[ethan-scitix]

Type

• feature — new benchmark, task, or capability

Summary

• New dialect: google_interactions for Gemini's Interactions API (POST /{v1,v1beta}/interactions), beside google_genai (generateContent), as [Feature]: google_interactions dialect for the Gemini Interactions API #152 proposes. Interactions has no candidate count, top_k, penalties or logprobs, so those requests are refused before any I/O and stay on google_genai.
• store: false on every request. Google otherwise retains prompt and reply server-side by default (55 days paid). The final HTTP body is re-checked for store: false, the bound model, and the absence of background / previous_interaction_id / agent. store is recorded in request_params, and no dialect option may override it.
• Stateless multi-turn, because store: false disables previous_interaction_id. History is resent as user_input / model_output / function_call / function_result steps. Signed thought steps round-trip verbatim through opaque_continuation.
• Supported: text and image input, function tools and tool choice, JSON output via response_format, thinking_level, thinking_summaries, temperature, top_p, max_tokens, stop, seed, and streaming (step.start / step.delta / step.stop, with streamed function arguments reassembled). Hosted tools are not lifted yet and are refused.
• Termination: the interaction status is the finish reason, recorded verbatim. incomplete is added to the n_truncated spellings in core/utils/meta.py. As the issue notes, the record cannot tell a budget stop from a policy stop. requires_action is a function call; failed fails the sample with Google's message.
• Usage: input = total_input_tokens + total_tool_use_tokens and output = total_output_tokens + total_thought_tokens, the same split as google_genai. reasoning_tokens stays a breakdown of output_tokens (cf. the feat(models): add native Google GenAI generateContent dialect #147 accounting fix).
• Per [Feature]: Gemini model-family limits via model profiles; split google_genai.py #151: no model-family request limits, and no provider_enforced_capabilities, since none was measured on this endpoint.
• Answers to the issue's open questions, from Google's own typed models in google-genai 2.33.0 (_gaos/types/interactions):
  • GenerationConfig does have temperature and top_p, though both are marked deprecated there.
  • Usage.total_output_tokens documents only "generated responses", with total_thought_tokens separate. Folding thoughts in mirrors generateContent's verified behaviour but is not live-verified here.
• No identity or on-disk format change. RESULT_SCHEMA_VERSION and IDENTITY_SCHEMA are unchanged. No existing dialect emits a bare incomplete finish reason (openai_responses records incomplete_details.reason), so existing counts are unaffected.

Risk: the dialect is not live-verified. The scitix gateway does not proxy /interactions: POST .../v1beta/interactions and .../v1/interactions both return 404, while generateContent on the same key returns 200; probed 2026-10-27. Wire shapes come from the API reference and the SDK models. A first run against generativelanguage.googleapis.com should confirm the usage split and the streaming event order.

Related Issues

Closes #152. Refs #147, #151.

Test Plan

Automated

• Lint/format clean (ruff check && ruff format --check): all checks passed, 687 files formatted
• Type check clean: bare ty check (sieval + tests), all checks passed
• Unit + integration: pytest tests/unit tests/integration -n 8, 10127 passed, 1 skipped
• New tests/unit/core/models/dialects/test_google_interactions.py: 153 tests; coverage of google_interactions.py is 96%
• Registry tests list the new dialect; the CLI dry-run reconcile test is parametrized over both Google dialects × {"", /v1, /v1beta}; the CLI api_key_env test asserts x-goog-api-key on the wire
• scripts/check_preflight.py, check_layer_imports.py, sync_package_stubs.py --check: all pass

Manual

• Probed the scitix gateway: /interactions 404 on both versions; generateContent control 200
• Live call against generativelanguage.googleapis.com with a Gemini key: not done, since no key is available in this environment

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring
• No new upper-layer dependencies added to core/
• Deleted code verified — no remaining call sites depend on it (none deleted)

🤖 Generated with Claude Code

[ethan-scitix]

Superseded by #263, which covers everything here (same pre-I/O refusals, store: false guard, stateless opaque continuation, usage split, and the incomplete truncation spelling) and adds hosted tools (Google Search, URL context), live validation against generativelanguage.googleapis.com, and the body.stream normalization in request_definition that this PR was missing (streaming and non-streaming Interactions requests would otherwise hash to the same request definition).

Note: this PR said Closes #152; #263 only says Refs #152, so #152 needs to be closed by #263 or manually.