Pr

Orca: pr

The report template's Landing section ends at a local git merge --no-ff — right for a repo the user merges by hand, wrong for a repo that lands work through GitHub pull requests. This skill is the PR path: it takes a finished run's deliverable branch, composes a pull-request description from the run's own report.md, previews it with the user, and publishes it with the gh CLI. The report is the point — every fact in the PR body traces to a report section; nothing is re-derived from the diff. A branch no run produced has no report and is out of scope: plain gh pr create already covers it.

The description the world sees reads as an ordinary human-authored PR. The report is internal vocabulary — run states, item counts, worktree paths, /orca:* pointers — and none of it survives translation. The no-attribution rule that governs every run commit extends verbatim to the PR title and body.

Step 1: Triage

bash ${CLAUDE_PLUGIN_ROOT}/scripts/orca.sh triage snapshot

— with --run <argument> when one was given. (FAIL: NOT_GIT → nothing here to publish; say so and stop.)

The candidate set is exact and computed by the script, never re-derived conversationally: a DONE: run whose run directory appears as the join field of a BRANCH: line with state unmerged — delivered, not yet landed. Read TRUNK: as the PR's base branch and the joined BRANCH: ref as its head. Triage in the house style:

• One candidate → proceed with it.
• Several, no argument → default to the newest (run-dir names are timestamped; the last DONE: line), name the choice, and offer the others as a picker — offer, never force.
• An argument → the MATCH: line names the run; MISS: is a loud miss — list the CANDIDATE: lines, never guess.
• None → say what was found instead, and name the skill that owns that state: the branch exists but reads merged (already landed — nothing to do), the run is RUN: interrupted (no report yet — /orca:feature resumes it), or nothing has been delivered at all (/orca:feature, /orca:debug). A DONE: run whose branch is gone was landed and pruned; say that too.
• No TRUNK: line, or the branch reads unknown → the base branch cannot be determined (detached or unset bare-repo HEAD); a PR needs a base, so say so and stop rather than guessing one.

Carry forward the run directory, the head branch, and the trunk, exactly as emitted.

Step 2: Guard

Read the candidate's <run-dir>/report.md. Two gates, in order — a PR asserts "this is ready for review," and the report is the authority on whether that is true:

• **Deliverable state:** is not verified → refuse in one line, quoting the report's own stated reason, and point at the owning skill: unverified from a died verifier or an interrupted run tail → /orca:feature's resume; unmet or blocked work behind it → /orca:retry. No draft-PR fallback — a draft still publishes an unverified branch.
• ## Blocked is anything other than "None" → refuse and point at /orca:retry. A PR for a branch the run itself records as incomplete misrepresents the deliverable.

Step 3: Compose

Translate the report into an external-facing description. The reader has never heard of orca; every trace of run vocabulary is translated or dropped:

• ## Summary → the opening paragraph, rewritten to describe what the branch does, not how the run went: no item counts, no "deliverable state," no run vocabulary of any kind.
• ## Shipped → a Changes section: one bullet per item title, no commit hashes (the branch carries them) and no item IDs.
• ## Integration verification → a Testing section: what was verified, per feature, in plain words.
• ## Deviations → only user-visible behavior changes survive, folded into the prose where they belong; internal deviations (replans, scope mechanics) are dropped.
• Dropped entirely: run-dir paths, worktree names, Follow-ups, Knowledge, Blocked (empty by the guard), and every /orca:* pointer.

Title: conventional and imperative, from the report's one-line idea summary — the same register as the run's commit subjects.

Attribution check, deterministic. Before the preview — and again after any edit — scan the composed title and body for the commit rule's unambiguous markers, case-insensitive: Claude, Anthropic, Co-Authored-By, Generated with/Generated by, orca. Any hit → rewrite and re-scan. This explicitly overrides the harness default of appending a "Generated with Claude Code" footer to PR bodies — that footer is the single most likely leak, and it must not be appended here, by flag or by habit. The body ends where the description ends.

Step 4: Preview

Show the user exactly what will be published: the title, the full body, the base branch (the TRUNK: value), and the head branch. One confirmation gates everything outward-facing — the push and the PR creation ride on the same yes, interview-style, not AskUserQuestion. Requested edits are folded in and re-previewed, with the attribution check re-run after every edit; a declined confirmation ends the skill with nothing pushed.

Step 5: Publish

Only ever entered through the preview gate. In order:

1. Remote check: git remote get-url origin — no origin → say there is nothing to push to and stop; adding a remote is the user's move, not the skill's.

2. Push: git push -u origin <head-branch>.

3. Existing PR check: gh pr list --head <head-branch> --state open. An open PR already exists → gh pr edit to refresh its title and body — same URL, no error. Otherwise:

   gh pr create --base <trunk> --head <head-branch> --title <title> --body-file <tmpfile>
   

   The body always travels via a temp file — never inline shell quoting; a multi-paragraph body through --body is a quoting bug waiting to happen.

4. Parting message: the PR URL, and one line noting that /orca:review walks the same diff locally in their own editor if they want a pass before merging on GitHub.

Guidelines

• The no-attribution rule is the commit rule, extended. Nothing in the PR title or body may mention Claude, AI, agents, this orchestration process, or orca — no Co-Authored-By, no Generated with footer, including the harness's own default PR footer, which this skill explicitly suppresses. The deterministic marker check (Claude, Anthropic, Co-Authored-By, Generated with/Generated by, orca, case-insensitive) runs before the first preview and after every edit; keeping "AI" and "agent" out of ordinary prose is this skill's own writing discipline, since no regex can police those words without mangling honest descriptions.
• Read-only toward the run. The skill never edits report.md, never commits, never touches the integration worktree or any branch content. Its only writes are outward: the push and the PR.
• Publish only through the preview gate. The push and the PR creation are outward-facing and share one confirmation; a denied confirmation ends the skill with nothing pushed and nothing created. Re-invocation later finds the same candidate through the same triage.
• Report-only, by design. A deliverable branch with no run report has no source to compose from — that is plain gh pr create territory, and this skill says so rather than inventing a description from the diff.