prism-mcp-server 20.22.0 → 20.22.2

package/README.md

@@ -99,8 +99,11 @@ Layer 1's rules only refuse here; the classifier reads everything else, and a
 prompt over 4,000 characters is read whole in 3,600-character windows (the most
 severe verdict stands), never as an excerpt. A classifier failure, an unreadable
 answer, or a missing classifier refuses the call; there is no keyword or Ollama
-fallback. A classifier read may take up to 90 seconds, so a classifier that
-macOS paged out during a long idle is waited for rather than refused.
+fallback. Each classifier read is bounded by 90 seconds and by the call's
+remaining `timeout_ms`, so a classifier that macOS paged out during a long idle
+is waited for rather than refused, and an oversize prompt read in windows
+cannot outlast the call budget; the first unrecoverable read ends the
+classification.
 
 Configured oMLX sampling and reasoning are preserved unless `temperature` or
 `think` is explicit, including through repairs. Generic thinking-disable retries
@@ -232,6 +235,33 @@ or by re-enabling after each run.
 <details>
 <summary>Release history (optional)</summary>
 
+## What's New in v20.22.2
+
+**Why it matters:** an agent that followed `session_task_route`'s advice was
+refused on explicit oMLX before anything ran.
+
+- **A complexity hint no longer refuses explicit oMLX.** `session_task_route`
+  recommends a `task_complexity` for every code task. That hint picks Ollama's
+  starting tier, but on explicit oMLX it lowered the model ceiling, so a call
+  that passed the recommended arguments was refused with `omlx_model_ceiling`.
+  Now only an explicit `model_ceiling` below the top tier, or a plan below it,
+  refuses. Thinking on oMLX stays with the server's settings.
+
+## What's New in v20.22.1
+
+**Why it matters:** on explicit oMLX, a long prompt could keep the device busy
+for minutes while being classified, and `timeout_ms` did not limit it.
+
+- **The call budget now covers classification.** `timeout_ms` (default 300 s)
+  starts before the safety screen, and each classifier read is bounded by the
+  remaining budget as well as its own 90 s.
+- **A failed read ends the classification.** An oversize prompt is read in
+  3,600-character windows; a read that fails already means refusal on this
+  path, so the remaining windows are no longer read. Before, a stalled server
+  could cost about four minutes per window.
+- **Tests:** the owned-job marker tests pin the size limit and the
+  non-blocking open directly.
+
 ## What's New in v20.22.0
 
 **Why it matters:** on Apple Silicon, `prism_infer` can now draft code on an

package/dist/tools/prismInferHandler.js

@@ -2206,12 +2206,14 @@ async function runInferAfterInputSafety(args, deps, t0, temperature) {
         // No trusted binding maps an oMLX model to a logical tier, and the
         // configured Flash model outweighs every Ollama tier. Below the top
         // ceiling, from the caller or the plan, refuse before any network,
-        // classification or residency change.
-        if (effectiveCeiling !== "27b") {
+        // classification or residency change. A task_complexity hint only picks
+        // Ollama's starting tier (session_task_route recommends one for every
+        // code task), so it is not a ceiling here; an explicit model_ceiling is.
+        if (clampCeiling(args.model_ceiling ?? "27b", ent.model_ceiling) !== "27b") {
             attempts.push({ tier: "omlx", reason: "omlx_model_ceiling" });
             if (wantReport)
                 return refusedResult("omlx_model_ceiling");
-            throw new Error("prism_infer: omlx_model_ceiling; explicit oMLX needs the top model ceiling");
+            throw new Error("prism_infer: omlx_model_ceiling; explicit oMLX needs the top model ceiling from model_ceiling and the plan");
         }
         if (!Number.isInteger(localMaxTokens) || localMaxTokens <= 0
             || !Number.isFinite(temperature) || temperature < 0
@@ -2240,6 +2242,9 @@ async function runInferAfterInputSafety(args, deps, t0, temperature) {
     const loaded = await deps.listLoaded();
     if (omlxTarget && installed === null)
         throw new Error("prism_infer: omlx_classifier_unavailable");
+    // Explicit oMLX: timeout_ms bounds the whole call, the Ollama unload and
+    // classification included.
+    const omlxDeadline = Date.now() + (args.timeout_ms ?? 300_000);
     // Ollama is unloaded during an explicit oMLX call (owner, 2026-10-15; this
     // replaces plan section 4's "never evict another Ollama user's weights").
     // Under the device lease, before classification: unload every resident
@@ -2328,7 +2333,7 @@ async function runInferAfterInputSafety(args, deps, t0, temperature) {
         const baseL1 = deps.callLayer1 ?? defaultCallLayer1;
         const generationTarget = omlxTarget, runtime = omlxRuntime;
         const omlxL1Fetch = generationTarget && runtime
-            ? omlxLayer1Fetch(async (request) => guardedOmlxGenerate(runtime, omlxModelTarget(generationTarget, OMLX_LAYER1_MODEL), request, deps.freemem))
+            ? omlxLayer1Fetch(async (request) => guardedOmlxGenerate(runtime, omlxModelTarget(generationTarget, OMLX_LAYER1_MODEL), request, deps.freemem), () => omlxDeadline - Date.now())
             : undefined;
         const l1fn = omlxL1Fetch
             ? (...[prompt, url, model, , images, opts]) => baseL1(prompt, url, model, omlxL1Fetch, images, opts)
@@ -2720,7 +2725,6 @@ async function runInferAfterInputSafety(args, deps, t0, temperature) {
     // Explicit oMLX classified on oMLX, so nothing of this call's is in Ollama.
     // Anything resident now belongs to another user: guardedOmlxGenerate
     // refuses generation (omlx_memory_contention) before any request.
-    const omlxDeadline = Date.now() + (args.timeout_ms ?? 300_000);
     const callLocal = omlxTarget && omlxRuntime
         ? async (_url, _model, prompt, system, maxTokens, temperature, timeoutMs, think, _images, history) => {
             const messages = [...(system === undefined ? [] : [{ role: "system", content: system }]),

package/dist/utils/layer1.js

@@ -550,7 +550,9 @@ images, opts) {
             const verdict = await callLayer1Unleased(window, ollamaUrl, model, fetchImpl, images, opts);
             if (VERDICT_SEVERITY[verdict] > VERDICT_SEVERITY[worst])
                 worst = verdict;
-            if (worst === "OBVIOUS_RESERVED")
+            // Reserved decides the call; under strict errors an ERROR does too (the
+            // call refuses), so the remaining windows are never read.
+            if (worst === "OBVIOUS_RESERVED" || (opts?.strictErrors && verdict === "ERROR"))
                 break;
         }
         return worst;

package/dist/utils/omlxLayer1.js

@@ -29,8 +29,13 @@ function chatMessages(value) {
  *    deadline.
  *  - Every request asks the provider not to keep its prompt cache: a classifier
  *    request is unique and never reused. */
-export function omlxLayer1Fetch(generate) {
+export function omlxLayer1Fetch(generate, remainingMs = () => OMLX_LAYER1_TIMEOUT_MS) {
     return (async (_input, init) => {
+        // Each read is bounded by the call's remaining timeout_ms as well, so an
+        // oversize prompt read in windows cannot outlast the budget the caller set.
+        const timeoutMs = Math.min(OMLX_LAYER1_TIMEOUT_MS, Math.floor(remainingMs()));
+        if (timeoutMs <= 0)
+            return Response.json({ error: "timeout" }, { status: 503 });
         let body;
         try {
             body = JSON.parse(String(init?.body));
@@ -46,7 +51,7 @@ export function omlxLayer1Fetch(generate) {
             return new Response(null, { status: 400 });
         const temperature = body.options?.temperature;
         const result = await generate({ messages, maxTokens, temperature: typeof temperature === "number" ? temperature : 0,
-            timeoutMs: OMLX_LAYER1_TIMEOUT_MS, think: body.think === true, cacheStore: false });
+            timeoutMs, think: body.think === true, cacheStore: false });
         return result.ok
             ? Response.json({ done: true, done_reason: result.doneReason, message: { role: "assistant", content: result.text } })
             : Response.json({ error: result.reason }, { status: 503 });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prism-mcp-server",
-  "version": "20.22.0",
+  "version": "20.22.2",
   "mcpName": "io.github.dcostenco/prism-coder",
   "description": "Persistent session memory for AI coding agents that never leaves your machine — including the on-device model that reasons over it. Restores your prior decisions, open TODOs, and changed files across sessions; adds associative recall of related past work, semantic drift detection, and local inference. Local-first by default. Works with Claude Code, Cursor, and Codex.",
   "module": "index.ts",