Skip to content

CHANGELOG.md

Latest commit

 

History

History
407 lines (268 loc) · 17.3 KB

CHANGELOG.md

File metadata and controls

407 lines (268 loc) · 17.3 KB

Changelog

0.18.2

Patch Changes

  • Test release

0.10.1

Patch Changes

  • Test release

0.7.0

Minor Changes

  • 60c8ba8: feat(google): Route OAuth requests through Cloud Code API for proper subscription support

    When Google OAuth is active, the agent now routes requests through Google's Cloud Code API (cloudcode-pa.googleapis.com/v1internal) instead of the standard Generative Language API (generativelanguage.googleapis.com).

    This is the same approach used by the official Gemini CLI and properly supports:

    • Google AI Pro/Ultra subscription users
    • The cloud-platform OAuth scope (which is all the OAuth client has registered)
    • Automatic subscription tier handling (FREE, STANDARD, etc.)

    The key insight is that the Gemini CLI doesn't call generativelanguage.googleapis.com directly. Instead, it uses Google's Cloud Code API which:

    1. Accepts cloud-platform OAuth tokens
    2. Handles subscription tier validation
    3. Proxies requests to the Generative Language API internally

    This fix includes:

    • URL transformation from Generative Language API to Cloud Code API
    • Request body transformation to Cloud Code API format (wrapping with model/project fields)
    • Response body transformation to unwrap Cloud Code API responses
    • Streaming response support with proper SSE chunk transformation
    • Fallback to API key authentication if Cloud Code API fails

    Fixes #100

    References:

0.6.3

Patch Changes

  • 2732471: fix: Resolve 'undefined is not an object (evaluating Log.Default.lazy.info)' error

    The error occurred because Log.Default.lazy was undefined - the Logger interface doesn't have a .lazy property. The logging methods (info, debug, warn, error) already support lazy evaluation through callback functions.

    Changes:

    • Fixed Log.Default.lazy.info() calls in src/index.js to use Log.Default.info() directly
    • Added backward-compatible lazy property to Logger (self-reference) for any external code using this pattern
    • Added unit tests for lazy logging callback support

0.6.2

Patch Changes

  • c2b3343: fix: Google OAuth 403 restricted_client error from unregistered scopes

    When using agent auth login with Google OAuth, users encountered a 403: restricted_client error with the message "Unregistered scope(s) in the request".

    Root cause: The generative-language.tuning and generative-language.retriever OAuth scopes were not registered for the Gemini CLI OAuth client (681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j).

    This fix removes the unregistered scopes and aligns with the official Gemini CLI implementation, which only uses:

    • https://www.googleapis.com/auth/cloud-platform
    • https://www.googleapis.com/auth/userinfo.email
    • https://www.googleapis.com/auth/userinfo.profile

    This fix resolves issue #93 ("agent auth login через Google не работает").

0.6.1

Patch Changes

  • 52692ba: fix: Google OAuth authentication race condition

    Previously, when running agent auth login with Google OAuth, the authentication would fail with a redirect to localhost:0 because of a race condition in port assignment. The redirect URI was built before the server port was actually assigned.

    Root cause: server.listen(0, callback) assigns the port asynchronously in the callback, but the redirect URI was built before the callback fired.

    Now, when authenticating with Google:

    • Port is discovered and assigned before building the redirect URI
    • Supports GOOGLE_OAUTH_CALLBACK_PORT environment variable for containers
    • Follows Google's recommended localhost redirect flow for installed applications

    This fix resolves issue #78 ("Не работает авторизация в Google").

0.6.0

Minor Changes

  • fe1fff2: feat: Implement JSON lazy logging with log-lazy library

    • Added log-lazy dependency from link-foundation for lazy evaluation pattern
    • Created src/util/log-lazy.ts module for standalone lazy logging with JSON output
    • Updated src/util/log.ts to support lazy logging with lazy.{debug,info,warn,error} methods
    • JSON output wraps all logs in { log: { ... } } structure for easy parsing
    • Logs are written to stderr to avoid mixing with stdout JSON data
    • Lazy evaluation prevents expensive string formatting and object serialization when logging is disabled
    • Converted existing log calls across the codebase to use lazy pattern for improved performance:
      • Session management files (agent.js, compaction.ts, processor.ts, prompt.ts, revert.ts, summary.ts)
      • Authentication (claude-oauth.ts, plugins.ts)
      • Provider and model handling (provider.ts, models.ts)
      • Configuration loading (config.ts)
      • File operations (watcher.ts, ripgrep.ts, time.ts)
      • MCP client handling (mcp/index.ts)
      • Server routing (server.ts)
      • Formatting, patching, and project state management
    • Added comprehensive tests in tests/log-lazy.test.js

    Closes #81

0.5.3

Patch Changes

  • 7386896: feat: Add echo and cache synthetic providers for robust dry-run mode

    Implements link-assistant/echo provider that echoes back user input, and link-assistant/cache provider with Links Notation format for caching API responses. Echo provider is automatically enabled in --dry-run mode for zero-cost testing of round-trips and multi-turn conversations.

    Closes #89

0.5.2

Patch Changes

  • c51e70b: docs: Add interactive mode multi-turn conversation example

    Add comprehensive example in docs/stdin-mode.md showing how interactive terminal mode works with multiple sequential inputs.

    The example demonstrates:

    • Initial status message when entering interactive mode
    • Complete JSON event stream for two user inputs ("hi" and "who are you?")
    • Session persistence across multiple messages
    • Plain text input auto-conversion to JSON
    • Streaming events (step_start, text, step_finish)
    • Token usage tracking in responses

    Fixes #86

0.5.1

Patch Changes

  • 1b83cef: fix: Enable interactive terminal mode when stdin is TTY

    Previously, when running agent in an interactive terminal (TTY), the CLI would immediately show help and exit, not accepting any input. This was a regression from issue #76 fix that treated all TTY connections as "no input expected".

    Now, when running agent in a terminal:

    • Outputs status message indicating interactive terminal mode
    • Accepts typed input from the user
    • Processes messages in continuous mode with session persistence
    • Respects --verbose and other flags

    To get the old behavior (show help and exit), use --no-always-accept-stdin.

    Fixes #84

0.5.0

Minor Changes

  • 5a09d94: feat: Enable continuous listening mode by default
    • Add --always-accept-stdin flag (default: true) for continuous input mode
    • Add --compact-json flag for machine-to-machine communication
    • Pretty-print status messages to stderr by default
    • Keep session alive between messages for multi-turn conversations
    • Handle SIGINT gracefully for clean shutdown

0.4.0

Minor Changes

  • f6f0756: feat: Improve CLI stdin handling with new options and user feedback

    • Add -p/--prompt flag to send prompts directly without stdin
    • Add --disable-stdin flag to explicitly disable stdin mode
    • Add --stdin-stream-timeout for optional timeout on stdin reading
    • Add --dry-run flag for simulating operations
    • Output JSON status message when entering stdin listening mode
    • Include CTRL+C exit hint and --help guidance
    • Show help immediately when running in interactive terminal (TTY)
    • Remove default timeout on stdin reading (wait indefinitely by default)

    This improves user experience by:

    1. Never hanging silently - always provide feedback
    2. Supporting multiple input modes (stdin, direct prompt, help)
    3. Following CLI best practices from clig.dev guidelines

0.3.1

Patch Changes

  • 5680f2e: fix: Add retry logic and serialized installation for reliable provider initialization

    Fixes issue #72 where version 0.3.0 appeared "completely broken" due to race conditions in parallel package installations causing Bun cache corruption.

    Root Cause:

    • When multiple provider packages (e.g., @ai-sdk/openai-compatible, @ai-sdk/openai) are installed concurrently, they can cause race conditions in Bun's package cache
    • This leads to "FileNotFound: failed copying files from cache" errors on first run after update

    Changes:

    • Add write lock to serialize package installations (prevents concurrent bun add commands)
    • Add retry logic with up to 3 attempts for cache-related errors
    • Improve error detection to catch ENOENT, EACCES, EBUSY errors
    • Add delay between retries to allow filesystem operations to complete

    Impact:

    • opencode/grok-code remains the default provider and works reliably
    • Agent handles transient cache issues gracefully with automatic retries
    • Better stability during first run after installation/update

0.3.0

Minor Changes

  • eb44ccf: Add support for link-assistant-agent branding and improve default model selection

    New Features:

    • Default model is now opencode/grok-code (Open Code Zen / Grok Code Fast 1)
    • Support for LINK_ASSISTANT_AGENT_* environment variable prefix (backwards compatible with OPENCODE_*)
    • Support for .link-assistant-agent config directory (backwards compatible with .opencode)
    • Global config paths now use link-assistant-agent name

    Breaking Changes:

    • Global config directory changed from ~/.config/opencode/ to ~/.config/link-assistant-agent/
      • Users should manually migrate global configs if needed

    Migration:

    • Environment variables: New LINK_ASSISTANT_AGENT_* prefix available; old OPENCODE_* still works
    • Project config: Both .opencode/ and .link-assistant-agent/ directories supported
    • Model override: Use "model": "provider/model-id" in config to override default

0.2.1

Patch Changes

  • d71920e: feat: Add --dry-run mode and enhanced debugging logging

    • Added --dry-run CLI flag to simulate operations without making actual API calls or package installations
    • Added OPENCODE_DRY_RUN flag and Flag.setDryRun() function for programmatic control
    • Enhanced error logging in BunProc.install() with detailed error messages including stderr/stdout
    • Enhanced provider initialization logging with detailed context
    • Added logging for package installation success/failure
    • Added unit tests for dry-run mode functionality
    • Improved debugging capabilities to help diagnose package installation failures

    Fixes #68

0.2.0

Minor Changes

  • 4f5f2a6: Add Google OAuth support for Gemini Pro/Ultra subscriptions
    • Implements OAuth 2.0 with PKCE for Google AI subscription authentication
    • Uses same public OAuth credentials as official Gemini CLI (installed app flow)
    • Adds local HTTP server to automatically capture OAuth redirect
    • Supports token refresh with 5-minute expiry buffer
    • Zeros out cost tracking for subscription users
    • Creates case study documentation in docs/case-studies/issue-66/
    • Fixes #66

0.1.4

Patch Changes

  • dd0dfdc: Fix empty system message handling for Anthropic OAuth credentials

    When using Claude Code OAuth credentials with an empty --system-message "", the agent now preserves the required "You are Claude Code" header message. This prevents:

    • cache_control cannot be set for empty text blocks errors
    • This credential is only authorized for use with Claude Code errors

    The fix ensures OAuth token authorization works correctly even when users explicitly set an empty system message.

    Fixes #62

0.1.3

Patch Changes

  • 50728b3: fix: Fixed agent --version command and added logging in --verbose mode
    • Fixed --version command that was showing "unknown" instead of the current package version
    • Added explicit import of package.json using createRequire with fallback via fs
    • Added logging of version, command, working directory and script path in --verbose mode

0.1.2

Patch Changes

  • de95398: fix: Pass API key to providers with multiple env var options

    Fixes #61 - Error when using google/gemini-3-pro model. When providers have multiple possible environment variables (like Google with GOOGLE_GENERATIVE_AI_API_KEY and GEMINI_API_KEY), the code was finding the API key correctly but then not passing it to mergeProvider.

0.1.1

Patch Changes

  • 09b6709: Fix GitHub release formatting to remove incorrect title for major/minor/patch versions and properly link related pull requests

0.1.0

Minor Changes

  • 2bcef5f: Add support for google/gemini-3-pro model alias

    • Added google/gemini-3-pro as an alias to gemini-3-pro-preview
    • Updated README.md with Google Gemini usage examples
    • Created comprehensive case study in docs/case-studies/issue-53/
    • Fixes ProviderModelNotFoundError when using google/gemini-3-pro

    This change allows users to use the commonly expected model name gemini-3-pro while maintaining compatibility with Google's official gemini-3-pro-preview identifier.

Patch Changes

  • 86f24ac: Add comprehensive image validation to prevent API errors
    • Added magic byte validation for PNG, JPEG/JPG, GIF, BMP, WebP, TIFF, SVG, ICO, and AVIF formats
    • Prevents "Could not process image" API errors from invalid files
    • Added VERIFY_IMAGES_AT_READ_TOOL environment variable for opt-out (enabled by default)
    • Enhanced error messages with hex dump debugging information
    • Comprehensive test suite with 6+ test cases
    • Fixes #38 and prevents session crashes from corrupted image files

0.0.17

Patch Changes

  • 8d3c137: Fix empty string system message override. When --system-message "" is provided, the system now correctly overrides with an empty string instead of falling back to the default system prompt. This was caused by a falsy check (if (input.system)) that evaluated to false for empty strings. Changed to explicit undefined check (if (input.system !== undefined)) to properly distinguish between undefined (use default) and empty string (override with empty).

0.0.16

Patch Changes

  • 1ace1a9: Add comprehensive Playwright MCP installation instructions to README.md. This includes step-by-step setup guide, complete list of available browser automation tools, usage examples, and links to official documentation. Fixes issue #54.

0.0.15

Patch Changes

  • 3a5fad7: Fix debug output appearing in CLI commands - logs are now suppressed by default and only shown with --verbose flag. This fixes the issue where commands like agent auth list displayed debug messages that broke the clean CLI UI.

0.0.14

Patch Changes

  • 07aef38: fix: Use valid placeholder API keys for OAuth providers (Anthropic, GitHub Copilot, OpenAI)

    When using --model anthropic/claude-sonnet-4-5 after authenticating with agent auth login (Anthropic > Claude Pro/Max), the command failed with ProviderInitError at line 732 in src/provider/provider.ts.

    The OAuth authentication plugin loaders in src/auth/plugins.ts were returning empty strings ('') for the apiKey parameter. The AI SDK providers (e.g., @ai-sdk/anthropic, @ai-sdk/openai) require a non-empty apiKey parameter even when using custom fetch functions for authentication. The empty string failed validation, causing provider initialization to fail.

    This fix changes the apiKey value from an empty string to a descriptive placeholder string ('oauth-token-used-via-custom-fetch') for all OAuth loaders (AnthropicPlugin, GitHubCopilotPlugin, and OpenAIPlugin). The placeholder satisfies AI SDK validation requirements while the actual OAuth authentication happens via Bearer tokens in the custom fetch's Authorization header.

    Fixes #47

0.0.13

Patch Changes

  • 4b4c0ea: Fix ESC key handling in CLI commands - pressing ESC now exits gracefully without showing error messages or help text

0.0.12

Patch Changes

  • c022218: fix: Replace prompts.autocomplete with prompts.select in auth login command

    The agent auth login command was failing with TypeError: prompts.autocomplete is not a function because @clack/prompts@0.11.0 does not have the autocomplete function (it was added in v1.0.0-alpha.0).

    This fix replaces prompts.autocomplete() with prompts.select() which is available in the stable version.

    Fixes #43

0.0.11

Patch Changes

  • e8e2d03: Fix system message override to use exclusively without additional context for low-limit models. When --system-message flag is provided, the system now returns only that message without adding environment info, custom instructions, or headers. This reduces token usage from ~9,059 to ~30 tokens, enabling support for models with low TPM limits like groq/qwen/qwen3-32b (6K TPM).

    Also adds --verbose mode to debug API requests with system prompt content and token estimates.

0.0.10

Patch Changes

  • 547d73f: Add CI/CD pipeline matching js-ai-driven-development-pipeline-template with changeset support, automated npm publishing, and code quality tools (ESLint, Prettier, Husky)

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.