Stabilize live chat tool loop

docs/how_to_run.md

@@ -16,6 +16,10 @@ cp .env.example .env
 
 The default `DUCK_MAIN_MODEL_PATH` points to `./models/Qwen3.6/nonMTP/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf`.
 
+`DUCK_ENABLE_REFLECTION=0` is the recommended default for the local single-slot
+stack. Set it to `1` only when you explicitly want critic reflection after each
+chat and accept that it can slow down the next request.
+
 3. Start DuckLM:
 
 ```bash

docs/web_api.md

@@ -51,6 +51,9 @@ Structured utility roles are validated locally before side effects:
 
 - `action` output must match `duck_core/schemas/action_directive.schema.json`; invalid directives are logged as `action_directive_failed` and no tool runs.
 - `memory_policy` output must match its JSON schema; invalid decisions fall back to `should_store=false`.
+- Large tool outputs are compacted before being sent back to the model; full outputs remain in task events and command audit.
+- Duplicate identical tool actions in one task are skipped and logged as `tool_call_skipped`.
+- Critic reflection is controlled by `DUCK_ENABLE_REFLECTION`; the default is off for responsive single-slot local chat.
 
 Chat requests accept optional `reasoning`: