1jehuang profile

Account failover, hibernation, and faster reloads

Highlights

• Automatically retry turns on the next pool account when one hits a usage limit, including for Desktop and SDK clients
• Hibernate idle sessions to free memory, with measured RAM status and transparent wake on the next interaction
• Reload and resume large sessions much faster by showing the local transcript immediately and sending only new history

Improvements

• Support OpenAI Ultrafast speed tiers for GPT-6.1 Sol, with Fast commands and hotkeys that apply promptly
• Show saved accounts, cached quotas, and available resets in the startup account router and usage cards
• Polish usage cards with responsive gauges and compact Mil and Bil token totals
• Offer additional OAuth accounts in the login picker and reauthenticate saved accounts from it
• Show cache misses inline with more accurate cache reuse diagnostics
• Show Jcode inference spending balance
• Compress OpenAI and Anthropic request bodies to reduce upload time
• Build the /model picker off the UI thread for large remote catalogs
• Reconnect only changed MCP servers instead of restarting all of them
• Simplify swarm mode by removing the task graph and defaulting swarm root effort to low

Fixes

• Keep startup headers and account rows on a single line at narrow widths
• Show usage-limit rejections clearly instead of stale quota bars
• Fix forked sessions failing on replayed Anthropic web search blocks
• Count Gemini thinking and tool-use tokens correctly in usage
• Keep streamed OpenAI output when the completion payload arrives empty
• Preserve token usage when requests are cancelled or fail
• Fix Windows named-pipe connections hanging or failing on startup probes
• Fix Firefox cold starts blocked by stale zombie processes
• Collapse expanded pinned todos and restore overall progress display

Full changelog: https://github.com/1jehuang/jcode/compare/v0.98.1...v0.99.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Windows executables are unsigned. Azure Artifact Signing is not configured and the repository's existing WINDOWS_SIGNING_REQUIRED=false override allowed unsigned publication. Windows may show publisher or SmartScreen warnings.

Faster idle sessions and richer responses

Highlights

• Reduce idle CPU and disk reads in repositories with many untracked files
• Share model catalog processing across equivalent clients while keeping usage and account state fresh
• Allow Desktop-capable clients to receive Markdown responses through the new user_response tool

Improvements

• Show compact context usage before and after compaction
• Open the resume picker immediately while session scanning continues in the background
• Disable automatic memory recall while retaining explicit memory search
• Display large recorded token and cost summaries compactly

Fixes

• Recover browser handoffs from runtime provider failures using direct browser actions
• Restore Windows companion connections over named pipes
• Expose global user-session connection presence to SDK clients
• Keep worker emoji placement consistent in the swarm status strip

Full changelog: https://github.com/1jehuang/jcode/compare/v0.97.0...v0.98.1

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Windows signature status

Windows x86_64 and ARM64 executables are unsigned. Azure Artifact Signing is not configured and the repository's existing WINDOWS_SIGNING_REQUIRED=false override allowed unsigned publication. Windows may show publisher or SmartScreen warnings. SHA-256 checksums are provided for download integrity.

Highlights

• Use custom HTTP JSON search providers with configurable authentication, result mapping, and fallback engines
• Rust and TypeScript SDK clients can invoke native runtime commands and transfer sessions without turning commands into agent prompts

Improvements

• Dismiss side panels directly from the TUI, and keep routine system notices on a compact single line
• Default opt-out usage telemetry can now include the active Anthropic or OpenAI OAuth account email when the matching server rollout is available, with a conditional one-time TUI notice, 90-day retention, and collection controls in /telemetry

Fixes

• Azure agent turns use the Responses API with correct route selection, explicit reasoning settings, and more reliable tool-call and retry handling
• Anthropic model catalog requests include the required authentication headers
• Browser handoff failures preserve reconnect timeout diagnostics
• Empty math no longer triggers LaTeX rendering, and repeated fallback failures avoid duplicate logging

Full changelog: https://github.com/1jehuang/jcode/compare/v0.96.0...v0.97.0

Release source recovery

Built from existing tag v0.97.0 with reviewed windows-listener repair from workflow commit f8c8e6c5db60aea1c40c317175c56759bb437bef. The tag and version are unchanged. This is patched tag source, not the exact tagged tree. Release-only installer tooling is pinned to the same workflow commit. See release-provenance.json for source hashes and build run.

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Highlights

• Connect to remote MCP servers over HTTP or SSE, including OAuth sign-in and command-line management
• Choose saved OAuth accounts per session from the model and login pickers, with email labels and cancelable account-switch countdowns
• SDK clients can organize sessions into collections and attach hidden reminders to individual turns

Improvements

• Swarm messages use compact, expandable one-row previews, while the activity dock shows session icons and each worker's current todo step
• Pinned todos span the full terminal width
• The /reset usage limits claude command supports resetting Claude session limits
• Prompts and pokes appear in every window attached to a session
• New-model usage costs receive estimates before published pricing becomes available

Fixes

• The first pinned todo update no longer hides the original session prompt, and hidden todo updates preserve transcript cache boundaries
• OpenAI native web citations render source links reliably without losing links already shown
• Provider failover honors account pools and model preferences, rotates rejected OAuth accounts, and stops retrying permanently rejected sessions
• Stale Anthropic usage is marked unavailable instead of showing misleading limits
• Browser control identifies the correct browser bridge and explains silent Chrome extension connections
• Remote MCP handshakes are bounded so unresponsive servers do not hold up startup indefinitely
• Copy-selection Enter and Y shortcuts work in client mode, and Ctrl+L keeps the fresh view when new messages arrive
• Escaped short prose lists no longer render as math, and system-notice emoji no longer leave ghost terminal cells
• SDK-launched sessions no longer flash console windows inside Windows GUI applications

Full changelog: https://github.com/1jehuang/jcode/compare/v0.95.0...v0.96.0

Release source recovery

Built from existing tag v0.96.0 with reviewed windows-listener repair from workflow commit 2f84206d98bba7b0e17bd784100703ed6aee5fc1. The tag and version are unchanged. This is patched tag source, not the exact tagged tree. Release-only installer tooling is pinned to the same workflow commit. See release-provenance.json for source hashes and build run.

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Highlights

• OpenCode Go models, including grok-4.7 and gpt-6-luna, now route over the Responses API
• Hooks expose turn-end token usage, and turn_start/turn_end now fire for one-shot jcode run turns
• /agents supports per-session worker model overrides

Improvements

• /resume can open scoped to the current project when display.resume_project_scope is enabled
• Swarm gains a stalled-worker watchdog, and headless run --socket sessions can join swarms
• The macOS menu bar helper uses a stable signing identity, draws the streaming icon green, and can be turned off with [menubar] enabled = false
• Inline images fit their reserved box and the info widget hugs its content
• Command risk assessment now inspects commands inside $(...), backticks, and heredocs
• SDK-launched private instances no longer auto-update mid-session and give shorter, more accurate connect errors

Fixes

• Ctrl/Alt/Cmd shortcuts work on the Korean 2-Set input source
• The terminal keyboard protocol and TUI modes are restored when jcode exits on a signal
• Rate-limit and provider reconnect auto-resumes are bounded instead of retrying indefinitely
• OpenAI streams recover faster from stalls and OAuth token refresh no longer hangs
• Tool calls emitted as DSML/invoke markup text are recovered instead of lost
• Kimi reasoning via OpenRouter no longer loses text from non-cumulative snapshots
• Resumed sessions heal a stale credential route so swarm workers inherit the live one
• Computer-use screenshots kept in transcripts are bounded in size
• The browser tool reuses persistent sessions on macOS without opening repeated windows
• Only one menu bar helper runs when jcode is started with an overridden HOME
• Swarm member attribution and restart notices are correct

Full changelog: https://github.com/1jehuang/jcode/compare/v0.94.0...v0.95.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Queued commands

Highlights

• Ctrl+Enter queues slash and ! commands to run after the current turn, and /clear drops anything still queued
• Alt+G opens the project's git commit log, and diff cycling moves to /diff
• Harness API clients can observe and cancel messages sent mid-turn

Improvements

• The startup frame settles immediately instead of popping in over about 200ms, with session facts shown on the first frame
• The Overview widget stays visible after startup
• Onboarding is cleaner: side-by-side start choice, no starter prompt cards, and the home directory is labeled '~ (home)'
• Onboarding detects accounts you are signed in with in Cursor, other AI tools, and stored Jcode agent logins
• Restored transcripts and the session picker preview no longer show a redundant 'tool:' summary line

Fixes

• Escaped single-word brackets like [workspace] are no longer rendered as display math
• Onboarding's Import option no longer commits when pressing Enter, and toggling preview no longer flickers the screen
• Sleep inhibition keeps its lock across refreshes and retries after helper failures

Full changelog: https://github.com/1jehuang/jcode/compare/v0.93.0...v0.94.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Instant startup

Highlights

• New sessions paint their first frame in about 20ms instead of about 135ms, and the screen no longer shifts while connecting
• Opening several sessions in quick succession is fast again, and resuming large sessions is about three times quicker
• New post_tool_feedback hook feeds a script's output back to the model after each tool call, and post_tool hooks now receive the tool input, call id, and touched paths

Improvements

• The jcode: model prefix and JCODE_OPENROUTER_MODEL are honored, and jcode run fails clearly when --model cannot be applied
• Onboarding predicts your default provider and model from existing logins
• Expired status notices clear on their own without waiting for a keypress
• Old build binaries and stale scratch files are pruned automatically
• Web search hints no longer recommend the retired Bing Search API
• jcode builds on FreeBSD

Fixes

• New sessions no longer show "Waiting for reload handoff" right after a reload
• --provider-profile works with Anthropic-compatible profiles
• Copy selections that end inside a wide character include that character
• The warning gate fails closed on cargo check errors and prints the diagnostics

Full changelog: https://github.com/1jehuang/jcode/compare/v0.92.0...v0.93.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Google Calendar and speed tiers

Highlights

• Google login lets you pick which services to connect and adds a built-in calendar tool for viewing, creating, updating, and deleting events
• Guided Google OAuth app setup that detects existing gcloud credentials, validates generated project IDs, and imports the app automatically
• Alt+Up/Down cycles the speed tier between Standard, Fast, and Ultrafast
• New /desktop command opens the current session in Jcode Desktop

Improvements

• MCP tool results and mcp_search output are fitted to token budgets instead of fixed caps, and accept_large_output returns the full result
• The mcp list action lists only tool names, and mcp_search ranks name matches first
• Todo quality gates share rows instead of taking one line each

Fixes

• Browser actions target the right browser when several run the bridge extension
• Interrupting a coordinator stops the swarm workers it spawned (thanks @SiavZ)
• A per-call swarm spawn_mode of auto defers to your configured mode (thanks @SiavZ)
• /restart and /reload from a remote client no longer wipe the server's transcript (thanks @SiavZ)
• The shared server reloads after an install repoints the link it started from (thanks @SiavZ)
• run --json and --ndjson report the local or direct profile instead of openrouter
• Fixed swarm plan and status handlers that could stall the server

Full changelog: https://github.com/1jehuang/jcode/compare/v0.90.1...v0.91.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Provider reliability and fixes

Highlights

• OpenAI-compatible providers that answer 529 or another 5xx "heavy usage" error are now retried automatically instead of failing the turn, and the error no longer blames your network
• /usage shows real quota windows for Kimi Code, Cursor, and the Z.ai Coding Plan (thanks @ghoker143)
• Pressing Esc while a follow-up prompt is pending stops the current turn and runs the follow-up (thanks @SiavZ)

Improvements

• Copilot accepts reasoning effort for every model whose catalog advertises it and sends it on both the Responses and Chat Completions APIs
• Saved API keys stay in their env files instead of the running process, so editing a key file takes effect without a restart
• The client shows the context window the server resolved instead of a placeholder
• The jcode log records which binary actually launched, to diagnose stale copies on PATH
• Gmail can update, list, and delete drafts
• Undo in the composer reverts a whole typing burst at once, including Korean and other IME input (thanks @mg-mg-mg)
• Cost estimates for OpenAI-compatible profiles fall back to a matching model from another provider's price catalog (thanks @ogodevonline)
• Edit tool rows show their intent, and the expand badge appears in every diff mode

Fixes

• DeepSeek enabled through --provider auto uses its own default model and 1M context instead of anthropic/claude-sonnet-4 with 200K
• In-app updates on Termux patch the glibc ELF interpreter like the installer does, and refuse to switch to a binary that cannot run
• Configured /colors roles now apply on 256-color terminals
• Images in formats providers reject are no longer stored or sent, and sessions self-heal from the resulting 400 errors
• Session saves are drained on SIGTERM, and an interrupted checkpoint can no longer duplicate history
• Anthropic history repair: orphaned or late tool results and orphaned native web searches no longer cause request errors
• Malformed tool calls are corrected within a bounded number of retries instead of looping
• Auto-poke re-arms when new open work appears after the completion gate trips
• /accounting is no longer parsed as /account
• The crash detector no longer reports a session that closed while idle as interrupted (thanks @SK-DEV-AI)

Full changelog: https://github.com/1jehuang/jcode/compare/v0.90.0...v0.90.1

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Provider-native web search

Highlights

• Web search now uses your model provider's own server-side search (Anthropic and OpenAI) by default instead of scraping from your machine, shown as regular web search rows live and on resume (thanks @johnnyfish)

Improvements

• Set websearch.prefer_native = false or JCODE_WEBSEARCH_PREFER_NATIVE=0 to keep using local search engines
• Provider-side search follows each session's allowed and disabled tool policy
• SDK apps that register their own websearch tool keep it instead of having it replaced by provider search
• Installing a skill mid-session no longer invalidates the prompt cache, so the next turn stays fast and cheap
• When switching providers, earlier native search results are summarized clearly so the model does not retract correct answers

Fixes

• OpenAI login validation no longer hangs when the account hits a usage limit
• Retrying a turn no longer leaves web search rows from the aborted attempt
• Interrupted Anthropic turns that used native search no longer risk a request error on the next message

Full changelog: https://github.com/1jehuang/jcode/compare/v0.89.3...v0.90.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Herdr support and /cloud session moves

Highlights

• jcode is now a first-class agent in herdr: its name and working/idle state show in the sidebar and herdr agent list, herdr agent wait works on it, and after a herdr restart the pane reopens the same jcode session
• /cloud and /local move a live session, including uncommitted code, to a cloud host and back

Improvements

• herdr reports run in the background with a short timeout, so herdr can never slow jcode down, and jcode releases the pane when you quit

Fixes

• Usage telemetry now counts tokens from responses that fail mid-stream, crashed sessions, compaction, and background agents

Full changelog: https://github.com/1jehuang/jcode/compare/v0.89.2...v0.89.3

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Cursor MCP tool calls complete on every model

Highlights

• Cursor MCP tool calls now complete on Composer, Grok 4.7, Grok 4.6, and Claude models instead of hanging after the tool is requested

Fixes

• Switching a session to Cursor with /model no longer leaves MCP tool calls stuck: tool results are sent back to Cursor
• jcode answers Cursor's MCP server state and tool allowlist checks, so the model's tool call actually reaches jcode
• Cursor model names resolve against Cursor's live model list, so Grok 4.6, Claude Opus 5/5.5, and GPT-5.6 Sol no longer fail with 'Model name is not valid'

Full changelog: https://github.com/1jehuang/jcode/compare/v0.89.1...v0.89.2

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Cursor provider fixes, Commits widget, and a dedicated diagram pane key

Highlights

• The Cursor provider works again: jcode tools run over Cursor's native MCP bridge, fast/high/thinking model ids resolve correctly, /fast works, and reasoning models no longer stop mid-answer
• A new Commits info widget shows recent commits with line counts and marks unpushed ones
• The swarm dock lists each agent with its status, current activity, and todo progress

Improvements

• Alt+Shift+M shows or hides the diagram pane, and Alt+M now only cycles the side panel (split, fullscreen, hidden)
• Info widgets put their headers, overflow counts, and legends on their borders
• The foot terminal is detected as sixel-capable for Mermaid diagrams
• The harness API can move a running tool to the background
• agentgrep 0.1.7 keeps partial results when ripgrep hits unreadable files

Fixes

• Anthropic no longer rejects turns with parallel tool_search calls (400 'tool_use ids were found without tool_result blocks')
• Images no longer render blank in Kitty inside Zellij
• Test binaries in custom Cargo target directories never open real browser windows
• Several flaky tests are now isolated from the developer's environment

Full changelog: https://github.com/1jehuang/jcode/compare/v0.89.0...v0.89.1

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Built-in voice input, agent applets, and a pinned session status line

Highlights

• Built-in voice input for the CLI, with native microphone capture on macOS and Windows and coding-aware dictation vocabulary
• The session status line (directory, branch, git status, context, provider, and model) stays pinned below the input, with info widgets showing the detail behind it
• Agents can show interactive applets, and swarms can message each other across swarms with unique labels

Improvements

• The Changes widget lists dirty files newest first with line counts and marks the files the agent edited
• MCP tools load through provider-native deferred loading, so adding or using MCP servers no longer busts the prompt cache
• Accounts have a persistent auto-switch pool, and failover rotates through it in your chosen order
• Show an upgrade card and daily feature allowances when a Jcode plan limit is reached
• The daemon detects KV cache misses and reports them to every client
• Untitled sessions are named after their first prompt, and save labels name the session
• Keep your reading position across terminal resizes
• Support embedded and multiple slash commands in one message
• The first Ctrl+C clears the input and the second quits
• Add pre-tool input transformers
• Add the Yolo-Auto provider, grok-4.7 and grok-4.6 on Cursor, and OpenRouter image support from the model catalog
• Block lid-close sleep on macOS and Windows while sessions are working
• Rust and TypeScript SDKs support per-connection SSH identities and pinned host keys

Fixes

• Antigravity consumer accounts no longer hit 429s from the wrong endpoint
• Claude usage limits are shared across processes and keep the last good values when throttled
• MCP servers no longer stall startup or spawn duplicates, and pending requests fail when a stdio server exits
• Reload exits with an error when the new server never becomes ready
• Switching models validates against the target provider, and configured default models with a provider prefix match correctly
• The effort chip follows model changes, and the alternate-enter hint respects queue mode and keymap conflicts
• Ctrl+Up into prompt history keeps your draft
• Hooks never flash a console window on Windows
• DeepSeek Flash and MiMo v2.6 models keep their full context window

Full changelog: https://github.com/1jehuang/jcode/compare/v0.88.0...v0.89.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Claude Opus 5.5 default, multi-browser support, and banked usage resets

Highlights

• Claude Opus 5.5 at medium effort is now the default Anthropic model, and anthropic-api: model specs use the native Anthropic runtime
• The browser tool detects your default browser and supports Chrome, Edge, Brave, Chromium, and Safari in addition to Firefox
• Reset exhausted OpenAI usage with banked resets through a confirmed command, with a hint shown when your account has one available

Improvements

• The edit tool now applies multiple replacements atomically, replacing the separate multiedit tool
• Saved sessions always appear in the session list even when they are older than the recency limit, and /save bookmarks persist through later writes
• Voice input starts recording immediately without waiting for the Nari connection, keeps your first words, and recognizes Jcode product vocabulary
• Unclear voice requests are routed to the coding agent instead of being dropped as uncertain
• Importing credentials from other tools never overrides logins Jcode already has
• Add OrcaRouter as a provider choice for jcode login and --provider
• Parallel tool calls stream their names and inputs as soon as they arrive
• Rust and TypeScript SDK clients can set per-session system prompts, enforce tool policies, and receive structured reasons for abnormal turn stops
• The auth badge updates immediately when switching between OAuth and API-key routes
• Add /merge-remote-release to merge the current branch and trigger a CI release in one step

Fixes

• Saving settings from the CLI no longer erases Jcode Desktop settings in the shared config file
• Visible swarm agents keep their chosen model, provider, and effort when their window attaches
• Proxy tool arguments are preserved until the stream completes
• Interleaved streaming tool calls from OpenAI stay matched to the correct call
• Todo quality checks stay quiet after the final response, and the usage reset review flow shows credit expirations
• Long TLDR swarm summaries are accepted, and null bash notification flags fall back to defaults

Full changelog: https://github.com/1jehuang/jcode/compare/v0.87.1...v0.88.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Always-on adaptive status line

Highlights

• The status line below the input is now always visible by default and adapts to the terminal width, keeping the directory, model, and context usage in view even on narrow terminals

Improvements

• Status line is right-aligned with the model at the right edge, and model and provider text is muted to keep the line quiet
• Status line shows git status counts next to the branch (modified, staged, untracked, ahead, behind)
• On narrow terminals facts compact step by step instead of disappearing, and the full token counts and context bar return when there is room
• Set display.overscroll_status = "overscroll" to keep the previous reveal-on-overscroll behavior

Full changelog: https://github.com/1jehuang/jcode/compare/v0.87.0...v0.87.1

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Claude Opus 5.5 and independent Anthropic model discovery

Highlights

• Use Claude Opus 5.5 directly through Anthropic, with explicit API-key and Claude-login routes in the model picker
• API-key and OAuth model catalogs refresh independently, so one account or authentication method cannot hide models available on the other
• Update Claude OAuth client compatibility to Claude Code 2.1.280 for the latest model release

Improvements

• Support Opus 5.5 mandatory adaptive thinking, progress summaries, and documented conversation-bound thinking controls
• Keep the last usable catalog during refresh failures and separate cached API-key catalogs by credential fingerprint
• Show distinct missing API key and missing Claude login labels rather than hiding the direct API route
• Add Opus 5.5 context, output-limit, reasoning-effort and pricing metadata without changing saved model selections

Fixes

• API-key long-context availability no longer depends on Claude subscription extra-usage settings
• Server replacement-model recommendations preserve decimal versions such as Opus 4.8 and Opus 5.5
• Share the bundled Anthropic model list between provider layers to prevent discovery fallback drift

Full changelog: https://github.com/1jehuang/jcode/compare/v0.86.0...v0.87.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Smarter task handling and clearer session feedback

Highlights

• Jev browser handoff now handles whole multi-step tasks, continuing through uncertain pages and returning to the parent only when finished or blocked
• Memory recall uses Jev to select relevant project and global memories without loading a local embedding model at startup
• Save Anthropic prompt-cache retention preferences with /cache extend and see clearer cache usage, retention estimates, and provider-specific limits

Improvements

• Use /merge to integrate the current branch into the repository primary branch with validation of the combined result
• Terminal tab titles show session edit-line counts and elapsed time for the current or most recent work period
• Jev features prefer the included subscription route when the gateway and account support it, instead of automatically charging a separate API key
• Adds gated remote-compilation client support on Linux and macOS, requiring a provisioned service and cloud-compute credits, with access checked before source uploads
• Native voice APIs add bounded microphone capture and Nari streaming for clients with a Nari API key, alongside clearer handling of requests to use the newest conversation
• Rust and TypeScript clients can distinguish assistant messages, final answers, and replacement text from retries
• Independent SDK clients can share an authenticated SSH connection without one client closing the others
• The pinned diff side panel has been removed while inline diffs remain available

Fixes

• OpenAI WebSocket conversations recover when earlier tool outputs change or a continuation reports a missing tool result
• Cache-write tokens are retained in usage reporting and OpenAI cost estimates no longer double-count them
• Estimated cache retention no longer triggers misleading expiry countdowns or warnings for inactive provider routes
• Drag-to-copy follows the visible text on wrapped lines, including narrow terminal layouts
• Todo quality checks avoid unnecessary ownership and confidence follow-ups for completed work
• Edit results preserve file-position diffs and distinguish a known unchanged file from an unavailable snapshot
• Fresh-session history stays available during active turns and socket messages survive interrupted reads
• Scheduled tasks reach attached sessions, and local client skills load before the initial history
• MCP calls preserve dynamic arguments, and completed non-image file drops are normalized correctly
• Source checkouts without an upstream no longer perform misleading update checks, and empty API session directories are handled gracefully

Full changelog: https://github.com/1jehuang/jcode/compare/v0.85.0...v0.86.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Configurable swarm effort and richer client workflows

Highlights

• Configure root-model reasoning effort independently for swarm and deep swarm, with the selected effort shown in the effort switcher
• Delegate multi-step browser tasks to a bounded Jev browser session that returns control to the parent agent
• Create and manage separate panels from Markdown or linked PDFs, with bounded PDF payloads and safe Markdown fallback for older TUI clients

Improvements

• Model pickers remember route usage and use shared usage-aware ranking across clients
• Gemini adds project and OAuth-route configuration, and context compaction can be capped below a model's full window
• Native SDK clients gain cancellable OAuth login, local Git worktree operations, per-turn response and edit statistics, swarm lifecycle metadata, and side-panel snapshots
• Light terminal themes have clearer contrast while preserving custom colors, with better-aligned resume previews and Markdown blockquote gutters
• OpenAI tool details begin streaming earlier, and DeepSeek Flash supports image input
• Desktop source checkouts receive dedicated self-development tools and instructions
• The initiative command is temporarily unavailable, while existing initiative data is retained

Fixes

• Active turns survive client reconnects, and restored image attachments retain their positions and batching
• Grok login and model switching preserve the selected authentication route
• OAuth-backed Bash calls honor their configured timeouts
• Ambient configuration changes take effect without restarting the session
• Development builds are protected against unintended backward updates
• Gemini retains tool thought signatures and retries transient Code Assist failures without confusing provider sessions with Jcode sessions
• MCP tools with colliding normalized names remain independently callable without losing their tool policies
• Compatible tmux sessions request extended keys so Shift+Enter reaches the input editor
• Project and global prompt guidance no longer appears twice when it resolves to the same source
• Logs, history and usage displays identify the active provider profile rather than an unrelated route

Full changelog: https://github.com/1jehuang/jcode/compare/v0.84.0...v0.85.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available

Clearer remote login onboarding

Highlights

• Connecting to an SSH host now asks a plain yes-or-no question before offering to import a local OpenAI or Claude login

Improvements

• The remote login picker lists the providers the host actually supports, with local imports shown as explicit choices
• Remote login status is sanitized before display, so host credential details are never echoed back

Fixes

• Pasted input while choosing a login stays in the active picker instead of leaking into the prompt
• Remote hosts configured with an unrecognized provider are reported as unknown rather than signed in

Full changelog: https://github.com/1jehuang/jcode/compare/v0.83.0...v0.84.0

Platform availability

• Linux x86_64: available
• Linux aarch64: available
• macOS Apple Silicon: available
• macOS Intel: available
• Windows x86_64: available
• Windows ARM64: available
• FreeBSD x86_64: available