Repository navigation
Releases
bstack ALL v0.2.0-rc.1
bstack ALL v0.2.0-rc.1
The first tagged public prerelease of ALL, the Agent Learning Loop. ALL connects significant coding-agent corrections to evidence-backed safeguards that you can review, verify against broken and corrected code, and approve for your project.
What's included
- Local, bounded observation capture and semantic handoff through your existing coding-agent session; no separate API key.
- One review command with two explicit decisions: authorise the exact verification plan, then approve the exact verified contents for application.
- Small reviewed existing-file changes with original hashes, conflict checks and preserved user configuration.
- Separate proposed, verified, applied, integrated and subsequently observed safeguard stages.
- Relevant safeguard delivery as tasks change, and recorded detection of recognised later failures.
- The independent
bstack-allexecutable and optionalbstack allnamespace dispatcher.
Install
Requires Node.js 24.13 or later, npm and Git. macOS is the tested primary platform.
Download both release assets into the same directory:
shasum -a 256 -c SHA256SUMS
npm install --global ./bstack-all-0.2.0-rc.1.tgz
bstack-all --version
cd /path/to/your/project
bstack-all init --agents claude-code --dry-run
bstack-all init --agents claude-code
bstack-all doctorRestart your agent, review its normal host trust prompts, and continue coding. Use bstack-all review when a safeguard is ready for review. The package is published as a GitHub release asset; it is not published to the npm registry.
For an existing 0.1 installation, read the migration instructions. Re-run init for each enrolled worktree after upgrading; schema migration is transactional, but 0.1 cannot open the upgraded database.
Verification evidence
The release checks cover 65 software tests, fourteen deterministic replay scenarios, package installation/uninstallation and an actual 0.1-to-candidate upgrade. The full controlled live Claude CLI trial captured an ordinary correction, generated a boundary test, verified actual broken/fixed commits after user authorisation, applied the separately approved contents, and detected a later deliberately reintroduced recurrence. The source was restored afterward.
These are separate evidence levels. The controlled trial is not a measurement of everyday productivity, long-term failure reduction or semantic classification accuracy. Detailed release report · Evaluation methodology
Known limitations
- Trusted Codex native dispatch and both official VS Code integrations remain blocked on host/UI trust verification. Installing configuration does not establish live compatibility. Host evidence
- Normal-check attribution currently supports recognised npm/Node command shapes. Unrecognised compound shell commands may not produce enforcement receipts; other language/CI runners need adapter work.
- Correction detection has known misses, particularly indirect and non-English corrections.
- Default verification executes repository code with your OS permissions. Git snapshots are not a security sandbox. The optional experimental macOS provider allows outside reads. Security boundaries
This is a release candidate for early adopters. Report reproducible problems through GitHub Issues, with private project content removed.