Release Notes

v1.0.17 — September 10, 2026

A stability release that fixes a DevBrowser crash on long-running page scripts.

Fixes

• TatsuCode crashed when a DevBrowser script ran long — If a script the agent ran in DevBrowser took more than five seconds (a page busy with a large upload, for instance), TatsuCode could exit without warning when it finished. The late result now lands safely and the script simply returns empty. The same fix covers page HTML capture and DevTools setup

v1.0.16 — September 9, 2026

A same-week model-support and diff-fidelity release. OpenAI's brand-new GPT Image 2.5 models — Sunburst and Flare, released September 8 — arrive in both the OpenAI direct and OpenRouter image pickers, and Claude Code's diffs finally show the file's real line numbers.

New Models

• GPT Image 2.5 Sunburst and Flare (OpenAI direct and OpenRouter) — OpenAI's new image models, released September 8. Sunburst is OpenAI's most capable generation and editing model, built for workflows where editing precision matters most — production creative, multi-turn edits, identity-sensitive changes. Flare is the fastest tier for everyday generation, at up to half the latency of GPT Image 2. Both add Extra High and Max quality presets above High, accept up to 16 reference images, and on the direct OpenAI provider produce transparent PNG or WebP output. Sunburst becomes the default for new OpenAI direct setups; an existing selection is left as it is

Improvements

• Transparent output on GPT Image 2 — OpenAI now supports transparent backgrounds on GPT Image 2 in preview, so the Transparent option appears for it in /image-generation on the direct provider. GPT Image models routed through OpenRouter remain opaque-only, and the agent's provider guidance says so
• Reference limits match OpenAI's documentation — Every direct GPT Image model now accepts up to 16 reference images per edit, as OpenAI's API reference states; the picker previously capped them at five
• Subscription mode reports what it really uses — When OpenAI access is set to Subscription, TatsuCode hands the prompt to ChatGPT's built-in image generation and OpenAI chooses the model without ever naming it in the response. The generated-image card, the /image-generation preview, and the model bar used to claim GPT Image 2; they now show ChatGPT Images (managed), and the provider information the agent sees says the same, so all four surfaces agree with what actually happens
• Quality help text follows the model — The quality hint in /image-generation now lists the presets the selected model actually supports instead of a fixed low/medium/high line

Fixes

• Claude Code diffs start at the real line number — Every Edit diff shown under Claude Code numbered its rows from 1, wherever the change sat in the file. The diff is now built directly from Claude Code's own patch rather than re-derived from the edited text, so both gutters show the file's real line numbers, an edit that changes several places at once (a replace-all, for instance) gets a separator between regions with each region numbered correctly, and the "no newline at end of file" marker no longer appears as a bogus line. Reopened sessions keep the same numbers
• Restored new-file diffs — A diff for a file the agent created rendered as two columns with an empty Original side after reopening a session. It now restores as the same single New column it showed live

v1.0.15 — September 6, 2026

An image-generation and chat-polish release. Image generation through OpenRouter grows from three models to a full catalog with output up to 4K, and the chat gets two long-overdue fixes for busy agent sessions: every edit now shows its diff right where it happened, and very long sessions stay responsive instead of landing updates in bursts.

New Models

• A full image catalog on OpenRouter — The OpenRouter image picker now offers Nano Banana 2, Nano Banana Pro, and the new Nano Banana 2 Lite; GPT Image 2; MAI-Image-2.6 and 2.6 Flash; Seedream 5.0 Pro and Lite; Grok Imagine Image 2.0; Qwen Image 3 Pro; Riverflow V2.5 Pro and Fast; FLUX.2 Max; and Recraft V4.1 Pro. /image-generation shows the sizes, aspect ratios, qualities, and backgrounds each one supports, and Riverflow V2.5 Pro brings transparent PNG output to OpenRouter
• Nano Banana 2 Lite (Google direct) — Google's lightweight image model (Gemini 3.1 Flash Lite Image) joins the Google direct picker for quick 1K generations

Improvements

• Claude Code's tool calls show up as they happen — When Claude Code worked through a long run of edits and reads in a single response, TatsuCode's chat could sit on "Claude thinking…" for minutes and then drop dozens of cards at once, even though the terminal showed each step live. Claude Code only hands its transcript a message's tool calls once that message is finished, so TatsuCode now listens to Claude Code's own tool hooks instead and renders each card the moment the call runs. The status banner cycles through its thinking phrases between calls instead of freezing on one
• Full-size images on OpenRouter — for edits too — Image requests now go through OpenRouter's dedicated image endpoint, so every size a model advertises, from 512px up to 4K where supported, works for fresh generations and for edits with reference images. Asking for 2K or 4K used to fail with an "image_size is not supported" error, and size and references couldn't be combined at all. A few models (GPT Image 2, MAI-Image, FLUX.2 Max, Recraft) pick their own resolution, so only the aspect ratio is offered for those
• Retired image models move over automatically — If the image model you had saved has been retired from the catalog (Riverflow V2 or Seedream 4.5, for instance), TatsuCode switches you to its successor on its own. A saved size the new model can't produce resets to Auto instead of quietly downgrading you
• Large generated images preview in chat — Very large results, such as a 4K PNG over 20 MB, used to show a broken preview because they were bigger than the chat can receive in a single message. Those images now stream from TatsuCode's own image server, the same way the gallery loads them, so the inline preview and the lightbox work at any size
• Long sessions stay responsive — On very long agent sessions, with thousands of tool cards in one window, each new card was getting slower to draw until edits started landing in bursts of fifty at a time instead of one by one. The chat now keeps the most recent stretch of the conversation on screen and quietly brings older messages back as you scroll up, the same way reopening a session already pages history in. Nothing is removed from the saved session, and the bottom of the chat, where you're watching the work happen, is never touched

Fixes

• Edits show their diff where they happened — When the agent edited several files in quick succession or in parallel, the chat showed all of the Edit cards first and then all of their diffs in a clump underneath, so you couldn't tell which change belonged to which file until it was over. Each diff now appears directly beneath its own Edit card the moment it lands. This applies to every provider, not only Claude Code

v1.0.14 — September 5, 2026

A model-support and workflow release: OpenAI's new GPT-6 Astra arrives in the OpenAI direct picker, and Claude Code stops missing your mid-task follow-ups.

New Models

• GPT-6 Astra (OpenAI direct) — OpenAI's most capable model for complex, demanding work is available as GPT-6 Astra, with its full 872K-token context (the Codex maximum) recognized from the first message, original-detail image input, native web search, and structured outputs. Availability is gated on OpenAI's side — Trusted Access Program enterprises first, broader rollout following — so if your account doesn't have it yet, requests will fail and GPT-5.6 Sol remains the flagship to use
• Max reasoning effort for OpenAI direct — The effort picker gains a Max level above Extra High. GPT-6 and GPT-5.6 accept it; older models keep their top supported level automatically

Improvements

• Claude Code sees queued follow-ups mid-task — When Claude Code is your main model and you type while it's working, it used to notice the queued message only when a TatsuCode tool ran. A long stretch of file reads, edits, or shell commands could run to the final response without ever seeing it, and you had to tell it to check the queue. A one-time reminder is now delivered right after the next tool call whenever new input is queued, so Claude reads the queue at its next safe break — once per queued message, never on every tool call
• Less waiting, faster starts with Claude Code — Send long prompts and get to work much faster. Detailed instructions, code blocks, and multi-paragraph messages now arrive in a fraction of the time, with your formatting intact
• Claude Code questions show their context first — When Claude Code asks you a question mid-task, the explanation behind it used to appear only after you answered, because Claude Code holds its narration until the question is resolved. The question tool now carries that background separately, so it renders as a normal assistant message directly above the question card, stays in the chat after you answer, and is restored with the session

Fixes

• GPT-5.6 Sol context window sized to the Codex maximum — Sol was previously configured for a 1-million-token context, which is larger than Codex actually allows. It now runs at Codex's 872K-token maximum, so the context bar and auto-compaction match the real limit
• "Claude Code never created its session transcript" on a long first message — Starting a fresh Claude Code session with a very long pasted prompt could trip TatsuCode's own startup watchdog and tear down a session that was actually running fine. The watchdog now accounts for how the prompt is delivered, and with the paste change above the delay it was guarding against is gone

v1.0.12 — September 1, 2026

A same-day model-support release. Anthropic's brand-new Claude Fable 5.1 — released today — is now selectable in Claude Code, both as your main model and for delegated agent runs. It also ships two important fixes — the video tool and screen recording work again in release builds, and switching between Claude Code models no longer loses your conversation context — plus visible cross-session messaging between TatsuCode windows.

New Models

• Claude Fable 5.1 (Claude Code) — Anthropic's latest and most capable model, released September 1, is available as Claude Code: Fable 5.1 in the model picker. It's built for demanding, long-horizon agentic coding and multistep research, and its full 1-million-token context window is recognized from the first message, so the context bar and the /usage panel stay accurate. The existing Claude Code: Fable entry keeps following whatever your subscription treats as the current Fable model; the new entry pins Fable 5.1 specifically, so you always know exactly which model you're on
• Fable 5.1 for delegated agents — The Claude Code Manage panel's Model dropdown gains the same Fable 5.1 option, so tasks the main agent hands off to Claude Code can run on it too

Improvements

• Cross-session messaging between TatsuCode windows (Claude Code) — Claude Code sessions can now talk to each other across TatsuCode windows, and the chat shows it. A message from another window renders as a distinct Message from <session> block (never as something you typed), the outgoing Send Message card collapses to a one-line summary with the full text a click away, and List Agents reports how many peer sessions it found. The footer shows this window's Claude session name — the address other sessions use to reach it — and the window title carries it too, so two windows on the same project are easy to tell apart. Renaming a session updates its address on your next message. This works with any Claude Code session on the machine, not only TatsuCode windows: a claude session in your terminal can list active TatsuCode agents and message them directly, and they can message it back
• Clearer video-tool errors — When a video can't be decoded, the error now includes the decoder's actual reason instead of a generic "no frames" message

Fixes

• Video tool and screen recording broken in release builds — Every public build since v1.0 shipped with built-in COM interop compiled out of the trimmed runtime, so the local video tool reported "No frames could be decoded" on perfectly good files, and screen recording, the taskbar badge, and drag-and-drop interception failed the same way. The runtime feature is now pinned on in the release pipeline
• Sheared or corrupted video frames — Some clips (for example 1142×954 sources) decode to a padded surface whose row pitch differs from the advertised stride, which produced diagonally sheared frames. The extractor now reads the surface's real pitch and infers padded layouts correctly
• Switching Claude Code models dropped the conversation — Changing between Claude Code models mid-conversation (for example Opus to Fable 5.1) rebuilt the session from a compact summary instead of resuming Claude's own transcript, so the context shrank and the most recent turns were lost. The native resume id now survives the model switch
• Broken tool-card headers — Commands containing HTML, XML, or C# doc-comment tags (a heredoc with <summary>, for instance) used to break the Bash card header into a stack of lines. Header text is now escaped at the source and stays on one line

Notes

• Plan requirements — Depending on your Anthropic plan, Fable-tier models may draw on pay-as-you-go usage credits rather than your subscription allowance. Nothing changes in TatsuCode — Anthropic bills you directly through the Claude CLI, exactly as before