Oh My Pi v17.3.3 Fixes Gemini Reasoning Loops

Oh My Pi — the open-source terminal coding agent that has been quietly shipping daily releases while bigger names take weeks — just crossed into version 17. The v17 series started with v17.0.0 on July 15, 2026, and the latest release, v17.3.3 (Aug 14), keeps the momentum going with fixes beginners will actually feel: it stops Gemini from looping on thought-only responses, recovers from malformed function calls, fixes hashline dangling range separators, and renders TUI markdown tables correctly — building on v17.3.2’s fixes for parent TUI stalls, hashline path collisions, Antigravity quota classification, and new models deepseek-v4-pro:preview and gemini-3.7-flash.

If you last looked at Oh My Pi during the v16 era, this is the update to care about: v17 is not a pile of small fixes. It reorganizes how the agent’s tools work under the hood, and two of its headline changes — the unified hub tool and the new xd:// virtual-file system — will change how you approve, watch, and share agent work. The August 9 hotfix batch (v17.2.12) fixed memory leaks and idle CPU spin. v17.2.14 and v17.2.15 added genuinely useful new capabilities like omp compress, natural-language omp cleanse, and MCP fixes. Now v17.3.2 polishes the daily-driver experience with reliability fixes that matter.

This guide explains what actually changed across the v17 series, what breaks when you upgrade, and what beginners should do next. No assumed knowledge: if you have never opened a terminal agent, the first section tells you what Oh My Pi even is.

What is Oh My Pi?

Oh My Pi is a free, open-source AI coding agent that lives in your terminal. You describe what you want in plain English — “refactor this function and add tests” — and it reads your code, plans edits, and applies them to your files with hash-anchored edits, so it can tell you exactly which lines changed and revert them cleanly. The project describes itself as an “AI coding agent for the terminal — hash-anchored edits, optimized tool harness, LSP, Python, browser, subagents, and more.” It sits at roughly 23,000 GitHub stars and supports macOS, Linux, and Windows.

Its defining trait is the “tool harness”: instead of assuming one workflow, Oh My Pi comes with a large catalog of tools (providers, browsers, subagents, file ops) and can be extended with plugins. That catalog is exactly what v17 reorganized.

Why v17 matters: the same agent, cleaner plumbing

Most of what a coding agent does happens through tools. Oh My Pi previously had several overlapping tools: irc for agent-to-agent messaging, job for background jobs, and launch for supervised long-running processes — plus a resolve tool that finalized plan approvals, and a “tool discovery” system that used full-text search to find tools on demand.

v17.0.0’s breaking changes:

The irc, job, and launch tools were merged into one hub tool (loaded as “essential”). Messaging, job control, and process supervision now route through a single interface.
The tool discovery system was removed, along with its settings (tools.discoveryMode, tools.essentialOverride, mcp.discoveryMode, mcp.discoveryDefaultServers).
The resolve tool was removed. Plan approvals now use three plain-text “virtual device” paths: xd://resolve (apply preview), xd://reject (discard preview), and xd://propose (submit a plan for approval).

The new xd:// protocol is the architectural centerpiece. v17 introduces the concept of mounting tools as virtual devices readable and writable through the normal read/write tools. Instead of hidden internal functions, approvals and previews become simple files: the agent writes to xd://propose, you approve via xd://resolve. This is much easier to reason about — you can see what the agent is doing because it is using the same read/write tools you see in the session.

What you get in the v17 series

The v17 line has shipped steadily since mid-July (v17.1.0 landed July 24, v17.2.0 on July 30, and patch releases almost daily since). Beyond the restructure, the series added several features beginners will actually feel:

Agent Plugins 1.0.0 standard (v17.2.x). Plugins are how you extend Oh My Pi. The new 1.0.0 standard brings automatic discovery, validation, and secure execution of compliant plugin packages — so plugins announce their capabilities and run in a validated way instead of relying on config inheritance. v17.2.11 fixed /reload-plugins and the Agent Control Center so agent definitions update without a full session restart.

omp share command. You can now share a saved session by ID or file path without launching the agent: omp share <session>. Useful if you want to hand a colleague (or another machine) a session.

AGENT=1 environment variable. Child processes spawned by the coding agent now see AGENT=1, so downstream tools can detect agent-driven execution. A small thing, but it helps hook watchers and safety tooling distinguish agent runs.

Better quota and provider handling (v17.2.x). v17.2.11 added reporting of Cursor’s personal monthly USD quota and remaining balance, honored ANTHROPIC_BASE_URL for custom gateways again (fixing broken agent proxies), raised the idle-timeout floor on Amazon Bedrock to 900 seconds for slow reasoning models like Claude, and improved error handling for Simplified Chinese quota messages and Copilot’s model_not_available_for_integrator error. Translation: fewer mysterious timeouts and credential rotation failures if you use Copilot, Cursor, Bedrock, or a custom gateway.

The v17.2.12 hotfix batch (August 9). The most recent patch in the series is aimed squarely at reliability, and three fixes matter to beginners:

Long sessions no longer leak memory. Previously, every background “scout”/task subagent that finished and got parked kept its whole session object — message list, provider transcript, debug buffers — pinned in memory until the process exited. Leave an agent running long enough and RAM slowly climbs no matter what the agent is doing. v17.2.12 releases that memory when the subagent is disposed.
No more idle CPU spin on WSL / Windows. The terminal-title spinner wrote updates in a way that consumed CPU while waiting on WSL, ConPTY, and slow Loader paints. On Windows/WSL dev machines, an idle agent now actually idles.
Web search results keep their sources. The Z.AI search integration was double-encoding some responses, which dropped source links and occasionally dumped raw JSON into the session. Fixed in this release.

The same batch fixed two correctness bugs you might have hit: shell “minimization” was rewriting short command outputs — including replacing meaningful rustc --print output with the single word OK — so outputs under 1,000 characters now pass through unchanged; and a /handoff with empty content was being shown as “cancelled” instead of a failed generation, which made debugging handoffs confusing. On the provider side, Codex sessions now try every configured account before surfacing a credential error, and Qwen3.8-Max’s discovered metadata was corrected so thinking controls and image input actually work through it.

v17.2.14 (August 11) — Reasoning controls and external thinking. This release adds granular control over model reasoning:

forceReasoningOff and disableReasoning options for OpenAI and Azure OpenAI models — turn off native reasoning entirely when you want faster, cheaper responses or when the model’s built-in reasoning hurts more than helps.
externalThinking setting — enables a private scratchpad via the new think tool, letting the agent reason externally before committing to an answer. Useful for complex multi-step tasks where you want to see the agent’s thought process.

v17.2.15 (August 12) — The feature-packed release. This is the biggest feature drop in the v17 series so far, and several changes directly improve the beginner experience:

New omp compress command. Rewrites single or multiple text files (supporting glob patterns and concurrent processing) into dense “prompt registers” using an isolated two-tool agent loop. Think of it as a smart summarizer that turns verbose source files into compact context you can feed back to the agent — perfect for reducing token usage on large codebases.
Smarter omp cleanse with natural language and interactive picker. The code-quality command now supports:
- Natural language requests: omp cleanse "fix all lint errors in the Python modules" launches a discovery subagent that inspects your project, determines the right commands, and runs them.
- Interactive picker: Run omp cleanse without arguments on a TTY and you get a menu to run all checkers, pick a specific one, or describe what to fix in plain English.
- Massively expanded tool discovery: Now supports staticcheck and golangci-lint (Go); mypy, pylint, flake8, ty, and basedpyright (Python); oxlint, deno lint, stylelint, and vue-tsc (JS/TS); and actionlint (GitHub Workflows).
- Higher subagent cap: Default for omp cleanse raised from 8 to 32 subagents for faster parallel checking.
MCP Streamable HTTP transport fixes. Updated the negotiated protocol version to 2025-11-25, resolving connection drops, SSE resumption gaps, and double-execution of tools during auth refreshes. If you use remote MCP servers over HTTP, this release is a significant stability upgrade.
/handoff preserves local artifacts. Plans, scratch files, and research notes now copy across the handoff session boundary instead of being lost. Handoffs between sessions (or machines) are finally practical.
Hardened tar parsing. Replaced libarchive-based tar reading with an in-process reader that safely handles complex archives, symlinks, and sparse metadata — no more crashes on weird tarballs.
--external-thinking CLI flag. Forces external thinking tool activation from the command line, useful for scripting or when you want consistent behavior across sessions.
Hashline edit safety. Added a post-apply parse advisory warning that alerts you when an applied edit fails to parse (despite the pre-edit content parsing successfully), catching balance-neutral misplacements that previously failed silently. Also fixed a bug where Rust lifetimes (e.g., 'static) were incorrectly parsed as string literals, which could lead to silent signature deletions.
Headless mode hang fix. Fixed a hang in omp -p (print mode) runs when plan.defaultOnStartup: true was enabled.
Tool activity hiding fixed. display.hideToolActivity now properly hides reminders, diagnostics, and completion blocks, not just the main tool output.
Anthropic model ID parsing is now case-insensitive, and the high-resolution 1932px frame tier extends to Claude Opus 5 and later, preserving full history per compaction.
SSE reconnection support. Extended parsed Server-Sent Events to include optional id and retry fields, enabling reconnecting transports to retain stream cursors and respect server-requested retry intervals.

v17.3.2 (August 13) — Reliability polish that matters daily. This patch fixes three issues that directly affect how the agent feels in daily use:

Parent TUI no longer stalls after subagent results. Previously, when a subagent submitted its result, the parent terminal interface would freeze until you switched focus or resized the window. The event loop now wakes correctly on subagent completion (#8462).
Hashline headers retain workspace-relative paths. A bug caused hashline headers (like [src/settings.json#0063]) to collapse to just the basename, which let a same-named file at the session root capture a follow-up edit and reject it with hash is not from this session. Paths now stay workspace-relative (#8482).
Antigravity quota errors classified with smart retries. HTTP 429 responses from Google’s Antigravity (Cloud Code Assist) are now parsed for structured google.rpc.ErrorInfo reasons — QUOTA_EXHAUSTED, RATE_LIMIT_EXCEEDED, INSUFFICIENT_G1_CREDITS_BALANCE — and retry delays of five minutes or longer are applied to distinguish rotatable quota windows from transient throttling. Unsigned thinking blocks are also dropped instead of causing HTTP 400 errors on session resume or model switch.
New models available: deepseek-v4-pro:preview and gemini-3.7-flash are now in the catalog. Antigravity client-version discovery is dynamic, so version-gated models appear without code changes.
omp update fixed for foreign npm/bun aliases. The update command no longer misclassifies globally linked checkouts while preserving package-manager ownership (#8468).

v17.3.3 (August 14) — Gemini reasoning loops, hashline edge cases, and TUI rendering fixed. This patch targets the daily experience with Gemini models and the edit/apply pipeline:

Gemini thought-only STOP responses no longer loop. Previously, when Gemini emitted a reasoning-only turn without final output, the agent could retry identically and spin up duplicate Antigravity endpoint streams. Now the agent distinguishes these from empty transports and uses a bounded final-answer reminder to recover the session cleanly (#8493).
Malformed function calls get smart retries. When every emitted tool call in a Gemini turn is proven unexecuted (MALFORMED_FUNCTION_CALL), the agent retries the turn while preserving real tool-result and visible-output replay guards — so you don’t lose legitimate progress.
Pinned retry banner. Current terminal retry errors now stay in one pinned banner with attempt context, surfacing local continuation failures instead of stale provider errors that confuse debugging.
Hashline dangling range separators recovered. Hunk headers like PUT 244.=:, CUT 5.= (which previously failed as orphan payload lines) are now normalized to single-line ranges N.=N so edits apply cleanly.
TUI markdown rendering fixed. When Gemini emits a lone closing Markdown fence without its opener, the TUI no longer renders final headings and tables as one raw code block — they display correctly.

What breaks when you upgrade

Breaking changes in v17.0.0 are the real upgrade hazard:

Anything that scripted the old irc, job, or launch tools needs to target the hub tool.
Config settings for the removed discovery system, and the old resolve tool, are gone. If your config references them, clean them up.
edit.enforceSeenLines (defaults to false) is a new opt-in safety flag: when enabled, the agent refuses to edit lines that have not been fully displayed in the session. If self-review quality matters to you, enable it in settings.

For a typical beginner workflow — omp in a repo, describe a change, approve the diff — nothing about the daily flow changes. The same commands, the same approvals; the reshuffle is mostly invisible unless you used the advanced messaging or job tools.

How to update

Oh My Pi generally self-updates, but the official upgrade paths are: macOS/Linux via curl -fsSL https://omp.sh/install | sh, Windows via irm https://omp.sh/install.ps1 | iex, Homebrew users via brew upgrade can1357/tap/omp, and npm users via bun update -g @oh-my-pi/pi-coding-agent. Run the version check after updating to confirm you are on v17.3.3. If you are automating agents, watch for config settings referencing the old discovery or resolve paths and replace them per the v17.0.0 release notes.

The bottom line

v17 is the biggest internal change to the Oh My Pi agent in months: one tool to rule messaging, background jobs, and long-running processes; transparent xd:// approvals instead of hidden ones; and a serious plugins standard that makes extension safer. For beginners the practical wins keep stacking: the stability fixes in v17.2.12 (no more memory leaks, idle CPU spin, or dropped search sources), reasoning controls in v17.2.14, v17.2.15’s omp compress, natural-language omp cleanse, MCP fixes, handoff artifact preservation, and hardened tar parsing, v17.3.2’s TUI stall fix, hashline path fix, and Antigravity quota smarts, and now v17.3.3’s Gemini reasoning-loop kill, malformed-call recovery, hashline edge-case normalization, and TUI markdown rendering fix. If you skipped the v16 series updates, the jump to v17.3.3 is absolutely worth it.

Oh My Pi v17.3.3 Fixes Gemini Reasoning Loops, Hashline Edge Cases, and TUI Rendering — What Beginners Need to Know