ALL turns significant coding-agent corrections into verified project safeguards.
Keep coding in your existing agent. Review the lesson, verify the proposed check, and decide what becomes part of your project.
Download prerelease · Get started · See a real example
Local capture · Your existing coding-agent session · Explicit approval
You fix an edge case, explain a requirement again, or restore behaviour that worked yesterday. The current session moves on. A later session can still make the same mistake.
ALL — Agent Learning Loop — connects that correction to the original requirement, the failure, and a proposed safeguard. A regression test is one possible result. Small validation and verification changes are also supported when they can be reviewed and checked safely.
The goal is a useful improvement in your repository, with evidence that it catches the original failure. You decide whether it belongs there.
Current status: 0.2.0-rc.1 prerelease. A complete controlled lifecycle has run through the actual Claude Code CLI, including a later recurrence caught by the applied test. Trusted Codex hook dispatch and both official VS Code integrations remain unverified and blocked on host/UI trust checks. macOS is the tested platform. Compatibility · Release evidence
flowchart LR
A[Code and correct normally] --> B[ALL captures a candidate]
B --> C[Your agent examines evidence]
C --> D[Review a proposed safeguard]
D --> E[Authorise broken/fixed verification]
E --> F[Approve exact contents and apply]
F --> G[Run through normal project checks]
1 · Code normally. Initialise ALL once in a repository and use your coding agent as usual. Selected corrections and verification failures can become local candidates.
2 · Examine the evidence. At a supported boundary, ALL asks your existing agent to inspect a bounded batch. A candidate can be dismissed, investigated, or confirmed against an established requirement. Capturing a correction does not automatically create a rule.
3 · Review and verify. Run bstack-all review. Inspect the evidence, exact changes, commands and execution risks. Authorise verification against the actual broken and corrected states.
4 · Keep the safeguard. After successful verification, separately approve the exact contents for application. ALL checks whether the safeguard is integrated into the project's normal verification and can record later native check observations.
No separate API key is required. Agent-assisted analysis uses your existing coding-agent session and can consume its usage allowance.
A shipping quote must charge 500 cents below 10000 cents and give free shipping at exactly 10000 cents or above. A comparison changes from >= to >, and the exact threshold breaks.
The ordinary correction is specific:
The previous version worked. I explicitly required free shipping at exactly 10000 cents; the quote now charges shipping at that boundary.
In the live Claude trial, ALL captured that correction, requested analysis, and the agent generated a test covering both sides of the boundary:
assert.equal(quote(9999).shippingCents, 500);
assert.equal(quote(10000).shippingCents, 0, 'FREE_SHIPPING_AT_EXACT_THRESHOLD');
assert.equal(quote(10000).totalCents, 10000, 'FREE_SHIPPING_AT_EXACT_THRESHOLD');
assert.equal(quote(10001).shippingCents, 0);After the user's separate verification and application decisions, the generated test:
- Failed on the actual broken commit and passed on the corrected commit.
- Ran through the existing
npm run testcommand without configuration or dependency changes. - Passed in a fresh Claude session, producing a native observation receipt.
- Detected a deliberately reintroduced regression in a later session. The product source was then restored.
This was one controlled live trial. It demonstrates detection of that recurrence; long-term failure reduction and productivity gains have not been measured. Full trial and limitations
Requires Node.js 24.13 or later, npm and Git. The package is distributed through GitHub Releases and is not yet published to the npm registry.
Download bstack-all-0.2.0-rc.1.tgz and SHA256SUMS from the release page, or run:
curl -fLO https://github.com/bstack-tools/all-agent-learning-loop/releases/download/v0.2.0-rc.1/bstack-all-0.2.0-rc.1.tgz
curl -fLO https://github.com/bstack-tools/all-agent-learning-loop/releases/download/v0.2.0-rc.1/SHA256SUMS
shasum -a 256 -c SHA256SUMS
npm install --global ./bstack-all-0.2.0-rc.1.tgz
bstack-all --versiongit clone https://github.com/bstack-tools/all-agent-learning-loop.git
cd all-agent-learning-loop
git checkout v0.2.0-rc.1
npm ci --ignore-scripts
npm run build
npm pack
npm install --global ./bstack-all-0.2.0-rc.1.tgz
bstack-all --versionThe executable is bstack-all. ALL is an independent bstack product. An optional ownership-safe dispatcher enables bstack all; it is not required. Namespace setup and migration
cd /path/to/your/project
bstack-all init --agents claude-code --dry-run
bstack-all init --agents claude-code
bstack-all doctorInit merges project hook definitions and installs two focused skills while preserving unrelated configuration. It does not edit AGENTS.md or CLAUDE.md. Restart your coding-agent session and review the host's normal trust prompts.
--agents codex installs the Codex adapter; its trusted live dispatch remains unverified. --agents manual supports an explicit manual workflow. Host setup and verification steps
Use ordinary coding and correction prompts. ALL's capture is local and bounded; supported host boundaries request semantic analysis through the current agent. Unsupported boundaries leave work pending. You can inspect it with:
bstack-all status
bstack-all review --jsonbstack-all reviewIn an interactive terminal, you make two separate decisions:
| Decision | What it authorises |
|---|---|
Type verify |
Execute the displayed immutable plan against broken and corrected snapshots. |
After successful results, type approve |
Approve the exact verified contents and apply them. |
Empty answers deny. Piped input cannot grant interactive consent. There is no --yes bypass. Low-level commands remain available for agent-mediated review and diagnostics with the same explicit user authorisation. Detailed workflow
A file on disk does not establish that a test runs. ALL distinguishes:
Proposed → Verified → Applied → Integrated → Observed in a later check
Normal-check integration requires a matching declaration and evidence that the normal command detects the broken state and passes the correction. The current integration adapter supports declared npm test/check/verify scripts and supported Node test discovery. Other runners need adapter work. Unrecognised compound shell commands may not produce observation receipts.
bstack-all integration --file path/to/regression.test.mjs --json
bstack-all metrics
bstack-all audit- Local capture. No telemetry or model call occurs in deterministic capture. Private state lives outside Git under
${XDG_DATA_HOME:-$HOME/.local/share}/bstack-allby default. - Existing agent account. Semantic work follows your host agent's provider and permission model. It is not necessarily local-model execution.
- Reviewed execution. Default verification runs repository code with your OS permissions. Temporary Git snapshots isolate files; they are not a security sandbox.
- Controlled changes. Existing-file edits require exact original hashes and reviewed diffs. Conflicting work is refused. Broad automatic refactoring and dependency changes are excluded.
The optional experimental macOS Seatbelt provider restricts tested network and outside-write operations; outside reads remain allowed. See privacy and execution boundaries before authorising verification.
bstack-all pause
bstack-all resume
bstack-all privacy
bstack-all uninstallUninstall removes owned integration, preserves user changes, and retains approved project safeguards and local history. Data deletion is a separate explicit operation.
ALL is MIT licensed, written in strict TypeScript, and uses Node's built-in SQLite. Zod is its sole npm runtime dependency. It has no daemon or separate database service.
npm ci --ignore-scripts
npm run verifyBuild and test · Contribute · Changelog · Security policy · Report a bug
Architecture · Evaluation · Troubleshooting · Third-party notices