September 5, 2026
The Pi coding agent added support for OpenAI’s GPT-6 Astra in version 0.85.1, published on September 5, two days after OpenAI began rolling out the model, according to the project’s changelog. The update covers OpenAI API keys and OpenAI Codex subscriptions.
Takeaway points
- Version 0.85.0, released a day earlier, added persistent Claude thinking effort and in-memory sessions that can be restored through the SDK, the changelog says.
- Version 0.85.1 fixed SDK import failures caused by unintentionally publishing internal experimental code and dependencies in 0.85.0, according to the changelog.
- On StrideNote’s test machine, 0.85.1 is installed and was still the latest release on the npm registry on September 10.
What is new in Pi coding agent 0.85
Version 0.85.1 is a small release. Besides GPT-6 Astra, it adds scrolling five times faster while holding Alt in fullscreen mode and fixes keybindings in the model and thinking selectors. Its main fix concerns the release before it: 0.85.0 had unintentionally published internal experimental code and dependencies, which broke imports for SDK users. The changelog says the experimental client and plugin paths are now available only from source, and that the supported local SDK and stdio RPC interface are unchanged. It also switched long prompt-cache requests for GPT-5.6 and later models to a 30-minute cache setting.
Version 0.85.0 carried the larger changes. Pi now preserves the thinking effort chosen for each turn on supported Anthropic connections and recovers when signed thinking blocks do not match. The fullscreen transcript gained a jump to the latest message, and SDK users can restore sessions stored elsewhere through an in-memory session manager. The same release fixed skills being unavailable when Bash was the only enabled tool, fixed file tools ignoring the working directory set by an extension, and removed an unavailable Grok Build 0.1 model from the model picker. Other fixes stopped concurrent session shares overwriting one another, kept session forks from losing their compaction boundary, stopped imported sessions overwriting an existing session with the same filename, and corrected proxy matching for root domains and subdomains, and repaired managed fd and ripgrep downloads on Linux musl systems. The same release made skills available when Bash is the only enabled tool and stopped branch summaries failing when reasoning used up the previous 2,048-token output limit. It also fixed plain-HTTP requests through a proxy that hung after a tool call, and made GitHub Copilot requests for Claude Fable 5 send the reasoning level the user selected. The Qwen Token Plan catalog gained Qwen3.8 Flash, and the built-in file tools now respect the working directory an extension sets. It also let developers set a vLLM scheduler priority and an output-token limit for OpenAI-compatible models.
Pi coding agent release history since version 0.83
The npm registry lists eight releases between 0.83.0 on July 29 and 0.85.1 on September 5. Version 0.83.0 added commands to export credentials for other clients and Claude Opus 5 through GitHub Copilot. Version 0.84.0, on August 6, added a fullscreen interface mode, Mermaid and LaTeX rendering and per-directory context overrides through an AGENTS.override.md file. It also added a built-in Baseten provider, authenticated through a BASETEN_API_KEY variable, with GLM-5.2 as its default model.
That release also changed Pi’s output format. JSON and RPC message_update events now carry only incremental deltas, and the cumulative message field was removed because it “caused quadratic output growth,” the changelog says. Clients that need partial messages must assemble the deltas between message_start and message_end.
Later releases filled in details. Version 0.84.1 added a Qwen subscription provider and a pi auth check command, which checks provider or model credentials in advance and can print the credential it resolves. Version 0.84.2 added fullscreen transcript search and a setting to choose which tools start by default. Version 0.84.3, on August 24, added a PowerShell tool for Windows and staged, verified updates for installer-managed installations, and renamed the GoogleThinkingLevel type to GoogleApiThinkingLevel, a breaking change for developers. Version 0.84.4 added terminal capability overrides and experimental support for a DeepSeek V4 Flash vision model, along with an RPC clear_queue command that retrieves and removes queued steering and follow-up messages.
Several of these changes matter for self-hosted models. Version 0.84.0 let users pass arbitrary sampling parameters to OpenAI-compatible servers and opt in to a vLLM thinking-token budget. Version 0.84.3 extended configurable thinking-token budget fields to vLLM, Qwen and SGLang, and llama.cpp servers, the engines that run most local models, and added an automatic fallback when Anthropic’s servers refuse a request. Version 0.84.2 added experimental strict JSON-schema constrained sampling for Pi’s default read, bash, edit and write tools, switched on only under the PI_EXPERIMENTAL=1 setting. Earlier, 0.83.0 had added headless OpenRouter sign-in over SSH and a breaking upgrade of the bundled TypeBox library that removed several deprecated interfaces.
Who owns the Pi coding agent now
Pi changed publishers this year. The npm packages moved to the Earendil name with version 0.74.0 on May 7, and 45 releases have been published under it, according to the npm registry. The source now sits in the earendil-works organization on GitHub.
How StrideNote installs and runs the Pi coding agent
On StrideNote’s test machine, Pi 0.85.1 is installed through npm under Node 22 and was the latest release on September 10. The studio’s shell command starts Pi with two house skills loaded through its --skill option. StrideNote runs Pi alongside Hermes Agent and OpenCode on its test stack. StrideNote’s guide to a local AI coding agent on a Mac covers the local side of that setup, and how the three agents share one voice file is covered in our brand voice setup for OpenCode, Hermes and Pi.
Other agents added GPT-6 Astra on the same day as Pi. OpenClaw’s release notes for version 2026.9.2, also published on September 5, list GPT-6 Astra support.
What the Pi coding agent changelog does not say
The changelog entry does not say which of Astra’s settings Pi exposes, such as its 1,050,000-token context window or its reasoning levels. It also does not say when the experimental remote-session client withdrawn from 0.85.0 will return to the published package.
Sources: Pi changelog; npm registry; OpenClaw release notes.