Skip to content

[v2]

The orchestration layer for selector-routed requests (on by default).

Key Type Default Description
v2.enabled bool true enabled runs selector-routed requests through the orchestration layer (default true; an absent key also means on). false returns to the v1 router.
v2.orchestration.default_strategy string "auto" default_strategy applies to orchestrated requests when no request, session, or global control supplies a strategy. Default “auto” (static balanced during cold start).
v2.orchestration.max_calls integer 0 max_calls caps total upstream model calls per request. 0 = no numeric cap.
v2.orchestration.max_cost_usd number 0 max_cost_usd caps total actual per-request orchestration cost. 0 = no cap.
v2.orchestration.max_duration_ms integer 0 max_duration_ms caps total orchestration wall time. 0 = no cap.
v2.orchestration.specialists array of string specialists is the ordered candidate list (“provider/model”) for the single bounded escalation stage. Empty disables specialist escalation.
[[v2.orchestration.specialist_rules]] array of tables specialist_rules are capability/tier selection rules that resolve escalation candidates dynamically from the live discovered catalog at plan time, so specialist targets survive provider model churn without config edits. Rules append to (never replace) the pinned specialists list; both forms may coexist. Empty plus empty specialists disables specialist escalation.
v2.orchestration.specialist_rules[].min_quality number min_quality is the routing quality-score floor within [0, 1]. 0 = no floor.
v2.orchestration.specialist_rules[].max_cost_usd number max_cost_usd is the per-request effective-cost ceiling. 0 = no ceiling.
v2.orchestration.specialist_rules[].require_tools bool require_tools / require_json demand the capability flag even when the request itself did not.
v2.orchestration.specialist_rules[].require_json bool require_tools / require_json demand the capability flag even when the request itself did not.
v2.orchestration.specialist_rules[].min_context_window integer min_context_window demands a known context window of at least this many tokens; candidates with an unknown (0) context window fail the rule. 0 = no constraint.
v2.orchestration.specialist_rules[].locality string locality constrains executor locality: “any” (or empty), “local”, or “cloud”. The privacy/budget local-only hard gate still binds first: “cloud” under a local-only constraint resolves nothing.
v2.orchestration.trace_retention integer 0 trace_retention bounds the orchestration trace store entry count. 0 selects the built-in default (512).
v2.orchestration.stream_buffer_max_bytes integer 0 stream_buffer_max_bytes / stream_buffer_max_ms bound the streaming acceptance buffer. The buffer is always bounded: 0 selects the built-in safety bounds (2 MiB / 120000 ms).
v2.orchestration.stream_buffer_max_ms integer 0 stream_buffer_max_bytes / stream_buffer_max_ms bound the streaming acceptance buffer. The buffer is always bounded: 0 selects the built-in safety bounds (2 MiB / 120000 ms).
v2.orchestration.session_ttl_ms integer 0 session_ttl_ms bounds the v2 routing session state lifetime. 0 selects the built-in 24h proposal; Expired state degrades to stateless service.
v2.orchestration.cache_warm_window_ms integer 0 cache_warm_window_ms bounds the own-traffic prompt-cache warmth estimate: an executor used within the window counts as warm. 0 selects the built-in 300000 ms proposal.
v2.orchestration.backend_affinity_min_cached_tokens integer 0 backend_affinity_min_cached_tokens is the materiality floor of aggregator backend affinity: the session’s estimated cached input must reach it before Prism asks the aggregator for the backend that served the last use. 0 selects the built-in 1024.
v2.orchestration.switch_friction_usd number 0 switch_friction_usd is the operator-configured switching-friction component of the switch penalty. Default 0 (no friction).
v2.orchestration.cache_savings_fraction number 0 cache_savings_fraction estimates the fraction of an executor’s per-request cost saved by a warm prompt cache (the cache-loss component of the switch penalty). 0 selects the built-in 0.4 proposal; values must stay within [0, 1].
v2.orchestration.effort_escalation_models array of string effort_escalation_models lists candidate IDs (“provider/model”) whose reasoning effort Prism may raise before switching models. Empty disables effort escalation.
v2.orchestration.outcome_retention integer 0 outcome_retention bounds the outcome event store entry count. 0 selects the built-in default (4096).