All automated tests, CI, synthetic scenarios, smoke checks, benchmarks, grading, calibration and retries use prepared data and mocked transports. Paid token/API testing is forbidden, including opt-ins, ambient credentials, subscription proxies, alternate keys/models and automatic fixture recording. Never load .secrets into a test runner.
The engine contains no provider/network transport. Unit tests deny socket connections and permit only deterministic Git subprocesses. Native host commands used for packaging validation must be read-only --help or plugin manifest validation, never inference/eval prompts.
Run:
PYTHONPATH=src python3 -m unittest discover -s tests -v
python3 scripts/build_bundle.py --check
python3 scripts/check_distribution.pyThe fixture intentionally includes a stale README, dead implementation, setting persistence/cache invalidation, frontend requests and an absent HTTP binding. Prepared expectations assert only what that fixture supports. Tests cover evidence forgery, invalid corroboration, stale snapshots, relation endpoints, explicit unknowns, contradiction retention, note preservation, symlink/secret exclusions, scan budgets, Git renames, worktree isolation and CLI exit behavior.
Prepared responses are contract evidence only. They do not establish provider availability, model quality, true production usage, settlement, live latency or launch qualification. Failed evidence must be preserved; do not regenerate expected values to conceal a regression.
Live-content dogfood is a separate manual activity requiring exact real-content scope, an explicit hard monetary/request cap and a stop condition before any call. This repository's implementation/release authorization does not authorize it. No live dogfood was used for this release.