A client could connect and send nothing, or only blank lines, and hold one of the 16 connection
seats for as long as it liked. All reads before the token now share one five-second deadline; the
connection is then closed without a reply. An authenticated connection may idle as before.
`solidcraft-cli mcp --connect` used to open a connection at start and hold it until the first tool
call. It now only checks that the app is there, and the first call connects.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #8315
* fix(control-server): release the watch seat deterministically (#8315)
The seat a subscribed connection holds on the DesignWatch lived in the
events thread alone, so it left the watch only when that thread ended:
on the first write a closed client refuses, which needs the design to
move, or when the thread noticed the end of the stream, which waits for
the connection's own thread to come out of its blocking read. A closed
connection could therefore keep counting on the watch for a second or
more, and the test that asks for the count to be gone could lose its
race with the ten seconds it waits.
The seat is now shared: it leaves the watch once, whichever thread sees
the connection's end first, the connection's own thread at the end of
the stream or the events thread at the first refused write. The count
of subscribed connections is right as soon as the connection is gone,
and the events a subscribed connection is told of do not change.
* claim: #9184
* Control channel: a request's own param_check level, applied and confirmed by the app (#9184)
`engine.execute`, `engine.batch` and `engine.script` take an optional
`param_check: "off"|"warn"|"strict"` in their params. The app runs the
request at that level through the same `run_checked` path and batch
preflight a headless session uses, puts its own level back, and every
reply says the level it ran at as `param_check` beside `ok`. A command
that switches designs keeps the app's level; any other word is refused.
Readings and exports answered on a copy off the window thread keep the
level too.
`control.capabilities`, answered at once on the control-server thread,
lists what the channel does beyond the methods every version has:
`{"param_check": ["off", "warn", "strict"]}`. An older app answers
"unknown method".
The MCP bridge asks it once per connection. When the app lists the
level, every call and batch goes with it and the app checks against its
own commands; otherwise the bridge checks before sending, as #9073 does.
At strict, a reply that doesn't say the app checked the call is an error,
and the next calls are checked by the bridge. Also: `add_param_warnings`
takes the reply by `&mut`, and `with_param_check` sits beside
`Remote::new`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Control channel: a design-switching command is checked at the request's level
Review fixes for #9184. `component.edit_in_place`, `component.finish_edit`,
`view.next_tab` and `view.previous_tab` ran at the app's own level and
echoed `off`, so a misspelt call went through at strict. They now have
their parameters checked at the request's level before they run (and
still run at the app's own level, as each design keeps its own), and the
reply echoes the requested level. The MCP bridge at strict now requires
the echo to match its level, not just to be there. `param_check` is read
only by `engine.execute`, `engine.batch` and `engine.script`; other
methods ignore it. The off-thread test now shows the design change
computing off the window thread before it is refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #8794
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019FAWPuFoh9qxn2uS2j53QZ
* Control token: a closing window removes only its own token and discovery files (#8794)
Under Flatpak each instance has its own loopback, so a second window
binds the same port number as the first and writes the same
<port>.token, <port>.pid and current.port in the shared data folder.
At exit the first window deleted them all, unconditionally, leaving the
second window running with no token file a client could read.
remove_token_file now takes this launch's token and removes the token
and process files only while the file still holds it; the process id
can't tell instances apart, as every Flatpak instance has the same one.
remove_current_port also takes the token and keeps the discovery file
while another launch's token file holds that port, and removes
current.pid with current.port (it was left behind).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019FAWPuFoh9qxn2uS2j53QZ
---------
Co-authored-by: Claude <noreply@anthropic.com>
The seat of a subscribed connection on the DesignWatch lived in the
connection's serve thread. It left only when that thread saw the end of
the stream, so one unbounded blocking read stood between a closed client
and the count. Under load (GPU backend probing and crash-dump children
on Windows) that read can sit for seconds; the test's 10 s deadline then
expired and `!WATCH.subscribed()` failed (#7226).
Move the seat onto the events thread. The thread ends on `stop`, or on
the first write a closed client refuses, so the next publish retires the
seat. The test now publishes while it waits, and its wait no longer
depends on the serve thread.
Until a connection had shown the token, its first line (up to 16 MiB)
was parsed whole into a JSON value. A line of [0,0,...] that size is
hundreds of MB once parsed, so 16 connections without a token could
make the app run out of memory and lose unsaved work. The token is now
read from the request object without building the rest of it: other
fields are only skipped, and anything but an object with a string
token is refused as before.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #6646
* Control channel: document.revision, a design revision and a document id (#6646)
A program that follows the open design asks document.revision and exports
only when the answer moves: document_id names the open design (kept through
edits, saves and renames, another for each design started or opened, never
shared between two tabs or two runs) and revision moves whenever the design,
what is hidden or an uncaptured component position changes. It stays put on
reads, selections and view changes, unlike the session's view revision.
The revision is read lazily: the session remembers the design it last
reported (held weakly, so Arc::make_mut can never change it in place and its
address can't be reused) and takes a new number from one counter for the
whole run when the design is another one. The control method answers at once,
also while a slow command computes; engine.execute reaches the same command
for scripts and MCP.
* document.revision: hash the run's id prefix with the process's random keys
A clock set before 1970 gave every run the prefix 0, so a restarted app could
answer a new design with the id and revision of the old one. The docs no
longer promise an answer at once: a screenshot waiting for an opened design to
be meshed still holds the requests behind it.
* document.revision: its label in the complete catalogs (ja, zh)
* claim: #6985
* Control channel: control.subscribe and design.changed events (#6985)
A connection that sends control.subscribe gets the active design's id and
revision in the reply (as document.revision answers), then a
{"event":"design.changed","params":{document_id, revision}} line whenever
either changes, so a live link need not poll.
The window publishes the design at the end of each frame's logic and
drawing while a connection is subscribed (one atomic load otherwise). Each
subscribed connection waits on that from a thread of its own and writes
the latest after a 100 ms gap, so a burst of changes is one line. Replies
and events share one lock per connection, each written whole (#6390).
* control.subscribe: the reply before any event, a second one refused
The events thread started before the subscribe reply was written, so a
change right after the window's answer could reach the client before the
reply that named the older revision. The subscription now starts holding
the connection's line lock and writes the reply under it.
A second control.subscribe on a subscribed connection is refused on the
connection itself: its answer could arrive behind a newer event. When no
thread or socket is left for the events, the reply says it did not
subscribe instead of ok.
* control.subscribe: a second one does not lose a finished late reply
The already-subscribed answer came before the busy check, so a request that timed out and had finished since lost its reply: control.pending then said the app ended it without one. It now comes after the busy check and drops a finished late reply as every other request does.
* control_server: declare the subscribe guard at the top of the loop
Clear of the lines #6643 changes (its token check moves to where the declaration was).
* control_server: the subscription's items beside serve
Away from the lines next to send, where #6643 adds token_of.
* control test: New on a blank design reuses its tab, so redo the box first
---------
Co-authored-by: MateuszSychXpl <151755914+MateuszSychXpl@users.noreply.github.com>
* claim: #6389
* perf: write each control reply to the socket in one piece
The control server wrote each reply with writeln!(out, "{reply}") on the unbuffered TcpStream.
Formatting a serde_json::Value hands the writer every string, number and punctuation mark
separately, so each became its own write system call. A reply of a few megabytes (component.list
on a design with thousands of components) took about half a second to send.
Every reply is now serialized to one buffer and written at once. Writing one large Value to a
loopback TcpStream: 1.79 MB took 1,052 ms with writeln! and 4.9 ms with one write_all.
The test counts the write calls for a 2000-component reply (one) and reads it back.
* claim: #4664
* Control: name the token file and spot a stale one when a call is unauthorized (#4664)
When the app refuses a token that the MCP client read from a token file,
the error now names that file, when it was written and how long ago,
without printing the token. The app also writes `<port>.pid` beside the
token file (its process id and, on Linux, its start time) and removes
both at exit; when that process has exited, the error says the token
file is stale. The process check runs only after a refusal, and the
token file itself stays one line, so existing clients keep working.
Document force-quit leftovers and the MSIX AppData redirection.
* Bench assistant: read the token out of the new ClientToken
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Rebased onto main as one commit (was 120d26f7, df488f9a, db9d1073 plus
merges of main).
- A request the app reports as still running (#1618) is waited for on the
same connection through `control.pending`; past the wait the error says it
was not cancelled and will still apply (or only write its file), later calls
are not sent behind it, and the new read-only `pending` tool returns its
result once it lands.
- A failed batch says what really happened: `rolled_back` only when the
design is back where the batch started (as #3288 / #3242 on main: the app's
`rolled_back` is believed; for an older app, every completed step undone;
only a rollback asked for is reported as one), plus `applied`, `not_run` and
`failed.still_running`.
- A batch step may name a tool (set_parameter, list_commands,
describe_command, execute, measure, inspect_design, undo, redo, export, save,
open, new_design, crash_reports); screenshot, body_topology, batch and
pending are refused before anything is sent.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The app opens its localhost control channel on port 27878 (or the next free
one) unless Preferences > Assistant turns it off; --control and
SOLIDCRAFT_CONTROL_PORT still win. A current.port discovery file lets
solidcraft-cli mcp find the running app with no arguments.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* claim: #3654
* Control channel: control.running lists what the app is running, across connections (#3654)
control.pending answers only for the connection that asks, so a second client,
the MCP bridge after a reconnect or a script saw {"pending": null} while the app
was busy. control.running answers for the whole app, the same on every
connection and at once (the connection's own thread answers it, also while it
is busy with a timed-out request): every request the app has been sent and not
answered yet, oldest first, with its connection, id, method, command id (or a
batch's step count), seconds since it arrived and whether it still waits its
turn. A timed-out request stays listed until the app answers it, also after its
connection closed. Parameters are never shown.
The window marks a request started when it takes it (ControlRequest.started),
which tells queued from running. The MCP server gets a read-only `running` tool.
---------
Co-authored-by: MateuszSychXpl <151755914+MateuszSychXpl@users.noreply.github.com>
* Agent chat prototype: drive an ACP agent the user already has from a side panel
A desktop-only chat panel (off by default) that starts a coding agent over the
Agent Client Protocol and gives it SolidCraft's MCP server, bridged to the open
window through the token-protected control channel with the default folder
limits. It finds installed agents (Claude Code and Codex adapters, Gemini CLI,
Copilot, Qwen, OpenCode, goose) with version and login state, streams text,
tool calls and plans, shows the agent's permission requests, and when an agent
isn't signed in shows its own login methods without handling any credential.
Design notes and findings: docs/design/in-app-agent-chat.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: say only what the end-to-end run showed
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Agent chat: gate, one undo per message, docking, process groups, discovery polish
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Agent chat: Settings → AI, several agents with a switcher, OpenAI-compatible endpoints, permission modes, toolbar button
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Agent chat: translations, tool cards, Add agent dialog fixes, test agent saves on its own
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Agent chat: third-party licences for the keychain crates; clippy
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Agent chat: badge text from the theme tokens
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Agent chat: Auto-allow model edits is the default permission mode
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Closing a socket with unread input resets the connection, and the reset could reach the client before the refusal did (requests_run_only_after_the_token failed 17 times in 100). The server now ends its side and drains, briefly and up to 64 KiB, what the client had sent before closing: 0 in 100.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The app's control protocol gets engine.batch: the steps run in one request (so one frame), each as engine.execute runs it, stopping at the first failure and rolling back to the checkpoint by default. The MCP batch tool sends it to a running app (with the control token), is never resent after a lost reply, and falls back to step-by-step when the app refuses (agent cursor shown, or a step loads a file). A batch may take up to an hour to answer.
solidcraft --control PORT writes a random per-launch token to a user-only file (<data>/control/PORT.token; 0700/0600 on Unix, icacls on Windows; SOLIDCRAFT_CONTROL_TOKEN overrides) and refuses any connection whose first request lacks it. mcp --connect, the demo and packaging smoke clients send it. MCP execute/batch/open/save/export/screenshot are annotated destructive and open-world, new_design destructive. Docs: control-protocol.md, mcp.md, AGENTS.md, pending release note in releasing.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>