Weisgnosis OS — Durable AI Process Runtime on Kubernetes

Typed · auditable Temporal · durable gVisor · isolated

01 / 08

failure.modes

Trapped in 3 places

The process exists only in spreadsheets, email, and one engineer's memory. Unversioned, unobservable, never run the same way twice.

0 lines of replay

Chat-AI returns prose with no types and no log. Ask "why did it decide that?" six months later — there is nothing to inspect.

1 shared kernel

Agent frameworks run model-written tool calls on the host kernel. One bad call reaches your infrastructure, so it never ships.

02 / 08

system.model

Typed graphs

18 node types · 3 edge kinds

Versioned and hash-pinned per run. Graphs compose and nest as subflows, so one reviewed flow drops into the next.

Eval-context flow

0 manual ports · 1 immutable context

Every node receives the full EvalContext as JSON and merges its output into the next snapshot. You design logic, not plumbing.

Durable execution

crash @ hour 2 → resume hour 2

A Temporal workflow interprets the graph deterministically. It survives worker restarts, network failures, and retries.

03 / 08

telemetry.live

thinking tool-calling complete dispatch · external

LIVE · 1s SSE REPLAY · 5,000-event ring

04 / 08

runtime.contract

01 · business_analyst

internal · in-process

→ RequirementSpec — scoped tasks + acceptance criteria, typed.

02 · developer / OpenClaw

external · gVisor sandbox

→ CodeArtifact — runs untrusted, returns a typed diff + build target.

03 · reviewer / QA

internal · in-process

→ ReviewVerdict — reads the same context, returns pass / fail + findings, typed.

One contract: both classes call submit_artifact with tool_choice=required. Every output is decoded as a typed artifact by construction — never regex-parsed.

05 / 08

isolation.proof

1

gVisor user-space kernel

/proc/version = 4.19.0-gvisor · syscalls intercepted

2

Hardened securityContext

runAsUser=1000 · readOnlyRootFilesystem · cap drop ALL

3

NetworkPolicy deny-floor

egress DNS + 443 only · blocks 10/8, 169.254.169.254

4

ghostunnel mTLS

FQDN-SAN · CA-pinned · loopback-only, bare-IP rejected

5

Per-tenant secrets

SecretKeyRef · deleted on teardown · never in logs

$ echo x > /rootfs/xPermission denied

$ nc 192.168.1.10 22BLOCKED

$ nc 169.254.169.254 80BLOCKED

$ capsh --print | grep sys_admindropped

$ nc 1.1.1.1 443OK

06 / 08

evidence.runs

run db3c9e8a

Research → OpenClaw → QA

5 steps, 22,135 tokens. QA caught a real burst-formula error in the brief and returned 10 authoritative citations.

23 steps · 95 events

Multi-topic research

Parallel branches with coordination rounds, fully replayed from nodeOutputs history.

spec → shipped

SDLC reference flow

6 subflows from spec to shipped, with 1 human-in-the-loop gate.

< 10 min

idea → running process

< 5 min

50-page SOP → workflow

hour 2

crash → resume

$0.47

per run, observable

07 / 08

08 / 08