Sidekick

11 slides · the Mission Supervisor for DevOps & data engineering

SIDEKICK1/11

Sidekick is a Mission Supervisor, not a chatbot

  • It runs governed missions — deploy, teardown, cost-audit, cross-app launch — through the kernel.
  • A DevOps and data-engineering copilot that acts, gated, and shows its work.
  • Every action is a planned, simulated, approved, replayable mission.
  • The human surface is the Projects Control Plane; the engine underneath is the Mission Runtime.
SIDEKICK2/11

The stack: three runtimes under every mission

  • Three runtimes underneath: the Discovery Runtime proposes work, the Mission Runtime executes, verifies, and replays it, and the Context Runtime plans what evidence each step needs.
  • Every question it asks — which stack? what's over-provisioned? is this allowed? — is a context question first.
  • Sidekick is an application on top: it states needs and delegates, never re-implementing retrieval, model choice, or any runtime primitive.
  • One governance model underneath, whether the mission is a deploy or a business process.
Mission — what you ask forDiscovery Runtimeproposes the workContext Runtimedecides every context needMission Runtimeplans · gates · runsproposeconsultexecute
SIDEKICK3/11

Deployment is a governed mission

  • A deploy, teardown, or cost-audit runs through the same runtime as any business process.
  • It inherits the same simulation gate, evidence ledger, approval, saga rollback, and replay.
  • One governance model, not a separate CI pipeline bolted on.
  • A failed deploy unwinds by the same saga that unwinds a failed launch.
SIDEKICK4/11

Preflight and the grounded stack advisor

  • Before acting, Sidekick analyses what a repo needs and checks dependencies.
  • A grounded five-tier stack advisor recommends where to run, priced from real data.
  • Credential onboarding guides vibecoders who lack cloud API keys.
  • No green light until preflight passes and the plan is understood.
SIDEKICK5/11

Ask at mission inception: security & observability

  • Security defaults and permission scoping are set at inception, not bolted on later.
  • The user is asked up front whether the deployment needs live monitoring afterwards.
  • If yes, Sidekick follows the continuous-operation routine.
  • Intent is captured before anything irreversible runs.
SIDEKICK6/11

Read with MCP, act with operators, teach with skills

  • Read/inspect -> MCP server: external system access, per-session tool schemas.
  • Procedure/know-how -> a Skill: SKILL.md, near-free until it triggers.
  • Policy-guarded mutation -> a custom operator: owned, gated, audited writes.
  • Irreversible actions live behind operators with dry-run, approval, and audit — never a raw connector.
SIDEKICK7/11

Live cluster reads

  • metrics-server for cluster inventory and utilisation.
  • Prometheus (PromQL) for performance signals; Loki (LogQL) for live logs.
  • Grafana dashboards embedded in the cockpit for the human view.
  • Sidekick inspects the real k3s cluster live during a mission, not a modeled stand-in.
SIDEKICK8/11

Continuous operations

  • It deploys as a standing container on your cluster (Helm) or a production server (compose) — not on your laptop, so the loop keeps running when your machine is off.
  • A standing monitor loop watches live signals — it is a daemon, not a one-shot mission.
  • On a real signal it spawns a governed response mission through the same kernel.
  • Each gate can alert a stakeholder for sign-off with a cockpit deep-link — the fast path from issue detected to human approves the fix.
PRODUCTION CLUSTER / SERVERworkloadsthe live deploymentSidekickstanding monitorread-onlystakeholderapproves via cockpitalertyour machine can power down — the loop lives on the infra
SIDEKICK9/11

The cost-audit discipline

  • cost_audit is a repeatable, schedulable mission — six priced dimensions, ordered.
  • Each finding is auto-applied where safe or gated for approval where it isn't.
  • Generalised from one $11k -> $1.9k anecdote into a standing practice.
  • The same EXPLAIN and evidence trail as any other mission.
SIDEKICK10/11

Why Sidekick

  • You describe the outcome; Sidekick plans, simulates, gates, executes, and can replay.
  • Every mission is explainable per node and reversible by saga.
  • One runtime for deploys, audits, and business processes — read the same way.
  • DevOps and data engineering as governed missions, not ad-hoc scripts.
SIDEKICK11/11

Whitepaper v10: the governed harness under the swarm

  • Every worker runs its model session through the governed harness (whitepaper v10) — the same ContextView and DereferenceEvent path — so swarm runs are replayable and auditable like any mission.
  • The manager reads worker outputs by reference from the shared object store, never by re-serializing them; each worker owns its work objects in an isolated worktree.
  • Teamwork is emergent coordination, not a predeclared orchestration topology: a supervisor stays the semantic authority and reconciles worker outputs via merge() on isolated worktrees when the integration context exceeds one window.
  • One normalized RuntimeEvent history covers live hooks and at-rest reconstruction — even for external coding agents — with a per-host coverage matrix marking where a pre-action gate makes enforcement real vs monitor-only.