Providers Models and AI Monitor

Providers, Models and AI Monitor

AI RV Harness routes model requests through explicit provider configurations and credential bindings. The selected identity remains bound to its trained model route unless the Profile contains another trained identity that the user chooses.

Provider transport

The current provider layer supports the provider families exposed by the application, including OpenRouter, OpenAI, Google, Anthropic, Z.AI, DeepSeek, Mistral, Blackbox and Custom OpenAI-compatible endpoints.

Transport retry is limited and classified. It handles retryable network and provider failures such as temporary rate limits. It is separate from semantic or analytical output recovery.

Continuation state

Contracted OpenRouter, Google and Anthropic paths can preserve provider-native continuation state. It is:

• stored separately from visible transcript text;
• bound to the exact message/event and route fingerprint;
• size and structure validated before persistence and replay;
• replayed only on a compatible route.

For OpenRouter streaming, consecutive compatible reasoning text/summary fragments are canonicalized into logical blocks. Opaque encrypted or signature-bearing blocks are preserved exactly. Safety block limits apply after canonicalization.

Reasoning controls

Known exact models use verified reasoning contracts. Unknown models receive only capabilities advertised by the provider; the application does not invent reasoning levels. Dynamic OpenRouter aliases use current provider metadata rather than a permanently pinned historical contract.

Output limits

Conversation and RV Session output limits are configured separately in Settings. Training and post-Reveal workflows use their own bounded analytical and learning contracts.

AI Monitor

The AI Monitor guides supported automatic sessions and evaluates its own work after Reveal. It cannot add new evidence to the sealed blind record. Monitor versioning remains separate from Viewer prompt versioning.

Common provider messages

• 429 Too Many Requests: temporary upstream/provider pool limit; retry later, use BYOK where supported or select another route.
• finish_reason: length: the provider reached its output limit; eligible analytical work may receive one controlled recovery attempt.
• connection reset while reading the body: the remote host closed the response stream; this is a transport failure, not proof of a model or session logic error.