okf-agentbus-ops

Opinionated SRE ops for AgentBus: edge-triggered health watchdog as a deterministic Python process (SREWatchdogAgent).

Why a separate package

Core (okf-agentbus) stays pure: routing, metrics, validate-config.
Opinionated policy (level transitions, cooldown, idempotency, hooks) lives here.

Install

pip install okf-agentbus-ops
# or path-install from monorepo during development:
# pip install -e packages/python/agentbus-ops

CLI (parity with bash P2)

# One-shot edge decision (no publish)
agentbus-ops watchdog --dry-run --json

# Live: publish SRE_STATUS only on healthy↔degraded↔critical edges
agentbus-ops watchdog --json

# Bootstrap seed silent (default); force first publish:
agentbus-ops watchdog --force-bootstrap-publish

# Attach compact metrics line
agentbus-ops watchdog --include-metrics

Exit codes match bash:

Health level is not mirrored as process exit (cron stays green on steady degraded).

Library

from agentbus_ops import SREWatchdogAgent

class MyWatchdog(SREWatchdogAgent):
    def on_critical_alert(self, cur, decision):
        # optional LLM / page — default is no-op
        pass

agent = SREWatchdogAgent(workspace="/path/to/okf")
decision = agent.run_once(dry_run=True)
print(decision.action, decision.level)

Probe (v0.1 MVP)

v0.1 wraps the coordination-root bash probe:

scripts/swarm_health_check.sh --json

Pure-Python probe port is deferred. Edge policy, state file, and publish path are native Python and share .agentbus/sre_last_state.json with the bash strangler so cron can switch without dual state.

State file

Same contract as sre-edge-triggered runbook:

.agentbus/sre_last_state.json under AGENTBUS_WORKSPACE.

Not in scope (v0.1)

• Autonomous restart daemon
• Default-on LLM every tick
• TypeScript package

License

MIT