prism-mcp-server 20.21.16 → 20.21.17

package/README.md

@@ -158,6 +158,15 @@ or by re-enabling after each run.
 <details>
 <summary>Release history (optional)</summary>
 
+## What's New in v20.21.17
+
+- Fixed: the on-device screen refused some ordinary coding requests. With a
+  Synalux account (free included), it now reads each request through a policy
+  that Synalux serves, pinned by its SHA-256. Without an account, or with
+  images attached, the screen is unchanged.
+- `prism_infer` route mode: set `allow_parallel_calls` to keep a reply that
+  calls several of the tools you offered in `allowed_tools`.
+
 ## What's New in v20.21.16
 
 ### Conversations: checked on your device, and free with an account

package/dist/tools/prismInferHandler.js

@@ -35,7 +35,7 @@ import { passesClinicalQualityGate, clinicalPlanScaffold, formatClinicalSections
 import { applyDeterministicCodingRepairs, buildCodingRepairPrompt, passesCodingQualityGate, } from "../utils/codingQualityPolicy.js";
 import { checkInputSafety, checkOutputSafety } from "../utils/safetyGate.js";
 import { callLayer1 as defaultCallLayer1, classifyDeterministicLayer1, keywordBackstop, reservedCategory, MAX_CLASSIFIER_PROMPT_LENGTH, layer1ClassifierContent, secondReadExclusion } from "../utils/layer1.js";
-import { getSecondReadPolicy, getAnswerCheckPolicy } from "../utils/inferencePolicy.js";
+import { getSecondReadPolicy, getAnswerCheckPolicy, getClassifierInputPolicy } from "../utils/inferencePolicy.js";
 import { pseudonymizeForCheck } from "../utils/pseudonymize.js";
 import { answerGroundingBytes, answerGroundingContent, parseGroundingVerdict, arithmeticSlips, arithmeticCorrection, ANSWER_GROUNDING_OUTPUT_TOKENS, ANSWER_GROUNDING_THINK, ANSWER_GROUNDING_THINK_TOKENS, ANSWER_GROUNDING_TIMEOUT_MS, ANSWER_GROUNDING_RETRY_TIMEOUT_MS, ANSWER_GROUNDING_FOLLOW_UP_TOKENS } from "../utils/answerGrounding.js";
 import { recordInference, recordThinkOnlyRetry, formatInferenceMetrics, estimateTokens } from "../utils/inferenceMetrics.js";
@@ -269,6 +269,8 @@ function layer1HistoryCached(model, window) {
     const hit = layer1HistoryCache.get(key);
     return hit !== undefined && hit.expiresAt > performance.now();
 }
+/** How long the screen waits for the classifier-input policy on its first load. */
+const CLASSIFIER_INPUT_LOAD_MS = 3_000;
 /** Tokens the classifier may generate (callLayer1's num_predict); they share
  *  the context with the request. */
 export const LAYER1_CLASSIFIER_OUTPUT_TOKENS = 16;
@@ -765,6 +767,11 @@ export const PRISM_INFER_TOOL = {
                     "Synalux deterministic route correction. 'local': skips that correction only.",
                 default: "auto",
             },
+            allow_parallel_calls: {
+                type: "boolean",
+                description: "Route: keep a reply of several calls only if every call is in allowed_tools.",
+                default: false,
+            },
             think: {
                 type: "boolean",
                 description: "<think> reasoning. Default true for chat/code, false for route; better on complex " +
@@ -772,10 +779,8 @@ export const PRISM_INFER_TOOL = {
             },
             strict_entitlements: {
                 type: "boolean",
-                description: "Fail loud instead of running with ASSUMED free-tier limits: when entitlements " +
-                    "fell back to free because the portal was unreachable (source='fallback_free'), " +
-                    "throw instead of silently applying free clamps. Portal-confirmed free plans and " +
-                    "unconfigured machines are unaffected.",
+                description: "Throw rather than apply assumed free limits when the portal was unreachable " +
+                    "(source='fallback_free'). Confirmed-free and unconfigured setups are unaffected.",
                 default: false,
             },
             escalation: {
@@ -874,6 +879,8 @@ export function isPrismInferArgs(args) {
     if (a.route_guard !== undefined &&
         !["auto", "local"].includes(a.route_guard))
         return false;
+    if (a.allow_parallel_calls !== undefined && typeof a.allow_parallel_calls !== "boolean")
+        return false;
     if (a.allowed_tools !== undefined) {
         if (!Array.isArray(a.allowed_tools) || a.allowed_tools.length > MAX_ROUTE_TOOLS)
             return false;
@@ -1878,6 +1885,7 @@ export async function runInfer(args, deps) {
             "Retry, or drop strict_entitlements to accept free clamps.");
     }
     const mode = args.mode ?? "route";
+    const gateOptions = { allowParallelCalls: mode === "route" && args.allow_parallel_calls === true };
     // Model choice belongs here—not in session_task_route—because this layer
     // owns every viability input and the explicit caller override contract.
     const requestedCeiling = resolveRequestedModelCeiling(args);
@@ -2156,10 +2164,15 @@ export async function runInfer(args, deps) {
         // kept: reserved and uncertain fail closed for text, error follows
         // the single-prompt error path), then each turn and the prompt in
         // context (raise only) — see below.
+        // The account's classifier-input policy; without one the classifier
+        // reads the prompt as written.
+        const classifierInput = await (deps.classifierInputPolicy ?? (() => getClassifierInputPolicy({ deadlineMs: CLASSIFIER_INPUT_LOAD_MS })))().catch(() => null);
         let l1;
         if (!args.messages?.length) {
-            // Single turn: the exact call it always was.
-            l1 = await l1fn(args.prompt, deps.ollamaUrl, l1Model, undefined, resolvedImages);
+            // Single turn: one call; the classifier-input policy is passed when there is one.
+            l1 = classifierInput
+                ? await l1fn(args.prompt, deps.ollamaUrl, l1Model, undefined, resolvedImages, { classifierInput })
+                : await l1fn(args.prompt, deps.ollamaUrl, l1Model, undefined, resolvedImages);
             if (l1 !== "OBVIOUS_NOT_RESERVED")
                 l1Layer = "prompt";
         }
@@ -2250,7 +2263,7 @@ export async function runInfer(args, deps) {
             // (review round 19: skipping it there bypassed that floor).
             const promptFastPath = promptRoutine && args.prompt.length <= MAX_CLASSIFIER_PROMPT_LENGTH && (resolvedImages?.length ?? 0) === 0;
             if (l1 !== "OBVIOUS_RESERVED" && !promptFastPath) {
-                l1 = raise(l1, await l1fn(args.prompt, deps.ollamaUrl, l1Model, undefined, resolvedImages, { deterministic: false }), "prompt");
+                l1 = raise(l1, await l1fn(args.prompt, deps.ollamaUrl, l1Model, undefined, resolvedImages, { deterministic: false, ...(classifierInput ? { classifierInput } : {}) }), "prompt");
             }
             // 3. Context, raise only: one window per turn and one for the
             // prompt (see contextWindows), cached like any window. Skipped
@@ -2844,7 +2857,7 @@ export async function runInfer(args, deps) {
                 let { stripped, thinkOnly } = stripThink(result.text);
                 let output = stripped;
                 // Quality gate — all modes. Route uses mode-aware empty floor (length===0).
-                let gate = passesQualityGate(output, thinkOnly, result.doneReason, mode);
+                let gate = passesQualityGate(output, thinkOnly, result.doneReason, mode, gateOptions);
                 if (gate.pass && mode === "code") {
                     gate = passesCodingQualityGate(args.prompt, output);
                 }
@@ -2883,7 +2896,7 @@ export async function runInfer(args, deps) {
                     const retried = await deps.callLocal(deps.ollamaUrl, ollamaName, args.prompt, effectiveSystem, tierTokens, temperature, timeout, false, resolvedImages, ...historyArgs(args));
                     if (retried.ok) {
                         const retriedStrip = stripThink(retried.text);
-                        const retriedGate = passesQualityGate(retriedStrip.stripped, retriedStrip.thinkOnly, retried.doneReason, mode);
+                        const retriedGate = passesQualityGate(retriedStrip.stripped, retriedStrip.thinkOnly, retried.doneReason, mode, gateOptions);
                         // Keep the retry only if it is actually better — a retry that
                         // truncates too must not overwrite the original with a
                         // shorter fragment.
@@ -2913,7 +2926,7 @@ export async function runInfer(args, deps) {
                     const deterministicRepair = applyDeterministicCodingRepairs(output, failedReason);
                     if (deterministicRepair.changes.length > 0) {
                         output = deterministicRepair.output;
-                        gate = passesQualityGate(output, false, result.doneReason, mode);
+                        gate = passesQualityGate(output, false, result.doneReason, mode, gateOptions);
                         if (gate.pass) {
                             gate = passesCodingQualityGate(args.prompt, output);
                         }
@@ -3181,6 +3194,7 @@ async function applyVerification(draft, args, deps, partial) {
     const mode = args.mode ?? "route";
     if (mode === "route") {
         const allowedTools = new Set(args.allowed_tools ?? DEFAULT_PRISM_ROUTE_TOOLS);
+        const contractOptions = { allowParallel: args.allow_parallel_calls === true };
         const parsed = parseRouteOutput(draft);
         const shouldUsePortal = args.route_guard !== "local" &&
             partial.plan !== "free" &&
@@ -3198,7 +3212,7 @@ async function applyVerification(draft, args, deps, partial) {
                 });
                 const portalOutcome = validatePortalRouteGuardOutcome(untrustedPortalOutcome, draft, allowedTools, args.prompt);
                 if (!portalOutcome) {
-                    const localCheck = applyLocalRouteContract(draft, allowedTools);
+                    const localCheck = applyLocalRouteContract(draft, allowedTools, contractOptions);
                     routeGuard = {
                         ...localCheck,
                         source: "local_fallback",
@@ -3220,7 +3234,7 @@ async function applyVerification(draft, args, deps, partial) {
                 }
             }
             catch (error) {
-                const localFallback = applyLocalRouteContract(draft, allowedTools);
+                const localFallback = applyLocalRouteContract(draft, allowedTools, contractOptions);
                 routeGuard = {
                     ...localFallback,
                     source: "local_fallback",
@@ -3239,7 +3253,7 @@ async function applyVerification(draft, args, deps, partial) {
             }
         }
         else {
-            routeGuard = applyLocalRouteContract(draft, allowedTools);
+            routeGuard = applyLocalRouteContract(draft, allowedTools, contractOptions);
         }
         routedDraft = routeGuard.output;
     }

package/dist/utils/inferencePolicy.js

@@ -1,7 +1,8 @@
 /**
- * Policies the on-device multi-turn features run with, served by Synalux to
- * plans with multi-turn: the second read's exclusion policy (the conversations
- * the 9b may not re-read after a 4b hedge) and the answer check's rules. Each
+ * Policies on-device features run with, served by Synalux to plans with
+ * multi-turn: the second read's exclusion policy (the conversations the 9b may
+ * not re-read after a 4b hedge), the answer check's rules, and the screen's
+ * classifier-input policy (layer1.ts classifierCopy). Each
  * release accepts exactly one artifact of each, pinned by its SHA-256, so the
  * policy a client runs is the one it was released and validated with.
  *
@@ -11,7 +12,8 @@
  * kept. Anything but the pinned, well-formed artifact is no policy: with
  * no second-read policy the second read does not run (the hedge stands); with
  * no answer-check policy a local answer to a conversation is unchecked (cloud,
- * else withheld).
+ * else withheld); with no classifier-input policy the classifier reads the
+ * request as written.
  */
 import { createHash } from "node:crypto";
 import { PRISM_SYNALUX_BASE_URL } from "../config.js";
@@ -25,10 +27,17 @@ export const SECOND_READ_POLICY_EVALUATOR = "second-read-exclusion/1";
 export const ANSWER_CHECK_POLICY_SHA256 = "ba12ab1f6858b68ed36b7c0551aa3381ffb45b6123eb0aacd09c9316efd27993";
 /** The mechanism this client implements (answerGrounding.ts, groundAnswer). */
 export const ANSWER_CHECK_POLICY_EVALUATOR = "answer-check/1";
+/** The classifier-input artifact this release runs. */
+export const CLASSIFIER_INPUT_POLICY_SHA256 = "6b215f9af8cb94c9467852d6bd20f93c3a5c33dd35c649bee77e5f69e0abe4b1";
+/** The mechanism this client implements (layer1.ts classifierCopy). */
+export const CLASSIFIER_INPUT_POLICY_EVALUATOR = "classifier-input/1";
 const MAX_ARTIFACT_BYTES = 64 * 1024;
 const MAX_PATTERN_CHARS = 1_024;
 const MIN_OPERATIONAL_TERMS = 8;
 const MIN_DEPLOY_DECISION = 2;
+const MIN_DROP_WORDS = 8;
+const MAX_WORD_CHARS = 40;
+const MAX_REQUIRED_GROUPS = 4;
 /** A list longer than this is not a policy this client was released with. */
 const MAX_LIST_ENTRIES = 1_000;
 /** A group whose body repeats may be repeated at most this many times. */
@@ -274,6 +283,66 @@ export function parseAnswerCheckPolicy(bytes, expectSha256 = ANSWER_CHECK_POLICY
         return null;
     }
 }
+/** The classifier-input policy from the artifact's exact bytes, or null for
+ *  anything but the expected artifact: another hash, schema or evaluator, a
+ *  word list below its floor or with an entry that is not one lowercase word,
+ *  no required group or a group or qualifier naming an unlisted word, no
+ *  words a kept sentence must offer, a token
+ *  pattern that is oversized, refers back, repeats a varying group or does not
+ *  compile. */
+export function parseClassifierInputPolicy(bytes, expectSha256 = CLASSIFIER_INPUT_POLICY_SHA256) {
+    if (Buffer.byteLength(bytes, "utf8") > MAX_ARTIFACT_BYTES)
+        return null;
+    if (sha256(bytes) !== expectSha256)
+        return null;
+    let a;
+    try {
+        a = JSON.parse(bytes);
+    }
+    catch {
+        return null;
+    }
+    const art = a;
+    if (art?.schema !== 1 || art.evaluator !== CLASSIFIER_INPUT_POLICY_EVALUATOR || typeof art.classifier_input !== "object" || art.classifier_input === null)
+        return null;
+    const s = art.classifier_input;
+    const words = s.drop_sentence_words;
+    const word = (v) => typeof v === "string" && v.length > 0 && v.length <= MAX_WORD_CHARS && v === v.toLowerCase() && !/\s/.test(v);
+    if (!Array.isArray(words) || words.length < MIN_DROP_WORDS || words.length > MAX_LIST_ENTRIES || !words.every(word))
+        return null;
+    const listed = new Set(words);
+    const subset = (v) => Array.isArray(v) && v.length > 0 && v.length <= MAX_LIST_ENTRIES && v.every(w => typeof w === "string" && listed.has(w));
+    const groups = s.require_each;
+    if (!Array.isArray(groups) || groups.length === 0 || groups.length > MAX_REQUIRED_GROUPS || !groups.every(subset))
+        return null;
+    const after = s.only_after ?? {};
+    if (typeof after !== "object" || after === null || Array.isArray(after))
+        return null;
+    const afterEntries = Object.entries(after);
+    if (!afterEntries.every(([w, prev]) => listed.has(w) && subset(prev)))
+        return null;
+    const needed = s.kept_needs_one_of;
+    if (!Array.isArray(needed) || needed.length === 0 || needed.length > MAX_LIST_ENTRIES || !needed.every(word))
+        return null;
+    const tokenPattern = (v) => v === undefined || (typeof v === "string" && v.length > 0 && v.length <= MAX_PATTERN_CHARS && !/\\[1-9]|\\k</.test(v) && !hasNestedRepetition(v));
+    const also = s.also_match, afterPattern = s.only_after_pattern;
+    if (!tokenPattern(also) || !tokenPattern(afterPattern))
+        return null;
+    try {
+        // The client sets the flags (none); the artifact supplies the source only.
+        return {
+            dropWords: listed,
+            requireEach: groups.map(g => new Set(g)),
+            onlyAfter: new Map(afterEntries.map(([w, prev]) => [w, new Set(prev)])),
+            onlyAfterPattern: typeof afterPattern === "string" ? new RegExp(afterPattern) : null,
+            alsoMatch: typeof also === "string" ? new RegExp(also) : null,
+            keptNeedsOneOf: new Set(needed),
+        };
+    }
+    catch {
+        return null;
+    }
+}
 /** One pinned artifact: loaded once, shared by concurrent callers, retried
  *  after RETRY_AFTER_MS when a load fails, and dropped by reset() when the
  *  account changes (a sign-in can switch the account and the portal). A load
@@ -358,15 +427,19 @@ async function load(o, sha, parse) {
 }
 const secondRead = pinned(SECOND_READ_POLICY_SHA256, parseSecondReadPolicy);
 const answerCheck = pinned(ANSWER_CHECK_POLICY_SHA256, parseAnswerCheckPolicy);
+const classifierInput = pinned(CLASSIFIER_INPUT_POLICY_SHA256, parseClassifierInputPolicy);
 /** The pinned second-read policy, or null. */
 export const getSecondReadPolicy = (o = {}) => secondRead.get(o);
 /** The pinned answer-check policy, or null. */
 export const getAnswerCheckPolicy = (o = {}) => answerCheck.get(o);
-/** Drops both cached policies; the next conversation loads them again. Called
+/** The pinned classifier-input policy, or null. */
+export const getClassifierInputPolicy = (o = {}) => classifierInput.get(o);
+/** Drops every cached policy; the next request loads them again. Called
  *  when the account changes (dashboard sign-in and sign-out). */
 export function clearInferencePolicies() {
     secondRead.reset();
     answerCheck.reset();
+    classifierInput.reset();
 }
 /** Tests only. */
 export function _resetSecondReadPolicyForTest() {

package/dist/utils/layer1.js

@@ -66,6 +66,62 @@ Answer (one word):`;
 export function layer1ClassifierContent(input) {
     return LAYER1_PROMPT.replace("{prompt}", () => input);
 }
+/** Question marks in the scripts prism serves: a sentence with one is never left out. */
+const QUESTION_MARK = new RegExp("[?" + String.fromCharCode(0xff1f, 0xfe56, 0x061f, 0x037e, 0x00bf, 0x203d, 0x2047, 0x2048, 0x2049, 0x2e2e, 0x055e, 0x1367) + "]");
+function words(sentence) {
+    return sentence.toLowerCase().split(/[\s,;:()]+/).map((w) => w.replace(TOKEN_EDGE, "")).filter(Boolean);
+}
+const TOKEN_EDGE = /^[^\w`'#.+-]+|[^\w`'#+-]+$/g;
+function droppable(sentence, p) {
+    // A sentence with a question mark is never left out: it may be what the request asks.
+    if (QUESTION_MARK.test(sentence))
+        return false;
+    const hit = p.requireEach.map(() => false);
+    let prev = null;
+    for (const token of words(sentence)) {
+        const pattern = !!p.alsoMatch?.test(token);
+        if (p.dropWords.has(token)) {
+            const after = p.onlyAfter.get(token);
+            if (after && !(prev !== null && (after.has(prev) || !!p.onlyAfterPattern?.test(prev))))
+                return false;
+            p.requireEach.forEach((group, i) => { if (group.has(token))
+                hit[i] = true; });
+        }
+        else if (!pattern) {
+            return false;
+        }
+        prev = token;
+    }
+    return hit.length > 0 && hit.every(Boolean);
+}
+/**
+ * The classifier's copy of a request under a classifier-input policy. A
+ * sentence the policy allows is left out with one adjacent separator; every
+ * other character is kept, and text with nothing to leave out is returned as it is.
+ */
+export function classifierCopy(text, p) {
+    // Even indexes are sentences, odd indexes the separators between them.
+    const parts = text.split(/((?<=[.!?])[^\S\r\n]+|[\r\n]+)/);
+    const keep = parts.map(() => true);
+    let dropped = false;
+    for (let i = 0; i < parts.length; i += 2) {
+        if (!droppable(parts[i], p))
+            continue;
+        keep[i] = false;
+        dropped = true;
+        if (i > 0 && keep[i - 1])
+            keep[i - 1] = false;
+        else if (i + 1 < parts.length)
+            keep[i + 1] = false;
+    }
+    if (!dropped)
+        return text;
+    // The sentences that stay must still say what is asked; otherwise the classifier reads it all.
+    if (!parts.some((part, i) => i % 2 === 0 && keep[i] && words(part).some((w) => p.keptNeedsOneOf.has(w))))
+        return text;
+    const out = parts.filter((_, i) => keep[i]).join("");
+    return out.trim() ? out : text;
+}
 const VALID = new Set([
     "OBVIOUS_RESERVED",
     "OBVIOUS_NOT_RESERVED",
@@ -408,7 +464,8 @@ images, opts) {
         // short-circuits to reserved handling.
         return "OBVIOUS_RESERVED";
     }
-    const classifierInput = oversize ? buildOversizeExcerpt(userPrompt) : userPrompt;
+    const excerpt = oversize ? buildOversizeExcerpt(userPrompt) : userPrompt;
+    const classifierInput = opts?.classifierInput && !hasImages ? classifierCopy(excerpt, opts.classifierInput) : excerpt;
     // A SYSTEM baked into the classifier model's Modelfile must not sit in front
     // of LAYER1_PROMPT. prism-coder:4b bakes a tool-routing prompt; with it the
     // private eval gate failed 5/5 runs (two hard negatives refused every run),

package/dist/utils/qualityGate.js

@@ -1,4 +1,14 @@
-import { parseRouteOutput } from "./routeContract.js";
+import { parseRouteCalls, parseRouteOutput } from "./routeContract.js";
+/** JSON with object keys sorted at every level: the same arguments in any order give one string. */
+function canonicalJson(value) {
+    if (Array.isArray(value))
+        return `[${value.map(canonicalJson).join(",")}]`;
+    if (value !== null && typeof value === "object") {
+        const record = value;
+        return `{${Object.keys(record).sort().map((k) => `${JSON.stringify(k)}:${canonicalJson(record[k])}`).join(",")}}`;
+    }
+    return JSON.stringify(value) ?? "null";
+}
 /**
  * Signal 5 — Tool-call bleed: pipe-delimited format leaking into non-tool turns.
  * Matches <|tool_call|> and <|tool_call_end|> only — NOT angle-bracket <tool_call> variants
@@ -11,8 +21,9 @@ export const TOOL_CALL_BLEED_RE = /<\|tool_call\|>|<\|tool_call_end\|>/;
  * @param thinkOnly  True if the response was only <think> blocks with no answer
  * @param finishReason  Ollama's finish_reason if available (e.g. "length" = truncated)
  * @param mode  Inference mode — "route": empty only when blank; "chat": empty only with no letter or digit; "code"/unset: 4 chars or fewer
+ * @param options.allowParallelCalls  Route mode: a reply of several complete calls is valid
  */
-export function passesQualityGate(stripped, thinkOnly, finishReason, mode) {
+export function passesQualityGate(stripped, thinkOnly, finishReason, mode, options = {}) {
     // Signal 1: Think-only — model reasoned but produced no answer (check before empty)
     if (thinkOnly) {
         return { pass: false, reason: "think_only" };
@@ -37,6 +48,22 @@ export function passesQualityGate(stripped, thinkOnly, finishReason, mode) {
     if (finishReason === "length") {
         return { pass: false, reason: "hard_truncation" };
     }
+    // Several complete calls, when the caller asked for them (route mode). The
+    // envelopes repeat by design, so the prose loop checks below would fail any
+    // three calls; a loop here is the same call again and again.
+    if (mode === "route" && options.allowParallelCalls) {
+        const several = parseRouteCalls(stripped);
+        if (several.kind === "tool_calls") {
+            const seen = new Map();
+            for (const c of several.calls) {
+                const key = canonicalJson([c.name, c.args]);
+                seen.set(key, (seen.get(key) ?? 0) + 1);
+                if ((seen.get(key) ?? 0) >= 3)
+                    return { pass: false, reason: "loop_detected" };
+            }
+            return { pass: true };
+        }
+    }
     // Signal 5: Tool-call bleed. The pipe envelope is invalid in chat/code,
     // but it is the canonical trained output in route mode. Route mode parses
     // the whole envelope and fails only when the contract is malformed.

package/dist/utils/routeContract.js

@@ -273,11 +273,76 @@ export function routeServesProse(output) {
     const last = ends.reduce((a, b) => (b.at > a.at ? b : a));
     return t.slice(last.at + last.e.length).trim() !== ""; // text after the envelope
 }
-export function applyLocalRouteContract(draft, allowedTools = DEFAULT_PRISM_ROUTE_TOOLS) {
+const OPENERS = [PIPE_START, ANGLE_START];
+const MAX_PARALLEL_CALLS = 64;
+/**
+ * A reply of complete tool-call envelopes one after another, with only
+ * whitespace between them. Every envelope must be closed and hold a valid
+ * call; any text, an unclosed envelope or a bad body makes the whole reply
+ * malformed. A single envelope is left to parseRouteOutput.
+ */
+export function parseRouteCalls(output) {
+    if (output.length > MAX_ROUTE_OUTPUT_CHARS)
+        return { kind: "malformed" };
+    const calls = [];
+    let rest = output.trim();
+    while (rest.length > 0) {
+        const opener = OPENERS.find(o => rest.startsWith(o));
+        if (!opener || calls.length >= MAX_PARALLEL_CALLS)
+            return { kind: "malformed" };
+        let end = -1;
+        let endToken = "";
+        for (const token of END_TOKENS) {
+            const at = rest.indexOf(token, opener.length);
+            if (at >= 0 && (end < 0 || at < end)) {
+                end = at;
+                endToken = token;
+            }
+        }
+        if (end < 0)
+            return { kind: "malformed" };
+        const body = rest.slice(opener.length, end).trim();
+        if (OPENERS.some(o => body.includes(o)))
+            return { kind: "malformed" };
+        const parsed = parseToolJson(body);
+        if (parsed.kind !== "tool_call")
+            return { kind: "malformed" };
+        calls.push({ name: parsed.name, args: parsed.args });
+        rest = rest.slice(end + endToken.length).trim();
+    }
+    return calls.length > 1 ? { kind: "tool_calls", calls } : { kind: "malformed" };
+}
+export function applyLocalRouteContract(draft, allowedTools = DEFAULT_PRISM_ROUTE_TOOLS, options = {}) {
     const parsed = parseRouteOutput(draft);
     if (parsed.kind === "plain_text") {
         return { output: draft, action: "plain_text", source: "local" };
     }
+    if (parsed.kind === "malformed" && options.allowParallel) {
+        const several = parseRouteCalls(draft);
+        if (several.kind === "tool_calls") {
+            if (several.calls.some(c => c.name === "NO_TOOL")) {
+                return { output: "NO_TOOL", action: "suppressed", source: "local", reason: "malformed_tool_call" };
+            }
+            const unadvertised = several.calls.find(c => !allowedTools.has(c.name));
+            if (unadvertised) {
+                return {
+                    output: "NO_TOOL",
+                    action: "suppressed",
+                    source: "local",
+                    original_tool: unadvertised.name,
+                    reason: "unadvertised_tool",
+                };
+            }
+            return {
+                output: draft,
+                action: "preserved",
+                source: "local",
+                original_tool: several.calls[0].name,
+                final_tool: several.calls[0].name,
+                calls: several.calls.map(c => c.name),
+            };
+        }
+    }
     if (parsed.kind === "malformed") {
         return {
             output: "NO_TOOL",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prism-mcp-server",
-  "version": "20.21.16",
+  "version": "20.21.17",
   "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",