Replace the state.cancel-based mid-turn injection with a break-after-stream
queue. When a new prompt arrives while the assistant is streaming, the current
LLM step now finishes cleanly, any pending suggest/question tool auto-dismisses
via Suggestion.dismissAll and the new Question.dismissAll, and runLoop breaks
out before the next LLM step via KiloSessionPromptQueue.hasFollowup. Queued
prompts still run in order, each getting a full turn unless a newer one
arrives during it.
When the plan agent pauses for the follow-up question, CLI users had no
way to send a free-text reply: the main prompt input is hidden while a
blocking question is active, and the "Type your own answer" row was
being forced off to avoid duplicating the VS Code prompt input. CLI now
shows the custom-answer row again, while VS Code keeps it hidden.
The prompt queue's scope() hid the user message injected by
PlanFollowup.inject() because its ID was newer than the queue target,
so the loop saw the same plan_exit messages and re-asked the question
in an infinite cycle. Retarget the queue after inject so the new
message is visible on the next iteration.
Restore custom: false on the plan follow-up question — the "Type your own answer" row was redundant because the main prompt already routes typed text as a question reply. Auto-submit single-question single-select option picks in the VS Code QuestionDock so the button behaves like the TUI instead of silently waiting for a second Submit click.
Add optional `labelKey`/`descriptionKey` to Question.Option and `questionKey`/`headerKey` to Question.Info, annotated with `kilocode_change` markers. Populate these keys from the plan follow-up so the "Ready to implement?" question, "Start new session", and "Continue here" buttons render in the sidebar language while the canonical English labels remain on the reply wire (unchanged server-side matching).
* fix: persist custom provider model and variant deletions
The CLI config.update endpoint deep-merges its payload with existing
global config, so removing a model or reasoning variant from a custom
provider in the UI had no effect on disk. Save payloads now emit null
sentinels for removed IDs, and the Provider schema accepts nullable
record values so stripNulls can delete them during the merge.
Closes#9186
* chore: update kilo-vscode visual regression baselines
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Restore custom: false on the plan follow-up question — the "Type your own answer" row was redundant because the main prompt already routes typed text as a question reply. Auto-submit single-question single-select option picks in the VS Code QuestionDock so the button behaves like the TUI instead of silently waiting for a second Submit click.
Lock in that a session with an open suggestion reports as idle (so reopening VS Code or switching worktrees does not show it stuck/running), that accepting flips it back to busy without an idle flash, and that a dismissed suggestion leaves it idle for the run loop to resume cleanly.
Replace raw async/Promise-based test helpers with Effect-based
equivalents to align with the upstream BashTool API changes. Introduce
a ManagedRuntime composed of CrossSpawnSpawner, AppFileSystem, Plugin,
Truncate, and Agent layers. Update metadata, ask, and execute call
sites to return Effect values instead of plain promises.
Add experimental workspace adaptor list and status endpoints to the
OpenAPI spec. Add --sanitize flag to CLI export command for redacting
sensitive data. Update SDK types to reorder session status types before
suggestion types. Remove deprecated /global/sync-event endpoint.
Includes formatting fixes across i18n files, opencode source, and
tests, plus a warpgrep import path correction.
* fix(vscode): restore sidebar diff viewer parity with Agent Manager
The sidebar diff viewer bundle was missing agent-manager-review.css, so
FullScreenDiffView rendered with a broken file tree (no flex layout, no
revert styles) since PR #7455. Co-locate the CSS imports with the
component so every consumer gets them automatically, and wire per-file
revert end-to-end via a shared WorktreeDiffClient used by both the
sidebar provider and the Agent Manager controller.
* fix(cli): skip flaky shell-completion test on Windows CI
The 'shell completion resumes queued loop callers' test relies on shell
process spawn timing that is unreliable on Windows CI and times out at
3s. Every other shell-process test in this file already uses the
existing `unix(...)` helper for the same reason — align this test
with that pattern.
* feat(agent-manager): enable /sessions command to browse and resume session history
* fix(agent-manager): prevent duplicate tabs when selecting already-open session from history
* fix(agent-manager): preserve tab memory when reopening local history sessions
* docs(vscode): add agent manager sessions changeset
* fix(cli): eliminate mock.module race in commit-message tests
git-context.test.ts and generate.test.ts both called mock.module on
the same path. Bun mock.module is process-wide and permanent, so
whichever file loaded second saw the other's mock. Replace with
injectable test seams (setGitRunnerForTest, setGitContextForTest) that
are cleaned up in afterEach.
* test(cli): retry model json reads during async persistence
* fix(agent-manager): hide review tab when picking session from history
Also revert the commit-message test seam refactor — upstream rewrote
git-context.test.ts to use real git repos and the test runner isolates
each file in its own process, so the mock.module race no longer exists.
* docs(vscode): bump agent manager sessions changeset to minor
Gate the agent-change auto-apply effect on model.json load and skip it when a per-agent pick already exists, so user-selected models for agents with a configured default no longer revert on agent switches or restarts.
- Break circular import: Worktree → bootstrap → remote-sender → app-runtime (Not to Kilo not being completely on effect)
- Bump tight test timeouts for Effect-native processing overhead
When a user queued a prompt while the previous turn was still streaming,
the message's time_created fell before later assistant steps of that
turn. Ordering by time_created alone left the queued prompt in the
middle of the prior turn's history, so the next request ended with an
assistant message and tripped Anthropic's prefill rejection.
Reorder inside KiloSessionPromptQueue.scope so the target user message
and any of its own turn's assistants are always placed at the end.
* fix(cli): continue queued follow-up prompts
* test(cli): stabilize local model persistence test
* fix(cli): preserve queued prompt order
* fix(cli): persist queued prompts before processing
* chore(cli): clarify prompt queue settling
* fix(cli): preserve queued prompt history order
* fix(cli): clear queue state on release and test abort
Re-introduce the lightweight TypeScript diagnostic mode that was lost
during upstream OpenCode merges. Fixes the buggy server.ts spawn (leaked
process, wrong args) and re-adds TsClient integration in index.ts.
Extract prompt state logic (isPromptBlocked, isPromptBusy, isSuggesting,
isQuestioning) into pure functions in prompt-input-utils.ts and unit test
them. Add contract tests verifying sendMessage/sendCommand dismiss
suggestions and reject questions. Add CLI contract test protecting
prompt.ts Suggestion.dismissAll from upstream merges. Add dismissAll
unit tests to the suggestion test suite.
The .gitignore generated in .kilo/ config directories was missing
pnpm-lock.yaml and yarn.lock patterns, causing these lockfiles to
appear as untracked files in users' projects.
Alex Alecu