Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

docs(agent-harness): clarify completion delivery ids

Steven Enamakel39e6040f1327
diff --git a/docs/gitbooks/en/developing/architecture/agent-harness.md b/docs/gitbooks/en/developing/architecture/agent-harness.md
index c8b76377e9..0bbb0cec2a 100644
--- a/docs/gitbooks/en/developing/architecture/agent-harness.md
+++ b/docs/gitbooks/en/developing/architecture/agent-harness.md
@@ -439,7 +439,7 @@ neutral TinyAgents lifecycle:
The model-facing orchestration surface has one child tool: `spawn_subagent`. It always returns asynchronously and accepts `agent_id` to select a specialist. A matching `task_key` reuses the compatible durable child session; when that child is paused on clarification, another spawn with the same key resumes its checkpoint and preserves the original task identity. The child result is delivered to the parent chat when a delivery thread exists. Headless callers remain asynchronous and record completion in the run log.
-The completion router deduplicates by task and holds a claimed batch until the delivery turn has been persisted. Failed delivery releases the batch for retry, and startup recovery drains completions that were not delivered before a restart. The OpenHuman host debounces nearby completions into one system turn and persists that reply under a deterministic run id before it publishes `chat_done`; the frontend can reuse that id without appending a duplicate row. If retries are exhausted, the host writes an undelivered notice into the thread. The completion store uses the caller's profile-scoped storage backend. A `workflow_proposal` in a child's history is also persisted as a parent-thread message so the frontend can restore the proposal after reload.
+The completion router deduplicates by task and holds a claimed batch until the delivery turn has been persisted. Failed delivery releases the batch for retry, and startup recovery drains completions that were not delivered before a restart. The OpenHuman host debounces nearby completions into one system turn and persists that reply under the deterministic `agent:<run_id>` message id, with `extraMetadata.requestId = run_id`, before publishing `chat_done`; the frontend reuses that id without appending a duplicate row. If retries are exhausted, the host writes an undelivered notice into the thread. The completion store uses the caller's profile-scoped storage backend. Interactive replies follow the same persist-before-`chat_done` rule under `agent:<request_id>`. Segmented delivery is the exception: the client owns one row per segment, so the core stores none. A `workflow_proposal` in a child's history is also persisted as a parent-thread message so the frontend can restore the proposal after reload.
### Spawn hierarchy and tiers