Earendil Works Release Notes

New Features

• Claude Opus 5 — Available on Anthropic and Amazon Bedrock with adaptive thinking (including xhigh), inference profiles, and prompt caching. See Providers.
• Anthropic gateway bearer auth — ANTHROPIC_AUTH_TOKEN authenticates against Anthropic-compatible gateways that require Authorization: Bearer, including compaction and branch summaries. See Environment Variables or Auth File.
• Faster, more resilient model catalogs — pi.dev catalogs revalidate with If-None-Match so unchanged providers answer with an empty 304, and llama.cpp models stay listed across restarts. See llama.cpp.

Added

• Exposed the outputPad setting to custom message renderers. See Extensions (#7045 by @xl0). …

New Features

• Constrained tool sampling — Tools can prefer or require strict JSON Schema sampling or use OpenAI Lark/regex grammars, with model capability metadata preventing unsupported requests. See Constrained Sampling for Tools.
• OpenRouter and Kimi Code sign-in — Use /login to authorize OpenRouter or a Kimi Code subscription without manually configuring API keys. See OpenRouter.
• Session-aware, streaming bash integrations — Bash tools receive current session/model metadata, while direct RPC bash commands stream correlated output. See Bash Tool Session Environment and RPC bash events.

Added

• Added inherited Tool.constrainedSampling with strict JSON Schema (prefer/require) and OpenAI Lark/regex grammar variants across OpenAI, Anthropic, Amazon Bedrock, Google Gemini, and Mistral. See Constrained Sampling for Tools. …

New Features

• Verifiable release source archives — GitHub releases now include deterministic, checksummed source archives with instructions for rebuilding standalone binaries. See Building standalone binaries from release source.
• Resilient compaction and branch summaries — Transient provider failures now follow the configured retry policy, with retry lifecycle events available to interactive, JSON, RPC, and SDK consumers. See Compaction & Branch Summarization and RPC retry events.

Added

• Added deterministic, checksummed source archives to GitHub releases with documented standalone binary rebuild instructions (#6913 by @christianklotz).

Fixed

• Fixed compaction and branch summarization to retry transient provider failures using the configured retry policy, with retry lifecycle events exposed to interactive, JSON, RPC, and SDK consumers (#6901 by @davidbrai). …

New Features

• Local llama.cpp model management — Connect to a llama.cpp router, search and download Hugging Face models, and explicitly load or unload models with live progress. See llama.cpp.
• Full provider extensions — Extensions can register complete pi-ai providers with authentication, model refresh, filtering, and custom streaming. See Register New Provider.
• Qwen Token Plan providers — Use the built-in international and China subscription providers with regional endpoints and API-key authentication. See API Keys.
• Expanded usage accounting — Tool, compaction, and branch-summary usage is persisted and included in session totals. See Compaction & Branch Summarization.

Added

• Added Qwen Token Plan and Qwen Token Plan China to built-in provider setup, default model resolution, and provider documentation (#6858 by @QuintinShaw). …

New Features

• Kimi Coding thinking compatibility — Kimi Coding models now use adaptive thinking correctly; K3 exposes its supported max level and supports replaying empty-signature thinking blocks. See Kimi For Coding setup and Model Options.

Fixed

• Fixed inherited Kimi Coding requests to use Anthropic adaptive thinking effort without token budgets, and enabled empty thinking signatures for K3 and kimi-for-coding.
• Fixed inherited Kimi K3 pricing metadata for Moonshot AI and Moonshot AI China.
• Fixed inherited Kimi Coding K3 thinking-level metadata to expose only the supported max level (#6737).
• Fixed inherited catalog generation restoring xAI models removed in 0.80.9 (#6736).

New Features

• Kimi K3 and deferred tool loading — Use Kimi K3 across built-in providers, including progressive extension tool activation through Kimi’s native protocol. See Dynamic Tool Loading, OpenAI Compatibility, and the kimi-deferred-tools.ts example.

Added

• Added inherited Kimi K3 support for Kimi Coding, Moonshot AI, Moonshot AI China, OpenRouter, and Vercel AI Gateway.
• Added Kimi deferred tool loading for extension-driven tool activation. See Dynamic Tool Loading, OpenAI Compatibility, and the kimi-deferred-tools.ts example.

Changed …

New Features

• Unified model runtime and provider authentication — ModelRuntime centralizes model configuration, provider-owned /login, and dynamic provider catalogs. See Providers.
• Live model catalog refresh — /model refreshes configured providers in the background, and pi update --models forces an immediate refresh. See Install and Manage.
• xAI device-code OAuth and Grok 4.5 Responses support — Sign in to xAI with a device code and use Grok 4.5 with low, medium, or high thinking. See xAI.

Breaking Changes

• Replaced the SDK's CreateAgentSessionOptions.authStorage and modelRegistry options with the async modelRuntime option. AuthStorage and its storage backends are no longer exported; use ModelRuntime (or a custom pi-ai CredentialStore), or readStoredCredential() for one-off reads of auth.json.
• Removed redundant ModelRuntime.getAll(), find(), getSnapshot(), and getAuthOptions() projections. Use the pi-ai Models methods getModels(), getModel(), getProviders(), and checkAuth() directly. …

Breaking Changes

• Removed the openai-responses compat.sendSessionIdHeader flag from models.json. Session-affinity behavior is now controlled by compat.sessionAffinityFormat ("openai", "openai-nosession", or "openrouter"). Replace sendSessionIdHeader: false with sessionAffinityFormat: "openai-nosession" (#6496 by @petrroll).

New Features

• Cache-friendly dynamic tool loading - Extensions can add tools during execution while supported Anthropic and OpenAI Responses models preserve prompt-cache prefixes. See Dynamic Tool Loading.
• Message copy shortcut - Ctrl+X copies the last assistant message in the transcript or the selected message in /tree, making older and branched messages directly copyable. See Display and Message Queue.
• Fable 5 xhigh and max thinking - Native xhigh and max thinking levels are available across generated provider catalogs. See Model Options.

Added …

New Features

• max thinking level - New opt-in thinking level above xhigh, natively supported on GPT-5.6 and adaptive Claude models, available across CLI (--thinking max), SDK, RPC, and model selection. Custom themes can define thinkingMax. See CLI Reference.
• Input-based pricing tiers - Request-wide input-token pricing tiers for accurate long-context cost accounting (e.g. GPT-5.4/5.5/5.6 long-context rates), also configurable for custom models in models.json and modelOverrides. See Model Configuration.

Added

• Added the opt-in max thinking level across CLI, SDK, RPC, model selection, and themes. Custom themes can define thinkingMax; existing themes fall back to thinkingXhigh.
• Added request-wide input-token pricing tiers to custom model costs in models.json, modelOverrides, and extension-registered providers.
• Added ~ (home directory) expansion for the shellPath setting (#6470 by @aaronkyriesenbach).

Fixed

• Fixed inherited post-compaction output-token budgeting to ignore stale assistant usage from before the compaction boundary (#6464). …