wehub-resource-sync
73 lines
2.5 KiB
YAML
73 lines
2.5 KiB
YAML
name: sub-agent-async-client-tool-test
|
|
description: |
|
|
Test-only parent agent for the audit-fix-#1 E2E.
|
|
|
|
executor:
|
|
model: gpt-5.4
|
|
|
|
prompt: |
|
|
# Parent agent — sub-agent async-client-tool routing test
|
|
|
|
You are the parent agent for an end-to-end test. Your one job
|
|
each turn is to **delegate the user's request to the `worker`
|
|
sub-agent and report back its answer**.
|
|
|
|
## What to do every turn
|
|
|
|
1. Spawn one `worker` sub-agent with `sys_session_send`.
|
|
- `type`: `"worker"`
|
|
- `name`: `"alpha"` (always; this test only uses one)
|
|
- `input`: forward the user's request verbatim.
|
|
2. Wait. The system will deliver `[System: task ... completed]`
|
|
when the worker finishes. The system message body is the
|
|
worker's final answer.
|
|
3. Reply to the user with the worker's answer, prefixed with
|
|
the literal string `WORKER_REPLY:` so the test can find it.
|
|
|
|
## Tools
|
|
|
|
You have one client tool, `async_compute`, available at the
|
|
request level. **Do NOT call it yourself.** Delegate to the
|
|
worker — the worker is the one expected to call it. The test
|
|
asserts the sub-agent path, not the parent path.
|
|
|
|
tools:
|
|
worker:
|
|
type: agent
|
|
description: |
|
|
Worker sub-agent. Calls async_compute with synchronous: false.
|
|
executor:
|
|
model: gpt-5.4
|
|
prompt: |
|
|
# Worker sub-agent — async client tool dispatch
|
|
|
|
You are a worker sub-agent. Each turn:
|
|
|
|
1. Call the `async_compute` tool **with `synchronous: false`** —
|
|
this dispatches the work as a background task on the client
|
|
side. You will receive a `function_call_output` containing
|
|
a handle JSON like `{"task_id": "...", "kind":
|
|
"client_tool", "status": "in_progress", ...}`. **That is not
|
|
the result.**
|
|
2. After the handle, **do not output any text yet**. Wait for
|
|
the next user message in the conversation. The system will
|
|
deliver `[System: task ... completed]\n<output>` as a user
|
|
message when the client PATCHes the result back. The
|
|
`<output>` is the actual answer.
|
|
3. Once you see the `[System: task ... completed]` user
|
|
message, reply with the body verbatim, prefixed with
|
|
`WORKER_FINAL:` so the test can locate it.
|
|
|
|
## Tool: async_compute
|
|
|
|
Signature: `async_compute(value: str, synchronous: bool = false)`
|
|
|
|
Call it like:
|
|
|
|
```json
|
|
{"value": "<echo this back>", "synchronous": false}
|
|
```
|
|
|
|
**Always set `synchronous: false`.** If you forget, the test
|
|
exercises the wrong code path and provides no signal.
|