Flux de PR et PR empilées

---

name: pr-flow description: | Manage branches, stacked pull requests and PR review on a GitHub repository the user owns, through the gh CLI and the gh stack extension. Use when asked to write a branch's plan file, open its pull request, or mark a draft PR ready; to stack a branch, add a branch to a stack, submit, sync, restack or view a stack; when asked what a branch is stacked on; when asked to review open PRs, review a numbered PR, or check what needs review; when asked to merge or land a pull request or a stack; or when the user says they replied to review comments on GitHub and wants those answered; when the user says the review is done, that you can merge, or asks to push changes for review; or when asked to watch a PR for new comments, poll for replies, or stop watching one. Also use when the user says a branch is done or ready for review, or types /gh-solo:pr-flow. Assumes one committer; for the issue tracker itself use tracker skill. argument-hint: "[open | plan | auto <issue-number> | go <pr-number> | ready [review] | review [pr-number] | watch [pr-number] | unwatch | discuss [pr-number] | reply [pr-number] | chat [pr-number] | resolve [pr-number] | merge [pr-number] | view | init | add | submit | sync | restack | help]" allowed-tools: Bash(gh:), Bash(git:), Bash(python3:*), Read, Write, Edit, Grep, Glob, Monitor, TaskStop, Agent, Skill, EnterWorktree

Tools used: Bash(gh:*) for the extension, PR queries and review posting, Bash(git:*) for branch and worktree state, Read / Grep / Glob for repository context, Write / Edit for plan files and PR-body scratch files, Bash(python3:*) for scripts/docs-check.py and scripts/post-review.py, Monitor / TaskStop for the watch in workflows/discuss.md, Agent for the reviewer subagent that workflows/review.md spawns, Skill for entering the implement skill from workflows/auto.md, EnterWorktree for moving the session into a sibling worktree in workflows/merge.md and workflows/stack.md.

The user invoked this skill with the argument: $ARGUMENTS

This is a routing skill. Read $ARGUMENTS and the conversation context, pick exactly one workflow, read that workflow file, and follow its instructions inline.

All paths below are relative to this skill's own directory. Resolve them against wherever this skill is installed rather than assuming a location.

Issue, ticket and task all mean the same thing: a GitHub issue. The tracker itself belongs to the tracker skill; this one starts where a branch does.

Every file here is written to the agent, so you is the agent reading it. The human is the owner, always in the third person. The distinction is load-bearing wherever authority is: a rule about who may approve a merge inverts if the referent slips. The exceptions are workflows/help.md, whose contents are printed to the owner, and README.md, which the owner reads rendered; both address them directly.

Who this is for

The owner owns the repository and is its only committer. Others may look — a mentor who comments, the reviewer agent a round spawns — but the owner is the only reviewer with authority: nobody else's approval gates a merge and nobody else is accountable for what ships. That is what makes workflows/review.md a round that prepares the owner's judgement rather than a substitute for it, and why REQUEST_CHANGES on their own PR is a note to themselves. If a second person commits to the repository, stop and say so: the confirmation gates and the assumption that a PR's author and its reviewer are the same person both stop making sense.

Standing conventions

Placeholders name what they hold. {issue-number} and <pr-number> are different values and a PR almost never shares a number with the issue it closes, so swapping them fails quietly rather than erroring: gh issue view <pr-number> returns a real issue, just the wrong one. Braces mark a field inside a format string, angle brackets mark a command argument, and <stack-number> is a third value again — the stack's own id, not a PR's. {owner} and {repo} are the one pair that only looks like the first kind: type them verbatim and gh substitutes them itself — in endpoints and -F values only, never inside -f strings, per the placeholder note in workflows/discuss.md.

• Branch format {type}/GHI-{issue-number}_{slug}, one issue per branch. Defined under Quick reference in the tracker standards, which owns it because the format encodes an issue key, alongside the commit header and the PR title; the {type} vocabulary itself is one section further on, under Branch and commit type. Nothing here chooses a type; this skill only ever parses the number out of a branch that already exists, and the parse is stated here for this skill: drop everything up to and including the first /, take everything before the first _, strip the GHI- prefix - feat/GHI-50_login-form gives 50. Follow the repository's existing names where they differ.

• A branch's PR is opened as a draft at the start of the work, not the end, carrying the plan file. workflows/open.md has the sequence; gh pr ready is what later admits it to the review loop.

• HARD RULE: never commit or push directly to main. gh stack merge lands PRs; pushing main yourself stays forbidden even when it would be faster.

• New dependent work is cut from the parent branch's tip, then gh stack add when tracked, or gh pr create --base <parent> when not.

• The remote's name is a per-repo fact, never assumed. origin is only git clone's default, and a repository whose one remote is named upstream fails every hardcoded origin command outright. Where a workflow writes <remote>, resolve it by the recipe stated here once: git remote printing exactly one name means that name; with several, use the one main tracks (git config branch.main.remote), and ask the owner if that is unset. Pushing outside the tool, the first push of a branch is git push -u <remote> <branch>. The trunk's name is the same kind of fact: these files write main because every repository this skill serves uses it, and where a repository's default branch is named otherwise, that name is a per-repo fact for .agents/gh-solo.md and every main in these files reads as it.

• After any push to a branch with an open PR, read gh pr checks <pr-number> before reporting the push as done. The ticked boxes are local evidence and CI is a different environment; a locally green gate and a red check are not a contradiction to resolve by preferring one - the disagreement is itself the finding. Report both and diagnose the difference; never re-run locally until it looks fine.

• Every gate workflow ends on a verdict line, and every refusal opens with one. The gate workflows are ready, both passes of review, and merge. Their Confirm step's first line is ✅ ALL PASS when nothing failed, ⚠️ PASSED WITH FINDINGS - {what} when the workflow completed but left something for the owner to read, and a stop anywhere prints ⛔ REFUSED - {reason} as its first line. It is the line the owner reads first; the detail above it is the record, not the message.

• A command printed for the owner to run sits alone on its own line, starting at column one. Never indented, never sharing a line with prose, never prefixed - the owner pastes the whole line into a shell, and a single leading space or trailing word breaks the paste. When a verdict line names a next command, the command goes on the line below it, flush left.

• Every PR gets --assignee @me at creation and Closes #{issue-number} in its body. GitHub sets neither for you. The first feeds the in-progress view described in tracker; the second closes the issue on merge and records the link permanently.

• Whoever runs a gate ticks its box; nobody else ever does. The PR body's ## Steps and ## Verification checkboxes are ticked by the agent that did the work or watched the command exit, at the moment it happened. A tick applied by anyone else is indistinguishable from evidence and is not evidence. workflows/ready.md audits that record and deliberately cannot write to it. The same rule reaches the issue's acceptance criteria: the implementing agent ticks each as it verifiably lands, per Writing good acceptance criteria in the tracker standards, and workflows/merge.md audits those at the door too.

• Editing a PR body is read, modify, write. gh pr edit --body replaces the whole body, so fetch it first (gh pr view <pr-number> --json body --jq .body), edit that file, and pass it back with --body-file — never --body "…", which mangles multi-line markdown. Keep the scratch file outside the working tree - the harness scratchpad is the place - so a copy of the body cannot get committed.

• Never invent a repository's check commands. They belong to the repo, named in the plan's ## Verification and otherwise in its per-repo agent config or its CONTRIBUTING. Whether a branch's gates all actually ran is audited by workflows/ready.md.

• The diff analysis is done by the reviewer agent this plugin ships, spawned by workflows/review.md with a PR number and nothing else. It is a separate agent so that the session which wrote the code is not the session that judges it, and that is the only reason it is a subagent rather than a skill run inline. It needs a harness that can spawn one, and a repository may appoint a different reviewer in its own config, per the repo-conventions bullet below - either another agent, or a capability invoked rather than spawned. What no appointment changes is who writes to the PR: every form hands its findings back to the orchestrator, which posts them through one script under one convention. A capability invoked here is therefore run without whatever flag makes it post for itself. What is never substitutable is that the analysis comes from something other than the author, per Who this is for above.

• The AI disclaimer opens everything posted under the owner's name - PR bodies, comments, Reviews, thread replies. Where the user's global instructions file defines a disclaimer line, that line owns the wording and the rules; absent one, this plugin's default is > 🤖 Written by AI --- read/modified by human! 🤓. Whatever the wording, it opens with > 🤖. That prefix is the whole of what the gates test - the watch filter's --jq in workflows/discuss.md tests it literally - so the tail after it is the owner's to change and the prefix is not.

• A via line follows the disclaimer on every comment, Review and thread reply, naming the process that posted it. Every post is made with the owner's credentials, so the disclaimer alone says only that an agent wrote it - an implementer's record, a divergence note and a review finding on the same PR are otherwise indistinguishable, to the owner included. It is the second paragraph of the disclaimer's own blockquote, so every startswith test on the disclaimer - the watch filter in workflows/discuss.md, the gates in workflows/merge.md, workflows/review.md's recognition of its own records - keeps working untouched:

  > 🤖 Written by AI --- read/modified by human! 🤓
  >
  > via `<skill>` <workflow>, <which post>
  

  So: via implement implement, the implementation record; via pr-flow review, convention check; via pr-flow resolve, the authorisation; via pr-flow discuss, thread reply. The one exception is the PR body, which carries the disclaimer alone: it is unmistakably itself, and per workflows/merge.md it lands in the squash commit on main, where a workflow tag would be noise.

• Repository-specific conventions live in the repository, in its agent config file (.agents/gh-solo.md, or .claude/gh-solo.md where that is what the repo uses). Read it when present and let it override the defaults here. Nothing repo-specific belongs in this skill. Which reviewer a round uses is one of those facts: a Reviewer agent: line there names an agent type to spawn instead of the bundled one, and a Reviewer command: line names a capability to invoke instead, whose findings come back to the orchestrator and go up through the same posting script. workflows/review.md Step 1 owns each form of appointment and what it must produce, and why an unregistered agent is a refusal rather than a fallback. What the reviewer costs is a per-repo fact too: a Reviewer model: line names the model the round asks the spawn for. That is not an appointment and produces nothing, so the same Step 1 owns it separately - what it is validated against, and why it does not reach the invoked-command form.

Preflight, once per session

Run gh auth status. If gh is missing, or it runs but exits non-zero, stop and say which of the two it was: the fixes differ, gh auth login for the second and an install for the first. Every later step depends on it.

Then gh repo view --json nameWithOwner,viewerPermission. Every command here infers the repository from the working directory, and this skill's writes include merges, repository settings and branch protection — the wrong directory aims all of them at the wrong repository. Confirm the resolved name is the intended one before any write, and say so rather than proceeding when it is not.

Issue context comes from the same gh, so a headless run under claude --print or cron gets the same issue context an interactive one does.

---

Routing

| The owner says | Workflow | |---|---| | "stack a branch", "add to the stack", "submit", "sync", "restack", "view the stack", "what is this stacked on" | workflows/stack.md | | "review open PRs", "review PR 60", "what needs review", "review unreviewed PRs" | workflows/review.md | | "write the plan", "open the PR", "start on this branch" | workflows/open.md | | "mark it ready", "it's done", "ready for review", "take it out of draft" | workflows/ready.md | | auto 50 - the literal command, with an issue number | workflows/auto.md from the top: plan and draft PR, the implementation, ready, the review round, stop at the owner with the watch armed | | go 60 - the literal command, with a PR number | workflows/auto.md at its go entrance: the same chain minus open, for a plan the owner has already read | | "resolve all and push" | workflows/resolve.md, the protocol's step 7 | | "merge it", "land this", "merge the PR", "merge the stack" | workflows/merge.md - but said while a review round has left fix commits unpushed, it means workflows/resolve.md first, which records the authorisation, resolves and pushes | | "I replied on the PR", "answer my comments", "I asked something on the review" | workflows/discuss.md | | "resolve all and push", "we are done", "you can merge" - said after a review round left fix commits unpushed | workflows/resolve.md, the protocol's step 7: stop the watch, post the authorisation, resolve what it covers, push, read the checks, then workflows/merge.md | | "stop watching", "you can stop polling now", "unwatch" | stop the monitor; see Stopping it in workflows/discuss.md |

Based on the argument above, do exactly one of the following:

• If the argument is exactly help → read workflows/help.md, output its contents, stop.
• If it starts with review → read workflows/review.md and follow it.
• If it starts with open or plan → read workflows/open.md and follow it.
• If it starts with auto followed by a number → read workflows/auto.md and follow it from the top. Only the literal command enters it - the chain skips the plan-reading stop, so no sentence routes here, however clearly it implies one.
• If it starts with go followed by a number → read workflows/auto.md at its go entrance. The same arming rule: literal command only.
• If it starts with ready → read workflows/ready.md and follow it. If the argument also contains review (ready review, ready review 60) → when ready ends green with the draft lifted, continue straight into workflows/review.md on the same PR, as if the owner had named it. A refusal in ready stops the chain, and the chain changes nothing inside either workflow - it only removes the wait between them.
• If it starts with discuss, reply or chat → read workflows/discuss.md and follow it. sync never means this, however the request is phrased: that word belongs to the cascade rebase in workflows/stack.md, and confusing the two rewrites history when someone asked for a conversation.
• If it starts with watch → read workflows/discuss.md and arm its watch. Only this literal command arms it — no sentence does, however clearly it implies one.
• If it starts with unwatch, or the owner says to stop watching in any words → stop the monitor with TaskStop and confirm it is gone. Arming needs the exact command; stopping deliberately does not, because a stop misread costs nothing and a stop missed leaves something running.
• If it starts with resolve, or the owner authorises the resolve and the push in any words - "resolve all and push", "we are done", "you can merge" - while a review round has left fix commits unpushed → read workflows/resolve.md and follow it, then continue into workflows/merge.md. Unlike auto, go and watch, this one is not literal-command-only: their sentence authorises it just as the typed word does, because what it needs is their decision rather than a precise instruction, and refusing to act on "we are done" would only make them type it twice.
• If it starts with merge → read workflows/merge.md and follow it, including when the PR is stacked. It is the single entry point for landing anything, and it routes to gh stack merge itself. Sending a stacked merge to workflows/stack.md instead skips the review gate.
• If it starts with view, init, add, submit, sync or restack → read workflows/stack.md and follow it.
• If the request is about a branch's relationship to another branch → read workflows/stack.md.
• If the request is to implement the plan, continue the implementation, or land review fixes → wrong skill; that work belongs to implement. Everything around it - open before, ready, review and merge after - stays here. The one path from this skill into that work is workflows/auto.md, whose chains enter the implement skill by name so it runs under its own tool grant.
• If it is about an issue rather than a branch or a PR → this is the wrong skill; use tracker.
• Otherwise → read workflows/help.md, output its contents, and say which argument failed to match anything above.

Supporting files

• workflows/open.md - plan file, first commit and draft PR, at the start of a branch's life
• workflows/auto.md - the auto and go chains: the lifecycle's workflows run back to back with the waits removed and every gate intact, ending at the protocol's step 6 with the watch armed. Literal commands only, because each skips a stop that is the owner's
• workflows/ready.md - the other end: auditing that every stated gate ran and that CI agrees with the record, then taking the PR out of draft. It runs nothing and writes nothing but the flag
• scripts/docs-check.py - verifies every backticked path resolves and every code fence closes; run it before pushing a docs or plan change, since no hook will. --ignore <glob>, repeatable, skips path spans that belong to another tree
• workflows/stack.md - the gh stack prerequisites, when a stack is wanted at all, the layers of stack state, the worktree trap, and the drift playbook. Everything stack-specific lives there rather than here, because it has exactly one reader
• workflows/review.md - the review round: scoping and gating a PR, the tracker checks, spawning the reviewer subagent, posting what it found, landing the fixes locally and stopping at the owner
• scripts/post-review.py - builds the one API call that lands a round's threads and its record Review together, refusing the whole round on any invalid finding, and afterwards reconciles what the PR carries against what was sent; scripts/test-post-review.sh is its bench
• workflows/merge.md - the end of a branch's life: the reviewed-or-not gate, the last checklist audit, the squash merge, branch cleanup, and confirming the issue actually closed
• workflows/discuss.md - answering the owner's replies on inline comment threads, in the thread; also the watch poll loop and what ends it
• workflows/resolve.md - the protocol's step 7: the authorisation comment and its marker line, the resolve mutation, and the round's only push
• workflows/help.md - the one file addressed to the owner rather than the agent; output it verbatim on help and add nothing to it
• references/review-protocol.md - the review round protocol, stated once: the cast, RF ids and severities, its steps 1 to 8, the owner's vocabulary, and the conclusions the review, resolve and merge gates enforce