WebOperator 1.5.0

Added

• Stopping a run now keeps its work. A stopped task gets its own stopped
  status instead of being recorded as a failure, and the agent writes a summary
  of what it collected and what the goal still does not cover — falling back to
  a summary built without the model when a provider is unreachable.
• Follow-up questions about a finished or stopped run, answered from that run's
  own evidence without revisiting any page.
• Resuming after a pause tells the model the page, tab, or login state may have
  changed while it was stopped, so a user's intervention is not mistaken for
  the agent's own progress.

Changed

• Migrated to @types/chrome 0.2.x, which types chrome.storage reads as an
  untyped bag rather than pretending the stored shape is known. Settings,
  memory and checkpoint reads now state what they expect and keep their
  existing runtime guards. No behaviour change.
• Updated ESLint, Knip and typescript-eslint.
• Moved to React 19. No source changes were needed: the panel already mounts
  through createRoot and uses none of the APIs React 19 removed.
• Contribution, security and conduct documents, issue and pull-request
  templates, Dependabot, CodeQL and a nightly npm audit, plus a coverage
  floor wired into CI.
• Tagging a release now builds and publishes the unpacked extension as a zip
  archive, so installing needs no Node and no build step.

---

Install

1. Download weboperator-1.5.0-chrome.zip below and unzip it.
2. Open chrome://extensions and enable Developer mode.
3. Click Load unpacked and select the unzipped folder.
4. Press Cmd+Shift+K (Ctrl+Shift+K on Linux) to open the side panel.

No Node.js or build step is needed — the archive is the built
extension. Verify it against .sha256 if you want to.

The default provider is Ollama at http://127.0.0.1:11434, so have a
tool-capable local model running, or paste an API key for one of the
remote providers under Settings.

WebOperator 1.4.0

Added

• CAPTCHA handling grew from detection into an autonomous solver: Cloudflare
  Turnstile, reCAPTCHA and hCaptcha checkboxes, slider puzzles, visual text
  challenges, and generic AWS WAF, GeeTest and Arkose flows, with retry polling
  and natural pointer events. Anything it cannot clear is handed back in the
  live tab with a sticky resume banner, and the task resumes automatically
  instead of failing silently.
• PDF export for task reports, replacing the raw JSON trace download.
• Smart scroll with container detection and boundary feedback.
• AssistantBench eval tasks and a fetch script for them.
• An Agent Connection section in the side panel, with Export moved next to Copy.
• CI runs the check gate on ubuntu-latest and macos-latest, and the README
  carries a check badge.

Changed

• Node is pinned through .nvmrc and engines.node, with the requirements
  documented in docs/requirements.md.
• The loop guard is stricter, cutting repeated no-progress actions earlier.

Fixed

• Page state is read from the main frame rather than whichever iframe answers
  first — an ad or embed could previously win the race and describe the wrong
  document.
• Transient LLM rate limits are retried, and duplicate subtasks no longer
  appear in the task view.
• CAPTCHA detection no longer fires on ordinary pages running background
  scripts.

---

Install

1. Download weboperator-1.4.0-chrome.zip below and unzip it.
2. Open chrome://extensions and enable Developer mode.
3. Click Load unpacked and select the unzipped folder.
4. Press Cmd+Shift+K (Ctrl+Shift+K on Linux) to open the side panel.

No Node.js or build step is needed — the archive is the built
extension. Verify it against .sha256 if you want to.

The default provider is Ollama at http://127.0.0.1:11434, so have a
tool-capable local model running, or paste an API key for one of the
remote providers under Settings.

WebOperator 1.2.0

WebOperator 1.2.0

Adds Server-Sent Events streaming for task progress through the Local Agent API.

Added

• GET /v1/tasks/:id/events SSE endpoint for live task events.
• Native bridge forwarding for compact task update, step, error, and skills events.
• README documentation for task event streaming.

Changed

• API event payloads omit large screenshots and snapshots; full trace remains available through /v1/tasks/:id/trace.

Verification

• ./scripts/check.sh
• npm audit --audit-level=moderate
• node --check weboperator-bridge/bridge.js
• Offline SSE smoke test via curl.

WebOperator 1.1.0

WebOperator 1.1.0 adds a local agent-facing HTTP API bridge.\n\nHighlights:\n- New optional weboperator-bridge/ Native Messaging host.\n- Local HTTP API on http://127.0.0.1:8765 for external agents.\n- Browser data endpoints for snapshot and screenshot.\n- Browser action endpoints for navigate, click, type, press, scroll, and extract.\n- Task endpoints for starting tasks, listing tasks, reading traces, waiting, pausing, resuming, confirming, and stopping.\n- README documentation with curl examples and endpoint list.\n\nVerification:\n- ./scripts/check.sh passed on weboperator@1.1.0.\n- npm audit reports 0 vulnerabilities.\n- node --check weboperator-bridge/bridge.js passed.\n- Bridge health endpoint returns online/offline state when run standalone.\n\nSecurity note:\n- The bridge is local-only by default. Do not expose port 8765 to a network.\n

WebOperator 1.0.0

WebOperator 1.0.0 is the first stable release of the Chrome browser-agent extension.\n\nHighlights:\n- Stable supported task surface for evidence lookup, product ranking, prompt-injection resistance, form fill verification, multi-tab collection, and long-task resume.\n- End-to-end extension eval runner and repeat runner for flakiness checks.\n- Local Ollama, xAI, OpenAI, OpenRouter, and MLX provider clients.\n- Runtime guardrails for visible planning, one-tool-at-a-time execution, grounded final answers, prompt-injection leakage, and long-task checkpoints.\n- Release documentation, known limitations, and deterministic local fixture suite.\n\nVerification:\n- ./scripts/check.sh passed.\n- Grok/xAI eval gate passed: 3 consecutive runs, 10/10 tasks each.\n- Ollama local eval gate passed with gemma4:26b: 3 consecutive runs, 10/10 tasks each, using --timeout-ms 600000.\n- npm audit reports 0 vulnerabilities after build toolchain updates.\n\nNotes:\n- Ollama gemma4:26b is correct but slow; use the longer timeout for local release-gate evals.\n- Remote providers send page observations to the configured provider.\n