[Pepps233]

Summary

Terminal-Bench evaluations now use the supplied fixed system prompt without the automatic <project_context> section.

The task user message ends with Do not cheat by using online solutions or hints specific to this task.

Reasoning

The pinned Diffex 0.3.1 CLI already supports --system-prompt and --no-context-files, so this change stays in the eval integration and requires no product release.

The pinned CLI does not propagate those flags to child agents; eval preflight therefore requires disabled subagents.

Changes

• Preserve the original instruction bytes and append the anti-cheating sentence after the native file wrapper in the same user message.
• Check in the exact eval prompt, suppress automatic context loading, and validate requested and observed prompt hashes.
• Require disabled subagents and update both controlled and full eval configurations.
• Update smoke tests, evidence checks, and documentation.

Testing

• All 188 Python tests passed; the seven execution-contract tests passed again after adding full-config coverage.
• node --test tests/launcher.test.cjs and all seven Linux cleanup regressions passed.
• The pinned-release Linux arm64 execution smoke passed three trials and nine loopback provider requests, including repeated root runs and UID 65534.
• Every captured provider request had the exact system prompt, no project context, and the anti-cheating sentence only in the user message.
• npm run check and the repository's isolated ./test.sh passed.

Smoke evidence: results/diffex-execution-b57c6f5b84a4/smoke.json.
No paid model calls or full benchmark rerun were performed.