Pi Web Access Files

Pi Web Access 0.25.0 makes web access work in more real-world environments. You can now route searches and fetches through an explicit proxy, fetch useful GitHub PR and issue summaries, and use Kimi Code Plan search when you are signed in through Pi. Gemini setups are also more flexible, with deterministic browser cookie profile selection and Application Default Credentials support for Vertex-backed generate-content calls. Extraction and retrieval paths are less fragile, so failures are clearer and stored-content lookups behave better.

Highlights

• web_search, source_check, and fetch_content can now use an explicit HTTP(S) proxy for restricted networks.
• fetch_content now gives useful GitHub PR and issue summaries, including comments, review threads, anchors, and truncation markers.
• Gemini users can choose browser cookie profiles more predictably and can use Google Application Default Credentials for Gemini generate-content calls.
• Kimi Code Plan users can run explicit Kimi web searches through Pi's kimi-coding login.
• HTML extraction, missing-page guidance, and stored-content lookup are more reliable.

Need to know

Highlights

• web_search, source_check, and fetch_content can now use an explicit HTTP(S) proxy for restricted networks.
• fetch_content now gives useful GitHub PR and issue summaries, including comments, review threads, anchors, and truncation markers.
• Gemini users can choose browser cookie profiles more predictably and can use Google Application Default Credentials for Gemini generate-content calls.
• Kimi Code Plan users can run explicit Kimi web searches through Pi's kimi-coding login.
• HTML extraction, missing-page guidance, and stored-content lookup are more reliable.

Added

• Added an optional proxy parameter to web_search, source_check, and fetch_content. When set, search APIs, page fetches, and content extraction use curl through that proxy. Localhost and NO_PROXY hosts still bypass the proxy. A default proxy can also be set with "proxy" in ~/.pi/web-search.json. Thanks to @mystery4f for PR [#307].
• Added deterministic Chromium cookie selection with browserCookies.browser and browserCookies.profile; arbitrary profile paths remain unsupported. Thanks to @lmilojevicc for issue [#297].
• Added GitHub PR and issue specialization in fetch_content, with gh-first metadata, bounded REST fallback, comment anchors, review threads, truncation markers, and the githubPrIssue.enabled opt-out (#294).
• Added explicit-only Kimi Code Plan search using Pi's refreshed kimi-coding OAuth credentials. Thanks to @lushangkan for PR [#275].
• Added opt-in searchRouting.useCurrentModel routing for automatic searches. Official OpenAI GPT Responses models can now fund Hosted Search with their current endpoint, credentials, and headers before configured fallbacks. Thanks to @nyankosama for PR [#293].
• Added strict Hosted Search response validation and better fallback classification for unsupported Hosted Search tools.
• Added current-model Hosted Search routing for official openai-codex GPT Responses models through the Codex Responses endpoint.
• Added Google Application Default Credentials support for Gemini generate-content calls with geminiAuth: "adc". This covers Gemini search, URL context, PDF, and inline-data extraction through Vertex AI. YouTube and local video analysis still require GEMINI_API_KEY because they use the Gemini Files API. Thanks to @smazurov for PR [#301].

Fixed

• Added Defuddle as a local fallback when Readability and RSC extraction cannot recover useful HTML content. Thanks to @tobru for issue [#300].
• Kept Gemini ADC config and credential errors, GitHub PR/issue REST fallback failures, and Defuddle fallback failures visible instead of silently downgrading them.
• Pointed definitive fetch_content 404/410 failures to search guidance instead of provider configuration. Thanks to @Daniishkhan for PR [#308].
• Improved Gemini Web browser-cookie diagnostics so /google-account shows sanitized attempted browser/profile entries and distinguishes missing required cookies, password-store access, and decryption failures. Thanks to @lmilojevicc for issue [#296].
• Made get_search_content tolerate bridge defaults when findText is supplied, while preserving ordinary pagination. Thanks to @ZacharyQin for PR [#295].