Change proc.once('exit') to proc.once('close') so all stdout/stderr
data events are delivered before StringDecoder.end() is called. The
'exit' event fires when the process exits but before stdio streams
finish draining, which could cause the decoder flush to miss trailing
chunks and corrupt the last multi-byte character.
decoder.write() can hold incomplete multi-byte sequences internally.
Without calling decoder.end() after streams close, those trailing bytes
are silently dropped. Flush all decoders after process exit so any
remaining bytes are emitted as replacement characters rather than lost.
Address review feedback: a shared StringDecoder across stdout and stderr
can mix partial UTF-8 bytes between streams. Use one decoder per pipe so
incomplete sequences on stdout are never completed by stderr bytes.
Replace chunk.toString() with StringDecoder in stream processing to
prevent garbled characters (arrows, tree-drawing chars appearing as
mojibake) when multi-byte UTF-8 sequences are split across chunk
boundaries. StringDecoder maintains state across chunks to correctly
reassemble incomplete characters.
Fixed in bash tool, prompt shell execution, and ripgrep output.
Every child_process.spawn() call was missing windowsHide: true, causing
a visible CMD console window to briefly appear on every shell command,
LSP server start, ripgrep search, clipboard operation, and process kill
on Windows. This option is a no-op on non-Windows platforms.
Add --cloud-fork flag to run, attach, and thread commands that imports
a cloud session to local storage before continuing. Extract shared
validateCloudFork helper to reduce duplication. Use graceful exit
patterns in attach.ts and thread.ts to ensure proper cleanup.
ensureTitle was sending small-model requests with the same HEADER_TASKID
as the agent session. The gateway saw two different models (kilo-auto/small
and kilo-auto/frontier) under the same task and routed to both. Prefixing
the title sessionID with "title-" isolates it from the agent task.
- Reduce SSE heartbeat timeout from 90s to 15s to detect dead connections faster
- Add independent health poll every 10s as a second detection channel
- Kill entire process group (SIGTERM + SIGKILL fallback) instead of just the child process
- Add orphan detection in serve.ts so the server exits when the extension host dies
- Race Instance.disposeAll() against a 5s timeout to prevent hung MCP subprocesses blocking shutdown
BatchTool rebuilds its tool map from ToolRegistry.tools() without
passing the caller agent, so agent-level permission filtering is
bypassed. Add plan_exit to the DISALLOWED set to prevent non-plan
agents from triggering it via batch.
A merge conflict resolution incorrectly gated PlanExitTool behind
KILO_EXPERIMENTAL_PLAN_MODE && KILO_CLIENT === "cli", preventing the
plan agent from signaling completion. This broke the "looks good,
implement" flow — the follow-up prompt never appeared and no new
code-mode session was created.
Regression of #6240.
Add a "Model per Mode" section in Provider settings where users can set
a default model for each primary agent (code, plan, debug, orchestrator,
ask). The per-mode model is used automatically when starting new sessions
in that mode, with the precedence: session override > per-mode config >
global default model.
Also fixes the "unset" bug where clearing a model default didn't persist
because undefined values were stripped during JSON serialization. Now
uses null as a delete sentinel throughout the config pipeline.
The fixDuplicateReasoning workaround was added to handle a bug in
@openrouter/ai-sdk-provider v1 that caused reasoning details to be
duplicated in messages. This was fixed upstream in
https://github.com/OpenRouterTeam/ai-sdk-provider/pull/344/ and the
fix has since been applied to @kilocode/kilo-gateway, making this
workaround no longer necessary.
* feat: add colorblind-friendly theme
Add a new colorblind-accessible theme using a scientifically-designed
palette that works for deuteranopia, protanopia, and tritanopia:
- Primary Blue (#0077BB) for main UI elements
- Secondary Orange (#EE7733) for variables and operators
- Success Teal (#009988) instead of green for success states
- Warning Yellow (#CCBB44) for warnings
- Error Vermillion (#CC3311) instead of red for errors
- Accent Sky (#33BBEE) for functions and links
Key design principles:
- No red/green combinations (most common color blindness)
- Blue/orange as primary contrasting colors
- Teal/vermillion for diff colors instead of green/red
- High luminance contrast between adjacent colors
* Update packages/opencode/src/cli/cmd/tui/context/theme.tsx
Co-authored-by: Marius <marius@kilocode.ai>
* Update packages/opencode/src/cli/cmd/tui/context/theme.tsx
Co-authored-by: Marius <marius@kilocode.ai>
* Update packages/opencode/src/cli/cmd/tui/context/theme/colorblind.json
Co-authored-by: kilo-code-bot[bot] <240665456+kilo-code-bot[bot]@users.noreply.github.com>
* Update packages/opencode/src/cli/cmd/tui/context/theme/colorblind.json
Co-authored-by: kilo-code-bot[bot] <240665456+kilo-code-bot[bot]@users.noreply.github.com>
* fix malformed block
---------
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
Co-authored-by: Marius <marius@kilocode.ai>
Co-authored-by: kilo-code-bot[bot] <240665456+kilo-code-bot[bot]@users.noreply.github.com>
These markers are only needed in shared upstream opencode files to prevent
merge conflicts. In kilo-specific packages (kilocode/, kilo-gateway/,
kilo-ui/, kilo-i18n/, kilo-vscode/, etc.) they serve no purpose.
Kevin van Dijk