Pi 编辑器接入 GPT-6 Astra 与 Codex

Howaboua @Howaboua原推发布于 2026.09.05

在 X 上打开原推

People of @pidotdev! GPT FUCKING 6 DROPPED :D (lol my Clanka took my style correction quite literally, I might need to edit this one heavily) Astra gets the Pi Codex treatment. Non-cache busting reasoning controls, context management tool (yes, that one, using Codex's primitves 1:1. It's a banger), TRUE developer role messages, and more. This is a chunky one. Any questions, ask but it's 2:30 AM, lol. Might be async :D - GPT-6 Astra: native Codex integration, plus /astra in GPT Switcher extension. Low reasoning by default, override it in the command when you want more. - Reasoning without rewriting the request prefix for Astra: your normal Shift+Tab changes preserve prompt-cache and WebSocket continuation eligibility, even across resume and native compaction. Not a promise of cache hits, but we stop breaking eligibility ourselves. - Auto reasoning tool in /codex tools: let Astra adjust effort by work phase in Structured, Code or Notebook mode. Opt-in, never below your starting level, restored after the run settles. The Clanka can think harder when it needs to without leaving your settings cranked up afterwards. - Fresh context, same workspace: experimental new_context rollover without an automatic summary. Shell and Notebook state survive, full history stays recoverable. Local mode uses Pi's JSONL with sneaky filtering, Tree archives context windows as branches (Pi style, I do not recommend - it uses a summarisation llm call and a lot of ctx), Remote uses Codex's OG encrypted history and notes service. Budget reminders included, no silent Remote fallback so test if it works with your acc/model. - Actual developer messages via Pi-Codex API (your extensions can send a dev role message mid-session): extensions can deliver persisted policy without pretending it's something you typed. Custom rendering and restoration stay intact. Incompatible models keep ordinary Pi delivery and sessions should be cross-compatible. This works thanks to the custom provider implementation. - Examples in the monorepo extensions: Ask answers get sent async if agent decided to use async mode, nested AGENTS.md now load as a dev message, Shepherdr events, review policy (subagent review) and voice lifecycle guidance use the compatible Responses path. - Luna Reserve: that gpt-reserve bucket in /codex usage is a separate limited allowance for eligible accounts after ordinary quota runs out. Pi Codex switches only with backend authorization and asks you to continue. No automatic retries or spending reset credits. Your original model and reasoning come back after confirmed recovery, including resumed sessions. And the Codex annoyances: - Custom models and models.json overrides survive refresh. Optional tool arguments stay optional, explicit strict sampling stays strict. - Streams finish even when the terminal SSE event lacks its final separator. Successful recovery clears stale incomplete-response errors. - Notebook's details toggle actually hides duplicate previews. Notes, history, Notebook and context actions get compact results with expandable detail. - Submitted messages show up without waiting for cached WebSocket warmup. Generation still waits its turn. It should all feel snappier. - Simpler Notebook metadata persistence and voice startup, same saved data, connection behaviour and tool output. Fixes for the other stuff in the monorepo: - Ask: steering questions keep the full response panel without freezing useful reversible work. Answers arrive at the next safe boundary, pending questions survive reloads. Gating decisions and human handoffs still wait (similar to new Codex async implementation). - Shepherdr: workers settle correctly after Pi expands skill or prompt-template invocations. No more waiting on work that's already finished because of that expansion. - Nested AGENTS.md: developer-message delivery when available, tool-result delivery otherwise, with branch-aware deduplication. - Browser: actual JavaScript exception details instead of the deeply informative "Uncaught". - Better Skills: batch independent reads in one cell. Stop repeating the entire inventory for a reference read, stop listing installed dependencies as skill content. - Imagegen: null edit selectors no longer turn a prompt-only generation request into an edit request. - Pi Pet: fixed the package-root entry so aggregate bundles can load the extension. Also went through all 79 README.md and AGENTS.md files. Fixed stale requirements and missing setup steps, cut the module tours, kept the useful boundaries. It was all one shotted by Astra low (multi prompts obvs), in one session, using the locally updated Pi-Codex extension with ctx management enabled. Pick what you use: pi install npm:@howaboua/pi-codex-conversion pi install npm:@howaboua/pi-gpt-switcher pi install npm:@howaboua/pi-ask pi install npm:@howaboua/pi-shepherdr pi install npm:@howaboua/pi-subagent-review pi install npm:@howaboua/pi-subdir-agents pi install npm:@howaboua/pi-browser pi install npm:@howaboua/pi-better-skills-tool pi install npm:@howaboua/pi-codex-imagegen https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-codex-conversion https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-gpt-switcher https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-ask https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-shepherdr https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-subagent-review https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-subdir-agents https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-browser https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-better-skills-tool https://github.com/IgorWarzocha/howaboua-pi-stuff/tree/main/packages/pi-codex-imagegen Happy Clanking! - Howaclawa