Symphony land

Land

install

source · Clone the upstream repo

git clone https://github.com/openai/symphony

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/openai/symphony "$T" && mkdir -p ~/.claude/skills && cp -r "$T/.codex/skills/land" ~/.claude/skills/openai-symphony-land && rm -rf "$T"

manifest: .codex/skills/land/SKILL.md

source content

Land

Goals

Ensure the PR is conflict-free with main.
Keep CI green and fix failures when they occur.
Squash-merge the PR once checks pass.
Do not yield to the user until the PR is merged; keep the watcher loop running unless blocked.
No need to delete remote branches after merge; the repo auto-deletes head branches.

Preconditions

```
gh
```
CLI is authenticated.
You are on the PR branch with a clean working tree.

Steps

Locate the PR for the current branch.
Confirm the full gauntlet is green locally before any push.
If the working tree has uncommitted changes, commit with the
```
commit
```
skill and push with the
```
push
```
skill before proceeding.
Check mergeability and conflicts against main.
If conflicts exist, use the
```
pull
```
skill to fetch/merge
```
origin/main
```
and resolve conflicts, then use the
```
push
```
skill to publish the updated branch.
Ensure Codex review comments (if present) are acknowledged and any required fixes are handled before merging.
Watch checks until complete.
If checks fail, pull logs, fix the issue, commit with the
```
commit
```
skill, push with the
```
push
```
skill, and re-run checks.
When all checks are green and review feedback is addressed, squash-merge and delete the branch using the PR title/body for the merge subject/body.
Context guard: Before implementing review feedback, confirm it does not conflict with the user’s stated intent or task context. If it conflicts, respond inline with a justification and ask the user before changing code.
Pushback template: When disagreeing, reply inline with: acknowledge + rationale + offer alternative.
Ambiguity gate: When ambiguity blocks progress, use the clarification flow (assign PR to current GH user, mention them, wait for response). Do not implement until ambiguity is resolved.
- If you are confident you know better than the reviewer, you may proceed without asking the user, but reply inline with your rationale.
Per-comment mode: For each review comment, choose one of: accept, clarify, or push back. Reply inline (or in the issue thread for Codex reviews) stating the mode before changing code.
Reply before change: Always respond with intended action before pushing code changes (inline for review comments, issue thread for Codex reviews).

Commands

# Ensure branch and PR context
branch=$(git branch --show-current)
pr_number=$(gh pr view --json number -q .number)
pr_title=$(gh pr view --json title -q .title)
pr_body=$(gh pr view --json body -q .body)

# Check mergeability and conflicts
mergeable=$(gh pr view --json mergeable -q .mergeable)

if [ "$mergeable" = "CONFLICTING" ]; then
  # Run the `pull` skill to handle fetch + merge + conflict resolution.
  # Then run the `push` skill to publish the updated branch.
fi

# Preferred: use the Async Watch Helper below. The manual loop is a fallback
# when Python cannot run or the helper script is unavailable.
# Wait for review feedback: Codex reviews arrive as issue comments that start
# with "## Codex Review — <persona>". Treat them like reviewer feedback: reply
# with a `[codex]` issue comment acknowledging the findings and whether you're
# addressing or deferring them.
while true; do
  gh api repos/{owner}/{repo}/issues/"$pr_number"/comments \
    --jq '.[] | select(.body | startswith("## Codex Review")) | .id' | rg -q '.' \
    && break
  sleep 10
done

# Watch checks
if ! gh pr checks --watch; then
  gh pr checks
  # Identify failing run and inspect logs
  # gh run list --branch "$branch"
  # gh run view <run-id> --log
  exit 1
fi

# Squash-merge (remote branches auto-delete on merge in this repo)
gh pr merge --squash --subject "$pr_title" --body "$pr_body"

Async Watch Helper

Preferred: use the asyncio watcher to monitor review comments, CI, and head updates in parallel:

python3 .codex/skills/land/land_watch.py

Exit codes:

2: Review comments detected (address feedback)
3: CI checks failed
4: PR head updated (autofix commit detected)

Failure Handling

If checks fail, pull details with
```
gh pr checks
```
and
```
gh run view --log
```
, then fix locally, commit with the
```
commit
```
skill, push with the
```
push
```
skill, and re-run the watch.
Use judgment to identify flaky failures. If a failure is a flake (e.g., a timeout on only one platform), you may proceed without fixing it.
If CI pushes an auto-fix commit (authored by GitHub Actions), it does not trigger a fresh CI run. Detect the updated PR head, pull locally, merge
```
origin/main
```
if needed, add a real author commit, and force-push to retrigger CI, then restart the checks loop.
If all jobs fail with corrupted pnpm lockfile errors on the merge commit, the remediation is to fetch latest
```
origin/main
```
, merge, force-push, and rerun CI.
If mergeability is
```
UNKNOWN
```
, wait and re-check.
Do not merge while review comments (human or Codex review) are outstanding.
Codex review jobs retry on failure and are non-blocking; use the presence of
```
## Codex Review — <persona>
```
issue comments (not job status) as the signal that review feedback is available.
Do not enable auto-merge; this repo has no required checks so auto-merge can skip tests.
If the remote PR branch advanced due to your own prior force-push or merge, avoid redundant merges; re-run the formatter locally if needed and
```
git push --force-with-lease
```
.

Review Handling

Codex reviews now arrive as issue comments posted by GitHub Actions. They start with
```
## Codex Review — <persona>
```
and include the reviewer’s methodology + guardrails used. Treat these as feedback that must be acknowledged before merge.
Human review comments are blocking and must be addressed (responded to and resolved) before requesting a new review or merging.
If multiple reviewers comment in the same thread, respond to each comment (batching is fine) before closing the thread.
Fetch review comments via
```
gh api
```
and reply with a prefixed comment.

Use review comment endpoints (not issue comments) to find inline feedback:

List PR review comments:

gh api repos/{owner}/{repo}/pulls/<pr_number>/comments

PR issue comments (top-level discussion):

gh api repos/{owner}/{repo}/issues/<pr_number>/comments

Reply to a specific review comment:

gh api -X POST /repos/{owner}/{repo}/pulls/<pr_number>/comments \
  -f body='[codex] <response>' -F in_reply_to=<comment_id>

```
in_reply_to
```
must be the numeric review comment id (e.g.,
```
2710521800
```
), not the GraphQL node id (e.g.,
```
PRRC_...
```
), and the endpoint must include the PR number (
```
/pulls/<pr_number>/comments
```
).
If GraphQL review reply mutation is forbidden, use REST.
A 404 on reply typically means the wrong endpoint (missing PR number) or insufficient scope; verify by listing comments first.
All GitHub comments generated by this agent must be prefixed with
```
[codex]
```
.
For Codex review issue comments, reply in the issue thread (not a review thread) with
```
[codex]
```
and state whether you will address the feedback now or defer it (include rationale).
If feedback requires changes:
- For inline review comments (human), reply with intended fixes (
```
[codex] ...
```
  ) as an inline reply to the original review comment using the review comment endpoint and
```
in_reply_to
```
  (do not use issue comments for this).
- Implement fixes, commit, push.
- Reply with the fix details and commit sha (
```
[codex] ...
```
  ) in the same place you acknowledged the feedback (issue comment for Codex reviews, inline reply for review comments).
- The land watcher treats Codex review issue comments as unresolved until a newer
```
[codex]
```
  issue comment is posted acknowledging the findings.
Only request a new Codex review when you need a rerun (e.g., after new commits). Do not request one without changes since the last review.
- Before requesting a new Codex review, re-run the land watcher and ensure there are zero outstanding review comments (all have
```
[codex]
```
  inline replies).
- After pushing new commits, the Codex review workflow will rerun on PR synchronization (or you can re-run the workflow manually). Post a concise root-level summary comment so reviewers have the latest delta:
```
[codex] Changes since last review:
- <short bullets of deltas>
Commits: <sha>, <sha>
Tests: <commands run>
```
- Only request a new review if there is at least one new commit since the previous request.
- Wait for the next Codex review comment before merging.

Scope + PR Metadata

The PR title and description should reflect the full scope of the change, not just the most recent fix.
If review feedback expands scope, decide whether to include it now or defer it. You can accept, defer, or decline feedback. If deferring or declining, call it out in the root-level
```
[codex]
```
update with a brief reason (e.g., out-of-scope, conflicts with intent, unnecessary).
Correctness issues raised in review comments should be addressed. If you plan to defer or decline a correctness concern, validate first and explain why the concern does not apply.
Classify each review comment as one of: correctness, design, style, clarification, scope.
For correctness feedback, provide concrete validation (test, log, or reasoning) before closing it.
When accepting feedback, include a one-line rationale in the root-level update.
When declining feedback, offer a brief alternative or follow-up trigger.
Prefer a single consolidated "review addressed" root-level comment after a batch of fixes instead of many small updates.
For doc feedback, confirm the doc change matches behavior (no doc-only edits to appease review).