* [docs] Add PR template for OSS contributions
Signed-off-by: Peter Wielander <mittgfu@gmail.com>
* Add line
Signed-off-by: Peter Wielander <mittgfu@gmail.com>
* Empty
Signed-off-by: Peter Wielander <mittgfu@gmail.com>
---------
Signed-off-by: Peter Wielander <mittgfu@gmail.com>
### Key Changes
**Generation Settings**
- Support for `temperature`, `maxOutputTokens`, `topP`, `topK`, `presencePenalty`, `frequencyPenalty`, `stopSequences`, `seed` at constructor and per-stream level
**Step Control**
- Added `maxSteps` to limit LLM calls (default: unlimited for backwards compatibility)
- Added `toolChoice` (`auto`, `none`, `required`, or specific tool)
- Added `activeTools` to filter available tools per-stream call
**Callbacks**
- `onChunk` - called for each stream chunk
- `onStepFinish` - called after each step completes
- `onFinish` - called when all steps complete
- `onError` - called on errors
- `onAbort` - called when aborted via `AbortSignal`
**Structured Output**
- Added `experimental_output` with `Output.object({ schema })` for parsing structured responses
**Context & Telemetry**
- Added `experimental_context` to pass data to tool executions
- Added `experimental_telemetry` for observability
**Stream Result**
- `stream()` now returns `{ messages, steps, experimental_output }` instead of just `{ messages }`
**Other**
- `prepareStep` can now override generation settings, toolChoice, and context
- Added `experimental_repairToolCall` for handling malformed tool calls
- Enabled file, source, and tool input streaming chunks
- Exported all new types from `@workflow/ai`
The current implementation of suspension handler processes the invocation queue sequentially which is really slow. Now we process is in parallel:
- Process hooks first, then steps and waits in parallel to prevent race conditions
- Refactor runtime.ts into modular files: `suspension-handler.ts`, `step-handler.ts`, `helpers.ts`
- nit: Add otel attributes for hooks created (`workflow.hooks.created`) and waits created (`workflow.waits.created`)
- nit: Update suspension status from `pending_steps` to `workflow_suspended`
The `idempotency` test was running **110 concurrent steps** (10 + 100), with each step requiring multiple filesystem operations in the local world:
1. **Step creation** - writes step JSON file
2. **Step update** - reads and overwrites step JSON file
3. **Event creation** - writes event JSON file
That's ~3-4 file operations per step × 110 steps = **330-440 total file operations**.
On Windows, this is especially slow because:
- **NTFS** is slower than ext4/APFS for small file writes
- **Windows Defender** real-time scanning adds latency to new file creation
- The **atomic write pattern** (write temp file → rename) is slower on Windows
Reduced the number of steps from **110 to 20** (5 + 15 instead of 10 + 100):
**`workflows/noop.ts`**:
- First batch: 10 → 5 steps
- Second batch: 100 → 15 steps
**`src/idempotency.mts`**:
- Updated assertion to expect 20 numbers instead of 110
This is ~5.5x fewer file operations while still testing the same concurrent step execution and idempotency behavior. The test should now complete well within the 60-second timeout even on slow Windows CI machines.
Vercel Release Bot