feat(interactions): add previous_interaction_id for multi-turn conversations

[leseb]

What does this PR do?

Adds previous_interaction_id support to the Google Interactions API, enabling server-side conversation state for multi-turn interactions. Without this, every request starts fresh and ADK agents cannot maintain context across turns.

When a request includes previous_interaction_id, the server fetches the stored conversation history from the SQL backend, prepends it (including the model's prior response) to the new input, and calls inference with the full context. Both streaming and non-streaming responses are persisted for future chaining.

Key additions:

• InteractionsStore class (src/ogx/providers/utils/interactions/) following the ResponsesStore pattern with AuthorizedSqlStore and access control policy
• InteractionsConfig with SqlStoreReference for configurable SQL backend (SQLite/Postgres)
• _build_messages() method for conversation history reconstruction, supporting multi-hop chaining
• Updated Google Interactions API conformance score (39.0% → 40.2%)

Test Plan

55 unit tests pass (37 existing + 5 new previous_interaction_id + 13 shape/passthrough tests). All 39 pre-commit hooks pass.

# Unit tests
uv run pytest tests/unit/providers/inline/interactions/ -x --tb=short -q

# End-to-end validation with Google GenAI SDK
OLLAMA_URL=http://localhost:11434/v1 uv run --extra starter ogx stack run starter --port 8321
uv run scripts/test_interactions_api.py --base-url http://localhost:8321 --model ollama/llama3.2:3b

Test 5 validates conversation chaining: creates a first interaction establishing context ("My name is Alice"), then chains from it with previous_interaction_id and verifies the model remembers the name.

[mergify]

This pull request has merge conflicts that must be resolved before it can be merged. @leseb please rebase it. https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[github-actions]

✱ Stainless preview builds

This PR will update the llama-stack-client SDKs with the following commit message.

feat(interactions): add previous_interaction_id for multi-turn conversations

⚠️ llama-stack-client-openapi studio · code

Your SDK build had at least one "warning" diagnostic.
generate ⚠️

⚠️ llama-stack-client-python studio · code

Your SDK build had at least one "warning" diagnostic.
generate ⚠️ → build ⏭️ → lint ⏭️ → test ✅

⚠️ llama-stack-client-node studio · code

Your SDK build had at least one "warning" diagnostic.
generate ⚠️ → build ⏭️ → lint ⏭️ → test ✅

⚠️ llama-stack-client-go studio · conflict

Your SDK build had at least one warning diagnostic.

---

This comment is auto-generated by GitHub Actions and is automatically kept up to date as you push.
If you push custom code to the preview branch, re-run this workflow to update the comment.
Last updated: 2026-05-12 08:00:20 UTC

[cdoern]

lgtm, one question on some removed tests

[cdoern]

lgtm

[devin-ai-integration]

Devin Review found 2 potential issues.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🔴 system_instruction silently dropped when previous_interaction_id is provided

In _build_messages at src/ogx/providers/inline/interactions/impl.py:174, when previous_interaction_id is set, self._convert_input_to_openai(None, request.input) is called with None for system_instruction, unconditionally discarding any system_instruction the user provided on the current request. If a user sends both previous_interaction_id and system_instruction, the system instruction is silently ignored with no error or warning. This leads to incorrect behavior where the model doesn't receive the user's intended system prompt.

Prompt for agents

In _build_messages (impl.py around line 163-175), when previous_interaction_id is provided, the system_instruction from the current request is unconditionally dropped by passing None to _convert_input_to_openai on line 174. This silently discards user-provided data.

Two possible approaches:
1. If overriding system_instruction during chaining is intentional behavior, validate that system_instruction is None when previous_interaction_id is set, and raise a ValueError if both are provided so the user gets an explicit error instead of silent data loss.
2. If system_instruction should be allowed during chaining, pass request.system_instruction to _convert_input_to_openai on line 174 instead of None. You may need to handle the case where the stored messages already contain a system message from the previous interaction (e.g., replace it or prepend a new one).

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔴 Error message violates AGENTS.md "Failed to ..." prefix rule

AGENTS.md mandates: "Error messages must be prefixed with 'Failed to ...'". The ValueError raised at src/ogx/providers/inline/interactions/impl.py:168-170 uses the message "Interaction '...' not found. Cannot chain from a non-existent interaction." which does not follow this required format.

Suggested change

	raise ValueError(
	f"Interaction '{request.previous_interaction_id}' not found. "
	"Cannot chain from a non-existent interaction."
	raise ValueError(
	f"Failed to chain interaction: '{request.previous_interaction_id}' not found."
	)

---

Was this helpful? React with 👍 or 👎 to provide feedback.