# Standalone live graph

Run `node tools/graph/run-live.mjs --session engagements/<fresh-session> --engagement engagements/<fresh-session>/engagement.yaml --azure-config engagements/<fresh-session>/.azure --concurrency 3 --max-reads-per-domain 3` after fresh scoped identity preflight creates `evidence/preflight-manifest.json`. The session and credential directory must be local ignored engagement artifacts. Successful sessions cannot be replayed as new live work.

Requires Node, PowerShell, Azure CLI/Resource Graph, and an authenticated Claude CLI. The installed native model default is used. Native contexts have no executable tools, no MCP servers, no session persistence, and no inherited settings sources. Structured output is validated locally. Model-selected reads are resource IDs from a finite host catalog, never shell commands. The host validates scope and applies the unchanged shared guard before every Azure read. Errors withhold raw subprocess payloads.

The canonical graph awaits actual asynchronous dispatch: scope, same-environment memory, preflight/inventory, shared security context, scoped bounded parallel specialists, reducer, native evaluator with at most two revisions, independent native judge with fresh ARM verification, correlation, native narrative/report rendering, and evidence-gated debrief. Missing or failed handlers cannot produce completion. Every candidate receives exactly one explicit verdict; missing verdicts fail closed.

This is a bounded configuration assessment. The catalog currently covers selected ARM fields for storage accounts, Key Vaults, web sites, registries, AKS, network security groups, VMs and AI accounts. Other in-scope domains still receive actual native dispatch and must disclose unavailable coverage. Default selection is at most three resources per specialist; `--max-reads-per-domain` accepts 1–20. It does not run all repository checks or certify the environment secure. Metadata nulls are unknown. No data-plane reads, probes, credentials, secret values, arbitrary URLs or mutations are available to models.

`evidence/security-context.json` persists the shared reference/status handoff passed to both specialist model calls. An inventory path is referenced, not verified configuration evidence; missing signal families remain unavailable. The canonical graph applies both domain and resource-type scope before dispatch. The standalone live scope loader accepts supported read-only domain selections, but rejects narrowed resource-group/type scopes and nonempty exclusions because its inventory cannot honor them. Explicit email-security, external-vuln, and authorization-attack-path domain selections are unsupported and fail closed; an omitted or empty domain list selects the default read-only roster. See the [scope parser contract](../powershell/README.md). Actual model exchange telemetry measures the requests containing that context without logging prompt bodies.

`runs/live-events.jsonl` contains actual lifecycle, handoff, guarded-tool and memory events with `mode: live` and a unique run ID. It never includes private reasoning or raw command output. `evidence/coverage-*.json` distinguishes available, inspected and unassessed scope. Native findings and judge evidence remain in the ignored session. Correlation is a separate narrative relationship artifact, not asserted exploit topology. `reports/report.html` renders schema-validated findings.

Memory retrieval is consumed in each matching specialist context. Debrief requires both raw and judged artifacts with matching finding attribution and hashes. Corroboration is not measured learning improvement; no improvement is claimed by this runner.

Tests: `node --test tools/graph/live-runtime.test.mjs tools/memory/workshop-memory.test.mjs`.
