replayd v0.1.3

Auto-instrumentation for OpenAI and Anthropic. Call instrument_openai(client) or instrument_anthropic(client) once and tool calls record themselves during capture ,no manual record_tool_call needed. Works with both sync (OpenAI, Anthropic) and async (AsyncOpenAI, AsyncAnthropic) clients.

replayd run CLI. Replay saved tests against an agent without writing a script: replayd run --agent module:agent_fn. Exits 1 on failure for CI.

Deeper grading. Assert on the arguments of an expected action and on call sequence, not just tool names.

uninstrument_openai / uninstrument_anthropic. Cleanly restore a client to its original state. Idempotent.

Real OpenAI example. examples/real_openai_agent.py runs an actual model call end to end.

Known gaps, documented: streaming responses (stream=True) are not auto-recorded yet and now warn loudly with the manual fallback. Async replay needs asyncio.run() around the agent. Both land in 0.1.4.

61 tests passing. Full details in CHANGELOG.md.

replayd v0.1.2 -- 6 fixes, 3 example agents, expanded tests

Six issues fixed, three real-world example agents added, test suite expanded from 34 to 41 tests.

Fixes

• README Python badge now correctly shows python-3.10+ (was 3.8+); the code uses str | None syntax which requires 3.10+
• Grading model is no longer hardcoded -- Replayd(grader_model=...) lets callers override the LLM judge without editing library source; default remains claude-haiku-4-5-20251001
• CaptureContext.__exit__ now emits a warnings.warn when run.output is None at exit, alerting developers who forget to assign it inside the with block; the run is still saved
• Structural grader now collects all forbidden violations before returning -- a replay that calls two forbidden tools reports both in the reason string instead of stopping at the first
• save_test() accepts an optional forbidden_call_args: dict parameter; when set, a forbidden action only triggers a FAIL if the call arguments contain every key/value pair in the dict
• replay_one() now has a docstring with a usage example

New examples

• examples/multi_step_planning_agent.py -- planning agent that skips check_constraints before finalize_plan
• examples/rag_policy_agent.py -- RAG support agent that approves a refund from a stale/deprecated policy chunk
• examples/incident_response_agent.py -- SRE agent that calls rollback_deploy without paging the on-call engineer first

Tests

• tests/test_examples.py -- 6 tests (buggy/fixed pair for each new example agent)
• tests/test_core.py -- 8 new tests covering grader_model, output=None warning, multi-violation reporting, forbidden_call_args matching, and replay_one by test ID
• Total: 41 tests (was 34)

replayd v0.1.1 -- bug fixes and performance improvement

Bug fixes

• Agent crash during replay now returns FAIL instead of crashing the entire suite
• Non-callable agent now raises a descriptive error instead of cryptic TypeError
• 100-test replay improved from 9.48s to 2.65s with run caching (3.58x faster)

replayd v0.1.0 — initial release

First release of replayd.
Core loop working: capture a failed agent run, save it as a regression test, replay it before you ship.
13 tests passing. pip install replayd.