fix: scope structured provider failures to request - #867
Conversation
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
b49d6ff
into
fix/provider-backed-embedding-batch
| admitted = self._request_admitted_agents.get() | ||
| if admitted is not None: | ||
| candidates = [agent for agent in candidates if agent.id in admitted] |
There was a problem hiding this comment.
🟡 Semantic-affinity embedding selection disabled during conduct
The request-admitted set, built only from structured chat readiness, is applied by _ranked_agents to the re-entrant embedding lookup reached via _embedding_agent_id. A dedicated embedding provider is excluded, so select_capability_agent("embedding") fails and semantic-affinity ranking silently falls back to declaration order for the whole conducted request.
Prompt for agents
The request-scoped admitted-agent filter added in _ranked_agents (contextual_orchestrator/orchestrator.py around lines 4414-4416) is intended to restrict chat provider selection during a conducted request to structured-chat-ready agents. However _ranked_agents is re-entrant: _semantic_affinities -> _embed_cached -> _embedding_agent_id -> select_capability_agent -> _capability_agents calls _ranked_agents again with capability="embedding", chat_only=False. Because the admitted set only contains structured-chat-ready (or mock) agents, a dedicated embedding provider is filtered out, making embedding selection fail and disabling semantic-affinity ranking for the whole conduct request. Consider scoping the admitted filter to chat-role ranking only (e.g. only apply it when chat_only is True, or when the capability/role is a chat role), so non-chat capability resolution like embeddings is not incorrectly gated by structured chat readiness.
Was this helpful? React with 👍 or 👎 to provide feedback.
| admitted = self._structured_admitted_agent_ids() | ||
| if not admitted: | ||
| raise NoViableAgentError( | ||
| retry_after_seconds=max(1, math.ceil(self.circuit_reset_seconds)) | ||
| ) | ||
| token = self._request_failed_agents.set(set()) | ||
| admitted_token = self._request_admitted_agents.set(admitted) | ||
| try: | ||
| return self._conduct_request( | ||
| messages, model_name=model_name, progress=progress | ||
| ) | ||
| finally: | ||
| self._request_admitted_agents.reset(admitted_token) | ||
| self._request_failed_agents.reset(token) |
There was a problem hiding this comment.
🔍 ****
The conduct gate raises NoViableAgentError only when the entire admitted set is empty. For an explicit model_name, _plan uses _requested_agent directly and _failover_candidates prepends that primary before the admitted filter shapes only the failover pool. A conduct request naming a real, not-structured-ready model can still reach it whenever any other admitted agent keeps the set non-empty.
Was this helpful? React with 👍 or 👎 to provide feedback.
1 participant
Summary
no_viable_agentretry contract when the declared set is exhaustedVerification
uv run pytest -q tests/test_provider_reliability.py tests/test_openai_passthrough.py tests/test_model_judge.py(75 passed)uv run python -m compileall -q contextual_orchestratorStacked on the delivered provider-backed embedding owner branch so the combined runtime tree remains reproducible.