kilocode/packages/opencode/CHANGELOG.md

@kilocode/cli

7.4.21

Minor Changes

• #12825 b692f1d - Add kilocode command-file endpoints so clients can list editable command/workflow files, inspect model and reasoning variant metadata, and remove them.

• #12991 0e1f11b - Add nested slash command suggestions for /review in VS Code and support staged, unpushed, and quick review modes.

• #12824 e87dc77 - Import conversation history from Claude Code and OpenAI Codex sessions with the /resume-claude and /resume-codex slash commands.

Patch Changes

• #12941 bddce1a - Automatically expose broad reasoning effort options for custom provider models and link saved providers to advanced JSON configuration.

• #12865 35801cb - Prevent VS Code sessions and Agent Manager worktrees from starting unused file watchers and defer file indexing until search is used.

• #12926 90ac91d - Prevent built-in skill documentation examples from triggering shell permission prompts.

• #13007 910f0f2 - Disable the suggest tool and auto-dismiss pending suggestions in non-interactive CLI runs to prevent hanging on benchmarks and automated pipelines.

• #13044 2e5199a - Exclude ChatGPT subscriptions from explicit prompt cache breakpoints.

• #12935 3917ed1 - Restore keyboard input for interactive terminal prompts when the CLI session uses a workspace.

• #12554 dcf9c5a Thanks @arimu1! - Prevent project MCP configs from resolving variable-backed headers or inheriting trusted headers when changing endpoints, while preserving unaffected servers.

• #12958 fdc4665 - Reduce noisy memory timeout warnings and retry transient background consolidation failures once.

• #12993 d3c50e6 - Route Agent Manager tool-launched sessions to the project that owns the tool event directory, keep sandboxed worktree sessions inside their active worktree, and wait for busy managed sessions before prompting them.

• #12937 4ea52f2 - Allow subagent tasks to be resumed after their parent session is forked.

• #12946 24da90f - Support the Agent Manager tool with llama.cpp servers that reject prefix-only JSON Schema patterns.

• #12882 9bfbc35 - Add bounded, context-aware grep controls without leaving agents waiting on completed searches.

• #13040 48c4a4a - Ignore negative pricing entries from model catalogs and handle unpriced models gracefully in UI price formatting.

• #13022 9dbf276 - Set explicit prompt cache breakpoints on stable prefixes for OpenAI GPT-5.6+ models.

• #12695 a606a91 - Changes from opencode v1.17.9 to v1.17.13 upstream:

  • Core Improvements: MCP servers can append their instructions to the model context, and MCP resources are available as tools with template listing.
  • Core Improvements: Model variants are generated from models.dev data, including modes exposed as models.
  • Core Improvements: Tool definitions pass strict through for Codex parity, and Gemini requests support video and audio media.
  • Core Bugfixes: Interrupted assistant steps settle instead of leaving sessions stuck busy.
  • Core Bugfixes: MCP OAuth reconnects after authorization even when the server is disabled, refreshes credentials on reauthentication, requests refresh token scope, surfaces completion errors, and binds its callback to the IPv4 loopback.
  • Core Bugfixes: MCP tool results prefer content over structured output, and denied resource template tools stay hidden.
  • Core Bugfixes: Stale GitHub Copilot Responses item IDs are no longer replayed, and OpenAI reasoning variants are forced where required.
  • Core Bugfixes: Adaptive thinking is enabled for Claude Sonnet 5, and expired promos were removed from the zen catalog.
  • Core Bugfixes: Preserve released prompt history during database replay and keep native event streams connected for all supported Kilo events.
  • Core Bugfixes: Remote skill manifests support optional per-skill versions; changing a version refreshes the cached skill atomically, and skill base directories are emitted as filesystem paths.
  • CLI Improvements: Ports increment from the default when busy.
  • CLI Improvements: Use --auto to start the TUI in a run-scoped auto-approve mode, and leave the mode mid-session from the command palette.
  • TUI Improvements: Redesigned crash screen, model picker sorted by release date, bindable diff viewer and Move Session commands, main-branch diff source, and inline skill load errors.
  • TUI Bugfixes: File autocomplete is scoped to the session, multi-day durations format correctly, and root sessions load in the session switcher.

• #12442 6b8c736 Thanks @IamCoder18! - Add a privacy mode that blurs PII in the TUI (personal balance, Kilo Pass usage, etc.) and requires confirmation before /profile reveals email, name, balance, and team. Toggle with the new /privacy command or by setting privacy_mode in kilo.json. The kilo profile CLI command is unaffected.

• #12897 e83b25e - Fix high CPU and runaway memory growth in the JetBrains background kilo serve process on macOS by no longer eagerly starting native file watchers, matching the VS Code backend.

• #12884 c9199cb - Show a concise retryable message when concurrent Kilo processes temporarily lock the SQLite database instead of printing the full server error trace.

• #12929 16deb19 - Prevent TUI config reload logs from corrupting the interactive terminal.

• #12950 d03d579 - Avoid printing an error when closing the TUI cancels in-flight startup refreshes.

• #12978 1406d71 - Remove unsupported kilo web CLI command.

• #12947 154b1ae - Prevent concurrent Kilo startups from rewriting unchanged credentials, retry transient database locks, and redact bound values from database errors.

• #12600 4e36297 - Apply saved sandbox settings to existing sessions and use the latest settings when enabling sandboxing

• #13047 e48c534 - Separate autocomplete item names from their descriptions in the TUI.

• Updated dependencies [fdc4665, 48c4a4a]:

  • @kilocode/kilo-memory@7.4.21
  • @kilocode/kilo-gateway@7.4.21
  • @kilocode/kilo-indexing@7.4.21
  • @kilocode/kilo-telemetry@7.4.21
  • @opencode-ai/server@7.4.21
  • @opencode-ai/tui@7.4.21
  • @opencode-ai/ui@7.4.21

7.4.20

Patch Changes

• #12847 9ed716a - Temporarily restore the default grep controls to prevent searches from stalling subagents.

• #12659 39f65ef Thanks @mjnaderi! - Skip API and telemetry lifecycle work for informational CLI commands and avoid profile requests when telemetry is disabled.

• #11961 fd60036 Thanks @mvanhorn! - Help models recover from invalid tool calls with clear, field-specific validation errors.

• #12846 8c84f8a - Support project agent and command directory symlinks explicitly allowed by global Markdown source permissions.

• Updated dependencies [39f65ef]:

  • @kilocode/kilo-telemetry@7.4.20

7.4.19

Minor Changes

• #12728 6b27a26 Thanks @bagatao-anaconda! - Show why a tool call was auto-approved or denied in the TUI, and record the denial reason on the tool call metadata (visible in kilo export) alongside the existing auto-approval reason.

• #12816 63a38f5 - A second identical large paste expands its collapsed prompt placeholder.

• #12729 ce7984f - Configure a model and reasoning variant for each workflow from Agent Behaviour settings.

Patch Changes

• #12792 7d3f50c - Prevent configured compaction thresholds from interrupting active tool sequences.

• #12652 c554409 Thanks @Hardik180704! - Keep config-defined subagents routable when an installed primary agent uses the same name.

• #12726 2fbd380 - Speed up local session recall searches across large conversation histories.

• #12811 989f7f0 - Add bounded, context-aware signal-to-noise controls to grep searches.

• #12790 8be3303 - Keep Kilo's persona out of generated conversation titles and Agent Manager branch names.

• #12802 5bd420a Thanks @bagatao-anaconda! - Stop treating !`cmd` shown as an inline code example in skill documentation as a live command, so it no longer triggers a shell permission prompt.

7.4.18

Minor Changes

• #12704 ace509d - Remote CLI session lifecycle: create_session accepts optional agent, model, and orgId (org claim rides session metadata); CLI adopts backend renames via system session.renamed and POSTs local title changes (generation-aware) to the ingest title route so auto-titles and explicit renames stay in sync.

• #12604 2d784a0 Thanks @bagatao-anaconda! - Support executing shell commands embedded in skill files. Commands written as !`command` in a SKILL.md run and their output is inlined into the skill. Only trusted skills can run commands and KILO_DISABLE_SKILL_SHELL disables the behavior; when the model loads a skill, the commands are shown in a single up-front approval before running.

Patch Changes

• #12682 ed9e132 - Reduce CLI startup time by deferring Kilo-specific module loading until commands actually run, caching the telemetry profile lookup across invocations, and uploading telemetry in the background so process exit is not delayed by a network round trip

• #12687 624b589 - Fix settings snapping back to their previous value after being cleared to "Not set" when multiple config files exist (e.g. both kilo.json and kilo.jsonc)

• #12701 89caab9 - Mark Kilo Console as deprecated and direct users to supported session and settings workflows.

• #12369 44f1373 - Allow users to enable web search for models from all providers through Kilo configuration, VS Code settings, and Kilo Console settings.

• #12647 5ad8b2f - Keep RC installations up to date when a newer stable CLI release is published.

• #12698 1d1630b - Report Bash commands terminated by a signal with the conventional 128 + signum exit code (e.g. 139 for SIGSEGV) instead of hanging until the command timeout.

• #12605 d579774 - Non-interactive kilo run no longer reports success for runs that did not complete. A plain headless run (neither --auto nor --dangerously-skip-permissions) in which the CLI auto-rejected at least one permission ask now exits 1 with a stderr diagnostic naming the cause, and a run whose session errors mid-stream now prints that diagnostic to stderr under --format json as well (previously the JSON branch swallowed it). Runs that complete their turn with no auto-rejected permission still exit 0. Under --format json the auto-reject path adds a new error event to the stream; existing event shapes are unchanged. The same exit-1 rule applies to a plain non-interactive --attach run that auto-rejects an ask (that run was equally crippled); interactive mode is untouched.

• #12727 1a34037 - Allow clearing a nested project setting when no project config file exists yet.

• #12700 f695425 - Show the provider's actual error message when an OpenAI or Azure Responses API stream fails (for example an upstream rate limit) instead of a generic retry notice.

• #12684 d7f8da9 - Include the underlying reason in search execution failures instead of showing a bare "ripgrep execution failed" message.

• #12723 a1ad65e - Retry transient locked-file errors (EPERM/EACCES/EBUSY) on Windows when atomically saving config and other files. Background plugin installs and Windows Defender/indexer can briefly hold the temp file during the rename step, which previously surfaced as a 500 error. A short backoff now retries the rename so config writes succeed without surfacing the contention.

• Updated dependencies [ed9e132, 0d923d0, 304c75e]:

  • @kilocode/kilo-telemetry@7.4.18
  • @kilocode/kilo-memory@7.4.18

7.4.17

Patch Changes

• #12544 b8d83fb - Support adaptive thinking levels for Claude Opus and Sonnet 5 and later.

• #12587 16f8e7e - Keep conversations and workspace files unchanged when a checkpoint cannot be fully restored.

• #12444 92076e7 Thanks @IamCoder18! - Add a /auto-approve slash command in the TUI for toggling auto-approve mode, with aliases /autoapprove, /approve-all, and /approveall. The command dispatches the existing palette entry, so behavior matches the Ctrl+P "Enable/Disable auto-approve mode" toggle.

• #11986 0abe474 Thanks @IamCoder18! - Let the Context section in the TUI session sidebar collapse and expand on header click, matching the existing collapsible pattern used by Token Usage, Models, and Terminal Bench 2.0. When collapsed, the header shows a one-line summary of percent used and total cost.

• #12601 dab2e79 - Exclude GPT-5.6 from models available through ChatGPT subscriptions while retaining access to variants such as GPT-5.6 Sol.

• #12592 8c88048 Thanks @noobezlol! - Keep Nix builds on the Bun version required by the repository.

• #12545 b2735bf - Fix session transcripts losing their final messages when the CLI exits — pending uploads are now flushed on shutdown and as soon as a session closes.

• #12470 c0ebf98 Thanks @IamCoder18! - Route the websearch tool's Exa requests through the Kilo proxy when signed into Kilo. The MCP-Exa transport is preserved as a fallback for users who set EXA_API_KEY or are not authenticated. A new KILO_WEBSEARCH_PROVIDER=kilo-exa env override forces the Kilo proxy path. Results are capped at 10.

• #12460 51d8031 - Changes from opencode v1.17.5 to v1.17.9 upstream:

  • Core Bugfixes: Improved MCP server compatibility by declaring Kilo's supported client capabilities.
  • Core Bugfixes: Plugin client requests now reuse the active server instead of assuming the default local port.
  • Core Bugfixes: ACP shell tool calls now show the command and working directory from the start.
  • Core Bugfixes: Plugin-provided shell environment variables now apply to PTY sessions.
  • Core Bugfixes: OpenAI-compatible providers now accept MCP tool schemas that previously failed validation. (@jquense)
  • Core Bugfixes: Cloudflare AI Gateway now receives the configured API key correctly. (@keefetang)
  • Core Bugfixes: MCP tools without declared schema properties now work with providers that expect object properties.
  • Core Bugfixes: Long-running MCP tools now keep their timeout alive when they report progress. (@Nomadcxx)
  • Core Bugfixes: The MCP OAuth callback server now shuts down once authorization finishes or is cancelled.
  • Core Bugfixes: MCP tool failures now surface the server's error text instead of a generic failure.
  • Core Bugfixes: MCP OAuth error pages now escape provider error text correctly.
  • Core Bugfixes: Honor configured agent step limits by forcing a final text response instead of failing mid-run.
  • Core Bugfixes: Queue steering prompts before dismissing pending questions so the previous turn cannot resume first.
  • Core Bugfixes: Prevent local server credentials from leaking into spawned PTY processes.
  • Core Bugfixes: Fix Devstral model detection when provider IDs use different casing. (@Robin1987China)
  • Core Bugfixes: Pass configured custom headers to Copilot model requests.
  • Core Improvements: MCP servers can now receive the current workspace as a client root.
  • Core Improvements: Session timelines load much faster and avoid flicker or scroll jumps.
  • Core Improvements: Add high and max thinking variants for GLM-5.2 across supported providers. (@imranshaiedi-byte)
  • Core Improvements: Stop wrapping follow-up user messages in a steering reminder so prompt caching stays effective.
  • TUI Bugfixes: MCP debug now uses the SDK's latest protocol version.
  • TUI Bugfixes: Only show the background subagent shortcut when the server supports it.
  • UI Bugfixes: Render completed Mermaid blocks from diagram source instead of fenced Markdown.

• #12585 a0a760e - Fix bash permission rules being bypassed on PowerShell for commands containing a bare -- such as git checkout -- <file>. Commands the shell parser cannot parse now get checked against their raw command text instead of executing without a permission check.

• #12505 bcf8b8b - Emit each agent event once from kilo run --format json.

• #12593 160b066 - Prevent the VS Code backend from eagerly starting native file watchers for every Agent Manager worktree.

• #12583 1310c12 Thanks @noobezlol! - Keep Windows snapshot diffs parseable and preserve valid files when a stored patch is malformed.

• #12588 deddf00 - Bound the wait for a provider's first response byte by the request timeout. A provider that accepts a request and returns headers but never sends body data now fails and retries instead of leaving the turn hanging after a tool call completes. The same timeout value now covers both the connection phase and the wait for the first byte as a single deadline; streaming responses that have already produced data are unaffected.

• #12514 a33493e - Stabilize cross-platform CLI subprocess tests under constrained CI runners

• #12639 8a47d8b - Stop flashing a "Turn interrupted" warning when a follow-up message is queued while the assistant is still working. The running turn now closes with a dedicated "superseded" reason instead of "interrupted" when it hands off to the queued prompt, so the premature-stop warning only appears for real interruptions.

7.4.16

Minor Changes

• #12392 16988a5 - Add a notify_user tool that lets an agent send a push notification to the user's phone (Kilo mobile app) for explicitly requested pings and significant mid-run milestones. The tool sends a single agent_notification item over the session's existing authenticated ingest channel with a bounded readiness wait, returns a friendly failure when the session is not connected to Kilo cloud, and never prompts for permission. Delivery may still be suppressed server-side by the user's notification preference, per-session rate limits, or active presence in the session.

• #12370 b367105 Thanks @hdcodedev! - Support deleting queued chat messages from the VS Code chat before they run.

• #12327 aa22680 - kilo remote instances now advertise themselves on the relay heartbeat. Each heartbeat carries the host's hostname, the project directory name, and the CLI build version, and each session entry advertises the platform it was created on. The cloud relay learns about a freshly-connected instance immediately (no 10s wait for the first timer tick), and the advertisement is race-safe across the explicit kilo remote command and bootstrap auto-enable (KILO_REMOTE=1 / remote_control config). Legacy CLIs that send neither field remain wire-compatible.

• #12394 e72238a - Support file attachments in remote CLI sessions.

• #11849 fe01f53 - Run asynchronous Cloud Agent tasks with repository, model, mode, and organization defaults through kilo cloud. Add --stream to kilo cloud start to print admission output and then stream WebSocket events as JSONL until completion or inactivity ends the stream.

• #12456 3d648d7 - Reference past chats inline with @ in the prompt. Typing @ now surfaces a "Past chats" option that opens a searchable picker of previous sessions (scoped to the current workspace/worktree, searched like the Agent Manager session search); selecting one attaches that session's transcript as context so the model can build on a prior conversation. Clicking the mention opens that session. Available in the CLI TUI and the VS Code extension.

• #12434 dcc0d64 Thanks @Githubguy132010! - Show tokens-per-second text-generation throughput (TG) on each assistant message and in the usage sidebar, computed from step duration and tokens. The toggle "Show Token Throughput" in Display settings controls both surfaces. PP (prompt-processing) support lands in a follow-up once the upstream llama.cpp metadata wiring ships.

Patch Changes

• #12475 c72817e Thanks @LCZcn96! - Load global skills reliably from projects outside Git repositories.

• #12497 23963e3 - Restore stream idle timeouts to opt-in provider configuration instead of aborting quiet model streams by default.

• #12393 9262f2b - Remote CLI sessions no longer appear frozen on mobile when the connection to the session relay stalls; they now recover on their own instead of staying read-only until the CLI is restarted. Token acquisition and connection attempts are bounded by deadlines with a single fenced retry owner, and heartbeat session gathers are bounded so one stuck gather can no longer silently kill every future heartbeat.

• #12485 079fd04 Thanks @rakshith1928! - Fix compaction failure against strict OpenAI-compatible providers during context compaction. The compaction path no longer leaks maxOutputTokens into provider options, which was rejected by strict upstreams with "Unsupported parameter(s)".

• #11940 0d830cb Thanks @rakshith1928! - Fix: inject $schema into config files using jsonc-parser, avoiding write-on-read for comment-first JSONC and preventing unnecessary file rewrites on every load

• #12508 0fe46ec - Fix a fatal startup crash ("attempt to write a readonly database") when the local database or its WAL sidecar files lost write permission. Kilo now repairs the permissions automatically when it safely can, and otherwise reports the exact file to fix instead of an opaque error.

• #12458 182d18b - Keep Plan and Architect mode source edits denied when agent-specific permissions request edit approval.

• #12496 2fcb137 - Preserve unexpected provider finish reasons and show the request and Gateway generation IDs when a response ends unexpectedly.

• #12488 c25f041 - Show the request ID when a model response ends without a finish reason.

• Updated dependencies [2fcb137, f715e2f, dcc0d64]:

  • @kilocode/sdk@7.5.0
  • @kilocode/kilo-gateway@7.4.16
  • @kilocode/plugin@7.4.16
  • @opencode-ai/tui@7.4.16
  • @opencode-ai/ui@7.4.16
  • @kilocode/kilo-indexing@7.4.16
  • @kilocode/kilo-telemetry@7.4.16
  • @kilocode/plugin-atomic-chat@7.4.16
  • @opencode-ai/server@7.4.16

7.4.15

Patch Changes

• #12249 cd205d8 - Prevent agent and subagent sessions from stalling indefinitely with a 60-second model-stream idle watchdog that pauses while local tools run. Set chunkTimeout: false to disable it.

• #12427 2dc1a52 - Improve xAI prompt cache hit rate by sending promptCacheKey by default for xAI models

• #12422 28d015f - Simplify project memory settings and activity visibility, replace direct editing with folder inspection, add nested memory slash-command completion and status views, improve empty-project handling, compact native tool-call summaries, and remove legacy memory audit logs.

• #12414 badf70d - Keep Linux sandbox setup working when a writable directory contains an unreadable subdirectory (for example a folder with mode 600); unreadable subdirectories are now protected with a read-only mount instead of failing every sandboxed tool call with an access error.

• Updated dependencies [28d015f]:

  • @kilocode/kilo-memory@7.4.14

7.4.13

Minor Changes

• #12271 38013f7 - Allow agents to stop and remove a targeted Agent Manager session.

• #12306 c081f58 - Configure a custom file extension allowlist for codebase indexing to limit scans to relevant project files and support additional text formats.

• #12224 df9e1fb - Remote CLI: expose slash command discovery, execution, and /new session creation over the relay. list_commands and send_command (including the built-in compact flow) are scoped to the current session's directory. create_session creates a root session in that directory, attaches it to the relay heartbeat set, and returns the new session id only after the heartbeat completes so the mobile client can navigate immediately. Failures are sanitized; the command is not auto-retried and the user may retry manually after a transient relay failure.

Patch Changes

• #12303 7c07a47 - Improve CLI sidebar usage sections with cent-formatted costs, collapsible details, and aligned model totals.

• #12304 79fe757 - Display here-document content as plain text in terminal approval prompts.

• #12302 ad66c90 - Clarify when reverting a conversation does not restore workspace changes and link disabled snapshots to the Checkpoints setting.

• #12329 084bcea - Fix Cloud Agent session imports in installed CLI builds and prevent malformed exports or write failures from leaving partial imports.

• #12314 520679a - Keep the CLI sidebar branch label in sync when Git branches change outside Kilo.

• #12378 b402cc2 - Include image-output models in the Kilo Gateway chat model list.

• #12335 99227f6 Thanks @shssoichiro! - Start the TUI from standalone CLI installations without requiring project-local OpenTUI dependencies.

• Updated dependencies [084bcea, c081f58, ff703fb]:

  • @kilocode/kilo-gateway@7.4.12
  • @kilocode/kilo-indexing@7.5.0
  • @kilocode/kilo-telemetry@7.4.12
  • @opencode-ai/server@7.4.12
  • @opencode-ai/tui@7.4.9
  • @opencode-ai/ui@7.4.12

7.4.11

Minor Changes

• #12255 e084ab7 - Improve CLI project-memory controls, status, activity indicators, and optional recall details.

• #12250 bd69158 - Support verbose project-memory settings and show recalled memory snippets in conversation markers when enabled.

Patch Changes

• #12242 06c2337 - Speed up local conversation recall searches on large histories.

• #12252 e67635d - Restore directory @-mentions by listing their entries without inlining child file contents. Untrusted external directory attachments remain denied.

• #12274 5180c10 - Show newly submitted messages immediately after reverting a conversation.

• #12086 c654f1e Thanks @rakshith1928! - Fix Grok 4.5 reasoning variants not showing up in the model picker.

• #12267 e3124d3 - Retry incomplete model responses that end without final output or tool activity while preserving partial answers and completed tools.

• Updated dependencies [319f159, 30e7ec4, bd69158]:

  • @kilocode/kilo-gateway@7.4.10
  • @kilocode/kilo-memory@7.5.0
  • @kilocode/sdk@7.5.0
  • @kilocode/kilo-indexing@7.4.10
  • @kilocode/kilo-telemetry@7.4.10
  • @kilocode/plugin@7.4.10
  • @opencode-ai/ui@7.4.10
  • @opencode-ai/server@7.4.10
  • @kilocode/plugin-atomic-chat@7.4.10

7.4.9

Patch Changes

• #12244 fe41426 - Fix Agent Manager tool calls through providers that require object-root input schemas without root combinators.

• #12243 e4ceeae - Prevent stalled operating system process queries from blocking background process management.

7.4.8

Minor Changes

• #12159 1083bb8 - Report active CLI and VS Code app and session presence.

Patch Changes

• #12160 ba6e5b9 - Allow persistent approval for shell access to a specific global skill directory while keeping other Kilo configuration protected.

• #12097 22d6edb - Release project file handles immediately after reads on Windows so editors and tools can replace existing files without restarting Kilo.

• #12175 bd08c13 - Preserve the selected model reasoning variant when forking a session.

• #12128 ad2cc71 Thanks @rakshith1928! - Surface an invalid Kilo indexing.model configuration as an indexing Error status instead of silently falling back to the default model.

• #11783 6a3e5f3 - Inherit sandbox state when a sandboxed agent starts new Agent Manager sessions.

• #12203 750b622 - Keep Agent Manager sessions running when concurrent branch-name generation times out during model refresh.

• #12174 3ba4c33 - Inspect managed Agent Manager sessions and send a targeted prompt to an idle existing session from the native Agent Manager tool. Require a separate explicit approval before prompting another managed session.

• #12156 6f11e35 - Preserve gateway and provider errors when chunked compaction fails instead of reporting every failure as a context overflow.

• #12205 2045190 - Temporarily disable free-model session and Git workspace data export.

• #12158 3b1e07c - Enforce read and ignore permissions when file mentions add content to a prompt.

• #12207 c49560a - Keep shared session databases writable by released Kilo clients after newer schema migrations run.

• #11424 3a4438e - Stop active Agent Manager sessions and their subagents when a session tab or the Agent Manager tab closes.

• Updated dependencies [6a3e5f3, 227c65d, 3ba4c33]:

  • @kilocode/sdk@7.4.8
  • @kilocode/kilo-indexing@7.4.8
  • @kilocode/plugin@7.4.8
  • @opencode-ai/ui@7.4.8
  • @kilocode/kilo-gateway@7.4.8
  • @kilocode/plugin-atomic-chat@7.4.8
  • @opencode-ai/server@7.4.2
  • @kilocode/kilo-telemetry@7.4.8

7.4.7

7.4.6

Minor Changes

• #12075 1e0b25a - Support configuring network destinations that sandboxed tools can reach while network access is otherwise restricted.

Patch Changes

• #12073 71aa54e - Inherit the current model and reasoning variant when Agent Manager starts sessions without explicit overrides.

• #12166 4618f1b - Preserve dynamic tool properties when removing unsupported regex lookarounds.

• #12164 039b73d - Wait for the primary codebase index before indexing a linked worktree, preventing large worktrees from consuming excessive CPU during startup.

• #12106 b6b55d1 - Make session model usage easier to scan with collapsible summary rows and aligned steps and cost columns.

• #12093 8b46601 - Speed up VS Code settings saves by draining pending prompts and disposing worktree instances concurrently.

• #12079 0a64070 - Resolve AWS Bedrock credentials from SSO profiles in packaged CLI builds.

• #12101 bf2b33b Thanks @Githubguy132010! - Use the correct filePath argument name in the Gemini system prompt.

• #12149 05dadaa Thanks @umi008! - Fix Gemma 4 models failing with "thinkingLevel not supported" when using Google AI Studio.

• #12148 77f7983 Thanks @umi008! - Install the latest stable CLI release when newer non-CLI or prerelease releases exist.

• #12167 988a92e - Fix kilo upgrade for curl installs resolving the wrong latest version

  The upgrade command's version resolution for curl-detected installations used GitHub's /releases/latest endpoint, which now returns JetBrains plugin releases (e.g. jetbrains/v7.0.4) instead of the latest CLI release. This caused kilo upgrade to fail for curl installs. Version resolution now uses the npm latest dist-tag, matching the install script fix.

• #11837 654e10e Thanks @mjnaderi! - Show the Kilo Gateway rate-limit message when login has too many pending authorization requests.

• #12162 3ee9144 - Show troubleshooting and migration guidance when Google Gemini rejects API credentials.

• #11955 cac82a3 Thanks @jstar0! - Prevent Gemini requests from failing when MCP tool schemas contain required fields without matching object properties.

• #12153 be15cf4 - Allow GPT-5.6 models to use tools whose JSON schemas contain regex lookarounds.

• #12168 032f3bb - Block environment and out-of-project file substitutions in project markdown configuration.

• #12040 93c209b Thanks @rakshith1928! - Hide gpt-5.5-pro from the model picker when using ChatGPT OAuth login, since Codex rejects it with HTTP 400.

• #12087 1f99fb2 - Stop explicitly directing GPT and Codex models to delegate tasks to subagents.

• #12105 e0bfed3 - Shut down the headless kilo serve process automatically when the editor client that launched it exits without a clean signal, preventing orphaned CLI processes.

• #12092 94b553b - Show detailed GPT-5.6 reasoning summaries and avoid expandable blank panels when a provider returns only a summary title.

• Updated dependencies [039b73d, 1e0b25a]:

  • @kilocode/kilo-indexing@7.4.6
  • @kilocode/sdk@7.5.0
  • @kilocode/plugin@7.4.6
  • @opencode-ai/ui@7.4.6
  • @kilocode/kilo-gateway@7.4.6
  • @kilocode/plugin-atomic-chat@7.4.6
  • @kilocode/kilo-telemetry@7.4.6

7.4.4

Minor Changes

• #12049 394af39 - Configure sandboxing through first-class sandbox settings, and show its controls in the dedicated Sandboxing page for all supported macOS and Linux users while keeping it disabled by default.

Patch Changes

• Updated dependencies [394af39]:
  • @kilocode/sdk@7.5.0
  • @kilocode/plugin@7.4.4
  • @opencode-ai/ui@7.4.4
  • @kilocode/kilo-gateway@7.4.4
  • @kilocode/kilo-indexing@7.4.4
  • @kilocode/plugin-atomic-chat@7.4.4
  • @kilocode/kilo-telemetry@7.4.4

7.4.3

Minor Changes

• #12067 ed36326 - Support task-aware pruning of agent-invoked Bash output with experimental SWE-Pruner.

Patch Changes

• #12052 61d90f1 - Exclude directory-scoped AGENTS.md instructions from SWE-Pruner context.

7.4.2

Minor Changes

• #11921 b976b5a Thanks @johnnyeric! - Add opt-in project memory commands, tools, automatic capture, and public API support.

• #12004 cef3dc7 - Add a reload action that reboots the per-directory instance, picking up config, skills, agents, commands, and MCP prompts changed on disk. Sessions and history are preserved. Surfaces: /reload in the CLI palette and editor chat, a reload button in the task header and settings panel, the Kilo Code: Reload Config and Skills command, and a POST /instance/reload HTTP endpoint. The endpoint returns 409 while a session is actively running.

• #11835 cd49ae6 - Support provider-aware model discovery and selection for remote Cloud sessions.

• #11980 adcbe0f Thanks @Drilmo! - Add experimental SWE-Pruner support (disabled by default). When enabled via experimental.swe_pruner or the Experimental settings tab in VS Code, the read and grep tools accept an optional context_focus_question parameter; when the agent provides it, large tool outputs are pruned by a small model down to the lines relevant to that question, with omitted sections marked inline and a SWE-Pruner · kept/total indicator on the tool row. The skimming model can be overridden via experimental.swe_pruner_model (defaults to the configured small model). Any pruning failure falls back to the full output.

• #11428 69f5b9d Thanks @drye! - Add vim modal editing to the CLI prompt input. Enable it with "vim": true in tui.jsonc, the Toggle vim mode command in the command palette, or the /vim slash command. Supports NORMAL-mode motions (h/j/k/l, w/b/e, 0/^/$, gg/G, counts), edits (x, dd, dw, cw, D, C, r, yy/p, u, Ctrl+r), insert transitions (i/a/A/I/o/O), and VISUAL / VISUAL-LINE mode (v/V with selection-extending motions, d/x/c/s/y, o to swap ends), with a mode indicator and matching cursor shape.

Patch Changes

• #11223 4104ab5 Thanks @maphew! - Fix cloud session fork commands so they import cloud sessions before validating the local session.

• #12033 9fc1a1d - Show a clear "No changes found to generate a commit message for" error instead of a generic "Unexpected server error" when there is nothing to commit. The endpoint now returns a typed 422, and the extension surfaces the real message directly.

• #11886 b793bf7 - Harden config credential substitution against untrusted project config. Environment references ({env:VAR}) now resolve only in trusted config (global config, KILO_CONFIG, KILO_CONFIG_CONTENT, and org/MDM-managed config); a project-committed kilo.json / opencode.json can no longer use them. File references ({file:...}) still work in project config but are confined to the project root, so absolute paths, ../ traversal, and symlink escapes are rejected. This closes a path where a malicious repository could exfiltrate local secrets to an attacker-controlled baseURL.

• #12002 885a994 - Defer Agent Manager automatic branch naming until the conversation shows a durable task. The first user message no longer renames the branch; naming waits for a second message (up to four) or for the worktree to contain changes, and renames only run while the session is idle. Read-only verification questions (for example "is X fixed?") no longer claim the branch name.

• #11968 7571508 - Fix Amazon Bedrock models returning no output. A smithy dependency version-skew made the Bedrock event-stream decoder silently fail under the browser build condition, so every Bedrock request completed with an empty response.

• #12042 22b9f7f Thanks @shssoichiro! - Respect nested .gitignore and .kilocodeignore files during codebase indexing.

• #11976 40790d8 - Show the Remote badge in the TUI prompt status area when remote session relay is enabled.

• #11999 61b9e09 - Use cloud account preferences to select the active Kilo organization and hide unavailable personal accounts.

• #11994 eefd891 Thanks @IOLOII! - Generate commit messages in the user's selected UI language instead of always using English.

• #11506 5135d2e Thanks @mvanhorn! - Show live session spend in the TUI sidebar while an assistant turn is still running.

• #12034 64c9b7e - Show a dismissible notification when a leftover opencode config directory is found. Kilo no longer falls back to opencode configuration, so the notice points you to move .opencode config into a .kilo directory (or the global kilo config dir). Dismiss it once and it won't return unless the directory is still present.

• Updated dependencies [b976b5a, 22b9f7f, 61b9e09, adcbe0f]:

  • @kilocode/sdk@7.5.0
  • @kilocode/kilo-memory@7.5.0
  • @kilocode/kilo-indexing@7.4.2
  • @kilocode/kilo-gateway@7.4.2
  • @kilocode/plugin@7.4.2
  • @opencode-ai/ui@7.4.2
  • @kilocode/kilo-telemetry@7.4.2
  • @kilocode/plugin-atomic-chat@7.4.2

7.4.1

Patch Changes

• #11887 51dc189 - Require authentication before enabling allow-everything permissions over HTTP.

• #11923 fda4e17 - Fail subagent permission prompts in headless kilo run immediately instead of hanging forever, and approve subagent permission prompts under --dangerously-skip-permissions

7.4.0

Minor Changes

• #11912 1f80fdf - Persist the /sandbox toggle across new CLI sessions per project directory, mirroring the VS Code extension's sandbox button. New sessions now inherit the last toggled state instead of resetting to the config default each time.

Patch Changes

• #11906 1d3a9e0 - Support adaptive reasoning presets for Claude Fable and Sonnet 5 models.

• #11084 69e5c58 Thanks @maphew! - Use /review as the single local review command, defaulting to staged, unstaged, and untracked changes while supporting guided uncommitted reviews, branch/base reviews, commits, and pull requests. Show deprecation notices for /local-review and /local-review-uncommitted that point to the matching /review modes.

• #11896 c36c293 Thanks @johnnyeric! - Report the plan file that was actually saved in Plan mode: point the "Plan is ready" link, the follow-up prompt, and the new-session handoff at the real file instead of a wrongly generated name, and fail plan_exit with a clear error when no plan was written.

• #11808 ce09eb3 - Show an interactive Implement / Keep refining panel when Plan mode is ready instead of asking users to type a numbered choice.

• #11891 9857c98 - Preserve model output capacity when requests contain encoded images. The output token cap now uses the provider-reported context size from the previous turn, so image and vision input is measured by the provider instead of by encoded payload size.

• #11838 eec075b - Retain the sandbox toggle state when forking a session or moving it to a worktree, instead of resetting it to the workspace default.

• #11898 067fcf5 - Keep sandboxing disabled by default unless the experimental sandbox setting or an explicit session toggle enables it.

• #11913 70a002d - Fix shell tool occasionally returning "(no output)" for fast-exiting commands

• #11496 bc0236b - Show the usable local IPv6 URL when the server binds to the IPv6 wildcard address.

• #11833 8cdd0aa Thanks @johnnyeric! - Add /cost-alert to get notified when a session's cost crosses a threshold you set.

• #11553 3847122 - Improve JetBrains agent, MCP, provider, and model settings so changes are staged until Apply, persist through the CLI, reload accurately, and hide unsupported removal actions.

• #11767 c94a097 - Fix non-default agents (Ask, Plan, and custom or organization agents) failing with a "Bad Request: Unsupported parameter(s)" error on some models and providers.

• #11701 61bc5d6 - Use model family metadata when selecting the apply_patch tool for GPT models.

7.3.63

Minor Changes

• #11714 7b2063f - Connect to a local Anaconda Desktop text-generation model server from the CLI or VS Code.

• #11786 123a939 - Allow Agent Manager chat tools to discover available models and reasoning variants by model name, then start each session with the chosen model and reasoning effort. Agent Manager resolves the provider for a named model automatically, preferring the provider behind the current default model and falling back to the Kilo Gateway.

• #11456 afa9633 - Allow background processes to transfer from subagents to parent sessions, or remain accessible from every session in their project after Kilo restarts.

• #11394 bbf3c5b - Run commands that require human interaction in an embedded CLI terminal dialog and return their output to the model when complete.

• #11729 7d64eb7 Thanks @johnnyeric! - Show personal credits, team credits, and Kilo Pass in the CLI sidebar, and refresh the balance immediately after switching teams.

• #11659 7f4702b - Enforce the sandbox network restriction for agent commands on Linux, including TCP, UDP, IPv4, IPv6, and descendant processes.

• #11603 9fbc456 - Block outbound network access from agent commands and in-process HTTP tools with the optional macOS sandbox, with a Sandboxing setting to allow network access when needed.

• #11548 c55e804 - Confine agent shell and file-tool writes to project and Kilo state directories with the optional macOS and Linux sandboxes.

• #11628 2638e06 - Add session-local macOS sandbox controls, show the effective active state, and confirm toggles in the CLI and VS Code extension.

Patch Changes

• #11762 d89b1b6 - Gate experimental agents on their declared skill, MCP, and VS Code extension requirements. VS Code shows requirement groups with Marketplace shortcuts, and the CLI stops before sending when requirements are unmet.

• #11594 f69d1cd - Keep turns responsive when snapshot infrastructure stalls and prevent transient snapshot progress from appearing in forked sessions.

• #11526 579a787 - Run Windows PowerShell tool commands without -EncodedCommand to reduce antivirus false positives.

• #11505 55203c3 Thanks @shssoichiro! - Create the default .kilo/plans directory automatically when Plan mode starts.

• #11601 2404009 - Fix kilo upgrade for curl installs by pointing at the install script instead of the install landing page.

• #11798 1d798a1 - Fix opening KiloClaw from the CLI and VS Code slash commands.

• #11744 6d25c1b - Allow the default TUI to import cloud-only sessions without rejecting their IDs as missing locally.

• #11721 be1f77d Thanks @johnnyeric! - Expose Kilo Pass state on the Kilo profile API contract.

• #11638 117a0d6 - Stop loading .opencode config directories and use .kilo instead, while retaining .kilocode as a legacy fallback.

• #11646 61bbc34 - Release disconnected event streams so long-running servers do not retain queued session diffs.

• #11696 be3ae82 - Remember sandbox choices per session and start new sessions with the last selected sandbox state.

• #11703 163aef5 - Prevent confined sessions and delegated agents from weakening their sandbox policy through configuration changes or unauthenticated server control.

• #11591 76a9d9b - Prevent sandboxed file tools from escaping project write roots through concurrent symlink replacement on macOS.

• #11584 588335e - Confine sandboxed worktree sessions to their active worktree instead of allowing writes to sibling or primary checkouts.

• #11556 9b0c45c - Show the concrete model reported for routed Kilo auto-model steps in CLI and VS Code session timelines, and break down TUI sidebar token usage, cache rate, and cost by model across subagent sessions.

• #11621 8ac629c Thanks @maoxin1234! - Surface the resumable task_id when a subagent stops on an error. Both foreground and background subagent failures now tell the parent agent that the session can be resumed via the task tool with task_id="<id>", so a stopped subagent can be continued instead of being lost.

• #11746 5080c78 - Include session-tree IDs in model usage API responses and show full task token usage with a provider-grouped model breakdown in the VS Code session header.

• Updated dependencies [7b2063f, 123a939, dcd2ae3, 1d798a1, be1f77d, be3ae82, 9b0c45c, 2638e06, 5080c78]:

  • @kilocode/sdk@7.4.0
  • @kilocode/kilo-gateway@7.3.55
  • @kilocode/plugin@7.3.55
  • @opencode-ai/ui@7.3.55
  • @kilocode/kilo-indexing@7.3.55
  • @kilocode/kilo-telemetry@7.3.55
  • @kilocode/plugin-atomic-chat@7.3.55

7.3.54

Patch Changes

• #11555 5c1dcdf - Use the correct High and Max thinking variants for GLM 5.2 on OpenCode Go and compatible providers.

7.3.53

Minor Changes

• #11468 27bd206 - Search titles and high-signal transcript content across all local sessions with the recall tool.

Patch Changes

• #11533 15f42d4 - Restore bounded text-file reads and keep zero-limit pagination and Unicode truncation from producing unusable tool output.

• Updated dependencies [6c55c28]:

  • @kilocode/kilo-gateway@7.3.53
  • @kilocode/kilo-indexing@7.3.53
  • @kilocode/kilo-telemetry@7.3.53
  • @opencode-ai/ui@7.3.53

7.3.52

Patch Changes

• #11450 cc924a6 - Changes from opencode v1.15.9 to v1.15.13 upstream:
  • Core Improvements: Added headerTimeout config for provider requests, with a 10s default for default OpenAI setups.
  • Core Improvements: Experimental background agents now push updates without polling.
  • Core Improvements: You can now set only modalities.input or modalities.output in config. (@robposch)
  • Core Improvements: Remote-backed projects now resolve a stable project identity.
  • Core Improvements: ACP integrations can now send prompts, slash commands, and usage updates through acp-next
  • Core Improvements: Added WebSocket transport for OpenAI responses on supported channels (set KILO_EXPERIMENTAL_WEBSOCKETS=true)
  • Core Improvements: Sessions can now store custom metadata through the API and SDK. (@shantur)
  • Core Improvements: Config now loads from the opened location upward, so directory-specific settings and provider policies apply more predictably.
  • Core Bugfixes: Dynamically added MCP servers now disconnect cleanly when removed.
  • Core Bugfixes: DigitalOcean inference now uses your OAuth token directly instead of creating a MAK. (@Spherrrical)
  • Core Bugfixes: Config loading now falls back cleanly when user info is unavailable.
  • Core Bugfixes: Fixed Google tool calling after the upstream tool ID regression.
  • Core Bugfixes: Experimental flags can now override the umbrella experimental flag.
  • Core Bugfixes: Resumed sessions no longer continue orphaned interrupted tools. (@edevil)
  • Core Bugfixes: OpenAI reasoning summaries now render as separate blocks.
  • Core Bugfixes: Updated Google Vertex support for reasoning signatures.
  • Core Bugfixes: The shell tool now advertises your configured timeout to the model.
  • Core Bugfixes: Enabled adaptive reasoning controls for Anthropic Opus 4.7+ models
  • Core Bugfixes: Allowed colons in passwords (@neriousy)
  • Core Bugfixes: Sped up warm acp-next model and config switches
  • Core Bugfixes: Improved first-session acp-next startup time
  • Core Bugfixes: Kept OpenAI WebSocket response timeouts active
  • Core Bugfixes: Retried failed OpenAI WebSocket streams before falling back
  • Core Bugfixes: Handled acp-next permission prompts correctly
  • Core Bugfixes: Used the persisted session directory for existing-session requests
  • Core Bugfixes: Forwarded remote workspace request bodies correctly
  • Core Bugfixes: Supported custom base URLs for OpenAI WebSocket responses (@Tarquinen)
  • Core Bugfixes: Gateway Anthropic Opus 4.7+ adaptive reasoning now keeps summarized thinking instead of returning empty thinking blocks.
  • TUI Improvements: Made the prompt resize with terminal width and added prompt size config. (@bjschafer)
  • TUI Improvements: Added a workspace management dialog
  • TUI Bugfixes: Accelerated diff viewer scrolling.
  • TUI Bugfixes: External editors now open from the worktree directory when available.
  • TUI Bugfixes: Kept session navigation working while prompt modes are open
  • TUI Bugfixes: Restored the thinking spinner
  • TUI Bugfixes: Surfaced subagent retry status
  • TUI Bugfixes: Fixed opening editors from non-Git project paths (@OpeOginni)
  • TUI Bugfixes: Wrapped inline tool rows now stay aligned, and failed inline tools can expand their error details in place.
  • Extensions Improvements: Added a dispose hook for plugins.
  • Extensions Bugfixes: Fixed Codex plugin requests to send the expected session ID header.

7.3.51

Minor Changes

• #11478 9611c8b - Support stopping the daemon with kilo console stop and keeping console or daemon commands attached with --foreground

• #10005 1d030dc - Support kilo run --command compact and --command summarize to compact the current session, matching the TUI's /compact and /summarize slash commands.

7.3.50

Minor Changes

• #11421 ccec216 - Show a BYOK badge for Kilo Gateway models that can use an enabled personal or organization provider key.

• #11028 a6ded9b Thanks @IamCoder18! - Display local and network URLs when the server binds to 0.0.0.0

Patch Changes

• #11412 2c9e72c - Deny provider data collection for Kilo Gateway requests when prompt-training models are hidden.

• #11301 081b653 - Add a privacy filter to the Console model explorer that hides Kilo Gateway models whose providers may use prompts for training.

• #11026 e2ebf8b Thanks @IamCoder18! - Skip automatic browser launch on Linux when no display is detected.

• #11212 8649ab6 Thanks @IamCoder18! - Show docs URL in dialog when no display server is detected on headless Linux systems

• #11319 fb37d9c Thanks @grandmaster451! - Show the docs URL in an alert dialog when the browser cannot be opened on headless systems instead of silently failing.

• #11455 4d09333 - Hide reverted provider errors so Redo controls remain visible after rewinding a session.

• #11475 3d4ccc2 - Preserve custom subagent tool permissions when tasks inherit restrictions from their parent agent.

• #11453 f7e68d1 - Limit completion sounds to parent agent sessions.

• Updated dependencies [ccec216, 2c9e72c, f7e68d1]:

  • @kilocode/kilo-gateway@7.4.0
  • @kilocode/sdk@7.3.50
  • @kilocode/kilo-indexing@7.3.50
  • @kilocode/kilo-telemetry@7.3.50
  • @kilocode/plugin@7.3.50
  • @opencode-ai/ui@7.3.50
  • @kilocode/plugin-atomic-chat@7.3.50

7.3.49

7.3.48

Minor Changes

• #11182 973d02c - Share the main codebase index with Agent Manager worktrees while indexing and searching only each worktree's changed files.

• #10781 66af690 Thanks @shssoichiro! - Add opt-in Unicode or emoji terminal title indicators for sessions that are working, need attention, or have finished.

Patch Changes

• #11242 9211000 - Support unauthenticated OpenAI-compatible endpoints for codebase indexing without requiring a placeholder API key.

• #11305 04ed322 - Prevent snapshot initialization progress from blocking conversations after the slow repository prompt.

• #11249 2c30dc7 - Show a clear, retryable provider rate-limit error instead of raw response JSON in chat.

• #11171 04ebc74 - Hide TUI news after they have been opened and add a button to close the news dialog.

• #10929 9329682 Thanks @shssoichiro! - Make /copy copy the latest agent response and use /copy-session for session transcripts.

• #10091 be234fa Thanks @shssoichiro! - Always deny tool calls for title, summarize, and compaction

• #11264 f78e54c - Fix upgrades to resolve Kilo CLI packages and releases instead of OpenCode packages and versions.

• #11347 b518a76 - Identify Kilo in provider request user-agent headers instead of OpenCode.

• #11279 e91eef2 - Show current-worktree sessions by default in the TUI sessions dialog and keep all/current scope toggling working when a scope has no sessions.

• #11158 8ff8371 - Add a shared model setting to hide Kilo Gateway models that may train on your prompts across Kilo clients.

• #11270 c5d39d0 - Replace remaining OpenCode-branded CLI and TUI copy with Kilo branding.

• #11279 2f69c13 - Show Agent Manager and other Git worktrees in the Kilo Console project view.

• #11291 4436139 - Load the bundled Atomic Chat integration without attempting to install an unpublished npm plugin.

• #11236 1511d13 Thanks @kapelame! - Add an instant/thinking reasoning toggle for MiniMax M-series models, matching the existing glm/kimi/qwen behavior.

• #11170 3845918 Thanks @johnnyeric! - Make native Plan mode follow Architect-style planning behavior while preserving Plan mode restrictions and repo-root plan files.

• #11257 f42789d - Changes from opencode v1.14.51 to v1.15.4 upstream:

  • Core Improvements: Clarified how to recover when the npm package is installed without its native binary.
  • Core Improvements: Reduced unnecessary prompting around shell, task, and todo flows.
  • Core Bugfixes: Ignored invalid exports in custom tool modules instead of failing tool loading.
  • Core Bugfixes: Ignored project instruction lookup errors so sessions keep loading when project instruction discovery fails.
  • Core Bugfixes: Fixed versioned event projector lookups so event replay uses the right handlers.
  • Core Bugfixes: Avoid duplicate consecutive entries in prompt history.
  • Core Bugfixes: Show full config validation errors during TUI startup instead of a generic failure.
  • Core Bugfixes: Fixed npm installs so the CLI can recover and fetch the right native binary on more setups.
  • Core Bugfixes: Fixed multiline @ mentions in prompts.
  • Core Bugfixes: Preserved custom tool metadata from Zod schemas.
  • Core Bugfixes: Preserved custom tool argument descriptions in generated schemas.
  • Core Bugfixes: Fixed file watching in repos where .git is a symlink. (@kagura-agent)
  • Core Bugfixes: Fixed sync events not reaching project-scoped subscribers in injected instances.
  • Core Bugfixes: Reduced wasted work when reading very large files after output truncation.
  • Core Bugfixes: Fixed project-scoped bus events so file watcher and update notifications reach the right instance.
  • Core Bugfixes: Fixed custom LSP servers not sending refresh events after they initialize.
  • Core Bugfixes: Hid background subagent task instructions unless experimental background mode is enabled.
  • TUI Improvements: Added a collapsed thinking view that can be expanded inline.
  • TUI Improvements: Added pinned sessions with quick-switch slots in the session picker.
  • TUI Improvements: Newly pinned sessions now stay at the end of the pinned list instead of jumping to the top.
  • TUI Improvements: Made Markdown H1 headings easier to distinguish.
  • TUI Bugfixes: Fixed thinking mode defaults so reasoning starts collapsed consistently.
  • TUI Bugfixes: Limited session quick-switching to pinned sessions.
  • TUI Bugfixes: Fixed Markdown table rendering in chat output.
  • TUI Bugfixes: Fixed kilo run --agent resolving project-local agents.
  • TUI Bugfixes: Fixed async commands losing the active instance context, which could break agent generation and GitHub-driven runs.

• #11356 326ff35 - Changes from opencode v1.15.4 to v1.15.9 upstream:

  • Core Improvements: Preview the native OpenAI runtime path behind an experimental flag
  • Core Improvements: Add --replay and --replay-limit to show recent history when resuming interactive runs
  • Core Improvements: Added a diff viewer in the TUI for reviewing changes.
  • Core Improvements: Collapsed single-child directories in the diff viewer file tree.
  • Core Improvements: Added shell mode to the run prompt.
  • Core Improvements: Replaced subagent tabs with an on-demand picker in run.
  • Core Improvements: Plugin file load errors no longer break the rest of plugin loading.
  • Core Improvements: Anthropic API-key models now use the native runtime.
  • Core Improvements: The v2 HTTP API now exposes structured public error schemas.
  • Core Improvements: Added Grok OAuth sign-in, including device-code login. (@Jaaneek)
  • Core Improvements: Redesigned the diff viewer with a file tree and refreshed layout.
  • Core Bugfixes: Fix plugin tools using ask so tool calls complete correctly
  • Core Bugfixes: Reduce missed /event updates caused by a subscription race
  • Core Bugfixes: Sort the v2 session list by most recently updated
  • Core Bugfixes: Zed editor context now only activates inside Zed terminals.
  • Core Bugfixes: Agent and command names now resolve correctly from relative config paths.
  • Core Bugfixes: Invalid KILO_PERMISSION JSON no longer crashes startup.
  • Core Bugfixes: Plugin tools with missing args no longer break tool loading.
  • Core Bugfixes: Restored legacy PgUp and PgDn TUI keybind aliases.
  • Core Bugfixes: Native runtime now prefers the console provider token for OpenCode models.
  • Core Bugfixes: V2 session APIs now return safe UnknownError responses with log reference IDs when stored messages are corrupt.
  • Core Bugfixes: Generic API 500s no longer expose config details from server errors.
  • Core Bugfixes: Unknown API errors now include reference IDs so you can match responses to server logs.
  • Core Bugfixes: V2 session APIs now return 503 ServiceUnavailableError for mutations that are not available yet.
  • Core Bugfixes: V2 session APIs now return SessionNotFoundError for missing sessions.
  • Core Bugfixes: Deduped concurrent Codex OAuth refreshes to avoid repeated refresh failures. (@cooper-oai)
  • Core Bugfixes: Restored native OpenAI OAuth requests.
  • Core Bugfixes: Tool schema failures now surface as friendly tool errors.
  • Core Bugfixes: Added PDF attachment support for Grok.
  • Core Bugfixes: Restored OpenAI reasoning streams.
  • Core Bugfixes: Return to the previous screen when closing the diff viewer.
  • Core Bugfixes: Show clearer errors when a default model is invalid or unavailable.
  • Core Bugfixes: Surface missing PTY session errors instead of failing generically.
  • Core Bugfixes: Improve diff viewer empty states and context handling.
  • Core Bugfixes: Show clearer errors when a skill invocation fails as expected.
  • Core Bugfixes: Show clearer errors when an installation upgrade fails.
  • Core Bugfixes: Show clearer project not found errors from the HTTP API.
  • Core Bugfixes: Return PTY error bodies from the HTTP API.
  • Core Bugfixes: Enable the diff viewer by default.
  • Core Bugfixes: Return MCP server not found errors from the HTTP API.
  • Core Bugfixes: Let MCP OAuth configs set a callback port and include configured scopes in client metadata. (@sebin)
  • Core Bugfixes: Use working Vertex Anthropic endpoints for us and eu multi-region setups. (@JPFrancoia)
  • Core Bugfixes: Return session busy error bodies from the HTTP API.
  • Core Bugfixes: Preserve native reasoning continuation metadata across turns.
  • TUI Improvements: Refresh the prompt layout after pasting content
  • TUI Improvements: The diff viewer now focuses the first file automatically.
  • TUI Improvements: Copy the current worktree path from the command palette.
  • TUI Bugfixes: Keep file references scoped to the current workspace
  • TUI Bugfixes: Preserve pasted prompt content when copying
  • TUI Bugfixes: Collapse very long tool output lines to keep the layout readable
  • TUI Bugfixes: Use a higher-contrast paste summary badge color in some themes (@kagura-agent)
  • TUI Bugfixes: Imported sessions now refresh their directory and relative path fields correctly. (@OpeOginni)
  • TUI Bugfixes: Collapsed thinking labels now use clearer punctuation.
  • TUI Bugfixes: New sessions now default to the local project.
  • TUI Bugfixes: Single-select question checkmarks no longer run into option labels.
  • TUI Bugfixes: Refine diff viewer keyboard shortcuts.
  • TUI Bugfixes: Restore question prompt key handling.
  • TUI Bugfixes: Keep the spinner color aligned with the active agent. (@OpeOginni)

• #11245 046b03a Thanks @johnnyeric! - Restore session timestamp prefixes for generated plan filenames while preserving descriptive model-chosen names.

• #9807 9394100 Thanks @truffle-dev! - Prevent unreachable telemetry endpoints from blocking or failing completed CLI commands.

• #11279 8c1cdf5 - Keep expanded Kilo Console file diffs open while resizing the context sidebar.

• #11279 2f69c13 - Keep Kilo Console terminal sessions open when changing diff layout and other console preferences.

• #11354 b2eef5c - Prevent the bundled Atomic Chat plugin from triggering an npm installation.

• #11373 f21a34a - Skip attention sounds when a session is manually interrupted.

• #11295 2fa0890 - Restore speech input when profile details are unavailable, move transcription model selection to the Models tab, and default transcription to Whisper Large V3 Turbo.

• #9758 8db7b68 - Restore files to their original paths when reverting a task that moved or renamed them.

• #11410 344a6a5 - Keep server controls and events connected to active sessions and subagents.

• #11221 987da27 - Show shared provider descriptions and provider icons in JetBrains and VS Code provider settings.

• #11262 0903183 - Expose the prompt-training model filter in the Kilo Console model settings.

• #10758 e511b23 Thanks @cooper-oai! - Prevent concurrent Kilo processes from reusing a ChatGPT Codex refresh token.

• Updated dependencies [9211000, 2fa0890, 973d02c, 66af690]:

  • @kilocode/kilo-indexing@7.4.0
  • @kilocode/sdk@7.4.0
  • @kilocode/plugin@7.3.47
  • @opencode-ai/ui@7.3.47
  • @kilocode/kilo-gateway@7.3.47
  • @kilocode/plugin-atomic-chat@7.3.47
  • @kilocode/kilo-telemetry@7.3.47

7.3.46

Patch Changes

• #11184 adf03a9 - Support model-specific reasoning overrides for task subagents, including custom subagents with their own model and variant settings.

• #11178 f63e771 - Accelerate initial snapshots for regular Git sessions while preserving existing changes and asynchronously storing snapshots independently from the source repository.

• Restore Kilo branding, fork-specific CLI commands, and CLI lifecycle initialization after upstream merges.

• #11240 f820e57 - Prevent skill removal from recursively deleting working directories.

• #11179 96a1610 - Validate GitHub attachments and language server release paths before downloading or executing them.

7.3.45

Patch Changes

• #11152 b23d3df - Allow Escape to stop Agent Manager prompts while their sessions are still starting.

• #11138 e354305 Thanks @IamCoder18! - Restart the daemon when kilo console or kilo daemon start receives explicit network options that don't match the running daemon, instead of silently ignoring the requested settings.

7.3.44

Minor Changes

• #11082 a16e82a - Use embedded LanceDB as the default semantic search vector store so indexing works without a separate Qdrant server. Existing Qdrant users and Intel Mac users can select qdrant with indexing.vectorStore.

Patch Changes

• #10922 bc3af9a - Prevent unnecessary repeat auto-compactions when providers report inconsistent token totals.

• #11160 78d83c0 - Preserve the calling model's reasoning effort when task subagents inherit that model.

• #10478 5bc8df8 - Allow hosted runtimes to cap shell command duration and explain environment-enforced timeouts.

• #11085 2a6596b - Indicate when no models are available in model-not-found errors.

• #11072 6920f37 - Speed up the first Agent Manager prompt in new worktrees by seeding snapshots from the checkout's Git index.

• #11075 e17ce0c - Speed up large session forks by retaining final task outcomes instead of duplicating resumable subagent histories, and load completed task details only when expanded.

• #11143 12144cf Thanks @IamCoder18! - Warn when kilo console or kilo daemon is invoked with an explicit --port outside the discovery range (4097–4116).

• #11006 69a0b38 - Refresh connected provider model lists when the models catalog updates.

• #11081 9c279a1 - Show model free and prompt-training indicators only when their explicit catalog metadata is enabled.

• #11101 294c532 - Prevent task subagents from asking questions that users cannot answer from the parent session.

• #11102 8a72708 - Prevent duplicate CLI attention alerts and route Kilo prompts through the configurable notification system.

• #10866 d5112ed - Stabilize code indexing workers, retry Kilo model catalog downloads, reduce progress log noise, and show indexing failures as TUI notifications instead of writing over the terminal interface.

• #11147 9a187d5 - Configure the project context sidebar width and default diff layout from Global Settings.

• #11091 57bef8a - Prevent streamed tool calls from executing twice and leaving answered questions disabled in VS Code.

• #11139 7226635 - Restore Kilo branding, fork-specific CLI commands, and CLI lifecycle initialization after upstream merges.

• #11031 bbfd59b - Remove the unsupported code search tool.

• #11117 b75af0d - Update the Vercel AI SDK providers for Cerebras, xAI, and OpenAI-compatible endpoints.

• #10866 d5112ed - Support configuring code indexing separately for global and project settings in Kilo Console, the CLI TUI, and VS Code.

• #11031 28a26b1 - Restore streamed responses in the CLI TUI and move code indexing status into the session sidebar.

• Updated dependencies [a16e82a, 9c279a1, 57bef8a, b75af0d]:

  • @kilocode/kilo-indexing@7.4.0
  • @kilocode/kilo-gateway@7.3.43
  • @kilocode/kilo-telemetry@7.3.43
  • @opencode-ai/ui@7.3.43

7.3.42

Patch Changes

• #11064 db7707d - Allow local review follow-up fix prompts to modify code after explicit user approval.

• #11050 8535d3d - Keep new Kilo Console terminals open in the TUI on macOS.

• #11011 9f072b0 - Re-enable free-model session and Git workspace data export.

• #10751 6e8d6f7 - Sync CLI sessions to Kilo session history when authenticated with KILO_API_KEY when no stored Kilo auth is present.

7.3.41

Minor Changes

• #10761 82b22f7 Thanks @IamCoder18! - Support reading .ods (OpenDocument Spreadsheet) files in the read tool

• #10879 b0a4f03 - Show Terminal Bench completion scores and per-attempt costs in supported model details.

• #10948 6ee090b - Restore cloud session filesystem changes from synced session diffs when importing sessions, including inherited changes across imported session forks.

Patch Changes

• #10996 cc03ffc - Preserve image attachments when Photon is unavailable, enforce attachment limits for user images, and correlate shell lifecycle events correctly.

• #10998 a59b255 - Restore automatic session titles for models that require reasoning without assuming a supported effort level.

• #11004 16e334f - Discover project-installed skills in Agent Manager worktree sessions.

• #11000 741b00f - Keep subagent sessions isolated when forking sessions through editor clients.

• #10991 ece8453 Thanks @shssoichiro! - Avoid copying visible planning chat into new sessions started from the plan follow-up prompt.

• #11034 0d76fa6 - Start forked sessions at zero cost instead of carrying over the source session's spend.

• #10109 df30123 Thanks @IamCoder18! - Prevent memory leak in KiloSessionPromptQueue.cancel for sessions without active tails

• #11010 a130641 - Compact sessions at the configured context percentage before sending an oversized provider request.

• Updated dependencies [b0a4f03]:

  • @kilocode/kilo-gateway@7.4.0
  • @kilocode/kilo-indexing@7.3.41
  • @kilocode/kilo-telemetry@7.3.41

7.3.40

Patch Changes

• #10925 881a451 Thanks @evanjacobson! - Display skills in CLI slash command autocomplete options

• #10952 be5f42f Thanks @johnnyeric! - Support custom plan file paths when exiting planning.

7.3.39

Patch Changes

• #10901 a8a8dd8 - Prevent icon images fetched from the web from causing provider request errors.

• #10933 a0eb3b7 - Write strict JSON when adding MCP servers to kilo.json configuration files.

• #10924 189f251 - Temporarily disable free-model session and Git workspace data export.

• #10949 78117d1 - Fail publication builds when the bundled models snapshot cannot be downloaded or validated, and load the snapshot as JSON data in compiled binaries.

7.3.33

Patch Changes

• #10935 6cab5f1 - Prevent the macOS Apple Silicon CLI from failing to start because of malformed bundled exports.

7.3.30

Patch Changes

• #10862 c4de1ac - Support reasoning with Mistral Medium 3.5 models, including the latest alias.

• #10895 2e1945c - Allow plan approval submissions to complete after planning finishes.

7.3.29

Patch Changes

• #10822 8b1ee66 - Preserve worktree routing for Kilo HTTP API clients and keep inherited task-subagent restrictions active.

7.3.28

Patch Changes

• #10847 cdf46c9 - Clarify when free-model data may be used for training and identify it with a brain circuit icon.

• #10833 8696edc - Keep Kilo Console terminals and worktree changes visible while refreshing diffs.

• #10833 fbacc31 - Use the updated favicon in Kilo Console.

• #10865 9c56107 - Show the animated Kilo logo while the console and dashboard finish loading.

• #10864 557d6ad - Preserve upstream error statuses for cloud session and KiloClaw gateway requests.

• #10831 837a875 - Keep post-compaction tool calls and follow-up messages ordered after the compaction summary in the CLI and VS Code transcript.

• #10849 a6b005d - Restore Cloud Agent transcripts in VS Code session previews and stop cloud session previews or continuation from loading indefinitely when a request stalls.

• #10883 1cdc398 - Restore kilo console startup in packaged CLI builds.

• #10863 35aa9bb - Restore Kilo Gateway-backed Mercury Next Edit completions.

• #10829 e64c1fb - Restore full-session forks in Agent Manager after the HTTP API migration.

• Updated dependencies [fc4cf10, a6b005d]:

  • @kilocode/sdk@7.3.23
  • @kilocode/kilo-gateway@7.3.23
  • @kilocode/plugin@7.3.23
  • @kilocode/kilo-indexing@7.3.23
  • @kilocode/kilo-telemetry@7.3.23

7.3.21

Minor Changes

• #10298 ac7e46d Thanks @Githubguy132010! - Add a kilo profile command for checking the active Kilo account or team balance.

• #10310 c265fa4 Thanks @IamCoder18! - Show running spinner in subagent footer to indicate when subagent is processing

Patch Changes

• #10191 b590f8c Thanks @IamCoder18! - Handle newlines in DialogAlert messages

• #10306 aca8aeb Thanks @IamCoder18! - Toggle export dialog checkboxes on mouse click

7.3.20

Patch Changes

• #10792 cb1fdb3 - Allow clearing agent model and variant overrides from settings.

• #10786 7dd8aab - Limit inferred background-process port discovery to the TUI and stop scanning after startup to avoid unnecessary Bun subprocess polling.

• #10735 593903f - Fix Claude Opus 4.8 reasoning on Amazon Bedrock by treating it as an adaptive thinking model like Opus 4.7. This resolves the "thinking.type.enabled is not supported for this model" error and exposes the full low/medium/high/xhigh/max reasoning effort range.

• #10789 316a662 - Fix queued plan prompts stalling in VS Code after a completed turn.

• #9499 c1c3af8 Thanks @truffle-dev! - Fix empty TUI session list when launching kilo from inside a git submodule. git worktree list --porcelain reports the submodule's gitdir (<repo>/.git/modules/<sub>) instead of the working tree, so the worktree-family filter dropped every session whose directory was the actual submodule path. Include Instance.worktree in the returned set so submodule sessions stay in scope.

7.3.18

Patch Changes

• #10736 57bc6ee - Use Kilo session share links when sharing conversations from the CLI.

• #10737 f574294 - Support reading text from DOCX files through the read tool.

• #10740 2081af2 - Support reading XLSX spreadsheets as labelled tabular text

7.3.17

Patch Changes

• #10721 2efa216 - Keep Agent Manager turns running while slow snapshot baselines initialize instead of stopping for an interactive question.

• #10703 eeff6d9 Thanks @barzhomi! - Fix LanceDB metadata corruption that caused a full re-index on every VS Code restart

• #10733 4967c22 - Read Jupyter notebooks as ordered markdown and code cell content instead of raw notebook payloads.

• #10669 0107a01 - Guide Agent Manager orchestration to recall completed session context only when needed.

• #10668 ef2390d - Access semantic indexing without an experimental feature toggle while keeping indexing disabled until enabled globally or for a project.

7.3.16

7.3.15

7.3.14

Patch Changes

• #8761 74e01b1 Thanks @brendandebeasi! - Fix packaged CLI startup crashes caused by duplicate OpenTUI/Solid renderer instances.

• #10648 9fbd547 - Keep the extension responsive while semantic indexing processes large workspaces.

• #10619 117691e - Use supported hosted model presets for Kilo indexing and clear obsolete model and dimension overrides.

• #10657 d883ad9 - Keep the extension usable on fresh startup when semantic indexing is enabled globally.

• #10618 dcfadac - Prevent saved global indexing provider changes from temporarily reverting in active workspaces.

• Updated dependencies [117691e, db38888]:

  • @kilocode/kilo-indexing@7.3.13
  • @kilocode/sdk@7.3.13
  • @kilocode/kilo-gateway@7.4.0
  • @kilocode/plugin@7.3.13
  • @kilocode/kilo-telemetry@7.3.13

7.3.11

Patch Changes

• #10485 7025c77 - Surface failed sub-agent tasks as tool errors so parent sessions can recover.

• #10443 8e76807 - Support configuring the default task subagent model and reasoning effort while safely inheriting the calling agent model when the override is unavailable.

7.3.10

Patch Changes

• #10302 8ba138d Thanks @IamCoder18! - Export all messages from TUI instead of truncated store

7.3.9

Minor Changes

• #10500 4ef3717 - Support xAI Grok OAuth and device-code login for SuperGrok users.

Patch Changes

• #10510 c076058 - Use the fallback logo in old Windows terminal emulators while keeping the Unicode logo available over SSH.

• #9951 0d12909 - Support optional review focus for /local-review and /local-review-uncommitted, optional base selection for /local-review, and focus both prompts on high-confidence security, performance, business logic, deploy safety, duplication, and dead-code findings.

• #10510 656572c - Avoid leaving mouse and advanced keyboard modes enabled after exiting the TUI in mintty and MINGW terminals.

7.3.8

Patch Changes

• #8403 42844e5 Thanks @saschabuehrle! - Accept env as an alias for environment in local MCP server configuration. Configurations using the more common env key (matching Docker, npm, and VS Code conventions) are now normalised on load instead of failing strict validation.

• #10495 ae0fbe8 - Show recent and favorited models in provider-specific model lists.

7.3.7

Patch Changes

• #10297 74e8604 Thanks @IamCoder18! - Vertically center TUI dialogs on screen

7.3.5

Patch Changes

• Updated dependencies [205e22e]:
  • @kilocode/kilo-indexing@7.3.5

7.3.3

Patch Changes

• #10155 371b7e8 - Resolve bundled tree-sitter WASM resources from the installed CLI layout so codebase indexing works in packaged CLI and VS Code builds.

7.3.2

7.3.1

Patch Changes

• #10285 d23e162 - Capture aggregate usage telemetry for experimental Morph-backed codebase search.

• #10358 413222f - Resume interrupted CLI turns automatically after network recovery while giving users 10 seconds to cancel.

• #10293 af115af - Harden Mermaid diagram rendering with upstream security fixes.

7.3.0

Patch Changes

• #10279 a3769d8 - Keep Enhance Prompt focused on rewriting draft prompts instead of answering question-shaped drafts directly.

7.2.54

Minor Changes

• #10218 4860e65 - Support setting an auto-compaction threshold percentage so long sessions can compact before the context window is full.

Patch Changes

• #10136 8af638e - Show ChatGPT sign-in again when Codex authentication expires.

• #8754 e498c02 Thanks @shssoichiro! - Fix TUI diff rendering when header-like content lines appear inside a unified diff hunk.

• #10158 d8245a0 - Fix Mermaid diagrams rendering with empty text inside every shape by restoring the foreignObject HTML integration point that DOMPurify dropped in 3.1.7.

• #10197 1ea86fb - Prevent Kilo Gateway Responses requests from replaying transient provider item IDs when request storage is disabled.

• Updated dependencies [4860e65, 1af7973, 1ea86fb, f5dc95b]:

  • @kilocode/sdk@7.3.0
  • @kilocode/kilo-gateway@7.3.0
  • @kilocode/plugin@7.2.53
  • @kilocode/kilo-indexing@7.2.53
  • @kilocode/kilo-telemetry@7.2.53

7.2.51

Patch Changes

• #10121 9963b02 Thanks @shssoichiro! - Auto-approve Task subagent tool permissions when running kilo run --auto.

• #10114 0676243 Thanks @shssoichiro! - Remove --dangerously-skip-permissions CLI flag which did nothing

• #10137 33a233f - Prevent subagents from spawning nested subagents.

• #10142 00313bf Thanks @truffle-dev! - Clarify that semantic search returns matching code snippets with paths, line ranges, and relevance scores.

7.2.50

7.2.49

Patch Changes

• #10076 c48b31c - Fix garbled diff and additions/deletions counts shown by apply_patch when updating a non-UTF-8 file.

• #10077 1cf0943 - Speed up reading large files: the read tool now streams UTF-8 content from disk and stops once the line/byte cap is reached, instead of loading the whole file into memory first.

7.2.48

Patch Changes

• #10051 2d50e1f - Harden git operations against malicious repositories and environment variables by upgrading the underlying git library.

• #10050 f1ae973 - Suggest local code reviews after more completed changes while still avoiding small edits and repeated suggestions.

• #10060 0cc0415 - Harden markdown rendering against malicious HTML by picking up the latest DOMPurify security fixes.

• Updated dependencies [924f034]:

  • @kilocode/kilo-indexing@7.2.48

7.2.47

Minor Changes

• #9851 9de7c98 - Support Kilo-hosted embeddings as a selectable code indexing provider.

Patch Changes

• #10016 d2ae16a - Support configuring Azure OpenAI resource names or endpoint URLs from the provider settings flow, and document using the native Azure provider for GPT-5 family deployments.

• #10014 4b88379 - Improved accuracy of Kilo Gateway cost reporting.

• #10012 0363006 - Recover compaction when large tool results or media attachments exceed provider payload limits.

• #9969 eb77fbc - Prevent an infinite agent loop when a provider ends the response stream without a terminal stop reason.

7.2.44

Minor Changes

• #9764 9886674 - Migrate KiloClaw chat to the new kilo-chat backend. Replaces the single-channel Stream Chat integration with a multi-conversation experience that matches the web UX at app.kilo.ai/claw/kilo-chat: conversation list, reactions, typing indicators, editing, and action approvals. The TUI continues to render a single chat view backed by the user's primary conversation.

• #9718 dcaccf3 - Rate assistant responses with thumbs up/down. Click the thumbs buttons next to the copy button on any assistant message, or press <leader>= / <leader>- in the terminal UI. Only shown when telemetry is enabled; feedback is sent to Kilo to help improve model and prompt quality.

Patch Changes

• #9915 bcb47be - Preserve the selected thinking level after compacting a session.

• #9997 de9f11e - Fix gpt-5 models failing with Unsupported parameter: max_tokens when accessed through custom OpenAI-compatible providers such as LiteLLM.

• #9993 98f5f65 - Support global and per-project codebase indexing enablement.

• #9975 c1ea810 Thanks @shssoichiro! - Honor configured permission overrides in Ask and Plan modes, including persisted always-allow rules.

• #10006 9e17137 - Recover sessions when providers end a response with an error finish but no error details.

• #9921 e5e9d0b - Remove custom providers from settings when disconnecting them so they do not reappear after being disabled and re-enabled.

• Updated dependencies [9886674, e5e9d0b]:

  • @kilocode/kilo-gateway@7.3.0
  • @kilocode/sdk@7.3.0
  • @kilocode/kilo-indexing@7.2.43
  • @kilocode/kilo-telemetry@7.2.43
  • @kilocode/plugin@7.2.43

7.2.42

Minor Changes

• #9909 9ffd047 - Detect and preserve UTF-32 (LE and BE) with BOM when reading and editing files. UTF-16 and UTF-32 without a BOM remain unsupported.

Patch Changes

• #9887 d9453f0 - Fix queued-turn auto-compaction so overflow recovery runs instead of exhausting compaction attempts.

• #9855 59e8eff - Respect project-specific semantic indexing decisions instead of enabling indexing globally across workspaces.

• #9928 520922f Thanks @shssoichiro! - Prevent VS Code empty windows from starting codebase indexing against the home directory.

• #9843 27d14d4 - Prompt before reading .env files even after broad read permissions were previously approved.

• #9924 914bbdf Thanks @shssoichiro! - Restore Skill tool access for Plan, Ask, Explore, and other non-system agents so skill workflows are available by default.

• #9907 d9d4dcd - Recover sessions left unable to continue after an assistant turn was created but never started.

7.2.39

Patch Changes

• #9840 db26be6 - Restore the KILO=1 environment variable so plugins and tooling can distinguish the Kilo CLI from upstream OpenCode.

7.2.36

Patch Changes

• #9869 d5fd42c - Fix question tool being unavailable in code mode

• #9838 f499257 - Honor approved external directory read access in Ask and Plan modes.

• #9778 33476e5 - Show an "Initializing snapshot…" line in the chat while the initial snapshot is running on very large repositories, and add an interactive prompt when it stalls. After 10 seconds (configurable via KILO_SNAPSHOT_TRACK_TIMEOUT_MS) the prompt asks whether to keep waiting or disable snapshots for the project; choosing to disable writes "snapshot": false to .kilo/kilo.json so future sessions skip snapshots entirely.

• #9833 614bca7 - Prevent macOS Spotlight from indexing Kilo-generated data directories.

7.2.35

Patch Changes

• #9820 a858f00 - Warn when a model hits its output limit before finishing a response.

• #8910 8472f90 Thanks @eolbrych! - Restore the Sign in action for MCP servers that require OAuth authentication in VS Code settings.

7.2.33

Minor Changes

• #9737 d5fb9eb - Support starting Agent Manager local sessions and worktree sessions from an experimental agent tool.

Patch Changes

• #9746 80535d4 - Avoid repeated command approval prompts when multiple sessions request the same saved command permission, without widening bash permission matching.

• #9460 26e4c11 - Scope the custom commit message prompt to the current project. Setting it in the VS Code settings now writes to the workspace's kilo.json so different repositories can have different conventions, instead of silently applying globally. Also fixes the project-level config update endpoint, which previously wrote to a file that wasn't loaded.

• #9626 5dbf91c - Honor allowed read-only external-directory access to Kilo config paths without repeated permission prompts.

• #9745 da3d79a - Use a GPT-5.5-specific coding prompt that improves autonomous task handling while keeping older Codex generations on their existing prompt.

• #9729 1493d65 - Keep Remote status visible in the TUI while remote control is connecting.

• #9669 0bf14eb - Stop emitting ai.* and gen_ai.* OpenTelemetry spans from AI SDK calls, and remove the PostHog bridge that forwarded them. Tool/session/indexing telemetry is unchanged.

7.2.31

Patch Changes

• #9687 9028174 - Show compact todo update cards when checking off items in long todo lists.

7.2.30

Patch Changes

• #9625 1e01ac3 - Respect configured agent models when reopening the CLI or switching projects.

• #9434 a995b94 - Fix sessions with large image attachments becoming unusable after compaction. When a conversation includes big inline images, the outgoing request can exceed the gateway's body-size limit even after a successful summary. The CLI now trims pre-summary messages for all successful summaries (including manual /compact) and strips media attachments from older turns once a summary exists, so follow-up prompts stay under the gateway limit and the session keeps working.

• #9450 2032fe4 - Fix a session hang that could occur when multiple Kilo panels showed the same permission prompt, or when a subagent's permission was replied to from the wrong worktree. Replies are now routed to the exact CLI instance that holds the pending permission, and stale/unknown permissions surface a clear error so the UI doesn't leave buttons permanently disabled.

• #9635 cbe5510 - Rename the published Docker image from ghcr.io/kilo-org/kilo to ghcr.io/kilo-org/kilocode so it lives alongside the active kilocode repo instead of the archived kilo one.

• #9628 6130a3e - Show paid Kilo models to signed-out users so selecting one prompts them to log in.

• #9556 eae081a - Prevent Ask and Plan modes, including saved or allow-all approvals, from editing files before an explicit implementation step.

• #9615 0907c6f - Keep interactive tools available when semantic indexing fails to load.

• #9603 4145e48 - Improve Windows worktree cleanup reliability when file handles are released slowly.

• Updated dependencies [28a0eae, 6130a3e]:

  • @kilocode/kilo-indexing@7.1.4
  • @kilocode/kilo-gateway@7.2.27
  • @kilocode/kilo-telemetry@7.2.27

7.2.26

Patch Changes

• #9549 a5bca01 - Prefer ChatGPT OAuth credentials over inherited OpenAI environment variables and make ChatGPT sign-in easier to find.

• #9448 73ab363 - Fix session cost display missing subagent costs. The TUI footer, sidebar, web context panel, and ACP usage reports now include the cost of every subagent the session spawned, including nested ones.

• #9484 dbf1135 - Prompt before agents access files outside the active directory when a workspace boundary resolves to a filesystem root.

7.2.25

Patch Changes

• #9526 c8113f2 - Fix multi-turn DeepSeek reasoning round-tripping on OpenRouter by bumping @openrouter/ai-sdk-provider to 2.8.1 in both the CLI and Kilo Gateway packages and letting the SDK handle reasoning details, plus pulling in upstream DeepSeek variant, reasoning-effort, and assistant-reasoning fixes. New DeepSeek conversations are fixed; existing sessions that already stored empty reasoning metadata may still need to be restarted.

• Updated dependencies [c8113f2]:

  • @kilocode/kilo-gateway@7.2.25
  • @kilocode/kilo-telemetry@7.2.25

7.2.23

Minor Changes

• #9418 12c2d86 - Show the open GitHub PR for the current branch in the session sidebar.

Patch Changes

• #9470 7fe4508 - Fix multi-turn tool calls with DeepSeek thinking mode by preserving empty reasoning_content in the interleaved transform.

7.2.22

Patch Changes

• #9455 567ca0d - Fix a 1-2 second startup delay before home content (agents, news, tips) appears in the TUI.

• #9425 6ee160f - Preserve typed text in the main prompt when a blocking question, suggestion, permission, or network overlay is shown and then dismissed.

7.2.21

Minor Changes

• #8587 010a946 - The agent now detects and preserves the original text encoding of files when reading and editing them, so non-UTF-8 files are displayed correctly to the model and written back in their original encoding. New files are still created as UTF-8 without BOM — detection only applies when overwriting or editing an existing file.

  Supported: UTF-8 (with or without BOM), UTF-16 with BOM, and common legacy Latin and CJK encodings (Shift_JIS, EUC-JP, GB2312, Big5, EUC-KR, Windows-1251, KOI8-R, ISO-8859, and others).

  Not supported: UTF-16 without BOM, UTF-32.

Patch Changes

• #9298 8d06a08 - CLI suggestions now render inline in the conversation at the position of the suggest tool call, instead of as a separate bar above the prompt input. The inline bar renders as a single full-width row with a subtle background and clickable action buttons, matching the VS Code extension. Dismissal happens automatically when you send a new prompt. Blocking suggestions still use the above-prompt overlay.

• #9298 2ba203b - CLI suggestions now render above an active input prompt. You can keep typing and submit a new message while a suggestion is on screen — sending a message auto-dismisses the pending suggestion, matching the VS Code extension behavior. The redundant "Dismiss" row has been removed; click an option to accept, or press Esc to dismiss.

• #9344 c032fc2 - Fix an infinite "busy" loop that could occur when a model kept reporting context overflow after every compaction. Each turn now caps compactions at three attempts and closes the turn with a visible context-overflow error instead of silently looping forever.

• #9408 c214d63 - Narrow when the CLI suggests a local code review so it no longer surfaces after PR-comment replies, reactive fixes (CI/lint failures, reported issues), trivial edits, non-implementation work (research, commits, docs), or review-adjacent turns.

7.2.19

Patch Changes

• Updated dependencies [3b73cf4]:
  • @kilocode/kilo-gateway@7.2.19
  • @kilocode/kilo-telemetry@7.2.19

7.2.18

Patch Changes

• #9300 0d0dabe - Fix the "Start new session" button on the plan follow-up prompt not switching the VS Code Agent Manager to the new session when handover generation is slow. The new session now opens immediately, shows the plan text right away, stays visibly busy while the handover summary is being prepared, and appends that summary once it finishes generating.

7.2.17

Patch Changes

• #9276 e6310c5 - Add Alibaba to kiloProviderOptions so thinking is enabled correctly when routing through the Kilo gateway with ai_sdk_provider: "alibaba".

• #9120 d40fc1c - Make the description parameter of the bash tool optional.

• #9239 2b17a7b - Fix custom provider model and variant deletions being silently reverted on save. Removing a model or reasoning variant from a custom provider now actually removes it from your config.

• #9193 f025e34 - Clarify suggest tool guidance so the assistant writes its final summary before offering a local review.

• #9164 448dba8 - Update @ai-sdk/anthropic to 3.0.71, adding xhigh effort for Opus 4.7 adaptive thinking (3.0.70) and fixing fine-grained tool streaming beta header for Opus 4.7 (3.0.71)

• #9170 297b988 - Update @ai-sdk/amazon-bedrock to 4.0.96 and @ai-sdk/google-vertex to 4.0.112, both of which include Opus 4.7 support with xhigh adaptive thinking effort

• Updated dependencies [8b90eec, 448dba8]:

  • @kilocode/kilo-gateway@7.3.0
  • @kilocode/kilo-telemetry@7.2.15

7.2.14

Patch Changes

• #9118 343455b - Respect per-agent model selections when an agent has a model configured in kilo.jsonc. Switching the model for such an agent now sticks across agent switches and CLI restarts. To pick up a newly edited agent default, re-select the model once (or clear ~/.local/share/kilo/storage/model.json).

• #9067 959a8b4 - Fix "assistant prefill" errors when a user queues a prompt while the previous turn is still streaming. The queued message no longer lands in the middle of the prior turn's history, so the next request always ends with the user prompt.

• #9023 5301258 - Support mentioning folders in the prompt with @ references, including top-level folder file contents.

7.2.12

Patch Changes

• #9068 e65c2d9 Thanks @kilo-code-bot! - Hide Kilo Gateway models that do not support tool calling from the model list.

• #9069 e60c326 Thanks @kilo-code-bot! - Support adaptive reasoning for Claude Opus 4.7 and expose the xhigh effort level for adaptive Anthropic models

• Updated dependencies [e65c2d9]:

  • @kilocode/kilo-gateway@7.2.12
  • @kilocode/kilo-telemetry@7.2.12

7.2.11

Patch Changes

• #8898 4a69a3e - Fixed default model falling back to the free model after login or org switch by invalidating cached provider state when auth changes.

• #8996 58ff01a Thanks @kilo-code-bot! - Include pnpm-lock.yaml and yarn.lock in the .kilo/.gitignore so lockfiles from alternative package managers don't appear as untracked files

• 4937759 - Merged upstream opencode changes from v1.3.10:

  • Subagent tool calls stay clickable while pending
  • Improved storage migration reliability
  • Better muted text contrast in Catppuccin themes

• 4937759 - Merged upstream opencode changes from v1.3.6:

  • Fixed token usage double-counting for Anthropic and Amazon Bedrock providers
  • Fixed variant dialog search filtering

• 4937759 - Merged upstream opencode changes from v1.3.7:

  • Added first-class PowerShell support on Windows
  • Plugin installs now preserve JSONC comments in configuration files
  • Improved variant modal behavior to be less intrusive

• #9047 bea8878 - Continue queued follow-up prompts after the active session turn finishes.

• Updated dependencies [4d2f553]:

  • @kilocode/kilo-gateway@7.2.11
  • @kilocode/kilo-telemetry@7.2.11