Models looked for generate_3d with mcp_search, which only searched MCP
servers, and concluded it did not exist (the chat had been in plan mode, then
switched to build). The search now lists matching built-in tools first:
"available – call it directly" or "not in the read-only mode plan – ask the
user to switch to build mode".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- New tool: image-to-3D (project file or an image attached to the chat,
optional back/left/right views) or text-to-3D; image + prompt refines the
image with the prompt first (the 3D models take no text with an image).
Styles textured / geometry / lowpoly, optional PBR, face count. Runs through
fal.ai's queue API with progress, saves GLB + preview, books the cost
(published prices, overridable per model). Request bodies for the Hunyuan3D,
Tripo and Trellis families; defaults [images] model_3d / model_3d_text.
- Chat card: "Open in the 3D view"; media overview: kind "model" (also
STL/3MF/OBJ written by agents) with previews.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- opencode Go reports plan requests as cost 0 with the usage in allowance;
ModelConfig::cost values them at the model's list prices (charged amounts
and free requests stay as reported). Rows booked at 0 are valued once on
start (idempotent).
- opencode's price list now has per-duration cache prices for some models
({"30m": ..., "12h": ...}): the shortest duration is used, also in tiers.
Saving the provider's models failed with "untagged enum ModelEntry".
- Each answer shows tokens sent/received and the cost of the turn's requests
(usage.cost filled when messages are saved and loaded).
- The chat title keeps a minimum width in a crowded top bar.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Palette from the Tintenblau desktop theme (colors.toml, neovim.lua, zed.json):
ink blue with cream text, paper with dark ink. Also in the terminal client's
palettes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A git wrapper (docker/git-coauthor -> /usr/local/lib/codecraft/bin/git) adds
--trailer "$CODECRAFT_COMMIT_TRAILER" right after the commit subcommand; the
agent shell puts it first on PATH and sets the trailer in the script itself
(works inside the sandbox, which clears the environment). Default text:
"Co-Authored-By: {model} ({provider}) via CodeCraft" (server.commit_trailer,
Settings -> General; empty = off). Git skips duplicates (amend, model wrote it).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- AgentConfig.reverse_tools (Settings -> Agents & commands) gives any mode
binary_info and decompile; the clean-room check of the tools applies to
craft only (ToolCtx.clean_room = clean room && craft).
- decompile: optional processor (Ghidra language id) and base_address for raw
images (BinaryLoader); separate cache per processor/address.
- Skill "reverse-engineering": identify, convert HEX, processor table for
common chips, decompile, explore, notes in analysis/.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Performance (measured with 5,900 messages: open 9 s -> 0.7 s, typing 440 ms -> 25 ms)
- GET /sessions/{id}?limit=N loads the end of a chat (from the last summary
on, at most N); /sessions/{id}/messages?before|after|around loads more
(scrolling up stops at a summary unless asked). Without parameters the
full history as before (terminal client).
- Chat view: memoised message list (typing, streaming and status changes no
longer re-render the history), finished turns fold into "N steps · M files"
with the answer visible, tool card bodies and summaries are built only when
opened, scroll-to-bottom only when content changes.
Search: /sessions/{id}/search over the whole chat; a hit loads its part
(around) and marks the message, "back to the latest messages" returns.
Media: /sessions/{id}/media lists images, videos and audio of a chat and its
subagents (generated images, tool output, files written, media in answers,
attachments); ⋯ -> Media shows them with filters, a large view and a jump to
the message.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Context window
- Requests asked for the model's full max_output, but providers count
input + output against the window: long chats with large-output models
(DeepSeek V4.1 Flash, 384k) failed with "input too long" before automatic
compaction (85 %) kicked in. max_tokens is now what the window leaves
(estimated input = last usage + messages since), the chat is compacted when
less than 32k would remain for an answer, and a context-overflow error
compacts and retries once.
- ModelConfig.max_input (catalogue limit.input, editable in the provider
settings) lowers the compaction threshold for models with an input limit.
Enforced reviews (modes with refine.rounds > 0, refine.after_todos default on)
- Every todo marked completed is reviewed against the work since the last
review; with findings it is reopened and the feedback goes to the agent
(max. rounds per todo). Passed reviews leave a visible note.
- Ending the turn with open todos sends the agent back to work while it makes
progress (max. 10 reminders in a row without progress); such modes may take
up to 2,000 steps per turn. Reaching the step limit is reported.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
User messages are rendered with the same Markdown component as the model's
answers; single line breaks are kept (hard breaks outside code blocks), so
plain multi-line messages look as before. The eye button next to the text
field shows the formatted message before sending. The text sent to the model
is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mobile browsers scroll the whole page up to show the focused input and do not
always scroll it back when the keyboard closes; the app never scrolls the page
itself, so the top bar (menu) stayed out of reach.
- Viewport interactive-widget=resizes-content: on Android the keyboard shrinks
the page instead of shifting it.
- Elsewhere (iOS Safari) the page offset is undone when the visual viewport
grows again or the input loses focus.
- The chat keeps following the latest message when its pane shrinks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agents tidy up folders after generating images (sort into subfolders, rename
revisions); the image records kept the old path, so /api/images/<id>/file
answered 404 and only the requesting device still showed them from its cache.
- Generated images get a copy in the image store (images.blob, kept by the
blob GC); existing ones are copied once on start-up.
- Records follow: after bash commands with mv/rename (missing images of the
owner from the last two weeks, found by name or by content hash), when moved
with the file manager (file or folder), and on access (served from the copy,
relocated in the background; without a copy found by file name).
- Chat gallery falls back to the preview the model saw when the file is gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Automations
- Job files hold only the instructions; name, schedule, mode, model, approval
mode and time limit are settings in the database (jobs.settings). Existing
files are adopted on start-up (frontmatter into the database and removed from
the file); plain Markdown files registered later get default settings and stay
switched off until scheduled.
- New automation: name, project and a starting point (blank example structure or
a template) create .codecraft/jobs/<name>.md and open the editor: settings with
a graphical cron builder (next runs in server time via /jobs/schedule), the
instructions in the Markdown editor and an assistant chat
(POST /jobs/{id}/assistant) whose model edits the file; the editor reloads it
after each turn and warns about conflicts with unsaved edits.
- The schedule tool writes prompt-only files with settings.
Chat list
- Folders and projects keep a user-defined order and can be pinned (drag and
drop, or the group's menu), stored per user via /account/sidebar. A new
message moves the chat to the top of its group, no longer the project itself.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Automations: the approval mode is a selection in the list; PATCH /jobs/{id}
writes permission_mode into the job file's frontmatter (rest of the file
unchanged, invalid modes rejected). "Allow everything" lets nightly runs work
without waiting for approvals via Telegram until the timeout.
- Session.telegram_quiet: approvals and questions of a chat (and its subagents)
are not sent to Telegram; the chat open in Telegram always reports. Toggle in
the chat's ⋯ menu when a bot is paired. New chats from the web UI and
automation runs start quiet; chats started from Telegram and existing chats
keep reporting.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Installed apps checked their service worker at /sw-<old build>.js. That file no
longer exists after an update and the static fallback answered with index.html,
so the update check failed and the old worker kept serving the previous app
shell. The old frontend crashed on the new image references (base64Size of a
missing `data`), leaving a blank page on phones. Old service worker URLs now get
the clean-up worker (/sw.js), which unregisters and reloads; missing scripts and
styles answer 404 instead of HTML.
- Task list above the composer folds to one line (progress, current task),
remembered per device.
- max-height: 560px layout (square phones such as the Clicks Communicator,
landscape): tighter top bar, composer and new-chat wizard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Images of messages (pasted images, screenshots, generated images) were base64
inside the message JSON, so opening a chat downloaded all of them at once
(66 MB for a chat with 120 screenshots). They now live in a `blobs` table,
addressed by SHA-256; messages keep a reference (blob, size). The UI loads
them from /api/blobs/<hash> lazily with an immutable cache header, the runner
fills the bytes in before model requests, JSON exports stay self-contained.
- Store::add_message moves inline images out; large opaque PNGs become JPEG
(q85) when smaller. Older chats are converted once at start-up in the
background, unused images are removed (also after deleting a chat) and the
database is vacuumed when that frees 32 MB or more.
- Responses are compressed with gzip/brotli (not images, SSE streams or
binary downloads).
Example (copy of production): chat load 66 MB -> 1.5 MB (370 KB brotli),
database 90 MB -> 27 MB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On x86_64 the dynamic loader lives in /lib64, so the probe failed with
"execvp /bin/true: No such file or directory" and shell commands of non-admin
users were disabled on amd64 hosts. The sandbox itself already mounted it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
scripts/deploy.sh rolls a published image out to an installation set up with
move-server.sh (DEPLOY_TARGET in .env): checks the image for the target's
architecture, waits until no agent or terminal runs (--force skips), pulls and
recreates, runs the smoke test and goes back to the previous version on failure.
move-server.sh: --exclude leaves paths out of the workspace copy; values in .env
may be quoted (WORKSPACE='/home/...').
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Set GIT_AUTHOR_*/GIT_COMMITTER_* from Settings → Account → Git (also inside the
sandbox), so a repository's own user.name/user.email no longer wins; tell agents
not to change the git identity.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Richard Zink
Richard Zink