@algosuite/vo-mcp 0.2.0-beta.30 → 0.2.0-beta.33

package/dist/agent-auth-probe-cli.mjs

@@ -1118,8 +1118,180 @@ var CursorRunner = class {
 };
 var cursorRunner = new CursorRunner();
 
+// ../../scripts/virtual-office/code-runner/ollama-agent-tools.mjs
+var MAX_READ_BYTES = 64 * 1024;
+var MAX_WRITE_BYTES = 512 * 1024;
+var TOOL_DEFS = [
+  {
+    type: "function",
+    function: {
+      name: "read_file",
+      description: "Read a UTF-8 text file inside the working directory. Returns up to 64 KiB.",
+      parameters: {
+        type: "object",
+        properties: {
+          path: { type: "string", description: "File path relative to the working directory." }
+        },
+        required: ["path"]
+      }
+    }
+  },
+  {
+    type: "function",
+    function: {
+      name: "list_files",
+      description: "List entries in a directory inside the working directory (files and subdirs).",
+      parameters: {
+        type: "object",
+        properties: {
+          path: { type: "string", description: 'Directory path relative to the working directory. Default ".".' }
+        }
+      }
+    }
+  },
+  {
+    type: "function",
+    function: {
+      name: "write_file",
+      description: "Create or overwrite a UTF-8 text file inside the working directory. Parent dirs are created.",
+      parameters: {
+        type: "object",
+        properties: {
+          path: { type: "string", description: "File path relative to the working directory." },
+          content: { type: "string", description: "Full new file contents." }
+        },
+        required: ["path", "content"]
+      }
+    }
+  }
+];
+var TOOL_NAMES = TOOL_DEFS.map((t) => t.function.name);
+var READ_ONLY_TOOL_NAMES = Object.freeze(["read_file", "list_files"]);
+var READ_ONLY_TOOL_DEFS = Object.freeze(
+  TOOL_DEFS.filter((tool) => READ_ONLY_TOOL_NAMES.includes(tool.function.name))
+);
+
+// ../../scripts/virtual-office/code-runner/ollama-agent-core.mjs
+var MAX_TURNS_DEFAULT = 20;
+var NUM_CTX_DEFAULT = 16384;
+function parseOllamaAgentEvent(line) {
+  const trimmed = String(line || "").trim();
+  if (!trimmed) return null;
+  let evt;
+  try {
+    evt = JSON.parse(trimmed);
+  } catch {
+    return null;
+  }
+  if (!evt || typeof evt !== "object") return null;
+  if (evt.type === "progress") {
+    const text = String(evt.text || "").trim();
+    return text ? { kind: "progress", text } : null;
+  }
+  if (evt.type === "tool") {
+    const via = evt.recovered ? " (recovered from text)" : "";
+    const label = `${evt.ok === false ? "tool failed" : "tool"}: ${evt.name}${evt.path ? ` ${evt.path}` : ""}${via}`;
+    return { kind: "progress", text: label };
+  }
+  if (evt.type === "result") {
+    const usage = evt.usage || null;
+    const tokenUsage = usage ? { inputTokens: usage.inputTokens ?? null, outputTokens: usage.outputTokens ?? null, totalTokens: usage.totalTokens ?? null } : void 0;
+    return {
+      kind: "result",
+      isError: Boolean(evt.isError),
+      costUsd: 0,
+      // sovereign local inference is free — no meter.
+      summary: String(evt.summary || (evt.isError ? "local run failed" : "completed")),
+      numTurns: Number.isInteger(evt.numTurns) ? evt.numTurns : null,
+      ...tokenUsage ? { tokenUsage } : {},
+      // Pass the sovereign receipt through to the daemon. Omitted entirely when
+      // absent so the event shape is unchanged for every other transport.
+      ...evt.receipt ? { receipt: evt.receipt } : {}
+    };
+  }
+  return null;
+}
+
+// ../../scripts/virtual-office/code-runner/ollama-native-transport.mjs
+import { fileURLToPath } from "node:url";
+import { dirname, join } from "node:path";
+var DEFAULT_LOCAL_TRANSPORT = "codex";
+var LOCAL_NATIVE_PROFILES = Object.freeze(["coding", "verification"]);
+var DEFAULT_LOCAL_NATIVE_PROFILE = "coding";
+function resolveLocalNativeProfile(env = process.env) {
+  const profile = String(env.VO_CODE_RUNNER_LOCAL_PROFILE || "").trim().toLowerCase() || DEFAULT_LOCAL_NATIVE_PROFILE;
+  if (!LOCAL_NATIVE_PROFILES.includes(profile)) {
+    throw new Error(`local-model runner (native): unknown profile "${profile}" (coding|verification).`);
+  }
+  return profile;
+}
+function resolveLocalTransport(env = process.env) {
+  return String(env.VO_CODE_RUNNER_LOCAL_TRANSPORT || "").trim().toLowerCase() === "native" ? "native" : DEFAULT_LOCAL_TRANSPORT;
+}
+function ollamaAgentScriptPath() {
+  return join(dirname(fileURLToPath(import.meta.url)), "ollama-agent.mjs");
+}
+function posIntOr(raw, fallback) {
+  const n = Number(String(raw ?? "").trim());
+  return Number.isInteger(n) && n > 0 ? n : fallback;
+}
+function buildOllamaAgentArgs({ model, numCtx, maxTurns, profile = DEFAULT_LOCAL_NATIVE_PROFILE } = {}) {
+  return [
+    ollamaAgentScriptPath(),
+    "--model",
+    String(model),
+    "--profile",
+    String(profile),
+    "--num-ctx",
+    String(posIntOr(numCtx, NUM_CTX_DEFAULT)),
+    "--max-turns",
+    String(posIntOr(maxTurns, MAX_TURNS_DEFAULT))
+  ];
+}
+function buildLocalNativeArgs(opts = {}, env = process.env) {
+  const provider = resolveLocalProvider(env);
+  if (provider !== "ollama") {
+    throw new Error(
+      `local-model runner (native): the native tool-loop executor speaks Ollama's /api/chat; provider "${provider}" is not supported on native transport. Set VO_CODE_RUNNER_LOCAL_PROVIDER=ollama, or use VO_CODE_RUNNER_LOCAL_TRANSPORT=codex for LM Studio.`
+    );
+  }
+  const model = resolveLocalModel(env);
+  if (!model) {
+    throw new Error(
+      "local-model runner (native): set VO_CODE_RUNNER_LOCAL_MODEL to a coding model your Ollama server already has (e.g. qwen2.5-coder:7b). Refusing to run with no explicit model (fail-closed)."
+    );
+  }
+  if (!isValidLocalModel(model)) {
+    throw new Error(`local-model runner (native): "${model}" is not a valid local model id.`);
+  }
+  const baseUrl = resolveLocalBaseUrl(env);
+  if (baseUrl && !isLoopbackBaseUrl(baseUrl)) {
+    throw new Error(
+      "local-model runner (native): VO_CODE_RUNNER_LOCAL_BASE_URL must be a loopback http(s) URL (localhost / 127.0.0.1 / [::1]). Remote endpoints are refused \u2014 use the Model Firewall lanes."
+    );
+  }
+  return buildOllamaAgentArgs({
+    model,
+    profile: resolveLocalNativeProfile(env),
+    numCtx: env.VO_CODE_RUNNER_LOCAL_NUM_CTX,
+    maxTurns: env.VO_CODE_RUNNER_LOCAL_MAX_TURNS
+  });
+}
+
 // ../../scripts/virtual-office/code-runner/local-model-runner.mjs
 var LOCAL_API_KEY_ENV = "VO_CODE_RUNNER_LOCAL_API_KEY";
+var FORBIDDEN_LOCAL_CHILD_CREDENTIALS = /* @__PURE__ */ new Set([
+  "ANTHROPIC_API_KEY",
+  "AWS_ACCESS_KEY_ID",
+  "AWS_SECRET_ACCESS_KEY",
+  "AWS_SESSION_TOKEN",
+  "FIREBASE_TOKEN",
+  "GH_TOKEN",
+  "GITHUB_TOKEN",
+  "GOOGLE_API_KEY",
+  "GOOGLE_APPLICATION_CREDENTIALS",
+  "OPENAI_API_KEY"
+]);
 var LOCAL_PROVIDERS = ["ollama", "lmstudio"];
 var DEFAULT_LOCAL_PROVIDER = "ollama";
 var LOCAL_PROBE_URLS = {
@@ -1196,6 +1368,9 @@ function buildLocalArgs(opts = {}, env = process.env) {
 }
 function applyLocalAuthEnv(baseEnv = process.env, configEnv = process.env) {
   const out = withAgentKey("local", baseEnv);
+  for (const key of Object.keys(out)) {
+    if (FORBIDDEN_LOCAL_CHILD_CREDENTIALS.has(key.toUpperCase())) delete out[key];
+  }
   const baseUrl = resolveLocalBaseUrl(configEnv);
   if (baseUrl && isLoopbackBaseUrl(baseUrl) && resolveLocalProvider(configEnv) === "ollama" && !String(out.OLLAMA_HOST || "").trim()) {
     out.OLLAMA_HOST = baseUrl.replace(/\/+$/, "");
@@ -1214,19 +1389,25 @@ var LocalModelRunner = class {
     this.env = env;
     this.fetchImpl = fetchImpl;
   }
-  /** Codex is the transport binary; the model/endpoint are the user's. */
+  /**
+   * Transport binary. codex transport → the codex CLI. native transport →
+   * this daemon's own node (process.execPath), which runs ollama-agent.mjs; no
+   * external CLI is involved on the native path.
+   */
   get binary() {
-    return this.resolveBinary();
+    return resolveLocalTransport(this.env) === "native" ? process.execPath : this.resolveBinary();
   }
   buildArgs(opts = {}) {
-    return buildLocalArgs(opts, this.env);
+    return resolveLocalTransport(this.env) === "native" ? buildLocalNativeArgs(opts, this.env) : buildLocalArgs(opts, this.env);
   }
   /**
-   * Codex JSONL events map identically, except a sovereign local model has no
-   * vendor bill. Codex omits total_cost_usd for OSS runs; turn that known fact
-   * into a measured zero at the producer so readers never have to guess.
+   * native transport → parse the executor's own JSONL contract. codex transport
+   * → codex JSONL maps identically, except a sovereign local model has no vendor
+   * bill: codex omits total_cost_usd for OSS runs, so turn that known fact into a
+   * measured zero at the producer. (The native parser already stamps costUsd:0.)
    */
   parseEvent(line) {
+    if (resolveLocalTransport(this.env) === "native") return parseOllamaAgentEvent(line);
     const event = parseCodexEvent(line);
     return event?.kind === "result" ? { ...event, costUsd: 0 } : event;
   }

package/dist/cli.js

@@ -1079,6 +1079,42 @@ function toEventPerModelVerdicts(src) {
     };
   });
 }
+function aggregateEventTokenUsage(src, engineUsage) {
+  if (engineUsage !== void 0) {
+    const hasIn = Object.keys(engineUsage.per_model_tokens_in).length > 0;
+    const hasOut = Object.keys(engineUsage.per_model_tokens_out).length > 0;
+    return {
+      per_model_tokens_in: hasIn ? engineUsage.per_model_tokens_in : null,
+      per_model_tokens_out: hasOut ? engineUsage.per_model_tokens_out : null,
+      total_cost_usd: engineUsage.cost_micro_usd === null ? null : engineUsage.cost_micro_usd / 1e6
+    };
+  }
+  const tokensIn = {};
+  const tokensOut = {};
+  let anyTokensIn = false;
+  let anyTokensOut = false;
+  let costMicroUsd = 0;
+  let anyCost = false;
+  for (const v of src) {
+    if (typeof v.input_tokens === "number") {
+      tokensIn[v.model] = (tokensIn[v.model] ?? 0) + v.input_tokens;
+      anyTokensIn = true;
+    }
+    if (typeof v.output_tokens === "number") {
+      tokensOut[v.model] = (tokensOut[v.model] ?? 0) + v.output_tokens;
+      anyTokensOut = true;
+    }
+    if (typeof v.cost_micro_usd === "number") {
+      costMicroUsd += v.cost_micro_usd;
+      anyCost = true;
+    }
+  }
+  return {
+    per_model_tokens_in: anyTokensIn ? tokensIn : null,
+    per_model_tokens_out: anyTokensOut ? tokensOut : null,
+    total_cost_usd: anyCost ? costMicroUsd / 1e6 : null
+  };
+}
 function toEventSynthesizedVerdict(src) {
   return {
     verdict: src.verdict,
@@ -2778,7 +2814,8 @@ async function handleCheckHollowTest(deps, rawInput, signal) {
     synthesized_verdict: synthForEvent,
     consensus_confidence: engineResult.synthesized_verdict.confidence,
     duration_ms: engineResult.duration_ms,
-    consensus_engine_version: engineResult.engine_version
+    consensus_engine_version: engineResult.engine_version,
+    ...aggregateEventTokenUsage(engineResult.per_model_verdicts, engineResult.token_usage)
   };
   const payload = {
     verdict: engineResult.synthesized_verdict.verdict,
@@ -2961,7 +2998,8 @@ async function handleVerifyAnswer(deps, rawInput, signal) {
     synthesized_verdict: synthForEvent,
     consensus_confidence: engineResult.synthesized_verdict.confidence,
     duration_ms: engineResult.duration_ms,
-    consensus_engine_version: engineResult.engine_version
+    consensus_engine_version: engineResult.engine_version,
+    ...aggregateEventTokenUsage(engineResult.per_model_verdicts, engineResult.token_usage)
   };
   const payload = {
     verdict: engineResult.synthesized_verdict.verdict,
@@ -2970,6 +3008,7 @@ async function handleVerifyAnswer(deps, rawInput, signal) {
     synthesized_verdict: synthForEvent,
     engine_version: engineResult.engine_version,
     degraded: engineResult.degraded,
+    ...engineResult.quorum_failed === true ? { quorum_failed: true } : {},
     gate_type: gateType,
     ...kbResult.error !== null ? { kb_unavailable: true } : {},
     ...kbTruncated > 0 ? { kb_rules_truncated: kbTruncated } : {}
@@ -3226,7 +3265,8 @@ async function handleConsensusJudgment(deps, rawInput, signal) {
     duration_ms: engineResult.duration_ms,
     consensus_engine_version: engineResult.engine_version,
     per_model_verdicts: perModelForEvent,
-    synthesized_verdict: synthForEvent
+    synthesized_verdict: synthForEvent,
+    ...aggregateEventTokenUsage(engineResult.per_model_verdicts, engineResult.token_usage)
   };
   const payload = {
     verdict: engineResult.synthesized_verdict.verdict,
@@ -3235,6 +3275,7 @@ async function handleConsensusJudgment(deps, rawInput, signal) {
     synthesized_verdict: synthForEvent,
     engine_version: engineResult.engine_version,
     degraded: engineResult.degraded,
+    ...engineResult.quorum_failed === true ? { quorum_failed: true } : {},
     gate_type: gateType,
     // ─── Consensus-engine feature outputs (additive; 2026-06-13) ─────────────
     // Feature 2 (calibrated-confidence) — ON by default; the engine attaches
@@ -3435,7 +3476,8 @@ async function handleArchitectureReview(deps, rawInput, signal) {
     synthesized_verdict: synthForEvent,
     consensus_confidence: engineResult.synthesized_verdict.confidence,
     duration_ms: engineResult.duration_ms,
-    consensus_engine_version: engineResult.engine_version
+    consensus_engine_version: engineResult.engine_version,
+    ...aggregateEventTokenUsage(engineResult.per_model_verdicts, engineResult.token_usage)
   };
   const escalationRequired = engineResult.escalation_required === true || engineResult.escalation_required === void 0 && engineResult.synthesized_verdict.dissent_summary !== null;
   const escalationReason = engineResult.escalation_reason ?? engineResult.synthesized_verdict.dissent_summary ?? "";
@@ -4922,7 +4964,8 @@ Produce the JSON dispatch plan now.`;
     duration_ms: engineResult.duration_ms,
     consensus_engine_version: engineResult.engine_version,
     per_model_verdicts: toEventPerModelVerdicts(engineResult.per_model_verdicts),
-    synthesized_verdict: toEventSynthesizedVerdict(engineResult.synthesized_verdict)
+    synthesized_verdict: toEventSynthesizedVerdict(engineResult.synthesized_verdict),
+    ...aggregateEventTokenUsage(engineResult.per_model_verdicts, engineResult.token_usage)
   };
   deps.events.append(enrichedEvent);
   return jsonContent(envelope);
@@ -5796,7 +5839,8 @@ async function handleReviewMerge(deps, rawInput, signal) {
     duration_ms: result.duration_ms,
     consensus_engine_version: result.engine_version,
     per_model_verdicts: perModel,
-    synthesized_verdict: synth
+    synthesized_verdict: synth,
+    ...aggregateEventTokenUsage(result.per_model_verdicts, result.token_usage)
   });
 }
 
@@ -7801,6 +7845,14 @@ function shadowEnabled(env) {
   const norm = raw.trim().toLowerCase();
   return !(norm === "0" || norm === "false" || norm === "no" || norm === "off" || norm === "");
 }
+var MIN_RESPONDERS_ENV_VAR = "VO_CONSENSUS_MIN_RESPONDERS";
+function resolveMinResponders(env) {
+  const raw = (env ?? {})[MIN_RESPONDERS_ENV_VAR];
+  if (raw === void 0 || raw.trim() === "") return 2;
+  const parsed = Number.parseInt(raw.trim(), 10);
+  if (!Number.isFinite(parsed) || parsed <= 0) return void 0;
+  return parsed;
+}
 function mapShadowSynthesis(s) {
   if (s === void 0) return void 0;
   return {
@@ -7950,9 +8002,11 @@ function createEngineConsensusClient(options) {
           ...options.agreement_gate_enabled !== void 0 ? { configEnabled: options.agreement_gate_enabled } : {},
           ...options.env !== void 0 ? { env: options.env } : {}
         });
+        const minResponders = resolveMinResponders(options.env);
         const engineOptions = {
           panel,
           ...options.per_model_timeout_ms !== void 0 ? { per_model_timeout_ms: options.per_model_timeout_ms } : {},
+          ...minResponders !== void 0 ? { min_responders: minResponders } : {},
           ...agreementGate !== void 0 ? { agreement_gate: agreementGate } : {},
           // Stage A7-shadow: run the adaptive verdict alongside the live one for grading.
           // Cheap (pure log-odds over already-fetched verdicts; no extra model calls),
@@ -8005,8 +8059,13 @@ function createEngineConsensusClient(options) {
           synthesized_verdict: response.synthesized_verdict,
           per_model_verdicts: response.per_model_verdicts,
           degraded: response.degraded,
+          ...response.quorum_failed === true ? { quorum_failed: true } : {},
           duration_ms: response.duration_ms,
           engine_version: response.engine_version,
+          // Cumulative cross-round inference usage (B44-3). Absent when no panel
+          // member reported usage; forwarded verbatim — the aggregator prefers it
+          // over summing final-round verdicts (which under-reports deliberation).
+          ...response.token_usage !== void 0 ? { token_usage: response.token_usage } : {},
           // Phase 2 Lane D-1 — forward escalation signal when present. The
           // source-grounded layer's own escalation (from the citation grade)
           // takes precedence when set, else the synthesizer's.