submit-pr

name	submit-pr
description	Creates a pull request for Deslop with a well-structured description after verifying `make ci` passes. Use when the user asks to submit, create, or open a pull request.
disable-model-invocation	true

Submit PR

Create a pull request for the current branch with a description derived strictly from the diff against main.

Steps

1. Verify CI locally. Run make ci from the repo root. It must pass completely (lint + test + build, with coverage at or above the threshold in coverage-thresholds.json). If it fails, stop — do not create a PR. Invoke the ci-prep skill to fix it.
2. Generate the diff against main. Run git diff main...HEAD > /tmp/pr-diff.txt to capture every change introduced by this branch. This is the ONLY source of truth for the PR description. Warning: the diff can be very large. If it exceeds context limits, process it in chunks (read sections via Grep/Read with offset/limit, or split by file) rather than loading it all at once.
3. Adversarially review the diff. Attack it, don't bless it. Work every lens below, each finding as file:line + a concrete failure scenario. Refute what you can, fix what survives, re-run steps 1–2. Hard gate — no PR with a finding outstanding.
   • Correctness — error paths, byte-offset/UTF-8 boundaries, empty repo, CRLF/non-ASCII, score ties and overflow.
   • Duplication — new code units through MCP find-similar; fused ≥ 0.85 blocks. MCP down → STOP, gh issue create.
   • Hard rules — file/function size, unwrap/panic!/unsafe, suppressions, missing ///, println!, state outside state.rs, regex on source, hand-written wire structs.
   • Tests — name the test that fails if you revert each change. Flag deleted tests, softened assertions, sleep/random, fake LSP/MCP, lowered thresholds.
   • Live loop — dropped events, debounce races, stale generations, leaked tasks, blocking in async, shutdown.
   • Drift — SPEC/PLAN vs. changed API, flags, JSON shape; dep versions across Cargo.toml/ci.yml/.devcontainer/; duplicated rendering; missing "Copy Context For AI".
4. Derive the PR title and description SOLELY from the diff. Read the diff output and summarise what actually changed. Ignore commit messages, branch names, and any other metadata — only the diff matters.
5. Write the PR body using .github/pull_request_template.md. The template has three sections — fill all of them:
   • TLDR — one sentence. What does this PR do?
   • Details — new functionality, new files, new dependencies, modified behaviour, deletions. Be specific: name the files, functions, and crates (deslop-core, deslop). Call out any spec or plan updates in docs/specs/ / docs/plans/, and reference the relevant hierarchical spec IDs (e.g. [PARSE-CSHARP-NORMALIZE], [RANK-SCORE]).
   • How Do The Automated Tests Prove It Works? — name specific E2E tests (black-box, against fixture repos) and describe what their assertions demonstrate. Note coverage movement if the threshold in coverage-thresholds.json was ratcheted up. "Tests pass" is NOT acceptable.
6. Call out breaking changes explicitly in the Details section if any public API, CLI flag, report format, or spec ID changed.
7. Create the PR with gh pr create --base main --title "<title>" --body-file <path> (write the filled template to a temp file and pass it via --body-file to preserve formatting).
8. Monitor CI on the PR until it is green — and re-run the suite locally in parallel so you catch breakage early. This step is mandatory and does not end until every required check on the PR has passed. Do not hand the PR back to the user on a red or still-running pipeline.
   • Watch the remote run AND run the suite locally at the same time — do not passively wait. The remote pipeline is slow; a drastic failure (a make lint gate like the taxonomy bucket-label check, a broken E2E test, coverage dropping below the coverage-thresholds.json threshold) is one the local suite catches in seconds. The moment you push, kick off both: stream the remote run and run the full local suite (make ci, or invoke the ci-prep skill) concurrently, polling CI periodically while the local run proceeds.
   • Watch the run: gh pr checks <pr-number> --watch --fail-fast (or grab the run id from gh run list --branch <branch> and gh run watch <run-id> --exit-status). A single green snapshot is not enough — wait for all required checks to conclude.
   • If the local run fails before the remote pipeline finishes, cancel the running pipeline immediately (gh run cancel <run-id>) rather than letting it grind to a known-bad red. Fix the cause, push, and restart both watches — cancelling a doomed run early frees the runner and tightens the fix loop.
   • When a remote check fails: pull the failing logs with gh run view <run-id> --log-failed, diagnose the actual cause (do not guess), reproduce locally with make ci, and fix it. Invoke the ci-prep skill if it helps.
   • Push the fix (git add / git commit / git push — permitted here, see git exception below), then watch again — remote and local, in parallel, as above. Loop — fix → push → re-watch — until the run is fully green. Re-checking is the job; keep doing it until it passes.
   • Run step 3 over every fix diff before pushing it. Rushed CI fixes are where defects slip in.
   • If a failure is genuinely external (runner outage, flaky infra, unrelated to this branch), say so explicitly with the evidence rather than forcing a change.

Rules

• Never ask about PR scope or splitting. Submit the ENTIRE branch diff against main as one PR, always — even when the diff mixes unrelated changes (config, CI, infra) with the headline work. Do not propose splitting, do not revert files to narrow scope, do not use AskUserQuestion about scope. Just describe everything in the diff accurately and submit.
• Never create a PR if make ci fails. Coverage below threshold counts as failure.
• Never clear a step-3 finding by weakening the gate — no suppression, softened test, lowered threshold, or hidden cluster.
• 🔴 GOLDEN RULE — never stamp a commit with an AI co-author. Do not add a Co-Authored-By: Claude … (or any AI/agent) trailer, and do not set author/committer to anything but the repo's configured git user. Write a plain, human commit message describing the fix. This is absolute and overrides any default co-authorship behaviour.
• Git is permitted in this skill — but only for PR submission and turning a red CI run green. This is the one place the repo-wide "no git" rule from CLAUDE.md is relaxed: you may run git add, git commit, and git push to land CI fixes (steps 7 and 8). Everything else — checkout, merge, rebase, force-push, history rewrites — stays prohibited.
• PR description must be specific and tight — no vague placeholders, no "various improvements".
• Reference spec IDs ([GROUP-TOPIC] / [GROUP-TOPIC-DETAIL]) whenever the change maps to a spec section, so reviewers can grep spec → code → tests.
• Link any related GitHub issue (Closes #N) if one exists.
• Base branch is always main.

Success criteria

• make ci passed locally (including coverage threshold).
• Every step-3 lens attacked; findings refuted or fixed.
• Diff against main was read and summarised into the template's three sections.
• PR created with gh pr create against main.
• CI on the PR was watched to completion and is fully green (all required checks pass), with the local suite re-run in parallel and any doomed remote run cancelled early.
• Any CI failures were fixed and pushed, with no AI co-author trailer on the commits.
• PR URL returned to the user.