Holon

Open-source AI assistants on your own computer or server. Assign tasks, track progress, and let them follow up on ongoing work.
Skip to content
Browse docs
On this page

Configure advisory decisions

AdvisoryDecision asks for a non-authoritative second opinion between explicit options. Its result cannot grant permission or approve an operation. For the routing model, see Models and providers.

Prerequisites

  • A running daemon and a working primary model; see Configure models.
  • Permission to change shared Decision settings.
  • For local inference, a build with the local-onnx feature and verified model files. Downloading a preset needs network access; local inference does not send decision prompts to a remote provider.
  • For remote inference, credentials and access to a route explicitly declaring Decision capability and a supported decision_protocol.

Step 1: Choose a Decision Provider

Inspect the existing selection before changing it:

holon config get decision.enabled
holon config get decision.model
holon config get decision.local_onnx.enabled
holon config models list

Choose one of the following paths. A remote decision.model and enabled local ONNX cannot be configured together. Ordinary chat capability alone is not sufficient for a Decision route.

Step 2: Configure the Provider

Option A: Use the Local ONNX Provider

In the Web GUI, open Settings → Decision Settings, choose the jev-selector-q4f16 preset, and click Download and verify preset. Wait until it reports ready before enabling local inference. For a pre-provisioned model, use decision.local_onnx.model_dir instead; the files must match the supported model/manifest format.

Then set the local route:

holon config unset decision.model
holon config set decision.local_onnx.preset jev-selector-q4f16
holon config set decision.local_onnx.num_threads 2
holon config set decision.local_onnx.enabled true
holon config set decision.enabled true

If switching back from a custom model directory to the managed preset, unset decision.local_onnx.model_dir so it does not override the preset location. The managed preset is stored under <HOLON_HOME>/models/decision/local-onnx/<preset>/.

Option B: Use a Remote Provider

Authenticate the provider using the model setup guide. For the built-in TypeSafe Jev route, use:

holon config set decision.local_onnx.enabled false
holon config set decision.model "typesafe@default/jev-latest"
holon config set decision.enabled true

For a custom service, configure its endpoint and credential source first. Register the service's actual model under models.catalog with capabilities.decision: true and its documented protocol: jev, cloudflare_clef, openai_decisions, or openai_compatible. Preserve existing catalog entries when editing that object. Then select the corresponding provider@endpoint/model route with decision.model.

These protocol names describe Holon's adapters; they do not establish that a particular model or endpoint is available to your account. In particular, openai_decisions uses the dedicated /v1/decisions endpoint through the API-key-based openai_responses transport, not an OAuth Codex transport. Exact metadata fields are in the Configuration reference.

Step 3: Enable the Advisory Tool and Set Guardrails

holon config set decision.tools.enabled true
holon config set decision.tools.max_calls_per_turn 3
holon config set decision.tools.min_confidence 0.65
holon config set decision.tools.timeout_ms 10000

The tool stays hidden until enabled. These settings cap calls, require a minimum confidence, and bound the tool's wait. Provider-level timeout, concurrency, and queue settings are separate; see the Decision key table.

Read configuration-change notes on stderr. If they require a restart, let active work settle and restart through your installation's normal procedure before testing. Saving a key does not prove the running daemon has applied it.

Step 4: Verify with a Test Prompt

Use an existing test agent; replace decision-test with its ID:

holon config doctor
holon run --agent decision-test "Call AdvisoryDecision once with question 'Which review step should come first?', options ['Read the diff', 'Publish a review'], and state 'The diff has not been inspected; publishing is not authorized.' Report the returned outcome and reason. Do not publish anything or change files."
holon transcript --agent decision-test --limit 30

Confirm it worked

The transcript contains an actual AdvisoryDecision call and its result. Inspect outcome, choice, confidence, abstain, reason, and provider/model metadata. A valid result can select an option, report a provider fallback, or abstain. An explicit abstention verifies the failure path; it does not prove successful inference. Record its reason rather than presenting it as a choice.

If the agent never called the tool, the test has not verified the Decision route. The recommendation remains evidence for the agent to weigh, not approval to publish or change state.

Troubleshooting

  • Tool absent: check both decision.enabled and decision.tools.enabled, the effective configuration, and the agent's allowed tools.
  • Conflicting routes: unset decision.model for local ONNX, or disable local ONNX for a remote model.
  • Missing preset or unsupported build: check the GUI download status and the build's local-onnx support. Do not assume selecting a preset downloads it.
  • Provider unavailable or rejected: check credentials, route capability, protocol, endpoint access, and any required daemon restart.
  • Abstention: inspect the reason for timeout, provider failure, call limit, invalid choice, or low confidence. Do not lower confidence limits merely to force a recommendation.

To hide the tool without deleting the provider configuration:

holon config set decision.tools.enabled false

Next Steps