The two start guard tests share the process-wide held mark, so one's opened() could drop the other's lock mid-test and fail main now and then. They now take a shared test lock.
A note sent while a build runs now carries the latest Just ask answers its mode may show, including through Claude Code's note hook, and they count as told only once the builder has them. A Clean room run never gets Deep answers.
SRV, SOA and CAA answers show their fields, address records of the wrong size are incomplete, a truncated reply says so, and unsolicited mDNS announcements keep the address under the record name.
Requests like "a low level disk tool", "an MPEG 2 player" and "a single player app for movies" stay apps, while other game cues in the request still win.
Rejected packet blocks share the packet budget and are reported once per reason, a timestamp fraction outside its second is ignored, and sub-nanosecond resolutions keep the interface offset.
A ClientHello with impossible vector lengths, repeated extensions or repeated host names is marked partial, and only registered GREASE values are dropped.
A tree and screenshot that would pass the helper's 16 MiB output cap now fit: the tree is cut to a prefix and reported as truncated, keeping the image, and an image that cannot fit alone becomes a screenshot limitation.
`runtime_ui_read` with `screenshot: true` returns one on-screen window of the pid as a PNG on macOS 14+, restricted by `window_id` when given. Missing permission, a missing window or a failed capture is a limitation with partial coverage, and no permission dialog opens.
Fixes#248
The DeepSeek Harness row says the macOS run renamed an existing craft-anything MCP registration, because the generated overlay uses the same server name.
`rebuild --continue-in`, `rebuild_start` on a project and `rebuild_question_ask` now say which AI works when the saved one is off or gone, as the app does.
The preamble and task go on stdin instead of the command line, so a multi-line or long task no longer fails in the Windows desktop app's batch launcher.
A saved AI that was turned off or removed no longer gets replaced silently: a notice says which AI builds instead, so a build can't quietly use one the person didn't choose.
The installed skill says that `setup --yes` writes the plain `mcp` command and so drops an `--all-tools` already in an entry; keeping the person's arguments is on the roadmap.
Someone signed in with `grok login` can build with their installed `grok` through ACP, with our approver deciding permission requests. Clean room, deep chat and questions are refused before it starts, because Grok can't keep to them; only Grok's sessions send `"type": "stdio"`.
When a server refuses a request and says how big its window is, the agent keeps that size for the backend, makes room to three quarters of it, measured with the server's own count, and asks once more; the error says how to give the model more context.
Closes#962.
A saved bare `claude`, or a saved path to the desktop app's own file, goes through the same lookup as the usual name: the real Claude Code when there is one, else a plain error.
The test that a stopped launch takes its descendants with it uses a parent and child started without a shell, so it runs the same on Unix and Windows, and a failed start still stops what it launched.
An Approvals menu beside the builder's-questions switch offers Default, Allow, Ask or Refuse for each kind of action, and `/approvals <kind> <choice>` does the same from the prompt box without sending anything to the builder.
The check for the project's records reads a command with quotes and escapes taken out, as the shell joins them, and any mention of `CRAFT_ANYTHING_AGENT` counts as a system change. docs/claude-code.md says these raise the bar and aren't a wall.
A project can allow, ask about or refuse each kind of builder action (web, commands, outside files, publishing, deletes outside the project, system changes, other tools); a kind it hasn't chosen follows Auto mode as before. `project_approvals_set` sets them, and an agent can only make a kind stricter, including through our CLI from its own shell. "Always" on a push, delete or system card allows that whole kind, and pushes are read word by word in every shell form.
A valid fenced tool call followed by a summary still isn't run, but the model is now told why and how to send it again, so the build doesn't stop with nothing written.
Without Screen Recording the capture is refused before the program starts, with where to allow it. The name fallback matches whole words and never the system's own windows, and a window that's late gets up to 10 more seconds.
Ghidra's public downloads leave out macOS's native programs, so a Deep build analysed for many minutes and then decompiled nothing. It now stops at once and says which command builds one; an install laid out in a way we don't recognise is left to Ghidra.
The block table parser checks every compressed block against the data region's size, so blocks can't reach into a trailing table or bytes after the declared bundle end.
The OpenCode remote setup test no longer expects one exit code from `command -v` for a missing helper, and runs with private PATH folders so a helper installed on the machine can't hide the failure.
STATUS.md no longer says the project feed, engine installs and launching exist only in the desktop app; it describes the shared tools and their limits.
A saved program path that's gone now goes through the same filtered lookup as every other path, and when only the desktop app's `claude` is on the search path the run stops with a plain error and how to install Claude Code.
Closing the app stops analyses and waits for them to end. A Deep build no longer waits on Ghidra for an Electron program, whose logic is in `app.asar`.
Closes#946.
Closes#947.
DeepSeek Harness (`dsh --profile headless`) as a coding agent; the tool server arrives as a `--patch` overlay in a private folder, so the person's profile is untouched.
Co-authored-by: 0231071 <747146906@qq.com>