0.9.18

Cumulative release of the 0.9.18-alpha.1 through 0.9.18-alpha.7 prereleases. Per-change details remain in the unchanged prerelease sections below.

Cumulative release of the 0.9.18-alpha.5 through 0.9.18-alpha.7 prereleases. Per-change details remain in the unchanged prerelease sections below.

Cumulative release of the 0.9.18-alpha.3 through 0.9.18-alpha.7 prereleases. Per-change details remain in the unchanged prerelease sections below.

Cumulative release of the 0.9.18-alpha.6 prerelease. Per-change details remain in the unchanged prerelease section below.

Breaking Changes

• Replaced /fast with selectable canonical -fast model IDs. Removed its selector, codexFastMode.chat/codexFastMode.workflow settings, ATOMIC_CODEX_FAST_MODE, scope inheritance, toggle-state exports, settings-manager methods, and the environment-sourced settings-override layer without a compatibility shim. Select a fast model explicitly; effective settings now merge global then project settings.
• Fast-routing APIs now live in core/fast-model-routing.ts and core/fast-model-routing-transport.ts, with package-root exports retained for the replacement APIs. usesChatGptCodexTransport and usesFirstPartyCodexRouting remain; withCodexFastRouteHeaders derives identity from route metadata rather than an enabled flag.
• Cloudflare AI Gateway binding transport requires binding.fetch() and no longer falls back to gateway(id).run(...).
• WorkflowPendingStageDelivery now requires fail(reason: Error): void so delivery owners can settle exhausted recovery instead of leaving stages parked forever.
• Removed the experimental remote-session lease API, transcript projection helpers, and harness factory. @bastani/atomic/client now re-exports the service-addressed @earendil-works/pi-client API.
• createGatewayBindingFetch requires binding.fetch() and forwards requests verbatim. The gateway(id).run(...) fallback is removed, baseUrl and gateway options are ignored, and bindings without fetch() fail at construction. Set model base URLs to https://workers-binding.ai/ai-gateway/gateways/{gateway}/{provider} (#8287).
• Fast inference is explicit Model.fastRoute metadata, not a name suffix. Responses adapters accept the canonical model, send its route's upstream ID, enforce its declared tier even against caller options, and price priority using its base model. A route without a tier sends none. Payload hooks that replace the body with a non-object or change the route's model or tier now fail. Models without routes retain caller-controlled tiers and hooks.
• GitHub Copilot exposes route-bearing fast models only when OAuth fastModelIds advertises the exact ID. Provider-owned names ending in -fast remain ordinary models governed by availableModelIds.
• Workflow-stage targets require root-anchored workflow:<rootRunId>/<segment>[/<segment>...] paths. Legacy <runId>:<stageKey> targets fail with a migration hint.
• Removed fastMode from result, progress, runtime, artifact, and run-history metadata. Select fast inference with a canonical -fast model ID in model or fallback fields, such as openai-codex/gpt-5.6-sol-fast:medium.
• Workflow-stage Intercom targets require root-anchored workflow:<rootRunId>/<segment>[/<segment>...] paths. Legacy <runId>:<stageKey> targets fail with a migration hint.
• Removed fastMode from task results, stage snapshots, fallback metadata, and durable records. Pin canonical -fast model IDs explicitly; normal and fast IDs remain separate ordered fallback candidates.

Added

• Added explicit fast model variants to /model, --list-models, workflow catalogs, scoped lists, session restore, fallback candidates, usage, and attempt records. Thinking suffixes remain supported. OpenAI and Codex variants send the base upstream ID with priority tier; Copilot exposes only account-advertised fast IDs and sends them without an OpenAI tier.
• Added fast-specific modelOverrides, derivation APIs, ModelRuntime.getFastModelVariantDiagnostics(), and getWarning(). Existing provider, custom, or extension-owned IDs win collisions and produce actionable startup/catalog warnings. Route metadata, not a name suffix, grants fast behavior.
• Added GPT-6-Astra across OpenAI, Codex, Bedrock, OpenRouter, and Vercel, including derived first-party fast variants, provider-specific pricing, long-context tiers, and provisional account-gated Copilot support. Added Gemini 3.8 Flash, Claude Fable 5.1, Copilot Claude Fable 5/5.1, and Baseten GLM-5.3-Fast with provider-owned metadata and availability.
• Embedded PostgreSQL now supports offline durable storage on Linux musl x64/ARM64 and Windows ARM64. Windows ARM64 uses PostgreSQL x64 through Windows 11 emulation; installs and standalone archives carry only the matching runtime.
• Added supportsMidConvoEffort and compat.vllmPriority custom-model settings, and transcript notices for dropped Anthropic thinking when cache-miss notices are enabled.
• Added a clickable fullscreen Jump to latest message overlay with the configured shortcut. Holding Alt makes fullscreen mouse-wheel scrolling move five times as far (upstream #9166).
• Documented Zed terminal keybindings for Kitty-protocol modified keys (#8828).
• Added Claude Fable 5.1 with provider-specific pricing, a 1M-token context window, 128K output, always-on adaptive thinking, and exactly low, medium, high, xhigh, and max efforts. Added its server-side fallback targets, Claude Opus 4.8 and Opus 5, and compatibility controls for thinking binding, forced tool choice, and temperature.
• Added Gemini 3.8 Flash across supported provider catalogs, including GitHub Copilot. Google and Vertex use a 1,048,576-token context window and 65,536 output tokens with low, medium, and high thinking; other providers retain their own metadata. Added Copilot Claude Fable 5/5.1 and Baseten zai-org/GLM-5.3-Fast.
• Added GPT-6-Astra on OpenAI, OpenAI Codex, Bedrock, OpenRouter, and Vercel AI Gateway, retaining provider-owned IDs, reasoning levels, pricing, and request-wide long-context tiers. Added a provisional Copilot entry with zero costs for unverified pricing; account availability still depends on Copilot's authenticated catalog.
• Added OpenRouter microsoft/mai-image-2.6 and microsoft/mai-image-2.6-flash image models.
• Added public DocumentContent PDF input for Anthropic Messages and Bedrock Converse. Unsupported models receive a visible placeholder; non-PDF MIME types are rejected, and token estimates measure the encoded document payload.
• Added public FallbackContent, per-turn Anthropic effort persistence, historical effort markers, signed-thinking recovery, and diagnostics for thinking blocks dropped at request start or after server-side fallback.
• Added ModelFastRoute, resolveRequestedServiceTier, and assertPayloadPreservesFastRoute, plus AssistantMessageFrameEncoder and reduceAssistantMessageFrames for compact, replayable stream progress.
• Added compat.vllmPriority for servers using priority scheduling, supportsMaxOutputTokens for Responses gateways that reject the output cap, and compatibility flags for unsupported temperature and forced-tool-choice fields.
• Added literal, run-ID, and glob stage-path routing, sticky delivery to future matches, and workflow:<rootRunId>/** broadcasts to live and future descendants until root termination.
• intercom list shows persisted possible future targets and queued counts. Valid unknown paths queue with notInKnownSet warnings and produce terminal undeliverable notices only if never delivered. Tool guidance, skills, and docs teach path discovery and live-only asks.
• Added checksum-pinned embedded PostgreSQL on Linux musl x64/ARM64 and Windows ARM64, retaining explicit database URL precedence, Docker fallback, privilege dropping, and process-owned shutdown. Windows ARM64 uses PostgreSQL x64 through Windows 11 emulation.
• Launch persists possible literal, patterned, and nested-child stage paths. Sticky sends reach future matches, root broadcasts reach live and future descendants until termination, and intercom list shows possible targets and queued counts. Valid unknown paths queue with warnings and settle as undeliverable only if never matched.
• Status, list/detail views, and the BACKGROUND panel identify materialized pending stages, expose exact targets only when delivery is available, and distinguish unavailable delivery explicitly.

Changed

• Updated Pi runtime dependencies to 0.85.1, preserving Atomic startup, client API, and footer watchers. Fullscreen search uses the cached index and visible-match highlighting introduced in Pi 0.85.0.
• Startup paints the themed identity and focused editor before isolated-engine readiness. First submissions wait for optional resources, Escape cancels a waiting submission, and /reload retries extension failures without publishing failed candidates' host-managed state.
• Compiled/bundled builds reuse native-imported builtin extension factories across reloads while editable extensions and workflows retain content-hash invalidation. Standalone builds use syntax-minified shared sidecars and bytecode launchers, including Windows.
• Interactive model, thinking, cycling, and scoped-list selections save as startup defaults immediately. Selectors keep active choices marked while browsing; scoped lists use consistent toggles and strike through unavailable models. Removed Ctrl+S save-default UI.
• The working indicator reads Working, preserving Atomic's animated status row. The centered jump overlay no longer consumes a transcript row, and workflow stage chat uses matching copy and shortcuts.
• ModelRuntime enforces first-party Codex routing for standalone stream/complete calls too. Fast derivation excludes unsupported adapters and extension-owned transports, Copilot restore requires account entitlement and a base catalog model, and all provider catalog accessors agree on derived variants. Labels show the canonical model ID without a redundant fast badge.
• Claude on Anthropic and Bedrock accepts PDF input through the bundled AI library, with visible placeholders elsewhere. Interactive sessions do not yet produce document blocks; Bedrock requires citations for full visual understanding.
• Updated hashline edit guidance with worked examples, rejected-shape warnings, and a specification of native block resolution. Model-selection references record the August 26 DeepSWE snapshot, corrected prices and Pareto choices, and Fable 5.1's unmeasured status. Compaction docs explain why preserve_recent preserves text and tool exchanges but resets signed reasoning.
• Capable Responses models use prompt_cache_options.ttl: "30m" for long retention; older models retain the 24-hour legacy control.
• Exported ./utils/* subpaths for Pi 0.85 runtime packages.
• Qwen3.8 Max and Flash thinking choices follow models.dev on Qwen Token Plan providers, offering low, medium, and xhigh without off.
• Responses models supporting prompt_cache_options, including Astra, use ttl: "30m" for long retention. Earlier models keep the 24-hour legacy field; short and explicit no-cache modes retain capability checks.
• Assistant-frame reduction reconstructs owned block values rather than deleting fields on externally supplied content, preserving output and field order.
• Labels show the complete selected model ID without a separate fast badge.
• Bundled agents use GPT-6 Astra at low, with debugger at xhigh. Added Astra and Claude Fable 5.1 fallbacks ahead of older models, and aligned locator fallback chains with ordinary agents, including Sol, GPT-5.5, and Opus 4.8 at medium.
• The qlty skill follows repository/user priorities, preserves authoritative checks and read-only boundaries, supports offline/manual setup, and distinguishes prepared configuration from executed checks.
• Default guidance honors task-scoped inline/no-workflow requests while retaining testing, review, safety, and reconciliation of active workflow effects. Model selection for unpinned agents consults task-specific measured evals alongside role guidance.
• Heartbeat and stage guidance use one authoritative path broadcast for shared scope changes, teach discovery/globs/live-only asks, and distinguish orchestration from cross-cutting extension policy with explicit companion dependencies.
• Goal/Ralph orchestration, Ralph prompt engineering/research, and Open Claude Design use Astra at high. Goal reviewers and Ralph reviewer B use Astra at xhigh; reviewer A uses Fable 5.1 at high. Added both model families to role-specific fallback chains and updated Ralph research/reviewer ordering.
• Guidance honors task-scoped inline requests, reconciles active runs without duplicate execution, chooses verification by browser/terminal/desktop environment, supports offline qlty setup, and requires verified hosted links for authorized GitHub media uploads. Model-pinning advice consults measured task-specific evals alongside role guidance.
• Expanded the prompt-engineering skill with sourced Astra prompting/API migration guidance and separate GPT-5.6, GPT-5.5, Fable 5.1, Fable 5, Opus 5, Opus 4.8, and Sonnet 5 guides. Shared instructions cover selective references, proportionate verification, permission boundaries, and completion criteria.
• Stage chat uses the same jump-to-latest copy and shortcut as fullscreen chat.

Fixed

• Fixed Windows x64 and ARM64 archives crashing before startup in 0.9.18-alpha.1 (#2781).
• Restored startup resource listings, source-path expansion, slash-prefixed prompt labels, isolated-child extension inventory, deterministic collision labels, overlap warnings, and consistent hidden-resource handling across platforms.
• Fixed in-memory forks during running tools, lost compaction boundaries after forks, session import filename collisions, RPC aborts not cancelling manual compaction, and concurrent /share exports overwriting temporary files.
• Builtin tools resolve relative paths against an extension's ctx.cwd, including matching search hashline tags. Write confirmations no longer mislabel UTF-16 counts as bytes or discard user prose merely because it begins with Successfully wrote to.
• Fixed Claude Fable thinking-level routing on Copilot, Fireworks GLM endpoint routing, Fable 5.1 prefix/model-switch recovery, exact reasoning-byte restoration, unsupported sampling fields and forced tool choices, fallback replay and serving-model billing, and persisted thinking-drop notices on resume. Custom Responses gateways can disable max_output_tokens.
• Fixed managed fd/rg downloads behind GitHub API quotas, selected static musl archives on Linux x64/ARM64, and pinned darwin/x64 fd to 10.3.0. Fixed proxied plain-HTTP requests hanging after tool calls and signal-killed subprocesses appearing successful.
• Fixed JPEG EXIF orientation detection, skills disappearing when bash exists without read, and model-refresh errors missing their status spacing. Jump-overlay clicks preserve other mouse reports in the same input chunk.
• Raised branch-summary output to 4096 tokens, clamped to model limits, to avoid reasoning exhausting the previous cap (#8845).
• Resumed workflows cannot report completion or start replacement work while skipping their exact unfinished durable tool. Completed work remains cached and cancellation retains precedence.
• Fixed Claude Fable temperature rejection across Anthropic, Bedrock, OpenRouter, and Copilot, including sampling defaults that reintroduced temperature, top_p, or top_k. Forced tool choices on Fable 5.1 now fail explicitly across supported APIs, including all OpenAI forcing shapes and Bedrock requests without tools; auto and none remain unchanged.
• Fixed first-party Anthropic model switches and Fable 5.1 prefix changes by preserving eligible signed reasoning and letting the API drop incompatible blocks. Disabled thinking no longer sends the interleaved-thinking beta, and recovery diagnostics survive empty later reports or failed streams.
• Fixed mid-stream Anthropic fallback attribution, replay, and billing. Handoff markers survive, serving-model prices apply, earlier output-producing attempts are charged once, and pre-handoff reasoning and unexecuted calls/results are dropped together. Faux-provider replay keeps content indices aligned around fallback markers.
• Fixed Fable 5.1 OAuth requests rejected as claude_code_version_too_old by advertising the verified minimum Claude Code version, 2.1.251.
• Fixed Copilot Claude Fable routing through Anthropic Messages and all Fireworks GLM models through OpenAI-compatible completions. Removed retired xAI grok-build-0.1 and restored Qwen3.8 Flash on Qwen Token Plan Individual.
• Fixed Codex SSE terminal events without a trailing blank line and assistant-frame start snapshots losing providerThinkingLevel.
• Fixed NO_PROXY matching for root domains, subdomains, IPv6, port-scoped entries, and wildcard entries.
• Fixed invocation control across isolated workflow subgroups, pending-stage roster discovery and delivery, correlated live replies, registration-based authorization, and sibling/cross-run isolation. Queued sends render as successful, not failed (#2784).
• Background reconnects retry with bounded backoff after releasing ownership, restore stage visibility/routing, and close failed accepted connections instead of accumulating stale registrations. Host sessions retain their startup home group across joins and reconnects, preserving multi-workflow control without granting workers cross-invocation authority.
• Recoverable disconnects during lazy initialization, stage warm-up, advisory subagent authorization, or event relays no longer produce misleading stage/UI failures. Registered transport resets enter typed recovery; unrelated protocol, auth, configuration, and terminal errors remain actionable.
• Stage warm-up owns bounded retries. Exhaustion settles queued delivery with a stage-scoped terminal failure instead of hanging or printing into the host transcript. No model retry/fallback is spent, queued messages remain available, late drains are safe, and stages without queued messages still start.
• Brokers retire ended sockets before routing more traffic. Delivery waits for socket-write acknowledgement, so failed writes cannot claim success, open reply authority, poison retry identity, or inflate sticky-broadcast receipts.
• Recoverable disconnects on send/ask/reply return opaque retryToken values for explicit exact-operation retries. Tokenless identical calls remain distinct. Tokens allow three claimed attempts within the original 11-minute deadline, retain identity after inconclusive nondelivery, preserve implicit reply correlation, and reject invalid or mismatched claims before side effects. Capacity is reserved before new operations begin.
• Retry signatures survive reconnects while preserving caller-visible target and argument distinctions. Local subagent result relays reserve deduplication identity before delivery and fail closed on conflicts, capacity, and uncertain acceptance.
• Broker acceptance records use keyed SHA-256 HMACs rather than plaintext signatures, keeping message and attachment text out of SQLite/WAL/SHM. POSIX directories/files use owner-only modes; key/database mismatch and corruption fail closed. The 12-minute durability and 10,000-record/64 MiB limits remain enforced.
• Closed stages suppress late messages from their own cancelled subagents without dropping other stages' traffic; already-submitted sends preserve receipts and retry identities (#2840).
• Completing workflow stages cancel their still-running single, parallel, and detached subagents. Children retain interrupted/abort outcomes and cleanup, late findings no longer reach parent chat, and exact ownership preserves other stages' traffic. Already-submitted Intercom sends keep receipts/retry identities; detached children still notify while their owning stage is live (#2840).
• Fixed invocation-owned stage groups, collision-free identities, durable-resume identity preservation, and isolated Goal/Ralph reviewers. Pending target/ID displays wrap without truncating usable addresses; ended runs no longer advertise delivery, and narrow widgets preserve tool/elapsed metadata with accurate overflow counts (#2784).
• Tool-only workflows no longer produce ZERO_STAGES warnings or advertise tool nodes as chat-stage targets.
• Structured-output exhaustion now advances the model fallback chain. Each candidate gets the prompt plus three corrective follow-ups; failed attempts remain recorded and successful model metadata persists immediately (#2812).
• Exhausted queued Intercom delivery fails its stage deterministically without model retries or fallback. The first failure settles readiness once, preserves queued steering, disposes the refused session, and handles pre-latched failures and late drains safely.
• Durability degradation uses display-only warnings in interactive/RPC sessions and actionable console diagnostics in headless mode. Pending-message sweeps skip unrelated runs, deduplicate interactive warnings, and preserve messages for recovery.
• In-flight resume, catalog preparation, and completed-run opening retain the backend selected during initialization rather than failing after concurrent mutable-state changes.
• Targeted durable-tool aborts retain exact unfinished-tool identity and an inspection-only cancellation frontier. Resume, DBOS hydration, and session restoration replay completed work and retry that tool without fabricating model stages. Invalid or ambiguous checkpoint state fails before callbacks; older state requires typed checkpoint proof, not transcript inference.
• Graph parent replacement rejects self-edges and cycles before mutation. Tool-frontier continuations reject completion that skips the unfinished tool and reject replacement model/task or child-workflow work before side effects, including worktree setup. Recovery evidence and completed callbacks remain intact.

Removed

• Removed /atomic. Use /workflow list, /changelog, and /hotkeys instead.
• Removed Ctrl+X copy-message/selection to avoid workflow-navigation conflicts. /copy and automatic mouse-selection copying remain.