hayatosc/dotfiles/subagent-orchestration

Subagent Orchestration

Delegate only when the user or an applicable instruction requests agent delegation. Choose a bounded task that can run independently while you make useful progress; handle short or tightly sequential work locally.

Handoff

Use the current harness's available agent tool and actual parameter schema. Supply:

• The outcome, relevant paths, known decisions, and acceptance criteria.
• File or responsibility ownership. Tell editing agents they share the workspace, must preserve others' edits, and should adapt to concurrent changes.
• Only the skill references relevant to the assigned task. Account for context already inherited by the agent instead of requiring every project skill again.
• Boundaries such as read-only work, no new dependencies, or no further delegation.
• A compact report contract: conclusion, changed files or path:line evidence, check results, and blockers.

Use configured roles where available: coder for implementation, explorer for read-only navigation, and reviewer for an assigned review. Use deep-reviewer only for an explicit user request for deep or advanced review. Preserve the configured model; an override requires a concrete stated reason.

Harness Differences

Select skills from the assigned files: TypeScript → typescript-best-practices, Go → golang-best-practices, Hono → hono-best-practices, and the corresponding framework skill for Vue, Svelte, or MoonBit. Include coding-style for coding/review and code-search for exploration/review. Do not scan unrelated modules solely to collect skills.

Coordination and Completion

Split by independent ownership, not arbitrary file counts. For contested or critical architectural decisions, use a parallel bakeoff: spawn 2-3 independent subagents with distinct approaches in isolated paths, evaluate their trade-offs against explicit criteria, and graft the strongest ideas from non-winning attempts into the chosen base. Reuse an existing agent for follow-up work that benefits from its context. Continue your independent work while agents run; use completion notifications or the harness's wait tool when their result is the next dependency.

Integrate reports against the requested outcome and existing check evidence. Resolve conflicting edits or unsupported claims before relying on them, without duplicating completed searches and checks. Send a focused follow-up when needed, or finish locally if a delegate fails.

Delegation is complete only when the assigned results are integrated and the parent task's acceptance criteria are met, or a concrete blocker is reported. Summarize the result for the user; do not assume they saw an agent's report.