Learning

Learning saves useful lessons so later investigations can start with what you already know. Open Learning in the browser sidebar.

Lessons and suggestions

• Lessons: source-linked notes about a codebase: useful files to inspect, trust boundaries, and investigation or testing approaches. Authorized local CLI and web chats can save these with remember_codebase. Later chats on the same codebase recall matching lessons and recheck the cited file hashes before each model request. Changed files and disabled lessons are excluded.
• Suggestions: changes to workflow instructions that you can review before applying. Run evidence is shown with each suggestion.

A finished chat is history, not a lesson. Completion events remain internal records; they do not populate the Lessons view or label vulnerabilities. Lessons are hints to check against the source, not verified findings or model training. Verifiers do not receive them. ZERO_DISABLE_HUNT_MEMORY=1 disables source memory; the engine API also accepts codebaseLearning: false.

Process local experience

0 learning status --json
0 learning process --limit 25 --json

Both commands use the local control database. ZERO_DB_PATH selects an alternate database. --project <id> filters processing and status; use the same project identifier for related improvements. Selecting a local repository with 0 learning process --project /absolute/path/to/repo also imports its existing current source notes into the control database. Completed authorized local source workflows do the same, even when scan state is stored in a separate run database. The browser engine processes pending work in bounded batches every five seconds. The durable queue resumes after restart; expired work leases can be reclaimed without duplicating retained knowledge.

The default worker records activity without model calls. A trusted host adapter can derive evidence-linked diagnoses or evaluate frozen candidates. Request bodies cannot supply passing evaluation receipts.

Run evaluated evolution

0 learning evolve --config ./evolution.json --project /absolute/path/to/repo --json

This runs the existing source evolution controller and mirrors its integrity-checked snapshots, proposal references, evaluation receipts and lifecycle into Learning. Existing source-access consent, budgets, sandbox requirements and promotion settings apply. It does not independently activate a different artifact or rerun a failed model request during reconciliation.

These evaluations are labeled output-fixture. Passing expected-output tests does not by itself establish better vulnerability discovery or verified exploit success. Real security capability requires independent security benchmarks and appropriate proof requirements. Candidate generation is explicit through this command; the default background worker does not continuously spend provider credits proposing changes.

Workflow versions

Saving a definition creates an immutable revision. Each execution keeps its original definition snapshot. Restoring an older revision creates a new revision and rejects stale edits through optimistic concurrency checks.

Existing definitions and available execution snapshots populate initial history; missing historical revisions are not invented. Schedules retain their reviewed revision and require review after definition changes. A restored definition does not authorize new targets, tools or credentials.

Automatic workflow suggestions

If the same step fails in three separate runs after its instructions change, Learning can suggest the instructions from an earlier version that finished successfully. It reads the last 200 runs and events; chat errors, cancellations and ordinary completions do not create suggestions.

Suggestions keep the same target, step graph, tools, budgets, inputs and fix settings. They never run or apply themselves. Review the earlier instructions, then apply the suggestion to save a new workflow version. If someone edits the workflow or dismisses the suggestion first, applying it fails rather than replacing their changes. Existing schedules still need their own revision review. A successful earlier run is operational evidence, not proof of better security.

Isolated workspace observations

SmolVM chats can save bounded source-linked lessons on their host engine before teardown. A fresh VM requests current lessons for its approved directory before model calls. Both sides check file hashes; lessons stay out of saved chat history and do not grant tools or permissions. Disabled lessons and changed files are excluded. Host memory stores, database handles and provider credentials are not serialized into the guest.

Standard mode needs host-approved local scope. YOLO mode can use an explicitly granted workbench workspace and its subdirectories; an implicit current directory is not a host scope grant. Learning opt-out and verifier isolation still apply.

The runtime qualification report records the tested distribution and runtime limits. Complete toolbox qualification is separate: the selected image still lacks SSH and needs a rebuilt archive.

Deployment boundary

The control database, queue and knowledge stay on the selected engine. Browser requests and caches are bound to that backend. Learning introduces no automatic customer-data export or hosted evaluator. Execution sandbox qualification and model-provider residency remain separate deployment concerns.

For the wider architecture and remaining work—fresh-task security evaluation, broader candidate generation, promotion policy and optional signed release distribution—see the repository’s continuous-learning design.