guard-shared-runtimelisted

# guard-shared-runtime — protect the shared environment The local runtime (Python env, vllm-mlx server, config files) is shared across sessions and dispatches. Blind upgrades or unchecked failures can break subsequent work. ## Check-don't-blind-upgrade Before installing or upgrading any dependency in the shared virtual environment: 1. **Dry-run:** Simulate the change to detect transitive churn. `--dry-run` resolves and reports what it *would* install without touching the environment; it needs **pip >= 22.2** (check with `pip --version`). ```bash pip install --dry-run <package> ``` Read the resolved set, not just the headline: a dependency being *downgraded* or *added* is the churn you are looking for. 2. **Verify stability:** Ensure no critical packages are downgraded or removed. 3. **Defer if risky:** If the dry-run shows significant churn, defer the upgrade to a supervised moment or use an isolated environment for that specific task. ## Verify fork patches survived reinstall If you maintain local fork patches (e.g., for vllm-mlx or Claude Code hooks): 1. **Check integrity:** After any reinstall or update, verify that your patches are still applied. 2. **Run diagnostics:** Use the plugin’s diagnostic harnesses (e.g., `direct-route-acceptance.py`) to confirm the server behaves as expected. ## Wedged single-slot servers A single-slot server (like vllm-mlx) can become **wedged**: the health endpoint returns `200 OK`, but generation is permanently stuck. Th