v15.13.3

@oh-my-pi/pi-agent-core

Added

• Added the interruptible tool field: when set, the agent loop may abort the tool mid-execution to deliver a queued steering message (honored only in immediate interrupt mode).


• Added support for gemini and gemma as valid owned tool syntax values in environment configuration

Fixed

• Fixed pruneToolOutputs blanking tiny tool results during overflow pruning: results below 50 tokens (MIN_PRUNE_TOKENS) are no longer replaced with the [Output truncated - N tokens] placeholder, which cost more tokens than the result itself and churned the prompt cache for zero savings.

@oh-my-pi/pi-ai

Added

• Added the gemini in-band tool-call syntax with Python-style tool_code blocks and default_api invocations


• Added the gemma token-delimited in-band tool-call syntax using <|tool_call> and <|tool_response> blocks


• Added gemini and gemma to owned stream tool-result token detection so their tool responses are recognized


• Fixed truncated Gemini and Gemma tool blocks from being emitted as plain text during streaming


• Added the Azure OpenAI provider definition (azure) to the registry; AZURE_OPENAI_API_KEY resolves as its env-var API key via the catalog provider table.

Changed

• Gemini tool-call examples now render without the default_api. namespace prefix, keeping <example> blocks concise. The live wire format still uses default_api. per the Gemini grammar.

Fixed

• Fixed duplicate tool call projections by deduplicating provider-native toolCall events against in-band tool_code calls and keeping only the first real channel


• Dropped nameless native toolCall events so they no longer appear as surfaced tool calls in owned-mode streams


• Fixed Gemini/Gemma in-band tool-call parsing around Python comments, raw/unicode string literals, and Gemma close-token text inside string values.

@oh-my-pi/pi-catalog

Added

• Added Azure OpenAI as a catalog provider (azure, default model gpt-5.5, env var AZURE_OPENAI_API_KEY), bundling the OpenAI-family models Azure serves over the Responses API (GPT-4/4.1/4o, GPT-5 family, o-series, Codex). Like Amazon Bedrock it is catalog-only — models ship in the bundle and become selectable once the env key is set, with the deployment base URL resolved at runtime from AZURE_OPENAI_BASE_URL/AZURE_OPENAI_RESOURCE_NAME.


• Added models.dev-backed bundled catalogs for providers that previously shipped no offline models: Hugging Face, Kilo, Moonshot, NanoGPT, Synthetic, Venice, Ollama Cloud, and the Xiaomi Token Plan regions (ams/cn/sgp). They still discover live when credentialed; the bundle is now a non-empty baseline.

Changed

• Updated stale provider default models to their latest bundled versions: OpenAI-family providers (azure, github-copilot, aimlapi) → GPT-5.5; Gemini providers (google, google-gemini-cli, google-vertex) → gemini-3.1-pro-preview; GLM providers (zai, zhipu-coding-plan) → glm-5.2, cerebras → zai-glm-4.7; Kimi providers (fireworks, opencode-go, moonshot) → kimi-k2.7-code, kimi-code → kimi-for-coding, together → moonshotai/Kimi-K2.7-Code; alibaba-coding-plan → qwen3.7-plus; and Claude-Sonnet defaults (cloudflare-ai-gateway, cursor, gitlab-duo, kilo, opencode-zen, vercel-ai-gateway) → Claude Opus 4.x.


• Restricted models.dev Azure discovery to OpenAI-family IDs (gpt-, o1, o3, o4, codex, chatgpt), excluding Foundry-hosted third parties (Claude/DeepSeek/Llama/Mistral/Phi) that Azure serves through non-Responses APIs.


• Detected the Azure OpenAI Responses compat surface (developer role, strict tool mode, strict tool-result pairing) by provider id as well as base URL, so bundled azure models whose deployment host is only known at runtime still get the right wire behavior.


• Renamed the Qwen3-ASR-Flash model label to Qwen3 ASR Flash

Fixed

• Fixed tool syntax selection for Gemini-family and Gemma model IDs by routing them to dedicated gemini and gemma formats instead of generic XML


• Fixed zhipu-coding-plan and together shipping no bundled models: their descriptors referenced non-existent models.dev keys (zhipu-coding-plan, together); pointed them at the real keys (zhipuai-coding-plan, togetherai) so they bundle their GLM and full catalogs respectively.


• Folded the azure-openai-responses API into the OpenAI Responses thinking-inference branches so Azure reasoning models (o-series, GPT-5, Codex) resolve the discrete effort vocabulary (including xhigh) and effort-control mode instead of falling through to generic defaults.


• Fixed ollama-cloud discovery inheriting an unsafe cross-provider contextWindow/maxTokens when /api/show returns no size metadata; it now falls back to the safe 128K context / 8K output caps.


• Dropped internal Fireworks control-plane resource ids (accounts/fireworks/{models,routers}/…) from the bundle; only the public request ids ship.

@oh-my-pi/pi-coding-agent

Added

• Unexpected stop detection: optional tiny/smol classifier that continues the turn when the assistant says it will act but emits no tool calls.


• Settings features.unexpectedStopDetection and providers.unexpectedStopModel.

Changed

• Changed the job poll to return early when a steering message is queued, draining the steer immediately instead of waiting out the poll window.


• Capped unexpected-stop auto-continuation to three retry attempts before giving up on repeated stops


• Updated the edit tool's hashline prompt, grammar, and docs to recommend the .= inclusive range separator (SWAP 1.=3:); the legacy .. form still parses.


• Normalized all internal worker argv selectors under the __omp_worker_ prefix, skipping the async worker dispatch check during normal CLI startup.

Fixed

• Filtered out whitespace-only and dot-only (. or …) assistant blocks so they are treated as empty and no longer appear as visible content in message rendering, streaming reveal counts, or session export output


• Filtered placeholder-only thinking content from ACP notifications and message visibility checks so dot-only reasoning_content no longer triggers turn completion or read/run updates


• Fixed ModelRegistry tests making outbound network calls by automatically stubbing fetch during test execution.


• Fixed eval JS cells (and browser-tab worker startup) always stalling for the full init timeout — typically the cell's whole 30s budget — before silently falling back to the slower inline worker. The self-dispatching CLI host imports the worker module dynamically from its argv dispatch, so the worker's own parentPort.on("message") attached only after Bun flushed the messages the parent posted before spawn; the synchronously-posted init handshake was dropped and never answered with ready. The host now installs a buffering parentPort inbox synchronously in the entry's sync prefix (before importing the worker module) and the worker binds it on load, replaying the buffered handshake. omp --smoke-test now also spawns the JS eval worker through the host entry and asserts it handshakes on a real worker thread.


• Fixed pre-prompt context-full compaction on OpenAI Responses sessions to use provider-anchored context usage when available, so large encrypted reasoning signatures no longer trigger automatic maintenance while the visible context percentage remains below threshold (
  Vollständiger Original-Bericht

  Ausführliche Details, Code-Beispiele & Hersteller-Stellungnahme auf github.com.

  ↗ Original-Artikel auf github.com lesen