@earendil-works/pi-durable

New Features

• Image input for classifiers: codemode models.classify() accepts images for TypeSafe's Jev and Cloudflare's Clef models on typesafe, cloudflare-workers-ai, and OpenRouter. See Use classifier models and Codemode.

• Models Configuration reference: every provider, model, model override, and compatibility field of models.json, generated from its schema. See Models Configuration.

• Summaries without a fixed output cap: compaction and branch summaries use the model's output limit, so long reasoning no longer makes them fail. See Compaction.

• Codemode waits for a free worker: parallel scripts and in-process subagents queue instead of failing, and PI_CODEMODE_MAX_WORKERS raises the limit. See Codemode limits.

• Rex terminal support: OSC 8 hyperlinks, truecolor, and inline images.

Added

• Added a Models Configuration reference generated from the models.json schema, listing all provider, model, model override, and compatibility fields

• Codemode models.classify() now accepts images for TypeSafe's Jev and Cloudflare's Clef models on typesafe, cloudflare-workers-ai, and OpenRouter, and lists Clef Omni and Jev 1.14 on Workers AI.

• Provider-level compat in models.json now also applies to classifier models, with the System One fields imageInput and sendImageOptInHeader for proxies that serve another service's classifiers. Extension classifier models accept compat too.

• Added MiniMax-M3.1-Flash to the minimax and minimax-cn providers, with thinking levels low through max

• Added support for the Rex terminal: OSC 8 hyperlinks, truecolor, and inline images (#11131, #11150

Changed

• Provider attribution headers now come from pi-ai's client identity. Extensions' before_provider_headers handlers no longer see the OpenRouter, NVIDIA NIM, and OpenCode attribution headers; returning null for one still removes it. Cloudflare requests send pi's detailed User-Agent instead of pi-coding-agent, as do GitHub release checks for tool downloads, and OpenCode requests send x-opencode-client whenever attribution is enabled, not only with a session.

• generateSummary() and generateSummaryWithUsage() now ignore their reserveTokens parameter. Summary requests use the model's output limit instead.

• Changed the default GitHub Copilot model from gpt-5.4, which Copilot retired, to gpt-6.1-sol.

• Changed the default NVIDIA model to nvidia/nemotron-3.5-super-vl-120b-a12b

• Changed the default Xiaomi model to mimo-v2.6-pro for the xiaomi and Xiaomi Token Plan providers, since mimo-v2.5-pro left the catalog.

• Grok 4.3 and Grok 4.5 on the xai provider now offer the xhigh thinking level.

• Sign in with ChatGPT: ChatGPT Business workspace usage-limit errors are not retried and tell the user to ask their workspace admin to increase the limit.

Fixed

• Fixed compaction and branch summaries failing with "generation hit the token cap" when reasoning used most of the fixed summary output limit. Summaries now use the model's output limit, and compaction.reserveTokens only controls when compaction starts (#9512, #9075).

• Fixed models.schema.json missing descriptions for provider and model fields, so editors showed no hints for them

• Fixed pi update failing with "Managed install marker is missing or invalid" when managed-install.json starts with a UTF-8 BOM, which Windows PowerShell 5.1 adds when it rewrites the file.

• Fixed HTML export showing literal asterisks for Markdown bold and italic next to CJK punctuation (#3353, #10154)

• Fixed the terminal staying in mouse-tracking, focus-reporting, Kitty keyboard, and alternate-screen modes when pi exits without its normal terminal teardown, which flooded the shell prompt with escape sequences

• Fixed concurrent extension select/confirm/input/editor dialogs and inline custom() components in interactive mode: a second dialog replaced the first and the first promise never settled; closing the visible dialog now shows the previous one again (#11166, #9909)

• Fixed pasting in the shell hanging after copying text in pi and suspending it with Ctrl+Z on Linux: the background wl-copy/xclip/xsel process that keeps serving the clipboard was suspended along with pi

• Fixed codemode scripts failing immediately when both codemode workers are busy, for example with parallel codemode calls or in-process subagents. Scripts now wait up to 60 seconds for a free worker, and PI_CODEMODE_MAX_WORKERS raises the limit of two workers

• Fixed extensions failing to load when importing @earendil-works/pi-ai/models (e.g. for clampThinkingLevel) by exposing the subpath through the extension loader aliases and virtual modules

• Fixed mcp.json files that start with a UTF-8 BOM failing to parse, which skipped all their MCP servers and made pi mcp add/pi mcp remove fail

• Fixed /compact <instructions> ignoring the instructions when summarizing the start of a turn that compaction splits. When the split turn was the only new content since the last compaction, the instructions were ignored entirely

• Fixed resumed sessions and /tree navigation leaving extension tools inactive when the branch recorded its tools before they were registered, such as tools from extensions installed after the session started. Extension tools the branch never added or removed are now active unless they are registered with defaultActive: false (#11062, #11075)

• Fixed sessions breaking after recovery from a response cut off at the output limit: the truncated tool call was omitted but its tool results were kept, so providers rejected every later request (#11227, #11233

• Fixed the grep tool following RIPGREP_CONFIG_PATH, where options such as --fixed-strings silently changed regex searches into literal ones. The tool now runs rg --no-config (#11203, #11230

• Fixed Windows shell detection rejecting runnable Microsoft Store shell aliases

• Fixed Windows shell detection missing PowerShell outside PATH and the legacy WSL bash.exe launcher when Windows is not installed in C:\Windows. pi now checks %ProgramFiles%\PowerShell\7, %SystemRoot%\System32\WindowsPowerShell\v1.0, and %SystemRoot%

• Fixed GitHub Copilot sessions with large request bodies, such as many images, failing with HTTP 413 failed to parse request instead of compacting and retrying

• Fixed the openai-codex provider offering gpt-5.5 and gpt-5.3-codex-spark, which always fail for ChatGPT accounts

• Fixed whitespace-only tool output (for example a bare \r\n from bash) making OpenAI Completions, OpenAI Responses, and Google requests fail with a 400

• Fixed automatic retry not triggering for more gateway upstream failures, stream disconnected before completion errors relayed by Codex CLI proxies, and transient Google Vertex OAuth token refresh failures, and retrying errors because of digits inside request ids or byte counts (#10920, #11057, #10996)

• Fixed OpenAI Responses models losing the prompt cache on every deferred tool load after a tool was removed

• Fixed Cloudflare Workers AI classifier pricing: Clef Flash costs $0.038 and Jev $0.042 per million input tokens instead of Jev being listed as free.

• Fixed an unclosed $$ display-math block swallowing or corrupting every Markdown table after it

• Fixed Markdown bold and italic not rendering next to CJK punctuation

• Fixed @ file autocomplete returning no results for queries containing regex metacharacters such as [ or (

New Features

• Choose where /share uploads: sessionShareDestination selects auto, gist, or radius. See Sessions and context and Export or share a session.

• Cache-preserving thinking level changes on OpenAI GPT-6: changing the thinking level mid-session with an API key keeps the prompt cache.

• Kimi Code over Chat Completions: tools added mid-conversation no longer invalidate the prompt cache. models.json base URL overrides must now include /v1. See Providers.

• Clean model switching: thinking from a previous model is dropped instead of being shown to the new model as text.

• Resumed sessions keep their active tools: --tools +name/-name applies on top of them. See Tools.

Added

• Changing the thinking level mid-session on OpenAI GPT-6 models (API key) now keeps the prompt cache

• Added the sessionShareDestination setting to choose automatic, GitHub gist, or Radius uploads for /share

Changed

• Changed the kimi-coding provider to use OpenAI Chat Completions at https://api.kimi.com/coding/v1, the protocol used by the official Kimi Code client. K3, K3-256K, and Kimi For Coding now accept tools added mid-conversation without invalidating the prompt cache. models.json overrides of the kimi-coding base URL must now include /v1

• Changed model switching to drop thinking blocks from earlier assistant messages produced by a different model, instead of sending them as plain text. The new model previously saw the old model's reasoning as part of its replies and could start writing its own reasoning into the visible answer

Fixed

• Pinned internal Pi dependencies to exact release versions to prevent fresh installs from mixing incompatible packages

• Fixed resumed sessions dropping tools activated during the session. Resume now restores the session's active tools, with --tools +name/-name entries applied to them, unless --tools with tool names, --no-tools, or --no-builtin-tools selects tools explicitly

• Fixed stopping a codemode script or reaching its deadline waiting for a non-interruptible built-in to return on Bun

• Fixed repeated codemode timeouts or aborts accumulating surviving workers; admission now limits live workers to two across sessions, counting cancelled workers until actual exit

• Fixed --model choosing a provider by catalog order when several providers offer the match; Pi now prefers the provider with credentials and otherwise reports the ambiguity. --models and enabledModels add a model for each provider that offers it and warn about the ambiguity

• Fixed read line-range labels for numeric-string or null arguments in the TUI, session tree, HTML exports, and examples

• Fixed the Nix flake build failing when a provider switched APIs: the model catalog pin is now refreshed from the newest published catalog when its model API groups differ, and releases pin the newest catalog and build with it before tagging

• Fixed the Nix package putting its Node.js first on PATH, which made bash tool shells use Pi's Node instead of the user's node/npm. Pi's Node is now only a fallback

• Fixed prompt templates dropping empty quoted arguments ("" or ''), which shifted later positional arguments and prevented positional defaults from applying

• Fixed the Nix package build failing in its install check because it still expected chord and esbuild, which Pi no longer depends on

• Fixed tool-call argument parsing dropping fields when a string contains an invalid \u escape such as \usepackage

• Fixed Anthropic and Bedrock Claude requests failing with a 400 when an error tool result contains images; the images are now sent after the tool results of the turn, so the model still sees them

• Fixed OpenCode Go Gemini models failing with a 404 by moving OpenCode Zen and OpenCode Go to the opencode.ai/inference endpoints, and added Google Generative AI support to the OpenCode Go provider

• Fixed OpenCode Go MiniMax M2.7 failing with "Model does not support this protocol" by using the Anthropic Messages API

• Fixed automatic retry not triggering for gateway Unable to forward upstream stream errors

• Fixed automatic retry not triggering for The service is temporarily unavailable errors, which also made compaction fail on the first attempt

• Fixed OpenAI-compatible responses showing thinking after the answer when a proxy such as LiteLLM sends reasoning and content in the same stream chunk

• Fixed Gemini 4 Argon requests on Google Vertex failing with "Thinking budget is not supported", and its thinking level being ignored on the Gemini API

• Fixed Xiaomi MiMo models failing with a 400 at the minimal thinking level; they now expose only the off and high levels

• Fixed openai-codex/gpt-6.1-sol using a 272K context window instead of the 500K window Codex now provides

• Fixed tool calls whose final arguments are malformed JSON silently running with only the fields parsed before the error; they now fail validation and show the model the raw arguments it sent

• Fixed Claude Haiku 5.5 and Claude Opus 5 not supporting the off thinking level

• Fixed done() of an extension's ctx.ui.custom({ overlay: true }) closing the topmost overlay instead of its own, which left the finished overlay on screen and another extension's dialog unresolved (#10648, #10999

New Features

• Per-key OpenRouter model lists: with an OpenRouter key, Pi lists only the models that key may use, with the key's context size, max output, and prices, including regional endpoints set in models.json. See OpenRouter.

• Published configuration schemas: add $schema to settings.json, models.json, keybindings, and themes for editor validation and completion. See Settings, Models, Keybindings, and Themes.

• Clipboard copies over remote sessions: PI_CLIPBOARD_OSC52=1 sends copied text to a remote client's clipboard when Pi cannot detect the remote session, such as a multiplexer attached from another machine. See Pi Process Configuration.

• Codemode improvements: tool declarations show input constraints such as integer, minimum/maximum, and default, and codemode rows show elapsed and total time. See Codemode.

• before_agent_start for extension-triggered runs: runs started by pi.sendMessage(..., { triggerTurn: true }) go through the same setup as prompts. See Events and concurrency.

Breaking Changes

• breaking

  Published generated configuration schemas for models, settings, keybindings, and themes at https://pi.dev/schemas/*.schema.json. The theme schema moved to https://pi.dev/schemas/theme.schema.json; themes loaded during startup now receive the same schema validation as later reloads. Update theme $schema references, define reusable colors under vars, and remove unsupported metadata (#9880, #10751

• breaking

  Changed before_agent_start to also fire for runs started by pi.sendMessage(..., { triggerTurn: true }) while idle. BeforeAgentStartEvent is now a union on messageType: "user" events carry prompt and images, and "custom" events carry the starting message and no prompt. Messages returned by handlers are added to both kinds of run, and pending deliverAs: "nextTurn" messages are delivered with the custom-message run. Handlers that should act only on user prompts must check event.messageType === "user" (#10267, #5581, #10739

• breaking

  Changed the session to count as busy while a run starts, including during before_agent_start handlers: ctx.isIdle() is false, pi.sendMessage() queues into the starting run instead of starting a second run that failed with "Agent is already processing", pi.sendUserMessage() requires deliverAs, and an abort requested meanwhile aborts the run once it starts

• breaking

  Changed pi mcp list --json to report resources: true for servers that offer resources instead of resources and resourceTemplates counts; Pi no longer lists resources when connecting

Added

• Added reachesTools to tool definitions, so tools that reach other tools (such as codemode and tool_search) declare which tools a call may reach; the MCP extension uses it to wait for the servers a call needs and to decide whether a replacement codemode or tool_search tool can reach MCP tools

• Added aliases to tool namespaces; codemode's describeNamespace() and the namespace option of searchTools() match them, and MCP namespaces use the server name

• Added per-key OpenRouter model lists: with an OpenRouter key, Pi lists only the chat models that key may use, with the key's context size, max output, and prices, including guardrails and regional endpoints set as the base URL in models.json. The list is saved and used offline (#10353, #10672

• Added PI_CLIPBOARD_OSC52=1 to send copied text to the terminal with OSC 52 after the local clipboard write, so copies reach a remote client's clipboard when Pi cannot detect the remote session, such as multiplexer attaches from another machine or remote IDE daemons (#10826, #10835, #10114)

• Added elapsed time while a codemode script runs and its total time once finished to the codemode tool row, like bash rows

Changed

• Changed the showHardwareCursor setting to use only the terminal cursor instead of also drawing Pi's reverse-video cursor

• Changed codemode tool declarations to show input constraints such as integer, minimum/maximum, length, pattern, and default as @ tags on property comments, so scripts see them without the tool repeating them in descriptions

• Changed codemode declarations of MCP tools to spell out the result type (content, structuredContent, isError) instead of referring to CallToolResult<T>, which describeTool(), searchTools(), and ALL_TOOLS showed without defining it; the codemode description no longer includes the shared MCP types

• Changed serializeConversation() (compaction, branch summaries, bug reports) to omit thinking blocks; replayed thinking is often not in the live context, so including it could make a summarization request larger than the context it summarizes and fail

Fixed

• Fixed codemode ignoring a // @options: line that is not the first line of the script, which ran the script without the requested timeout_ms; the call now fails with an error before the script runs

• Fixed Mistral streams being aborted mid-generation (e.g. during long thinking) after the request timeout; the timeout now applies only to waiting for response headers

• Fixed a symlinked AGENTS.md in a git worktree nested inside its main repo causing no project context file to load

• Fixed the code mode tool description claiming every failed tool call rejects; a tool that declares an output schema resolves with its value on failure too (e.g. bash with a nonzero exit)

• Fixed runs started by pi.sendMessage(..., { triggerTurn: true }) removing system prompt sections added in before_agent_start after their first tool call, which changed the prompt mid-run and broke the provider prompt cache twice per run (#10267, #10739

• Fixed codemode blocking the process for seconds to minutes after scripts with many text() calls; joining the output items was quadratic

• Fixed print mode (-p) exiting 0 without a warning when the final response was cut off at the model's output token limit; it now prints the partial text, writes a warning to stderr, and exits 1

• Fixed pi mcp login --help and the CLI docs describing --timeout as only the wait for the browser; it limits the whole sign-in

• Fixed ~ not being expanded in !pattern, +path, and -path entries of the extensions, skills, prompts, and themes settings, so entries like !~/skills/pdf matched nothing

• Fixed MCP servers with resources registering no tools until every resource was listed; Pi no longer lists resources when connecting, so /mcp and pi mcp list show that a server offers resources instead of counting them

• Fixed MCP tool calls failing when the server returns structured content that is not an object (an array, null, or a primitive, as MCP 2026-07-28 allows); codemode declarations of MCP tools without an output schema now type structuredContent as unknown instead of an object

• Fixed reading or writing auth.json on Windows failing with EPERM when the previous lock folder was still being deleted; Pi now retries briefly before reporting the error

• Fixed OpenRouter router models such as openrouter/auto showing a cost of -$1,000,000 per 1M tokens; OpenRouter's negative placeholder prices are now 0

• Fixed Anthropic requests failing when the provider rejects the strict tool set for grammar size, tool count, or union limits; Pi now retries once with preferred tools non-strict

• Fixed "The upstream provider returned an error" not being retried as a transient provider error

• Fixed focused fullscreen overlays blocking mouse-wheel scrolling of the transcript

• Fixed fragmented terminal color replies leaking hex values into the prompt and opening the external editor (related reports: #10256, #10657)

• Fixed inline images not rendering in Herdr since 1.1.0

• Fixed key bindings that name a shifted symbol, such as + or ctrl++, not matching under the Kitty keyboard protocol or modifyOtherKeys

• Fixed the compaction status indicator embedded in a custom editor using the default muted colors instead of the editor border or thinking-level color, as the working indicator does (#10809, #10810

• Fixed MCP tool calls that were cancelled or timed out while waiting for an OAuth refresh still being sent to the server once the refresh finished

• Fixed MCP OAuth sign-in accepting an authorization response with an iss parameter that could not be checked because no authorization server metadata was discovered; sign-in now fails with an issuer mismatch. Reported

• Fixed codemode failing with a broken bridge error when Pi runs under node --watch on Node 24 or 26

Removed

• breaking

  Removed the unreleased experimental server, client, and plugin stack (src/experimental, src/cli/experimental) and its source-only ./client and ./experimental/plugin exports; none of it was part of the published package. The @earendil-works/pi-client, @earendil-works/pi-protocol, and @earendil-works/pi-server packages it used are removed and no longer released

• breaking

  Removed the dependency on @earendil-works/chord

New Features

• Program status reporting: terminals and agent dashboards that support OSC 7501 see whether Pi is working, blocked on a dialog or login, done, or failed. See Program status.

• Claude Haiku 5.5: anthropic/claude-haiku-5-5, with adaptive thinking up to xhigh/max effort.

• Adjust default tools with +name/-name: --tools entries like pi -t +codemode,-write change the default selection instead of replacing it. See Tools.

• GPT-6 Luna and image classification: OpenAI's GPT-6 Luna is available as a classifier model through the Decisions API, and codemode's models.classify() accepts images for classifiers that support them. See Use classifier models.

• Native llama.cpp decision models: Julia-1, Laya, Kev, lev, and OpenJev served by llama.cpp 0.6.0 or later run natively as classifiers through /v1/systemone. See Classification.

Added

• Added +name and -name entries to --tools, which change the default tool selection instead of replacing it, for example pi -t +codemode

• Added durationMs to the tool render context and to tool_execution_end extension events: the recorded execution time of a final tool result

• Added outputPad to the tool render context

• Added OpenAI's GPT-6 Luna as a classifier model through the Decisions API, available with OPENAI_API_KEY (see Use classifier models)

• Added images to codemode's models.classify() context, so classifiers that accept images, such as GPT-6 Luna, can judge them

• Added program status reporting with OSC 7501: terminals and agent dashboards that support it see whether Pi is working, blocked on a dialog or login, done, or failed. PI_PROGRAM_STATUS=1|0 overrides detection (see Terminal setup)

• Added aborted to agent_settled session, extension, and JSON events, so integrations can tell a cancelled run from a finished one

• Added Claude Haiku 5.5 (anthropic/claude-haiku-5-5), with adaptive thinking up to xhigh/max effort and prompt caching on Bedrock

• Added native llama.cpp decision models: Julia-1, Laya, Kev, lev, and OpenJev served by llama.cpp 0.6.0 or later are listed only as classifiers through /v1/systemone instead of as chat models (see Classification)

Changed

• Changed outputPad to also apply to ! command output, tool output, and summary blocks (#9946, #10557

• Changed pi mcp login --timeout to limit the whole sign-in, including requests to the authorization server, instead of only the wait for the browser

Fixed

• Fixed bash and PowerShell results losing Took after reloading a session, and the live Took including wall-clock steps; both now show the recorded execution time

• Fixed managed installs keeping every old release; pi update now keeps only the new release and the one it updated from (#10392, #10511

• Fixed standalone binaries loading .env, .env.local, and .env.development from the launch directory into Pi's environment

• Fixed !! command headers losing their dim color once output arrives

• Fixed the codemode description not marking searchTools(), describeTool(), and describeNamespace() as async, which led models to serialize the unawaited promise as {}

• Fixed codemode output items running together, so models could not tell where one text() or console.log() output ended and the next began. With several text items, each now starts with a ==> text N/M <== line, and console calls follow the other output in one <console_output> block with one line per call

• Fixed /mcp waiting for all servers to connect before opening; the manager now updates live and remains usable while enabling, reconnecting, or disabling servers

• Fixed images being dropped as "could not be resized" when running under node --watch on Node 24.19+ and 26.x, where Node posts its own messages on the image resize worker channel

• Fixed clipboard paste doing nothing in Termux, and failed copies there omitting the Termux:API install hint

• Fixed ! and RPC bash output keeping fragments of color codes, such as a stray m, when a code was split across output chunks

• Fixed MCP OAuth sign-ins that could not be cancelled while waiting on the authorization server and kept running after the session ended. The sign-in screen now cancels with Esc at every step, session shutdown aborts a running sign-in, and each request to the authorization server times out after 15 seconds

• Fixed shutdown waiting up to 15 seconds to refresh an MCP OAuth token that was about to expire, only to close the server's session

• Fixed the fullscreen text selection surviving session switches and other transcript rebuilds, which highlighted unrelated text in the new transcript (#9311, #10567

• Fixed OpenAI models on Bedrock ignoring the thinking level and always running at Bedrock's default reasoning effort (#9331, #10142

• Fixed model headers in models.json not overriding the originator and User-Agent headers of Codex requests

• Fixed server_busy and servers are currently busy provider errors ending the turn instead of being retried

• Fixed Mistral responses that end with finish_reason: "error" not being retried

• Reduced context-limit request failures by estimating input at 3.5 characters per token instead of 4 when calculating output limits

• Fixed Radius models disabled by an organization owner still being listed

• Fixed Anthropic browser login failing with "localhost refused to connect" when port 53692 is reserved or in use, for example by Hyper-V/WSL port exclusions on Windows: login now falls back to a free loopback port

• Fixed session costs undercounting long prompts on models with prompt-length pricing tiers, such as Claude Haiku 5.5, Gemini 3.1 Pro, and GPT-5.4, through OpenCode, OpenCode Go, OpenRouter, Vercel AI Gateway, Google, MiniMax, and other providers

• Fixed Markdown links not being clickable in Herdr

New Features

• Tool patterns and --no-mcp: --tools and --exclude-tools accept * patterns, for example --tools read,codemode,'mcp__radius__*' keeps only one MCP server's tools. --tools now keeps MCP tools unless an entry starts with mcp__, and --no-mcp turns off MCP for one run. See Tools and MCP tools.

• Codemode persists images: tools.read() on an image file now gives back an image block that image() can show. See Call tools.

Added

• Added * patterns to --tools and --exclude-tools, for example --tools read,codemode,'mcp__radius__*'

• Added --no-mcp to disable the built-in MCP support for one run

Fixed

• Fixed syntax highlighting losing colors after the first line of multiline strings and comments in fenced code blocks

• Fixed codemode scripts not receiving images from read: tools.read() now resolves to an image block for image files, which image() shows

• Fixed MCP OAuth sign-in failing with invalid_redirect_uri on servers with OpenID Connect client registration, such as mcp.modem.dev: pi now registers as a native client

• Fixed --tools removing MCP tools, which left pi --tools codemode without any MCP servers. --tools now keeps MCP tools unless an entry starts with mcp__

• Fixed MCP session shutdown returning while a server was still connecting, leaving its transport open until the server answered or timed out

• Fixed system prompt rules and the skills hint naming tools hidden by prepareLoadout. Hidden tools are left out of the rules, the skills hint names no tool when the file reader is hidden, and codemode shows each tool's prompt guidelines with its declaration; ToolLoadout gains getPromptGuidelines()

• Fixed Bedrock requests that fail with The pending stream has been canceled after a stalled HTTP/2 connection not being retried automatically

• Fixed codemode scripts that patch built-ins (for example Array.prototype.toJSON = ...) crashing pi and leaving the tool call unsettled. Built-ins are now frozen before the script runs, so such patches have no effect

New Features

• Azure Foundry Chat Completions — The azure provider (renamed from azure-openai-responses) now also serves Foundry Chat Completions deployments, starting with azure/deepseek-v4-pro. See Azure OpenAI.

• Codemode images saved to files — image() also writes each image to a temp file and names the path in the result, so later turns can copy or move generated images. See Generate images.

Breaking Changes

• breaking

  Renamed the Azure provider from azure-openai-responses to azure. Rename the provider key in auth.json (or run /login again), in models.json, and in settings.json (defaultProvider, enabledModels patterns, and modelThinkingLevels keys). Sessions that used the old provider fall back to another model when resumed, and their prompt cache is not reused. The AZURE_OPENAI_* environment variables are unchanged

Added

• Added Azure Foundry Chat Completions deployments, starting with azure/deepseek-v4-pro (#9645, #9714

Changed

• Codemode image() now also saves each image to a temp file and names the path in the result, so later turns can copy or move generated images

• Output files (full text of truncated tool output, binary MCP resources, codemode images) are now readable only by the user

• Home/End now always move the editor cursor to the line start/end; fullscreen transcript top/bottom moved to Ctrl+Home/Ctrl+End, which no longer move the editor cursor

Fixed

• Fixed subscription logins such as Sign in with ChatGPT failing with refresh_token_invalidated after a request was cancelled during an OAuth token refresh

• Fixed codemode failing for the rest of a session after a pnpm global update removed the running install, and added a restart hint when errors occur after pi was updated or removed on disk

• Fixed interactive sessions reporting a read EIO or setRawMode EIO crash (and asking to run /bug) when the terminal went away, e.g. after closing the window or resuming a suspended pi in a closed terminal

New Features

• Nix flake — nix run github:earendil-works/pi/stable runs the latest release, and nix profile add github:earendil-works/pi/stable installs it. See Install pi.

• Project overrides for MCP servers — .pi/mcp.json and /mcp can enable, disable, or change the exposure of a user-level server for one project. See Configure servers.

• MCP Client ID Metadata Documents — oauth.clientRegistration: "cimd" lets authorization servers allow pi by its document URL instead of dynamic registration. See Authenticate with OAuth.

• Tool renderers for any tool — pi.registerToolRenderer() draws calls to tools that are not registered yet, such as MCP tools in resumed sessions. See Tool rendering.

• Cloudflare Clef classifiers — @cf/cloudflare/clef and @cf/cloudflare/clef-flash are usable from codemode scripts and extensions. See Use classifier models.

Added

• Added a copy key (app.message.copy, default ctrl+x) to OAuth sign-in screens in /login, /mcp, and /mcp login, which copies the sign-in URL when the browser cannot be opened or the wrapped link cannot be selected.

• Added oauth.clientRegistration: "cimd" for MCP servers, which identifies pi with its Client ID Metadata Document on pi.dev instead of dynamic client registration, so authorization servers can allow pi by URL

• Added project overrides for user-level MCP servers: a .pi/mcp.json entry without command or url sets only enabled, exposure, and toolExposure of the user-level server, and /mcp can enable or disable a server for the current project

• Added Cloudflare's Clef and Clef Flash classifier models to cloudflare-workers-ai, usable from codemode scripts and extensions

• Added pi.registerToolRenderer(), which chooses how calls to a tool are drawn, including tools that are not registered

• Added a Nix flake for macOS and Linux: nix run github:earendil-works/pi/stable runs the latest release, and nix profile add github:earendil-works/pi/stable installs it. See Install pi

Changed

• pi update on global npm installations now recommends migrating to the managed installation from the pi.dev installer, which pins all dependencies.

• Anthropic tools added or redefined mid-conversation are now defined inline in the conversation, so redefining a tool under the same name keeps the prompt cache instead of resending the full tool list.

Fixed

• security

  Fixed installations resolving vulnerable brace-expansion 5.0.9 by pinning brace-expansion 5.0.12 as a direct dependency (GHSA-q2hr-2g5m-vwhr, GHSA-qhr7-859c-m2p7, GHSA-6j4f-fj2g-mc7p)

• Fixed a trailing comma in --models adding an extra model to the model cycle

• Fixed a codemode script that prints in a loop crashing pi by running out of memory: a script fails once its output passes 16 Mi characters or 100000 items

• Fixed JPEG, GIF, and WebP images rendered by extensions through Image not appearing in Kitty, Ghostty, WezTerm, and Warp

• Fixed MCP tool calls in resumed sessions and HTML exports rendering fully expanded until their server connected, or for good if it never did

• Fixed fullscreen Kitty images collapsing to a one-row strip after scrolling in WezTerm

• Fixed "Selected model is at capacity" provider errors ending the turn instead of being retried

• Fixed Cloudflare AI Gateway Claude models failing with a 404 by using dashed model IDs (claude-opus-5-5 instead of claude-opus-5.5)

• Fixed Sign in with ChatGPT continuing when its callback port is taken by another login, which made the browser show "OAuth state mismatch"; it now fails with a port-in-use error

• Fixed Amazon Bedrock OpenAI models costing requests above 272k input tokens at the short-context rate; Bedrock models now include the pricing tiers listed on models.dev

• Fixed Amazon Bedrock Claude requests failing with "Invalid signature in thinking block" after the system prompt or tools changed

• Fixed Together DeepSeek V4 Pro losing its thinking level controls after Together renamed it to deepseek-ai/DeepSeek-V4-Pro-0813

• Fixed the default NVIDIA model pointing at nvidia/nemotron-3-super-120b-a12b, which NVIDIA no longer serves; the default is now nvidia/nemotron-3-ultra-550b-a55b

Removed

• breaking

  Removed npm-shrinkwrap.json from the published package. npm installations no longer pin transitive dependencies, and library consumers can now override them. Use the pi.dev installer for pinned installations

New Features

• Fullscreen by default — The TUI now runs fullscreen. Set tuiMode to "regular" to keep the terminal's normal scrollback. See Terminal and display.

• Leaner codemode — About 40% fewer prompt tokens, and errors that tell the model how to recover. See Codemode.

• Image generation in codemode — Scripts call models.generateImages() with the session's credentials. See Generate images and Use image models.

• Radius in /login — Sign in with Radius and set up its MCP server in one step. See Radius.

• Anthropic copy code login — Sign in when the browser runs on another machine. See Authenticate interactively.

• MCP OAuth hardening — oauth.authServerMetadataUrl, RFC 9207 iss checks, credentials per server, and step-up sign-in that keeps granted scopes. See Authenticate with OAuth.

• Header-only quiet startup — quietStartup: "header" keeps the version and key hints and hides the rest. See Terminal and display.

Added

• Added an oauth.authServerMetadataUrl setting for MCP servers that advertise a wrong OAuth authorization server or none. Pi uses the configured metadata document instead of discovery (#10172).

• Added quietStartup: "header", which keeps the startup header with version and key hints but hides the model scope line and loaded-resource listing.

• Added models.generateImages() to codemode scripts. It runs image models such as OpenRouter's with the session's credentials and returns base64 image blocks that image() attaches to the result; usage counts toward the session cost like models.classify(). Extensions can call ctx.modelRegistry.generateImages(). See Use image models.

• Added a copy code login method to Anthropic /login for headless setups where the browser runs on another machine

Changed

• Changed the default TUI mode to fullscreen. Set tuiMode to "regular" or pass --tui-mode regular to keep the terminal's normal scrollback.

• /login now offers "Sign in with Radius" at the top level, as the last option, with its status. After a Radius sign-in, /login offers to configure the Radius MCP server in the global mcp.json with "auth": { "provider": "radius" } and reloads. Cancelling a login returns to the menu it was started from.

• The provider docs page is renamed to Providers, its "Cloud Providers" section is now "Provider Specific Config", and it documents Radius first.

• MCP OAuth credentials are now stored per server name and URL, so MCP servers with the same URL can sign in with different accounts. Credentials stored by URL alone move to the first server that uses them (#10252).

• Codemode costs far fewer prompt tokens: with the default tools and codemode active, a GPT-5.6 request shrinks from about 5,300 to 3,300 tokens. The codemode description lists the script globals in one line each and points to the new Codemode reference for the models API, which the model reads when it needs it. Declared tools say in one line how scripts call them and what the call resolves to, instead of repeating their full declaration, and the system prompt's codemode guidance and MCP server section are shorter.

• Codemode errors now say how to recover: reading a tool or models member that does not exist names the close matches (tools.Bash suggests tools.bash), models.classify() and models.generateImages() reject malformed arguments with the expected shape, an unknown model points to models.getAvailableOfType(), an oversized store() value explains what the store is for, and a script that generates images without showing them gets a note. Scripts that probed for a tool with typeof tools.name must use "name" in tools.

• /login and /logout now label providers without credentials as "not configured" instead of "unconfigured".

• OAuth browser pages now show the color Pi logo.

Fixed

• Fixed MCP OAuth sign-in accepting an authorization response whose iss parameter names another authorization server; the code is now rejected before it is exchanged (RFC 9207).

• Fixed MCP OAuth sign-in failing with Invalid scope when the token response contains "scope": "", and similar failures for other empty or null optional OAuth fields (#10266).

• Fixed the sign-in URL printed by /mcp login not being clickable when it wraps (#10186).

• Fixed --provider without --model being silently ignored and running the default model from another provider; it now fails with an error (#10236).

• Fixed MCP servers that ask for more scope (insufficient_scope) requesting sign-in over and over. The new sign-in requested only the missing scopes, so the new token lost access the previous one had; it now keeps the granted scopes.

• Fixed user messages in the transcript keeping two full-width copies of every rendered line; they keep one, with identical output.

• Fixed /login and /logout labeling every OAuth sign-in, including Radius, as a subscription; only subscription-backed providers say "subscription", other OAuth sign-ins say "account".

• Fixed the startup header logo rendering with gaps in Apple Terminal; it now shows a colored "Pi" with the version instead.

• Fixed deferred MCP tools that tool_search loaded being dropped on resume and /reload even when their server reconnected before the next prompt, because the session restored its tools before the MCP servers reconnected.

• Fixed the system theme making pastel palettes such as Catppuccin Frappe much more vivid; palette colors now keep their chroma (#10255, #10293

• Fixed slash command autocompletion not triggering when the input starts with whitespace

• Fixed color bleeding past mouse selections and search highlights in fullscreen mode when a styled token ends at the highlight boundary (#10169).

• Fixed memory retained per rendered message in the transcript; a long assistant message keeps about a fifth of the heap it kept before.

New Features

• MCP servers stay out of the way: servers with the default codemode exposure are no longer listed in the codemode description and no longer block the first prompt. They appear in a short system prompt section, and scripts find their tools with searchTools() and describeNamespace(). See Control tool exposure.

• More MCP authentication options: oauth.clientName for servers that only accept known OAuth clients, and "auth": { "provider": "<provider>" } to authenticate HTTP servers with a provider's /login token. See Authenticate with OAuth.

• Anthropic workload identity federation from the Anthropic SDK environment variables. See Use an API key from the environment.

• /reload enables tools newly added to the defaultTools setting. See Tools.

Added

• Added a description field for MCP servers (pi mcp add --description), shown with the server in the system prompt and used to rank its tools in tool search, and a describeNamespace(name) codemode helper that returns a namespace's instructions and tool names. describeNamespace() and searchTools() accept a namespace as mcp__dev-radius, mcp__dev_radius, dev-radius, or dev_radius.

• Added an oauth.clientName setting for MCP servers (pi mcp add --oauth-client-name) to change the client name sent during OAuth client registration, for servers that only accept known clients (#10226).

• Added "auth": { "provider": "<provider>" } for HTTP MCP servers to send a provider's current /login token as the bearer token instead of using MCP OAuth. The token is read on every request, so provider refreshes apply. Only allowed in the global mcp.json and from extensions, and requires https except on loopback hosts.

• Added Anthropic workload identity federation from the ANTHROPIC_FEDERATION_RULE_ID, ANTHROPIC_ORGANIZATION_ID, and ANTHROPIC_IDENTITY_TOKEN_FILE environment variables (see Providers) (#10177, #10242

• /reload now enables tools newly added to the defaultTools setting. Tools removed from it stay enabled, tools turned off during the session stay off unless newly added, and --tools, --no-tools, and --no-builtin-tools still override the setting (#10245).

Changed

• MCP servers with the default codemode exposure no longer appear in the codemode description; scripts find them with searchTools(). codemode-deferred is now an alias for codemode. Use direct exposure for tools the model should see without searching (#10212).

• The codemode description no longer includes deferred tools, tool counts, or MCP server instructions, so it no longer changes when MCP servers connect or change their tools. The tool_search description no longer lists the servers whose tools it can load, for the same reason. Servers are listed instead in an mcp_servers system prompt section with a one-line summary, updated at the start of each prompt; a changed section is appended to the conversation. Scripts read server instructions with describeNamespace() (#10212).

• The first prompt no longer waits for MCP servers without direct tools. They connect in the background and are waited for when a codemode script names them, a script searches tools, or tool_search runs (#10212).

Fixed

• Fixed new sessions intermittently ignoring the saved default model, or warning that no models are available, when it belongs to an extension-registered native provider with a stored credential (#9962, #10190

• Fixed the /mcp sign-in URL not being clickable when it wraps across lines, by emitting it as a terminal hyperlink with a Cmd/Ctrl+click to open line like /login (#10186).

• Fixed codemode image() accepting malformed base64 data or unsupported image types, which persisted an invalid image block that made every later provider request fail with HTTP 400 (#10215).

• Fixed codemode failing to start its script worker from the standalone Windows executable (#10204).

• Fixed prompt submission slowing down with session length, because resolving the session's model selection looked up the model catalog once per assistant message (#10198).

• Fixed model lookups slowing down for providers with a refreshed pi.dev catalog, because merging remote catalog models took quadratic time.

• Fixed the built-in-tool-renderer.ts and minimal-mode.ts extension examples removing the built-in tools' summaries and guidelines from the system prompt (#10072, #10193

• Fixed context overflow detection for Z.AI CN endpoint Prompt exceeds max length errors (#10208).

• Fixed Anthropic requests failing when a tool schema uses keywords Anthropic strict tool use rejects, such as minimum/maximum; such tools are now sent non-strict (#9953).

• Fixed provider retries firing immediately when a Retry-After header contains an unparseable date; they now use exponential backoff (#9571).

• Fixed extension commands registered without a string name or handler crashing pi when typing /; the extension now fails to load with an error instead (#10054).

• Fixed collapsed codemode and MCP tool results filling the screen when the output is one long line, such as minified JSON. Like bash output, the preview is now limited to wrapped lines instead of logical lines.

• Fixed codemode.mode: "only" listing read, bash, edit, and write in the system prompt's tool list although requests only declare codemode (#10192).

• Fixed codemode scripts calling the wrong MCP tool when two tool names differ only in - and _, such as read-file and read_file. Like in Codex, MCP tool and namespace names now replace - with _ (mcp__my-server__x is now mcp__my_server__x), colliding tools of a server all get a hash suffix, and server names that differ only in - and _ are rejected (#10239).

New Features

• GPT-6.1 Sol — Available on OpenAI, Azure OpenAI, and OpenAI Codex, and now the default OpenAI Codex model. See Select a model.

Added

• Added GPT-6.1 Sol (gpt-6.1-sol) to the OpenAI, Azure OpenAI Responses, and OpenAI Codex providers.

Changed

• Changed the default OpenAI Codex model to GPT-6.1 Sol (gpt-6.1-sol).

Fixed

• Fixed /login with OpenAI failing in the bundled release with a missing openai-chatgpt.js module error.

New Features

• Codemode and MCP — Connect MCP servers and let models run JavaScript that calls tools in parallel. See MCP Servers and Enable codemode.

• System theme — Pi's colors now come from your terminal's own palette by default. See Use your terminal's colors.

• Sign in with ChatGPT — Use a ChatGPT subscription with the OpenAI provider through /login openai. See Authenticate interactively.

• Virtual models — Extensions can route each request to a different physical model. See Virtual Models.

• Classifier models — Run Jev classifiers from codemode scripts, or use any llama.cpp model as a classifier. See How codemode works and Classification.

Added

• Added codemode, tool search, and MCP support as built-in extensions. The codemode tool runs model-written JavaScript in a QuickJS sandbox that calls pi's tools; enable it with defaultTools or --tools and configure it with codemode.mode and codemode.inlineBudget. tool_search finds tools that are not declared to the model and declares them. MCP servers over stdio or streamable HTTP, with OAuth, come from mcp.json (global, or per project once trusted) or pi.registerMcpServer() and are managed with /mcp and pi mcp add|remove|list|login|logout. See MCP Servers and Enable codemode (#10040).

• Added extension tool APIs for orchestrating tools: exposure (direct, model-only, codemode, deferred, or hidden), namespace, annotations, outputSchema with structuredContent, isError results, prepareLoadout(), and ctx.executeTool() for nested tool calls, which emit events with parentToolCallId and are recorded as bounded nestedCalls on the calling tool's result. See Tool exposure.

• Added a warning when an extension that registers the same tool, command, or flag replaces a built-in extension

• Added experimental virtual models: extensions register them with pi.registerVirtualModel() and pick a physical model and thinking level for each request. The footer shows the routed model, /session lists cost per physical model, and examples/extensions/jev-router.ts routes with the Jev classifier. See Virtual Models.

• Added Sign in with ChatGPT to the OpenAI provider in /login, which uses a ChatGPT subscription with the OpenAI API. Pi stores a stable deviceId in the global settings for this login and omits it from bug reports.

• Added the system theme, now the default, which derives pi's colors from the terminal's reported foreground, background, and ANSI palette and rebuilds them when the terminal switches between light and dark. See Use your terminal's colors.

• Added #rgb, oklch(), and okhsl() colors and an optional appearance field to theme files, and theme.style(), theme.colors, and theme.appearance for extensions. See Themes and TUI.

• Added a classifier model for every llama.cpp chat model, answered from next-token label probabilities. See Classification.

• Added inherited Jev classifier models on OpenRouter, Cloudflare Workers AI, Vercel AI Gateway, and OpenCode Zen.

• Added the fullscreenWheelScrollLines setting and /settings entry for fullscreen mouse-wheel scrolling. The default "auto" accelerates fast wheel spins outside local macOS terminals (#9758).

• Added per-input disposition to successful RPC prompt, steer, and follow_up responses, AgentSession.steer()/followUp(), and RpcClient.prompt()/steer()/followUp(); RpcClient.prompt() also accepts streamingBehavior (#9098, #9803).

• Added image generation to ModelRuntime: generateImages() with runtime-resolved auth (stored credentials, OAuth, runtime API keys, models.json headers), plus getModelsOfType(), getModelOfType(), getAvailableOfType(), getAllModels(), and getAllAvailable(). OpenRouter image models are listed under the openrouter provider and share its credential; an upstream ID can have separate chat and image entries. models.json providers and extension registrations without a model list keep built-in image generation. Extension model lists can include discriminated chat, image, and classifier entries with operation implementations; when supplied, they replace the provider catalog across every operation. Chat-facing reads (getModels(), getAvailableSnapshot(), the model picker) are unchanged.

• Added classifier support to ModelRuntime, including classify(), classifier model accessors, runtime-resolved authentication, and the built-in TypeSafe jev-latest model.

• Added types=chat,image,classifier to pi.dev model catalog requests so remote refreshes overlay every supported model type; entries of unknown model types are ignored.

• Added the provider_stream_event extension event for observing parsed provider events before normalization, with an opt-in /debug-provider example viewer (#9784, #9901

• Added a show/hide toggle (H) in HTML exports for custom messages marked display: false. Messages remain hidden by default and can also be revealed from the sidebar (#8896, #10020

• Added inherited Claude Sonnet 5.5 support for Anthropic with adaptive thinking and a 1M context window.

• Added a Built-in section in pi config to disable the built-in mcp, llama.cpp, codemode, and tool-search extensions globally or per project, stored as -builtin:<name> in the extensions setting. SDK inline extensions opt in with builtin: true.

• Added +name and -name entries to the defaultTools setting to add or remove tools without repeating the defaults, for example "defaultTools": ["+codemode"]. Project entries of this form apply on top of the user setting. Documented how to enable codemode without MCP and how to use classifier models such as Jev from codemode scripts.

• Added the token usage and cost of codemode models.classify() calls to the codemode tool result, so they count toward the session cost; the codemode result shows each call's cost.

Changed

• Switched the build from the TypeScript native preview to TypeScript 7.0 with an ES2024 target, and replaced tsx with Node's built-in type stripping for running from source (#9965).

• Removed the [Themes] section from the startup banner. Custom themes remain available in /settings, and theme conflicts are still reported.

• Changed the startup header to show the pi logo with the version instead of the app name.

• Changed the built-in dark and light themes to the revised pi colors, written in OKHSL.

• Changed light/dark terminal detection to use the reported background color first, then the terminal's light/dark report, then COLORFGBG. The first-time setup no longer shows the detected appearance.

• Renamed the inherited OpenAI Codex provider to "OpenAI Codex (legacy)"; Sign in with ChatGPT on the OpenAI provider supersedes it.

• Changed inherited terminal detection to treat TERM=*-direct as truecolor.

• Built-in extensions and tools are named builtin:<name> (for example builtin:mcp and builtin:read) in errors, diagnostics, RPC source info, and bug reports, instead of <inline:name> and <builtin:name>. Their slash commands no longer carry a [t] autocomplete tag.

• --no-extensions also disables the built-in extensions, including the llama.cpp provider. Load one explicitly with -e builtin:<name>, for example pi -ne -e builtin:mcp.

• Tool calls without a custom call renderer, including direct MCP tool calls, now show their arguments: as key=value pairs on the title line when collapsed and one key: value line per argument when expanded. MCP calls are titled server/tool and their results collapse to 5 lines.

• bash and powershell structured results, which codemode scripts receive, now hold up to 1 MiB of output instead of the model-facing 2000 lines or 50KB, and add truncated and full_output_path. Longer output keeps its first and last 512 KiB. Empty output is "" instead of (no output).

Fixed

• Fixed X11 clipboard text being misidentified as an image when the clipboard owner accepts unadvertised image targets (#9786).

• Prevented managed git packages from automatically installing Pi peer dependencies and added warnings for extension packages that list host-provided modules in dependencies (#9863).

• Fixed pinned git extensions loaded with -e continuing to use the first downloaded commit after the ref changes (#9982).

• Fixed RpcClient skipping the next event listener when a listener unsubscribes while handling an event, which could make waitForIdle() time out after collectEvents() (#9990).

• Fixed full-file read calls rendering as :1 when models send null for omitted offset and limit (#9996).

• Fixed new sessions being lost when pi exits before the first assistant response. The session file is now created when the first user message is sent (#10000).

• Fixed unloaded llama.cpp autoload presets overwriting a cached runtime context window with the GGUF training context (#10077, #10158

• Fixed custom themes ignoring terminal.trueColor and other terminal capability overrides and rendering with 256 colors (#9973, #10039

• Fixed pasting files copied in Finder inserting the file icon image instead of the file paths; paths are quoted in bash mode (#9999, #10136

• Fixed the startup header, loaded resources, and chat notices keeping their old colors after a theme change.

• Fixed the Fireworks default model pointing at the removed Kimi K2.6 model; it now defaults to Kimi K3.

• Fixed the OpenCode Go default model pointing at the removed Kimi K2.6 model; it now defaults to Kimi K3.

• Fixed the Together default model pointing at the removed Kimi K2.6 model; it now defaults to Kimi K3.

• Reduced CPU use while streaming in long sessions and when previewing themes: the footer caches session usage totals, collapsed bash results cache their preview, and sanitizeBinaryOutput() no longer splits output into per-character arrays.

• Fixed the usage of tools called through ctx.executeTool(), for example from codemode scripts, being dropped from the session cost; it is now added to the calling tool's result usage.

• Fixed inherited /skill autocomplete appearing empty when loaded skill names did not contain the letters in skill (#9944).

• Fixed inherited path and @ autocomplete not working after opening wrappers such as (, [, {, <, or a backtick.

• Fixed inherited image stretching in terminals that use the Kitty graphics protocol (#8938, #9957

• Fixed inherited shell cursor staying hidden after exit when an extension closed an overlay during shutdown (#10026).

• Fixed inherited keyboard input being lost after a mouse click in a /settings submenu closed it.

• Fixed inherited 1-hour Anthropic cache writes through Vercel AI Gateway being priced at the 5-minute rate (#9210).

• Fixed inherited model-level samplingParams being dropped by direct stream()/complete() calls on OpenAI-compatible APIs (#9506).

• Fixed inherited Mistral GLM requests failing with "Expected at most one leading ThinkChunk" after empty content deltas (#9674).

• Fixed inherited OpenAI Fast mode requests being priced at the standard rate (#10034).

• Fixed inherited Mistral reasoning models ignoring the requested thinking level (#9678).

• Fixed inherited OpenCode Zen and OpenCode Go qwen3.8-flash thinking being replayed as plain text on later turns (#10047).

• Fixed inherited OpenAI Responses streams from servers that omit output_index, such as llama.cpp, running mixed-up tool calls; such streams now end with an error (#9974).

• Fixed inherited Anthropic and OpenAI Codex browser sign-in waiting indefinitely after the provider redirected with an authorization error, and Anthropic sign-in failing when its callback port is in use.

• Fixed inherited GitHub Copilot Claude Opus 5.5 offering unsupported thinking levels when upstream model metadata is incomplete.

New Features

• Latest frontier models — Use Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna through supported providers, including GitHub Copilot. See Choose a Model.

• Grok 4.7 by default for xAI — New xAI sessions now default to Grok 4.7. See Provider Authentication.

Added

• Added inherited Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna support for GitHub Copilot.

• Added inherited GPT-6 Sol and GPT-6 Luna support for OpenAI API keys and OpenAI Codex subscriptions.

• Added inherited Claude Opus 5.5 support for Anthropic with adaptive thinking and a 1M context window.

Changed

• Changed the default xAI model to Grok 4.7.

Fixed

• Fixed split-turn compaction summaries being refused by Claude Fable 5.1 by clearly separating the conversation and using continuation-oriented instructions

• Fixed missing or invalid --mode values being silently ignored instead of reporting an error and exiting with a nonzero status (#9045).

• Fixed inherited image-only user messages being rejected by some OpenAI-compatible providers because they included an empty text part (#9797).

• Fixed inherited Anthropic OAuth requests reporting an outdated Claude Code version.

New Features

• Canonical session context and extension boundaries — Edit model context without rewriting history and add actionable lifecycle hooks. See ContextEditEntry and extension events.

• Full-transcript context extensions — Use context_with_system for per-request system-message transformations. See context_with_system.

• Per-model image input limits — Configure cache-safe image resizing per model for attachments, read, and tool-result images. See Image Input Limits.

Breaking Changes

• breaking

  Removed the inherited shouldStopAfterTurn agent option. Use finishTurn and return { action: "end" } instead. finishTurn runs before turn_end but applies the decision afterward, and it also receives error and aborted responses; migrate normal-response predicates by returning undefined for those hard exits. See the @earendil-works/pi-agent-core changelog for a complete before-and-after example.

• breaking

  Added ContextEditEntry to the exported SessionEntry union. TypeScript consumers with exhaustive entry switches must handle context_edit; use replacement: null for omission and a content replacement otherwise.

• breaking

  Made SessionManager canonical for AgentSession provider context. Assigning session.agent.state.messages no longer replaces future request history; restore with SessionManager.inMemory(cwd, { id }, entries), navigate with session.navigateTree(), or append through session.sessionManager and call session.refreshContext().

• breaking

  Expanded TurnEndEvent with required boundary fields and added AgentBeforeSettleEvent to the exported ExtensionEvent union. Consumers constructing events or exhaustively switching on ExtensionEvent must handle the new shapes. ExtensionRunner.emit() no longer accepts turn_end; host integrations dispatch actionable boundaries with emitBoundary(baseEvent, buildContext).

• breaking

  Deferred runs requested from agent_settled handlers until all settled handlers finish. Handlers still observe ctx.isIdle() === true, but no longer see a reentrant agent_start during the same notification dispatch.

Added

• Added append-only model-context edits. For example, sessionManager.appendContextEdit(entryId, null) omits one message from future provider context without changing raw history, usage, or UI history.

• Added actionable turn_end and agent_before_settle extension boundaries. Return { entries: [...event.entries, draft], continue: true } to persist structural entries in order and ensure one next provider request without changing steering or follow-up scheduling.

• Added retain-none compaction input: sessionManager.appendCompaction(summary, null, tokensBefore) stores the compaction's own ID as its kept boundary.

• Added the context_with_system extension event, which runs after context handlers on the full transcript including system messages and sends its result verbatim. See context_with_system.

• Added per-model image resize profiles through inputLimits.images.resize in models.json, applied to file attachments, image reads, and tool-result images (#9631).

Fixed

• Fixed string context-edit replacements producing invalid assistant and tool-result message content instead of text blocks.

• Fixed context-invisible boundary metadata and replacement edits causing newly appended or replaced input to be summarized before its first provider request.

• Fixed edited-context accounting both discarding valid assistant usage captured after the latest context edit and reusing that usage after a later compaction made it stale.

• Fixed selected error retries and final length/overflow recovery retaining abandoned model attempts in future provider context; post-run recovery omissions are now persisted without hiding raw transcript history or changing queue scheduling.

• Fixed context handlers that filter or slice messages dropping the prompt and tool declarations, which after extension-driven compaction left requests without built-in tools or made Codex emit raw tool-call text. Handlers no longer see system messages; Pi restores the prompt and tool state after they run. See context (#9789, #9822).

• Fixed /bug allowing uploads in offline mode while preserving local zip exports

• Fixed idle prompt-cache warming rebuilding expired caches when its timer or an extension decision is delayed.

• Improved crash diagnostics with hints identifying loaded extensions that appear in the stack trace.

• Fixed text files beginning with GIF being misclassified as images and omitted from read and CLI @file input (#9755).

• Fixed malformed prompt template frontmatter being silently ignored instead of reported as a resource warning

• Fixed inherited unknown OpenAI-compatible Chat Completions endpoints receiving strict tool schemas unless they explicitly advertise support (#9816).

New Features

• Meta Muse provider — Sign in with Meta using /login meta or use META_API_KEY to access Muse Spark models. See Meta (Muse subscription).

Added

• Added Meta (Muse subscription) login via /login meta with automatic Model API key refresh, plus META_API_KEY support

Changed

• Enabled Node's persistent compile cache before loading the bundled CLI runtime, reducing repeat launch time.

Fixed

• Fixed /bug descriptions dropping line breaks from pasted diagnostics.

• Fixed /bug hints appearing for user cancellations and retryable provider failures such as service unavailability.

• Fixed clipboard copy failing in containers and WSL without WSLg by restoring the OSC 52 fallback when no display is available, and added a verified Windows clipboard backend for WSL (#9688).

• Fixed inherited z.ai Prompt too long errors not being recognized as context overflow (#9805).

• Fixed inherited Cerebras models advertising unsupported strict tool schemas, which caused HTTP 400 errors when strict and non-strict tools were mixed

New Features

• Prompt cache warming — Keep valuable prompt caches alive during long tool runs and optionally while idle using cost-aware refreshes. See Cache Warming.

• Bug reporting — Report problems with /bug using redacted diagnostics, optional transcripts, or exported ZIP archives. See Reporting Bugs.

• Transcript-aware prompt and tool updates — Preserve instruction and tool changes across resume and branch navigation while retaining cached prefixes. See before_agent_start.

• Offline Radius model catalog — Select Radius models immediately, with cached and live catalogs overlaid when available. See Radius.

• Per-model compaction budgets — Configure reserved and recent-token budgets by model. See Per-model overrides.

Breaking Changes

• breaking

  Changed inherited pi-ai provider stream inputs from Context to normalized TranscriptContext values. Custom providers must read system prompts and tool declarations from context.messages with getCurrentSystemPrompt() and getCurrentTools(). See Custom Streaming API.

• breaking

  Restricted inherited ToolCall.arguments and ToolResultMessage.details to JSON-compatible values, changed ToolResultMessage into a conditional type, and made JsonValue arrays readonly.

• breaking

  user_bash now fails closed: errors or invalid defined results abort the command without invoking later handlers or executing locally. Return undefined to continue propagation; otherwise return { operations } or { result } (#9068).

Added

• Added transcript-backed mid-conversation system prompt and tool changes so instruction and tool updates survive resume and branch navigation while preserving cached prefixes on supported models. See before_agent_start and Entry Types (#9548).

• Added inherited native deferred tool loading for Fireworks Messages models. Use ToolSearch or tool_search as the loader name for prompt-prefix deferral (#9323).

• Added click toggling for branch summaries, compaction summaries, and skill invocation entries.

• Added the public Radius model catalog for immediate and offline model selection, with cached and live gateway catalogs overlaid when available.

• Added ctx.modelRegistry.stream() and streamSimple() for extension model calls through configured providers with resolved authentication (#8964).

• Added per-model reserveTokens and keepRecentTokens settings through compaction.modelOverrides, with ordinary compaction settings as fallback (#8133).

• Added compat.allowedFallbackModels configuration for overriding or disabling Anthropic server-side fallback models (#9294).

• Added an unsubscribe function from pi.on() so extensions can drop event handlers. Handlers added or removed during a dispatch apply to later dispatches, not the current one (#8967).

• Exported extension hook event and result types that were previously omitted from the package entry points (#9642).

• Added /bug [description] to report a bug to the Pi developers. The report bundles environment, model, provider, extension, and settings metadata (secrets redacted), assistant message diagnostics from the session, optionally the session transcript, or a model-written summary of what went wrong instead. It is uploaded to Radius (no login required; attributed when logged in) or exported as a zip archive, and the report id is recorded in the session as a pi.bug-report entry. Crashes are recorded in ~/.pi/agent/crashes.json, announced once on the next start, and attached to the next report; unexplained errors and exhausted retries point at /bug once per session.

• Added cost-aware prompt-cache warming during long tool runs and optionally while idle, with configurable modes, model cache-lifetime metadata, /session diagnostics, transcript notices, and the cache_warming_decision extension event. See Cache Warming (#9668).

Changed

• Made --resume session results appear progressively, using file modification times to prioritize all-folder loading and cancelling outstanding transcript reads after selection.

• Reduced --continue startup time by checking candidate session headers in modification-time order and stopping after the newest matching session.

• Replaced the external native clipboard dependency with bundled asynchronous macOS, Windows, and X11 helpers while preserving platform command and OSC 52 fallbacks (#9163).

• Reduced inherited fuzzy search latency for long texts by using native substring search instead of scanning each character in JavaScript (#9267).

• Moved compaction, branch summarization, and retry spinners into the editor border alongside the working indicator. Custom editors use the same embedding opt-in for all status spinners.

• Enabled strict-prefer JSON-schema sampling by default for built-in read, bash, powershell, edit, and write tools, without requiring PI_EXPERIMENTAL. Extensions can re-register tool definitions with constrainedSampling: false.

• Formatted Bash and PowerShell tool durations of at least one minute as minutes and seconds, with hours when needed (#9628).

• Deferred the extension compiler and bundled virtual modules until a filesystem extension is loaded, reducing the baseline SDK import cost (#9540).

Fixed

• Fixed GitHub Copilot GPT models, including GPT-6 Astra, using the Chat Completions adapter instead of the required Responses adapter

• Fixed inherited DeepSeek V4.1 thinking levels on OpenRouter and OpenCode Go preserving provider effort metadata (#9485).

• Fixed inherited bodyless HTTP 400/413 errors from non-Cerebras providers being misclassified as context overflow (#9482).

• Fixed inherited Vercel AI Gateway replaying unsigned thinking as assistant text (#9676).

• Fixed inherited Google Generative AI and Vertex AI using unsupported thinking levels when reasoning is omitted or when model capabilities differ within a Gemini family (#9455).

• Fixed inherited Anthropic-compatible relays breaking signed thinking replay when they report a different response model, while preserving fallback pricing (#9188).

• Fixed inherited Amazon Bedrock one-hour cache writes being priced at the five-minute rate (#9457).

• Fixed inherited quadratic CPU usage when draining buffered EventStream events (#9055).

• Fixed inherited Mistral Medium reasoning requests to use reasoning_effort for all reasoning-capable mistral-medium-* model IDs instead of the unsupported prompt_mode (#8700).

• Fixed inherited OpenCode and OpenCode Go requests to send x-opencode-session from sessionId across all supported API adapters (#9326).

• Fixed inherited OpenAI Codex requests to send the model's Off reasoning effort instead of omitting it, while respecting unsupported Off mappings (#9191).

• Fixed inherited Fireworks unsigned thinking replay and reasoning effort selection using catalog metadata, with verified DeepSeek V4 and Qwen3.8 fallbacks and removal of redundant GLM 5.2 and Kimi K3 effort aliases (#9323).

• Fixed inherited OpenRouter requests to send x-session-id from sessionId for Chat Completions and Anthropic Messages models when prompt caching is enabled (#9102).

• Fixed the inherited DeepSeek catalog to advertise deepseek-flash for DeepSeek V4.1 Flash instead of retired Flash aliases, and refreshed DeepSeek pricing metadata (#9423).

• Fixed inherited Mistral-hosted GLM-5.2 reasoning requests to use reasoning_effort instead of the ignored prompt_mode (#9375).

• Fixed inherited OpenAI-compatible Responses errors to identify the actual provider instead of always labeling them as OpenAI errors (#9298).

• Fixed inherited Baseten requests to send session-affinity headers from sessionId for automatic prompt-cache routing (#9629).

• Fixed inherited retry classification for Cloudflare 520 responses (#9627).

• Fixed inherited retry classification for transient Azure peak-load capacity errors (#9669).

• Fixed session tree navigation racing with active compaction and replacing its progress UI

• Fixed exact session ID lookup scanning complete transcript bodies instead of reading session headers

• Fixed repeated Anthropic thinking-drop notices being shown for the same dropped blocks, and shortened notices while retaining details in the session (#9391).

• Fixed mid-run threshold compaction silently skipping oversized trailing tool results (#9740).

• Fixed signal-terminated local shell commands being reported as successful with partial output

• Fixed local clipboard failures reporting success when the terminal ignored the fallback OSC 52 write, and added platform-specific setup guidance when no clipboard backend works (#9618).

• Capped agent-level retry backoff at retry.maxAgentDelayMs (60s by default) so long retry runs stay responsive during prolonged transient outages (#8826).

• Fixed direct RPC steer and follow_up commands bypassing extension input handlers (#8718).

• Fixed premature missing-model errors after login by waiting for catalog discovery. Radius now defaults to balanced, falling back to the first available Radius model when needed.

• Fixed fullscreen mode reserving a blank row for custom footers that render zero rows (#8919).

• Fixed extension tools without parameter schemas to be rejected during registration instead of breaking provider requests (#9300).

• Fixed before_agent_start handlers returning systemPrompt (and forceSystemPrompt) on models with mid-conversation system messages: the forced prompt is now sent as the provider's leading system prompt instead of being appended as a section patch after the original prompt.

• Fixed loaded llama.cpp models with enable_thinking chat templates ignoring Pi's thinking level (#9528).

• Fixed cancellation races that could start automatic compaction, leave stale retry state, or miss cancellation while waiting for summarization authentication (#9340, #9777).

• Fixed asynchronous Kitty image conversion replacing newer partial tool output images

• Fixed inherited skill slash-command autocomplete ranking the skill: prefix instead of the bare skill name

• Fixed inherited file autocomplete boundaries and path quoting around CJK punctuation

• Fixed inherited LaTeX legacy font switches falling back to raw source, centered cases layouts around surrounding equations, and vertically laid out unsupported and nested display scripts (#8827, #9564, #7929).

• Fixed inherited fullscreen Kitty images being erased by later row clears in WezTerm (#9169).

Removed

• breaking

  Removed unavailable inherited GPT-5.4 and GPT-5.4 mini models from OpenAI Codex selection (#9394).