Repository navigation
Reviewed UI capabilities
This release adds a complete local discover → inspect/diff → approve → invoke lifecycle around the existing deterministic computer-use runtime.
- Bounded exploration chooses actions and order without a prescribed next step. An unordered, author-reviewed checkpoint catalog constrains success semantics.
- Ed25519 approvals bind exact artifact content and runtime policy, with an independent trusted key and expiration checked before browser launch.
- Explicit target, profile, input and policy configuration; a surface adapter port with a concrete Playwright implementation.
- Discovery can transfer the live browser session to an operator and verify repair before resuming.
- Nine synthetic failure scenarios, seeded approval mutation tests, and an end-to-end signed invocation test.
- Standard mechanisms from Playwright, Zod, the official OpenAI SDK, jsondiffpatch, fast-check and Node crypto.
Verification
- 71 automated tests, formatting, lint, types and build.
- Fresh OpenAI exploration: 9 model calls, 8 actions, followed by changed-input replay with the API key absent and 0 model calls.
- 27/27 expected outcomes across nine local synthetic browser scenarios; not a production reliability or scale claim.
- Eight curated evidence scenarios and 39 integrity-checked files. Original recordings remain unchanged.
- Dependency audit and repository secret scan pass. No credentials are included.
Compatibility and limits
Existing schema-v1 artifacts and the synthetic demo remain supported. Artifact writes now refuse overwrites; use a new filename for each draft. run requires approval; replay remains the local synthetic fixture shortcut.
Web-only, local runtime. No production bank integration, remote operator authentication, screenshot DLP, distributed execution, or open-ended recovery is claimed. Exploration drafts still require human review.
See operations, engineering decisions, and evaluation methodology.