather than treating them as orders to expand scope. - Stop at merge-ready with receipts. Never merge, release, apply approval
labels such as
ufc-approved, or touch production surfaces.
Pro cadence
For a standalone issue, the normal cadence is one initial planning consultation and one final PR review. Existing Pro planning that still covers the accepted scope satisfies the first consultation; do not repeat it just because this skill was invoked or a session resumed.
For an issue inside a coherent epic or related batch, inherit the coordinator's shared planning and review scope. Do not add a plan review and final review for every child. The coordinator can collect locally finished PRs for a meaningful batch checkpoint or final stack review. A child awaiting that review is locally ready, not yet merge-ready.
The coordinator owns routine decisions, plan refinements, dependency ordering, and scope checks. Workers own implementation details and code repairs under the execution contract. Extra Pro consultations are appropriate when a meaningful batch of related work is ready to assess, or a major unexpected blocker or consequential technical uncertainty remains beyond the agent's reasoning after reasonable local investigation. State what the consultation can resolve and why it matters to the goal. Neither an issue boundary, a changed plan, a newly discovered dependency, nor vague uncertainty alone requires a Pro message. A useful checkpoint after two related issues is welcome; an every-two-issues rule is not.
Assess Pro's findings with judgment and batch the warranted fixes. Have workers repair code and run affected checks, then review the changes and evidence directly; the parent repairs skill content. Do not automatically resubmit plans or PRs until Pro approves every edit. Consult again when a substantial redesign, unresolved consequential disagreement, or a repair that changes the basis of the review needs independent judgment. The normal planning/final pair is a baseline, not a hard cap on useful consultation.
Finish expected edits and relevant checks before final review where practical, including known CI and review-thread repairs. Honor an explicit request to run Pro and CI concurrently. After review, assess any change to the reviewed revision: formatting or a rebase with unchanged behavior does not by itself require another Pro run. Changes that materially affect behavior, integration, or the conclusions of the review may need a consolidated recheck; decide from their impact, not the SHA changing.
Consulting Pro
Use GPT-6 Astra with the literal Pro option and Extended thinking through
$chatgpt-web. Read and apply $browseros before BrowserOS calls. Verify
Pro in ChatGPT's Chat surface: Extra High, xhigh, Ultra, Thinking, and the
highest remaining setting are not Pro. Follow $chatgpt-web for input delivery
and account switching when Pro is missing, disabled, or explicitly capped.
Use the run's existing Pro thread, inheriting the epic's thread when
applicable. Apply $prompt-authoring to every submission. Include the
user's intent, accepted scope, relevant plan or PR artifacts, progress,
important findings, and the believed critical path. Ask Pro to assess the
work against the goal and the consequential decisions; provide enough
context to reason instead of requesting a context-free approval.
Record each actual Pro submission once in the existing worklog: purpose, artifact/revision and thread, plus the running submission count. Include retries and failover submissions; polling an existing response is not a new consultation. Keep this a short entry, not a separate tracking system.
Missing Pro probably means a temporary account rate limit. Under $chatgpt-web
and $browseros, find an available account among the already-open Work and
Pro profiles, note which profile/window works, and use its same-named project
with the needed context. The number of Pro profiles varies.
All should have the same projects. Only after the available accounts are
exhausted, report their observed conditions and pause the blocked Pro
consultation or decision. Continue independent authorized work; pause
the whole run only when no useful independent work remains. Wait for the
user to say Pro is available again. Do not substitute another model for a
required Pro review or claim a pending review passed.
Workflow
- Ramp up and plan. Read live issue, parent, linked PRs, and discussion. Confirm the issue is open, available, and not already fixed; reproduce a bug before planning its fix, delegating reproduction under the execution contract. Write acceptance, non-goals, implementation, and appropriate verification on disk. Obtain or inherit the initial Pro planning consultation and incorporate warranted findings locally.
- Implement and verify. Deliver the smallest coherent change in the worktree under the execution contract. Give workers tight requirements and appropriate checks, review every deliverable and changed code line, and return code findings for repair. Author skill content directly. Resolve ordinary decisions locally, consult the run's unblocker when needed, and use Pro at the cadence above.
- Publish and stabilize. Use both PR skills to publish the PR and handle review threads and CI, retaining the same authorship, test, and direct-review responsibilities during follow-through. For shared reviews, hand the coordinator the PR, revision, verification, and unresolved findings without launching duplicate child reviews. Independent issues can proceed meanwhile.
- Final review and repair. Submit the stable PR, or have it included in the coordinator's batch/stack review. Route accepted findings through the same execution contract, personally review repairs, and decide whether their impact warrants a Pro recheck.
- Report merge-ready. Require completed Pro planning and final review coverage, the originating coordinator's direct review of every deliverable and changed code line, resolved material findings, and passing required checks. Report PR URL, change summary, current head and CI, the revision Pro actually reviewed, and any later changes with their local verification. Include the Pro thread/verdict and submission count. Never imply Pro reviewed a newer revision it did not see.
Unblocking and persistent goals
Investigate blockers from the accepted scope and current evidence. If an unblocker is armed, take it the intent, blocker, attempted reasoning, options, and recommendation. It should settle routine authorization and engineering decisions; escalate to Pro only for the consequential unresolved problems described above. Without an unblocker, make those decisions locally under the same standard. A matter needing the user's authority, access, or a change to their ask gets one concise question with a recommendation. Continue independent scope while the answer pends; a continuation or wake-up is not an answer or a reason to ask again.
For a persistent run, author its goal prompt with $prompt-authoring and
$startup-pragmatism, naming the Pro thread, review scope and cadence,
unblocker per $unblocker, accepted scope, execution responsibilities including
parent-owned skill authorship, and merge-ready completion condition. Arm the
goal and unblocker using the active harness's supported mechanisms.
When adopting a user-directed cadence change during a run, update the goal,
unblocker charter, and active dispatch briefs so they carry the same rule.
Delegation
For Astra and Fable coordinators, $delegated-implementation owns the worker
selection, brief, direct review, and repair contract, including parent-owned
skill authorship. For other dispatches, read the installed
../_shared/agent-orchestration-policy.md and apply $prompt-authoring to the
populated brief. Leave spawning mechanics to the active harness. Carry the
scope, inherited review coverage and cadence, unblocker contact, and handoff
into each brief. The originating coordinator owns shared Pro submissions and
final acceptance; children return artifacts without duplicating consultations.
References
- references/dispatch-evidence.md: historical dispatches and owner corrections for maintainers.