Decision Machine — Settings

OPERATOR ⌂ Home Changes → Cockpit →

Operator-facing runtime, governance and connectivity settings. All current values are fetched live at page load. This page never authors or edits business rules or governed policy — those route through the change process. Help text for each setting is hidden by default; click ? to reveal it.

Enter your token, then Connect

1. Context and Execute

1a. Context allowlist
Files the developer role can read when authoring a patch during Execute. If Execute is HELD with a missing file error, add that file here and retry. Removing a file will cause Execute to block on any CR that needs to modify it.
  • Connect to load.
1b. Per-file context cap
Maximum characters read from any single file in the developer context. Files larger than this are truncated. If the developer blocks because a file appears incomplete, raise a CR to increase this cap. Higher values increase Execute cost and latency. api/main.py is exempt from this cap.
value 64,000 chars
1c. Total context cap
Maximum combined characters across all files in the developer Execute context. If the total exceeds this limit, lower-priority files are dropped.
value 460,000 chars
1d. Always-included lifecycle files
These files are always included in Execute context regardless of the allowlist. They define lifecycle and deploy infrastructure. Changes require a CR.

    2. Operational Controls

    2a. System mode
    RUN allows new submissions and council rounds. HOLD blocks them immediately without affecting in-flight rounds. Use HOLD before maintenance or when a critical finding requires investigation. Pre-authorized — no CR required.
    2b. Gate state
    The gate controls whether real-money execution is permitted. Emergency Stop halts downstream execution immediately. Disarm returns the gate to safe default. Both are pre-authorized — no CR required. Re-arming requires a new authorization via the governed gate process.
    2c. Execution mode
    dry_run runs the full pipeline without placing real orders. real_money places live orders. Switching is blocked while the gate is armed. Disarm the gate first to switch.

    3. Config Governance

    3a. Config fingerprint
    The engine computes a fingerprint of its runtime configuration and compares it to the last operator-approved fingerprint. If they differ, a config_drift alert fires. This is normal after any deployment that changes config. Click Approve after verifying the deployment is intentional — this clears the alert. Never approve if you did not expect a config change.
    current
    approved

    4. Connectivity

    4a. Apply and Deploy webhook URL
    The URL the engine calls when the operator clicks Apply and Deploy in changes.html. Points to dm-webhook.service on the host. Change only if the webhook service moves to a different host or port. An incorrect URL silently breaks Apply and Deploy.
    4b. AI API endpoint
    The upstream AI API endpoint used by all role calls. Change requires a redeploy and a CR.
    endpoint https://api.perplexity.ai/v1/responses
    4c. Execute timeout
    Maximum seconds the engine waits for a single Execute role call. If Execute is consistently timing out on large changes, raise a CR to increase this.
    value 600s

    5. Model Assignments

    5a. Per-role model assignments
    Each AI role has a model assignment. Pinned roles halt the engine if their assigned model is unavailable. Changing model assignments requires a CR because incorrect changes can silently degrade output quality or halt execution.

    Connect to load.