An undo layer for AI agents. Provides tools to inspect, preview, and rollback agent actions through a journal-based system.
Synartesis demonstrates solid tool design with clear naming conventions (verb_noun pattern), comprehensive descriptions (avg 150+ chars), and well-structured schemas using Zod. All 11 tools have explicit descriptions and input schemas with type definitions. Tool composition is excellent, each tool has a single responsibility and outputs chain naturally (e.g., list_sessions → show_session → undo_session). Error handling includes recovery guidance (e.g., 'Use list_sessions to see them'). However, output schemas are not formally documented in the code, responses are inferred from implementation rather than declared. Parameter descriptions are present but some lack constraint details (e.g., email format, plan enum values not restated in param descriptions). Risk annotations (READ_ONLY, WRITE, DESTRUCTIVE, IRREVERSIBLE) are excellent for security but not exposed as tool annotations in the MCP schema.
Create a customer and return the created record, including its assigned id.
Delete a customer and return the record as it was immediately before deletion.
Fetch a single customer record by id.
Every session an agent has worked in, newest first, with how many actions each holds. Use this to find the session id for the other tools.
Plan the undo of a session without doing any of it: which actions would be reversed, which cannot be, and where it would stop. Nothing is written and no undo is sent. Use this before undo_session so you can tell the person what will happen.
Recreate a previously deleted customer under its original id. Intended as the inverse of delete_customer.
Output schemas not formally documented. Tool responses are inferred from implementation (e.g., say() returns text content) but not declared in tool registration. LLMs cannot plan downstream calls without knowing response structure.
Risk annotations (READ_ONLY, WRITE, DESTRUCTIVE, IRREVERSIBLE) are tracked in code comments but not exposed as MCP tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot infer safety properties from the schema.
Parameter descriptions lack constraint details. 'plan' enum values (free, pro, enterprise) are defined in schema but not restated in param description. Email format, customer ID format, and session ID prefix matching rules are not documented in param descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | 2026-07-28+ | v2 |
Send an email to a customer. This cannot be recalled once sent.
What happened in one session: every action, what it acted on, and whether an undo was recorded for it. Reads the journal only; tells you nothing about the world now.
Actually reverse a session, newest action first. Stops rather than writing over anything somebody has changed since. This changes the world, so a person is asked first. Run preview_undo and tell them what it will do before you call this.
Apply a partial patch to a customer. Omitted fields are left unchanged.
Read every resource a session touched as it is NOW, and say which still match what the session left. This is how you find out whether somebody has edited something since. Writes nothing and sends no undo. Costs a real read per resource.
Confirmation/dry-run pattern not implemented for irreversible operations. undo_session and send_email are irreversible but lack a confirmation step. preview_undo exists but is a separate tool call, should be integrated as a confirmation gate.
Error responses are text-only. resolve() returns {problem: string} but no structured error codes or recovery hints. Agents cannot programmatically distinguish 'not found' from 'ambiguous prefix' to decide retry vs. user input strategy.