Skip to content

WebCodex v0.3.8

Choose a tag to compare

View all tags
@yyjeqhc yyjeqhc released this 20 Aug 08:23
· 214 commits to main since this release
v0.3.8
477c1f7
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

WebCodex 0.3.8 is a V1 capability and reliability release focused on guarded cross-platform Computer Use, safer large-artifact transfer, durable Runner/Job recovery, and tighter model-facing bounds and authorization.

Highlights

  • Windows and macOS Computer Use. The native Windows UI Automation and macOS Accessibility backends now share a semantic-first surface for exact window observation/activation, element discovery/state, press/focus, bounded text input, semantic scroll, and closed key input, with stale-handle fencing and outcome_unknown reconciliation instead of blind retries.
  • Bounded application, display, pointer, and clipboard control. Windows and macOS add opaque-ID application discovery and launch, exact full-display discovery/snapshots, snapshot-generation-fenced pointer move/single-left-click, and bounded plain-Unicode clipboard read/write. The higher-risk operations remain separately capability- and scope-gated rather than being implied by ordinary Computer observation/control.
  • Computer image delivery. Snapshot output is more tightly bounded, and MCP Computer snapshots can be surfaced through authenticated resource links instead of forcing large image payloads through the model-facing structured result.
  • Streaming artifact data plane. Artifact upload/import/export paths now support bounded streaming-oriented transfer so large files do not require model-context-sized Base64 aggregation.
  • Durable Job recovery. Detached Job ownership handoff survives Runner restart, including macOS parity, while process lifecycle and transport failure handling preserve explicit not-started/unknown semantics.
  • Sessions, audit, and authorization. Hosted Workflow Session observability, authenticated caller attribution, unified bearer-scope enforcement, and the documented manual multi-window collaboration flow improve evidence and operator control without automatic transcript sharing.
  • Tighter coding-tool bounds. Model-facing tool context is smaller, guarded text edits handle CRLF correctly, Git path mutations use structured argv, generated helper programs are isolated from configured shells, and sparse project-search output is reflected consistently in E2E/eval expectations.
  • Release engineering hardening. Exact-source release-readiness, five-platform pre-tag native gates, guarded failed-tag recovery, durable release-build correlation, same-run bundle collection, and digest-based draft verification make publication recovery safer and more reproducible.

Compatibility and upgrade notes

There is no intentional breaking protocol change. New Runner capabilities are additive, default false when absent, and fail closed when an older peer cannot prove support.

Upgrade Server and Runner together before relying on the new Computer Use or detached-Job restart behavior. Full-display observation, pointer control, clipboard access, and application launch use explicit scopes/capabilities; existing clients do not receive those broader authorities automatically.

Known limitations

  • Native desktop Computer Use is currently implemented for Windows and macOS; browser DOM/ARIA-native automation remains a separate future surface.
  • Coordinate pointer control is intentionally narrow: move and one left click only, with no drag, wheel, right/double click, or generic global-coordinate API. The macOS pointer path currently requires an unrotated display.
  • MCP App snapshot rendering can still be host-dependent; an intermittent blank/gray card has been observed even when the underlying capture succeeds.
  • macOS x64 is not a published native artifact.
  • The npm wrapper requires Node.js 18+ independently of native binary compatibility.

Validation

  • Release source: 477c1f754e8b5c7d9f0e8b1c073487532a749101 at immutable annotated tag v0.3.8.
  • Exact-source release-readiness run 32339887099 completed successfully, including the canonical release check, locked workspace suite, frontend checks, WebSocket and polling zero-config E2E, coding-loop comparison eval, and all five native pre-tag validation lanes.
  • Coding-loop comparison eval: 6 passed / 0 failed.
  • Authoritative release-build run 32343102712 completed successfully for Linux x64/ARM64, macOS ARM64, Windows x64/ARM64, plus the same-run assemble job.
  • Both Linux targets passed the manylinux2014 ABI/dependency gates: x64 requires at most GLIBC 2.16 and ARM64 at most GLIBC 2.17, with only the reviewed dynamic dependencies.
  • Windows release candidates passed the workflow's native package/install smoke.
  • The staged npm package passed the retained-binary staging/smoke flow using the exact CI-produced candidate bytes.
  • The retained candidate archives are published with SHA256SUMS; all five local retained hashes were rechecked immediately before draft upload.
Assets 10
Loading