subagent-execution - SKILL.md Agent Skill

name: subagent-execution description: Delegate bounded implementation, investigation, and review tasks to focused subagents with explicit reasoning effort. Use inside worker or orchestrator sessions when independent subtasks, spec checks, or risk-based reviews can be done in parallel.

Subagent Execution

Use subagents to reduce cost, isolate context, and improve review quality. The controller remains responsible for verifying results.

When To Use

independent implementation subtasks inside one Linear issue
focused investigation or debugging domains
spec-compliance review against Linear issue/PRD
code quality or risk review before production-ready
multiple unrelated failures or flaky tests

Do not use subagents when tasks would edit the same files concurrently without a clear owner, or when the full system context is required in one head.

Reasoning Defaults

Set reasoning_effort explicitly:

low: focused exploration, bounded implementation, mechanical fixes
medium: spec compliance review, fix agents, integration tasks
high: code quality review, auth/security, persistence, architecture
xhigh: broad/high-risk review across auth, data, infra, migrations, or cross-package contracts

Prompt Shape

Every subagent prompt should include:

Linear issue key/title and the relevant acceptance criteria
parent PRD/project summary when needed
exact scope and files, if known
constraints and out-of-scope boundaries
expected output format
whether it may edit files or is read-only
required verification commands for edit agents

Subagents do not inherit the controller's context. Inline the context they need or point to exact local files they must read.

Review Loop

For implementation subtasks:

Dispatch implementer.
Read report and diff.
Dispatch or perform spec-compliance review.
Dispatch or perform code-quality/risk review.
Ask the implementer to fix targeted findings.
Verify materially, not just by trusting the subagent report.

Stop Conditions

Escalate to the user or orchestrator when:

subagent reports NEEDS_CONTEXT twice
subagent reports BLOCKED
implementation requires out-of-scope files
tests fail repeatedly without root cause
the Linear issue appears wrong or stale