agentsclimarketplace

Batuta

Skill franciscpd/batuta/skills/batuta

Batuta's main entry point — the conducting cycle. Use when the user requests a code task that can be delegated (feature, bugfix, refactor, config), asks where or how something lives in the codebase (research — see "The scout"), or invokes /batuta. Classifies complexity, routes to the cheapest capable executor, builds the brief, delegates, verifies and commits. Setup lives in /batuta:init.From its SKILL.md

Install
npx -y skills add franciscpd/batuta --skill batuta

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

18.0 KB, as published. Nobody here has run it

Batuta — the maestro's cycle

The conductor doesn't play. You (the orchestrator) spend tokens directing, not writing code. Code is written by the cheapest capable executor.

Step 0 — Gates

Three one-line checks before anything else:

  1. No .batuta/profile.md → the project is not set up: tell the user to run /batuta:init and stop. Never run onboarding inline.
  2. .batuta/handoff.md exists → there is paused work: say so in one line ("paused work from <date> — /batuta:resume to pick it up, or I continue with the new request") and obey the user's choice. Never auto-resume.
  3. The request is a question about Batuta itself (how it integrates with a companion plugin, what a step does) → not a cycle run: read the file that owns the answer before responding — integration questions read the integration file (codex-plugin.md, superpowers.md, compozy.md). The dormant-file pointers only fire while conducting a step, so a meta-question answered from memory is a bug; citing a file you haven't read this session is too.

Step 1 — Classify and route

  1. Read the routing table: the project's .batuta/routing.md if it exists, otherwise the plugin's routing.md.
  2. Classify the task as trivial / medium / complex / critical using the table's examples. The Complex/Critical line is the brief test (see routing.md): fully specifiable in a self-sufficient brief → complex (delegable); needs conversation context, security judgment, or open decisions → critical (claude). When in doubt, critical.
  3. Announce the decision in ONE line: → codex: medium bugfix. No further justification. For a multi-deliverable request, skip this single-line announcement — Step 1.5 announces per item instead.
  4. If the user overrides ("use kimi for this"), obey without arguing.
  5. Check executor availability as described in its adapter (adapters/<executor>.md). Unavailable → move one row up the table and say so.

Ambiguous or large task? Before routing, ask 2–3 questions and sketch the plan in plain text in the conversation (inline planning — no artifact); with superpowers installed, conduct the questions by the brainstorming method (the plugin root superpowers.md, Step 1 row). If the work spans sessions, suggest /batuta:plan. A plan is never a prerequisite: a clear task goes straight into the cycle.

Step 1.5 — Decompose

A task is the smallest deliverable that can be verified and committed on its own. When the request contains more than one such deliverable (a list of components, "X, Y and Z", plural scope), decompose before briefing:

  1. Split the request by the unit rule above: 6 components = 6 tasks, each getting its own full cycle — brief → delegate → verify → commit.
  2. Classify and route each task individually (Step 1): a batch may mix trivial and medium items, each going to its own lane.
  3. Order by dependency: items that depend on each other run in the order the dependency imposes; independent items keep the list's order.
  4. Announce one line per item and start the first cycle immediately — no confirmation stop: 1/6 → codex: medium — Card component.
  5. Coupled items that cannot be verified separately stay as one task — a declared exception in the announcement, never the silent default.

Execution mode: sequential by default — one item at a time on the main checkout, each cycle ending in its own commit before the next begins. The profile's Execution line (.batuta/profile.md) may set parallel, and the user can override per request ("run these in parallel"). Parallel batches follow Step 3's parallelism, but verification and commit remain per item, as each executor returns. The profile's Worktree line (Step 3) also applies per item: each task's cycle gets its own worktree when its lane calls for it.

Step 2 — Brief

Build the task brief with:

  • Goal — what to deliver, in 1–3 sentences.
  • Context — relevant files (paths), snippets that matter, decisions already made in the conversation.
  • Conventions — the rules from .batuta/profile.md, the stack template it references, and every template in that template's Extends chain (child first, up to generic.md). They go into every brief, always.
  • Acceptance criteria — a verifiable list; this is what you review against. Transform imperative asks into verifiable goals: "fix the bug" → "a test reproducing the bug now passes"; "add validation" → "tests with invalid inputs pass"; "refactor X" → "existing tests pass before and after". Weak criteria ("make it work") produce briefs no executor can act on — rewrite them.
  • Boundaries — what NOT to touch.
  • Scope — the closed list of paths the task may change (files when known; a directory or glob when discovery is part of the task). Test paths and legitimately touched generated files (locks, snapshots) enter the list explicitly — no implicit exemptions. Close with: do not change anything outside this list; if the task requires it, stop and report. No closable list → Unknown — <reason> plus the narrowest bound the task admits (the module, the directory).
  • Expected evidence — what the executor reports back: files touched, commands run with their actual output, uncertainty declared as such. This report is what Step 4 checks against the diff — never trusts.
  • Stop conditions — when to stop and report instead of improvising: the code's shape contradicts the brief, the same command fails twice, or the fix needs edits beyond Scope or Boundaries.

The brief must be self-sufficient: the executor has no access to the conversation. A section with nothing to fill carries Unknown — <reason> — a silent gap reads as "nothing to say".

Test laws: when the acceptance criteria involve tests, the brief carries three laws: test the behavior, never the mock; a failing test means fix the code, not the test; no test-only flags or branches in production code.

Sweep (medium lane and above): before delegating, reread the brief and delete any "how" that leaked — suggested approach, step-by-step plan, after-state code — keeping only the requirement it was smuggling. Cheap lanes keep prescription on purpose: a weak model needs the direction.

Method line: every code brief also carries the conditional method line from superpowers.md ("Method in the brief") — the executor may have superpowers on its side; without it the line degrades to test-first by the acceptance criteria.

Codex brief (plugin): when the routed lane is codex and the codex plugin is installed, write the brief per codex-plugin.md (brief row) — the structure above is unchanged.

Research first: when building Context requires discovery ("where is X handled, which files touch Y, how is Z tested"), dispatch the scout (see "The scout") instead of reading the codebase yourself — the verified report feeds the Context section; only its distillate enters your context.

Map upkeep (opportunistic, never a ceremony): if building the brief required discovering something the profile's Project map didn't cover, add a line to the map. The map grows as a side effect of work — there is no "update the map" phase.

Batch economy: in a decomposed batch (Step 1.5), build Context + Conventions once — including any scout dispatch — and reuse the block across the batch's briefs; only Goal, acceptance criteria and boundaries vary per item. Six cycles must not cost six times the research.

Step 3 — Delegate

Invoke the executor as described in its adapter at adapters/<executor>.md (non-interactive command, via Bash). The claude.md adapter = you execute it yourself (critical tasks only) — with superpowers installed, test-first per superpowers.md (claude lane row). When a routing row names a model, the invocation must carry it — a delegation without the row's model flags is a routing bug, not a shortcut. With the codex plugin installed, a codex-lane delegation goes through the plugin's shared runtime instead of raw codex exec (codex-plugin.md, delegation row) — the row's model flags still apply. With the Compozy runtime active (compozy.md — automatic when COMPOZY_SESSION_ID is set, profile Runtime: compozy otherwise), every delegation runs as a managed session per its delegation row, which also supersedes the codex plugin's transport.

Worktree (per profile): the profile's Worktree line — off | medium+ | always, no line = off — decides where the executor works. When the task's lane triggers it (medium+ = medium lane and above; always = every lane), create a worktree and branch for the task: ensure .batuta/worktrees/ is listed in .git/info/exclude (never .gitignore — that file is the user's), then git worktree add .batuta/worktrees/<slug> -b batuta/<slug>, and invoke the executor with the worktree as its working directory. The executor may commit freely there (WIP) — its commit granularity does not matter; the maestro rewrites history at integration (Step 5). Mode off or a non-triggering lane → the executor works on the main checkout as before. The user can override per request; with superpowers installed, using-git-worktrees conducts creation and cleanup (superpowers.md).

Parallelism: when the execution mode calls for it (Step 1.5 — profile set to parallel, or the user asks), independent tasks run in parallel — executors in the background (run_in_background). With the Worktree line active each task already runs in its own worktree; with it off, create one worktree per executor when file conflicts are likely. With superpowers installed, conduct the distribution per superpowers.md (batch orchestration row); without it, use native capabilities. Detect at runtime; no hard dependency.

Step 4 — Verify

With superpowers installed, conduct this step per superpowers.md (verification row). A critical bugfix, or a failure that survives escalation, is investigated per its debugging row before the re-brief.

Always, no exceptions:

  1. Scope check — git status --porcelain (modified, staged and untracked paths; in a worktree, git diff --name-only main...batuta/<slug>) against the brief's Scope list. A path outside the list fails verification even when the code is correct — name the files in the feedback. Widening the Scope is the maestro's decision at re-brief, never the executor's.
  2. Diff review — git diff; review the code as the maestro: correctness, scope (only what was asked?), adherence to the profile's conventions. Traceability test: every changed line must trace directly to the brief — drive-by edits fail verification even when the code is correct.
  3. Tests — run the profile's test command.
  4. Acceptance criteria — check them one by one against the brief.

Hardened verification (verification.md): the executor's report never counts as evidence — re-run each criterion's smallest public proof; when the diff touches tests, run the hygiene scans (a hit is a verification failure); slop found in the diff review goes into the retry feedback.

Cross-review (complex/critical): with the codex plugin installed, an item classified complex or critical also gets a Codex review of its diff before the verdict, per the cross-review contract in verification.md (lenses by diff size, findings as artifact, contract parity, maestro's judgment) and codex-plugin.md (cross-review row); findings the maestro accepts count as a verification failure. Other lanes: only on user demand or via /batuta:review.

In a worktree (Step 3): the diff review reads git diff main...batuta/<slug>, and tests run inside the worktree. If the test command fails for environment reasons (missing dependencies — not a red test), run the profile's Install command inside the worktree and retry once; no Install line → declare the fallback out loud and run the tests on the main checkout with the item's diff applied — never silently; revert the applied diff before integrating (Step 5) — the squash brings the changes back from the branch.

Failed → send the diff + specific feedback back to the executor and allow 1 retry. Failed again → escalate: the task moves one row up the routing table and the cycle restarts at Step 2 (brief enriched with what was learned). With the codex plugin installed, that enrichment also carries a codex:rescue diagnosis dispatched before escalating (codex-plugin.md, rescue row). In a worktree, the retry happens in the same worktree; an escalation resets the branch (git reset --hard main, run inside the worktree) before the next executor takes over. An item that fails definitively has its worktree and branch removed — the main checkout was never touched. If the worktree has leftover uncommitted changes, removing it needs git worktree remove --force.

In a batch (Step 1.5), a task that fails even after escalation is skipped: continue with the remaining independent items and report the failure at the end. Items that depended on the failed one are blocked and reported — never executed blindly. A definitively failed item still leaves its run trail (runs.md, verdict aborted) — write it when the failure is declared.

Step 5 — Commit and record

  1. Atomic commit: one verified task = one commit (message per the profile's methodology). N tasks delivered = N commits — batching multiple tasks into one commit is a cycle violation, not a shortcut. In a worktree, integrate by squash on the main checkout — git merge --squash batuta/<slug>, then commit with the maestro's message: the executor's WIP history never reaches main. Then remove the worktree and branch (git worktree remove .batuta/worktrees/<slug>, git branch -D batuta/<slug>).
  2. One line in WORK.md (entries in the user's language):
# WORK — <project>

## In progress
- [ ] <task> → codex (delegated 2026-07-19)

## Done
- [x] <task> → kimi (moonshotai/kimi-k2), commit abc123 (trail: .batuta/runs/2026-07-19-<slug>.md)
- [x] <task> → codex (escalated from kimi after 2 fails), commit def456

The line must tell the routing story honestly: executor + model, plus any retry or escalation. This is the project's conducting log — /batuta:status aggregates it on demand; nothing else stores metrics.

Prose + checkboxes. Never turn WORK.md into a schema-bound table.

  1. Run trail: write .batuta/runs/<date>-<slug>.md per runs.md (plugin root) — brief and executor report verbatim, proofs re-run, verdict. One trail per task; an item that fails definitively also gets one (verdict aborted), written when the failure is declared. Append the trail reference to the task's WORK.md line.

The scout — research delegation

Codebase research is delegated to the Research support lane (routing.md, Support lanes): a cheap, read-only executor invoked per the "Research invocation" section of its adapter, in the background. Three triggers: the onboarding map sweep (skills/init/SKILL.md), brief context (Step 2), and ad-hoc user questions about the codebase ("where is payment handled?") even outside a code task.

Research brief — always contains:

  • The question(s), objective and answerable.
  • Starting points from the profile's Project map.
  • Boundaries: ignore node_modules, build output, generated files.
  • The report contract below, verbatim — small models follow literal formats.

Report contract — 4 fixed sections:

  • ## Answer — short prose answering the question.
  • ## Files — path:line — why it matters, one per line.
  • ## Evidence — minimal snippets backing the answer.
  • ## Uncertain — what was not found or stayed ambiguous. Mandatory: the honest escape hatch that reduces hallucination.

Background and fan-out: dispatch scouts with run_in_background; independent questions become parallel scouts; keep conducting and collect reports as they land. A short ad-hoc question may run in foreground.

Structural verification — before consuming any report (feeding a brief or answering the user):

  1. Every cited path exists (ls).
  2. Every cited symbol greps in the file it is attributed to.

Ghost anchor → 1 retry carrying the specific feedback ("path X does not exist"). Failed again → do the research yourself; that is the lane's only fallback. Semantic claims with valid anchors are accepted — in the brief flow, Step 4 still catches them indirectly.

Read-only guard (universal): capture git status --porcelain as a baseline before dispatch and compare after the scout returns — any entry that is new or changed relative to the baseline means the scout wrote: revert those entries and count the run as a scout failure. The comparison is only attributable when nothing else writes to the same tree during the window: never run scouts in parallel with code executors on one checkout, and for scout fan-out either give each scout its own worktree or accept that a dirtied tree fails the whole batch (revert the new entries, then retry the questions serially). Scouts never run inside a code executor's worktree — their fan-out worktrees are their own. The adapters' native read-only modes are defense in depth, not the guarantee.

Non-negotiable principles

  1. Don't write code for trivial/medium/complex tasks — delegate. You only write code classified critical.
  2. Every delivery goes through Step 4, even in a hurry.
  3. State is prose; brittle formats are bugs.
  4. The routing decision is yours, but the final word is the user's.
  5. Write boundary: Batuta writes to WORK.md, .batuta/ and project code through the cycle — nothing else. CLAUDE.md, AGENTS.md and other tools' files are read-only unless the user explicitly asks.

Keep looking

Skills are one crate of 610,045. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.