跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
开发与测试

parallel-execution

Use when two or more problems can each be solved without the others' results or shared state, such as unrelated test failures in separate subsystems, independent investigations, or separate modules with no common files.

NoobyGains · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add NoobyGains/godmode --skill parallel-execution

技能说明


name: parallel-execution description: "Use when two or more problems can each be solved without the others' results or shared state, such as unrelated test failures in separate subsystems, independent investigations, or separate modules with no common files."

Parallel Execution

When several problems are genuinely independent, working them one after another wastes wall-clock time and fills one context with unrelated detail. Give each problem its own agent with a focused brief, let them run at the same time, and reconcile the results yourself.

Core principle: one agent per independent problem domain, independence proven before dispatch, and a full test run after reconciliation. Parallelism is only safe when nothing is shared; the check comes first, not after the merge conflict.

The Prime Directive

NO PARALLEL DISPATCH UNTIL INDEPENDENCE IS PROVEN AND ISOLATION IS IN PLACE

No exceptions. No workarounds. No shortcuts.

When to Use

Required:

  • Two or more problems, each understandable without the others' context.
  • No shared files, no shared mutable state (database, ports, caches, generated files), no ordering between them.
  • Examples: three test files failing for different reasons in different subsystems; researching three unrelated libraries; two modules with disjoint file sets.

Skip:

  • The failures may share a cause (fixing one may fix the others): investigate together first with godmode:fault-diagnosis.
  • You do not yet know what is broken: explore before splitting.
  • The work is a sequential plan: use godmode:delegated-execution or godmode:inline-execution.
  • Workers must exchange findings mid-work: use godmode:team-orchestration.
  • Only one problem, or two tiny ones: do them yourself; dispatch overhead exceeds the gain.

The Protocol

  1. Enumerate the problems. One line each: what is broken or wanted, and the evidence (failing test names, error text, the request).

  2. Map each to its domain. For every problem, list the files it will likely read and write, the commands it will run, and any state it touches (test database, port, cache directory, lockfile, generated code).

  3. Prove independence. Build a small table: rows are problems, columns are write sets and shared resources. Any overlap in written files or touched state means those problems are not independent; merge them into one domain or run them in sequence. Write the table down; it is the evidence for the decision.

digraph independence {
    "Problems listed" -> "Write sets overlap?";
    "Write sets overlap?" -> "Merge into one domain or run in sequence" [label="yes"];
    "Write sets overlap?" -> "Shared state (DB, port, cache)?" [label="no"];
    "Shared state (DB, port, cache)?" -> "Isolate the state or run in sequence" [label="yes"];
    "Shared state (DB, port, cache)?" -> "Read-only work?" [label="no"];
    "Read-only work?" -> "Dispatch in parallel, same checkout" [label="yes"];
    "Read-only work?" -> "One worktree per agent, then dispatch" [label="no"];
}
  1. Put isolation in place. Read-only investigations can share the checkout. Any agent that writes gets its own worktree (godmode:workspace-isolation, or the harness's built-in per-subagent worktree option where it has one) so that a mistaken edit outside the expected files cannot clobber another agent's work. Confirm each worktree exists and its baseline is known before dispatch.

  2. Write one brief per agent using godmode:agent-briefing. Each brief names: the single problem; the evidence (error text, test names, verbatim); the files in scope and the files out of scope; constraints ("do not change production code", "do not raise timeouts"); the exact verification command; the report-file path; and "do not dispatch subagents". Choose each agent's model and effort per godmode:model-routing and name both on the dispatch.

  3. Dispatch all of them in the same turn, so they run concurrently. One per turn is sequential execution with extra overhead.

  4. Wait without burning context. Do local work while they run. When idle, wait in bounded stretches; between stretches, check for agents that finished without reporting.

  5. Read the short replies, then the report files only where a reply raises a concern. Each report says: root cause, what changed, files, verification command and output.

  6. Reconcile. For each agent:

    • Check its diff touches only the files its brief allowed. Anything outside is a finding.
    • Check no two diffs touch the same file or the same interface. If they do, resolve it by hand and say so.
    • Spot-check one change per agent against its stated root cause. Agents make systematic errors (a timeout raised instead of a race fixed).
    • Merge the worktree branches into the working branch one at a time, running that agent's verification after each merge.
  7. Run the full suite on the integrated result, not the per-agent subsets. Green per agent proves nothing about the combination. Read the output.

  8. Report to the user: per problem, the root cause and the fix; the reconciliation (conflicts found, how resolved); the full-suite result with counts. Then hand off to godmode:code-review or godmode:merge-protocol as the work requires.

Brief Skeleton

Role: You are fixing one problem in [repo], in [worktree path].
Problem: [one sentence]
Evidence: [failing test names and error text, verbatim]
In scope: [files you may change]
Out of scope: [everything else; do not change]
Constraints: [for example: find the real cause; do not raise timeouts; do not edit tests to pass]
Verify with: [exact command]; it must pass and its output must be clean.
Do not dispatch subagents.
Report: write root cause, changes, files, the command and its output to [REPORT_FILE];
reply with status (DONE | BLOCKED | NEEDS_CONTEXT), commits, one-line test result, concerns.

Cognitive Traps

RationalisationTruth
"These look unrelated, n
阅读完整来源 ↗