CHANGELOG.md

Changelog

v0.19.15

Features

• New media-generation subcommands aivo video, aivo audio, and aivo speak join the existing aivo image, sharing a services::media_io module for output path parsing, overwrite policy, atomic writes, and error extraction. Hidden from --help for now while the surface stabilizes.

Internal

• CI: scope concurrency cancellation to ci.yml and run the test workflow on main pushes / PRs so the matrix validates before tagging.

v0.19.14

Fixes

• Launcher: skip extension-less PATH entries when looking up tool binaries on Windows — npm drops both claude.cmd and a bash-style claude (no extension) into %APPDATA%\npm; the lookup matched the unspawnable bash shim first. aivo claude / aivo codex / aivo gemini now resolve to the .cmd shim and spawn correctly even when the tool was already installed

(v0.19.12 and v0.19.13 carried the same fix but failed CI — v0.19.12 on a PATHEXT casing mismatch in tests, v0.19.13 on a clippy::needless_return lint that only fires on the newer Rust shipped to windows-latest. No binaries were published for either.)

v0.19.11

Fixes

• Launcher: pin the resolved binary path before spawning so npm .cmd shims (claude.cmd, codex.cmd, etc.) launch on Windows — CreateProcessW does not honor PATHEXT for non-.exe files, so spawning the bare claude/codex name failed after install

v0.19.10

Fixes

• Chat TUI / keys add secret prompt: filter Windows key Release events so typed characters aren't doubled — crossterm emits both Press and Release on Windows for every keystroke; we now only process Press

v0.19.9

Fixes

• Build: statically link the MSVC C runtime for Windows targets so aivo.exe no longer depends on VCRUNTIME140.dll — fixes silent load failure (STATUS_DLL_NOT_FOUND) on Windows ARM64 systems without the Visual C++ Redistributable installed, where aivo --version printed nothing

v0.19.8

Improvements

• Bridge: propagate cached input tokens from OpenAI-shape upstreams — Claude Code now records cache_read_input_tokens correctly so cached usage shows up in aivo stats instead of being silently dropped
• Bridge: coalesce parallel function_calls into one assistant message
• Run: skip model picker on first run when only the starter key exists
• Launcher: fall back to installer drop dirs when PATH lookup misses
• MCP/router: pin nickname session resolution; trust proven route on attempt 0
• Removed cross-tool MCP peer awareness
• Build: add win32-arm64 target

v0.19.7

Improvements

• aivo update and install.sh now fetch binaries from getaivo.dev (R2) instead of GitHub Releases — faster in regions with poor GitHub connectivity
• Fixed text selection and session list in chat TUI
• Bridge: filter Anthropic server-side tools (web_search_*, code_execution_*, etc.)
• Bridge: propagate input_tokens via message_delta on OpenAI streams — fixes Claude Code status-line percent stuck at 0%
• Router: apply learned requires_reasoning_content to in-flight requests, eliminating per-request 400 + retry until relaunch
• MCP: stabilize session discovery and nickname mapping
• Removed Chat TUI thinking toggle and think-tag detection

v0.19.6

Fixes

• Fix handling reasoning_content in deepseek learn the config
• Reorder --help command listing (keys, models, chat, serve, image, stats)

v0.19.5

Improvements

• Extended alias support to launch tool, not only model
• Added Poolside provider preset
• Improved --max-context flag parsing to handle more formats and edge cases
• Force non-streamed upstream for Inception Mercury when tools are in use
• Fixed alias rm <name> parsing

v0.19.4

Improvements

• Fallback: persistent learned protocol/path routing with smarter Responses-API detection
• Bridge: forward image/file parts and dropped sampling params; safer tool_use IDs, stop-reason mapping, and cache_control handling
• Gemini: persist thoughtSignature across restarts; group parallel tool responses into one user turn
• Claude: --1m/--2m flags (and --max-context=1m|2m) append the canonical [Nm] suffix to set max context window
• aivo keys reset-route <name> to clear learned routing; --debug rewritten as JSONL HTTP logger

v0.19.3

Improvements

• aivo stats --since DURATION: time-windowed reports (e.g. --since 7d, --since 24h)
• Claude Code: per-task model overrides for individual slots
• CLI: drop custom short-flag pre-expander, rely on clap POSIX bundling
• Gemini: handle thoughtSignature for gemini-3

v0.19.2

Features

• aivo image: experimental image generation command

Fixes

• Router: try native /v1/messages when target_protocol is anthropic; respect learned PathVariant on fast paths
• Router: prefer CLI-native protocol and force aivo-starter through the router
• Fix Linux ETXTBSY flake in Claude setup-token spawn tests
• aivo chat --json now prints the provider's raw response body instead of aivo's envelope
• aivo run: skip the model picker on non-TTY and under --dry-run
• aivo chat no longer injects max_tokens: 8192 for DeepSeek / aivo-starter; matches curl
• opencode: default to an OpenAI-style model instead of claude-sonnet

v0.19.1

UX

• Improve key pickers and status prints for --as.

Bug Fixes

• Preserve Codex trust settings across launches
• Reject for disable OAuth keys for models and serve instead of error

v0.19.0

Features

• Multi-account OAuth for Codex, Gemini and Claude Code

Bug Fixes

• Align ctx, out, and price columns in aivo models

Refactors

• Drop reserved-name shortcuts in aivo keys; preselect in the picker instead
• Rework Ctrl+C / Ctrl+L handling and drop the /clear command

v0.18.1

• Fix build on windows.

v0.18.0

Features

• Cross-tool MCP communication via --as <name>: Run a tool under a custom identity so peers can query it live
• aivo context and --context injection: Export recent session context as Markdown and inject it into any launched CLI for cross-tool handoffs
• Copilot premium-request multiplier: Surfaced in aivo models so you can see which Copilot models cost more
• Chat TUI keybinding swap: Ctrl+C and Ctrl+L swapped to match other agents

Bug Fixes

• Bypass HTTP proxy for the loopback router when launching CLIs
• Cancel in-flight chat requests when the user exits
• Close tool-calling parity gaps in the serve bridge layer
• Plug correctness gaps across crypto, bridges, and TUI
• Tighten bridge-layer fragility across serve, launch, and cache_control
• Make Windows first-class in PID liveness checks and Pi binary reuse
• Preserve tool_result images through Anthropic-to-OpenAI conversion and the typed OpenAI round-trip
• Extend multimodal preservation through the Responses API and recover sniffed media_type
• Close post-review gaps across serve auth, tool_result images, and the OpenAI round-trip
• Surface unreadable session files in debug builds during --context ingestion
• Align aivo stats Claude totals with Claude Code's /stats UI
• Prevent update download from timing out on slow connections

v0.17.0

Features

• Improved the ux of adding keys: Interactive provider picker backed by the full known-provider catalog
• Added JSON output via --json: Enables scripting and | jq pipelines

Refactors

• Streamline logs command and redesign the status output

v0.16.2

Bug Fixes

• Normalize input_tokens to fresh-only across all stats parsers for consistent accounting
• Exclude subagent sidechain files from Claude session count
• Add Claude Code env vars for timeouts, attribution, and subagent model
• Handle Ollama pull errors instead of silently succeeding
• Refresh PATH from login shell after tool install
• Move --version and --help before service init and improve smoke test diagnostics

v0.16.1

Bug Fixes

• Fix ETXTBSY on Linux during self-update: close write file descriptor before smoke test

v0.16.0

Features

• Google Gemini native API: Direct support for generativelanguage.googleapis.com as a provider across all tools
• Open aivo-starter model list: aivo-starter users can now access the full model catalog
• Enriched aivo models: Show context window, max output tokens, and pricing from the provider API
• Prompt to install missing tools: Interactively offer to install tool binaries when not found on PATH, with cross-platform support
• R2 download mirror: GitHub-first binary downloads with Cloudflare R2 fallback for faster installs and updates

Bug Fixes

• Clear last_selection on key add so the newly added key is actually used
• Resolve aivo-starter sentinel URL in serve router and model picker
• Normalize Pi stats token counting to include cached input tokens
• Reduce GitHub download timeout and deduplicate mirror fallback messages
• Fix mirror fallback for self-update flow

v0.15.0

Features

• aivo-starter: Zero-config provider — start using aivo without any API key setup
• Update rollback: Automatically roll back failed updates; added config migration tests and CI clippy gate
• Local session logging: SQLite-backed aivo logs command for browsing session history
• Native top session view: Opt-in aivo stats --top for a live session overview
• Combined short flags: Support Unix-style combined flags like -xr, -nar
• Ollama lifecycle management: Auto-stop Ollama on exit using PID-file refcount for safe concurrent instances
• DeepSeek reasoning streaming: Stream reasoning_content through routers for DeepSeek-reasoner models
• Conditional default model option: Only show "default model" in the picker when the selected tool supports it

Bug Fixes

• Cap max_tokens for aivo-starter and DeepSeek in chat requests
• Remove last two production unwrap() calls for safer error handling
• Fix device auth for starter provider across all tools
• Hide default model option in chat mode since it requires a concrete model
• Support Responses API-only Copilot models (e.g. gpt-5.4) for Claude and Gemini
• Resolve tilde paths and add PDF/binary support for chat attachments
• Remove tool name from active key display, show only key and model

Performance

• Avoid PBKDF2 decryption when displaying active selection label
• Warm model cache in background after adding API key

Refactors

• Redesign key/model selection: per-directory → global last-selection
• Replace sqlite3 CLI with rusqlite for OpenCode stats reading
• Route OpenCode through router for providers with quirks

v0.14.5

Major update with stats aggregation, better tool support

Improvements

• Global stats aggregation across all AI tools (Claude, Codex, Gemini, OpenCode, Pi).
• Mask API key input with asterisks during entry
• Show install guide when a tool is not found on PATH
• Support Pi tool with Copilot subscription
• Rename ls command to info (keep ls as alias)
• Embed provider registry as JSON with table-driven tests
• Remove redundant token stats recording from run tool
• Bump GitHub Actions to v5 for Node.js 24 compatibility

Fixes

• Remove custom User-Agent headers from API requests
• Use Codex model_provider config to bypass auth.json and OPENAI_BASE_URL deprecation
• Wire --refresh flag through run command for model picker cache bypass
• Auto-strip anthropic-beta headers for Bedrock/Vertex providers

v0.14.4

Stability hardening. Fixed panics from char-boundary slicing and API response handling. Switched Linux builds to musl targets for better portability.

v0.14.3

Added Responses API fallback for models that require the /v1/responses endpoint. Fixed /attach command autocomplete.

v0.14.2

Bug fixes and CI improvements.