* feat: add isolated pure-inference completion * fix(google): block ambient system prompt writes * docs: refresh generated map
7.1 KiB
summary, read_when, title
| summary | read_when | title | ||
|---|---|---|---|---|
| JSON-only LLM tasks for workflows (optional plugin tool) |
|
LLM task |
llm-task is a bundled optional plugin tool that runs a single JSON-only
LLM call and returns structured output, optionally validated against a JSON
Schema. It gives workflow engines like Lobster an LLM step without custom
OpenClaw code per workflow.
Enable
- Enable the plugin:
{
"plugins": {
"entries": {
"llm-task": { "enabled": true }
}
}
}
- Allow the tool:
{
"tools": {
"alsoAllow": ["llm-task"]
}
}
alsoAllow adds llm-task on top of the active tool profile without
restricting other core tools. Use tools.allow only if you want a restrictive
allowlist mode instead.
Config (optional)
{
"plugins": {
"entries": {
"llm-task": {
"enabled": true,
"llm": {
"allowModelOverride": true,
"allowedCompletionModels": ["openai/gpt-5.6-sol"],
"allowAuthProfileOverride": true
},
"config": {
"defaultProvider": "openai",
"defaultModel": "gpt-5.6-sol",
"defaultAuthProfileId": "main",
"maxTokens": 800,
"timeoutMs": 30000
}
}
}
}
}
The llm block is host-owned authorization. allowedCompletionModels restricts every
completion, so include the resolved agent default as well as any override targets.
allowAuthProfileOverride permits defaultAuthProfileId and the per-call
authProfileId parameter. The config keys are selection defaults used when a
tool call omits the corresponding parameter.
Run openclaw doctor --fix once for llm-task entries created by older releases.
Doctor grants the shipped model/profile selection permissions and moves any
legacy config.allowedModels value into llm.allowedCompletionModels without widening it.
Tool parameters
| Parameter | Type | Notes |
|---|---|---|
prompt |
string | Required. Task instruction for the LLM. |
input |
any | Optional payload; serialized to JSON and appended to the prompt. |
schema |
object | Optional JSON Schema the parsed output must validate against. |
provider |
string | Overrides defaultProvider / the agent's default provider. |
model |
string | Overrides defaultModel; accepts bare model ids, aliases, or a provider/model ref (a duplicate provider prefix is stripped automatically). |
thinking |
string | Reasoning level (e.g. low, medium); must be one supported by the resolved model. |
authProfileId |
string | Overrides defaultAuthProfileId. |
temperature |
number | Best-effort; not all providers honor it. |
maxTokens |
number | Best-effort cap on output tokens. |
timeoutMs |
number | Run timeout; default 30000. |
Output
Returns details.json (the parsed, schema-validated JSON) plus details.provider
and details.model naming what actually ran.
Each call starts a fresh prompt-only inference operation. It does not reuse the calling agent's transcript or native runtime session, run agent lifecycle hooks, or deliver model output to a channel. OpenClaw uses the selected provider, model, auth profile, and runtime exactly once; it does not fall back to another route when that owner cannot provide a literal zero-tool call.
A selected agent harness must implement isolated completion. Otherwise the call
fails before inference with a does not support isolated completion error.
This fail-closed behavior prevents a JSON task from silently becoming a normal
tool-capable agent turn.
CLI runtimes must provide the equivalent isolated preparation guarantee. The bundled Claude and Gemini CLI runtimes do; a different CLI runtime that has not adopted this internal contract fails before its process starts.
Gemini CLI isolated completion supports Gemini API-key and Vertex auth. Google
OAuth and compute/Code Assist auth are rejected because managed-account policy
can add administrator-required tools after local CLI settings are loaded.
Gemini prompts containing native @path includes or a leading /command also
fail before inference because Gemini CLI has no literal raw-input mode.
Example: Lobster workflow step
Important limitation
The example below assumes the standalone Lobster CLI is running where
openclaw.invoke already has the correct gateway URL/auth context.
For the bundled embedded Lobster runner inside OpenClaw, this nested CLI pattern is not currently reliable:
openclaw.invoke --tool llm-task --action json --args-json '{ ... }'
Until embedded Lobster has a supported bridge for this flow, prefer either:
- direct
llm-tasktool calls outside Lobster, or - Lobster steps that do not rely on nested
openclaw.invokecalls.
Standalone Lobster CLI example:
openclaw.invoke --tool llm-task --action json --args-json '{
"prompt": "Given the input email, return intent and draft.",
"thinking": "low",
"input": {
"subject": "Hello",
"body": "Can you help?"
},
"schema": {
"type": "object",
"properties": {
"intent": { "type": "string" },
"draft": { "type": "string" }
},
"required": ["intent", "draft"],
"additionalProperties": false
}
}'
Safety notes
- JSON-only: the model is instructed to return only a JSON value, no code fences, no commentary.
- No tools: the selected runtime must expose a literal empty model-callable tool surface. OpenClaw rejects tool-shaped results instead of treating them as task output.
- Isolated: the run has no agent transcript, session reuse, lifecycle hooks, channel delivery, or provider fallback.
- Treat output as untrusted unless you validate it with
schema. - Put approvals before any side-effecting step (send, post, exec) that consumes this output.