Agent tool calls (MCP, and bridged control requests carrying allow_paths) reach files only in the open design's folder (strictly inside home or the temporary folder, not in a hidden folder, and never its hidden files), the app's recovery folder, the server's private scratch folder and each --allow-path; --allow-any-path lifts the limit. Paths are resolved first, so .. and symlinks that lead out (including a save through a link, #845) are refused. On Windows, UNC, device, NT, verbatim non-drive, drive-relative, alternate-stream, reserved-device and 8.3 short-name paths are refused before any file system call, and the resolved path must be on a local drive. doc.recovery_discard and doc.recover {discard} delete only real recovery entries. Docs: mcp.md, control-protocol.md, README.
The MCP server instructions now point to `body_topology` for finding edge points (there is no `list_edges` tool), with a test that every tool the instructions name exists.
The control channel read each request with no length cap, so a client sending
gigabytes without a newline made the server thread allocate until the process
was killed. Read lines against the 16 MiB cap the MCP server already uses, now
one shared constant, answer a longer one with an error and close; close a
connection after eight lines that are not JSON, serve at most sixteen
connections at once, and run at most sixty-four queued requests per frame so a
flood cannot starve drawing.
Every file command checked the size with the metadata and then read to the end.
A character device such as /dev/zero reports a length of 0, so the cap passed
and the read was unbounded; a FIFO blocked the calling thread, which over the
control channel is the thread drawing the window. vfs::read now takes the
caller's cap, refuses anything that is not a regular file before opening it,
re-checks on the handle to close the window between the two, and reads no more
than the cap plus the byte that proves it was exceeded. Callers pass the limits
they already had, and the recovery read and the CLI's own reads go through it
too. `solvecraft-cli eval /dev/zero` and a FIFO now fail at once instead of
never returning.
Fixes#147
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
(cherry picked from commit 4156e8acd11c90c435a54a3b0ccc59991db71b12)
The old ids live only in crates/engine/src/legacy_ids.rs (the legacy migration table): find_command maps them, so old scripts, recipes and control calls keep working, and saved shortcuts, recent and pinned commands are migrated on load (designs store no command ids). Parity maps the observed catalog ids to ours through xtask/data/fusion-ids.tsv. The agent rule now reads: our own dotted id, never another product's. Cargo.toml's repository points at github.com/bherbruck/solvecraft.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/mcp (L5, engine only): JSON-RPC 2.0 over stdio with protocol version negotiation, a
headless Session backend and a bridge to the app's control port, 14 tools (list_commands,
execute, batch with rollback, inspect_design, measure, body_topology, set_parameter,
screenshot as an image block, export, undo/redo, new_design/open/save) and tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AS5Tn3xieewUSg3uxrMVdL
Kevin Ravensberg