feat(deployments): registry, deterministic probes, breach signals to Joe

[alari76]

Summary

Phase C2 of the Joe value-expansion plan (§4, §5.1 partially): Joe's first non-repo value surface. Registered deployments are probed deterministically; the agent enters only on a breach.

Design

• Registry (~/.codekin/deployments.json, 0600): deployments with http (status/latency/TLS days), pm2 (status/restarts/memory), and disk (free %) probes. repoPath links a deployment to its source repo for future incident→recent-merges correlation. GET /api/deployments/discover proposes pm2 processes — never auto-enrolls.
• Two-layer architecture: probes are plain code sampled into deployment_samples (runs.db, 30-day retention); no LLM in the hot path. Breaches flow to Joe as notifications via the durable signal queue from C1 (feat(workflows): durable signals — at-least-once event queue in the trigger engine #605) — at-least-once, deduped while pending.
• Transition-based alerting: ok → breached publishes one probe-breach signal; breached → ok publishes probe-recovered; a pm2 restart-count increase publishes a one-off probe-event. A persistent condition alerts once, not every 5 minutes.
• registerTickTask on the engine: periodic work (probe sampling here) rides the dispatch tick + heartbeat instead of adding another setInterval — non-overlapping, error-isolated. This is the A0 promise of folding scattered loops onto one clock, now available as a mechanism.
• Sudo-free by policy (per the proposal §5.2): fetch, pm2 jlist, df as the unprivileged user; probe failures are themselves breaches so a broken probe is visible.
• Joe integration: list_deployments / get_deployment_samples MCP tools (allowlisted, template v6 tells Joe to check state+history before reacting and reiterates propose-only for privileged host actions).

Phase D (breach → trust-gated diagnostic child + incident reports) builds directly on the probe-breach signal payload.

Testing

• server/deployment-monitor.test.ts (7 tests): sampling+metrics, disabled skip, breach-only-on-transition + recovery, first-sample breach, restart events, history ordering + retention pruning.
• trigger-dispatch.test.ts +1: tick tasks run at interval, non-overlapping, error-isolated.
• Full suite: 3043 tests pass; build + lint clean.

🤖 Generated with Claude Code