When a run's time limit or a cancel fired, the program was stopped through a call that then waited for it with no limit (and on Windows for a job message the system doesn't promise to send), so a program that outlived its stop held the run past every deadline, Pause included. The program and whatever it started are still stopped, but the run now waits at most 5 seconds for them to be gone, then logs a warning and carries on.
Refs #1438
When some arguments of a Swift function, initialiser or subscript have labels, the others now read _: as swift-demangle prints them, so append(_:length:) reads append(_: Swift.UnsafeRawPointer, length: Swift.Int) instead of leaving the first argument bare. A single argument not wrapped in a tuple prints bare, as there. About 21,000 symbols from system binaries now match.
A Swift symbol ending Ml is the cache the runtime fills the first time it lays out a type's metadata whose layout is only known at run time, and one ending ML is the type's lazy metadata cache variable. Both were named "lazy cache variable for type metadata for"; Ml symbols now read "type metadata singleton initialization cache for".
A job of unknown length drew its progress bar with egui's animated ProgressBar, which asks for a frame every frame, so the window redrew at the display's rate for as long as the job ran, even in the background (as the spinner did before #579). The jobs list now uses our own progress bar: filled to a known fraction with no frames asked for, or sweeping at the animation's pace (once a second in the background) when the fraction isn't known.
On Unix a folder's name may hold a backslash, and turning it into / named another package (cmd/foo\bar became ./cmd/foo/bar); Go reads a backslash as a separator, so keeping it wouldn't help either. The package is now written from the path's folders joined with /, the same on every system, and a command folder whose name holds a backslash gets no inferred launch.
A Go project's command package was written with the system's separator, so on Windows the launch test expected ./cmd\viewer while the launch gave ./cmd/viewer (right for go run only by accident, from a mixed path). {package} now always writes the folder with /, as Go names packages, and the test expects ./cmd/viewer.
Look for AI suggested the first chat model a local server listed. LM Studio lists every model downloaded and FastFlowLM every model it can download, and asking for another one loads or downloads it. A local server's suggestion is now a chat model it says is loaded, from Ollama's /api/ps or LM Studio's /api/v0/models (embedding models left out), and the first chat model listed only when nothing says what's loaded. Adding LM Studio by address gets the same.
Refs #1434
Adding a server by its address with no model named saved the first chat model its /models listed. On a server whose list is a download catalogue (FastFlowLM), that started a 5.5 GB download and unloaded the model in use. The server is now asked which models are loaded (/api/ps, which Ollama and FastFlowLM answer) and a chat model from there is used; otherwise the first chat model in /models, as before. A model the person names always wins.
Refs #1434
Adding a model server by its address showed nothing while it connected and gave up after 10 seconds, then said to check the address even when the server had been reached. The form now shows "Connecting to <name>…" with a spinner and the seconds waited, and Add can't be pressed again meanwhile. The server has a minute to list its models once reached, and one that was reached but didn't answer in time is told it may still be starting or loading a model, to wait and add it again.
Fixes#1429
A build on an SSH host started Craft Anything's tool server by the bare name craft-anything-cli. The pre-build check found it, but Claude Code starts the tool server with its own PATH, so the server failed and the build stopped with "Craft Anything's tools did not start in Claude Code". The check now prints where the helper is, and the tool server is started by that absolute path; when the host gives no absolute path, the bare name stays.
Since a turn now ends shortly after its stream's result line, that line alone decides whether a lingering process counts as a finished turn. Gemini CLI's result event wasn't checked for a failure; one whose status isn't "success" now fails the turn with its error, as Claude Code's already does.
A coding agent's turn ended only when its process exited, so when Claude Code printed its final result and then kept running (a shutdown hook, logging), a finished build stayed on "Building" or "Pausing" for as long as the idle limit allowed, hours for a builder turn. The turn now ends a short grace (5 s) after the stream's own result, without being reported as stopped or failed; a deadline hit before any result still stops the turn as before.
Each request starts a new run, and a run after the first skips the study, so its session had nothing open and the builder's first native tool calls said "Nothing is open yet". In Deep mode, a run that skips the study now opens the program the study looked at (found the same way, so a game given as a folder gets its executable), so the builder's tools start with it. Clean room, a builder over SSH and a program that's gone open nothing.
Fixes#1421
The run removed its mark on a builder's check right after answering, so an answer that landed as rebuild_check's wait ended made a taken check read as "No build is running". The mark now stays until rebuild_check has read it. A check interrupted by stopping the build now answers that the build was stopped, so a builder still waiting isn't told the check is still going.
A coding agent's turn killed mid-turn leaves a record of its session, which was removed only at the next agent turn or on uninstall; meanwhile the agent's own background service could carry on a session nobody was watching. The desktop app now sweeps those records when it starts, on a background thread, with the same rules as any sweep: a record a running turn holds is left alone, and a delete that fails is kept for the next sweep.
Two edits to the README's AI paragraph had landed on top of each other, so it broke off mid-sentence, repeated the Copilot and ACP text, and said both that only Maki and that Maki and Devin had been run live. It now has one version with both updates: Devin among the ACP agents and xAI among the providers.
Started from the Dock, the app has a short PATH, so a Codex installed inside the ChatGPT or Codex app wasn't found. It's now looked for in those apps' known bundle locations (never the app's own executable), after an explicit program and PATH. Testing a local coding agent in Settings → AI (or `ai test`) now also asks its adapter whether it's signed in, without sending a prompt, and Codex only gives sign-in advice for real authentication failures.
A coding agent's saved session counted as the project's history whatever mode it ran in, so after switching between Clean room and Deep a run could skip its study and build from the other mode's capture. A saved session or transcript now counts only in its own mode (one saved before the mode was kept counts as Deep); one from the other mode makes the run study again and replace the capture before the builder is briefed, including an API follow-up whose transcript was rejected for its mode.
rebuild_check said "no build is running" when the run had taken the check and one of its steps hadn't finished, and the builder gave up after its default 20 minutes, below the run's own 30-minute limit on a check. The run now marks a check request it takes, the tool says "still going" apart from "no run took it", and its default wait (35 minutes) outlasts the run's limit, so the run reports a stuck check itself.
Whether a build turn gets the Deep-mode decompile reminder is now one function the run calls, with a test that reopens the project from disk as a later request's run does: reminded in Deep mode on a program with nothing decompiled yet, and not in Clean room, after an earlier run's decompile, or without a program target. Behaviour is unchanged.
The English README explains how to check that a Mac download is signed and notarized; the Japanese, Korean and Chinese READMEs stopped before that section. Each now has it, translated, with the same commands.
The Traditional Chinese README's Craft App Suite table had no MaterialCraft row. It now lists it after TextureCraft, as the English README does, with the same links and its description in that language.
The Portuguese README's Craft App Suite table had no MaterialCraft row. It now lists it after TextureCraft, as the English README does, with the same links and its description in that language.
The French README's Craft App Suite table had no MaterialCraft row. It now lists it after TextureCraft, as the English README does, with the same links and its description in that language.
The Spanish README's Craft App Suite table had no MaterialCraft row. It now lists it after TextureCraft, as the English README does, with the same links and its description in that language.
46 strings shown in the app had no Japanese line, so those screens showed English in Japanese. They're now translated in the catalogue's existing wording; only the product name, a test string and the ellipsis stay as they are.
In Deep mode, a builder that hadn't decompiled anything was reminded to, but whether it had was kept only for one run, and each request from the person starts a new run. So a builder that had read the original's functions earlier was told "You haven't decompiled anything yet" after every request. It's now kept in the project's craft.json, so a decompile in any run counts.
A question answered from a fork of the builder's Claude Code session reported the whole session's running total, so its cost was hidden. It's now what the total added since the builder's last turn, read before the question starts, and unknown when there's no kept total, never the session's total. The question's cost is added to the project's usage, and a build turn and a question saving their costs at the same time no longer overwrite each other.
A GNU long-name or long-link record whose text was complete but whose 512-byte padding was cut off, or whose padded size overflowed, stopped the listing without saying why, so the engine installer could still write the files listed before it and report success. Both cases now count as a cut-short GNU record, and the installer refuses the archive before writing anything.
An ASAR manifest whose nodes aren't objects, are both a folder and a link, have a link that isn't text, have no size, a non-boolean `unpacked` flag, or a packed file with no offset is now reported as damaged instead of being listed with made-up values.
When a tar listing stopped at a GNU long name or link it refused (too long, or cut short), the engine installer still wrote the files listed before it and reported success, which could replace a working engine with part of one. The listing now says why it stopped, and the installer refuses such an archive before writing anything, with a plain-English error. An archive missing only its end marker still installs.
ZIP entries compressed with LZMA (method 14) were listed and not read. They're now extracted with the same output limit, size and CRC checks as other methods, whether the stream gives its size or ends with a marker; the dictionary is held to the output limit, and properties of another size are reported as unsupported.
A GNU sparse member stores only some extents of its file, and was extracted as if those were the whole file. It's now listed as another kind of member with a warning, under the name GNU.sparse.name gives when there is one.
A versioned ELF import names its library while its relocation slot doesn't, so since slots were matched by name and library, such imports lost their slot address. An import now takes a slot of its own library first, then a slot with no library only when the name belongs to one library and has one address.
A standalone LZMA stream's declared size is now checked against the output limit before decoding, and its dictionary is held to the limit, so a damaged stream can't grow a large window before failing.
ZIP entries compressed with bzip2 (method 12) or XZ (method 95) were reported and not read. They're now extracted with the same output limit, size and CRC checks as deflated entries.
A gzip member in a packed bundle was inflated without its trailer being checked, and only its first gzip member was read. Every member is now read, within the size limit, and each trailer's checksum and length must match.
A gzip header with the optional header checksum (FHCRC) is now checked against it. The extra-flags and operating-system bytes, which are only informational, no longer make a valid header be refused, while reserved flags still do.
A member whose stored bytes reach past the start of the archive's table, into the table, cookie or a signature after it, could be read as that member. Such members are now unreadable, with a note counting them, and a PYZ there supplies no nested members.
A method list whose record size was smaller than a method record, or whose last record was only partly in the file, was still read, so methods came from whatever bytes followed. A list now needs records of at least 24 bytes (12 for relative ones), each wholly in the file, and longer extended records still read.
A Godot pack embedded in an executable ends before its trailer, and a member's data could run into the trailer. The file table and members are now read only before it, and a member that would reach past it isn't readable.
A `long` or `double` constant takes two constant-pool slots, and one in the last slot was read past the pool's declared count. Such a constant is now refused as a damaged class, with no overflow at the largest pool size.
Cooked Unreal textures in half- and full-precision float formats (R16F, G16R16F, FloatRGBA, R32_FLOAT, G32R32F, A32B32G32R32F and the filtered half aliases) were reported as unsupported. They're now decoded with the shared float reader: colour becomes 8-bit sRGB, alpha stays linear, and values outside the range are clipped.
The metadata reader took a header offset past the end of a short input and read past the declared metadata into object bytes. The offset and size are now checked against the object data, the declared file length and the bytes supplied, and only that range is parsed.
Each repeated name or producers section got its own allowance, so together they could keep more than the item limit, and the cutoffs left metadata out without saying so. The function-name and producer limits now apply to the whole module, names already kept can still be updated at the limit, and the report says when anything was left out.
A resource chunk could fit in the file while reaching past the size its table or package declares, so string pools, library records and short headers could read unrelated bytes. Each chunk is now read only within its parent, package and known child headers must be complete, package string pools stay inside the package, and only complete library records are kept.
A class-data group cut at the member limit left its unread records where the next group would be read, so they were taken as that group's members, and a typed-handler list that stopped early could turn its remaining bytes into a catch-all address. The class data now ends as incomplete before an unread group, and a catch-all is read only after every typed handler.
A private browser or driven Electron app was stopped only when its process handle was dropped, which a crash, a kill, a force quit or an MCP server stopped by its client never does, so the browser was left running and its profile stayed in the temp folder. On Unix a small watcher now stops the browser's process group and deletes its `craft-*` profile once the app is gone, and ends by itself as soon as the browser does. Windows children are held by a Job Object; whether one outlives an abrupt exit is left as a roadmap line.
Refs #810
The indirect symbol limit applied to each section on its own, so many sections could still read an unbounded number of slots. It now applies across all sections, with a warning when slots are left out.
An `LC_UNIXTHREAD` entry point was read from the first state record whatever its CPU or length. It's now read from the record whose flavour matches the file's CPU (x86_64, arm64 or i386) and that holds a complete state, so a short or foreign record no longer gives a made-up entry point.