| name | beans-multi-worker |
| description | Use this skill when multiple workers (parallel checkouts, separate machines, or human + agent) share a single beans tracker. Covers session close protocol, when to pull, and resolving git conflicts on bean files. |
Beans Multi-Worker Sync
When This Skill Applies
Use this skill when multiple worker checkouts, separate machines, or a mix of human and agent collaborators share one beans tracker. Beans live as plain markdown in .beans/ under version control, so multi-worker discipline reduces to normal git discipline — but the failure modes and conflict shapes are worth calling out.
The Rule
Before acting on another worker's output, git pull. After producing your own, git push.
Pushes are cheap. Pulls are situational. The asymmetry matters — push proactively, pull at moments where you're about to act on what someone else produced.
Unlike a SQL-graph tracker that needs a separate sync step, beans state moves with the code. One git pull brings both. One git push sends both. No second protocol to remember.
Push after every bean write
Bean state changes touch real files. Push them so other workers can see what's happening.
- After
beans create, beans update, beans delete, beans archive → git add .beans/ && git commit -m "..." && git push.
- The bean change usually piggybacks on the related code commit. Claiming a bean (
status=in-progress) is fine as a standalone commit. Completing a bean ships with the code change.
Pull at handoff points
Pulls are situational — only at moments where you're about to act on what someone else produced. Do not pull on every beans show/beans list — that's wasteful.
Verification handoff
Running /verify or otherwise reviewing a bean tagged unverified:
git pull --rebase
…before reading source or bean state. One pull covers both.
Verification is the canonical worker→reviewer handoff. Skipping the pull risks reviewing stale code or overwriting concurrent reviewer notes.
Resuming after external change
If you've been told "the bean was reopened," "the verifier left notes," "the user closed a dependency," or any signal that another actor touched your bean since you last looked:
git pull --rebase
…before beans show <id>. Your in-memory model of the bean is stale.
Session start
git pull --rebase
You don't know what landed while you were away.
Sequence pull → read → write — never parallelize
git pull --rebase, beans show, beans list, and bean writes (beans update, beans create) are state-dependent. They MUST run sequentially.
The trap: agent harnesses that support parallel tool calls make it tempting to batch git pull --rebase and beans show <id> in one parallel block. That is unsafe. The bean read can complete against the pre-pull state, giving you a stale view that contradicts the post-pull truth. Symptom: "first check says not ready, second check says ready, with no apparent change between them."
Rule:
- Run
git pull --rebase alone. Wait for it to finish.
- THEN run
beans show / beans list / any verification reads.
- THEN run bean writes, commits, pushes — one at a time.
Do NOT issue these in the same parallel batch, even when the harness allows it. The performance saving is small; the cost of acting on stale bean state is real.
Session Close Protocol
When ending a work session, you MUST complete ALL steps below. Work is NOT complete until git push succeeds.
[ ] 1. git status (check what changed)
[ ] 2. git add <files> (stage code AND bean changes)
[ ] 3. git commit -m "..." (one commit covers both)
[ ] 4. git pull --rebase (rebase on remote)
[ ] 5. git push (push to remote)
[ ] 6. git status (must show "up to date with origin")
Critical rules
- Work is NOT complete until
git push succeeds.
- NEVER stop before pushing — that leaves work stranded locally.
- NEVER say "ready to push when you are" — YOU must push.
- If push fails, resolve and retry until it succeeds.
- File new beans for remaining work before pushing — capture follow-ups in beans, not in your head.
Conflict resolution
When two workers update the same bean concurrently, git pull --rebase may report a merge conflict on .beans/<id>--*.md. Resolve like any other text-file conflict:
git pull --rebase
git add .beans/isaac-abc--something.md
git rebase --continue
git push
Common shapes:
- Status race — two workers claim or close the same bean. Pick the right end state (usually the later one wins; if the work was actually done by one, their state should stick).
- Tag drift — one worker added a tag, another removed it. Decide which is correct based on intent (the verifier removing
unverified should stick; a claim that bypassed it should not).
- Body drift — both workers appended verification notes. Concatenate both.
After resolving, commit the merge and push. Do not force-push.
What's gone vs the beads era
- No separate sync command.
bd dolt push/pull is gone. git push/pull does everything.
- No silent stranding via failed sync. If your
git push succeeds, your bean state is visible. If it fails (non-fast-forward), you know immediately and pull/resolve.
- No
--theirs / --ours SQL conflict resolution. Conflicts are text-file conflicts.
- No Dolt credentials, no remote service to be down, no out-of-band push step. One credential surface: git.
When this skill does NOT apply
Single-worker projects with one machine and one human. The discipline is the same shape as ordinary git hygiene; no extra protocol needed.
