Agent Skills: Lambda Workflow

Lambda Workflow

Use this skill whenever you touch delivery end-to-end - from grabbing a br issue through merge/closure. Each phase builds on the last; do not skip ahead unless a human explicitly says so.

Lifecycle Map

1. Select & Claim Work - pick an unblocked br issue, understand scope, and claim it
2. Kick Off & Draft – baseline tests, branch, and open a draft PR immediately so CI starts.
3. Build & Validate – implement with tests-first habits, keeping trackers and notes fresh.
4. Land the Plane – ready-for-review PR with full test/QA + repo hygiene.
5. Monitor & Respond – use the Dumbwaiter MCP to wait on GitHub signals (checks, reviews, comments, merge) and react.
6. Close the Loop - sync br/git, close the issue, and record proof the change stuck.

1. Select & Claim the br Issue

• If .beads/beads.db is missing (fresh clone, new worktree, etc.), run br sync --import-only --db .beads/beads.db from the repo root so the local database hydrates from .beads/issues.jsonl before listing work.
• If the beads-sync worktree is missing, recreate it with a sparse checkout of only /.beads/:

  git fetch origin beads-sync:beads-sync || git branch beads-sync
  git worktree add --no-checkout .git/beads-worktrees/beads-sync beads-sync
  git -C .git/beads-worktrees/beads-sync sparse-checkout init --cone
  git -C .git/beads-worktrees/beads-sync sparse-checkout set /.beads
  git -C .git/beads-worktrees/beads-sync checkout beads-sync
  

• You may be handed an issue: if so, choose that one. Otherwise, run br ready --json --limit 0 before asking for work; respect blockers/dependencies, pick the first ready issue and claim it. br update <id> --status in_progress --notes "Starting work on ${issue title}".
• Default selection mechanism: bv -robot-next (or bv -robot-triage for full context). Map the selected id to br show <id> and br update <id> --status in_progress. Ignore embedded bd claim/show commands in the bv output.
• Export and commit the .beads/issues.jsonl changes immediately on main and push. This avoids multiple agents pulling the same issue. If you get a conflict, revert the .beads/issues.jsonl changes, pull with rebase, and try again.
• Read the issue (and linked docs) end-to-end. Confirm acceptance criteria, implicit contracts, and dependent tasks. It is possible it is in a partially complete state: if so, pick up where it was left off.
• Clarify gaps before coding. Update the br issue with questions or new discoveries so history lives in .beads/issues.jsonl. Again, push these changes (and only these changes) directly to main.

2. Kick Off & Draft the PR

1. Baseline the repo
   • Sync with git fetch --all and git pull --rebase origin main.
   • Run the project’s test suite. If it fails on main, stop and escalate instead of piling on.
2. Branch + tooling
   • Create a fresh branch (git checkout -b <short-task-name>). Never work directly on main.
   • Verify gh auth status (or equivalent) so PR automation works later.
3. Immediate draft PR
   • Push the branch and open a draft PR sourced from the br issue summary/acceptance criteria, including any questions or underspecified areas.
   • Preferred helper: gh pr create --draft --title "..." --body-file body.md. Be wary of quoting issues, it's easy to end up with "\n" characters in the PR.
   • Capture acceptance criteria + planned tests in the PR body so reviewers know how you’ll prove success.
4. Plan validation
   • Decide which automated + manual tests will prove the work. Note the plan in br or the PR so it is reviewable before implementation.

3. Build & Validate Continuously

• Keep git status -sb clean; commit coherent checkpoints and push often so the draft PR reflects reality. Pull and rebase over origin/main before pushing each time.
• Treat failing feedback as a test design task:
  1. Write or extend a failing test that reproduces the bug/regression (default to proptest! or integration coverage when stateful sh loops).
  2. Run the suite, commit/push the red test alone.
  3. Fix the bug in a follow-up commit, push, and reply to the review thread with the fixing hash/summary.
• Strip debug aids (dbg!, println!, temporary flags) before moving to landing.
• For transactional upgrade/apply loops, add coverage that proves: atomic rollback on failure, lockfile sync after success, cross-device safety (rename EXDEV), and symlink preservation.
• You will commonly discover unexpected tasks while building; if they are directly necessary for the success of this task, add a br subtask blocking this task and work on it first, on the same PR. If they are not, add them as discovered-from tasks in br and continue; they can be done in the next round.

4. Land the Plane (Ready for Review)

1. Quality gates

   • Run the full suite (tests, linters, formatters) on the feature branch to a green state after your last rebase.
   • Add or expand tests until every change path is covered.

2. Repo hygiene

   • Remove throwaway files, logs, and manual scripts. Ensure .gitignore keeps artifacts out.
   • Rebase on the latest main, squash/reorder into meaningful commits, and confirm no stray stashes remain (git stash list). Make sure you only squash/reorder your commits! Commits must all be rebased on top of origin/main.

3. PR

   • Flip the draft PR to Ready using gh pr ready once all our tests pass, and we think the acceptance criteria are met. Close the br issue now, update br notes with the PR URL, testing evidence, and any deviations from the original plan, export with br sync --flush-only, and push.
   • Update the PR body with: summary, testing notes, linked br issue, and known follow-ups.
   • Trigger automated review with a comment containing exactly @codex review; reply inline to every Codex comment with the commit hash that fixes it.

4. Tracking

5. Monitor & Respond with Dumbwaiter MCP

Once the PR is Ready, hand monitoring to the Dumbwaiter MCP so you don’t poll GitHub manually.

1. Start a wait (via MCP tools or mcp__dumbwaiter__wait.start):

   {
     "provider": "github",
     "selector": { "owner": "ORG", "repo": "REPO", "pr": 123 },
     "condition": "checks_succeeded"
   }
   

   Capture the returned wait_id.

2. Await completion

   • Call wait.await with that wait_id to stream progress notifications (check statuses, workflow runs, etc.).
   • Use other conditions as needed: pr_merged, checks_failed, comment_received (plus filters), changes_requested, or workflow_completed.
   • If you only need polling, call wait.status on an interval; cancel via wait.cancel if superseded.

3. React to outcomes

   • On green checks → post the success + wait_id back to br/PR notes.
   • On failures or change requests → surface the failing context and return to step 3.
   • For comment streams, enable condition: "comment_received" with filters.since so every new review/comment notifies you in real time.
   • On merge, we're done, report success to user.
   • On approval: if we got approval and we are in ready-for-review and all the checks are passing and we are rebased on top of origin/main, we are also done, but in this case we should report back to the user that the PR can now be merged, along with the URL of the PR for easy access. If everything but the rebasing is done, we can rebase and try again, again commenting "@codex review" after the commit lands and the tests pass in CI.

4. Background durability

   • Set DUMBWAITER_DB and DUMBWAITER_WATCHER=1 if you need waits to survive process restarts. Always log the wait_id in br so another agent can resume with wait.status.

6. Close the Loop

• After Dumbwaiter reports pr_merged, archive local branches/stashes so the next effort starts clean.

Following this workflow keeps the entire Lambda lifecycle observable: br reflects intent, GitHub shows work-in-progress via draft PRs, Ready PRs meet the landing checklist, and Dumbwaiter MCP watches the PR until it merges.