Summary

The init wizard seeds an AGENTS.md "Subagent Delegation" section that tells the frontline LLM when to delegate to spawn_agent. The current guidance sets the bar too high — "deep web research requiring multiple searches and synthesis" as the threshold excludes the bulk of research tasks that would benefit from delegation. Live smoke testing earlier in the sub-agent work confirmed this: Qwen performed ~50k tokens of in-process research on a clearly delegable task because the guidance never reached the delegation trigger.

This PR rewrites the section end to end, applying the proposal from #522 verbatim with two adjacent additions I think close a gap in the original proposal:

1. A context-window-protection bullet in "When to delegate". Delegation's biggest structural benefit isn't specialization or parallelism — it's keeping raw file/web content out of the main session's context window. The existing guidance never said that out loud, and Phase 0.2 smoke testing showed it's the real problem (Qwen burned ~15k tokens reading GitHub HTML into its own context when research-assistant would have returned a 1KB cited summary).
2. A new "Per-call specialization" paragraph introducing the optional context argument that shipped in feat(subagents): single-file markdown format + contextual prompt + project-scoped discovery #647 / feat(subagents): single-file markdown format with contextual prompt #652. The discovery context layer advertises it, but AGENTS.md was still telling the model to call spawn_agent with only agent + task.

Expanded "When to delegate" list

• Research requiring 2+ sources or multiple searches (was: "deep research requiring multiple searches and synthesis")
• Parallelizable tasks (multiple independent queries can run concurrently)
• Any work that would otherwise pull large files or web pages into this session's context — the subagent reads them, you get the synthesis (new)
• Background prep work that doesn't block immediate response
• Code analysis on large files or multiple files
• Summarization of long documents or web pages
• Preliminary passes on topics before diving deep

Expanded "When NOT to delegate" list

• Simple single searches (use web_search directly)
• Tasks requiring MCP tools
• Interactive browser tasks
• Tasks where coordination overhead outweighs parallelization benefits (new)

Per-call specialization paragraph (new)

Teaches the model that spawn_agent takes an optional context argument for per-invocation background — workspace details, the user's broader goal, facts the subagent would otherwise have to rediscover — and explicitly warns against duplicating the agent's built-in instructions in it. The runtime side of this shipped in #647 / #652 but the AGENTS.md text was stale.

Scope notes

• Seed template only. This updates the init-wizard seed at src/Netclaw.Cli/Tui/Wizard/Steps/IdentityStepViewModel.cs:205-239. New installs pick it up via netclaw init. Existing users' ~/.netclaw/identity/AGENTS.md files are not mutated automatically — operators who want the new guidance should copy the new block into their identity file manually.
• Not coupled to SubAgentActor: tool file paths don't contribute to parent session WorkingContext #600. Earlier design discussion considered whether this should land with SubAgentActor: tool file paths don't contribute to parent session WorkingContext #600 (sub-agent file paths → parent WorkingContext). After reviewing WorkingContext.cs in detail, SubAgentActor: tool file paths don't contribute to parent session WorkingContext #600 was closed as not planned — the proposed mechanism would misrepresent the parent's actual file awareness (RecentFiles carries an implicit "contents are fresh in my mind" semantic that a subagent-read file doesn't satisfy), and the preservation concern it raised is already handled by the conversation history and the structured findings pipeline (Add structured findings envelopes to SubAgentResult and parent-session acceptance flow #183 / checkpoint heuristics). This PR ships alone.
• Related: ToolIndexUpdater.UpdateSubAgentDiscovery() was enriched in feat(subagents): single-file markdown format with contextual prompt #652 to include per-agent description, tool lists, and a context-aware example call, so the runtime-side discovery block and the static AGENTS.md guidance are now consistent.

Test plan

• dotnet build src/Netclaw.Cli/Netclaw.Cli.csproj — clean
• dotnet slopwatch analyze — 0 issues
• Reviewed the rendered raw-string output by reading the file back at lines 205-239 — indentation matches surrounding block, markdown renders correctly
• Live delegation retest post-merge — rerun the Phase 0.2 smoke prompt against a fresh install with the new seed and see whether Qwen now reaches for spawn_agent organically (this is follow-up validation, not a blocker for merge)

Related issues

• Closes Expand subagent delegation guidance #522
• Built on feat(subagents): single-file markdown format + contextual prompt + project-scoped discovery #647 (single-file markdown format) and feat(subagents): single-file markdown format with contextual prompt #652 (context parameter) — the AGENTS.md text now references the context argument those PRs shipped
• Adjacent to docs(subagents): rewrite runbook for single-file markdown format #655 (runbook rewrite) — both keep the stale JSON+MD references out of user-facing guidance, landing on different files so no conflict