Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 293 Agent status: 2 with issues Updated Oct 7, 2026
AgentGid

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. 2 months covered; the current month updates daily.

October 2026 · so far

12 releases: 18.4.6 → 18.7.0

October brought significant improvements to agent workflow management, model compatibility across multiple providers, and stability fixes. The month included enhanced follow-up handling, better support for Claude models on various platforms, and new RPC capabilities for host integration.

Action needed
  • task.completionProbeMs replaced by task.completionProbe (on/off setting); existing values migrate automatically.
  • SessionStorage.claimSessionFile(sessionPath) replaced by claimSession(sessionId, sessionPath).
  • sessionOwnerLeasePath() replaced by new method (exact name not specified in notes).
  • Agent.withdrawLiveSteering() replaced by Agent.withdrawUndeliveredQueuedMessages() with different return structure.

Highlights

  • New APIs for managing agent follow-ups and steering workflows, with queue-change notifications for improved control.
  • Fixed forced tool calls for Claude Opus 5.5 and Sonnet 5.5 on Amazon Bedrock, resolving a critical compatibility issue.
  • Added live Background Jobs view to monitor running jobs without interrupting the transcript.
  • New RPC controls to cancel or steer individual subagents, and word-completion support for web and IDE hosts.
  • Fixed multiple context overflow and model catalog caching issues affecting session stability.

New

  • Post-tool guidance via afterToolCall results to include additional context after tool calls, including failures.
  • Configurable catalog cache-warming lifetimes for Claude models on Amazon Bedrock with 5-minute defaults.
  • Agent lineage navigation in the composer header to move between subagents and main session.
  • Live Background Jobs view for monitoring without interrupting the transcript.
  • RPC ask-dialog mode allowing hosts to render all questions together with checkbox or radio-button controls.
  • Optional terminal section in theme JSON for naming the terminal a theme was made for.
  • Goal management for RPC hosts with automatic goal continuation, plus --goal flag for interactive launches.
  • Per-call model selector on task items, eval agent(), and workpool() using provider/model[:level] patterns.

Improved

  • Background job completion estimates now use adaptive polling (after 2, 5, 10, 30 minutes, then hourly) instead of fixed intervals.
  • Tool-result supersession improved to track path reads with selectors independently from bare-path reads.
  • Model picker autocomplete now uses the same relevance ordering as the full model selector.
  • Subagent follow-up messages now appear as background jobs that can be waited on or canceled.
  • /models Roles view shows which saved preset is in effect and allows switching between presets.
  • Parallel tool calls now execute and report independently instead of reusing or merging results.
  • OpenAI Responses requests now omit sampling parameters for incompatible models to avoid HTTP 400 errors.
  • Reduced CPU use while streaming with multiple agents active through lighter message-update queue handling.

Fixed

  • Forced tool calls failing for Claude Opus 5.5 and Sonnet 5.5 through Amazon Bedrock.
  • Startup segmentation fault on Apple silicon Macs running macOS older than 27 due to Apple Foundation Models.
  • HTTP 400 rejected requests consuming unbounded disk space by enforcing cleanup and size limits.
  • Context-overflow recovery for Strata requests with no usage information.
  • Model catalog caching causing unnecessary rewrites and incorrect handling of offline snapshots for endpoint-less models.
  • Cursor turns aborting with "Provider stream stalled" after long local tools by providing full idle window.

September 2026

46 releases: 18.1.0 → 18.4.5

September brought major provider expansion, improved authentication and streaming capabilities, and significant enhancements to context management and model routing. The month culminated in version 18.4 with telemetry updates and better credential failover handling.

Action needed
  • Removed Ruby and Julia eval backends; eval now supports Python and JavaScript only.
  • Removed eval parallel() and pipeline() helpers; agent() and completion() return handles immediately.
  • Python eval tool calls are now asynchronous coroutines requiring await syntax.
  • Renamed claudeCodeSessionId to sessionId in AnthropicClientOptionsArgs.

Highlights

  • Added support for numerous new providers including DeepInfra, LiteLLM, Qwen Portal, SiliconFlow, StepFun, Muse Code, and Factory Droid, plus Apple Foundation Models.
  • Implemented live steering support allowing users to provide input during active streaming responses across multiple model providers.
  • Introduced Anthropic server-side compaction as a remote backend, improving context management for long conversations.
  • Expanded service offerings to include text embeddings, document reranking, video generation, image generation, speech synthesis, and transcription across multiple providers.
  • Added unified predictive text engine with multiple providers and background model downloads for improved text completion.

New

  • ClinePass API-key authentication and Devin router-model support with credit usage reporting.
  • Amazon Bedrock requestMetadata support for cost and attribution in AWS invocation logs.
  • Native Windows ARM64 binaries with architecture-aware installation.
  • Recoverable native custom-scheme OAuth callbacks for macOS, Linux, and Windows with manual fallback.
  • Opt-in bash.allowCompoundCommands approval for conservative literal && chains.
  • Notes-backed context windows with persistent branch-local notes and searchable session history.
  • Omp stream for livestreaming terminal sessions with viewer chat and screen redaction.
  • Browser automation tools for accessibility auditing, React inspection, console monitoring, and semantic DOM queries.

Improved

  • Provider behavior now driven by model-resolved compatibility, identity, thinking, and behavior policies rather than name matching.
  • Devin integrations use released CLI identity and support parallel tool calls.
  • Dynamic mid-conversation updates for Anthropic system prompts, tools, and reasoning effort.
  • Improved steering behavior preserving already-emitted non-interruptible tool calls.
  • Extended-context catalog rebuilds now faster by resolving each model's window once per process.
  • Anthropic prompt-cache breakpoints now historical decimation every 15 user turns for stable prefixes during branching.
  • Better browser automation supporting already-running Chrome with isolated profiles and no process killing.
  • Improved context-window handling with automatic output-token limit adjustment and provider-reported usage.

Fixed

  • OpenAI remote-compaction replay for persisted sessions now works correctly.
  • Cursor Fable requests failing with JSON Schema composition keywords.
  • Gemini 3 sessions getting permanently stuck on 400 INVALIDARGUMENT after parallel tool calls.
  • DeepSeek DSML markup healer leaking orphan close tags into visible text.
  • Anthropic thinking context preservation when conversations are compacted or tool results pruned.
  • GitHub Copilot sign-in permissions issues for Enterprise organizations.

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