Add sub-agent delegation guidance

2026-03-19 16:00:34 +08:00
parent dd8da4aa58
commit b448d98e71
1 changed files with 34 additions and 0 deletions
@@ -54,3 +54,37 @@ For `skills/`:
 - if a skill depends on an executable or other runtime asset, prefer bundling the required artifact with the skill instead of relying on ad hoc rebuild instructions as the primary workflow
 - when updating an existing skill, keep `SKILL.md`, `agents/openai.yaml`, and bundled assets consistent with each other
 - if you add or materially change a project skill, update [docs/implementation-roadmap.md](/home/kurihada/project/ai-workflow-skill/docs/implementation-roadmap.md) in the same change
+
+## Sub-Agent Delegation
+
+This repository allows sub-agent delegation for parallel implementation work.
+
+### Default Sub-Agent Configuration
+
+When spawning sub-agents for this repo:
+
+- model: `gpt-5.4`
+- reasoning effort: `xhigh`
+
+### Context Strategy
+
+Sub-agents should use this context policy:
+
+1. first attempt with `fork_context: true`
+2. if the sub-agent fails due to context-window pressure, retry with `fork_context: false`
+3. when retrying with `fork_context: false`, provide a compact task brief and explicitly list the local files the sub-agent must read first
+
+### Practical Guidance
+
+- prefer fewer, higher-quality sub-agents over many shallow ones
+- give each sub-agent a clearly isolated write scope
+- keep shared integration files owned by the main thread whenever possible
+- if a sub-agent needs a shared dependency or a shared export change, have it report that back rather than editing unrelated files
+
+### Priority
+
+If there is any conflict between older default delegation habits and this file:
+
+- follow this file for sub-agent model selection
+- follow this file for sub-agent reasoning effort
+- follow this file for context fallback behavior