Enterprise Readiness

What regulated teams get today (policy controls, audit trail, observability) and what is explicitly missing (SSO, SCIM, VPC).

Honest accounting first: NIKI has no SSO/SAML, no SCIM provisioning, no VPC/on-prem control plane, and no role-based seats. Those are major builds with external infrastructure, not configuration flags, and this page will say so until they exist. What follows is what regulated teams can rely on today.

What exists

Policy controls (`[permissions]`)

Modes (manual default, auto, dontask, bypass) plus per-run override: niki run "<description>" --permission-mode <mode>. Unknown mode strings fail closed to manual.
disable_worktree = true refuses the unisolated worktree backend outright, the governance kill-switch for fleets where local-process execution is unacceptable.
fail_closed_headless = true turns headless Ask into Deny instead of Allow-with-warning, for unattended runs where silent auto-approval is unacceptable. Default false (behavior-preserving); enable it in CI.
Command deny-lists cover the narrow global defaults (force-push forms, root-filesystem rm, mkfs, dd, pipe-to-shell patterns, and --no-verify) plus role-specific Git/filesystem denies. They do not generally block git reset --hard, rm -rf ~, fdisk, or git branch -D main.

Audit trail (per run, no extra setup)

Task output is stored under .niki/tasks/<id>/: report.md, changes.patch, per-agent artifacts/*.json, trace.jsonl, and manifest.json when snapshots are enabled. safety_proof.json appears only for a non-empty eligible branch, while artifacts/test_execution.json appears only when a test command resolves. Eval runs write eval-manifest.json under the eval output directory, not the task directory.

niki audit <task-id> emits partial or complete JSON bundles. Its complete flag checks only for a task record, report.md, and safety_proof.json; it does not require test execution, a trace, or a patch. Default snapshot retention is 14 days, after which old task directories are pruned.
niki run "<description>" --output-format json gives scripts the verdict/branch/cost contract; --bare strips ambient inputs (memory, MCP discovery, knowledge URLs) for reproducible runs.

Observability

--otel-endpoint (or OTEL_EXPORTER_OTLP_ENDPOINT) exports run spans as OTLP/HTTP to any collector (Jaeger, Grafana Tempo, Honeycomb). Every span carries niki.timeline = "derived"; offsets come from recorded latencies, never fabricated wall-clock.
Nightly eval gate (.github/workflows/nightly-eval.yml): keyless replay + disclosure manifest on every schedule, so regressions page you instead of your customers.

Supply chain

Tight deny.toml license allow-list, golangci-lint in CI, checksum-verified installers, and release channels via goreleaser. Outbound traffic depends on features: provider APIs, niki research, knowledge URLs, optional OTLP export, and remote MCP client traffic can all leave the machine. The current pipeline does not connect configured MCP servers.

What is missing (roadmap, not features)

Need	Status	Workaround today
SSO/SAML login	Not built	BYOK keys via supported env vars or `niki.toml`; short-lived tokens rotated externally
SCIM provisioning	Not built	niki.toml is a file; manage it with your existing config management
VPC / private runners	Not built	`DOCKER_HOST=ssh://…` runs containers on your own hosts (see Remote Daemon)
RBAC seats	Not built	OS users + file permissions on niki.toml and `.niki/`
Data residency	Provider-dependent	Use region-pinned `base_url`s; Ollama for fully local runs