Submit GitHub PR

Boundaries

• Active repository: NVIDIA/TensorRT-Model-Connect.
• Use the local github remote and target GitHub main.
• Never push directly to main.
• Publish only the change the user placed in scope.
• Open a draft PR by default; omit --draft only when the user explicitly asks for a ready-for-review PR and the evidence supports that state.
• Do not add labels, reviewers, milestones, projects, or auto-merge unless the user explicitly requests them. The authorized Internal CI trigger is the exception described below.
• Do not merge from this skill. Use $pr-babysitter for monitoring and merge.
• Use $write-git-messages for commit, PR, squash, and rebase text.

1. Authenticate And Verify The Remote

gh auth status
gh repo view NVIDIA/TensorRT-Model-Connect \
  --json nameWithOwner,url,defaultBranchRef
git remote get-url github
git fetch github main

The remote URL must resolve to https://github.com/NVIDIA/TensorRT-Model-Connect.git or its authorized SSH equivalent. If gh cannot read the private repository, stop for safe reauthentication. Do not inspect or replay stored credentials.

2. Confirm Branch And Exact Scope

BRANCH=$(git branch --show-current)
test -n "$BRANCH"
test "$BRANCH" != main
BASE_SHA=$(git rev-parse github/main)
HEAD_SHA=$(git rev-parse HEAD)
git status --short --branch
git merge-base --is-ancestor "$BASE_SHA" "$HEAD_SHA"
git diff --stat "$BASE_SHA"...HEAD
git diff --name-status "$BASE_SHA"...HEAD
git log --oneline "$BASE_SHA"..HEAD

Review focused diffs, staged changes, and untracked files. Do not publish:

• unrelated user changes;
• generated artifacts or secrets not intended for review;
• an empty branch;
• a branch based on the wrong repository or target;
• commit messages containing Claude.

If the branch is behind or diverged from current github/main, rebase only when that history rewrite is within the user's request. Revalidate after any rebase.

3. Validate The Publishable Head

Run git diff --check plus checks appropriate to the changed surface. Record the exact command and pass/fail/skip/not-run result. Do not present:

• a dry run as execution;
• compilation as model parity;
• static docs checks as target-hardware proof;
• a check on a prior SHA as evidence for the current head.

Confirm the worktree contains no unsaved in-scope change before committing or pushing. If a new commit is needed, use $write-git-messages and a GitHub-safe author email; do not change global Git identity implicitly.

Create or update an architectural record only when an existing repository policy, template, maintainer request, or user instruction requires it. This skill does not invent ADR policy or files merely from the size/type of a diff.

4. Draft Reviewer Text

Give $write-git-messages:

• exact base and head SHA/ref;
• changed files and behavioral purpose;
• exit criteria and non-goals;
• validation commands with outcomes;
• unrun CI, GPU, model, performance, or qualification gates;
• dependency and merge order;
• compatibility, security, migration, rollback, and review-order notes.

The PR title should match the expected squash/rebase title and follow repository convention. Keep the body focused on one purpose.

5. Push The Named Branch

git push -u github "$BRANCH"
PUSHED_SHA=$(git rev-parse "github/$BRANCH")
test "$PUSHED_SHA" = "$(git rev-parse HEAD)"

Do not force-push unless the user authorized the history rewrite and the exact remote branch was resolved first. When rewriting an existing branch, prefer an explicit lease for that branch and its previously observed SHA.

6. Create And Verify The PR

Pass the reviewed body on stdin to avoid leaving an unnecessary local file:

gh pr create \
  --repo NVIDIA/TensorRT-Model-Connect \
  --base main \
  --head "$BRANCH" \
  --draft \
  --title "<type(scope): concise summary>" \
  --body-file -

Verify the immutable facts:

gh pr view <number-or-url> \
  --repo NVIDIA/TensorRT-Model-Connect \
  --json number,url,isDraft,baseRefName,headRefName,headRefOid,title,state

The base must be main, head name must match, and headRefOid must equal PUSHED_SHA. Then inspect initial checks:

gh pr checks <number-or-url> \
  --repo NVIDIA/TensorRT-Model-Connect

Queued, missing, skipped, or failed checks are not green.

7. Start Exact-Head Premerge When Authorized

Creating or pushing the PR does not start premerge. When CI execution is authorized, resolve the current PR API head before adding the one-shot trigger:

REPOSITORY=NVIDIA/TensorRT-Model-Connect
PR_NUMBER=<number>
pull=$(gh api "repos/$REPOSITORY/pulls/$PR_NUMBER")
PR_HEAD_SHA=$(jq -er '.head.sha' <<<"$pull")
test "$PR_HEAD_SHA" = "$PUSHED_SHA"

gh api \
  "repos/$REPOSITORY/commits/$PR_HEAD_SHA/status" \
  --jq '[.statuses[] |
    select(.context == "TRTMC Internal CI / Automated premerge gate")][0]'

If the exact head already has PENDING or PASS for TRTMC Internal CI / Automated premerge gate, do not trigger it again. Otherwise an actor with maintain or admin permission may run:

gh pr edit "$PR_NUMBER" \
  --repo "$REPOSITORY" \
  --add-label run-internal-ci

If authorization fails and retains the label, remove run-internal-ci before adding it again after the reported prerequisite is satisfied. Re-adding an already-present label is a no-op and does not trigger the bridge.

Never use the legacy run-ci label. The bridge consumes the trigger, verifies the exact head, and dispatches private Internal CI. A successful Source bridge run is not a passing premerge result: only TRTMC Internal CI / Automated premerge gate=PASS on the current head is a pass. Raw Internal CI logs, artifacts, packages, runner details, and URLs stay private and must not be copied to the Source PR.

8. Report And Hand Off

Report the PR URL and number, draft state, pushed head SHA, validation evidence, initial exact-head CI state, and any dependency or unrun gate. Hand monitoring and any merge work to $pr-babysitter; this skill never merges.