run-parallel-agents-feature-build（运行并行 Agent 功能构建）

Run Parallel Agents Feature Build

Inputs

• $request: Optional plan path, task selection, or user direction such as execute layer 0 from.ulpi/plans/foo.json or build these three features in parallel.

Goal

Exploit Claude Code's real parallel agent runtime to execute independent build tasks concurrently without file conflicts, missing context, or fake orchestration.

Step 0: Resolve the candidate work set

Identify the source of truth for the work:

• an approved DAG plan
• an explicit user task list
• a clearly decomposed set of independent implementation lanes

Then determine:

• task count
• dependency edges
• write scope overlap
• whether each lane is build work rather than research or debugging

Do not use this skill if:

• there are fewer than 3 independent lanes
• any lane depends on another lane's result
• multiple lanes write the same file or the same narrow subsystem surface
• the task is still at planning or exploration stage

Use AskUserQuestion only if the decomposition is ambiguous and safe parallelization depends on a user choice.

Success criteria: You have a concrete set of parallel-safe build lanes.

Step 1: Prove independence, do not assume it

For each candidate lane, verify:

• required inputs are already available
• no output from another lane is needed first
• write scope is disjoint
• validation can be run per lane or after merge without ambiguity

If a plan exists, treat its dependency graph and write scope as the primary source.

If there is no plan, derive equivalent temporary structure:

• lane name
• scope
• expected files
• acceptance criteria
• validation command

Load references/agent-runtime-semantics.md for the runtime rules that make or break parallel execution.

Success criteria: Every lane is independently executable without merge-race risk.

Step 2: Match each lane to the right agent

Choose the most appropriate specialized agent for each lane based on:

• repository technology
• target files and directories
• framework-specific patterns
• whether the work is implementation or general support

Load references/agent_matching_logic.md for detailed matching rules and edge cases.

Practical defaults:

• use the most specific framework/domain agent available
• use general-purpose only when no stronger fit exists
• do not invent agent types

Success criteria: Every lane has a justified agent assignment.

Step 3: Build full agent briefs

Because fresh subagent_type agents start without the main thread's task context, every prompt must contain:

• exact scope
• why this task exists
• files or directories to inspect
• write scope
• constraints and patterns to follow
• acceptance criteria
• validation command
• explicit instruction whether the lane should write code or only investigate

If the work came from a DAG plan, preserve:

• task ID
• acceptance criteria
• write scope
• validation command
• review requirement

If the task is identity, routing, registry, or isolation sensitive, require the lane to use references/task-exit-gate.md before declaring completion.

Success criteria: Each lane prompt is complete enough for a fresh agent to act without asking basic clarifying questions.

Step 4: Launch all agents in one parallel batch

Launch every lane in a single assistant message containing multiple Agent(...) tool uses.

Required launch semantics for code-writing lanes:

• subagent_type: matched agent type
• description: short 3-5 word description
• prompt: full task brief
• run_in_background: true
• isolation: "worktree"

Runtime-specific rules:

• do not send the agent launches across multiple assistant messages if the user asked for parallel work
• do not rely on agent defaults for async or isolation when code-writing safety matters
• do not use read-only forks here; this skill is for fresh specialized build agents

Load references/agent-runtime-semantics.md for the exact source-backed reasons behind these rules.

Success criteria: All independent lanes are launched concurrently with explicit, safe runtime settings.

Step 5: Aggregate completions without polling

After launch:

• continue with non-overlapping coordination work
• wait for completion notifications rather than polling background transcripts
• trust the returned agent summaries unless there is a concrete reason to inspect deeper

When each lane completes, capture:

• status
• files touched
• validation result
• blockers or deviations from brief

If a lane is blocked, do not hide it inside a merged success summary.

Success criteria: Every spawned lane has a tracked outcome without unnecessary transcript noise.

Step 6: Review, conflict check, and closeout

Before treating a lane as merged:

• check for write-scope conflicts across finished lanes
• if the plan specifies a review field, invoke the corresponding review skill:

- codex → codex-review - claude → claude-review - kiro → kiro-review

• if the lane came from a structured task plan, require the lane to satisfy references/task-exit-gate.md

If conflicts exist:

• stop automatic aggregation
• identify the overlapping files
• merge intentionally or escalate to the user

Success criteria: Finished lanes are reviewed and conflict-checked before being treated as complete.

Step 7: Report the consolidated result

Summarize:

• which lanes ran
• which agents handled them
• completion status per lane
• files or areas changed
• reviews performed
• remaining blockers or follow-up sequencing

Be explicit about partial success. Parallel execution is only a win if the user can see which lanes are actually done.

Success criteria: The user gets a concise but accurate parallel-execution summary.

Guardrails

• Do not use this skill for fewer than 3 genuinely independent lanes.
• Do not parallelize tasks with shared write scope or dependency edges.
• Do not keep giant agent catalogs or launch templates inline in SKILL.md.
• Do not assume fresh agents know the surrounding task context.
• Do not poll or tail background agent output unless the user explicitly asks.
• Do not rely on default agent background or isolation settings when explicit launch parameters are safer.
• Do not add disable-model-invocation; this skill should remain available when the user asks for parallel execution.
• Do not add context: fork; this workflow coordinates fresh specialized agents, not read-only skill forks.
• Do not add paths:; this is a generic orchestration skill.

When To Load References

• references/agent-runtime-semantics.md Use for source-backed Agent runtime rules: fresh context, single-message parallel launch, background behavior, worktree isolation, and explicit parameter precedence.
• references/agent_matching_logic.md Use for detailed agent-selection rules and framework-specific edge cases.
• references/task-exit-gate.md Use when the task came from a structured plan with acceptance criteria, write scope, and validation requirements.

Output Contract

Report:

1. resolved parallel lane set
2. agent assignment per lane
3. launch status
4. completion and review status per lane
5. conflicts, blockers, or remaining sequential work