Thoth Files

v4.8.0 - Reasoning Controls, Context Safety & Native OpenCode Discovery

This release builds on v4.7.1 with provider-aware reasoning controls, safer context-capacity handling, more resilient rolling compaction, native OpenCode gateway discovery, and a responsive desktop composer. It lets each chat keep an exact model-specific reasoning choice, prevents custom endpoints from silently inheriting an invented context window, recovers more long conversations without a compaction failure loop, discovers newly listed OpenCode models with their real transport metadata, and strengthens Google, Anthropic-routed, mobile, and streaming compatibility without weakening local-first, approval, credential, or transcript boundaries.

Provider-Aware Reasoning Controls

• Exact model controls - adds Provider default, supported effort levels, thinking On or Off, and bounded token-budget choices only when the active provider-qualified model exposes an actionable capability.
• Thread-and-model persistence - stores reasoning selections locally for one thread and one canonical model reference, restores a valid selection when that model is revisited, and does not leak it into another chat or an incompatible model.
• Desktop and mobile access - adds a Thinking picker beside the desktop model control, exposes the same choices inside mobile Chat controls, and shares the behavior with normal Chat, Designer Studio, and Developer Studio.
• Slash and channel control - adds /reasoning to Chat and Telegram, Discord, Slack, WhatsApp, and SMS conversations; the bare desktop or mobile command opens the visible control while explicit effort, toggle, budget, and default arguments are validated against the active model.
• Provider-native requests - maps supported selections to OpenAI and Codex, Anthropic and Claude Subscription, Google, xAI, Ollama and Ollama Cloud, OpenRouter, OpenCode, and compatible-endpoint request formats without applying a provider-wide guess to an unknown model.
• Custom endpoint controls - extends Custom/Self-hosted provider settings with reasoning Auto, On, or Off behavior, optional thinking budgets, returned-reasoning preservation, explicit replay capability, and advanced request JSON for endpoints that require compatibility tuning.
• Separate thinking presentation - streams returned reasoning separately from the final answer and retains it as a collapsed Thinking section when the provider returns replayable content.
• One safe compatibility fallback - when a provider rejects a valid-looking explicit reasoning option before producing output, retries once with Provider default, clears only that rejected model selection, and shows a local notice; authentication, rate-limit, timeout, cancellation, server, and mid-stream failures are not silently replayed.
• Single stream callback path - isolates callbacks inside the reasoning fallback wrapper so normal and fallback streaming cannot duplicate visible tokens, tool events, or completion callbacks.

Context Capacity And Compaction Recovery

• Custom endpoint capacity ownership - treats every Custom/Self-hosted endpoint as server-managed even when it runs locally, uses only detected or manually declared model capacity in Auto, and no longer presents the generic remote 128K application fallback as evidence about a custom server.
• Fail-closed unknown custom context - keeps a custom model unavailable when Row-Bot cannot determine its context window and no cap is set, with guidance to refresh or probe the endpoint, declare its native limit, choose a verified Custom cap, or select another model.
• Exact advanced context values - adds validated Custom context entries from 16,384 through 4,194,304 tokens for local and provider models while retaining the common 16K through 1M presets and preserving exact saved values across the context-policy v3 migration.
• 64K local Auto target - raises the recommended Ollama Auto request from 32K to 65,536 tokens, still capped by known native metadata or the observed loaded allocation; a fixed 32K choice remains available for smaller-memory or reduced-tool configurations.
• Server-safe request behavior - treats the app setting as a planning cap, not permission to reconfigure the server, and stops sending the undocumented n_ctx chat-completions field to llama.cpp; server context remains a startup setting such as --ctx-size.
• Model-scoped endpoint probes - records which model a Custom endpoint probe tested and applies chat, tool-round-trip, streaming, and context evidence only to that model, preventing a successful sibling model from promoting or changing transport behavior for an untested one.
• Readiness aligned with evidence - promotes a probed custom model to Agent mode only after a successful tool round trip, keeps verified chat-only models in Chat Only, blocks failed chat probes, and preserves the minimum context floor even after a successful tool probe.
• Fixed-envelope preflight - measures the non-compactable system prompt and bound tool schemas before attempting rolling compaction and produces an exact estimated-versus-usable token error when the selected context cannot fit them.
• Newest-turn recovery - when retaining two recent groups cannot create enough slack, safely ages one more complete atomic group while keeping the newest group intact, including tool-call and tool-result pairs.
• Exact rebuild fallback - validates the rebuilt prompt before persistence, makes one additional bounded summary pass when necessary, and saves only the final successful summary state so an oversized intermediate result cannot trap the unchanged conversation in a repeated failure loop.
• Private failure diagnostics - logs bounded known compaction reasons while reducing unexpected exceptions to their class name, keeping transcript and provider text out of warning logs.

Native OpenCode Gateway Discovery

• Live gateway intersection - refreshes the OpenCode Zen and OpenCode Go /models catalogs and intersects each gateway's actual availability with the public native routing metadata used by OpenCode.
• Per-model transport routing - derives OpenAI Chat, OpenAI Responses, Anthropic Messages, or Google GenAI transport from explicit model or provider SDK metadata instead of relying only on a maintained model-name classifier.
• Richer model metadata - persists provider-qualified display names, context windows, input and output modalities, tool calling, streaming, and reasoning capability so newly listed supported models can flow through the catalog, readiness checks, context policy, and runtime after refresh.
• Google gateway support - enables OpenCode models routed through Google GenAI with the gateway's v1 endpoint and provider-native reasoning request mapping.
• Durable dynamic routes - restores cached OpenCode route metadata after a restart, retains the legacy static classifier for older cache rows, and fails closed with a sanitized diagnostic when a cached or newly advertised native protocol is unsupported.
• Failure-safe refreshes - coalesces the shared native-registry request across a full provider refresh, preserves the last known good gateway rows on network or registry failure, uses the static catalog only for a cold failed refresh, clears stale rows after a valid empty gateway response, and keeps Zen and Go caches isolated.

Composer And Provider Compatibility

• Responsive desktop composer - uses component-width breakpoints to compact Model, Thinking, and Approval labels progressively while keeping each icon, picker, tooltip, keyboard target, and accessible state available.
• Compact secondary controls - collapses active Skill chips into a counted Skills button, shortens the context meter without losing its threshold or tooltip, moves transient voice status above the toolbar, and gives Send and Stop one stable action slot so controls do not wrap or jump.
• Clearer desktop invitation - updates desktop Chat, Designer, and Developer composer copy to “Do anything…” while retaining the compact mobile wording.
• Google adapter modernization - moves to the consolidated langchain-google-genai 4.x adapter and langchain-core 1.6 line, removes the retired google-ai-generativelanguage runtime dependency, and updates effective tool-schema inspection for the new SDK declarations.
• Google feature preservation - verifies typed array and union tool schemas, streaming, multimodal input, thought-signature replay, reasoning defaults, and asynchronous cancellation through the consolidated adapter.
• Routed Claude history repair - recognizes Claude routes through OpenRouter and Requesty, consolidates late system messages for Anthropic Messages, keeps tool-call/result groups valid, and leaves the durable checkpoint transcript unchanged.
• Google history repair - normalizes replayable Google reasoning or thinking blocks and drops private incompatible blocks only in the provider-facing copy so switching from another transport cannot fail before the Google request.
• Awaited mobile submission - keeps mobile send handling inside the UI callback lifecycle instead of spawning an untracked task, improving error propagation and deterministic command behavior.

Documentation, Dependencies And Release Validation

• Reasoning user guide - adds a dedicated public Reasoning Controls page, navigation, provider/settings links, /reasoning examples, persistence and fallback behavior, privacy guidance, and troubleshooting, then regenerates the searchable documentation artifacts.
• Context documentation sync - updates generated settings references and published pages for the 64K local Auto target, exact custom values, custom server capacity semantics, and model-scoped endpoint evidence.
• Post-4.7.1 site completion - aligns the already published landing-page downloads and contracts with v4.7.1 before the v4.8 documentation work.
• Provider and agent coverage - adds deterministic coverage for reasoning capability resolution, persistence, commands, transport payloads, fallback classification, duplicate callback prevention, custom endpoint context and probe isolation, fixed-envelope failures, compaction fallback, and dynamic OpenCode refresh-to-runtime behavior.
• UI and compatibility coverage - adds responsive desktop composer, accessible compact-control, context-meter, mobile reasoning, Google adapter, routed-Claude normalization, tool-schema, Gmail, Goal, Requesty, and cancellation contracts, and updates source-to-test ownership for the new provider modules.

Breaking Changes And Caveats

• No application data migration or public CLI break is introduced by v4.8.0; model and thread settings migrate in place.
• Existing local Auto context settings now target 64K instead of 32K and can therefore use more Ollama memory. Select a fixed 32K context if the larger allocation is unsuitable for the machine or model.
• A Custom/Self-hosted model with unknown native context no longer inherits a generic fallback. Refresh or probe it, declare the server's native limit, or set a verified Custom cap before use.
• A Custom server context cap limits Row-Bot's planning and requests; it does not change the server's loaded allocation. Configure llama.cpp, vLLM, SGLang, LM Studio, or another server independently.
• Reasoning options, latency, token use, and billing are model- and provider-specific. Provider default remains the compatibility choice, and reasoning replay for a custom endpoint should be enabled only when that endpoint's format and trust boundary are understood.
• OpenCode catalog refresh now consults both the selected gateway and the public native routing registry. A failed refresh preserves the last known good rows or uses the bundled static fallback on a cold cache.

:::manifest
schema: 1
files:
  Row-Bot-4.8.0-Linux-x86_64.tar.gz: sha256=469e6d6dac43b65b65cbfb11376bd143726e09abfd6bfb84dd3da51a75a9b97b
  Row-Bot-4.8.0-Windows-x64.exe: sha256=43440e564e5cc5b859fe6065563ed5327cb3669e385600e4ec66d41b3c1e788b
  Row-Bot-4.8.0-macOS-arm64.dmg: sha256=da3a056b0b6c7bc114fb6e881b5d0e3003f2d90be3e52ff1c5d635392644f7ce