[jiahuei]

Summary

• Added explicit provider configuration for Together AI, Gemini, OpenRouter, Fireworks, and Bedrock, with protocol-specific destinations, credential isolation, and no silent endpoint switching.
• Added strict reasoning/history extensions, native Google attachment and OpenRouter format preservation, unambiguous Gemini tool-index repair, balanced tool-stream lifecycles, and Together semantic-error recognition. Fireworks native continuation preserves observed Chat null placeholders and explicit nullable Messages content; Responses function tools may omit strict. Unsupported translation and invalid history fail explicitly.
• Repaired recorder error-byte fidelity, partial-stream evidence, sanitization, and staging. Independent native SDK drivers cover all three protocols, including Messages bearer authentication, explicit provider controls, actual-output continuation, and completion validation. A reasoning-detail check stops when SDK accumulation corrupts repeated metadata. Fresh captures and source snapshots remain separate from unchanged historical matched fixtures from PR #76.
• Added SDK parity, exact next-request bodies, translated semantic checks, corruption checks, credential/URL tests, and recorder/matcher regressions. Fireworks contributes 36 complete native conversations; OpenRouter contributes 13. Failed attempts remain visible, and original HTTP captures/manifests were not edited to make tests pass.

Related to #78. This remains a draft because the full issue is not complete. The support matrix distinguishes recorded, matched, synthetic, unsupported, and unknown coverage:

• OpenRouter's public corpus contains 40 exchanges in 19 groups: 13 complete conversations and six unsuccessful attempts. nvidia/nemotron-3.5-lightning:free covers text and sequential tools across all three protocols and Chat parallel calls. All requests use explicit free IDs, zero price caps, disabled provider fallbacks, and no plugins; every reported cost was zero. Ten native histories replay fully; three non-streaming Chat histories reject provider reasoning:null. The 24 translated first-turn cases include explicit rejection of raw Responses reasoning under the summary profile. liquid/lfm-2.5-2.6b:free returned recorded HTTP 400 errors when reasoning was disabled. Two additional low-reasoning groups hit HTTP 429 and remain outside the repository because their responses contain an account identifier. The total live run used 43 requests.
• Fireworks accounts/fireworks/models/nemotron-lightning-3p5-30b-a3b completed text, sequential tools, and parallel tools across Chat, Responses, and Messages, each with both native modes and reasoning disabled/low. Its 47 groups contain 101 exchanges, including 11 incomplete attempts. All 36 completed conversations replay with exact native history; 72 translated first-turn cases preserve representable semantics or explicitly reject unsigned summaries.
• Together Qwen/Qwen3.5-9B completed native SDK text and sequential tools in both modes, plus non-streaming parallel calls. Full Molto continuation is verified for non-streaming sequential/parallel tools. Recorded null reasoning/output-only histories reject; its reasoning_content output is unsupported by Together's translated reasoning dialect. Streaming parallel attempts returned only one call.
• Gemini live requests returned HTTP 400 because this environment's location is unsupported. Its history/index behavior is covered offline. Bedrock live cells remain unknown because credentials/profile were unavailable.
• Translated replay does not establish live translated requests or translated multi-turn continuation. Images and additional models remain gaps. Anthropic extra_content.anthropic continuation remains owned by feat: support hosted web search across Responses and Messages #74/fix: Preserve Anthropic continuation through extra_content #92.

Verification

• uv run --extra test pytest: 2,195 passed on Python 3.12.9, offline/replay-only.
• uv run pytest tests/core/test_openrouter_conversations.py tests/core/test_fireworks_conversations.py tests/core/test_recorded_provider_conversations.py -q --tb=short: 156 passed, including all 40 OpenRouter cases (native/rejection, translated, HTTP error, and SDK accumulation-failure parity).
• Focused recorder suites: 22 native-driver and 29 Chat-driver tests passed.
• uv run --extra lint ruff check --output-format github --config pyproject.toml .: passed.
• uv run --extra lint ruff format --diff --config pyproject.toml .: passed.
• npx prettier@3.8 --check .: passed. An initial formatting failure in authored catalog-selection metadata was corrected; original captures and manifests were unchanged.
• git diff --check: passed. Original OpenRouter captures/manifests/source snapshots remained byte-identical to baseline 30f8b6d after replay.
• Standalone native SDK driver runs retained exact settings and successful/incomplete outcomes in manifests. Public captures passed credential/account-identifier scans, decoded-byte inspection, source-hash checks, and byte comparisons against staging. No pytest recording mode was used. Bedrock live recording was not run because credentials/profile were unavailable.
• Remote CI at 6002ba6: unit tests passed on Python 3.12, 3.13, and 3.14; both lint/formatting jobs passed.