feat(tools/wot_engine): add Web-of-Thought multi-agent reasoning

skills/coordination/web-of-thought/SKILL.md

@@ -0,0 +1,103 @@
+---
+name: web-of-thought
+description: "When to call wot_chat; design agents; modes & opt-in."
+version: 1.1.0
+author: Hermes Agent
+license: MIT
+metadata:
+  hermes:
+    tags: [debate, perspectives, critique, red-team, multi-agent, synthesize, wot, web-of-thought]
+    related_skills: []
+---
+
+# Web-of-Thought (`wot_chat`)
+
+**Opt-in only.** Unavailable unless `HERMES_ENABLE_WOT=1` (or `true`/`yes`/`on`).
+Even then, include the `wot` toolset deliberately.
+
+## Provider
+
+Inner agents use **Hermes active session credentials** by default
+(`resolve_hermes_endpoint` → runtime / config / `resolve_provider_client`).
+
+Override with env (lab multi-backend):
+
+- `LLM_BASE_URL` + optional `LLM_API_KEY` + `LLM_DEFAULT_MODEL`
+- else `OLLAMA_URL` / localhost
+
+Transcript `backend.endpoint_source` shows which source won.
+
+## When to use
+
+Reach for `wot_chat` when **at least two** apply:
+
+- Multi-perspective question (tradeoffs, orthogonal review axes)
+- Single answer likely brittle/biased
+- Real downside cost if wrong
+- You would otherwise simulate “another angle” in one reply
+
+## When NOT to use
+
+Simple facts, casual chat, time-sensitive tiny asks, tight token budgets
+(`agents × rounds` multiplies spend).
+
+## Designing agents
+
+Minimal system prompts (1–2 sentences). No long role-play bios.
+
+```json
+{"name": "alpha", "system_prompt": "Argue the case for. Brief."}
+{"name": "beta",  "system_prompt": "Argue the case against. Brief."}
+{"name": "gamma", "system_prompt": "Synthesize alpha and beta."}
+```
+
+Names: alphanumeric / `_` / `-` (spaces auto-sanitized).
+
+## Modes
+
+| Mode | Behavior |
+|---|---|
+| `parallel` (default) | Concurrent rounds; peers see completed messages next round |
+| `streaming` | Concurrent + chunked peer partials absorbed **next turn only** — **not** mid-token injection into an in-flight peer request |
+| `sequential` | Round-robin full transcript |
+| `queue` | Wake on `@Name` DM, or `#tag` matching `interests`, or any inbox if interests empty. Round 1 wakes all |
+
+Routing in content:
+
+- `@Peer: …` → direct message (Channel delivers only to Peer)
+- `#topic` → tags for queue interests
+
+## Cost
+
+Prefer `max_rounds: 2–3`. Set `token_budget` for hard caps. 3×3 is usually enough.
+
+## Result fields
+
+1. `errors` — surface failures
+2. `agents_done` / `rounds_run` / `stop_reason`
+3. `transcript` — synthesize the user answer from this
+4. `backend.endpoint_source` / `default_model` — verify provider wiring
+
+## Example
+
+```json
+{
+  "agents": [
+    {"name": "alpha", "system_prompt": "Argue the case for. Brief."},
+    {"name": "beta",  "system_prompt": "Argue the case against. Brief."},
+    {"name": "gamma", "system_prompt": "Synthesize. Brief."}
+  ],
+  "task": "Should we migrate internal services to gRPC this quarter?",
+  "mode": "parallel",
+  "max_rounds": 3,
+  "token_budget": 12000
+}
+```
+
+## Anti-patterns
+
+1. Calling WoT for everything
+2. Per-agent `model` in tool args (stripped at tool boundary)
+3. Long ceremonial prompts
+4. Ignoring `errors`
+5. Treating multi-agent output as manufactured truth