[achildress83]

What

Implements the medium-priority items of BACKLOG.md: 13, 15, 16, 17 and 18, and records the live check for 14. Stacked on #13, so the base is feature/high-priority; retarget to main once #13 merges.

Why

• A transcript saved mid-turn (unanswered tool calls) was sent as-is and rejected with a 400 that is never retried.
• A handler that is not thread-safe could not opt out of parallel execution.
• Timeouts were hard-coded per adapter, and the adapter guessed whether a supplied client had chosen its own.
• OpenAI's o-series models reject parallel_tool_calls, which was sent on every request.
• A second configure_tracing(), or trace=True with no provider, silently discarded spans.

Changes

• Resume: unanswered tool calls are run before the first model call, then check_invariants() runs, so a bad transcript raises TranscriptError before any model call.
• ToolSpec.parallel_safe and @tool(parallel_safe=...): a turn calling an unsafe tool runs all its calls sequentially, in call order.
• Settings.timeout_seconds and LLM_TIMEOUT_SECONDS: replaces the read timeout of the OpenAI-compatible providers; for Anthropic it is the SDK's overall timeout. An explicit timeout now beats a supplied http_client's.
• Model-aware quirks: providers.openai.MODEL_QUIRKS and quirks_for_model (o1/o3/o4 omit parallel_tool_calls), plus a quirks= override on build_openai_client.
• Tracing: one warning for a second configure_tracing() and one for a traced run with no provider installed.
• Item 14: checked live. o4-mini returns encrypted_content without include, and replay is accepted. No code change.
• Changelog, docs and backlog updated.

Not done: resuming a turn that paused and then failed (13), other OpenAI families and OpenRouter/Modal in the quirks table (17), and the tracing detection is untested on older OpenTelemetry SDKs (18). ruff and mypy are clean, 586 tests pass, coverage 94.8%.

🤖 Generated with Claude Code