Co-authored by: Kieran Klukas <kieran@dunkirk.sh>
Co-authored by: Tai Groot <tai@taigrr.com>
Co-authored by: Andrey Nering <andreynering@users.noreply.github.com>

Includes fix for nil pointer crash when writing to a closed WrapWriter
and glamour writer chain close ordering fix.

Load ~/.config/crush/CRUSH.md and ~/.config/AGENTS.md by default as
global context files, separate from project context. Global context is
presented to the model as user preferences that apply across all
projects. Paths are configurable via global_context_paths.

Some models like Qwen take this intentation too literally and add them
to the commit message. The new system prompt fixes that.

Co-Authored-By: Charm Crush <crush@charm.land>

Adds tests confirming that a leftover socket is detected and cleared
while a live socket is left untouched, and that the socket is placed
in the expected per-user location.

Co-Authored-By: Charm Crush <crush@charm.land>

The client now checks whether a leftover socket has a live server
behind it. If it does not, the client clears the dead socket and
starts a fresh server instead of getting stuck trying to reach one
that is no longer there. The detection is shared across platforms.

Co-Authored-By: Charm Crush <crush@charm.land>

When the server starts and finds a socket file with no process behind
it, it now removes the dead socket before binding instead of failing
with an address-already-in-use error. A live server is never disturbed,
and the cleanup is recorded in the log.

Co-Authored-By: Charm Crush <crush@charm.land>

Sets up the persistent log file before the client checks the socket,
so any message about detecting or clearing a leftover socket is saved
to disk instead of being lost.

Co-Authored-By: Charm Crush <crush@charm.land>

Moves the Crush socket out of the shared temp folder and into the
per-user runtime directory on Linux and macOS, falling back to the
temp folder when needed. Guards against socket paths that are too
long to bind.

Co-Authored-By: Charm Crush <crush@charm.land>

Adds a helper that recognizes when a unix socket file is left on disk
but no process is listening on it, so later code can clean it up safely.

Co-Authored-By: Charm Crush <crush@charm.land>

💘 Generated with Crush

Assisted-by: Crush:qwen3.7-max

fix(server): avoid closing other clients' connections during cancellation

Publish a terminal completion for submitted background prompts that fail
before normal agent completion can run. The error still reaches
observers, and duplicate completion events remain suppressed.

Co-Authored-By: Charm Crush <crush@charm.land>

Ensure queued prompts with their own submitted identity emit one
terminal completion when canceled or dropped. Callers no longer wait on
a completion that was folded into another prompt's turn.

Co-Authored-By: Charm Crush <crush@charm.land>

Check queued prompt cancellation while holding the same session handoff
lock used for accepted and active prompts. Queue draining now follows
the same ordering guarantees as prompt dispatch.

Co-Authored-By: Charm Crush <crush@charm.land>

Exercise the real server-to-agent path for a cancel that lands after a
prompt is accepted but before it becomes active. This protects the race
fix from passing only in mocked unit paths.

Co-Authored-By: Charm Crush <crush@charm.land>

Carry the submitted prompt identity through error events so command-line
runs fail only for their own prompt. Older session-scoped errors remain
supported for compatibility.

Co-Authored-By: Charm Crush <crush@charm.land>

Apply a cancel only to prompts that were already accepted when the
cancel request arrived. Immediately canceled accepted prompts also
publish completion so callers waiting on that prompt do not hang.

Co-Authored-By: Charm Crush <crush@charm.land>

Lock in that prompt cancellation is handled by background completion
events rather than a synchronous HTTP error. This prevents canceled
prompts from being reported as server failures.

Co-Authored-By: Charm Crush <crush@charm.land>

Add end-to-end coverage for cancellation from another client, request
disconnects, accepted-prompt races, and workspace handoff under the new
asynchronous prompt contract.

Co-Authored-By: Charm Crush <crush@charm.land>

Treat HTTP 202 as a successful prompt submission and preserve
server-provided error messages for rejected prompts. Callers no longer
see generic status errors when the server explains the failure.

Co-Authored-By: Charm Crush <crush@charm.land>

Return as soon as a prompt is accepted instead of keeping the request
open for the full agent turn. Workspace shutdown is reported as a
conflict so clients can distinguish rejection from accepted work.

Co-Authored-By: Charm Crush <crush@charm.land>

Accept validated prompts quickly, then run them on workspace-owned
background work. Failures are published through events while normal
cancellation continues to be represented by the canceled assistant turn.

Co-Authored-By: Charm Crush <crush@charm.land>

Add a prompt failure event that can be delivered after the HTTP request
has already returned. Remote clients and subscribers now receive the
same failure message through the event stream.

Co-Authored-By: Charm Crush <crush@charm.land>

Track prompts that have been accepted but have not started running yet
so a cancel issued right after acceptance applies to that prompt.
Idle-session cancels remain a no-op, preventing one client's cancel
from poisoning the next prompt.

Co-Authored-By: Charm Crush <crush@charm.land>

Use the same prompt and session checks before accepting background work
that synchronous runs already use. This keeps rejected prompts
consistent as the server begins accepting requests asynchronously.

Co-Authored-By: Charm Crush <crush@charm.land>

Give each workspace its own run lifetime and shutdown gate so agent
work is not tied to the HTTP request that submitted it. Final state can
still be saved during shutdown, avoiding lost completion or error
records when a workspace is canceled.

Co-Authored-By: Charm Crush <crush@charm.land>

The UI for invoking skills duplicated skill discovery, but missed some
of the nuances like following symlinks. This removes some of that
duplicative code and wires other bits together so the UI shares the same
discovery logic as the backend, fixing at least the symlink behaviour.