0.142.2

• /context now shows where auto-compaction triggers: a 🔲 marker on the threshold cell of the grid plus an Auto-compaction at N% line.
• Include autoCompactPercentage in the usage chat content so clients can mark the auto-compaction threshold on their context-usage bar.
• Bugfix: kill-job! commits :killed before destroying the process, fixing a race where the exit monitor overwrote it with :failed (exit 143). (#510)
• Bugfix: enabling trust mode while a subagent is running now applies to it; subagents (including nested) inherit the parent chat's trust. (#504)

0.142.1

• Bugfix: postRequest hook no longer fires for subagent chats; subagents use subagentPostRequest only, preventing duplicate hook notifications. (#505)
• Bugfix: clearer rejected tool-call result wording so models no longer assume a rejected edit was applied; it now states the call did not run and changed nothing. (#507)
• Bugfix: ask_user normalizes the options argument so a malformed value (e.g. a stringified options an LLM sometimes emits) no longer reaches clients as broken choices. Accepts string/object arrays, recovers a JSON-encoded string, and drops unusable input.

0.142.0

• Add /context command showing context-window usage with a fixed 10x10 proportional emoji grid to the left of a two-section legend: Instructions (system prompt, rules, skills, AGENTS.md, tool definitions) and Chat (tool calls, conversation), plus free space. Each category has a canonical distinct color/emoji swatch.
• Include contextBreakdown in the usage chat content so clients can render a live context-usage bar. Each category carries a canonical color and emoji (plus freeColor/freeEmoji for the unused portion) so all clients render the same colors/legend.

141.0

• Add message pagination over a shared cursor core: new JSON-RPC chat/history method and optional limit/before/after window params on chat/open, plus opt-in pagination on remote HTTP GET /api/v1/chats/:id. Opaque cursors with a lastCompaction sentinel; meta exposes before/after/compaction cursors.
• Support declaring image input for custom models via a per-model imageInput config flag; openai-chat now round-trips tool-result images. (#503)
• Bedrock: honor Anthropic thinking variants (effort + adaptive thinking) for Claude models, fixing Claude 4.7+ failing with thinking.type.enabled not supported. (#502)

0.141.1

• OpenAI OAuth (ChatGPT/Codex) models now report Codex context windows instead of inflated direct-API limits (e.g. gpt-5.5 272k, not 1.05M).
• Agent defaultModel (and explicit model ids) resolve alias-first: a bare model id is matched against the currently selected provider's models before being treated as a full id, enabling provider-relative subagent models. (#501)

0.140.1

• MCP: recover remote servers with stale connections (e.g. after suspend): tool-call timeouts now probe and re-initialize; add mcpKeepAliveSeconds pings; 404 triggers re-init.
• Update github-copilot default model to gpt-5.5.

0.140.0

• Add preCompact, postCompact and subagentStart hooks; subagents no longer trigger chatStart.
• Add /hooks command with optional description field in hook config.
• Tool hooks (preToolCall/postToolCall) now include tool_call_id in input data.
• Expand hook contracts: response not prompt, plain-text tool_response, continue:false everywhere, followUp, replacedOutput, standalone systemMessage, exact-string matchers.
• Fix postToolCall continue:false leaking across turns, chatStart additionalContext dropped, and preRequest exit-2 naming the blocking hook.
• Add Claude Fable 5 and Mythos 5 support.

0.139.2

• MCP OAuth: add authScope (global default / workspace / custom bucket) to namespace stored tokens, so signing into different accounts per project no longer clobbers a shared server's token.
• OpenAI: inline completion no longer fails with a NullPointerException on provider/auth errors (e.g. OAuth login); the underlying error is surfaced instead. (#495)
• Surface a clear error when a prompt exceeds the model's context window instead of silently finishing; skip auto-compaction when there's no conversation to compact. (#491)
• Resolve home via $HOME/$USERPROFILE when user.home is invalid (e.g. native image, email-style username), fixing broken cache paths and login persistence. (#490)

0.139.1

• Model fetch: skip providers without configured credentials, include the response body in 4xx/5xx logs, and quiet down per-provider logging.
• Model fetch: renew near-expiry OAuth tokens before syncing catalogs, fixing spurious token expired 401s (e.g. GitHub Copilot) on startup and config reloads.
• GitHub Copilot: present the expected VS Code editor identity (Editor-Version/Editor-Plugin-Version/Copilot-Integration-Id) on model-catalog and chat requests, fixing 400 ... Editor-Version header for IDE auth.

0.139.0

• Remote mode: ask_user now reaches both the editor and connected SSE/web clients simultaneously, with the first answer from either winning, instead of only the web client when one is connected.
• Remote REST API: GET /api/v1/chats/:id pendingToolCalls now includes ask_user tool calls while waiting for an answer, with a requestId field for POST /api/v1/answer.
• Add OpenAI Responses API for GitHub Copilot models that require it (gpt-5.5, gpt-5.4-mini).
• MCP OAuth: persist and reuse the dynamically-registered client on token refresh, so servers with non-idempotent DCR (e.g. RunLayer) refresh instead of forcing a browser re-login, and recover from expired-token tool errors automatically.
• Deliver cursor context per-turn in the user message instead of the system prompt, and only re-send it when the position changes, so a moving cursor no longer invalidates the cached system prefix (avoiding llama.cpp full prompt re-processing). (#464)
• Remote endpoint now lists resumed and forked chats immediately, without waiting for the first prompt.
• Fix resumed chats being permanently missing from the remote GET /session and GET /chats endpoints.