agent-valley 3.0.0 → 3.0.1

package/dist/assets/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
   "name": "av",
-  "version": "3.0.0",
+  "version": "3.0.1",
   "description": "Delegate repository goals to Agent Valley and inspect mission progress and verification evidence.",
   "author": {
     "name": "Agent Valley"

package/dist/index.js

@@ -16084,17 +16084,12 @@ function buildDarwinSandboxCommand(req, sandboxExecPath = "/usr/bin/sandbox-exec
   };
 }
 function buildSeatbeltProfile(req, home) {
-  const tmp = tmpdir();
+  const tmpPaths = pathAliases(tmpdir());
   const git = linkedWorktreeGitPaths(req.workspacePath);
   const agentPaths = agentHomeAccess(req.agentType, home, req.env);
-  let tmpReal = tmp;
-  try {
-    tmpReal = realpathSync4(tmp);
-  } catch {}
   const writablePaths = [
     req.workspacePath,
-    tmp,
-    ...tmpReal !== tmp ? [tmpReal] : [],
+    ...tmpPaths,
     ...agentPaths.active,
     `${home}/.cache`,
     `${home}/.npm`,
@@ -16120,7 +16115,7 @@ function buildSeatbeltProfile(req, home) {
   ];
   const denyReadLiterals = [
     `${home}/.git-credentials`,
-    ...new Set([process.cwd(), req.workspacePath].map((root) => join7(root, PROJECT_CONFIG_FILENAME)))
+    ...new Set([process.cwd(), req.workspacePath].flatMap((root) => pathAliases(root).flatMap((path) => pathAliases(join7(path, PROJECT_CONFIG_FILENAME)))))
   ];
   const denyReadRules = [
     ...denyReadSubpaths.map((p) => `(deny file-read* (subpath ${seatbeltString(p)}))`),
@@ -16171,6 +16166,14 @@ ${denyWriteRules}
 (allow network-outbound (remote tcp "*:80"))
 `;
 }
+function pathAliases(path) {
+  try {
+    const canonical = realpathSync4(path);
+    return canonical === path ? [path] : [path, canonical];
+  } catch {
+    return [path];
+  }
+}
 function seatbeltString(value) {
   return `"${value.replace(/\\/g, "\\\\").replace(/"/g, "\\\"")}"`;
 }
@@ -16396,11 +16399,14 @@ class BaseSession {
   _process = null;
   terminalEventReceived = false;
   processFailed = false;
+  sandboxPlatform = null;
   get process() {
     return this._process;
   }
   set process(proc) {
     this._process = proc;
+    if (!proc)
+      this.sandboxPlatform = null;
     if (proc) {
       this.processFailed = false;
       this.terminalEventReceived = false;
@@ -16417,6 +16423,7 @@ class BaseSession {
     }
   }
   async spawnSandboxed(options) {
+    this.sandboxPlatform = null;
     if (!this.config) {
       this.emitError("CRASH", "spawnSandboxed() called before start()", false);
       return false;
@@ -16437,6 +16444,7 @@ class BaseSession {
       this.emitError("CRASH", `${err}`, false);
       return false;
     }
+    this.sandboxPlatform = plan.sandboxed ? plan.platform : null;
     this.process = spawn3(plan.command, plan.args, {
       detached: process.platform !== "win32",
       cwd: this.config.workspacePath,
@@ -16646,7 +16654,8 @@ var init_codex_session = __esm(() => {
       this.threadId = threadResult.thread.id;
       const turnResult = await this.rpc("turn/start", {
         threadId: this.threadId,
-        input: [{ type: "text", text: prompt }]
+        input: [{ type: "text", text: prompt }],
+        ...this.sandboxPlatform === "darwin" ? { sandboxPolicy: { type: "externalSandbox", networkAccess: "restricted" } } : {}
       });
       if (!this.terminalEventReceived && turnResult.turn)
         this.turnId = turnResult.turn.id;
@@ -68804,6 +68813,47 @@ file '${this.frames.at(-1)?.attachment.name}'
 // ../../packages/core/src/chief/schemas.ts
 init_container_observation_policy();
 
+// ../../packages/core/src/chief/review-response.ts
+function reviewResponseFormat(summary, criteria) {
+  const passing = {
+    passed: true,
+    summary,
+    findings: [],
+    ...criteria ? {
+      criteria: criteria.map((criterion) => ({
+        criterion,
+        passed: true,
+        evidence: "actual inspected path or check"
+      }))
+    } : {}
+  };
+  const rejected = {
+    ...passing,
+    passed: false,
+    summary: "An inspected requirement or evidence gap needs attention.",
+    findings: [
+      "Concern: describe the gap. Evidence: cite the inspected path or failed check. Action: give the concrete repair or next check."
+    ],
+    ...passing.criteria ? { criteria: passing.criteria.map((entry, index) => ({ ...entry, passed: index !== 0 })) } : {}
+  };
+  return [
+    "Return ONLY one JSON object under 8,000 characters, without Markdown fences or text outside JSON.",
+    `Use exactly these top-level keys: passed, summary, findings${criteria ? ", criteria" : ""}. Do not add other top-level keys or wrap the response in another object.`,
+    "passed must be a boolean, summary must be a nonempty string, and findings must be an array of nonempty strings with at most 50 entries. Do not put objects, arrays, numbers or null in findings. Put each concern, inspected basis and concrete recommendation together inside one finding string; do not use concern/basis/recommendation objects.",
+    "Use findings:[] for passed:true. For passed:false, include at least one actionable finding. Do not omit required keys.",
+    ...criteria ? [
+      "criteria must assess every requested criterion exactly once. Each criteria entry has exactly criterion, passed and evidence: criterion and evidence are nonempty strings, and passed is a boolean. Use the exact criterion text and actual inspected evidence. Overall passed must equal all criteria passing."
+    ] : [],
+    "The examples show response structure only. Replace example descriptions with actual inspected evidence; never claim an example check was performed.",
+    `Passing response example:
+${JSON.stringify(passing)}`,
+    `Rejected response example:
+${JSON.stringify(rejected)}`
+  ].join(`
+
+`);
+}
+
 // ../../packages/core/src/chief/technical-lead.ts
 import { createHash as createHash14 } from "node:crypto";
 var TECHNICAL_DIRECTOR_ROLE = "Serve as the Chief Director's permanent Technical Director counterpart. Rigorously minimize total cost and wasted work while maximizing execution efficiency. Protect the user goal and quality requirements; standardize on the existing stack, reuse maintained modules and native integrations, keep dependencies justified and consistent, and favor the smallest maintainable solution.";
@@ -68857,7 +68907,8 @@ function technicalReviewPrompt(mission) {
     "Flag unjustified dependencies, unnecessary migrations, duplicate implementations, or complexity that materially harms maintainability. Give each concern an inspected basis and a concrete smaller adequate alternative or required evidence so the Chief Director can plan or revise. A passing assessment has no unresolved findings; passed:false includes at least one actionable finding.",
     "This consultation is advisory, not an execution veto or an operator approval gate. The Chief Director makes the final business and execution tradeoffs, including when a justified higher cost improves profit or the goal outcome. Do not discard an opportunity solely because of uncertainty: identify assumptions, evidence gaps and a concrete economical validation step or alternative. Do not invent profits, prices or evidence.",
     "The Chief Director retains execution judgment, supervision and the final explanation; the Technical Director supplies independent technical advice. The operator retains ultimate operational responsibility. Resolve routine technical choices from evidence without asking for microapprovals or inventing approval pauses.",
-    'Return ONLY one JSON object under 8,000 characters: {"passed":true,"summary":"evidence-based technical advice with cost assumptions or unknowns","findings":[]}. For concerns use passed:false and actionable findings; the Chief Director decides how to act on them.',
+    reviewResponseFormat("evidence-based technical advice with cost assumptions or unknowns"),
+    "For concerns use passed:false and actionable findings; the Chief Director decides how to act on them.",
     [
       "Technical review input (JSON data):",
       JSON.stringify({
@@ -68900,7 +68951,8 @@ function designReviewPrompt(mission) {
     "If actual data is absent or insufficient, state what is unknown and propose the smallest useful instrumentation or user/field test: the specific journey or hypothesis, relevant user persona, observation or metric, comparison and decision rule. Propose these steps as advice; do not execute them in this read-only stage. Avoid false certainty and do not discard an opportunity solely because evidence is not yet available.",
     "Recommend the smallest adequate design approach that preserves the operator goal, any immutable success criteria, existing acceptance obligations and fixed verification command. Flag usability or evidence gaps with inspected examples and concrete alternatives; do not weaken requirements or demand an unrelated redesign.",
     "This consultation is advisory, not an execution veto or operator approval gate. The Chief Director owns the final business and execution judgment, supervision and explanation; the operator retains ultimate operational responsibility. Resolve routine choices from evidence without asking for microapprovals. Existing execution permissions remain in force.",
-    'Return ONLY one JSON object under 8,000 characters: {"passed":true,"summary":"evidence-based design advice: alternatives, recommendation, data or unknowns, and retention tradeoffs","findings":[]}. Use passed:false with actionable concerns when evidence or usability needs improvement; the Chief Director decides how to act on them.',
+    reviewResponseFormat("evidence-based design advice: alternatives, recommendation, data or unknowns, and retention tradeoffs"),
+    "Use passed:false with actionable concerns when evidence or usability needs improvement; the Chief Director decides how to act on them.",
     [
       "Design review input (JSON data):",
       JSON.stringify({
@@ -68945,7 +68997,8 @@ function marketingReviewPrompt(mission) {
     "If real data is missing, state the gap and propose minimal measurement before claiming ROI improved. Do not discard a promising route solely because it is uncertain; distinguish a sourced fact, hypothesis and economical validation opportunity. Compare short-term revenue with costs, retention and payback instead of equating activity or clicks with money.",
     "Keep recommendations tied to the operator goal, any immutable success criteria, existing acceptance obligations and fixed verification command. Do not replace the goal with unrelated promotion or weaken requirements to report a better ROI. Recommend a concrete money-making option or improvement with its evidence and tradeoffs.",
     "This consultation is advisory, not an execution veto or operator approval gate. The Chief Director owns the final business and execution judgment, supervision and explanation; the operator retains ultimate operational responsibility. Resolve routine choices from evidence without asking for microapprovals. Existing execution permissions remain in force; this Actor preference does not grant outreach, publication or spending permission.",
-    'Return ONLY one JSON object under 8,000 characters: {"passed":true,"summary":"evidence-based marketing advice: audience, revenue routes, recommendation, unit economics or unknowns, and measurable next test","findings":[]}. A passing assessment has no unresolved findings. Use passed:false with actionable concerns or missing evidence; the Chief Director decides how to act on them.',
+    reviewResponseFormat("evidence-based marketing advice: audience, revenue routes, recommendation, unit economics or unknowns, and measurable next test"),
+    "A passing assessment has no unresolved findings. Use passed:false with actionable concerns or missing evidence; the Chief Director decides how to act on them.",
     [
       "Marketing review input (JSON data):",
       JSON.stringify({
@@ -69856,6 +69909,7 @@ ${JSON.stringify({
 `);
 }
 function reviewPrompt(mission, task, state) {
+  const criteria = !task && (mission.supervision || mission.operatingPolicy?.metricTargets?.length) ? finalCriteria(mission) : undefined;
   return [
     task ? "Independently review this task's changes and acceptance evidence." : "Perform the Chief Director's final review of the complete mission.",
     ...!task ? [CHIEF_DIRECTOR_ROLE, CHIEF_DECISION_RESPONSIBILITY] : [],
@@ -69870,7 +69924,10 @@ function reviewPrompt(mission, task, state) {
     ] : [],
     "Return actionable findings for the Actor to repair. A passing review must have no unresolved findings.",
     "Keep the complete JSON response under 8,000 characters; prioritize actionable blockers.",
-    !task && mission.supervision ? `Assess every goal success criterion in this exact JSON list once with a concise nonempty description of inspected files/checks: ${JSON.stringify(finalCriteria(mission))}. Every task acceptance obligation also needs its fresh independent task review. passed must equal all goal criteria passing; failed criteria need actionable findings. Return ONLY JSON: {"passed":true,"summary":"evidence-based assessment","findings":[],"criteria":[{"criterion":"exact criterion text","passed":true,"evidence":"actual inspected file or check"}]}` : 'Return ONLY JSON: {"passed":true,"summary":"evidence-based assessment","findings":[]}',
+    ...criteria ? [
+      `Assess every goal success criterion in this exact JSON list once with a concise nonempty description of inspected files/checks: ${JSON.stringify(criteria)}. Every task acceptance obligation also needs its fresh independent task review. passed must equal all goal criteria passing; failed criteria need actionable findings.`
+    ] : [],
+    reviewResponseFormat("evidence-based assessment", criteria),
     `Review input (JSON data):
 ${JSON.stringify({
       mission: JSON.parse(context(mission)),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-valley",
-  "version": "3.0.0",
+  "version": "3.0.1",
   "type": "module",
   "engines": {
     "node": ">=26.10.0",