Thoth Files

v3.19.0 — Provider Runtime Foundation & ChatGPT / Codex

Thoth's model layer has been rebuilt around a first-class provider runtime. API-key providers, local Ollama models, custom OpenAI-compatible endpoints, media providers, and ChatGPT / Codex subscription access now flow through one provider-aware catalog and picker system instead of a mix of legacy cloud lists, starred models, and per-screen dropdown logic.

This release also adds ChatGPT / Codex as a distinct subscription-backed provider. It is intentionally separate from OpenAI API-key access: Codex uses an in-app ChatGPT sign-in, keeps Thoth-owned tokens in the OS credential store, and labels duplicate model names as OpenAI API versus ChatGPT / Codex so users always know which route they are using.

🧠 Provider Runtime Foundation

• New providers/ subsystem — provider definitions, metadata-only config, keyring-backed provider secrets, catalog normalization, runtime construction, status summaries, error normalization, Quick Choices, custom endpoint support, and routing-profile foundations now live in one dedicated package
• Provider runtime facade — OpenAI, OpenRouter, Anthropic, Google AI, xAI, custom OpenAI-compatible endpoints, Ollama catalog rows, and ChatGPT / Codex all route through a shared runtime layer while preserving the public models.py compatibility API
• Stable model refs — provider-backed picker values use refs such as model:openai:gpt-5.5 and model:codex:gpt-5.5, keeping identical raw model IDs distinct across providers
• Provider-aware labels — duplicate model names now show route labels such as GPT-5.5 — OpenAI API and GPT-5.5 — ChatGPT / Codex in chat, Designer, workflow, status, and settings pickers
• Metadata-only provider config — providers.json stores provider state, Quick Choices, catalog cache, fingerprints, and status metadata; raw API keys and OAuth tokens stay in the OS credential store when available
• Status and insight awareness — Thoth Status now exposes provider-aware model/runtime context and an insights category, while Dream Cycle Phase 5 includes model/provider/media context in its system snapshot before generating actionable insights
• Custom endpoint foundation — custom OpenAI-compatible endpoints can be saved, refreshed, and surfaced as provider catalog rows without overloading the built-in OpenAI provider

⚙️ Settings → Providers & Settings → Models

• Providers tab cleanup — the old Cloud surface is now Providers. It focuses on provider connection state, API keys, ChatGPT / Codex sign-in, health, refresh, setup guidance, and custom endpoint management
• Models tab ownership — model browsing, raw provider catalogs, local Ollama catalog rows, pin/unpin actions, defaults, and Quick Choices now live in Settings → Models
• Consolidated Model Catalog — a category-first catalog groups Brain, Vision, Image, and Video-capable rows by provider, with inline actions for pinning, setting defaults, downloading local models, and clearing disabled reasons
• Polished Defaults panel — Brain, Vision, Image, and Video defaults use compact provider/local badges, context controls, enable switches, and empty states that point users to the catalog instead of scattering model controls across tabs
• First-run setup alignment — setup now offers migration before model setup, supports the Providers path for API-key users, and points users to Settings → Models for exact model pinning after launch

💬 Picker Unification

• One picker source — chat header overrides, live chat model override, background workflow model override, Designer inline model selection, Telegram /model, and Thoth Status model updates all use the same provider-aware Quick Choice helpers
• Surface-specific choices — Brain, Vision, Image, and Video surfaces filter models by capability so media-only models do not leak into normal chat and chat-only models do not appear as image/video options
• Legacy compatibility — existing starred cloud models and bare model IDs are migrated or resolved without breaking saved settings, while new provider-backed selections preserve their provider route
• Runtime banner cleanup — chat status now uses dynamic provider display labels, so custom providers and ChatGPT / Codex show accurate route names instead of hardcoded cloud labels

🔐 ChatGPT / Codex Subscription Provider

• In-app ChatGPT sign-in — direct Codex runtime requires Thoth's device-flow ChatGPT sign-in and stores Thoth-owned OAuth tokens in the OS credential store
• CLI auth boundary — external Codex CLI auth files are display-safe metadata/reference hints only. Thoth can show that a CLI login exists, but it does not copy runnable tokens from ~/.codex/auth.json
• Live Codex catalog — ChatGPT / Codex catalog discovery uses https://chatgpt.com/backend-api/codex/models?client_version=1.0.0 when OAuth runtime credentials are present, caches display-safe metadata, filters hidden/internal rows, and falls back to documented subscription models when live discovery is unavailable
• Responses transport — ChatCodexResponses handles the ChatGPT/Codex Responses SSE backend, bearer/account headers, streaming text, function-call chunks, tool-call replay, and 401 refresh retry behavior
• Tool-call parity — Codex streaming now emits LangChain tool-call chunks, so normal chat can execute tools instead of ending with empty assistant messages when Codex asks for workspace/tool context
• Current-turn fallback — checkpoint fallback only uses an AI answer from the current submitted turn, preventing stale prior assistant text from being replayed after an empty streaming turn

🖼️ Media Providers & Model Catalog

• Image/video model routing — image generation and video generation models participate in provider-aware selection, catalog pinning, and surface filtering
• Provider media status — Thoth Status and Models settings can report media provider availability and selected image/video models without treating media rows as Brain models
• Ollama catalog parity — downloadable Ollama rows appear as non-runnable catalog entries until installed, with local download actions in the Models catalog
• Vision reuse — provider models with image capability can be detected and reused by the Vision feature alongside local vision models

🎨 Designer & Streaming Reliability

• Detached stream cleanup — long Designer/browser sessions now clear terminal active-generation bookkeeping when the graph finishes, even if the browser client disconnects during streaming
• Final-response hydration — detached completions reload active thread messages from LangGraph checkpoints before rebuilding the UI, so final assistant prose appears after reconnect instead of being hidden behind stale in-memory state
• Stored HTML normalization — Designer project HTML no longer persists render-time data:image/...base64 payloads; stored projects keep canonical asset://... references while preview/export resolves assets at render time
• Preview timer cleanup — Designer preview polling timers deactivate on client disconnect or deleted-parent errors instead of continuing to touch removed NiceGUI clients
• Stale-run recovery — sending a new Designer/chat message can drop stale terminal generation entries while still blocking truly live runs

💻 Claude Code Delegation Skill

• New bundled skill — bundled_skills/claude_code_delegation/SKILL.md teaches Thoth how to coordinate Claude Code CLI as an external coding worker for implementation, review, refactor, and larger repository tasks
• Thoth remains coordinator — the skill keeps Thoth responsible for scoping the request, checking local state, choosing the narrowest Claude Code tool permissions, inspecting diffs, running verification, and explaining results to the user
• Approval-gated shell workflow — Claude Code runs through Thoth's shell workflow with explicit working-directory checks, bounded print-mode commands, --allowedTools, --max-turns, optional budget limits, and no permission bypass unless the user explicitly asks
• Secret and safety boundaries — the skill warns not to forward API keys, Thoth memory, private notes, or sensitive user data to Claude Code unless explicitly requested, and it forbids destructive git, deploy, production migration, and secret-handling delegation without clear user approval
• Interactive mode guidance — print mode is preferred; interactive/tmux-style Claude Code orchestration is documented as advanced and best suited to macOS/Linux/WSL2 with explicit cleanup

🧪 Tests & Release Checks

• Focused provider suites — new provider tests cover config normalization/masking, keyring namespace storage and chunking, provider catalog inference, model selection refs, media model filtering, custom endpoints, runtime construction, and ChatGPT / Codex OAuth/catalog/transport behavior
• Bundled skill coverage — the main suite validates claude_code_delegation as a bundled skill and checks the skill parser/discovery path that loads it
• Designer regressions — test_suite.py covers detached finalization cleanup, stale terminal generation recovery, deleted-client detach detection, Designer asset canonicalization, preview timer cleanup, and checkpoint hydration for detached final answers
• Release smoke — test_suite.py validates v3.19.0 version consistency across version.py, Windows installer, macOS app plist, CI release workflow, bug report template, and install dependencies
• Packaging smoke — Windows installer coverage includes recursive providers/ plus ui/model_catalog.py and ui/provider_settings.py; macOS app packaging includes providers and the full ui package
• Clean first-run smoke — a temporary THOTH_DATA_DIR import/config check confirms setup wizard and provider config load cleanly before any provider state exists
• Final validation — direct test_suite.py passes with the release-smoke checks, and full pytest -q passes with 159 passed, 1 skipped

⚠️ Release Notes & Risk Notes

• Codex runtime sign-in — ChatGPT / Codex models only run after an in-app ChatGPT sign-in stores Thoth-owned OAuth tokens in the local OS credential store
• Subscription backend risk — ChatGPT / Codex uses ChatGPT's subscription/internal Codex backend rather than the public OpenAI API. The endpoint, catalog shape, auth requirements, rate limits, and model availability may change upstream without the same stability guarantees as the public API
• Privacy — when a ChatGPT / Codex model is selected, the current conversation plus model-visible tool context and tool results are sent to ChatGPT / Codex for that turn. Durable Thoth data such as memories, documents, files, and other conversations stay local unless explicitly included in the active conversation or exposed through a tool result
• Manual smoke still required — before publishing installers, run clean-machine Windows/macOS smoke for first launch, Settings → Providers, Settings → Models catalog/pinning/defaults, ChatGPT / Codex sign-in/status, shared model pickers, and a long Designer/browser task with reconnect

📁 Files Changed

:::manifest
schema: 1
files:
  Thoth-3.19.0-macOS-arm64.dmg: sha256=694bc76dbec60789bc15cf585e49dd17b470458faa64fdbd1123b4ba8f465f64
  ThothSetup_3.19.0.exe: sha256=465db780985cb3a4f25620854758b81184a861ea23a0c87027d5dbe03ae65fc4