@yawlabs/mcp 0.60.3 → 0.62.0

package/CHANGELOG.md

@@ -2,6 +2,29 @@
 
 All notable changes to `@yawlabs/mcp` (formerly `@yawlabs/mcph`) are documented here. This project uses [semantic versioning](https://semver.org) and a script-gated release flow: `./release.sh <version>` runs lint + tests + build, bumps, tags, publishes to npm, and creates the GitHub release.
 
+## 0.62.0 -- verifiable-signal routing: graded reward, miss tracking, and an eval foundry
+
+Lands #25, #26, and #27. The dispatch router's learning signal moves from a binary "any non-error reply counts as success" to a sound, quality-graded reward, plus the surrounding machinery to manufacture and verify that signal. All of the new behavior is additive and the new knobs are off by default, so existing setups are unchanged.
+
+- **Dispatch reward is now graded, not binary.** An empty body or an error-shaped 200 (e.g. `{isError:false, text:"not found"}`) no longer banks full credit toward a server's `mcp_connect_dispatch` boost -- the learning signal is a quality-weighted reward in [0,1], so a server that "replies" but does not actually help stops accruing a nudge. When every reply is a clean success or a hard error this collapses to the old behavior.
+- **Re-dispatch miss signal.** When an intent routes to server A, A replies cleanly but is abandoned, and a token-similar intent then routes to a different server B within ~2 minutes, A is penalized as the wrong route. Designed multi-server flows (curated bundles, detected packs) are excluded so an intentional A-then-B chain is not mistaken for a miss.
+- **Step-level (process) reward in `mcp_connect_exec`.** Each pipeline step is graded on its own; a step that fails on bad `$ref` input it consumed from an upstream step splits the blame with the producer instead of being fully blamed, so a flaky producer does not hide behind a healthy consumer (or vice versa).
+- **Routing effort dial -- `YAW_MCP_ROUTE_EFFORT=off|auto|aggressive`** (also a per-call `routeEffort` arg on `mcp_connect_dispatch`). Controls how much LLM sampling dispatch spends to disambiguate close rankings. `auto` (default) preserves the prior fixed top-2 tiebreak exactly -- no latency change on the default path; `aggressive` samples best-of-3 on milder ambiguity; `off` never samples.
+- **Opt-in LLM reward grader -- `YAW_MCP_REWARD_GRADER=1`** (off by default). On the uncertain reward bands only, it asks the client's own LLM (via MCP sampling) whether a call accomplished the goal and revises the credit in the background. Non-blocking (the tool result never waits on the grade), capability-gated, and never-throwing -- a missing capability / timeout / unparseable reply just leaves the heuristic reward standing.
+- **Opt-in routing-eval harvest + CI gate -- `YAW_MCP_FOUNDRY=1`** (off by default). Writes a privacy-safe `(intent -> chosen server)` corpus to `~/.yaw-mcp/foundry.jsonl`: the intent is reduced to a redacted, secret-stripped, order-shuffled token bag, never the raw text. New `yaw-mcp foundry export` folds the harvest into a checked-in regression corpus (snapshotting the local server catalog), and a BM25-floor gate scores the ranker against real dispatches -- it skips cleanly until a corpus is committed.
+
+## 0.61.0 -- audit-fix wave: exec binding payloads (BREAKING), live-bug fixes, 50+ hardening items
+
+- **BREAKING: `mcp_connect_exec` step bindings now hold the step's semantic payload, not the raw MCP wire wrapper.** A single-text-item JSON response binds as the parsed value, a non-JSON text response binds as the string, and anything else binds as the content array. `$ref` paths written against the old wire shape -- e.g. `"stepId.content[0].text"` -- now throw a RefError; migrate to `"stepId"` (whole value) or `"stepId.field"` (a specific field). This matches what the tool description always promised.
+- `mcp_connect_activate` / `mcp_connect_dispatch` set `isError` whenever a real activation failure occurs (partial successes no longer mask failures); concurrent-server-cap refusals are flagged internally and stay informational in both. Deactivating an already-unloaded namespace is now an idempotent success.
+- ~30 further fixes from a full-implementation audit: bundles.json write serialization + ENOENT-only absence handling, `list` surfaces parse warnings, analytics 401 no longer clears the team session, `secrets pull` refuses cross-passphrase overwrites without `--force`, uv probe/retry fixes on Windows, stable array positions in pruned tool output, `--help` exits 0 on stdout across subcommands, ASCII-safe terminal output, and assorted copy corrections.
+
+## 0.60.3 -- npm-prefix refinement + pnpm/bun self-upgrade
+
+- `refineInstallMethod` now probes `npm prefix -g` (3s timeout) and normalises the result through `realpathSync` so junctioned prefixes (scoop's `current` symlink, Volta shims) resolve to the real path before comparison. When the running entrypoint lives under the npm global prefix, ambiguous `local-node-modules` / `unknown` detections are promoted to `global-npm` -- fixes exotic prefix setups the path-marker list doesn't know.
+- The probe is wired into both `yaw-mcp upgrade` and `yaw-mcp doctor` via the shared `refineInstallMethod` call in `runUpgrade` / `runDoctor`.
+- `maybeAutoUpgrade` (the fire-and-forget startup check) now acts on stale pnpm and bun global installs with `pnpm add -g` / `bun add -g @yawlabs/mcp@latest` in addition to the existing `npm install -g` path. The background spawn log messages interpolate the actual tool name instead of hardcoding `npm`.
+
 ## 0.60.2 -- pnpm/bun global stores upgrade with their owning tool
 
 - `yaw-mcp upgrade` now detects pnpm global stores (`<pnpm-home>/global/<n>/node_modules/...`) and bun global installs (`~/.bun/install/global/...`) as their own install methods. `--run` spawns `pnpm add -g` / `bun add -g @yawlabs/mcp@latest` instead of misclassifying them as local node_modules trees -- which would have npm-installed a foreign package-lock + node_modules into the tool manager's internal store.

package/README.md

@@ -367,7 +367,7 @@ Rotate a credential in one place (the dashboard), every machine picks up the new
 | `LOG_LEVEL` | No | Log verbosity: `debug`, `info`, `warn`, `error` (default: `info`) |
 | `YAW_MCP_POLL_INTERVAL` | No | Config-poll interval in seconds. `0` disables polling (config fetched once at startup). Default: `60` |
 | `YAW_MCP_AUTO_ACTIVATE` | No | When `discover` is called with a context string and one server clearly wins, auto-load it. Set to `0` to disable. Default: enabled |
-| `YAW_MCP_AUTO_UPGRADE` | No | When yaw-mcp starts as a server, runs a non-blocking background check for a newer global-npm install and upgrades quietly. Set to `0` to disable. Default: enabled. |
+| `YAW_MCP_AUTO_UPGRADE` | No | When yaw-mcp starts as a server, runs a non-blocking background check for a newer npm/pnpm/bun global install and upgrades quietly with the owning tool. Set to `0` to disable. Default: enabled. |
 | `YAW_MCP_SERVER_CAP` | No | Hard cap on concurrently activated servers. Default: `6`. Set to `0` to disable. |
 | `YAW_MCP_PRUNE_RESPONSES` | No | Conservative response pruning (redact large file blobs etc. before returning to the client). Set to `0` or `false` to disable. Default: enabled. |
 | `YAW_MCP_DISABLE_PERSISTENCE` | No | Set to `1` or `true` to keep learning + pack-history scoped to the current process -- nothing loaded at start, nothing written on shutdown. Intended for ephemeral / shared environments (CI, containers). Default: cross-session persistence enabled at `~/.yaw-mcp/state.json`. |
@@ -410,7 +410,7 @@ If you find a security issue in yaw-mcp itself, report it via [GitHub's private
 ## Requirements
 
 - Node.js 18+
-- A [yaw.sh/mcp](https://yaw.sh/mcp) account
+- No account required for core features. A Yaw Team license key is needed only for sync and shared bundles (yaw.sh/mcp).
 
 ## Links
 

package/dist/{chunk-C3WU6HAG.js → chunk-WORQOSXT.js}

@@ -27,7 +27,7 @@ var LOG_LEVELS = { debug: 0, info: 1, warn: 2, error: 3 };
 var minLevel = LOG_LEVELS[process.env.LOG_LEVEL?.toLowerCase()] ?? LOG_LEVELS.info;
 function log(level, msg, data) {
   if (LOG_LEVELS[level] < minLevel) return;
-  const entry = JSON.stringify({ level, msg, ts: (/* @__PURE__ */ new Date()).toISOString(), ...data });
+  const entry = JSON.stringify({ ...data, level, msg, ts: (/* @__PURE__ */ new Date()).toISOString() });
   process.stderr.write(`${entry}
 `);
 }
@@ -115,10 +115,10 @@ function invalidateState() {
   cachedState = null;
 }
 async function loadStoredState(filePath) {
-  if (cachedState) {
+  if (cachedState && cachedState.filePath === filePath) {
     const s = cachedState.state;
     if (s && expMs(s.session) < Date.now()) {
-      cachedState = { state: null };
+      cachedState = { filePath, state: null };
       return null;
     }
     return s;
@@ -135,14 +135,14 @@ async function loadStoredState(filePath) {
     parsed = null;
   }
   if (parsed && expMs(parsed.session) < Date.now()) {
-    cachedState = { state: null };
+    cachedState = { filePath, state: null };
     return null;
   }
-  cachedState = { state: parsed };
+  cachedState = { filePath, state: parsed };
   return parsed;
 }
 async function saveStoredState(filePath, state) {
-  cachedState = { state };
+  cachedState = { filePath, state };
   try {
     await atomicWriteFile(filePath, JSON.stringify(state, null, 2));
     if (process.platform !== "win32") {
@@ -328,7 +328,6 @@ async function postAnalyticsEvent(event, opts = {}) {
     baseUrl: opts.baseUrl
   });
   if (res.status === 401) {
-    await clearStoredState(filePath);
     return { ok: false };
   }
   return { ok: res.status === 200 };
@@ -353,18 +352,23 @@ async function listAnalyticsEvents(opts = {}) {
   }
   return { events: res.body.events ?? [], cap: res.body.cap ?? 0, order_id: res.body.order_id ?? "" };
 }
+function getCachedCookie(home = homedir2()) {
+  void home;
+  const s = cachedState?.state ?? null;
+  return s ? s.cookie : null;
+}
 function _resetForTests() {
   invalidateState();
 }
 
 export {
+  atomicWriteFile,
   log,
   cacheDir,
   CONFIG_DIRNAME,
   userConfigDir,
   findProjectConfigDir,
   guidePath,
-  atomicWriteFile,
   SESSION_STATE_FILENAME,
   TeamSyncAuthError,
   TeamSyncForbiddenError,
@@ -377,5 +381,6 @@ export {
   putResource,
   postAnalyticsEvent,
   listAnalyticsEvents,
+  getCachedCookie,
   _resetForTests
 };