cradle-observability-debuggerlisted

# Observability Debugger Use this skill when a chat turn behaves unexpectedly and you need concrete local evidence. ## Data sources - `observability_events` — structured error/warn/info events - `observability_incidents` — deduplicated grouped events; `TURN_STREAM_FAILED` aggregates by code unless a producer supplies a narrower dedupe key - `backend_run_snapshots` — Cradle-owned runtime-neutral run envelope with millisecond lifecycle timestamps - `backend_run_snapshot_events` — ordered snapshot event stream for stable harness phases such as model text/reasoning boundaries, tool input/output availability, usage, and finalization - `GET /observability/runtime-snapshot` — live server/runtime resource view that includes server process health, active chat runs, replay buffers, provider runtime hosts, PTY resources, Chronicle daemon resources, latest desktop samples, renderer/browser drill-downs, and observability queue health - OpenTelemetry metrics — low-cardinality gauges/counters derived from runtime samples; use only as trend/correlation evidence during debugging - **Server logs** — pino-structured JSON written to `{CRADLE_DATA_DIR}/server.log` (or `$CRADLE_LOG_FILE`) Observability and run snapshot rows are forensic records. Session/run/message foreign keys may be `NULL` after source rows are deleted; do not treat a null FK as malformed data. Snapshot retention is controlled by `CRADLE_CHAT_RUN_SNAPSHOT_RETENTION_DAYS` (default 30, `0` disables pruning). Keep these snapsh