[jack-scitix-ai]

Type

• feature — new capability

Summary

• Add the native google_interactions dialect alongside google_genai, using the existing provider-neutral HTTP connection and dialect registry.
• Support text, images, function tools, structured JSON output, Google Search, URL context, citations, and streaming.
• Require Boolean store: false on every request. Support stateless, client-side opaque continuation through signed native-step history, system-instruction replay, and typed persistence.
• Validate tool-result content and annotations before HTTP, infer function-result names from matching calls, and independently audit serialized input and opaque-history replay.
• Preserve provider text and thought summaries verbatim. Accept signature-only thoughts without weakening content validation or terminal-step reconciliation.
• Normalize usage accounting, preserve cumulative delta-usage fallback, and reject malformed provided counters rather than silently treating them as missing.
• Normalize HTTP-envelope body.stream in shared request-definition hashing, distinguishing Interactions streaming from non-streaming requests while preserving existing GenerateContent path/query handling.
• Keep model-profile capability gates intact and expose consistent legacy capability metadata for function calling, hosted tools, and structured output.
• Introduce no new dependencies or result-directory layout changes. RESULT_SCHEMA_VERSION remains 1; upstream identity schema 5 is preserved.

Related Issues

Closes #152. Refs #25.

Related implementation: #249.

Test Plan

Automated

• Lint/format clean (ruff check && ruff format --check)
• Type check clean (ty check or mypy --strict)
• Unit tests pass (pdm run pytest)

Manual

• Complete the real-service validation matrix against generativelanguage.googleapis.com on October 28, 2026.

Tested model: gemini-3.6-flash, reported as 3.6-flash-07-2026 in model metadata. Both POST /v1beta/interactions and POST /v1/interactions returned successful responses.

The final matrix passed 30 verification scenarios, representing 33 generation requests, excluding earlier diagnostic/probe iterations. Guard-only scenarios deliberately made no provider request.

Area	Verification
Text and streaming	Non-streaming and streaming replies, completed and budget-limited incomplete responses, and explicit stop sequences
Sampling and reasoning	Temperature, top-p, and seed accepted; minimal, low, medium, and high effort; summary none and auto; verbatim multi-item summaries and signature-only thoughts
Opaque continuation	Two-turn non-streaming and streaming replay, exact native-history prefixes, and system-instruction preservation
Persistence	Serialize a live reasoning result, reload it with type metadata, and use its opaque continuation in a subsequent real request
Function calling	Auto, required, none, and named-function choices; multiple streamed calls with separate tool-result messages
Function results	String, object, annotated text, type-only citation annotation, image content, and is_error: true
Structured output and images	Schema-shaped JSON in both response modes; inline and URI image inputs; image content returned by a function
Hosted tools	Google Search and URL context in both response modes, with actual call/result evidence and citations
Accounting	Observed token arithmetic includes thoughts in IR output and tool-use tokens in IR input; cache counters remain breakdowns
Public binding and guards	Ordinary YAML binding followed by Model.agenerate(ChatInput(...)) in both response modes, without injected capability outcomes; invalid storage/stateful/background settings rejected before HTTP

Every transmitted request was checked for Boolean store: false and the absence of previous_interaction_id, background, and agent.

Detailed evidence, accounting examples, and validation limits are recorded in tests/unit/core/models/dialects/fixtures/google_interactions/live_validation.md:1.

Rebase preserved the validated Interactions dialect source and captured JSON/SSE bytes. The automated checks above were rerun afterward; no additional live API requests were made for the rebase verification.

Live-test findings and validation scope

• Live responses sometimes omit visible summaries under thinking_summaries: auto while providing a thought signature. This is accepted and covered by four live SSE fixtures, while contradictory content still fails.
• One response after an errored tool result emitted a function call despite tool_choice: none. The client correctly rejected it; the sanitized response is retained as a negative regression fixture.
• Live terminal events omitted steps. Optional terminal-step reconciliation and metadata-only usage fallback are covered by deterministic tests, not claimed as observed live behaviors.
• Accepted sampling parameters do not establish statistical seed determinism. Live validation is specific to the tested model and service versions.
• Capability-dependent public Model requests retain the existing model-profile requirements. Standalone dialect tests do not bypass those gates.

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

[ethan-scitix]

#249 has been closed as superseded by this PR. It carried Closes #152; could you change Refs #25, #152. in the description to Closes #152. Refs #25. so #152 closes when this merges?