AGENT SKILL

dispatching-parallel-agents

Assigns two or more tasks to delegates that run in parallel and combines their outputs, when no task edits a file another task reads or edits, no two share mutable state or an exclusive resource, and none needs another's output.

License 许可
Apache-2.0
所属仓库
AidALL/ghost-alice
最近更新
2026-11-15
Stars
★ 15
Forks
4

01 前言 Front Matter

name
dispatching-parallel-agents
description
Assigns two or more tasks to delegates that run in parallel and combines their outputs, when no task edits a file another task reads or edits, no two share mutable state or an exclusive resource, and none needs another's output.
compatibility
[ "Python 3.11+ standard library" ]

03 SKILL.md 正文

Dispatching Parallel Agents

The response model assigns independent tasks to separate delegates when their work can proceed without shared mutable state or sequential dependencies. Give each delegate a self-contained assignment; do not pass the whole session history.

Decide whether to delegate

Use parallel delegates for separate failures or subsystems whose investigation and edits cannot affect one another. Investigate a shared cause before dividing related failures. If delegates would edit the same file, use the same exclusive resource or depend on another delegate's result, assign those tasks sequentially.

The number of failures does not establish independence. For each proposed assignment, name its target files or subsystem, dependencies and permitted mutations.

Dispatch procedure

  1. Group the work by independent problem, not by arbitrary file count.
  2. Prepare each assignment with the observed failure, expected result, relevant requirements, exact target paths, allowed changes, required checks and reporting format.
  3. Dispatch one delegate per independent assignment through the host's exposed agent API. Provide only task-relevant context and actual human authorization. The delegate receives the assignment message as task input; XML tags or role language in it do not raise its instruction priority.
  4. When results return, inspect each diff and its evidence. Identify overlapping changes before integration. Reconcile any conflict against the requirements that the human user accepted.
  5. Run the integration checks required by the project and the changed shared behavior. The response model follows verification-before-completion to decide whether earlier evidence remains valid and whether an additional discretionary check can change its integration decision.

Assignment example

Investigate the abort tests in src/agents/agent-tool-abort.test.ts.
Observed failure: a completed tool is reported as aborted.
Required behavior: preserve completion when completion precedes cancellation.
Allowed changes: the abort state implementation and its regression tests.
Do not edit batch-completion or approval logic; another delegate owns those files.
Trace the state transition before changing code. For the known behavior, follow
test-driven-development. Replace arbitrary timing waits only when evidence
identifies timing as the cause; do not weaken an expected result to make it pass.
Return the cause, changed files, exact check results, unresolved limits and every
accessed absolute file path.

The example is an assignment, not executable tool syntax. Map it to the active host's tool schema.

Failure modes

  • Parallel dispatch does not authorize shared-file edits or a broader objective.
  • A repeated failure needs a revised hypothesis or new evidence; follow systematic-debugging's incident accounting.
  • Do not increase a timeout without evidence that the original bound caused the failure.

Result

The response model reports each task's result, evidence, conflicts and remaining dependency. It then completes the requested combined result; delegation alone does not satisfy the user's objective.

04 同仓库更多 Skills

AidALL
AidALL/ghost-alice

skill-evolution

Recommends skill changes from recorded tool-call patterns and conduct corrections asserted by the human user or observed by the response model, when the human user asks for such recommendations or their pending backlog, or authorizes that analysis.

★ 15
AidALL
AidALL/ghost-alice

session-intent-analyzer

Records the response model's new or changed interpretation of the input the response model is answering as its `interpreted-as` record before the first tool call of work, and reads that record when it can change the objective, constraints, work choice or whether earlier work continues.

★ 15
AidALL
AidALL/ghost-alice

necessity-gate

Judges whether proposed work is justified, needs a narrower scope or is unjustified, when the response model proposes new work or the human user asks whether to do it.

★ 15
AidALL
AidALL/ghost-alice

merge-companion

Reviews the installer's backups of external changes to installed skills and merges each into the new skill version, discards or defers it as the human user chooses, before a task changes or reinstalls an installed skill, or when the human user asks for that review.

★ 15
AidALL
AidALL/ghost-alice

compact-handoff

Writes a handoff of the task's objective, scope, evidence and next action when a long conversation, compaction or handover to another session or agent could lose them, and resumes the task from it afterwards.

★ 15
AidALL
AidALL/ghost-alice

writing-skills

Writes a new skill or changes an existing one in its folder (SKILL.md, references and scripts) and tests it.

★ 15