@vitest-agent/mcp 4.3.1 → 5.0.0

package/README.md

@@ -10,12 +10,13 @@ The `vitest-agent-mcp` MCP server bin, built on Effect's native `McpServer` (`ef
 
 ## Features
 
-- **30 action-keyed tools** — one `Tool.make` per file, assembled into a single `Toolkit`; per-CRUD families collapse into single tools dispatching on an `action` discriminator; covers `test_status`, `test_overview`, `test_coverage`, `test_errors`, `run_tests`, `note`, `hypothesis`, `tdd_task`, `tdd_goal`, `tdd_behavior`, `tdd_progress_push` and more
+- **30 action-keyed tools** — one tool per file, assembled into a single `Toolkit`; per-CRUD families collapse into single tools dispatching on an `action` discriminator; covers `test_status`, `test_overview`, `test_coverage`, `test_errors`, `run_tests`, `note`, `hypothesis`, `tdd_task`, `tdd_goal`, `tdd_behavior`, `tdd_progress_push` and more
 - **Six framing prompts** — `triage`, `why-flaky`, `regression-since-pass`, `explain-failure`, `tdd-resume`, `wrapup`
 - **Idempotent writes** — `tdd_task`, `tdd_goal`, `tdd_behavior` and `hypothesis` create-actions are idempotent on derived keys via the `withIdempotency` combinator; a replay carries `_idempotentReplay: true`
-- **Strict tool inputs** — every served `inputSchema` is strict at every object level; an unknown key is rejected with an error naming it and listing the accepted params, instead of being stripped and running a wider query than the caller asked for
-- **Session-surviving error handling** — a handler that dies returns a structured `UnexpectedToolError` envelope, every log line goes to stderr (stdout is the JSON-RPC wire), and a stray unhandled rejection after the transport connects is logged rather than killing the server mid-session
-- **Programmatic API** — `ServerLayer({ version })` is the whole server as an Effect `Layer` over any `Stdio` implementation (the test harness runs it over in-memory queues); `Kit`, `toolHandlers`, `PromptsLayer`, `registerStrictToolkit` and `McpSession` are exported for embedding or extension
+- **Strict tool inputs** — every served `inputSchema` is strict at every object level; an unknown key is rejected with an error naming every unrecognized key and the accepted params, instead of being stripped and running a wider query than the caller asked for
+- **Typed results** — every tool serves an object-rooted `outputSchema`, and every successful result carries the typed object in `structuredContent`
+- **Session-surviving error handling** — a handler that dies returns a generic `isError` result (the detail is logged on stderr, never on stdout, the JSON-RPC wire), and a stray unhandled rejection after the transport connects is logged rather than killing the server mid-session
+- **Programmatic API** — `ServerLayer({ version })` is the whole server as an Effect `Layer` over any `Stdio` implementation (the test harness runs it over in-memory queues); `Kit`, `toolHandlers`, `PromptsLayer` and `McpSession` are exported for embedding or extension. Built on [`@effected/mcp`](https://github.com/spencerbeggs/effected)'s `McpStdio` and `McpToolkit`
 - **No MCP SDK, tRPC or zod** — the wire protocol, JSON Schema generation and input validation all come from `effect`; tool inputs and outputs are Effect `Schema` values end to end
 
 ## Install

package/idempotency.js

@@ -93,7 +93,7 @@ const withReplayMarker = (parsed) => parsed !== null && typeof parsed === "objec
 *    `DataStore.recordIdempotentResponse` best-effort: a persistence
 *    failure (or a cache-lookup failure) never surfaces to the caller, it
 *    is swallowed so a transient DB failure doesn't turn into a tool
-*    error.
+*    error. A failed `handler` persists nothing, so a retry re-runs it.
 *
 * The combinator reads no ambient state — `path` and `params` are its
 * only inputs beyond the `DataReader` / `DataStore` services it adds to