feat: add web_search and web_fetch tools by dpavlin · Pull Request #2 · citrolabs/pi-chrome-use

dpavlin · 2026-07-20T17:31:10Z

Summary

Port web search and fetch functionality from ds4 WebFetch to CDP. Adds two new Pi agent tools registered through a dedicated extension.

New Tools

web_search: Google search via Chrome with consent dialog handling, up to 20 visible result links and text snapshot as markdown
web_fetch: URL page fetch with dynamic scrolling, semantic HTML extraction (headings, paragraphs, lists, code, blockquotes), up to 80 links, 900KB truncation

Both tools share Chrome profile for session continuity and use profileDir to reuse an existing Chrome session.

Files Changed

src/web-fetch.ts: core web fetch/search logic (418 lines)
extensions/browser-execute-web.ts: Pi extension adapter (142 lines)
test/web-fetch.test.ts: 20 unit tests
test/browser-execute-web.test.ts: 9 integration tests
README.md + README.zh-CN.md: documentation updates
package.json: extension registration
src/cdp/session.ts: minor type fix

Tests

59/60 tests pass (1 pre-existing failure unrelated to this work)
Typecheck clean

…the agent

…ng snippet execution

…et ReferenceErrors

… error handling

When DevToolsActivePort is absent (Chrome started with --remote-debugging-port=9333), probe common fixed ports via /json/version to discover running instances. This makes session.connect() auto-detect existing browsers without manual wsUrl.

Port web search and fetch functionality from ds4 WebFetch to CDP. Adds two new Pi agent tools registered through a dedicated extension: - web_search: Google search via Chrome with consent dialog handling, up to 20 visible result links and text snapshot as markdown - web_fetch: URL page fetch with dynamic scrolling, semantic HTML extraction (headings, paragraphs, lists, code, blockquotes), up to 80 links, 900KB truncation Both tools share Chrome profile for session continuity and use profileDir to reuse an existing Chrome session. Includes: - src/web-fetch.ts: core web fetch/search logic (418 lines) - extensions/browser-execute-web.ts: Pi extension adapter (142 lines) - test/web-fetch.test.ts: 20 unit tests - test/browser-execute-web.test.ts: 9 integration tests - Updated README (EN + ZH), package.json, minor type fix in session.ts 59/60 tests pass (1 pre-existing failure: Chrome-on-port-9333)

… curl Make tool descriptions explicitly tell the LLM to prefer these tools over curl for web searches and page fetching. Add stronger guidance in promptGuidelines to use web_search/web_fetch instead of curl.

Change Target.createTarget from background: true to background: false so the user can see Chrome navigating during web search and fetch operations, matching the 'visible Chrome' design goal.

The promptSnippet was changed from 'Chrome' to 'curl' to emphasize the tool is a curl replacement. Update test accordingly.

Integration tests connect to Chrome on port 9333, navigate to pages, extract content, and leave tabs visible for 15 seconds so the user can visually verify tabs actually open during web_search/web_fetch.

The web_search and web_fetch tools now keep the Chrome tab visible for 15 seconds after extracting content, so the user can see what the extension did. The default is 15 seconds but can be configured via the keepTabVisibleMs option. Unit tests pass keepTabVisibleMs: 0 to avoid artificial delays.

- Update ensureConnected() to call session.connect() when profileDir is not provided, allowing detectBrowsers to find Chrome on common ports. - Fix whitespace-only query validation in webSearch. - Update tests to mock Session.connect for error cases to avoid connecting to real Chrome during unit tests. This ensures Chrome tabs open and stay visible for 15 seconds after extraction, even when profileDir is not explicitly provided.

The pi extension loader requires explicit file extensions in the manifest. Added .ts to browser-execute and browser-execute-web paths.

Add onProgress callback to web_search and web_fetch tools so the agent session can display status updates (query URL, navigating, extracting) to the user in real-time. This helps distinguish multiple concurrent browser calls in the tool execution log.

Add custom renderCall functions to web_search and web_fetch tools so the URL/query appears inline with the tool name in the agent UI, similar to how bash shows '$ command' inline. This helps distinguish multiple concurrent browser calls in the tool execution log.

Previously runPage() awaited a 15-second delay before closing the tab, blocking the entire tool call and preventing the pi agent from continuing to the next step. Now the tab closes in the background via setTimeout, so the tool returns immediately after content extraction. The Chrome tab stays visible for the full duration so the user can still see what happened, but the agent can proceed without waiting.

section9-lab · 2026-07-27T07:13:44Z

Thank you again for the continued work here. I really appreciate the energy and the concrete implementation. The idea of making Chrome-backed fetch/search visible to the user is interesting, and several infrastructure pieces here could be valuable for the project.

I did a first pass review and also tested the branch in a temporary worktree.

Validation results:

npm run typecheck passes.
npm test currently fails because the new integration tests require Chrome to be running at 127.0.0.1:9333.
GitHub reports this PR as conflicting with current main, mostly because the project has since been renamed to pi-chrome-use and the release workflow/package metadata changed.

Main blockers before this can be merged:

Please rebase on current main and keep the current package metadata: pi-chrome-use, citrolabs/pi-chrome-use, current image URL, and current release workflow.
Please make real-Chrome integration tests opt-in, for example behind RUN_CHROME_INTEGRATION=1. The default npm test must pass on a clean machine and in CI without a local Chrome debug port.
Please split the PR. The fixed-port/browser reuse work is much easier to review and likely fits the core package. The user-facing web_search / web_fetch tools are a larger product decision and should be separate.
Please avoid changing browser_execute error semantics from thrown tool failures into successful text responses. That is a behavior change for existing users and needs a separate discussion.
Please check resource lifecycle in web_search / web_fetch. They create new Session instances and should make sure WebSocket/browser resources are closed or intentionally reused.
Please remove unused imports and local-only/debug files before resubmitting.

For the user-facing web_search / web_fetch feature direction: I want to be transparent that I personally will not have time over the next two weeks to take ownership of that feature work. I am still happy to help review, test, and merge smaller focused PRs, especially the CDP detection/reuse improvements. If you want to continue with this, the fastest path is probably:

PR A: fixed-port CDP detection only
PR B: optional Chrome auto-launch/reuse behavior
PR C: web_fetch/web_search as an optional extension, after the API shape is agreed

Thanks again. This is useful work; I just want to keep the merge path small and safe for existing users.

dpavlin added 19 commits July 11, 2026 12:19

fix(execute): catch snippet execution rejections to prevent crashing …

76c90cd

…the agent

fix(execute): intercept unhandledRejection and uncaughtException duri…

9ec5912

…ng snippet execution

test(execute): add automated test cases for async and unawaited snipp…

f3c0f93

…et ReferenceErrors

fix(execute): strip verbose stack traces from tool error responses

f42d6f7

feat(cdp): support auto-launching Chrome and concurrency-safe process…

1cdf76d

… error handling

test(cdp): add unit tests for Chrome auto-launching and reuse logic

a192d8a

fix: improve web_search and web_fetch tool descriptions to discourage…

bdc1c60

… curl Make tool descriptions explicitly tell the LLM to prefer these tools over curl for web searches and page fetching. Add stronger guidance in promptGuidelines to use web_search/web_fetch instead of curl.

fix: open visible Chrome tabs for web_search and web_fetch

78d57f5

Change Target.createTarget from background: true to background: false so the user can see Chrome navigating during web search and fetch operations, matching the 'visible Chrome' design goal.

fix: update web_fetch test assertion for new promptSnippet

941589d

The promptSnippet was changed from 'Chrome' to 'curl' to emphasize the tool is a curl replacement. Update test accordingly.

test: add integration tests with real Chrome browser

3554d68

Integration tests connect to Chrome on port 9333, navigate to pages, extract content, and leave tabs visible for 15 seconds so the user can visually verify tabs actually open during web_search/web_fetch.

fix: add .ts extensions to package.json manifest entries

fd24957

The pi extension loader requires explicit file extensions in the manifest. Added .ts to browser-execute and browser-execute-web paths.

chore: remove debug logging from browser-execute-web.ts

75bec39

section9-lab mentioned this pull request Jul 27, 2026

improvements to this very useful extension #1

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: add web_search and web_fetch tools - #2

feat: add web_search and web_fetch tools#2
dpavlin wants to merge 19 commits into
citrolabs:mainfrom
dpavlin:web-fetch-search

dpavlin commented Jul 20, 2026

Uh oh!

section9-lab commented Jul 27, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

dpavlin commented Jul 20, 2026

Summary

New Tools

Files Changed

Tests

Uh oh!

section9-lab commented Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

section9-lab commented Jul 27, 2026 •

edited

Loading