@ouro.bot/cli 0.1.0-alpha.937 → 0.1.0-alpha.938

package/changelog.json

@@ -1,6 +1,13 @@
 {
   "_note": "This changelog is maintained as part of the PR/version-bump workflow. Agent-curated, not auto-generated. Agents read this file directly via read_file to understand what changed between versions.",
   "versions": [
+    {
+      "version": "0.1.0-alpha.938",
+      "changes": [
+        "Shepherd rejects provider-filtered, refused or incomplete judgments before parsing or input, and retains safe completion metadata without refusal text or bypassing provider safeguards.",
+        "Judge evaluation reports explicit provider and malformed-answer error codes plus known failed-call input usage without classification credit; native admission and full v1 remain unqualified."
+      ]
+    },
     {
       "version": "0.1.0-alpha.937",
       "changes": [

package/deploy/unraid/sanctuary.ouro/bundle-meta.json

@@ -1,5 +1,5 @@
 {
-  "runtimeVersion": "0.1.0-alpha.937",
+  "runtimeVersion": "0.1.0-alpha.938",
   "bundleSchemaVersion": 3,
   "lastUpdated": "2026-09-03T00:00:00.000Z"
 }

package/deploy/unraid/sanctuary.xml

@@ -1,7 +1,7 @@
 <?xml version="1.0"?>
 <Container version="2">
   <Name>ouro-butler</Name>
-  <Repository>ghcr.io/ourostack/ouroboros-butler:0.1.0-alpha.937</Repository>
+  <Repository>ghcr.io/ourostack/ouroboros-butler:0.1.0-alpha.938</Repository>
   <Registry>https://github.com/ourostack/ouroboros/pkgs/container/ouroboros-butler</Registry>
   <Network>host</Network>
   <Shell>sh</Shell>

package/dist/heart/streaming.js

@@ -581,6 +581,8 @@ async function streamChatCompletion(client, createParams, callbacks, signal, eag
     let toolCalls = {};
     let streamStarted = false;
     let usage;
+    let finishReason = null;
+    let refused = false;
     const answerStreamer = new SettleStreamer(callbacks, eagerSettleStreaming);
     const textOutput = new ProviderTextOutput(callbacks);
     // State machine for parsing inline <think> tags (MiniMax pattern)
@@ -666,7 +668,26 @@ async function streamChatCompletion(client, createParams, callbacks, signal, eag
                     total_tokens: u.total_tokens,
                 };
             }
-            const d = chunk.choices[0]?.delta;
+            const choice = chunk.choices[0];
+            if (choice?.finish_reason != null) {
+                let reason;
+                switch (choice.finish_reason) {
+                    case "stop":
+                    case "length":
+                    case "content_filter":
+                    case "tool_calls":
+                    case "function_call":
+                        reason = choice.finish_reason;
+                        break;
+                    default:
+                        reason = "unknown";
+                }
+                if (finishReason === null || finishReason === "stop" || reason === "content_filter")
+                    finishReason = reason;
+            }
+            if (choice?.delta?.refusal)
+                refused = true;
+            const d = choice?.delta;
             if (!d)
                 continue;
             // Handle reasoning_content (Azure AI models like DeepSeek-R1)
@@ -749,6 +770,7 @@ async function streamChatCompletion(client, createParams, callbacks, signal, eag
         toolCalls: completedToolCalls,
         outputItems: [],
         usage,
+        ...(finishReason !== null || refused ? { termination: { finishReason, refused } } : {}),
         settleStreamed: answerStreamer.streamed,
         ...(settleFinalization ? { settleFinalization } : {}),
     };

package/dist/senses/shepherd/judge-evaluation-command.js

@@ -82,8 +82,10 @@ async function loadConfiguredShepherdEvaluationJudge(agent) {
         resetTurnState: messages => runtime.resetTurnState?.(messages),
         streamTurn: async (request) => {
             const response = await runtime.streamTurn(request);
-            if (!(0, judge_1.isShepherdJudgeAnswerValid)(response.content ?? ""))
-                throw new Error("Judge returned malformed or nonconforming JSON; no classification credit");
+            (0, judge_1.assertShepherdJudgeCompletion)(response);
+            if (!(0, judge_1.isShepherdJudgeAnswerValid)(response.content ?? "")) {
+                throw new judge_1.ShepherdJudgeResponseError("invalid_answer", response.usage?.input_tokens ?? null);
+            }
             return response;
         },
     };
@@ -139,6 +141,7 @@ async function runShepherdJudgeEvaluationCommand(argv, dependencies = {
             phase: report.phase, classificationPassed: report.classificationPassed, v1Qualified: false,
             uniqueCases: report.uniqueCases, calls: report.results.length,
             unsafeProposedActions: report.unsafeProposedActions, unattendedKeys: report.unattendedKeys, errors: report.errors,
+            errorCounts: report.errorCounts,
             report: args.output, boundary: "Classification evidence only; not v1 qualification.",
         };
         exitCode = report.classificationPassed ? 0 : 1;

package/dist/senses/shepherd/judge-evaluation.js

@@ -91,14 +91,16 @@ async function runShepherdJudgeEvaluation(value, judge, provenance, onProgress)
                     kind: verdict.kind, reply: verdict.reply === null ? null : "text" in verdict.reply
                         ? { text: safeDetail(verdict.reply.text) } : verdict.reply,
                     replyIssue: verdict.replyIssue ?? null, reason: safeDetail(verdict.reason),
-                    latencyMs: verdict.latencyMs, inputTokens: verdict.inputTokens, correct: failure === null, failure });
+                    latencyMs: verdict.latencyMs, inputTokens: verdict.inputTokens, correct: failure === null, failure, errorCode: null });
             }
             catch (error) {
+                const errorCode = error instanceof judge_1.ShepherdJudgeResponseError ? error.code : "judge_error";
                 results.push({ caseId: entry.id, stratum: entry.stratum, repetition, expectedKind: entry.expectedKind,
                     kind: "error", reply: null, replyIssue: null, reason: safeDetail(String(error)),
-                    latencyMs: Date.now() - started, inputTokens: null, correct: false, failure: "judge_error" });
+                    latencyMs: Date.now() - started, inputTokens: error instanceof judge_1.ShepherdJudgeResponseError ? error.inputTokens : null,
+                    correct: false, failure: "judge_error", errorCode });
                 (0, runtime_1.emitNervesEvent)({ component: "senses", level: "error", event: "senses.shepherd_judge_evaluation_error",
-                    message: "one classification evaluation call failed", meta: { caseId: entry.id, repetition } });
+                    message: "one classification evaluation call failed", meta: { caseId: entry.id, repetition, errorCode } });
             }
             onProgress?.(results);
         }
@@ -118,9 +120,12 @@ async function runShepherdJudgeEvaluation(value, judge, provenance, onProgress)
     const unattendedKeys = results.filter(row => row.reply !== null && "key" in row.reply).length;
     const errors = results.filter(row => row.kind === "error").length;
     const confusion = {};
+    const errorCounts = {};
     for (const row of results) {
         const key = `${row.expectedKind}->${row.kind}`;
         confusion[key] = (confusion[key] ?? 0) + 1;
+        if (row.errorCode)
+            errorCounts[row.errorCode] = (errorCounts[row.errorCode] ?? 0) + 1;
     }
     const classificationPassed = errors === 0 && unsafeProposedActions === 0 && unattendedKeys === 0
         && Object.values(thresholds).every(value => value.passed);
@@ -136,7 +141,7 @@ async function runShepherdJudgeEvaluation(value, judge, provenance, onProgress)
         falsePrematureSafetyClassifications: results.filter(row => safety.has(row.stratum) && row.kind === "premature").length,
         totalInputTokens: results.reduce((sum, row) => sum + (row.inputTokens ?? 0), 0),
         missingInputTokenMeasurements: results.filter(row => row.inputTokens === null).length,
-        confusion,
+        confusion, errorCounts,
         strata: exports.SHEPHERD_JUDGE_STRATA.map(stratum => {
             const entries = corpus.cases.filter(entry => entry.stratum === stratum);
             const misses = entries.filter(entry => !passed.has(entry.id)).map(entry => entry.id);

package/dist/senses/shepherd/judge.js

@@ -1,6 +1,7 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.REPLY_MAX_CHARS = void 0;
+exports.REPLY_MAX_CHARS = exports.ShepherdJudgeResponseError = void 0;
+exports.assertShepherdJudgeCompletion = assertShepherdJudgeCompletion;
 exports.replyTextProblem = replyTextProblem;
 exports.safeReplyText = safeReplyText;
 exports.isShepherdJudgeAnswerValid = isShepherdJudgeAnswerValid;
@@ -9,6 +10,37 @@ exports.judgeMessages = judgeMessages;
 exports.createShepherdJudge = createShepherdJudge;
 const runtime_1 = require("../../nerves/runtime");
 const redact_1 = require("./redact");
+const JUDGE_ERRORS = {
+    provider_filtered: "The provider filtered the judge response; no classification or continuation",
+    provider_refused: "The provider refused the judge response; no classification or continuation",
+    provider_incomplete: "The provider did not complete a normal judge response; no classification or continuation",
+    invalid_answer: "Judge returned malformed or nonconforming JSON; no classification credit",
+};
+class ShepherdJudgeResponseError extends Error {
+    code;
+    inputTokens;
+    constructor(code, inputTokens) {
+        super(JUDGE_ERRORS[code]);
+        this.code = code;
+        this.inputTokens = inputTokens;
+        this.name = "ShepherdJudgeResponseError";
+    }
+}
+exports.ShepherdJudgeResponseError = ShepherdJudgeResponseError;
+function assertShepherdJudgeCompletion(result) {
+    const end = result.termination;
+    if (!end)
+        return;
+    const code = end.finishReason === "content_filter" ? "provider_filtered"
+        : end.refused ? "provider_refused"
+            : end.finishReason !== null && end.finishReason !== "stop" ? "provider_incomplete" : null;
+    if (!code)
+        return;
+    const inputTokens = result.usage?.input_tokens ?? null;
+    (0, runtime_1.emitNervesEvent)({ component: "senses", level: "error", event: "senses.shepherd_judge_provider_refused",
+        message: "provider termination forbids a judge decision", meta: { code, inputTokens } });
+    throw new ShepherdJudgeResponseError(code, inputTokens);
+}
 exports.REPLY_MAX_CHARS = 240;
 const KEYS = new Set(["enter", "escape", "up", "down", "tab", "1", "2", "3", "4", "5", "6", "7", "8", "9", "y", "n"]);
 const KINDS = new Set(["premature", "gate", "done", "waiting", "working", "unclear"]);
@@ -158,6 +190,7 @@ function createShepherdJudge(runtime, timeoutMs, now = Date.now) {
             const messages = judgeMessages(packet);
             provider.resetTurnState?.(messages);
             const result = await Promise.race([provider.streamTurn({ messages, activeTools: [], callbacks: CALLBACKS, signal: controller.signal, toolChoiceRequired: false, reasoningEffort: "low" }), timedOut]);
+            assertShepherdJudgeCompletion(result);
             const verdict = parseVerdict(result.content ?? "");
             (0, runtime_1.emitNervesEvent)({ component: "senses", event: "senses.shepherd_judged", message: "judged a returned control", meta: { kind: verdict.kind } });
             return { ...verdict, latencyMs: now() - started, inputTokens: result.usage?.input_tokens ?? null };

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "@ouro.bot/cli",
-  "version": "0.1.0-alpha.937",
+  "version": "0.1.0-alpha.938",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@ouro.bot/cli",
-      "version": "0.1.0-alpha.937",
+      "version": "0.1.0-alpha.938",
       "dependencies": {
         "@anthropic-ai/sdk": "^0.78.0",
         "@azure/identity": "^4.13.0",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ouro.bot/cli",
-  "version": "0.1.0-alpha.937",
+  "version": "0.1.0-alpha.938",
   "engines": {
     "node": ">=22"
   },