Repository navigation
bstack 3.4.1
Setup shows the models you're choosing, headless Claude calls retain user authentication settings, and PR checks stop treating ordinary application code as instructions to the reviewer.
Changes
- Full setup shows the effective PR review models and effort before asking you to keep or change them. You can save and test grilling models for later while keeping work human-guided. PR visuals are a separate choice, and a model check does not start a workflow. PR #38
- Arena and Interrogate explain the result, scope, participant count, models, and extra usage before they start. Separate attempts using the same model no longer imply different model families.
- Setup asks for working headless commands before live checks and reuses saved launchers. Claude calls load user authentication and provider settings while keeping reviewer tools restricted. Errors distinguish login, launcher, model access, quota, and connection failures without telling users to change unrelated shell or certificate settings. PR #35
- PR preparation captures the model making the call. The creating agent writes the briefing and prepares it in its own session; missing hook delivery is repaired there. Development hooks now share consumer hook definitions, and publication preserves the captured identity. PR #36
- The PR safety check looks for attempts to direct the reviewing agent. Hooks, scripts, network calls, and credential references in application code remain subjects for the full review. Direct overrides in agent instructions and unreadable executable changes still pause review. Invalid evidence produces an incomplete check, not a security finding. PR #37
- The setup instructions preserve existing host selections during upgrades, including Amp, Copilot, and OpenCode. New installs still enable automatic routing; upgrades keep the saved choice.
Upgrade
Refresh the installer to 3.4.1, then run it in each checkout and start a fresh agent session. Updating a downloaded folder or installer skill alone does not update .bstack/package/. Omit --hosts, --auto, and --verification unless you intend to change those settings.
For model workflows, keep the working CLI launch commands saved in .bstack/model-clis.json and rerun the relevant live doctor after upgrading. Full setup can configure or check grilling models without starting an autonomous interview. Each autonomous run still needs an explicit request or scoped delegation.
In a BStack development checkout, run python3 -B bstack/shared/router/install.py --configs-only --apply to repair native hook configuration. Consumer projects get their hooks through the installer. Codex hook trust still needs the host's normal approval. Model attribution depends on native metadata for the preparation call; unavailable identity remains explicit.
Verification
The generated catalog, customization layers, and committed package checks passed. The six test suites completed with 361 tests and two skips; the archive checksum passed. A fresh extracted-ZIP installation passed doctor with all seven supported hosts configured. An upgrade from the published 3.4.0 ZIP preserved the selected hosts, automatic-routing and verification choices, Python interpreter, saved launcher, review settings, and grilling models.
Live checks apply to the configured CLI commands and models. They do not establish compatibility with every custom launcher or CLI version. A passing to-pr text check does not verify image generation, and package doctor does not establish native hook delivery or trust.
Install
Download bstack-3.4.1.zip and SHA256SUMS, verify the checksum, and extract the ZIP. From the extracted bstack folder, run:
python3 install.py --project /path/to/my-projectUse the same command to upgrade. Requires Python 3.11+ on macOS or Linux. To distribute an extracted release through a folder-based skill installer, select its root bstack skill and explicitly run setup. Installation guide