[iAmir97]

Summary

• Added provider-hosted web search across Responses and Messages, including supported search options, tool selection, calls, results, URL citations, assistant history, and reported search usage in streaming and non-streaming responses.
• Preserved native passthrough behavior and explicit rejection of unsupported translated request semantics. Chat Completions callers can request mandatory search through Responses or Messages upstreams when the returned reasoning is absent or unsigned.
• Added Messages citation policies: drop streams text without citations by default; buffer preserves attribution by waiting for completed text parts; auto buffers when normalized tools or history indicate search or citations.
• Preserved Anthropic search continuation through extra_content.anthropic, implemented by merged PR #92. Serialized Responses output retains source ciphertext/details and citation indices/excerpts for the next turn against the same Anthropic Messages upstream, with ordering and call/result linkage preserved.
• Added matched fixtures, independent Anthropic search and continuation recordings, SDK parity and actual-output continuation coverage, matching-tool updates, and documentation.

Limitations

• Translation supports basic hosted search. Normalized search is rejected on translated Chat upstream routes; native Chat search remains supported. Dynamic filtering and fetch remain unsupported for translation.
• Translated Chat output omits structured citations and search activity, and rejects signed or encrypted reasoning.
• Messages citation buffering increases text latency and changes chunk boundaries. Native Messages passthrough preserves citations under every policy.
• Cross-protocol citations and history use the documented reductions. Anthropic continuation state is not portable across providers; missing or invalid required state and incompatible routes are rejected.
• extra_body and extra_headers remain opaque upstream transport extensions, with credential-boundary header filtering. They do not control translation or citation policy. Broader provider extension support remains tracked in issue #78.

Verification

Verified at dcbf1a9d2d146087e3c9a3d0667e7de50a8ab4ea:

• PYTHONPATH=python /home/akk/jiahuei/molto/.venv/bin/python -m pytest -q -o addopts='': 2,294 passed on Python 3.12.9, using the pinned SDK versions. Offline replay only; no live provider calls or cassette changes.
• uv run --extra lint ruff check --output-format github --config pyproject.toml .: passed.
• uv run --extra lint ruff format --diff --config pyproject.toml .: passed; 99 files already formatted.
• npx prettier@3.8 --check .: passed.
• git diff --check origin/main...HEAD: passed.
• GitHub CI: all five required checks passed, including tests on Python 3.12, 3.13, and 3.14.

[jiahuei]

Two confirmed P2 findings are attached inline. Reviewed commit c01c79a. The full offline suite passed (1,880 tests), focused transport-override/history tests passed (24 tests), and Ruff and Prettier checks passed. Both findings were reproduced locally; no live provider calls were made.

[jiahuei]

Review at 995304f

Reviewed the full diff against acf4abe (merge base with main). The two inline findings from my previous review (utils/web_search.py multi-query history, and effective history after extra_body.input) are addressed in 79b0c7a and are not repeated below.

Verification in a clean worktree at 995304f: python -m pytest (1920 passed, Python 3.12), ruff check, ruff format --diff and npx prettier --check . all clean, and tests/cassettes unchanged after the run. Every behavioural claim in the inline comments was reproduced at this commit and compared against origin/main (acf4abe) with identical request bodies over MockTransport.

Inline findings:

Severity	Location	Issue
P1	upstream/openai.py:1035	Native Chat passthrough rejects web_search_options (regression on an untranslated route)
P1	upstream/openai.py:932	incomplete_details: null silently becomes pause on every route
P1	utils/web_search.py:32	Citation splitting emits empty and whitespace-only text blocks that Anthropic rejects
P1	adapters/openai.py:1419	Search-gated reasoning drop hides an unrepresentable signature
P2	upstream/openai.py:936	Search count belongs on Usage, not three extra_fields channels
P2	utils/web_search.py:157	UnsupportedInput raised while normalizing provider output
P3	upstream/openai.py:676	tool["type"] KeyError on raw extra_body tools
P3	upstream/anthropic.py:440	Typo in an error message, plus duplicated range validation
P3	utils/test.py:55	Unrelated churn (_reject_input_extras as well)
P3	utils/web_search.py:1	Module ownership not stated in CONVENTIONS

Not a code finding: the four web_search upstream requests are genuine recordings, but per tests/cassettes/README.md the Messages response is rewritten from the Responses recording by scripts/match_anthropic_cassette_*.py. The Messages-upstream path therefore rests on the Responses recording for citations_delta, server_tool_use input_json_delta and usage.server_tool_use. One real Anthropic search recording would de-risk that direction, and would also confirm that the reduced history ("N/A" placeholders, empty titles) is accepted by the API.