parallel-execution（并行执行）

Parallel Execution

Overview

When multiple distinct failures appear across separate subsystems, tackling them one after another wastes time. Each investigation is self-contained and can run simultaneously.

Core principle: Assign one agent per isolated problem domain. Let them operate concurrently.

Prime Directive

NO CONCURRENT DISPATCH WITHOUT CONFIRMING ISOLATION

No exceptions. No workarounds. No shortcuts.

When to Use

digraph dispatch_decision {
    "Multiple problems?" [shape=diamond];
    "Are they isolated?" [shape=diamond];
    "Single agent handles all" [shape=box];
    "One agent per domain" [shape=box];
    "Can they run concurrently?" [shape=diamond];
    "Sequential agents" [shape=box];
    "Parallel dispatch" [shape=box];

    "Multiple problems?" -> "Are they isolated?" [label="yes"];
    "Are they isolated?" -> "Single agent handles all" [label="no - coupled"];
    "Are they isolated?" -> "Can they run concurrently?" [label="yes"];
    "Can they run concurrently?" -> "Parallel dispatch" [label="yes"];
    "Can they run concurrently?" -> "Sequential agents" [label="no - shared state"];
}

Appropriate when:

• Three or more test files failing for distinct reasons
• Multiple subsystems broken independently
• Each issue can be understood without knowledge of the others
• No shared mutable state between the work items

Not appropriate when:

• Failures are connected (resolving one may resolve others)
• Understanding requires a holistic system view
• Agents would interfere with each other's work

The Approach

1. Map Independent Domains

Classify failures by root cause area:

• Module X tests: Permission validation flow
• Module Y tests: Queue drain behavior
• Module Z tests: Cancellation handling

Each domain is self-contained — fixing permissions has no bearing on cancellation tests.

2. Craft Focused Agent Briefs

Each agent receives:

• Defined scope: One test file or subsystem
• Concrete objective: Get these tests green
• Boundaries: Do not modify code outside your scope
• Deliverable: Summary of root cause and changes made

3. Launch Concurrently

// Dispatch parallel agents using the Agent tool in Claude Code
Agent("Fix permission-validation.test.ts failures")
Agent("Fix queue-drain-behavior.test.ts failures")
Agent("Fix cancellation-handling.test.ts failures")
// All three execute at the same time

4. Reconcile and Verify

When agents complete:

• Study each summary
• Confirm changes do not overlap
• Execute the full test suite
• Integrate all results

Agent Brief Structure

Effective agent briefs share these qualities:

1. Narrow — One clear problem area
2. Self-sufficient — Everything needed to understand the issue is included
3. Output-specified — The agent knows exactly what to report

Resolve the 3 failing tests in src/workers/queue-drain.test.ts:

1. "should drain remaining items on shutdown" - expects empty queue, finds 2 items
2. "should handle mixed priorities during drain" - high-priority item processed last
3. "should report drain metrics accurately" - expects 3 metrics but receives 0

These stem from async drain ordering. Your assignment:

1. Read the test file and understand the intended behavior
2. Determine root cause - ordering bug or stale test assumptions?
3. Fix by:
   - Implementing deterministic drain ordering
   - Correcting production bugs if found
   - Updating test expectations if behavior legitimately changed

Do NOT just add delays or increase timeouts - identify the underlying issue.

Return: Root cause analysis and description of changes made.

Pitfalls

Too broad: "Fix all the tests" — the agent loses focus Focused: "Fix queue-drain.test.ts" — narrow scope

No context: "Fix the ordering bug" — the agent has no starting point Contextual: Paste error messages and failing test names

No boundaries: Agent may restructure unrelated code Bounded: "Do NOT modify production code outside src/workers/"

Vague output request: "Fix it" — you have no idea what changed Specific output request: "Return root cause summary and list of modified files"

Cognitive Traps

When NOT to Use

Coupled failures: Resolving one may cascade fixes to others — investigate holistically first Holistic understanding required: Comprehension demands seeing the full system Exploratory debugging: The problem space is unknown Shared resources: Agents would conflict (same files, same databases, same services)

Worked Example

Situation: 8 test failures across 3 files after a major refactor

Failures:

• permission-validation.test.ts: 3 failures (auth token issues)
• queue-drain-behavior.test.ts: 3 failures (ordering bugs)
• cancellation-handling.test.ts: 2 failures (incomplete cleanup)

Judgment: Isolated domains — auth validation, queue drain, and cancellation logic are unrelated

Dispatch:

Agent 1 -> Fix permission-validation.test.ts
Agent 2 -> Fix queue-drain-behavior.test.ts
Agent 3 -> Fix cancellation-handling.test.ts

Outcomes:

• Agent 1: Replaced hardcoded tokens with dynamic fixtures
• Agent 2: Fixed drain ordering to respect priority field
• Agent 3: Added cleanup hooks for partial cancellation state

Reconciliation: All fixes touched separate files, zero conflicts, full suite green

Value: Three problems solved concurrently instead of sequentially

Advantages

1. Concurrency — Multiple investigations proceed at once
2. Concentration — Each agent manages a small context
3. Non-interference — Agents stay out of each other's way
4. Throughput — N problems solved in the time of one

Post-Dispatch Checks

After agents return:

1. Study each summary — Understand what changed and why
2. Detect overlaps — Did any agents touch the same code?
3. Run the full suite — Confirm all fixes coexist
4. Spot-check — Agents can introduce systematic errors
5. Finalize — REQUIRED SUB-SKILL: Use godmode:merge-protocol to land the work

Guardrails

Prohibited:

• Dispatching concurrent agents without verifying task isolation
• Allowing agents to modify the same files simultaneously
• Integrating without reviewing agent summaries
• Using parallel dispatch for exploratory diagnosis
• Dispatching without defined scope, boundaries, and deliverables

Mandatory:

• Confirm isolation before dispatching
• Give each agent a narrow, explicit scope
• Include error messages and test names in agent briefs
• Review all agent summaries before integration
• Execute full test suite after integrating all changes
• Check for conflicts between agent outputs

Integration

godmode:team-orchestration:

• Parallel execution is a foundational pattern for team-based work
• Use team orchestration when agents have partial dependencies

godmode:delegated-execution:

• Launch parallel agents for independent tasks within a development plan
• Each agent owns one task from the plan

godmode:agent-messaging:

• Agents dispatched in parallel should not need inter-agent communication
• If they need to communicate, they are not isolated enough for parallel dispatch