Security middleware for AI agents. See Everything. Undo Anything. Test Anything Safely. — Provides policy enforcement, event recording, rollback capabilities, and ghost mode (dry-run proxy) for AI agent tool calls.
UNWIND exposes 25 tools across sidecar, CLI, dashboard, and cadence subsystems. Most tools lack complete input schemas and descriptions are generic or missing. Tools are poorly named (ghost_toggle, api_undo, cadence_reset lack verb clarity). Parameter descriptions exist but are minimal (avg ~40 chars, well below 72-char baseline). No output schemas documented. Error handling guidance absent. Security-critical tools (policy_check, ghost_approve) have minimal descriptions. This is a typical community server with significant definition gaps.
Create CR-AFT chain anchor. CLI command that creates a cryptographic anchor point in the event hash chain.
Get paginated event timeline. Dashboard API endpoint that returns events with optional filtering and pagination.
Get snapshot metadata. Dashboard API endpoint that returns filesystem snapshots for rollback capability.
Get current trust state and summary statistics. Dashboard API endpoint that returns trust state, event counts, taint level, and ghost mode status.
Perform undo action via API. Dashboard API endpoint that executes rollback of filesystem changes.
Natural language query interface. CLI command that accepts natural language questions about events and trust state.
Remove events before a given date. Cadence CLI command that deletes historical rhythm data before a specified cutoff.
Weak tool naming: 'ghost_toggle', 'api_undo', 'cadence_reset', 'api_trust_state' lack clear action verbs. LLMs cannot infer intent from names alone. Should be 'enable_ghost_mode', 'rollback_changes', 'clear_rhythm_data', 'get_trust_state'.
Missing output schemas: No tool documents what fields it returns. LLMs cannot plan downstream calls or extract required data. E.g., does 'log' return event_id, timestamp, session_id? Does 'export' return file_path or file_content?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | 2026-07-28+ | v2 |
Clear all rhythm data. Cadence CLI command that resets temporal-awareness state and removes all pulse log events.
Show current rhythm state. Cadence CLI command that displays temporal-awareness and user-rhythm modelling state.
Launch web dashboard. CLI command that starts the Flask web dashboard for viewing events, trust state, and managing approvals.
Export events in various formats. CLI command that exports recorded events as JSON, JSONL, or HTML.
Toggle Ghost Mode on/off. CLI command that enables or disables ghost mode (dry-run proxy) for the current session.
Approve a pending ghost mode action. POST /v1/ghost/approve endpoint that approves a specific pending action to execute it for real.
Discard a pending ghost mode action. POST /v1/ghost/discard endpoint that discards a specific pending action without executing it.
Get pending ghost mode events/actions. GET /v1/ghost/events endpoint that lists all pending actions in ghost mode for a session.
Get ghost mode (dry-run proxy) status. GET /v1/ghost/status endpoint that returns whether ghost mode is active for a session and pending action counts.
Toggle ghost mode on/off for a session. POST /v1/ghost/toggle endpoint that enables or disables dry-run proxy mode.
Liveness and readiness probe. GET /v1/health endpoint that returns sidecar health status, engine version, and operational state.
Show recent events with optional filtering by time range or session. CLI command that displays event timeline with optional --since and --session filters.
Evaluate tool call against enforcement pipeline. POST /v1/policy/check endpoint that checks if a tool call is allowed, blocked, mutated, or requires challenge approval based on configured policies.
Show current trust state and last 5 high-risk events. CLI command that displays UNWIND status, event counts, chain integrity, and recent high-risk events.
Run tamper detection. CLI command that performs tamper detection checks on the event database and hash chain.
Ingest after_tool_call telemetry (best-effort). POST /v1/telemetry/event endpoint that records tool execution telemetry including tool name, parameters, results, errors, and duration.
Undo state-modifying actions with optional targeting by event ID, time range, or session. CLI command that rolls back filesystem changes from recorded events.
Verify CR-AFT hash chain integrity. CLI command that validates the cryptographic hash chain of all recorded events.
Generic/minimal descriptions: Most tools have descriptions under 50 chars (e.g., 'Show current trust state and last 5 high-risk events' for 'status'). Baseline is 194 chars. Descriptions lack WHEN to use, WHAT it returns, and prerequisites.
No error handling guidance: Tools lack recovery hints. E.g., 'policy_check' can return 'blocked' or 'challenge_required' but no description explains what the agent should do next or how to request approval.
Destructive tools lack confirmation pattern: 'cadence_reset', 'cadence_forget', 'undo' modify or delete state but have no dry-run or confirmation step. Agents can accidentally destroy data.