forked from retoor/devplacepy
Docs: update root CLAUDE.md, DEFINITION_OF_DONE, README and gitignore
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -774,6 +774,7 @@ Configuration on the Services tab:
|
||||
| `gateway_rsearch_cost_per_call` | 0.0 | Flat cost attributed to each external `rsearch` call (web search / AI answer / chat / image describe), recorded under backend `rsearch` so external AI spend appears in AI usage |
|
||||
| `gateway_max_retries` / `gateway_retry_backoff_ms` | 2 / 250 | Retry attempts and linear backoff on timeout, connection error, or upstream 5xx |
|
||||
| `gateway_circuit_threshold` / `gateway_circuit_cooldown_seconds` | 5 / 30 | Consecutive failures before the circuit breaker opens, and its cooldown |
|
||||
| `gateway_target_circuit_enabled` | on | Per-model circuit breaker: one upstream error opens it for the cooldown and notifies every admin with the error details |
|
||||
| `gateway_usage_retention_hours` | 720 | How long per-call usage rows are kept before pruning (30 days) |
|
||||
| `gateway_model_context_map` | JSON map | Model to max-context-tokens map for context-window utilization tracking |
|
||||
|
||||
@@ -798,7 +799,7 @@ token usage with averages and percentiles (p50/p90/p95/p99), latency (upstream r
|
||||
gateway overhead, semaphore queue wait, connection establishment), error rates by category,
|
||||
cost (per model, per caller, input vs output, projected monthly burn, caching savings),
|
||||
caller behavior, and an hourly breakdown - divided and combined. The gateway also retries
|
||||
transient upstream failures and opens a circuit breaker after repeated failures. Time to
|
||||
transient upstream failures, opens a circuit breaker after repeated failures, and opens a per-model breaker on the first error of a model and notifies every admin. Time to
|
||||
first token and inter-token latency are not reported because the gateway forwards
|
||||
non-streaming to the upstream.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user