v0.preview.72

v0.preview.72 - 2026-08-09

Added

• Onboarding "Use API key" now takes a pasted key directly: masked input with provider auto-detect from the key prefix, validated against the provider, and stored owner-only under ~/.ante/auth (env vars still take precedence)
• On the API-key step with Anthropic selected, Tab signs in to the Anthropic Console in the browser and provisions an API key automatically — the key never touches the clipboard
• Compaction results are now visible: a collapsed * Compacted marker appears in the conversation (manual and auto compaction), with the full summary in the ctrl+o transcript view; CompactEnd now carries the summary text on the wire, and trimmed oversized tool results are reported with an info line

Changed

• Improved Bash command execution
• Antix OAuth is listed ahead of OpenAI in /connect and provider auto-detection

Fixed

• A fresh install with no credentials no longer lands in a session silently wired to the unreachable built-in localhost:8080 fallback — the not-connected state points at /connect, and a successful sign-in restarts the session on the newly connected provider
• Standalone ante update no longer panics with a broken pipe when its launcher closes stdout, and a closed pipe can no longer cancel an update

v0.preview.71

v0.preview.71 - 2026-08-07

Added

• ante doctor checks that the Ante home directory (~/.ante) is writable

Changed

• Dependency updates

Fixed

• Cancelling a running Bash tool call now kills its whole process group, so grandchild processes no longer keep running after Ante reports execution stopped
• Streaming model calls that produce no usable output within 5 minutes fail with a timeout instead of hanging the turn indefinitely
• The installer refuses sudo installs that would leave ~/.ante root-owned, and startup shows a visible notice when the Ante home is unwritable
• Proxy auth_unavailable credential failures are classified as terminal auth errors instead of consuming the reconnect budget
• Parameterless tool schemas keep an explicit empty properties object on OpenAI-compatible requests, fixing rejections from strict endpoints

v0.preview.70

v0.preview.70 - 2026-08-03

Added

• --profile <name>: named settings profiles as whole-file replacement settings files (<name>.settings.json), plus the ANTE_PROFILE env var and a built-in file-less bare profile
• Bare /term opens a terminal picker — list, attach/detach, kill, and create sessions

Changed

• Updates now ride the normal TUI flow: when a new version is available an in-chat notice shows its release notes, and /update installs it on exit (/update skip dismisses that version, /update cancel unschedules); the blocking startup modal is gone, and running sessions now check for updates hourly
• Use DeepSeek V4 Flash 0731 on OpenRouter
• Nightly builds log at debug level by default
• SKILL.md reads are labeled as skill loads in the transcript

Fixed

• ante catalog exits cleanly instead of panicking when its output pipe closes early
• Brighter agent response text; thinking text pinned to foreground gray
• Hidden hardware cursor parked at the composer so OS IME popups appear at the caret
• Z.ai five-hour usage-limit errors (code 1308) are classified as terminal quota instead of being retried

v0.preview.69

v0.preview.69 - 2026-08-02

Added

• Shift+Enter and Ctrl+Enter insert a newline in the composer on terminals with kitty keyboard protocol support (kitty, ghostty, WezTerm, iTerm2 3.5+, Alacritty 0.13+)

Changed

• Diff View V2 across all diff surfaces — chat Edit/Write details, the /diff pager, the approval Tab preview, and the theme dialog: numbered line gutter, full-width added/removed background bands, and +N -M change counts inline in headers

Fixed

• Normalize MCP tool schemas to the portable provider subset at ingestion, so servers using advanced JSON Schema keywords (exclusiveMinimum, $ref, oneOf, …) no longer get requests rejected by strict providers such as Gemini
• Merge system prompt fragments into a single leading system message for unrecognized OpenAI-compatible models, fixing strict local chat templates that error on multiple system messages
• Spinner token count now shows only tokens generated this turn (it previously re-added the full context every generation, inflating wildly), with a ↑/↓ request/stream phase arrow
• Gemini quota errors that mention token limits are classified as quota, not context overflow, so they no longer trigger futile compaction

v0.preview.68

v0.preview.68 - 2026-07-31

Added

• /term <name> [args...] launches the named binary in the fresh session — /term claude --continue opens a split already running claude; existing sessions still just attach
• terminals status-line chip (on by default) showing live ante-* tmux sessions by name
• --no-skills run flag: skip skill discovery entirely (nothing advertised in the system prompt or dispatchable as commands); /resume preserves the choice

Changed

• Repeated /term calls stack viewer splits in one right-hand column, replacing the composed ante-view-* viewer — every agent keeps a full-width column and Ante keeps its size
• Two-row footer: identity items (model, dir, branch, …) on top, permission mode and activity chips beneath
• Default DeepSeek V4 Flash to max reasoning effort on all routes (Antix, direct, OpenRouter)
• Dependency updates

Fixed

• Fully typed commands that take arguments (e.g. bare /term) no longer swallow the first Enter

v0.preview.67

v0.preview.67 - 2026-07-29

Added

• /term <a> <b> composes a side-by-side viewer over several sessions in one window — watch or compare two terminal agents at once
• Bundled tmux skill: session reuse, interactive-program input, and driving other CLI agents such as Claude Code or Codex
• Discover skills from user-level ~/.claude/skills, and highlight recognized commands in the composer
• @-mention completion for ~/, absolute, ../, and ./ paths
• Delete sessions from the /resume picker

Changed

• Replace /pty with /term: the agent drives durable named tmux sessions through ordinary Bash (namespaced ante-*), and /term <name> opens a native terminal split/window to watch or type; sessions survive Ante restarts
• No subprocesses or file writes before the TUI's first frame (faster, quieter startup)
• Bump the bundled llama.cpp engine to b10107

Fixed

• Redact the OAuth callback query from logs
• Parse model weight_class values case-insensitively

v0.preview.66

v0.preview.66 - 2026-07-26

Added

• Per-category context-window usage reporting over the protocol (Op::ContextReport)

Changed

• Speed up Grep further by parallelizing the directory walk and search (large-corpus searches 40–70% faster)
• Update and prune the built-in model catalogs

Fixed

• Don't abort OAuth login when stdin closes

v0.preview.65

v0.preview.65 - 2026-07-25

Added

• Add a /pty window that attaches a native terminal side panel
• Render LaTeX math in the TUI via a delimiter normalizer and TeX→Unicode converter
• Add an auto compact setting

Changed

• --tools now sets the base tool set and --include-tools adds on top of it
• Serialize concurrent Edit/Write calls that target the same file instead of only warning
• Speed up Grep by dropping per-file parallel fan-out (~2× faster)
• Use catalog keys as provider IDs
• Cap skill catalog descriptions rendered into the system prompt
• Extend short-prompt coverage to the system prompt and more tools
• Harden release publication identity and atomicity

v0.preview.64

v0.preview.64 - 2026-07-24

Added

• Add an Antix auth login command

Changed

• Wrap tool header lines at the viewport edge instead of clipping them
• Dependency updates

Fixed

• Detect when iTerm2 blocks scrollback clearing: purge by default when verified safe, show a notice when blocked
• Exit the TUI cleanly on broken stdout pipes
• Omit absent optional arguments from tool call displays

v0.preview.63

v0.preview.63 - 2026-07-21

Added

• Add Tencent Hy3 to the OpenRouter model catalog and pricing table
• Include step count in turn end events

Changed

• Markdown render improved by 93% for long streams
• Remove Qwen 3.7 Max/Plus from the OpenRouter catalog (endpoints no longer accessible); they remain available via Antix
• Render markdown lists directly instead of via tui-markdown
• Inset all TUI rows two columns from the window edge
• Consolidate the ante-guide into a bundled, model-only skill
• Warn when concurrent Edit/Write mutations target the same file path
• Prune older built-in model catalog entries

Fixed

• Reject tool calls with malformed arguments before execution and return the parse error so the model can retry
• Handle tool calls truncated at the output-token limit gracefully across all providers, continuing the turn instead of surfacing a hard error
• Drop malformed ambient replies instead of leaking model reasoning