nx_watch.nx

nx_watch.nx -- SOVEREIGN JOB / LIVENESS MONITOR (operator 2026-07-16: "our monitors are garbage -- WSL-polling Claude-Code Monitor flakes; build a STATE-OF-THE-ART Nishi monitor as MCP + RACI + API + workstream capability"). ROOT CAUSE of the flake: watching = polling WSL for "is it done" through fragile shell `until` conditions (screen-noise, exit-code weirdness, compound tests). THE SOVEREIGN FIX: a job writes ITS OWN terminal state into the sovereign store; nx_watch reports it DETERMINISTICALLY from the store -- zero polling, zero WSL dependence, works cross-session. mark <ns> <id> <state> job records its state (RUNNING at start, GREEN/RED/DONE at end) + ts beat <ns> <id> heartbeat for a long job (liveness; staleness = crash signal) status <ns> <id> print state + age_ms; EXIT CODE = verdict (0 GREEN/DONE, 1 RED, 2 RUNNING, 3 STALLED[running+stale], 4 ABSENT) -- scriptable, no grep wait <ns> <id> <max_ms> block until terminal (GREEN/RED/DONE) or timeout; sovereign sleep poll Store = seg_store via nx_sov_ledger (knowledge/store/watch_<ns>, key=id -> "STATE|us"). The staleness verdict is PURE (stored ts vs now) = gate-testable without a real clock race. RACI: supervisor|monitor (owns this). MCP: status/list = read verbs (broad-expose tier). license_tier: ORIGINAL