Headless CLI harness

Rescriptum features a dual-entry architecture: the exact same binary that renders the desktop GUI also functions as a headless automation and evaluation harness.

Invoking the Harness

You can run Rescriptum from PowerShell, bash, or CI pipelines using the harness subcommand:

``bash

``

Key Subcommands

• rescriptum harness chat --rounds : Runs a headless agent conversation session against a workspace, outputting full timeline events and tool execution traces to JSONL.
• rescriptum harness eval : Executes a standardized benchmark scenario from evals/scenarios, scoring agent performance against objective file diffs and test suites.
• rescriptum harness friction : Computes session health scores, tool failure rates, and friction penalty metrics without manual inspection.
• rescriptum harness suite: Runs the complete batch of evaluation scenarios in parallel for model bake-offs and regression testing.

Outcome-First Health Scoring

The harness scores agent sessions using an **outcome-first formula**:

• **Completed turns** that modify files and pass verification anchor in the **85–100 tier**.
• Tool iteration counts and diagnostic retries act as minor tie-breaker deductions capped at 10 points.
• This ensures that agents tackling hard, high-effort problems are rewarded for delivering green code rather than penalized for thorough exploration.