@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