Agent OS provides kernel-level governance for AI coding assistants running inside VS Code. It enforces policies in real time, audits every AI suggestion, and visualizes the health of your agent mesh through a set of sidebar panels and detail views.
Displays Service Level Objective health for the governance kernel. Four metric groups:
| Metric | Meaning |
|---|---|
| Availability | Percentage of successful governance evaluations over the current window. |
| Latency P50 / P95 / P99 | Response time percentiles for policy evaluation calls (milliseconds). |
| Compliance | Percentage of tool calls that passed policy evaluation without violations. |
| Trust Score | Mean and minimum trust scores across all registered agents (0--1000 scale). |
Click the panel header to open the SLO Detail view with burn-rate sparklines and error budget gauges.
Shows the agent mesh as a list of registered agents, protocol bridges (A2A, MCP, IATP), and delegation chains. Each agent entry displays its DID, trust score, and execution ring.
Click the panel header to open the Topology Detail view with a force-directed graph.
Scrollable list of recent governance events: tool calls evaluated, blocked, warned, or allowed. Each entry shows timestamp, action, agent DID, affected file, and severity badge.
Lists all active policy rules with their action (ALLOW / DENY / AUDIT / BLOCK), match pattern, evaluation count, and violation count for the current day.
Aggregate counters: total tool calls blocked, warnings issued, CMVK reviews triggered, and total log entries. Refreshes on the same tick as all other panels.
Live view of kernel internals: registered agents, active violations, saga checkpoints, and kernel uptime. Useful for diagnosing why a tool call was blocked or escalated.
Virtual filesystem browser showing the episodic memory kernel (EMK) contents. Navigate directories and inspect files stored by agents during execution.
Composite view combining SLO, topology, audit, and policy data in a tabbed interface. Provides a single-pane-of-glass overview of governance health. Tabs: Overview, SLO, Topology, Audit, Policy.
Full SLO view with:
- Availability and latency gauges against their targets.
- Error budget remaining bars for availability and latency.
- 24-point burn-rate sparkline showing consumption trend.
- Trust score distribution histogram (4 buckets: 0--250, 251--500, 501--750, 751--1000).
Force-directed graph of the agent mesh. Nodes are agents colored by trust tier. Edges represent delegation chains labeled with the delegated capability. Bridge status indicators show connected protocol bridges.
Table of all policy rules with columns: name, action, pattern, enabled, evaluations today, violations today. Sortable and filterable.
| Term | Definition |
|---|---|
| SLO | Service Level Objective -- a target for a measurable reliability metric. |
| SLI | Service Level Indicator -- the measured value that an SLO tracks. |
| P50 / P95 / P99 | Latency percentiles. P99 = 99% of requests are faster than this value. |
| Burn Rate | How fast the error budget is being consumed. 1.0 = on pace to exhaust exactly at window end. |
| Error Budget | Allowed unreliability. If target is 99.9%, the budget is 0.1% of total requests. |
| Trust Score | Numeric reputation of an agent (0--1000). Derived from behavioral signals via reward scoring. |
| Trust Ring | Concentric tiers grouping agents by trust level for visualization (high, medium, low). |
| DID | Decentralized Identifier. Format: did:mesh:<hash> (toolkit) or did:myth:<persona>:<hash> (FailSafe). |
| CMVK | Constitutional Multi-Model Verification Kernel. Cross-checks AI output with multiple models. |
| Delegation Chain | A directed trust relationship where one agent grants a capability to another. |
| Bridge | Protocol adapter connecting Agent Mesh to external systems (A2A, MCP, IATP). |
| CSP | Content Security Policy. HTTP header restricting resource loading in webviews. |
| Policy Action | Evaluation result: ALLOW (permit), DENY (reject), AUDIT (permit + log), BLOCK (reject + alert). |
| Execution Ring | Privilege tier from hypervisor: Ring 0 (root), Ring 1 (supervisor), Ring 2 (user), Ring 3 (sandbox). |
| Agent Mesh | The network of registered agents, their identities, trust scores, and interconnections. |
| Saga | A multi-step workflow with checkpoints and compensating actions managed by the hypervisor. |
| Symptom | Cause | Fix |
|---|---|---|
| Status bar shows Disconnected | WebSocket connection to local governance server dropped. | Check that the server is running (Agent OS: Start Server command). Verify port 9845 is not blocked. |
| Panel header shows Stale | Last data refresh was more than 2 tick intervals ago. | Click the refresh icon on the panel. If persistent, restart the extension host. |
| Panel shows Waiting for data... | First data fetch has not completed yet. | Wait 10 seconds for the first broadcast cycle. If it persists, the mock backend may have failed to initialize. |
| Topology graph is empty | No agents are registered in the mock or live backend. | Ensure the topology data provider is configured. In dev mode, the mock backend seeds 4 agents automatically. |
| SLO shows 0% availability | The SLO provider returned a zeroed snapshot. | This usually means the provider has not received any evaluation events. Trigger a policy evaluation or restart. |
| Browser dashboard not loading | Server failed to bind to 127.0.0.1. | Run Agent OS: Start Server and check the Output panel for port conflict messages. |
| Decision | Rationale | Risk Level |
|---|---|---|
'unsafe-inline' for style-src in CSP |
Required for VS Code theme CSS variable injection (var(--vscode-*)). Scripts remain nonce-gated. |
Low -- style-only; no script injection vector. |
retainContextWhenHidden: true on Topology Detail |
Preserves force-simulation state across tab switches (~120 animation frames). | Low -- adds ~2 MB memory when backgrounded. |
Session token via Sec-WebSocket-Protocol subprotocol |
Token sent as subprotocol (not URL query string) to avoid proxy/debug logging. 128-bit crypto.randomBytes. |
Low -- server binds to 127.0.0.1; token never leaves loopback. |
| Rate limiter Map with TTL eviction | Stale entries evicted on each request. Server is loopback-only, so the map holds at most one entry (127.0.0.1). | Negligible -- no memory growth risk. |
Math.random() for burn-rate sparkline jitter |
Synthetic demo data only, not used for any security or cryptographic purpose. | None -- replaced by real SRE data when backend connects. |
axios not used; http module for server |
The governance server uses Node built-in http. No external HTTP client dependency. |
N/A |
All extension settings are documented in the README under the Extension Settings section.
Key settings are prefixed with agent-os. in VS Code's Settings UI.
For policy file configuration, see the Agent OS documentation on policy schemas:
agent-governance-python/agent-os/src/agent_os/policies/schema.py.