[jiahuei]

Summary

• Added explicit provider/protocol clients: Responses and Chat Completions for Bedrock, Together AI, and Fireworks; Responses for OpenRouter; Chat Completions for Gemini. Each class implements one protocol, reusing the shared implementation where sufficient. The registry selects clients without automatic protocol switching. Provider classes own destinations and credential requirements; Bedrock requires an explicit regional URL and accepts BEDROCK_API_KEY alongside the standard AWS credential name.
• Added OpenRouter Responses handling for explicitly labeled raw reasoning and signed summaries, plus Gemini thought-signature translation and unambiguous streamed tool-index repair. Gemini preserves each Google signature in ordered IR thinking parts, projects it as Responses encrypted_content or Messages signature, and restores its text/tool attachment on continuation. Unsupported reasoning or opaque fields fail explicitly. Shared Chat streaming closes tool parts before subsequent content and rejects resumed noncontiguous calls.
• Extended tests/upstream/test_multi_turn.py through its CLIENTS inventory, covering independent SDK parity, actual-output history, normalized continuation, and public translation or rejection. Added 87 conventionally named provider cassettes under upstream_ir/. Forty Chat Completions exchanges cover two-turn text and three-turn tools in both native modes for Bedrock Luna, Together Qwen3.5-9B, Fireworks Nemotron, and Gemini 3.1 Flash-Lite. Responses coverage includes those providers and OpenRouter's free Nemotron model. Captures retain their original bytes; tests use documented input-copy adjustments where SDK output contains fields outside request schemas.
• Fixed recording-proxy staging, retry policy, credential masking, non-JSON error-byte fidelity, and partial-stream evidence. Removed duplicate conversation suites, the PR WIP: Add more providers #76 reference corpus, manifests, source snapshots, and standalone recording drivers. Updated architecture, conventions, and the provider support matrix.

Related to #78. This remains a draft: recordings establish selected models and scenarios, not complete catalogs. Gemini native, normalized, and Responses/Messages paths preserve the recorded text/tool thought signatures. Parallel-call signature association also has synthetic regression coverage. Other models, images, and some parallel/reasoning conversations remain unverified. Fireworks unsigned Responses summaries reject translation; OpenRouter raw output projects to Messages, but unsigned Messages reasoning input still rejects. Bedrock coverage uses a global inference profile through the Singapore runtime endpoint, which does not establish the actual inference region.

Verification

• Merged main at 70ad86c without conflicts. A clean core-only environment passed import molto with PyYAML absent, matching the new CI import check.

• .venv/bin/python -m pytest -q -o addopts='' --tb=short: 2,633 passed, Python 3.14.7, offline replay only.

• .venv/bin/python -m pytest tests/upstream/test_multi_turn.py -q -o addopts='' --tb=short: 89 passed, including four Gemini native conversations, two normalized signature continuations, and eight Responses/Messages conversations across streaming and non-streaming output.

• 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.

• .venv/bin/python -m pytest tests/test_providers.py tests/upstream/test_openai.py tests/test_hosted_providers.py tests/core/test_provider_history_routes.py tests/upstream/test_multi_turn.py -q -o addopts='' --tb=short: 480 passed. After strengthening the synthetic streaming helper, the 30-test provider-history suite also passed.

• git diff --check: passed. All 40 Chat Completions cassettes are byte-identical to their reviewed sources, with complete responses and no configured credentials. Existing cassette content was unchanged. Recording used native SDKs outside pytest, zero retries, and bounded request deadlines.

• Bedrock Chat Completions recording: ten successful exchanges with global.openai.gpt-5.6-luna, reasoning disabled. Ministral 3B attempts returned runtime HTTP 400 and a 75-second Mantle timeout; unsuccessful captures remain outside the committed corpus and recorder processes exited.

• Together Chat Completions recording: ten successful exchanges with Qwen/Qwen3.5-9B, reasoning disabled. Caller history restores the single assistant role after SDK accumulation of repeated role deltas. Fireworks: ten successful exchanges with accounts/fireworks/models/nemotron-lightning-3p5-30b-a3b, reasoning disabled and null output-only history fields omitted.

• Retained OpenRouter recordings use an explicit free model, disabled fallbacks, zero price caps, and no plugins.

• Gemini recording: run 42160882628 completed ten exchanges using the existing Actions secret, native OpenAI SDK calls, minimal reasoning, zero retries, and bounded request deadlines. Artifact YAML was imported unchanged and passed credential/completion checks. Caller history restores repeated assistant roles and omits top-level null fields while preserving nested Google state. Independent SDK references repair only the recorded single-call missing index; actual Molto output is not repaired by tests. Temporary recording tooling remains on a separate branch; no credential upload or secret changes were made.

• gh pr checks 96 --repo EmbeddedLLM/molto --watch --interval 15: all five checks passed at 56c7a22: tests on Python 3.12, 3.13, and 3.14, Ruff, and Prettier.

[github-actions]

Coverage Report (Python 3.14) •

File	Stmts	Miss	Cover	Missing
configs.py	41	4	90%	45, 49, 57, 61
core.py	146	3	98%	119, 132, 186
providers.py	12	0	100%
adapters
anthropic.py	461	46	90%	147, 157, 195, 223, 260, 311, 316, 320, 327, 331, 335, 341, 347, 351–353, 356–358, 361, 369, 371, 376, 382, 389, 413–414, 521, 583, 609, 633, 635, 637, 699, 706, 729, 755, 761, 828, 848, 873, 877, 882, 888, 896, 950
openai.py	656	64	90%	112, 192–194, 196, 205, 362, 367, 379, 420, 422, 459, 468, 483, 494, 519, 534, 559, 591, 595, 732, 743, 747, 778, 840, 843, 848, 909, 923, 959–962, 971, 975, 982, 986, 1042, 1117, 1187, 1227, 1263, 1269, 1280, 1290, 1304, 1313, 1325, 1398, 1400, 1458, 1460, 1469, 1475, 1528, 1535, 1541, 1548, 1552, 1566, 1577, 1584, 1590, 1594
types
openai_chat_completions.py	119	0	100%
openai_responses.py	145	0	100%
upstream
base.py	61	4	93%	113, 118, 132, 137
bedrock.py	7	0	100%
fireworks.py	7	0	100%
gemini.py	81	2	98%	67, 74
openai.py	809	34	96%	156, 158, 191, 203, 213, 225, 289–290, 299, 439, 468, 626, 634, 636, 639, 673–674, 860, 950, 1029, 1050, 1218, 1250, 1286, 1288, 1292–1293, 1477, 1523, 1527, 1540–1541, 1558, 1560
openrouter.py	19	0	100%
registry.py	11	0	100%
together.py	7	0	100%
vllm.py	5	0	100%
utils
test.py	364	163	55%	196, 245–248, 258–263, 273–277, 308, 313, 354–356, 366–369, 379–380, 411, 440–443, 459, 469–470, 472–474, 485–486, 495–499, 508–510, 517, 527, 570, 579–581, 583–586, 594–595, 606–607, 609–613, 615–617, 630–632, 651, 660, 691–692, 707–709, 715, 730–731, 738–740, 742, 750–751, 753, 755–756, 777–784, 798, 803, 811, 819–826, 828–829, 836, 838, 845, 847, 852–853, 900, 905, 909–911, 917–919, 925–926, 928–931, 933–934, 940–941, 945–948, 953–956, 961, 963–967, 986–987, 989–994, 1022, 1043–1045
TOTAL	4145	474	89%

Tests	Skipped	Failures	Errors	Time
2633	0 💤	0 ❌	0 🔥	31.005s ⏱️

[github-actions]

Coverage Report (Python 3.12) •

File	Stmts	Miss	Cover	Missing
configs.py	41	4	90%	45, 49, 57, 61
core.py	146	3	98%	119, 132, 186
providers.py	13	0	100%
adapters
anthropic.py	461	46	90%	147, 157, 195, 223, 260, 311, 316, 320, 327, 331, 335, 341, 347, 351–353, 356–358, 361, 369, 371, 376, 382, 389, 413–414, 521, 583, 609, 633, 635, 637, 699, 706, 729, 755, 761, 828, 848, 873, 877, 882, 888, 896, 950
openai.py	656	64	90%	112, 192–194, 196, 205, 362, 367, 379, 420, 422, 459, 468, 483, 494, 519, 534, 559, 591, 595, 732, 743, 747, 778, 840, 843, 848, 909, 923, 959–962, 971, 975, 982, 986, 1042, 1117, 1187, 1227, 1263, 1269, 1280, 1290, 1304, 1313, 1325, 1398, 1400, 1458, 1460, 1469, 1475, 1528, 1535, 1541, 1548, 1552, 1566, 1577, 1584, 1590, 1594
types
openai_chat_completions.py	119	0	100%
openai_responses.py	145	0	100%
upstream
base.py	62	4	94%	113, 118, 132, 137
bedrock.py	7	0	100%
fireworks.py	7	0	100%
gemini.py	81	2	98%	67, 74
openai.py	809	34	96%	156, 158, 191, 203, 213, 225, 289–290, 299, 439, 468, 626, 634, 636, 639, 673–674, 860, 950, 1029, 1050, 1218, 1250, 1286, 1288, 1292–1293, 1477, 1523, 1527, 1540–1541, 1558, 1560
openrouter.py	19	0	100%
registry.py	11	0	100%
together.py	7	0	100%
vllm.py	5	0	100%
utils
test.py	364	163	55%	196, 245–248, 258–263, 273–277, 308, 313, 354–356, 366–369, 379–380, 411, 440–443, 459, 469–470, 472–474, 485–486, 495–499, 508–510, 517, 527, 570, 579–581, 583–586, 594–595, 606–607, 609–613, 615–617, 630–632, 651, 660, 691–692, 707–709, 715, 730–731, 738–740, 742, 750–751, 753, 755–756, 777–784, 798, 803, 811, 819–826, 828–829, 836, 838, 845, 847, 852–853, 900, 905, 909–911, 917–919, 925–926, 928–931, 933–934, 940–941, 945–948, 953–956, 961, 963–967, 986–987, 989–994, 1022, 1043–1045
TOTAL	4190	474	89%

Tests	Skipped	Failures	Errors	Time
2633	0 💤	0 ❌	0 🔥	59.703s ⏱️

[github-actions]

Coverage Report (Python 3.13) •

File	Stmts	Miss	Cover	Missing
configs.py	41	4	90%	45, 49, 57, 61
core.py	146	3	98%	119, 132, 186
providers.py	13	0	100%
adapters
anthropic.py	461	46	90%	147, 157, 195, 223, 260, 311, 316, 320, 327, 331, 335, 341, 347, 351–353, 356–358, 361, 369, 371, 376, 382, 389, 413–414, 521, 583, 609, 633, 635, 637, 699, 706, 729, 755, 761, 828, 848, 873, 877, 882, 888, 896, 950
openai.py	656	64	90%	112, 192–194, 196, 205, 362, 367, 379, 420, 422, 459, 468, 483, 494, 519, 534, 559, 591, 595, 732, 743, 747, 778, 840, 843, 848, 909, 923, 959–962, 971, 975, 982, 986, 1042, 1117, 1187, 1227, 1263, 1269, 1280, 1290, 1304, 1313, 1325, 1398, 1400, 1458, 1460, 1469, 1475, 1528, 1535, 1541, 1548, 1552, 1566, 1577, 1584, 1590, 1594
types
openai_chat_completions.py	119	0	100%
openai_responses.py	145	0	100%
upstream
base.py	62	4	94%	113, 118, 132, 137
bedrock.py	7	0	100%
fireworks.py	7	0	100%
gemini.py	81	2	98%	67, 74
openai.py	809	34	96%	156, 158, 191, 203, 213, 225, 289–290, 299, 439, 468, 626, 634, 636, 639, 673–674, 860, 950, 1029, 1050, 1218, 1250, 1286, 1288, 1292–1293, 1477, 1523, 1527, 1540–1541, 1558, 1560
openrouter.py	19	0	100%
registry.py	11	0	100%
together.py	7	0	100%
vllm.py	5	0	100%
utils
test.py	364	163	55%	196, 245–248, 258–263, 273–277, 308, 313, 354–356, 366–369, 379–380, 411, 440–443, 459, 469–470, 472–474, 485–486, 495–499, 508–510, 517, 527, 570, 579–581, 583–586, 594–595, 606–607, 609–613, 615–617, 630–632, 651, 660, 691–692, 707–709, 715, 730–731, 738–740, 742, 750–751, 753, 755–756, 777–784, 798, 803, 811, 819–826, 828–829, 836, 838, 845, 847, 852–853, 900, 905, 909–911, 917–919, 925–926, 928–931, 933–934, 940–941, 945–948, 953–956, 961, 963–967, 986–987, 989–994, 1022, 1043–1045
TOTAL	4190	474	89%

Tests	Skipped	Failures	Errors	Time
2633	0 💤	0 ❌	0 🔥	1m 24s ⏱️