| Files | Last commit | Last update |
|---|---|---|
fix(fetch): preserve JSON number precision (#4131) | 1 day ago | |
perf(ansiext): grow string builder to optimize allocations | 1 year ago | |
refactor: replace swaggo annotations with a Go-native endpoint registry API routes and docs now share one source of truth: a fluent registry (internal/apigen) whose entries drive both mux registration and the docs served at /v1/docs/ (plain HTML page plus an OpenAPI 3.1 spec). Body schemas are derived from the real proto types via a small stdlib reflection walker, so renamed types break the build and stale annotation comments can no longer drift from the routes. Removes the generated internal/swagger artifacts, the task swag codegen step, and the swaggo dependencies. π Generated with Crush Assisted-by: Crush:kimi-k3 | 1 month ago | |
fix(permissions): harden the dangerous-command check Addresses PR review on sysadmin mode: - keep the block list on at exec time for commands the static check did not flag, so yolo's auto-approval still has a backstop; only a warned and approved command (or sysadmin mode) runs unguarded - flag a command whose name is built at runtime ($CMD, $(which curl)) - treat env, eval, nohup and xargs as dangerous: they run a name the static check cannot see - match Windows executable names case-insensitively (CURL.EXE) - tell the user what tripped the check instead of just 'potentially dangerous', and carry that reason over the client/server boundary, which never sent it at all - one constant for the duplicated sysadmin warning; name the permission mode instead of comparing against 0 | 7 days ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
feat(permissions): add sysadmin mode Yolo mode still refuses commands the block list flags as dangerous at run time. Sysadmin mode drops that last check too, for people who know what they are doing and are tired of being second-guessed. | 7 days ago | |
fix(clipboard): stop reporting successful copies as failures (#3952) The copy path wrote the clipboard twice: once through the terminal via OSC 52, once natively, and then read the clipboard back to confirm. The terminal applies its write whenever it gets around to it, so the check often read the clipboard in the middle of that write and called a good copy lost. Verify before handing the text to the terminal, and retry a few times, since any other program can land in the same window. | 1 month ago | |
ui: theme the exit banner (#4239) Fixes: CHARM-2645 | 1 day ago | |
refactor(commands): centralise skill discovery 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. | 5 months ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
fix(agent): close dispatch completion-boundary cancel race Run's deferred activeRequests.Del fired unconditionally on function return. When a finishing run handed off to a queued prompt (or a concurrent caller registered in the window between the explicit Del and the deferred Del), the stale cleanup wiped the new run's cancel function β breaking Cancel() and potentially allowing two concurrent turns on the same session. Wrap the cancel in an activeCancel struct with pointer identity, and replace the unconditional Del with CompareAndDelete so the deferred cleanup only fires when the stored entry is still ours. Also adds CompareAndDelete to csync.Map and a regression test for the completion-boundary transition. | 3 months ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
chore: make diff config independent | 1 year ago | |
feat: generally render output that looks like a diff as a diff (#2607) For example, the GitHub MCP server can fetch diffs. This makes those diffs render as such, rather than just showing their raw output. | 6 months ago | |
chore: bump catwalk Signed-off-by: drew <me@andrinoff.com> | 8 days ago | |
fix(dns): fix tmux dns resolver | 5 months ago | |
chore: return empty slices instead of nil for safety π Generated with Crush Assisted-by: GLM-4.7 via Crush <crush@charm.land> | 9 months ago | |
telemetry: differentiate invocation types (#3533) I suspect lots of `crush run` is just from people telling their interactive Crush to use `crush run` for write-capable subagents, since our built in subagents are currently hobbled. This checks for nesting on startup by looking for an env var the shell tool already sets for all shell invocations. If `crush run` sees the env var, it knows it's a non-interactive instance launched by another possibly interactive, possibly noninteractive instance. | 2 months ago | |
fix(glob): keep file search fast and bounded on large directories Pointing the glob tool at a huge or symlink-heavy directory (a home folder, a module cache) could peg every CPU core, balloon memory to several GB, and freeze the session until manually interrupted. It could also wander outside the project by following symlinks. Glob now applies a timeout, streams results instead of buffering the whole listing, scopes the walk to the pattern's literal prefix, and no longer follows symlinks. A search over a giant tree now returns quickly or times out cleanly instead of hanging. | 4 months ago | |
feat(lsp): start LSPs on demand, improve auto-start (#2103) * feat(lsp): start LSPs on demand, improve auto-start Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * refactor: simplify Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * chore: fmt Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * feat: handle new session/load session Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * refactor: manager holds client list reference Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: do io in cmd Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: better manager Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: nil Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: err Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: properly handle restart Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: add stopped state Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: root markers Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * fix: load for read files as well Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> * chore(go.mod): move indirect dependency to the right block --------- Signed-off-by: Carlos Alexandro Becker <caarlos0@users.noreply.github.com> Co-authored-by: Andrey Nering <andreynering@users.noreply.github.com> | 8 months ago | |
ui-perf: drive all spinners from one shared animation clock Every anim.Anim used to schedule its own 20Hz tea.Tick chain, so N running spinners meant N full Update+View passes per frame, each also resetting the frame cache and re-pinning follow mode. Replace the per-item chains with a single animTickMsg owned by Chat: one tick per frame advances every visible spinner, the clock stops itself when nothing visible spins, and Update's tail re-arms it whenever a spinner appears. Ticks carry a generation so a delayed tick can never run alongside a replacement clock, and frames that change nothing are served from the frame cache. Animatable becomes Spinning/Advance; pausedAnimations, RestartPausedVisibleAnimations and startItemAnimations are gone. Co-authored-by: Tai Groot <amittai.groot@gsa.gov> | 1 month ago | |
chore: gofmt | 3 months ago | |
test: remove signing off of the branch test (#3999) Signed-off-by: drew <me@andrinoff.com> | 1 month ago | |
chore: gofmt | 3 months ago | |
chore(tools/edit): various fixes (#1921) | 9 months ago | |
fix(config): keep tests from rewriting the real global config Loading config is not read-only: it can save migrations, refreshed tokens and removed credentials. Tests that loaded it without redirecting the global paths wrote to the developer's own files, and one run on a checkout that drops a Claude subscription on load signed every open Crush window out. Under go test the home directory is now a scratch dir and the inherited variables that override config, data and cache locations are cleared, so every path built from them stays out of the real home. The ad-hoc isolation in the config tests is no longer needed. | 9 days ago | |
refactor(hooks): run every event through one Event and one Runner Replace the per-event entry points, payload builders and env builders with a single hooks.Event that callers fill in; only the fields that are set appear in the stdin payload. The runner reads hook config from a source closure on every event, so a config reload is picked up with no extra wiring, and a nil runner runs nothing so callers never guard it. Payload field names follow Claude Code where an equivalent exists (hook_event_name, tool_use_id, tool_response, last_assistant_message, source, reason, trigger) and CLAUDE_PROJECT_DIR is exported next to CRUSH_PROJECT_DIR. Output accepts Claude Code's decision:block, continue:false and updatedToolOutput forms, and a JSON object on stdout counts even when the exit code is a non-blocking error. tmp-resolve | 11 days ago | |
refactor(lock): add canonical internal/lock package, migrate db and cmd callers Introduce internal/lock as the single file-locking helper for Crush, replacing the duplicated flock implementations in internal/db and internal/cmd. - lock.File(ctx, path) blocks until ctx cancels or lock acquired (context-aware deadline replaces hardcoded 5s timeout) - lock.TryFile(path) returns ErrContended immediately if held - Migrate internal/db.acquireDataDirLock to lock.TryFile - Migrate internal/cmd.spawnAndWaitReady to lock.File(cmd.Context(), ...) (now respects Ctrl+C during spawn wait) - Delete internal/db/datadirlock_{unix,windows}.go - Delete internal/cmd/spawnlock_{other,windows}.go - Add tests: acquire/release, contention, deadline, cancellation, concurrent mutual exclusion Config store uses lock.File with a 5s context deadline, preserving existing behavior while gaining context cancellation support. | 5 months ago | |
fix: never leave a tool call waiting for an answer that cannot come π Generated with Crush Assisted-by: Crush:claude-opus-5-5 | 21 days ago | |
fix: keep browser output off the terminal (#4235) Co-authored-by: Drew Smirnoff <drew@charm.land> Signed-off-by: Drew Smirnoff <drew@charm.land> | 1 day ago | |
oauth-cli: adjust inline logout TUIs | 1 month ago | |
fix: stop intermittent CI test failures across all three platforms (#4034) | 1 month ago | |
fix: replay data correctly for anthropic (#4237) Signed-off-by: Drew Smirnoff <drew@charm.land> | 1 day ago | |
chore: bump catwalk Signed-off-by: drew <me@andrinoff.com> | 8 days ago | |
fix(shell): say what is dangerous in plain words The warning shown before a risky command read "it uses sudo", and the error when one was stopped read "command is not allowed for security reasons" followed by two sentences of advice. The advice told the reader to run the command directly rather than from inside a script, which was untrue whenever they had typed it directly, and then suggested a flag that switches the checks off. Both now name the thing and stop: "sudo", and "blocked: sudo is a dangerous command". Where no command can be named, the warning names the shell construct that hid it instead, since that is the useful fact: expansion in command name, substitution in arguments, process substitution in redirect position. Recognising a command through a wrapper is a fact about shells, not a policy either list owns, so it now lives in one place beside the matching and both lists consult it. The warning names the command recognised rather than the wrapper in front of it, so `nice curl` reads "it uses curl" instead of asking without saying why. A command the shell cannot parse gets no warning at all. It never runs, so the old "it could not be parsed" dressed a typo up as a threat while the real parse error said where the quote was left open. The warning wraps rather than clipping, because the longest of these phrases was being cut off in the middle. | 7 days ago | |
fix(password-prompt): address visual review (#4227) Signed-off-by: Drew Smirnoff <drew@charm.land> | 2 days ago | |
fix: stop intermittent CI test failures across all three platforms (#4034) | 1 month ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
feat(permissions): add sysadmin mode Yolo mode still refuses commands the block list flags as dangerous at run time. Sysadmin mode drops that last check too, for people who know what they are doing and are tired of being second-guessed. | 7 days ago | |
fix(ui): render question answers from structured data, not parsed text (#3903) | 1 day ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
feat(config): let a project configure the commands it runs all day The dangerous-command list could be added to and subtracted from, but the list of commands quiet enough to run without asking could not be touched. So a project where `go build` runs fifty times a day had two options: answer the prompt every time, or hand over the shell entirely by allowing the whole bash tool. permissions.safe_commands takes exact command lines, and covers only what it says: "go build" means that and not `go build -o` anywhere you like. The built-in entries can afford a per-command flag policy because each one was checked by hand; a setting that widened itself to flags nobody reviewed would grant far more than its reader intended. Configured entries compose with the built-in ones, so `go build && cargo check` is quiet if both halves are. Command names in configuration are also now read the way the lists match them, so an entry written CURL, /usr/bin/curl or with stray spaces lines up with the curl already on the list. Allowing a command by a slightly different spelling used to do nothing whatsoever, which is the worst way for a permission setting to be wrong: it looks applied. All three lists are set from crushrc, the format crush.json is giving way to, through the permissions builtin: permissions safe "go build" "cargo check" permissions block kubectl terraform permissions unblock curl | 7 days ago | |
feat(config): let a project configure the commands it runs all day The dangerous-command list could be added to and subtracted from, but the list of commands quiet enough to run without asking could not be touched. So a project where `go build` runs fifty times a day had two options: answer the prompt every time, or hand over the shell entirely by allowing the whole bash tool. permissions.safe_commands takes exact command lines, and covers only what it says: "go build" means that and not `go build -o` anywhere you like. The built-in entries can afford a per-command flag policy because each one was checked by hand; a setting that widened itself to flags nobody reviewed would grant far more than its reader intended. Configured entries compose with the built-in ones, so `go build && cargo check` is quiet if both halves are. Command names in configuration are also now read the way the lists match them, so an entry written CURL, /usr/bin/curl or with stray spaces lines up with the curl already on the list. Allowing a command by a slightly different spelling used to do nothing whatsoever, which is the worst way for a permission setting to be wrong: it looks applied. All three lists are set from crushrc, the format crush.json is giving way to, through the permissions builtin: permissions safe "go build" "cargo check" permissions block kubectl terraform permissions unblock curl | 7 days ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago | |
fix(password-prompt): address visual review (#4227) Signed-off-by: Drew Smirnoff <drew@charm.land> | 2 days ago | |
fix: preserve first line indentation in tool call previews (#3652) Assisted-by: Crush:kimi-k3 | 2 months ago | |
ui: theme the exit banner (#4239) Fixes: CHARM-2645 | 1 day ago | |
chore: fix more typos (#1863) | 9 months ago | |
fix: detect stale server during development with BuildID The version check compared Version strings, but during development both client and server report "devel" β so stale servers were never restarted. BuildID is now derived from the executable's modification time, which changes on every recompilation (including go run). The server exposes it via /v1/version and the client checks both Version and BuildID match. Assisted-by: Crush:AWS Claude Opus 4.6 | 5 months ago | |
chore: call them subagents everywhere a person or model reads it | 2 days ago |
| Files | Last commit | Last update |
|---|---|---|
| 1 day ago | ||
| 1 year ago | ||
| 1 month ago | ||
| 7 days ago | ||
| 2 days ago | ||
| 7 days ago | ||
| 1 month ago | ||
| 1 day ago | ||
| 5 months ago | ||
| 2 days ago | ||
| 3 months ago | ||
| 2 days ago | ||
| 1 year ago | ||
| 6 months ago | ||
| 8 days ago | ||
| 5 months ago | ||
| 9 months ago | ||
| 2 months ago | ||
| 4 months ago | ||
| 8 months ago | ||
| 1 month ago | ||
| 3 months ago | ||
| 1 month ago | ||
| 3 months ago | ||
| 9 months ago | ||
| 9 days ago | ||
| 11 days ago | ||
| 5 months ago | ||
| 21 days ago | ||
| 1 day ago | ||
| 1 month ago | ||
| 1 month ago | ||
| 1 day ago | ||
| 8 days ago | ||
| 7 days ago | ||
| 2 days ago | ||
| 1 month ago | ||
| 2 days ago | ||
| 7 days ago | ||
| 1 day ago | ||
| 2 days ago | ||
| 2 days ago | ||
| 7 days ago | ||
| 7 days ago | ||
| 2 days ago | ||
| 2 days ago | ||
| 2 months ago | ||
| 1 day ago | ||
| 9 months ago | ||
| 5 months ago | ||
| 2 days ago |