feat: wire agent.delegate tool with sub-agent configs

- Export createAgentDelegateTool through builtin/index.ts → tools/index.ts
- Register agent.delegate in routing.ts with lazy orchestrator pattern
- Add agent.delegate + agents.list to messaging and coding policy profiles
- Add group:agents tool group to policy.ts
- Add research/code/comms agent config examples to default.yaml
- Add research/code/comms agent configs to user config.yaml
- Add 11 tests for agent-delegate tool (all pass)
- Typecheck clean, no regressions

config/default.yaml

@@ -186,6 +186,35 @@ models:
   # Then reference them in fallback_chain:
   #   fallback_chain: [ollama, llamacpp, local]
 
+# Optional: external CLI backends for non-tool conversation turns.
+# - native remains available unless you disable it.
+# - set `default` to route normal text turns to a backend by default.
+# - per-agent override is available via `agent_configs.<name>.backend`.
+# backends:
+#   default: codex                  # claude_code | opencode | codex | gemini
+#   native:
+#     enabled: true
+#   claude_code:
+#     enabled: false
+#     # path: /usr/local/bin/claude
+#     # args: ["--print", "{prompt}"]
+#     # timeout_ms: 120000
+#   opencode:
+#     enabled: false
+#     # path: /usr/local/bin/opencode
+#     # args: ["-p", "{prompt}"]
+#     # timeout_ms: 120000
+#   codex:
+#     enabled: false
+#     # path: /usr/local/bin/codex
+#     # args: ["-p", "{prompt}"]
+#     # timeout_ms: 120000
+#   gemini:
+#     enabled: false
+#     # path: /usr/local/bin/gemini
+#     # args: ["-p", "{prompt}"]
+#     # timeout_ms: 120000
+
 # Optional: Kubernetes / homelab awareness tools (k8s.pods, k8s.deployments, k8s.logs)
 # k8s:
 #   enabled: false
@@ -422,3 +451,37 @@ hooks:
 #     wake_phrase: "hey flynn"
 #     timeout_ms: 120000
 #     allow_manual_toggle: true
+
+# ── Sub-Agent Configs ────────────────────────────────────────────────
+# Named agent configurations for delegation via agent.delegate tool.
+# Each agent gets a focused system prompt, model tier, and tool profile.
+#
+# agent_configs:
+#   research:
+#     model_tier: default
+#     tool_profile: messaging
+#     system_prompt: |
+#       You are a research agent. Your job is to find, verify, and synthesize
+#       information from the web. Be thorough but concise. Cite sources when
+#       possible. Return structured findings — not conversational filler.
+#       Use web.search to find sources, web.fetch to read them, and file.write
+#       to save findings when asked.
+#
+#   code:
+#     model_tier: complex
+#     tool_profile: coding
+#     system_prompt: |
+#       You are a code agent. Your job is to read, write, debug, refactor,
+#       and review code. You have full access to the filesystem and shell.
+#       Be precise. Read files before modifying them. Run tests after changes.
+#       Use file.edit/file.patch for surgical edits, not full file rewrites.
+#       Commit conventions: conventional commits, small atomic changes.
+#
+#   comms:
+#     model_tier: fast
+#     tool_profile: messaging
+#     system_prompt: |
+#       You are a communications agent. Your job is to draft messages,
+#       summarize emails, triage inbox items, and prepare quick replies.
+#       Be concise and match the operator's tone. Skip marketing emails.
+#       Never send messages without explicit instruction — draft only.