Unattended harness recipes

dvalincode run is the bounded primitive for schedulers and CI. It never opens an approval prompt, applies the org policy's unattended ceiling when stdin is not a TTY (or when --unattended is explicit), and leaves one audit chain per turn. Use --report to persist a human-readable evidence artifact beside the harness logs.

Recommended machine or repository policy:

{
  "unattended": {
    "maxPermissionMode": "auto",
    "maxIterations": 40,
    "maxWallMinutes": 30
  }
}

An explicitly requested permission mode, iteration limit, or timeout that exceeds this block fails with exit code 3. Organization provider, model, path, command, tool, MCP, and network policy continues to apply inside the run.

Nightly dependency maintenance with cron

Keep the prompt in a reviewed file, create one timestamped report per attempt, and let cron capture stdout/stderr separately:

15 2 * * * cd /srv/acme-app && mkdir -p reports && dvalincode run --prompt-file ops/nightly-deps.md --permission-mode auto --report "reports/deps-$(date +\%s).md" --output-format json >> reports/deps.jsonl 2>> reports/deps.log

Example ops/nightly-deps.md:

Check for patch-level dependency updates. Apply only compatible patch updates,
run the focused test suite, and stop without changing files if tests fail.

CI post-merge fixup

The JSON result is emitted even on failure, so the job can archive it and the report regardless of the exit code:

set +e
dvalincode run \
  --prompt-file .ci/post-merge-fixup.md \
  --permission-mode auto \
  --timeout 20 \
  --report artifacts/dvalin-report.md \
  --output-format json > artifacts/dvalin-result.json
status=$?
set -e

test "$status" -eq 0

Treat exit 2 as bad harness input, exit 3 as a policy mismatch, and exit 4 as a timeout or interrupt. Exit 1 represents an agent, tool, or provider failure.

Claude Code /loop delegation

Register the task-level stdio server:

{
  "mcpServers": {
    "dvalincode": {
      "command": "dvalincode",
      "args": ["mcp-serve", "--workspace", "/srv/acme-app", "--max-permission-mode", "auto"]
    }
  }
}

Then instruct the loop to call dvalin_run_task for each code-changing iteration and dvalin_get_evidence afterward. The external agent owns the schedule; DvalinCode retains the internal model/tool loop and policy chokepoint. dvalin_run_task is synchronous, so configure a caller timeout longer than its timeout_minutes input.