Stacked on #1395 — review/merge that first. The diff here includes both PRs; the incremental diff is only the provider-specific changes and integration tests.

Summary

Builds on #1395 to implement full reasoning trace preservation across all providers, ensuring reasoning content survives multi-turn conversations including tool call loops.

• OpenAI Responses API: Threads msg_ IDs through streaming/non-streaming paths; reverts strip-reasoning workaround; emits MessageId from OutputItemDone; includes reasoning_encrypted_content automatically when reasoning is configured
• Anthropic: Multi-block Reasoning → Thinking/RedactedThinking conversion with signature preservation; streaming RedactedThinking support via get_or_insert_with
• Gemini: thought_signature roundtrip preservation; emits full Reasoning block (not just delta) when signature is present in streaming
• xAI: Structured Reasoning ↔ Chat Completions reasoning_content conversion with encrypted/redacted mapping; fixes #[serde(rename)] → #[serde(rename_all)]
• OpenRouter: Groups reasoning_details by ID preserving ordering; maps Encrypted/Redacted content types through Chat Completions format
• Streaming agent loop bug fix: ReasoningDelta events now accumulate in a separate text buffer instead of being converted to unsigned Reasoning blocks — fixes Anthropic 400 "signature required" errors in multi-turn tool call loops

Closes #1147, closes #1146

Integration Tests

Four test suites covering both streaming and non-streaming code paths with per-provider coverage:

Streaming Tests

reasoning_roundtrip.rs (streaming) — Text-only 2-turn streaming conversation:

• Turn 1: model.stream() → collect reasoning blocks/deltas + text
• Turn 2: follow-up with reasoning in chat history → provider accepts

reasoning_tool_roundtrip.rs (streaming) — Reasoning + tool call via agent.stream_chat().multi_turn():

• Agent receives weather prompt → model thinks → calls get_weather tool
• Agent loop preserves [Reasoning, ToolCall] in chat history
• Agent sends tool result → model responds with text referencing tool output
• 11 universal assertions per provider (no errors, tool invoked, reasoning before tool call, etc.)
• Provider-specific assertions (OpenAI: Encrypted content, Anthropic: signatures)

Non-Streaming Tests

reasoning_roundtrip.rs (non-streaming) — Text-only 2-turn non-streaming conversation:

• Turn 1: model.completion() → verify CompletionResponse.choice contains Reasoning blocks
• Turn 2: follow-up with reasoning in chat history → provider accepts
• Directly inspects reasoning content types in the structured response

reasoning_tool_roundtrip.rs (non-streaming) — Reasoning + tool call via agent.chat():

• Non-streaming agent loop internally calls model.completion(), preserves reasoning via resp.choice.clone()
• Verifies: no errors (400 = dropped reasoning), tool invoked, text references tool output
• 4 universal assertions per provider

Results

Test plan

• cargo fmt clean
• cargo clippy -p rig-core --tests — 0 warnings
• cargo test -p rig-core --lib — 271 tests pass
• Streaming reasoning_roundtrip integration tests pass (4/4 providers)
• Streaming reasoning_tool_roundtrip integration tests pass (4/4 providers)
• Non-streaming reasoning_roundtrip integration tests pass (4/4 providers)
• Non-streaming reasoning_tool_roundtrip integration tests pass (4/4 providers)
• xAI tests pending API key validation with grok-4-0725