Skip to content

arbiterForge/codeArbiter / plugins/ca/skills/dispatching-parallel-agents

dispatching-parallel-agents

Unverified

Parallel fan-out for authors, scouts and reviewers. Collect every result; only explicit review batches use finding-triage and verdict-aggregator.

dispatching-parallel-agents

Fan out and collect every result. The caller's existing agent kind and completion contract determine whether this is author/scout work or an explicit review batch. This routing adds no receipt, approval mode or extra review requirement. The primitive owns parallel discipline, not domain logic.

Pre-flight

The caller supplies the work. Confirm it, or STOP and surface the gap — never invent units or pick an agent on a hunch:

  • The unit list — N discrete, independently-completable units of work, each with its scope and target paths.
  • The agent kind — which agent kind to dispatch per unit (e.g. backend-author, frontend-author, scout, security-reviewer at ${CLAUDE_PLUGIN_ROOT}/agents/<name>.md). One kind per batch.
  • The completion contract — what each agent must return: a structured result, its unit ID, pass/fail/error status, required checks and any findings or unfinished obligations. Self-reports are claims, not evidence (see Phase 4). Preserve whether the caller requested authors/scouts or an explicit reviewer batch.

Phase 1 — Partition · gate: BLOCK

Split the work into units that do not collide. A unit owns a distinct file or path set; two units in one batch MUST NOT mutate the same path.

  • Independent paths — units touch disjoint files. Dispatch directly. This is the common, fast case.
  • Shared paths unavoidable — units must mutate the same file or tree. Do NOT dispatch into the conflict. Route to using-git-worktrees (${CLAUDE_PLUGIN_ROOT}/skills/using-git-worktrees/SKILL.md) to give each unit an isolated working tree, or serialize the colliding units into one sequential unit. A shared-path collision in a parallel batch is a corruption, not a merge.

Tag each unit with its ID, scope, target paths, and the agent kind. Read-only batches (review, scout) skip the collision check — they mutate nothing.

Gate: every unit has a disjoint path set, or its overlap is resolved by worktree isolation or serialization. An unresolved collision does not pass.

Phase 2 — Dispatch · gate: BLOCK

Dispatch one agent per unit, bounded.

  • Concurrency bound — dispatch in waves of at most a fixed width (default 4 concurrent agents; the caller MAY override). Do not fan out unbounded; a hundred agents at once starves the batch and corrupts nothing useful.
  • Fresh and isolated — each agent gets only its own unit's scope and the completion contract. No agent sees another's working state. No shared mutable context between units in flight.
  • Hard-stop propagation — a tdd BLOCK, a commit-gate failure, a security CRITICAL, or a [CONFIRM-NN] inside any unit halts that unit and surfaces immediately; it does not get silently swallowed by the batch.

Gate: every unit is dispatched or explicitly deferred. No unit is dropped silently at dispatch.

Phase 3 — Collect · gate: BLOCK

Wait for the full wave, then the next, until every unit has terminated. Record each unit's terminal state: DONE (returned a result), ERRORED (the agent failed), or DEFERRED (a hard-stop surfaced).

An ERRORED unit is not complete. Preserve its diagnostic output and any still-inspectable changed paths for diagnosis; do not apply its unreviewed partial changes as a finished unit. One errored unit MUST NOT corrupt or discard the batch. The remaining units' results stand. Record errors and deferrals for the caller, and for the review funnel when this is an explicit review batch; do not retry blindly.

Gate: every unit has a terminal state and DONE results are captured intact. A still-running or unaccounted unit prevents a complete batch result; partial diagnostic evidence remains visible.

Phase 4 — Dedupe · gate: BLOCK

Parallel agents over adjacent units may report the same defect or contradict each other. Reconcile the actual overlap before the coordinator uses the result.

  • Merge findings that name the same defect at the same location into one, keeping the highest severity reported.
  • Reconcile contradictory results across units (two units that disagree on the same fact) into a single surfaced conflict, not a silent pick.
  • Inspect any "done" claim that gates real work: command, cwd, relevant source/input identity, output and exit code. Reuse accessible applicable passing evidence when those inputs still match; run a missing, changed, insufficient or unresolved check. A success sentence without accessible evidence is not proof. Follow ${CLAUDE_PLUGIN_ROOT}/includes/fresh-verification.md.

Gate: overlapping findings are deduped, contradictions surfaced, and completion claims inspected against applicable evidence. No failed or unknown result becomes a pass because another unit succeeded.

Phase 5 — Return by the caller's contract · gate: STOP

Author/scout batches return inspected results directly to their coordinator: unit IDs, changed paths or source observations, accessible checks, errors, deferrals, contradictions and remaining obligations. The coordinator reviews implementation and applies only accepted units through its normal task and worktree contract. It does not dispatch finding-triage or verdict-aggregator merely to collect author/scout work. This result is not an independent reviewer certificate.

Explicit review batches use finding-triage then verdict-aggregator in order:

  1. Dispatch finding-triage (${CLAUDE_PLUGIN_ROOT}/agents/finding-triage.md) over the complete deduped reviewer result set and batch contract. It classifies severity and marks out-of-scope items with inline [NEEDS-TRIAGE]. Every reviewer finding reaches the verdict; only non-finding transport metadata may be omitted. ERRORED, DEFERRED and missing reviewer units remain incomplete-result findings.
  2. Hand the complete triage report to the read-only verdict-aggregator (${CLAUDE_PLUGIN_ROOT}/agents/verdict-aggregator.md). It returns PASS, BLOCKING_FINDINGS, or INCOMPLETE; incompleteness takes precedence without suppressing blocking findings.

Gate: the caller gets its selected completion result. An explicit review cannot bypass the funnel; an author/scout batch cannot claim a review verdict.

Hard rules

  • MUST NOT dispatch two units that mutate the same path in one batch; isolate via using-git-worktrees or serialize.
  • MUST NOT fan out unbounded; dispatch within the concurrency bound.
  • MUST NOT let one ERRORED unit discard or corrupt the rest of the batch, or apply its unreviewed partial changes as complete.
  • MUST use the finding-triage → verdict-aggregator funnel for every explicit review batch and preserve every finding, contradiction and incomplete unit.
  • MUST NOT require that reviewer funnel for an author/scout batch whose caller requested direct results.
  • MUST NOT trust a subagent's self-reported completion; inspect accessible proving evidence and rerun only missing or affected checks.
  • MUST NOT silently drop a unit — every unit terminates with a visible result, error or deferral.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 463kUpdated Claude CodeCodexVerifiedInstall
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 463kUpdated Claude CodeCodexVerifiedInstall
discernment-nudgeanthropics/skillsAfter you give a substantive answer or draft that the user may act on — advice or recommendations, drafted artifacts such as goals, plans, pitches, proposals, or emails, estimates or projections, analysis or interpretation of data, factual claims they may rely on, or a multi-step argument — invoke this skill BEFORE finalizing your reply and then, if it applies, append 2-3 short follow-up questions, each tied to something specific in what you just produced, that help the user check key facts, probe the reasoning or assumptions, and notice missing context. Do this at most once per conversation. Skip it when the user asked a trivial how-to or simple lookup, wants a purely educational explanation, asked you only to format, convert, or assemble a file from content they provided, is writing code they will run, is doing creative writing or casual chat, or already asked you to double-check, cite, or review — the skill file explains these boundaries and the exact output format.Stars: 195kUpdated Claude CodeCodexVerifiedInstall
internal-commsanthropics/skillsA set of resources to help me write all kinds of internal communications, using the formats that my company likes to use. Claude should use this skill whenever asked to write some sort of internal communications (status reports, leadership updates, 3P updates, company newsletters, FAQs, incident reports, project updates, etc.).Stars: 195kUpdated Claude CodeCodexVerifiedInstall