fix bug encountered when running GLM-4.7-Flash hosted by LM Studio, in
which the provider sends tool call arguments via
`response.function_call_arguments.done` events instead of streaming them
via `response.function_call_arguments.delta` events. The final
`response.output_item.done` event then contains empty `{}` arguments.
The code only handled delta events, so tool calls failed with validation
errors like `"must have required property 'command'"`.
Full disclosure, Opus triaged the bug and provided the fix (by adding
logging statements to the req/resp to the upstream provider (LM
Studio)). I'm to provide prompts/transcripts, and acknowledge that I'm
not an expert in Pi internals at this time.
When switching between OpenAI models (e.g., gpt-5-mini to gpt-5.2-codex),
function_call IDs with fc_ prefix trigger pairing validation errors because
OpenAI tracks which fc_xxx IDs were paired with rs_xxx reasoning items.
The fix omits the id field for function_calls from different models, which
avoids the pairing validation while keeping call_id for matching with
function_call_output.
Fixes#886
- Added headers field to base StreamOptions interface
- Updated all providers to merge options.headers with defaults
- Forward headers and onPayload through streamSimple/completeSimple
- Bedrock not supported (uses AWS SDK auth)
Fixes OpenAI Responses 400 error 'reasoning without following item' by
skipping errored/aborted assistant messages entirely rather than filtering
at the provider level. This covers openai-responses, openai-codex-responses,
and future providers.
Removes strictResponsesPairing compat option (no longer needed).
Closes#838
Split OpenAICompat into OpenAICompletionsCompat and OpenAIResponsesCompat
for type-safe API-specific compat settings. Added strictResponsesPairing
option to suppress orphaned reasoning/tool calls on incomplete turns,
fixing 400 errors on Azure's Responses API which requires strict pairing.
Closes#768
- Replace hardcoded pi->CC tool mappings with single CC tool name list
- Case-insensitive lookup: if tool name matches CC tool, use CC casing
- Remove broken find->Glob mapping (round-trip failed)
- Add test coverage for tool name normalization
When an assistant message has stopReason 'error', its tool calls are now
excluded from pending tool tracking. This prevents synthetic tool results
from being generated for calls that will be dropped by provider-specific
converters (e.g., Codex drops tool calls from errored messages).
Previously, this mismatch caused OpenAI to reject requests with 'No tool
call found for function call output with call_id ...' errors.
fixes#812
Instead of skipping unsigned tool calls entirely (which lobotomizes context),
convert them to text with an explicit note telling the model this is historical
context from a different model and not a format to mimic.
This preserves tool call/result context when switching from providers without
thought signatures (e.g. Claude via Antigravity) to Gemini 3.
When 429/500 errors occur during tool execution, empty assistant messages
with stopReason='error' get persisted. These break the tool_use -> tool_result
chain for Claude/Gemini APIs.
Added centralized filtering in transformMessages to skip assistant messages
with empty content and no tool calls. Provider-level filters remain for
defense-in-depth.
Previously, OpenAI-compatible provider settings (like developer role support)
were detected only from the baseUrl. When using custom URLs (e.g., proxies),
detection failed. Now checks model.provider first, then falls back to URL.
Fixes#774
- align Codex Responses provider with Pi static instructions
- simplify Codex request/stream handling and cleanup exports
- keep legacy OpenCode Codex prompt for testing until Pi prompt is allowlisted
* Add Amazon Bedrock models test suite for agent package
Tests basic prompts, multi-turn conversations with thinking, and
synthetic thinking signatures across all Bedrock models.
Known issues are categorized and skipped:
- Models requiring inference profile (5)
- Invalid model IDs for us-east-1 region (6)
- Max tokens config exceeds model limit (2)
- No signature support in reasoningContent (10)
- Rejects reasoning content in user messages (25)
- Validates signature format - Anthropic newer models (7)
* Fix Bedrock signature support for non-Anthropic models
Only include the signature field in reasoningContent.reasoningText for
Anthropic Claude models. Other models (OpenAI, Qwen, Minimax, Moonshot,
etc.) reject this field with:
"This model doesn't support the reasoningContent.reasoningText.signature field"
This fix enables multi-turn conversations with thinking content for
10 additional Bedrock models that previously failed.
https://buildwithpi.ai/session?7e39c05f66ea358da3f993c267fe3e29
* Add a CHANGELOG entry
Z.ai uses thinking: { type: "enabled" | "disabled" } instead of
OpenAI's reasoning_effort. Added thinkingFormat compat flag to handle
this. Thinking is now explicitly enabled/disabled based on user setting.
- Add minimax to KnownProvider and Api types
- Add MINIMAX_API_KEY to getEnvApiKey()
- Generate MiniMax-M2 and MiniMax-M2.1 models
- Add context overflow detection pattern
- Add tests to all required test files
- Update README and CHANGELOG with attribution
Also fixes:
- Bedrock duplicate toolResult ID when content has multiple blocks
- Sandbox extension unused parameter lint warning
Improve Gemini CLI provider retries and headers
- Add Antigravity endpoint fallback (tries daily sandbox then prod when baseUrl is unset)
- Parse retry delays from headers (Retry-After, x-ratelimit-reset, x-ratelimit-reset-after) before body parsing
- Derive stable sessionId from first user message for cache affinity
- Retry empty SSE streams with backoff without duplicate start/done events
- Add anthropic-beta header for Claude thinking models only
* Improve Cloud Code Assist error messages
- Extract just the message from verbose JSON error responses
- Extract cause from generic 'fetch failed' errors for better diagnostics
* Make 'other side closed' network error retryable
* Make 'other side closed' network error retryable
Adds cache points to system prompt and last user message for:
- Claude 3.5 Haiku
- Claude 3.7 Sonnet
- Claude 4.x models (Opus, Sonnet, Haiku)
Uses Bedrock's cachePoint blocks with 5-minute TTL.
Adds support for Amazon Bedrock with Claude models including:
- Full streaming support via Converse API
- Reasoning/thinking support for Claude models
- Cross-region inference model ID handling
- Multiple AWS credential sources (profile, IAM keys, API keys)
- Image support in messages and tool results
- Unicode surrogate sanitization
Also adds 'Adding a New Provider' documentation to AGENTS.md and README.
Co-authored-by: nickchan2 <nickchan2@users.noreply.github.com>
Claude models accessed through Google Cloud Code Assist API require
explicit id fields in both functionCall and functionResponse parts.
Without these IDs, the API returns 'tool_use.id: Field required' error.
Add requiresToolCallId() helper to centralize the Claude model detection
and include IDs in both tool call and tool result message conversions.
* Avoid cross-provider thought signatures
* Fix Google thought signature replay
Filter thought signatures to same provider with base64 validation and rename the transform helper for clarity.
- Store thoughtSignature on text blocks during streaming (all 3 providers)
- Replay textSignature as thoughtSignature in convertMessages
- Remove redundant conditional since retainThoughtSignature handles undefined
Per Google docs, text part signatures are optional but recommended for
high-quality reasoning in multi-turn conversations.
- isThinkingPart now only checks thought === true, not thoughtSignature
- thoughtSignature is for context replay and can appear on any part type
- Store thoughtSignature on text blocks as textSignature for proper replay
- Remove id from functionCall/functionResponse (unsupported by Vertex/Cloud Code Assist)
Refs: https://ai.google.dev/gemini-api/docs/thought-signatures
Co-authored-by: Amp <amp@ampcode.com>
Renamed from supportsStreamOptions to clarify this controls stream_options: { include_usage: true }.
Defaults to true (no behavioral change for existing providers).
Providers like gatewayz.ai that reject this parameter can set supportsUsageInStreaming: false in model config.
Based on #596 by @XesGaDeus
OpenRouter with models like openai/gpt-5.2 enforces strict mode which
requires all properties in the required array. Setting strict: false
allows optional parameters without null unions, matching the approach
already used in openai-responses.ts.
lm-studio hosted openai-like api endpoint requires this parameter to either be a defined boolean, or not specifying this option entirely. null will fail the API validation.
