Beads Parallel Grinding

Grinding Beads in Parallel Worktrees

The serial grind (the /bg command) does one bead at a time in the current worktree. This variant keeps the same queue discipline but overlaps the implementation of several independent beads, giving each subagent a private git worktree so their edits, builds, and commits cannot collide.

Load and apply the beads-best-practices skill throughout. Its honest-WIP, comment, and human-attention rules remain authoritative in this parallel workflow.

What is parallel and what is not:

• Parallel: implementing beads. Each runs in a subagent, in its own worktree, on its own branch.
• Serial: everything touching shared state — reading the queue, claiming and closing beads, merging branches back, and verifying each merge. The orchestrator (you, the main thread) does all of it.

That split is the whole design. Get it wrong and you get concurrent Dolt writes and simultaneous mutations of the user's working tree.

Use wt for all worktree operations

Worktree creation, merging, and removal go through worktrunk (wt), never through raw git worktree or git merge. A repository can configure lifecycle hooks — installing dependencies, copying env files, running the test suite before a merge — and wt is what runs them. Reaching for plain git skips the repo's own setup and gates, which is exactly the failure this skill exists to avoid.

Both prerequisites are required

This skill deliberately does not restate how wt works. It covers only the orchestration — which beads to run in parallel, who merges, and in what order — and delegates everything about wt itself to upstream. So it needs two things present, and checks both before doing any work:

1. The wt CLI. Verify with wt --version.
2. The upstream worktrunk skill. Load it before the first wt command. It is the authority on wt configuration, hook types and timing, template variables, and troubleshooting — all of which this skill assumes rather than explains.

If either is missing, stop and say which. Do not fall back to git worktree, and do not improvise wt usage from memory. A silent fallback produces worktrees with no dependencies installed and no lifecycle hooks run, and the damage is not obvious until later. Point the user at https://worktrunk.dev (the CLI, and the plugin providing the skill) and let them decide.

Consult the worktrunk skill, rather than guessing, whenever a run needs more than the handful of commands below: reading or changing .config/wt.toml or the user config, understanding why a hook fired or did not, resolving a template-path question, or debugging any unexpected wt behaviour.

Two consequences worth internalising:

• wt decides where worktrees live. Paths come from a configurable template, so never invent one. Get the real path from the JSON output of wt switch (see below) or from wt list.

• Hooks may already do the work you were about to do. wt hook show lists what the repo has configured. If a pre-merge hook already runs the tests, do not run them a second time by hand; if there is none, you run them yourself after each merge.

• How a merge is shaped is not this skill's business. Squashing, rebasing, and whether a merge commit is created are matters of the user's general git and wt preferences, not of grinding beads in parallel. Never pass flags to force a particular shape.

Decide the concurrency limit first

The arguments may already name a maximum ("3 at a time", "max 4 in parallel", "-j2"). If so, use it, and do not ask.

If they do not, ask before starting any work — use AskUserQuestion with 2 / 3 / 4 as the options, noting that the user can type another number. Do not pick a default and proceed; the right number depends on how heavy the project's test suite is and on what else the machine is doing.

The limit is a cap on concurrently running subagents, not a batch size. Whenever a slot frees, refill it from the queue.

Refilling is part of finishing a merge, not a separate step to report and await approval on. After merging and verifying, check the queue and dispatch into every free slot in the same turn. A cap is a ceiling, not a request for permission each time — "do not exceed N" is not "do not launch without being asked", and a slot left idle while beads are unclaimed is wasted wall-clock time.

Hold a slot empty only for a stated reason, and say the reason in a clause: the user asked you to hold, or a running subagent is benchmarking and a second CPU-heavy job would corrupt its measurements. Otherwise fill it.

Never state how many subagents are running from memory. Check first — see Establishing state below.

Determine scope

The rest of the arguments restrict which beads are eligible, exactly as for the serial grind. Translate them into bd ready filter flags:

A bare word with no other context means a label. The scope is authoritative — apply it to every bd ready call for the whole run, never widen it, never fall back to the unfiltered queue. If it is ambiguous or matches nothing, ask rather than guessing; verify an epic ID with bd show <id> before using --parent.

Keep work that needs a person out of the agent queue

Run bd human list --json during preflight and again before the final report.

Every automatic queue read must exclude the label:

bd ready --exclude-label=human

A bead waiting on a person is usually blocked by nothing, so it appears in a plain bd ready like any other issue — bd ready has no knowledge of the label. Miss the flag once and a subagent gets dispatched to decide something that was explicitly reserved for the user. Treat it as an invariant on every call.

When a subagent or the orchestrator finds that progress needs a person's judgement, access, hardware, credentials, or observation, split the work at that point. Their part becomes its own bead; the agent remainder stays in the original:

ask=$(bd create --title="Decide: <the question>" \
                --description="<context, options, what turns on each>" \
                --type=task --json | jq -r .id)
bd label add "$ask" human
bd dep add "$blocked" "$ask"
bd update "$blocked" --status=open
bd human list --json

Never label the in-flight bead and leave it at that. bd human respond closes whatever it answers, so a bead holding both their decision and unfinished implementation gets closed with the implementation undone. One bead, one doer — see beads-best-practices.

Only the orchestrator does this. A subagent that hits such a blocker adds a comment describing what is needed and says so in its report; it changes no labels and no statuses. Preserve its coherent partial work, clean up the worktree safely, and refill the slot with ready work.

Keep the new bead inside the active scope, with the same label or --parent as the bead it came from.

The user answers these with /blockers (beads-blocker-review), which closes each bead and automatically releases whatever depended on it. Since every dispatch re-reads the queue, released work reappears as ordinary ready work with no action from you. Do not answer them yourself, and never call bd human respond or bd human dismiss on the user's behalf.

An answer can land mid-run, so never exclude a bead from dispatch because you remember creating it — the bead's current state is the only authority.

Preflight

Before dispatching anything:

0. Run bd human list --json and record the existing human-attention queue.
1. Confirm both prerequisites: run wt --version, and load the upstream worktrunk skill. Stop if either is unavailable.
2. Note the current branch. It is the base branch: every worktree branches from its tip, and every merge lands on it. Pass it explicitly to wt rather than relying on the default, which is the repository's default branch and may not be where you are.
3. git status — uncommitted work in the base worktree is not overwritten by wt merge (it advances the branch rather than merging into your working tree), but a dirty tree makes it hard to tell your own changes from the merged ones. If it is dirty, say so and ask whether to commit, stash, or proceed.
4. wt hook show — see which lifecycle hooks the repo configures, so you know what will run on create and merge, and what you still have to do yourself.
5. wt list — check for leftover bgp/* worktrees from an interrupted earlier run, and clean them up before starting.

Do not pick worktree paths yourself; wt derives them from its own config.

Selecting a batch

Run bd ready --exclude-label=human with the scope flags, freshly, every time you are about to dispatch — at the start, and again each time a slot frees. Never reuse an earlier listing or a remembered ordering. The human can reprioritise, close, add, or block beads at any moment while the grind runs, and a run that is overlapping work by design spends long stretches between queue reads. A listing taken before the last merge may already be stale.

Consider only the top few highest-priority issues — enough to fill the free slots, not the whole queue. Then filter for parallel safety, and this is the one place worth spending a little thought:

• Prefer beads that plainly touch different files or subsystems.
• Never dispatch two beads that you expect to edit the same file. Run those serially instead, one after the other.
• Never dispatch a bead whose work is mostly beads bookkeeping (creating, restructuring, or relabelling issues) to a subagent — it has no usable beads database. Do that work yourself in the main worktree.
• A bead whose description is vague enough that you cannot predict its blast radius is a poor parallel candidate. Run it serially.

If fewer beads are safely parallelisable than there are free slots, run fewer. Underfilling is always correct; a merge conflict storm is not.

Dispatching a bead

For each selected bead, in the main worktree:

1. bd show <id> — read the full issue. The subagent cannot.

2. bd update <id> --status=in_progress

3. Create the worktree and branch with wt, from the main worktree:

   wt switch --create bgp/<id> --base <base-branch> \
             --no-cd --format=json -y
   

   --no-cd keeps your own shell where it is — you are orchestrating, not moving in. --format=json prints the result as structured output; read the worktree path from it rather than guessing, since the path comes from wt's configured template. -y skips approval prompts, which nothing is present to answer.

   This is the step that runs the repo's pre-start hooks, so the subagent starts in a worktree with dependencies installed and env files in place. Let it finish before dispatching.

   If the bead needs gitignored files that hooks do not provide (local config, caches), copy them over with wt step copy-ignored <worktree-path> before dispatching.

4. Spawn a subagent with an Agent call. Send all dispatches for a batch in a single message so they actually run concurrently.

   If the harness lets you title or label a subagent — Claude Code's Agent tool takes a short description, other harnesses have their own equivalent — prefix it with the bead ID, as in <id>: rename config loader. With several subagents running at once, the label is often the only thing distinguishing them in a progress display, and an untitled or generically-titled batch makes it impossible to tell which bead is which when one stalls or fails. It also matches the bgp/<id> branch and worktree names, so a label, a branch, and a bead line up on sight. If the harness offers no such field, skip this — it is presentation, not correctness.

The subagent's prompt must be self-contained, because it cannot see this conversation. Include:

• The absolute path of its worktree, and an instruction to do all work there and nowhere else. It must not cd into the main repository.

• The bead's ID, title, description, design notes, and acceptance criteria — pasted in full, not referenced.

• Bead bookkeeping rules — include these verbatim in every dispatch, and do not drop them when a brief gets long.

  A subagent must keep its bead current as it works. bd comment is mandatory, not optional: findings, evidence, measurements, refuted hypotheses, dead ends, and anything the next person would otherwise have to rediscover — recorded as they are found, not saved for the end.

  This is the only durable record. A final report is lost entirely if the subagent is stopped, crashes, or loses its connection mid-run, and that is not hypothetical: in one run three subagents were interrupted, and their discoveries survived only because someone happened to notice and transcribe them by hand. A comment written at the moment of discovery survives anything.

  A subagent must never change a bead's status — above all, never bd close. Closing marks the work done before it is merged, from a worktree whose branch may still be rejected at the gate, and a closed bead with unmerged work is worse than an open one because nothing will bring you back to it. Closure means integrated into the trunk and verified there, which only the orchestrator can know. Status transitions and closure are the orchestrator's alone.

  Nor should a subagent create beads for work it discovers. Have it report those in its output or as a comment on its own bead; the orchestrator files them, so scope labels and parents stay consistent and duplicates are caught against the queue the subagent cannot see.

  If only a person can unblock the bead, the subagent must immediately add a comment describing exactly what they must do and call it out in its report. It must not change labels or status; the orchestrator splits the work into the person's own bead after preserving the branch.

  Some setups back the database with a single server process (a lockfile, PID or port under .beads/). That is a reason to expect occasional contention on a write, not a reason to stop subagents commenting — have them retry. Never respond to it by banning bd outright: doing so trades a recoverable retry for the permanent loss of everything an interrupted subagent had learned.

• The project's development expectations: tests, linting, and the repository's commit conventions.

• An instruction to commit its work to its branch and not to push at all, and not to merge, rebase, or otherwise touch other branches.

  Say this as a flat prohibition. Phrasing like "push only your own branch" is meant as don't push to the base but reads as an instruction to push, and subagents will follow it. There is no reason for them to: wt merge integrates from the local branch, so committing is sufficient, and every pushed branch outlives its worktree as remote litter that no teardown removes.

• An instruction not to run wt at all, and not to merge. Worktree lifecycle and integration belong to the orchestrator; a subagent merging would write to the shared base branch concurrently with its siblings.

• A request to report back: what changed, which files, what it ran to verify, and anything it could not finish or discovered along the way.

Gitignored files do not exist in a fresh worktree unless a pre-start hook creates them or you copied them with wt step copy-ignored. If a bead needs something neither provides, run it serially instead.

Merging back

Merge one branch at a time, in whatever order the subagents finish.

Use wt merge, driven at the subagent's worktree from where you are. It rebases the branch onto the base before merging, and that is why it is the right tool for a parallel grind: a conflict stops with the rebase left open in the bead's own worktree, leaving the base branch untouched. The mess stays isolated where it belongs, instead of sitting half-merged in the shared base worktree and blocking every sibling behind it.

Note that wt merge works in the opposite direction to git merge: it merges the worktree's branch into the target. For each finished bead:

1. Read the subagent's report. If it failed or stopped short, do not merge; go to When a subagent fails below.

2. Merge it:

   wt -C <worktree-path> merge <base-branch> -y
   

   This runs the repo's pre-merge hooks (its own quality gate), rebases the branch onto the base, merges it, then removes the worktree and branch. One command covers merge and teardown. Pass no shape flags — squash, rebase and merge-commit behaviour come from the user's wt config, not from this skill.

3. If a pre-merge hook fails, the merge aborts and nothing lands. Fix the problem in the worktree, or treat the bead as failed. Never pass --no-hooks to force it through — the hook is the repo's gate, and this skill is not authority to bypass it.

4. If the rebase conflicts, the branch is left mid-rebase in its worktree and the base is untouched. Resolve it there, git rebase --continue, then re-run the merge. You have the full picture; the subagent does not, and re-dispatching to it will not help. If it is not worth resolving, git rebase --abort and treat the bead as failed.

5. Run the project's tests and linters after each merge unless a pre-merge hook already ran them — no point duplicating the gate. The reason to verify per-merge is to know which merge broke what: a branch that passed alone can still break once combined with a sibling's work.

   A skipped gate is not a passing gate. Some gates disable themselves when a prerequisite is missing — an absent build tree, an uninstalled tool — and report a cheerful skip rather than a failure. Read what the gate actually did, not just its exit status, and treat "did not run" as the same severity as "failed". The same applies to a subagent reporting "all gates pass": check whether any of them skipped.

   Where a subagent's central claim is checkable, check it rather than accepting the report — inject the defect its fix prevents and confirm the gate rejects it, then restore. A fix that cannot be shown to fail without it has not been verified.

6. bd close <id> once merged and green. Close it now, in this step — closing is part of finishing the merge, and the moment attention moves to the next subagent is exactly when a bead gets left in in_progress.

7. Refill the free slot from the queue and dispatch again, in this same turn.

Do not batch the merges. Do not close a bead before its branch is merged and verified — a closed bead with unmerged work is worse than an open one.

Establishing state — never infer it

Every question about what is running, where, and how far along has a command that answers it. Use the command. Inferring from memory or from a remembered path is how a long run drifts away from reality, and the errors compound silently because a wrong answer looks exactly like a right one.

That last row has a trap worth knowing. wt merge rebases the branch onto the base before merging, so every commit gets a new SHA. Asking whether the original commit is an ancestor — git merge-base --is-ancestor <sha> <base> — correctly answers no for work that landed perfectly. Ask whether the content arrived instead: check the file exists on the base and diff it against the branch. A subagent that checks its own SHAs will conclude its work was lost when it was not.

Specifically:

• Never glob a remembered worktree path. wt derives paths from its own config, and a repository can hold worktrees under several parents. A relative glob is worse still: it resolves against your current directory, which may itself have drifted into a worktree, and then silently returns nothing.
• Never count bgp/* branches to count subagents. Branches outlive the agents that made them — a paused bead, a merged-but-unpruned branch, and a live agent all look identical in git branch.
• A query that returns nothing about state you believe exists is evidence the query is wrong, not that the state is gone. Verify the query before acting on an empty result.

When a subagent is paused or interrupted

Distinct from failure, and more dangerous, because the work usually still exists and is easy to destroy by accident.

There is no pause primitive. Stopping a subagent terminates it. If asked to pause, say so and ask whether to stop them or let them finish; do not silently substitute one for the other.

Before stopping anything, have it commit. A subagent's uncommitted work lives only in its worktree's working tree — not in the reflog, not in the index, not in dangling objects, not in any stash. If the worktree is later removed, the work is genuinely gone. So instruct subagents to commit work-in-progress to their own branch early and often, and prefer letting one finish over stopping it mid-edit.

If a subagent has already stopped without committing, commit on its behalf, in its worktree, labelled clearly as unverified:

git -C <worktree-path> add -A
git -C <worktree-path> commit -m "wip(<scope>): preserve interrupted work"

Do this only once the subagent is confirmed finished. Committing under a live subagent moves its branch pointer mid-edit and races it.

A transient error is not death. A notification reporting a stalled stream, a network drop, or an API error describes a broken connection; the subagent may still be running and may resume. Neither silence nor an error justifies dismantling a worktree. Ask it directly if the harness supports messaging, or wait for a real completion notification.

Never re-dispatch a bead into an existing worktree. If a subagent was stopped or interrupted, its worktree and branch usually survive. Sending a second subagent to the same bgp/<id> puts two of them in one working tree, committing over each other — one will sweep the other's uncommitted changes into a commit whose message describes something else entirely, and history stops matching content.

Before re-dispatching, run git worktree list and decide explicitly:

• Resume the same subagent (if the harness can message it) — best, since it keeps its accumulated context.
• Reuse the worktree with a new subagent — only after confirming the old one is finished. Tell the new one exactly what state it will find and that any existing commits are a predecessor's, not its own.
• Start clean — remove the worktree first, having preserved anything uncommitted.

What you must not do is dispatch and hope. If a subagent reports finding changes it did not make, treat that as a coordination failure on your side, not a curiosity: stop dispatching into that worktree until you know who else is in it.

Before concluding any work is lost, look in the worktree. Searching the reflog, index, dangling objects and stashes is a sound search aimed at the wrong place: uncommitted changes live in none of those stores.

When a subagent fails

Do not retry blindly, and do not leave the bead in limbo:

• Reported a blocker only a person can clear: preserve coherent partial work, then split it as above — create the person's bead, label it human, keep it in scope, and make the original depend on it. Return the original to open and verify the new bead appears in bd human list --json.
• Reported an agent-resolvable blocker: create a bead for the blocker (bd create ...), add the dependency (bd dep add ...), and return the original bead to open with bd update <id> --status=open. Give the new bead the same label or --parent so it stays inside the scope.
• Produced partial work worth keeping: merge it if it is coherent and green on its own, then reopen the bead with a note about what remains. Otherwise discard the branch.
• Produced nothing usable: discard the branch, reopen the bead, and either take it serially yourself or move on.

Always remove the worktree afterwards, whatever the outcome. Stale worktrees accumulate and confuse the next run. A successful wt merge already removed it; otherwise discarding abandoned work needs both force flags:

wt remove bgp/<id> --force --force-delete -y

--force covers a dirty worktree, --force-delete an unmerged branch. Both are needed to discard abandoned work, so be sure that is the intent — this throws the subagent's commits away. To keep the branch for inspection, drop --force-delete and use --no-delete-branch.

When priorities shift under a running batch

A parallel grind can have several beads in flight for a long time, so the backlog it was dispatched from may no longer be the backlog the human cares about. Each time you re-read bd ready, compare the top of the fresh queue against what is currently in flight.

Never discard or roll back in-flight work. A subagent's commits are finished effort; losing them to a reshuffled backlog is strictly worse than landing something the human deprioritised. Merge what completes, as normal.

If, and only if, there is a large discrepancy — the top of the fresh queue is markedly higher priority than everything currently in flight, e.g. a new P0/P1 sitting behind a batch of P3s — draw the human's attention to it and carry straight on working:

• Say it once, briefly, in your normal output: which in-flight beads, which higher-priority ones now waiting, and that they can tell you to pause the batch if they would rather you switched.

• Fire an out-of-band notification, so it lands even when nobody is reading the transcript:

  ai-notify "Beads priority discrepancy" \
            "<in-flight beads> running while <higher-priority beads> wait"
  

  ai-notify picks whichever mechanism the machine actually has, makes the notification persistent where it can (the human may be away from the screen), and always exits 0 — so there is nothing to check and no need to probe for herdr, notify-send, or anything else yourself. It never waits for input. If it is not installed, skip the notification; the in-transcript mention is enough.

This must never block. Do not use AskUserQuestion, do not wait for a reply, do not slow the loop down. Mention it, notify, and continue exactly as before. Pause only if the human explicitly asks you to — an unanswered question here would stall the grind for hours, which is far worse than finishing some lower-priority work first.

Say it once per discrepancy, not once per loop iteration. Repeating the same alert every cycle is noise, and noise gets ignored.

Continuing and stopping

Keep refilling slots until bd ready --exclude-label=human returns nothing within scope. Then wait for the in-flight subagents, merge their branches, and stop.

Run bd human list before the final report. State that the agent-ready queue is empty, then list the beads waiting on the user separately, with what each one asks of them. When a scope was given, name it, so it is clear the agent-ready queue is empty within that scope rather than overall; do not imply the human queue shares that scope unless it was filtered separately.

If that queue is non-empty, end by telling the user they can run /blockers to work through it one bead at a time. Beads flagged during the run are the run's main unfinished business, and a count alone does not tell them how to clear it.

Between merges, do not pause to ask what to do next and do not summarise progress — keep the loop running. The exception is the concurrency question at the very start, and anything that genuinely needs a decision you cannot make (a dirty working tree, an ambiguous scope, a merge conflict whose correct resolution is a judgement call about intent). A shifted backlog is not such a case — flag it and keep going, as above.

Pushing

Push after each successful merge by default, but the repository's own rules win. Do not push when the project's agent instructions, its Beads profile, or a current user or orchestrator instruction prohibits or restricts it. A conservative or minimal Beads profile, or an explicit "do not push", means commit only.

This skill is not authority to override such a rule. When pushing is blocked, keep grinding and committing as normal, and say once that pushes are being held back and why.

Subagents never push, regardless.

Cleaning up

At the end of the run, and after any interruption:

wt list                # any surviving bgp/* worktree is a leak

Remove leftovers with wt remove bgp/<id> -y. Without --force-delete it declines to delete a branch holding unmerged work, which is the behaviour you want here: leave those alone, and say they exist and which beads they belong to, so nothing is silently lost.

Check the remote too. wt merge removes the local branch and worktree, but anything a subagent pushed survives on the remote and no teardown step removes it. A long run leaves one stale bgp/* branch per bead:

git ls-remote --heads <remote> 'refs/heads/bgp/*'

Verify with git cherry, not ancestry. Because wt merge rebases, merged branches have different SHAs, so git log <base>..<branch> and git merge-base --is-ancestor both report unmerged work that landed perfectly. git cherry compares by patch content and sees through the rebase:

git cherry <base> <remote>/bgp/<id>   # lines starting '+' are genuinely unmerged

Delete only branches with no + lines, and report any that do rather than forcing them.