Decision models

A decision model answers typed questions about an event instead of generating text. An SOP can ask one, after its trigger matches and before a run starts:

• Gate: should this event start this SOP at all?
• Mode: which of the execution modes the author listed fits this run?

This lets one trigger (a webhook path, a channel, an MQTT topic) feed several SOPs, with the model deciding which of them the event is really for and how much human supervision each run gets.

ZeroClaw speaks the “System One” typed-decision API (POST <base_url>/v1/systemone). Two models implement it:

Configure models

Declare each model under an alias in config.toml and pick its provider:

[decision_models.jev]
provider = "jev"
# api_key is a secret; set it with `zeroclaw config` or
# ZEROCLAW_decision_models__jev__api_key.

[decision_models.laya]
provider = "laya"

[decision_models.gpu_box]
provider = "custom"
base_url = "http://10.0.0.5:8000"
model = "laya"

base_url and model override the provider defaults. A custom entry without a base_url is ignored, and SOPs selecting it fall back to their strictest mode. The aliases can also be managed from the dashboard’s Config page.

Select a model in an SOP

In the web dashboard’s SOP editor, tick Decision model under Execution mode and pick a configured model from the dropdown; the gate, modes, and thresholds are edited in the same panel. Saving rejects an invalid [decision] table with the reason.

In a file-backed SOP, add a [decision] table to SOP.toml and name the alias:

[decision]
model = "jev"
gate = "Is the customer asking for money back?"
gate_threshold = 0.7
modes = ["auto", "supervised", "step_by_step"]
mode_instructions = "auto: under $50 and routine. step_by_step: over $500, legal threats, or anything unusual."
min_confidence = 0.7

A [decision] table needs a gate, modes, or a step with decide. modes cannot be used with a deterministic SOP.

Conditional steps

A step can carry its own yes/no question. The model answers it in the same request as the gate and mode, so one call decides which steps a run includes. This composes one SOP from parts instead of splitting it into several SOPs that each ask the model whether to start.

## Steps

1. **Hold for a maintainer** - Draft a comment that tags a maintainer to decide on scope.
   - decide: Does this PR add a new feature outside the roadmap and existing features?

2. **Security review** - Review the auth, secret, and sandbox changes.
   - decide: Does this PR touch authentication, secrets, or sandboxing?
   - unless_decided: 1

3. **Test review** - Check that the changed behavior is tested.
   - decide: Does this PR change runtime behavior?
   - unless_decided: 1

4. **Compose the review** - Combine the findings into one review comment.
   - unless_decided: 1

• - decide: <question> runs the step only when the answer’s “yes” probability reaches part_threshold.
• - unless_decided: N skips the step when step N’s question was answered yes. Step N must have a decide question.
• A skipped step is recorded with status skipped and the reason, and the run continues with the next step in order. The next step receives the input the skipped step would have received.
• When the model cannot answer, or answers a step’s question with a malformed value, that step runs. Skipping is never the fail-safe.
• Routing guards can read the answers: $.decisions.2 is step 2’s “yes” probability in a when: condition.
• Another step cannot depends_on a conditional step, because a skipped step produces no output. Read $.steps.N in a when: condition instead.
• A step is skipped before its approval gate, so a skipped step never asks for approval.

Safety

• The model only chooses among what the author wrote. It cannot add triggers, pick an unlisted mode, or select priority_based or deterministic.
• A chosen mode applies to that run only and is stored with the run, so it survives a restart. Step-level requires_confirmation and checkpoints still gate regardless of the chosen mode.
• Fail-closed: an unconfigured alias, a network or HTTP error, a malformed answer, or a low-confidence answer never reduces supervision. The run uses the strictest of its listed modes, its authored execution_mode, and supervised.
• Event payloads are sent to the model framed as untrusted content, capped at 8,000 characters. Choose a self-hosted model when event content must not leave the host.

Each decision is logged with the SOP name, outcome, chosen mode, each conditional step’s answer, and input tokens.