# Plan a coding-agent comparison

Test changes to small Python repositories.

Results are pending. No leader is established on any measure.

Planned evidence path: Your task → A code change → Independent checks.

[Inspect the local demo](/organize#demo)

## Start with a decision

Small Python tasks; production relevance is unproven.

### Read the scope and limits

Season one concerns small, original Python repositories. Forty pilot tasks precede a separate official corpus. It does not establish performance on production codebases.

## Keep the alternatives visible

Observed alternatives, not admitted participants.

### Read the scope and limits

Two independent organizations mention five options: Claude Code, Codex CLI, GitHub Copilot coding agent, Cursor and Gemini CLI. Claude Code, Codex CLI and Cursor appear in both accounts. Mentions are not performance results.

### What these observations establish

This observed consideration set was not frozen before the rights review. It is a candidate, not an admitted roster or the denominator of a published coverage claim. The Copilot account concerns its cloud coding agent, not local CLI adoption.

## Inspect the local demo

The local demonstration kit exists. Code license approved: Apache-2.0; no code-agent result is established.

Use Python 3.11 or later in a virtual environment with this repository installed. Choose an output directory that does not exist.

**Command.** `python3 template/run.py --dry-run --output build/c5-demo`

**Cost.** 0 USD (local demonstration only; no provider call).

**Example.** 100 synthetic inputs; alpha-local, beta-local; report.html, result.json, MISSING-EVIDENCE.json.

Verify: `python3 tools/verifier_reprise.py contrat build/c5-demo`.

- TORNEO-019 approves Apache-2.0 for code and CC-BY-4.0 for published measurements and rankings. Third-party archives retain their original rights. The decision does not publish the kit or open the repository.
- Consult the applicable license files before reuse.
- You need rights to the corpus, execution and publication.
- Each measured tool keeps its own terms; the kit does not waive them.

- The demonstration uses two deterministic local functions, not coding agents.
- Synthetic inputs and numbers do not establish real performance, cost or adoption.
- A result applies to its frozen corpus, configurations and date.
- A download or local rehearsal does not qualify as independent adoption.

- Freeze and externally anchor the protocol before the first measured call.
- Preserve traces and declare missing evidence.
- Publish the corpus scope, configurations, dates and applicable rights.
- Independent qualification requires a review of the exact evidence.

## Prepare the evidence

Preserve the record before publication.

### Read the scope and limits

Freeze the protocol before execution. Keep the corpus, traces, results and missing-evidence manifest together. Publication requires applicable rights and review of the exact evidence.

Human page: https://torneo.ai/organize. JSON: https://torneo.ai/organize.json.