- Remove setApiKey, resolveApiKey, and global apiKeys Map from stream.ts
- Rename getApiKey to getApiKeyFromEnv (only checks env vars)
- Remove OAuth storage layer (storage.ts deleted)
- OAuth login/refresh functions now return credentials instead of saving
- getOAuthApiKey/refreshOAuthToken now take credentials as params
- Add test/oauth.ts helper for ai package tests
- Simplify root npm run check (single biome + tsgo pass)
- Remove redundant check scripts from most packages
- Add web-ui and coding-agent examples to biome/tsgo includes
coding-agent still has compile errors - needs refactoring for new API
When a user interrupts a tool call flow (sends a message without providing
tool results), APIs like OpenAI Responses and Anthropic fail because:
- OpenAI requires tool outputs for function calls
- OpenAI requires reasoning items to have their following items
- Anthropic requires non-empty content for error tool results
Instead of filtering out orphaned tool calls (which breaks thinking signatures),
we now insert synthetic empty tool results with isError: true and content
'No result provided'. This preserves the conversation structure and satisfies
all API requirements.
- Fix tool result format for Gemini 3 Flash Preview compatibility
- Use 'output' key for successful results (not 'result')
- Use 'error' key for error results (not 'isError')
- Per Google SDK documentation for FunctionResponse.response
- Improve type safety in google.ts provider
- Add ImageContent import and use proper type guards
- Replace 'as any' casts with proper typing
- Import and use Schema type for tool parameters
- Add proper typing for index deletion in error handler
- Add comprehensive test for Gemini 3 Flash tool calling
- Tests successful tool call and result handling
- Tests error tool result handling
- Verifies fix for issue #213Fixes#213
Previous test used compressed 8k images (0.01MB) which was meaningless.
Now tests with actual large noise images that don't compress.
Realistic payload limits discovered:
- Anthropic: 6 x 3MB = ~18MB total (not 32MB as documented)
- OpenAI: 2 x 15MB = ~30MB total
- Gemini: 10 x 20MB = ~200MB total (very permissive)
- Mistral: 4 x 10MB = ~40MB total
- xAI: 1 x 20MB (strict request size limit)
- Groq: 5 x 5760px images (5 image + pixel limit)
- zAI: 2 x 15MB = ~30MB (50MB request limit)
- OpenRouter: 2 x 5MB = ~10MB total
Also fixed GEMINI_API_KEY env var (was GOOGLE_API_KEY).
Related to #120
Tested max 8kx8k images per provider:
- Anthropic: 100 (explicit limit, fails at 101)
- OpenAI: 100-200 (100 works, 200 times out)
- Mistral: 8 (explicit limit, fails at 9)
- xAI: 100-150 (100 works, 150 times out)
- Groq: 0 (8k exceeds 33M pixel limit)
- zAI: 400 (context window limited at 500)
- OpenRouter: 40 (context window limited at 50)
- Gemini: untested (no API key in test env)
Key finding: Anthropic's 'many images' rule does NOT cause API errors.
100 x 8kx8k images work fine. Anthropic likely auto-resizes internally.
Related to #120
Tests max image count, size, dimensions, and 8k stress test for:
- Anthropic, OpenAI, Gemini, Mistral, OpenRouter, xAI, Groq, zAI
Key finding: Anthropic's 'many images' rule (>20 images = 2000px max)
does NOT cause API errors. 100 x 8k images work fine. Anthropic likely
auto-resizes internally.
Related to #120
Tests provider-specific image limitations across all supported providers:
- Maximum number of images in context
- Maximum image size (bytes)
- Maximum image dimensions
Discovered limits (Dec 2025):
- Anthropic: 100 images, 5MB per image, 8000px max dimension
- OpenAI: 500 images, >=25MB per image
- Gemini: ~2500 images, >=40MB per image
- Mistral: 8 images, ~15MB per image
- OpenRouter: ~40 images (context limited), ~15MB per image
- Allow branch selector to open with single user message (changed <= 1 to === 0 check)
- Support in-memory branching for --no-session mode (no files created)
- Add isEnabled() getter to SessionManager
- Update sessionFile getter to return null when sessions disabled
- Update SessionSwitchEvent types to allow null session files
- Add branching tests for single message and --no-session scenarios
fixes#163
- Skip empty assistant messages (no content, no tool calls) to avoid
Mistral's 'Assistant message must have either content or tool_calls'
error
- Remove synthetic assistant bridge message after tool results (Mistral
no longer requires this as of Dec 2024)
- Add test for empty assistant message handling
Follow-up to #165
- Add Mistral to KnownProvider type and model generation
- Implement Mistral-specific compat handling in openai-completions:
- requiresToolResultName: tool results need name field
- requiresAssistantAfterToolResult: synthetic assistant message between tool/user
- requiresThinkingAsText: thinking blocks as <thinking> text
- requiresMistralToolIds: tool IDs must be exactly 9 alphanumeric chars
- Add MISTRAL_API_KEY environment variable support
- Add Mistral tests across all test files
- Update documentation (README, CHANGELOG) for both ai and coding-agent packages
- Remove client IDs from gemini.md, reference upstream source instead
Closes#165
- Add agentLoopContinue() to pi-ai for resuming from existing context
- Add Agent.continue() method and transport.continue() interface
- Simplify AgentSession compaction to two cases: overflow (auto-retry) and threshold (no retry)
- Remove proactive mid-turn compaction abort
- Merge turn prefix summary into main summary
- Add isCompacting property to AgentSession and RPC state
- Block input during compaction in interactive mode
- Show compaction count on session resume
- Rename RPC.md to rpc.md for consistency
Related to #128
- Add 'xhigh' to ThinkingLevel type in ai and agent packages
- Map xhigh to reasoning_effort: 'max' for OpenAI providers
- Add thinkingXhigh color token to theme schema and built-in themes
- Show xhigh option only when using codex-max models
- Update CHANGELOG for both ai and coding-agent packages
closes#143
- read: actionable notices with offset for continuation
- First line > 30KB: return empty + bash command suggestion
- Hit limit: '[Showing lines X-Y of Z. Use offset=N to continue]'
- bash: tail truncation with temp file
- Notice includes line range + temp file path
- Edge case: last line > 30KB shows partial
- grep: pre-truncate match lines to 500 chars
- '[... truncated]' suffix on long lines
- Notice for match limit and line truncation
- find/ls: result/entry limit notices
- '[N results limit reached. Use limit=M for more]'
- All notices now in text content (LLM sees them)
- TUI simplified (notices render as part of output)
- Never return partial lines (except bash edge case)
- Added totalTokens field to Usage interface in pi-ai
- Anthropic: computed as input + output + cacheRead + cacheWrite
- OpenAI/Google: uses native total_tokens/totalTokenCount
- Fixed openai-completions to compute totalTokens when reasoning tokens present
- Updated calculateContextTokens() to use totalTokens field
- Added comprehensive test covering 13 providers
fixes#130
Tool results now use content blocks and can include both text and images.
All providers (Anthropic, Google, OpenAI Completions, OpenAI Responses)
correctly pass images from tool results to LLMs.
- Update ToolResultMessage type to use content blocks
- Add placeholder text for image-only tool results in Google/Anthropic
- OpenAI providers send tool result + follow-up user message with images
- Fix Anthropic JSON parsing for empty tool arguments
- Add comprehensive tests for image-only and text+image tool results
- Update README with tool result content blocks API
- Add handling for message_start event to capture initial token usage
- Fix message_delta to use assignment (=) instead of addition (+=)
since Anthropic sends cumulative token counts, not incremental
- Add comprehensive tests for all providers (Google, OpenAI Completions,
OpenAI Responses, Anthropic)
- Document OpenAI limitation: token stats only available at stream end
Fixes issue where aborted streams had zero token counts despite
Anthropic sending input tokens in the initial message_start event.
Fixes issue where unpaired Unicode surrogates in tool results cause JSON serialization errors in API providers, particularly Anthropic.
- Add sanitizeSurrogates() utility function to remove unpaired surrogates
- Apply sanitization in all provider convertMessages() functions:
- User message text content (string and text blocks)
- Assistant message text and thinking blocks
- Tool result output
- System prompts
- Valid emoji (properly paired surrogates) are preserved
- Add comprehensive test suite covering all 8 providers
Previously only Google and Groq handled unpaired surrogates correctly.
Now all providers (Anthropic, OpenAI Completions/Responses, Google, xAI, Groq, Cerebras, zAI) sanitize text before API submission.
- Changed ToolRenderer return type from TemplateResult to ToolRenderResult
- ToolRenderResult = { content: TemplateResult, isCustom: boolean }
- isCustom: true = no card wrapper, false = wrap in card
- Updated all existing tool renderers to return new format
- Updated Messages.ts to handle custom rendering
This enables tools to render without default card chrome when needed.
- Add 'aborted' as a distinct stop reason separate from 'error'
- Change AssistantMessage.error to errorMessage for clarity
- Update error event to include reason field ('error' | 'aborted')
- Map provider-specific safety/refusal reasons to 'error' stop reason
- Reorganize utility functions into utils/ directory
- Rename agent.ts to agent-loop.ts for better clarity
- Fix error handling in all providers to properly distinguish abort from error
- Added partial-json package for parsing incomplete JSON during streaming
- Tool call arguments now contain partially parsed JSON during toolcall_delta events
- Enables progressive UI updates (e.g., showing file paths before content is complete)
- Arguments are always valid objects (minimum empty {}), never undefined
- Full validation still occurs at toolcall_end when arguments are complete
- Updated all providers (Anthropic, OpenAI Completions/Responses) to use parseStreamingJson
- Added comprehensive documentation and examples in README
- Added test to verify arguments are always defined during streaming
- Switch from Zod to TypeBox for tool parameter schemas
- TypeBox schemas can be serialized/deserialized as JSON
- Use AJV for runtime validation instead of Zod's parse
- Add StringEnum helper for Google API compatibility (avoids anyOf/const patterns)
- Export Type and Static from main package for convenience
- Update all tests and documentation to reflect TypeBox usage
- Replace JSON Schema with Zod schemas for tool parameter definitions
- Add runtime validation for all tool calls at provider level
- Create shared validation module with detailed error formatting
- Update Agent API with comprehensive event system
- Add agent tests with calculator tool for multi-turn execution
- Add abort test to verify proper handling of aborted requests
- Update documentation with detailed event flow examples
- Rename generate.ts to stream.ts for clarity
- Add 'zai' as a KnownProvider type
- Add ZAI_API_KEY environment variable mapping
- Generate 4 zAI models (glm-4.5-air, glm-4.5v, etc.) using anthropic-messages API
- Add comprehensive test coverage for zAI provider in generate.test.ts and empty.test.ts
- Models support reasoning/thinking capabilities and tool calling
- Anthropic API requires tool call IDs to match pattern ^[a-zA-Z0-9_-]+$
- OpenAI Responses API generates IDs with pipe character (|) which breaks Anthropic
- Added sanitizeToolCallId() to replace invalid characters with underscores
- Fixes cross-provider handoffs from OpenAI Responses to Anthropic
- Added test to verify the fix works
- Implement QueuedGenerateStream class that extends AsyncIterable with finalMessage() method
- Add new types: GenerateStream, GenerateOptions, GenerateOptionsUnified, GenerateFunction
- Create generateAnthropic function-based implementation replacing class-based approach
- Add comprehensive test suite for the new generate API
- Support streaming events with text, thinking, and tool call deltas
- Map ReasoningEffort to provider-specific options
- Include apiKey in options instead of constructor parameter
- Test handling of empty content arrays
- Test handling of empty string content
- Test handling of whitespace-only content
- All providers handle these edge cases gracefully
