Oh My Pi changelog: what's new each month
Every stable Oh My Pi release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.
November 2026 · so far
18 releases: 18.15.0 → 18.20.3
November brought major provider and model additions, improved conversation handling for several platforms, and numerous reliability fixes across authentication, reasoning, and tool execution. The month saw support for Liquid AI, Otari, Modular Cloud, and Antigravity CLI alongside significant improvements to Gemini, Claude, and OpenAI-compatible model support.
- Removed PI_TOKENIZER_ACCURATE environment variable; all models now counted with native tokenizer.
- Deprecated formatError on embeddings, images, rerank, speech, transcriptions, video, pi-native and SystemOne server modules.
- Removed public isXdDevicePath export from providers/cursor-pi-args.
- Changed browser open API: ignore_https_errors renamed to insecure, allow_file_access removed.
Highlights
- Added support for Liquid AI decision models and Google Decisions API for judgment tasks.
- Introduced Antigravity CLI provider with managed installation and native thinking controls for Gemini, Claude, and GPT models.
- Improved Gemini 4 response handling across multiple platforms to support long outputs that exceed single segments.
- Added opt-in standard Responses WebSockets for incremental turns with automatic HTTP/SSE fallback.
- Redesigned ask forms to use HTML with support for images, free-text "Other" options, and form notes.
New
- Modular Cloud (modular) added as a provider with per-model reasoning controls.
- Devin CLI seat key and Windsurf Enterprise sign-in logins now supported.
- Otari gateway provider added with live model discovery per workspace.
- Bash tool cards now show hoisted cd commands and run commands in the correct working directory.
- Subagents receive task instructions with formatting preserved including blank lines and table spacing.
- LiteLLM adapter added for gateway key and user USD budget reporting.
- Tern gains long-press menu on phone for editing, retrying, rewinding, and forking sessions.
- Eval cells can run shell commands using sh() in Python and JavaScript.
Improved
- SambaNova Cloud multi-turn conversations now reuse prefix cache, avoiding reprocessing earlier context.
- Anthropic official-API sessions use message threads, sending only new messages per turn for faster first tokens.
- Claude auto-redeem now spends a full saved reset when account is blocked by weekly limit if reset clears Fable cap.
- Factory Droid Beam requests now accept tool-call turns without reasoning.
- Context estimates no longer charge readers for another model's thinking signatures or redacted thinking.
- MCP headers and stdio env values using !command now rerun on 401/403 and /mcp reconnect.
- Browser role selectors in Chrome tabs now match accessibility tree reports and skip hidden elements.
- Google Antigravity quota recovery improved for Gemini 4 Argon access resumption.
Fixed
- Fixed OpenAI remote compaction replaying backend-only metadata that Codex rejects.
- Fixed Google Vertex requests with Workforce Identity Federation ADC failing token refresh.
- Fixed Anthropic OAuth accounts being signed out when background token refresh got invalid_grant.
- Fixed Gemini CLI and Antigravity responses losing or misordering thought signatures and reply text.
- Fixed Gemini 4 Argon sharing general Gemini quota scope instead of dedicated weekly bucket.
- Fixed concurrent OAuth requests rejected as claude_code_version_too_old failing to retry with required version.
October 2026
48 releases: 18.4.6 → 18.14.2
October brought substantial improvements to model support, caching efficiency, and agent steering workflows. The month featured new Claude models, prompt cache enhancements, and numerous fixes for tool execution and provider compatibility.
- Removed display() from Python and JavaScript eval; use print and console.log instead.
- Removed per-call model overrides from task, eval agent(), and workpool(); subagents use configured models.
- task.completionProbeMs replaced by on/off setting task.completionProbe; existing values migrate automatically.
- SessionStorage.claimSessionFile() replaced by claimSession(); sessionOwnerLeasePath() replaced by new function.
Highlights
- Added Claude Haiku 5.5 with adaptive thinking support and 1M-token context window.
- Implemented warm prompt-cache awareness for conversation pruning and native compaction on Bedrock Claude models.
- Improved tool-call parsing and execution reliability across multiple providers, including fixes for forced tool calls on Amazon Bedrock.
- Added per-session worktree support and session mobility with pack/unpack and auth export/import commands.
- Enhanced request compression with gzip for Anthropic and zstd for GitHub Copilot, reducing upload sizes significantly.
New
- Live Background Jobs view to monitor running jobs without interrupting the transcript.
- Agent lineage navigation for moving between subagents and main session from composer header.
- RPC controls to cancel or steer individual foreground and background subagents.
- Goal management for RPC hosts with optional automatic goal continuation.
- Native judge support for OpenAI Decisions API and Cloudflare Workers AI decision models.
- Cloudflare Workers AI authentication and automatic question batching for native judges.
- /fork support for creating new Tern pane while a turn is streaming.
- Local IPC endpoints now run on native transport with owner-only permissions on all platforms.
Improved
- Context caching for repeat turns, reducing redundant prompt processing on unchanged system prompts.
- Anthropic prompt caching for long conversations with optimized memory usage and cache breakpoint selection.
- Devin tool calls now use native CLI sampling settings by default.
- Completion estimates for running subagents with configurable polling and adaptive scheduling.
- Question handling in eval with support for DataFrames, dicts, lists, and images printing as tables and JSON.
- Model picker and agents views no longer stack duplicates; existing views are brought forward instead.
- Reduced CPU usage while streaming with multiple active agents through lighter message queue.
- Improved auth broker account selection and usage-limit enforcement consistency across sessions.
Fixed
- Forced tool calls failing for Claude Opus 5.5 and Sonnet 5.5 on Amazon Bedrock.
- Tool calls with cut-off or trailing JSON no longer executing from auto-closed previews.
- Native judgment requests accidentally exposing credentials when credential redaction was enabled.
- Parallel same-name tool results being paired with wrong calls in Gemini when history stored out of order.
- Factory Droid Gemini dropping MCP tool calls with names over 64 characters from later requests.
- Cursor cached prompt token accounting reporting duplicate input-token and cost on cached turns.
September 2026
33 releases: 18.1.13 → 18.4.5
September brought major updates to model support, authentication, and streaming infrastructure. The month included new providers, improved error handling, and significant performance optimizations across the platform.
- Removed getCatalogProviderEntry and the CATALOGPROVIDERS constant in favor of providerEntry and providerEntries.
- Anthropic streaming and provider request helpers must be imported from @oh-my-pi/pi-ai/providers/anthropic instead of package root.
- Moved the public NOAUTHSENTINEL export from providers/openai-shared to auth-retry.
- Removed stream-level Anthropic prompt-cache keep-alive: StreamOptions.anthropicCacheRefresh and StreamOptions.anthropicCacheRefreshRequest.
Highlights
- Added support for new providers including DeepInfra, SiliconFlow, Cerebras, StepFun, and Apple Foundation Models with on-device processing.
- Introduced live steering support allowing users to send messages and interact with models during active streaming responses.
- Implemented comprehensive authentication improvements including credential rotation, multi-account support, and OAuth policy routing.
- Added new modalities including text embeddings, document reranking, video generation, image generation, and audio services across multiple providers.
- Launched omp stream for livestreaming terminal sessions with viewer chat and screen redaction.
New
- Agent.getPendingToolResults() for reconstructing live displays before buffered tool results are persisted.
- Opt-in experimental notes-backed context windows with persistent branch-local notes and searchable session history.
- /loop accepts --until and --while flags to gate iterations on shell command exit status.
- advisor.maxNotesPerUpdate setting and WATCHDOG.yml configuration for batching reasoning verifier findings.
- Judgment module for typed questions over JSON state including choice, yes/no, and score judgments.
- TypeSafe provider support with System One authentication and credential rotation.
- omp predict command for evaluating completion performance and bootstrapping predictions.
- wait tool for background-process coordination available to agents using task or bash.
Improved
- Browser automation now works alongside already-running Chrome using isolated profiles and keeps requested profiles separate.
- Improved agent performance by reducing redundant tool-schema processing during repeated model calls.
- Anthropic prompt-cache head now anchors on the last stable segment instead of volatile recall suffix.
- Fixed streaming CPU usage on long turns by deep-cloning only touched blocks instead of entire accumulated messages.
- SmolLM word completion performance improved about 25x on Windows x64 and Intel Macs.
- Syntax highlighting now uses 5x less memory and runs 3-5x faster using Oniguruma instead of fancy-regex.
- Context-window handling automatically adjusts output-token limits and supports models that truncate at context-window limit.
- Improved model cache invalidation efficiency with deep equality checks when database version signals change.
Fixed
- Fixed Anthropic OAuth requests capping output at 64k tokens; they now request the model's full ceiling of 128k on Opus 5.5.
- Fixed agents looping for hours when every turn spends whole output limit on reasoning; length-stop retries now tell the model reasoning was discarded and to act in smaller steps.
- Fixed concurrent usage refreshes repeatedly probing providers after failures, reducing unnecessary usage-endpoint requests.
- Fixed Codex sessions continuing to spend account credits after plan limit reached by switching to accounts with available plan usage.
- Fixed tool calls that put payload in the intent field silently running with leftover arguments; they now fail with an error.
- Fixed background job completions interrupting foreground Bash and eval calls, causing repeated background movement.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Oh My Pi: pricing, features and alternatives · All changelogs