Build Verification and Automatic Repair

After a mutating tool succeeds in a project task, the Agent runs BuildVerifier before completing its answer. A failed verification appends the command and error output through the builtin/repair prompt in the same Session and continues within the configured repair limit.

Flow

flowchart TD
    A["Mutating tool succeeds"] --> B{"Project task with AutoVerify?"}
    B -- No --> Z["Complete turn"]
    B -- Yes --> C["Resolve verification command"]
    C --> D["Run with a 10-minute limit"]
    D --> E{"Exit code is zero?"}
    E -- Yes --> Z
    E -- No --> F{"Below MaxRepairAttempts?"}
    F -- Yes --> G["Add command and errors through repair prompt"]
    G --> A
    F -- No --> Z

Global tasks never run build verification. plan and readonly do not expose mutating tools and therefore cannot trigger it.

Default commands

Python and Unity currently have no built-in default command. Configure a workspace override when verification is required.

Configuration

Global ~/.seekclaw/config.json:

{
  "agent": {
    "autoVerify": true,
    "maxRepairAttempts": 3
  }
}

Workspace .seekclaw/config.json:

{
  "autoVerify": true,
  "verifyCommand": "dotnet test"
}

verifyCommand is a workspace field, not a global agent.verificationCommand. Commands run from the workspace root. Windows prefers Bash, then PowerShell, and finally cmd.exe; other systems use /bin/bash -c.

Standard output and error are combined. When output exceeds 8,000 characters, the verifier keeps the end because compiler errors usually appear there. Verification events are rendered in Desktop or CLI, and reaching the repair limit stops the loop while preserving the actual failure details.