name	agent-recorder-fixtures
description	Record real OpenAI/Anthropic HTTP back-and-forth (requests + responses, including streaming text/event-stream) and print paste-ready Swift fixtures for SwiftAgent unit tests (ReplayHTTPClient) using the AgentRecorder CLI or HTTPReplayRecorder. Use when adding/updating any provider adapter tests (text, streaming, structured outputs, tool calls), when payload formats change, or when debugging agent loop mismatches by inspecting recorded JSON/SSE payloads.

name

agent-recorder-fixtures

description

Record real OpenAI/Anthropic HTTP back-and-forth (requests + responses, including streaming text/event-stream) and print paste-ready Swift fixtures for SwiftAgent unit tests (ReplayHTTPClient) using the AgentRecorder CLI or HTTPReplayRecorder. Use when adding/updating any provider adapter tests (text, streaming, structured outputs, tool calls), when payload formats change, or when debugging agent loop mismatches by inspecting recorded JSON/SSE payloads.

Agent Recorder Fixtures

Overview

Use AgentRecorder to capture real provider payloads and generate paste-ready Swift fixtures for tests using ReplayHTTPClient. This is the fastest loop for keeping SwiftAgent’s unit tests (and agent-loop behavior) in sync with real OpenAI/Anthropic traffic.

Workflow

1) Pick the right scenario - Existing scenarios live in AgentRecorder/AgentRecorder/Scenarios/ (grouped by provider + one file per scenario). - List them quickly with AgentRecorder --list-scenarios. - If no scenario matches your test, add one (keep it small and deterministic).

2) Set API keys - Preferred: use a local Secrets.plist in the repo root (not committed). - Run with --secrets-plist Secrets.plist - Plist keys: OpenAI_API_Key_Debug / Anthropic_API_Key_Debug

Optional: - Env vars fallback (useful for CI or quick runs): OPENAI_API_KEY, ANTHROPIC_API_KEY - AGENT_RECORDER_SECRETS_PLIST is also supported (fallback if env vars are missing).

Legacy: - If you already have Secrets.plist under Examples/Example App/ExampleApp/Secrets.plist, either move/symlink it to the repo root or pass the full path via --secrets-plist.

3) Run the recorder (Xcode or Terminal) - Xcode: select AgentRecorder scheme, set env vars, Run → copy stdout from Debug console. - Terminal:

# Run from repo root.
# Apple Silicon: add `arch=arm64` to avoid “multiple matching destinations” warnings.
xcodebuild -workspace SwiftAgent.xcworkspace -scheme AgentRecorder -destination "platform=macOS,arch=arm64" -derivedDataPath .tmp/DerivedData build -quiet
./.tmp/DerivedData/Build/Products/Debug/AgentRecorder --list-scenarios
./.tmp/DerivedData/Build/Products/Debug/AgentRecorder --secrets-plist Secrets.plist --provider openai --scenario openai/streaming-tool-calls/weather --no-include-headers

4) Paste fixtures into tests - Output is already formatted for ReplayHTTPClient(recordedResponses:). - Paste into the relevant test file (common locations): - Tests/SwiftAgentTests/OpenAISession/... - Tests/SwiftAgentTests/AnthropicSession/...

5) Run tests - xcodebuild -workspace SwiftAgent.xcworkspace -scheme SwiftAgentTests -testPlan SwiftAgentTests test -quiet

Notes

Headers may include secrets. HTTPReplayRecorder redacts common auth header fields, but always review before committing.
Prefer --no-include-headers unless you’re explicitly testing header handling.
Always pass -quiet to xcodebuild to keep logs readable.
xcodebuild may print noisy destination warnings; ignore if the build succeeds.
Streaming fixtures are raw text/event-stream. If the SDK stops consuming early, the recorded payload may be partial (this is usually fine for replaying the SDK’s behavior).
If you need request bodies for debugging, re-run with --include-requests.
OpenAI scenarios: keep models pinned/stable (scenario-defined). Change only when you’re updating fixtures/tests.
Cleanup: if you used .tmp/DerivedData (and/or wrote capture files like .tmp/AgentRecorderOutput/*.txt), delete them after you’ve pasted fixtures into tests.

来源：SwiftedMind/SwiftAgent · MIT 许可 · 许可原文 · MD 为来源页保存的历史版本