create-agdf 0.14.1 → 0.14.2

package/generated/plugins/copilot/agdf/runtime/create-agdf/lib/control-state/run-state-resolver.js

@@ -0,0 +1,45 @@
+import { discoverRuns } from "./run-state-repository.js";
+export function resolveRuns(
+  root,
+  { runIdArg, runIdEnv, allActive = false } = {},
+) {
+  if (
+    (allActive && (runIdArg || runIdEnv)) ||
+    (runIdArg && runIdEnv && runIdArg !== runIdEnv)
+  )
+    throw Error("AGDF_SELECTOR_CONFLICT");
+  const runs = discoverRuns(root),
+    active = runs.filter((r) => r.valid && r.meta.lifecycle === "active");
+  const findings = runs
+    .filter((run) => !run.valid)
+    .flatMap((run) =>
+      run.findings.map((finding) => ({ ...finding, run_id: run.run_id })),
+    );
+  if (allActive)
+    return { selection_source: "all_active", runs: active, findings };
+  const id = runIdArg || runIdEnv;
+  if (id) {
+    const run = runs.find((r) => r.run_id === id);
+    if (!run?.valid) {
+      throw Error(
+        `AGDF_RUN_NOT_SELECTABLE:${id}; discovered=${runs.map((item) => item.run_id).join(",")}; next=select a valid run`,
+      );
+    }
+    return {
+      selection_source: runIdArg ? "cli" : "environment",
+      run,
+      active_run_ids: active.map((r) => r.run_id),
+    };
+  }
+  if (active.length !== 1)
+    throw Error(
+      active.length
+        ? `AGDF_ACTIVE_RUN_AMBIGUOUS:${active.map((r) => r.run_id)}`
+        : "AGDF_ACTIVE_RUN_MISSING",
+    );
+  return {
+    selection_source: "single_active",
+    run: active[0],
+    active_run_ids: [active[0].run_id],
+  };
+}

package/generated/plugins/copilot/agdf/runtime/create-agdf/lib/control-state/run-state-writer.js

@@ -0,0 +1,97 @@
+import { randomUUID } from "node:crypto";
+import {
+  closeSync,
+  fsyncSync,
+  existsSync,
+  lstatSync,
+  openSync,
+  readFileSync,
+  renameSync,
+  unlinkSync,
+  writeFileSync,
+} from "node:fs";
+import { dirname } from "node:path";
+
+import { parseRunState } from "./run-state-parser.js";
+
+function fsyncDirectory(path) {
+  if (process.platform === "win32") return;
+  const descriptor = openSync(path, "r");
+  try {
+    fsyncSync(descriptor);
+  } finally {
+    closeSync(descriptor);
+  }
+}
+
+export function atomicWrite(path, content) {
+  if (existsSync(path)) {
+    const destination = lstatSync(path);
+    if (destination.isSymbolicLink() || !destination.isFile()) {
+      throw new Error("AGDF_RUN_PATH_INVALID");
+    }
+  }
+
+  const parent = lstatSync(dirname(path));
+  if (parent.isSymbolicLink() || !parent.isDirectory()) {
+    throw new Error("AGDF_RUN_PATH_INVALID");
+  }
+
+  const temp = `${path}.tmp-${process.pid}-${randomUUID()}`;
+  let descriptor;
+
+  try {
+    descriptor = openSync(temp, "wx");
+    writeFileSync(descriptor, content, "utf8");
+    fsyncSync(descriptor);
+    closeSync(descriptor);
+    descriptor = undefined;
+
+    renameSync(temp, path);
+    fsyncDirectory(dirname(path));
+  } catch (error) {
+    if (descriptor !== undefined) closeSync(descriptor);
+    try {
+      unlinkSync(temp);
+    } catch (cleanupError) {
+      if (cleanupError.code !== "ENOENT") throw cleanupError;
+    }
+    throw error;
+  }
+}
+
+export function writeRun(path, content, expectedRevisionId) {
+  const lockPath = `${path}.lock`;
+  let lockDescriptor;
+  try {
+    lockDescriptor = openSync(lockPath, "wx");
+  } catch (error) {
+    if (error.code === "EEXIST") throw new Error("AGDF_RUN_WRITE_LOCKED");
+    throw error;
+  }
+
+  try {
+    const current = parseRunState(readFileSync(path, "utf8"));
+    if (!current.valid) throw new Error("AGDF_RUN_STATE_INVALID");
+    if (current.meta.revision_id !== expectedRevisionId) {
+      throw new Error("AGDF_STALE_RUN_REVISION");
+    }
+
+    const next = content
+      .replace(
+        /^- revision:\s*.*$/m,
+        `- revision: ${Number(current.meta.revision) + 1}`,
+      )
+      .replace(/^- revision_id:\s*.*$/m, `- revision_id: ${randomUUID()}`);
+    const candidate = parseRunState(next);
+    if (!candidate.valid || candidate.meta.run_id !== current.meta.run_id) {
+      throw new Error("AGDF_RUN_STATE_INVALID");
+    }
+
+    atomicWrite(path, next);
+    return candidate;
+  } finally {
+    closeSync(lockDescriptor);
+    unlinkSync(lockPath);
+  }
+}

package/generated/plugins/copilot/agdf/runtime/create-agdf/lib/delivery-path-search/candidate-policy.js

@@ -0,0 +1,99 @@
+import { validateCandidate } from "./contracts.js";
+
+function normalized(value) {
+  return String(value).normalize("NFKC").toLowerCase().replace(/[`*_]/g, "").replace(/[^\p{L}\p{N}\s-]/gu, " ").replace(/\s+/g, " ").trim();
+}
+
+const STOP_WORDS = new Set(["a", "an", "and", "der", "die", "das", "ein", "eine", "the", "to", "und", "zu"]);
+
+function tokens(value) {
+  return new Set(normalized(value).split(/\s+/).filter((item) => item && !STOP_WORDS.has(item)));
+}
+
+function jaccard(left, right) {
+  const union = new Set([...left, ...right]);
+  if (!union.size) return 1;
+  let intersection = 0;
+  for (const item of left) if (right.has(item)) intersection += 1;
+  return intersection / union.size;
+}
+
+function signature(candidate) {
+  return JSON.stringify({
+    gate_action: normalized(candidate.gate_action ?? candidate.action),
+    affected_boundaries: [...(candidate.affected_boundaries ?? [])].map(normalized).sort(),
+    evidence: [...(candidate.expected_evidence ?? [])].map(normalized).sort(),
+    tests: [...(candidate.tests ?? [])].map(normalized).sort(),
+    risk_strategy: normalized(candidate.risk_strategy ?? ""),
+    reversibility: normalized(candidate.reversibility ?? ""),
+  });
+}
+
+export function candidateLegality(candidateValue, input) {
+  const candidate = validateCandidate(candidateValue);
+  const action = normalized(candidate.gate_action ?? candidate.action);
+  const forbiddenMatch = input.forbidden_actions.find((item) => {
+    const forbidden = normalized(item);
+    return forbidden && action === forbidden;
+  });
+  if (forbiddenMatch) return { legal: false, reason: `forbidden_by_gate: ${forbiddenMatch}`, candidate };
+
+  const allowedMatch = input.allowed_actions.some((item) => {
+    const allowed = normalized(item);
+    return allowed && action === allowed;
+  });
+  if (!allowedMatch) return { legal: false, reason: "not_in_allowed_actions", candidate };
+  return { legal: true, reason: "allowed_by_current_gate", candidate };
+}
+
+export function candidatesFromInput(input) {
+  return input.allowed_actions.slice(0, input.budgets.max_candidates).map((action, index) => ({
+    id: `candidate-${index + 1}`,
+    action,
+    expected_evidence: input.evidence_refs,
+    tests: [],
+    assumptions: [],
+    depth: 0,
+    parent_id: null,
+  }));
+}
+
+export function generatedCandidatesFromResponse(response, input, existingCandidates = []) {
+  const accepted = [];
+  const rejected = [];
+  for (const proposal of response.proposals) {
+    const candidate = {
+      id: `generated-${proposal.proposal_id}`,
+      action: proposal.intent,
+      source: "generated",
+      gate_action: proposal.gate_action,
+      intent: proposal.intent,
+      expected_evidence: proposal.expected_evidence,
+      tests: proposal.tests,
+      assumptions: proposal.assumptions,
+      affected_boundaries: proposal.affected_boundaries,
+      risk_strategy: proposal.risk_strategy,
+      reversibility: proposal.reversibility,
+      generator_proposal_id: proposal.proposal_id,
+      depth: 0,
+      parent_id: null,
+    };
+    const legality = candidateLegality(candidate, input);
+    if (!legality.legal) {
+      rejected.push({ proposal_id: proposal.proposal_id, reason: legality.reason });
+      continue;
+    }
+    const peers = [...existingCandidates, ...accepted].filter((item) => normalized(item.gate_action ?? item.action) === normalized(candidate.gate_action));
+    const duplicate = peers.find((item) => {
+      const same = normalized(item.intent ?? item.action) === normalized(candidate.intent);
+      const sameDecision = signature(item) === signature(candidate);
+      return same || sameDecision || jaccard(tokens(item.intent ?? item.action), tokens(candidate.intent)) >= 0.8;
+    });
+    if (duplicate) {
+      rejected.push({ proposal_id: proposal.proposal_id, reason: "cosmetic_or_material_duplicate" });
+      continue;
+    }
+    accepted.push(candidate);
+  }
+  return { accepted, rejected };
+}

package/generated/plugins/copilot/agdf/runtime/create-agdf/lib/delivery-path-search/contracts.js

@@ -0,0 +1,237 @@
+export const CONTRACT_VERSION = "1";
+export const GENERATOR_CONTRACT_VERSION = "1";
+export const SCORING_POLICY_VERSION = "1";
+export const SCORE_DIMENSIONS = [
+  "scope_fit",
+  "gate_readiness",
+  "risk_reduction",
+  "evidence_gain",
+  "testability",
+  "reversibility",
+  "cost",
+];
+
+function requireObject(value, label) {
+  if (!value || typeof value !== "object" || Array.isArray(value)) throw new Error(`${label} must be an object`);
+  return value;
+}
+
+function requireString(value, label) {
+  if (typeof value !== "string" || !value.trim()) throw new Error(`${label} must be a non-empty string`);
+  return value.trim();
+}
+
+function stringArray(value, label) {
+  if (!Array.isArray(value) || value.some((item) => typeof item !== "string")) throw new Error(`${label} must be a string array`);
+  return value.map((item) => item.trim()).filter(Boolean);
+}
+
+function optionalBoundedString(value, label, maxLength = 1000) {
+  if (value === undefined) return undefined;
+  const result = requireString(value, label);
+  if (result.length > maxLength) throw new Error(`${label} must be at most ${maxLength} characters`);
+  return result;
+}
+
+function rejectUnknownKeys(value, allowed, label) {
+  const unknown = Object.keys(value).filter((key) => !allowed.includes(key));
+  if (unknown.length) throw new Error(`${label} contains unknown fields: ${unknown.join(", ")}`);
+}
+
+function generationConfig(value, wholeRunBudgets) {
+  if (value === undefined) return undefined;
+  const generation = requireObject(value, "generation");
+  if (typeof generation.enabled !== "boolean") throw new Error("generation.enabled must be a boolean");
+  const limits = {
+    max_calls: 1,
+    max_proposals: 5,
+    max_duration_ms: 30000,
+    max_cost_units: 5,
+  };
+  for (const [key, maximum] of Object.entries(limits)) {
+    if (!Number.isInteger(generation[key]) || generation[key] < 1 || generation[key] > maximum) {
+      throw new Error(`generation.${key} must be an integer from 1 to ${maximum}`);
+    }
+  }
+  if (generation.max_calls !== 1) throw new Error("generation.max_calls must equal 1 in this contract version");
+  if (generation.max_duration_ms > wholeRunBudgets.max_duration_ms) throw new Error("generation.max_duration_ms cannot exceed budgets.max_duration_ms");
+  if (generation.max_cost_units > wholeRunBudgets.max_cost_units) throw new Error("generation.max_cost_units cannot exceed budgets.max_cost_units");
+  return structuredClone(generation);
+}
+
+export function validateSearchInput(value) {
+  const input = requireObject(value, "search input");
+  if (input.contract_version !== CONTRACT_VERSION) throw new Error(`unsupported search input contract_version: ${input.contract_version}`);
+  requireString(input.scope_key, "scope_key");
+  if (!/^[a-z0-9][a-z0-9._-]*$/i.test(input.scope_key) || input.scope_key.includes("..")) {
+    throw new Error("scope_key must be a safe path segment");
+  }
+  requireString(input.objective, "objective");
+  requireString(input.current_gate, "current_gate");
+  stringArray(input.allowed_actions, "allowed_actions");
+  stringArray(input.forbidden_actions, "forbidden_actions");
+  stringArray(input.evidence_refs ?? [], "evidence_refs");
+  stringArray(input.risks ?? [], "risks");
+  validateEnforcement(input.enforcement);
+  const budgets = requireObject(input.budgets, "budgets");
+  for (const key of ["max_candidates", "max_depth", "max_evaluations", "max_duration_ms", "max_cost_units", "stability_window"]) {
+    if (!Number.isInteger(budgets[key]) || budgets[key] < 1) throw new Error(`budgets.${key} must be a positive integer`);
+  }
+  generationConfig(input.generation, budgets);
+  return structuredClone(input);
+}
+
+export function validateCandidate(value) {
+  const candidate = requireObject(value, "candidate");
+  requireString(candidate.id, "candidate.id");
+  requireString(candidate.action, "candidate.action");
+  stringArray(candidate.expected_evidence ?? [], "candidate.expected_evidence");
+  stringArray(candidate.tests ?? [], "candidate.tests");
+  stringArray(candidate.assumptions ?? [], "candidate.assumptions");
+  if (candidate.source !== undefined && !["deterministic", "generated", "expanded"].includes(candidate.source)) {
+    throw new Error("candidate.source must be deterministic, generated or expanded");
+  }
+  optionalBoundedString(candidate.gate_action, "candidate.gate_action");
+  optionalBoundedString(candidate.intent, "candidate.intent");
+  stringArray(candidate.affected_boundaries ?? [], "candidate.affected_boundaries");
+  optionalBoundedString(candidate.risk_strategy, "candidate.risk_strategy");
+  optionalBoundedString(candidate.reversibility, "candidate.reversibility");
+  optionalBoundedString(candidate.generator_proposal_id, "candidate.generator_proposal_id", 200);
+  return structuredClone(candidate);
+}
+
+export function validateGeneratorRequest(value) {
+  const request = requireObject(value, "generator request");
+  rejectUnknownKeys(request, ["contract_version", "scope_key", "objective", "scope_summary", "current_gate", "allowed_actions", "forbidden_actions", "artefact_refs", "evidence", "missing_evidence", "risks", "constraints", "enforcement", "budgets"], "generator request");
+  if (request.contract_version !== GENERATOR_CONTRACT_VERSION) throw new Error(`unsupported generator contract_version: ${request.contract_version}`);
+  requireString(request.scope_key, "generator scope_key");
+  optionalBoundedString(request.objective, "generator objective", 2000);
+  optionalBoundedString(request.scope_summary, "generator scope_summary", 2000);
+  requireString(request.current_gate, "generator current_gate");
+  for (const key of ["allowed_actions", "forbidden_actions", "artefact_refs", "evidence", "missing_evidence", "risks", "constraints"]) {
+    const items = stringArray(request[key] ?? [], `generator ${key}`);
+    if (items.length > 20 || items.some((item) => item.length > 500)) throw new Error(`generator ${key} exceeds context bounds`);
+  }
+  validateEnforcement(request.enforcement);
+  const budgets = requireObject(request.budgets, "generator budgets");
+  generationConfig({ enabled: true, ...budgets }, { max_duration_ms: 120000, max_cost_units: 20 });
+  const serialized = JSON.stringify(request);
+  if (/(secret|credential|environment|raw_prompt|hidden_reasoning|source_snapshot)/i.test(serialized)) {
+    throw new Error("generator request contains disallowed context");
+  }
+  return structuredClone(request);
+}
+
+export function validateGeneratorResponse(value) {
+  const response = requireObject(value, "generator response");
+  rejectUnknownKeys(response, ["contract_version", "proposals", "cost_units", "metadata"], "generator response");
+  if (response.contract_version !== GENERATOR_CONTRACT_VERSION) throw new Error(`unsupported generator contract_version: ${response.contract_version}`);
+  if (!Array.isArray(response.proposals) || response.proposals.length > 5) throw new Error("generator proposals must be an array with at most 5 items");
+  if (!Number.isInteger(response.cost_units) || response.cost_units < 0 || response.cost_units > 5) throw new Error("generator cost_units must be an integer from 0 to 5");
+  if (response.metadata !== undefined) {
+    const metadata = requireObject(response.metadata, "generator metadata");
+    rejectUnknownKeys(metadata, ["name", "runtime", "model"], "generator metadata");
+    for (const [key, item] of Object.entries(metadata)) optionalBoundedString(item, `generator metadata.${key}`, 500);
+  }
+  const proposals = response.proposals.map((proposal, index) => {
+    const item = requireObject(proposal, `proposal ${index}`);
+    rejectUnknownKeys(item, ["proposal_id", "gate_action", "intent", "expected_evidence", "tests", "assumptions", "affected_boundaries", "risk_strategy", "reversibility"], `proposal ${index}`);
+    const normalized = {
+      proposal_id: requireString(item.proposal_id, `proposal ${index}.proposal_id`),
+      gate_action: requireString(item.gate_action, `proposal ${index}.gate_action`),
+      intent: requireString(item.intent, `proposal ${index}.intent`),
+      expected_evidence: stringArray(item.expected_evidence ?? [], `proposal ${index}.expected_evidence`),
+      tests: stringArray(item.tests ?? [], `proposal ${index}.tests`),
+      assumptions: stringArray(item.assumptions ?? [], `proposal ${index}.assumptions`),
+      affected_boundaries: stringArray(item.affected_boundaries ?? [], `proposal ${index}.affected_boundaries`),
+      risk_strategy: requireString(item.risk_strategy, `proposal ${index}.risk_strategy`),
+      reversibility: requireString(item.reversibility, `proposal ${index}.reversibility`),
+    };
+    if (JSON.stringify(normalized).length > 5000) throw new Error(`proposal ${index} exceeds size limit`);
+    if (/(^|\s)(rm|sudo|curl|wget|bash|sh|exec)\s/i.test(normalized.intent)) throw new Error(`proposal ${index} contains executable instructions`);
+    return normalized;
+  });
+  return {
+    contract_version: response.contract_version,
+    proposals,
+    cost_units: response.cost_units,
+    metadata: response.metadata ? structuredClone(response.metadata) : {},
+  };
+}
+
+export function generatorOutputSchema() {
+  return {
+    type: "object",
+    additionalProperties: false,
+    required: ["contract_version", "proposals", "cost_units"],
+    properties: {
+      contract_version: { type: "string", const: GENERATOR_CONTRACT_VERSION },
+      cost_units: { type: "integer", minimum: 0, maximum: 5 },
+      proposals: {
+        type: "array", maxItems: 5, items: {
+          type: "object", additionalProperties: false,
+          required: ["proposal_id", "gate_action", "intent", "expected_evidence", "tests", "assumptions", "affected_boundaries", "risk_strategy", "reversibility"],
+          properties: {
+            proposal_id: { type: "string", minLength: 1 }, gate_action: { type: "string", minLength: 1 }, intent: { type: "string", minLength: 1 },
+            expected_evidence: { type: "array", items: { type: "string" } }, tests: { type: "array", items: { type: "string" } },
+            assumptions: { type: "array", items: { type: "string" } }, affected_boundaries: { type: "array", items: { type: "string" } },
+            risk_strategy: { type: "string", minLength: 1 }, reversibility: { type: "string", minLength: 1 },
+          },
+        },
+      },
+    },
+  };
+}
+
+export function validateEvaluation(value, candidateId) {
+  const evaluation = requireObject(value, "evaluation");
+  if (evaluation.contract_version !== CONTRACT_VERSION) throw new Error(`unsupported evaluator contract_version: ${evaluation.contract_version}`);
+  if (evaluation.candidate_id !== candidateId) throw new Error(`evaluation candidate_id must be ${candidateId}`);
+  for (const dimension of SCORE_DIMENSIONS) {
+    if (!Number.isInteger(evaluation[dimension]) || evaluation[dimension] < 0 || evaluation[dimension] > 5) {
+      throw new Error(`${dimension} must be an integer from 0 to 5`);
+    }
+  }
+  if (!Number.isInteger(evaluation.uncertainty) || evaluation.uncertainty < 0 || evaluation.uncertainty > 5) {
+    throw new Error("uncertainty must be an integer from 0 to 5");
+  }
+  requireString(evaluation.rationale, "evaluation.rationale");
+  stringArray(evaluation.risks ?? [], "evaluation.risks");
+  stringArray(evaluation.assumptions ?? [], "evaluation.assumptions");
+  stringArray(evaluation.child_actions ?? [], "evaluation.child_actions");
+  if (/(^|\s)(rm|sudo|curl|wget|bash|sh|exec)\s/i.test(evaluation.rationale)) {
+    throw new Error("evaluation rationale contains executable instructions");
+  }
+  return structuredClone(evaluation);
+}
+
+export function validateEnforcement(value) {
+  const enforcement = requireObject(value, "enforcement");
+  if (!["full", "tool_enforced", "instruction_only"].includes(enforcement.level)) {
+    throw new Error("enforcement.level must be full, tool_enforced or instruction_only");
+  }
+  const evidence = stringArray(enforcement.evidence ?? [], "enforcement.evidence");
+  if (enforcement.level !== "instruction_only" && evidence.length === 0) {
+    throw new Error(`${enforcement.level} requires enforcement evidence`);
+  }
+  return { level: enforcement.level, evidence };
+}
+
+export function evaluatorOutputSchema() {
+  const scoreProperties = Object.fromEntries(SCORE_DIMENSIONS.map((key) => [key, { type: "integer", minimum: 0, maximum: 5 }]));
+  return {
+    type: "object",
+    additionalProperties: false,
+    required: ["contract_version", "candidate_id", ...SCORE_DIMENSIONS, "uncertainty", "rationale", "risks", "assumptions", "child_actions"],
+    properties: {
+      contract_version: { type: "string", const: CONTRACT_VERSION },
+      candidate_id: { type: "string" },
+      ...scoreProperties,
+      uncertainty: { type: "integer", minimum: 0, maximum: 5 },
+      rationale: { type: "string", minLength: 1, maxLength: 1000 },
+      risks: { type: "array", items: { type: "string" } },
+      assumptions: { type: "array", items: { type: "string" } },
+      child_actions: { type: "array", items: { type: "string" }, maxItems: 5 },
+    },
+  };
+}

package/generated/plugins/copilot/agdf/runtime/create-agdf/lib/delivery-path-search/evaluators/claude.js

@@ -0,0 +1,35 @@
+import { execFileSync } from "node:child_process";
+import { evaluatorOutputSchema, validateEvaluation } from "../contracts.js";
+import { guardedExecFileSync } from "../transports/read-only-guard.js";
+import { buildEvaluatorPrompt } from "./prompt.js";
+
+export function claudeEvaluator(options = {}) {
+  const claudeBin = options.claudeBin ?? "claude";
+  let runtime = "claude version unavailable";
+  try {
+    runtime = execFileSync(claudeBin, ["--version"], { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] }).trim();
+  } catch {}
+
+  return {
+    name: "claude",
+    metadata: { name: "claude", runtime, model: options.model ?? "configured-default" },
+    async evaluate(input, candidate) {
+      const prompt = buildEvaluatorPrompt(input, candidate);
+      const args = [
+        "-p", "--disallowedTools", "Edit,Write,Bash",
+        "--output-format", "json", "--json-schema", JSON.stringify(evaluatorOutputSchema()),
+      ];
+      if (options.model) args.push("--model", options.model);
+      args.push(prompt);
+      const raw = guardedExecFileSync(claudeBin, args, {
+        cwd: options.cwd,
+        encoding: "utf8",
+        stdio: ["ignore", "pipe", "pipe"],
+        timeout: options.timeoutMs ?? 120000,
+      });
+      const response = JSON.parse(raw);
+      if (response.is_error) throw new Error(`claude evaluator returned an error: ${response.result ?? "unknown error"}`);
+      return validateEvaluation(JSON.parse(response.result), candidate.id);
+    },
+  };
+}

package/generated/plugins/copilot/agdf/runtime/create-agdf/lib/delivery-path-search/evaluators/codex.js

@@ -0,0 +1,44 @@
+import { execFileSync } from "node:child_process";
+import { mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { evaluatorOutputSchema, validateEvaluation } from "../contracts.js";
+import { guardedExecFileSync } from "../transports/read-only-guard.js";
+import { buildEvaluatorPrompt } from "./prompt.js";
+
+export function codexEvaluator(options = {}) {
+  const codexBin = options.codexBin ?? "codex";
+  let runtime = "codex version unavailable";
+  try {
+    runtime = execFileSync(codexBin, ["--version"], { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] }).trim();
+  } catch {}
+
+  return {
+    name: "codex",
+    metadata: { name: "codex", runtime, model: options.model ?? "configured-default" },
+    async evaluate(input, candidate) {
+      const temp = mkdtempSync(join(tmpdir(), "agdf-dps-"));
+      const schemaPath = join(temp, "schema.json");
+      const outputPath = join(temp, "output.json");
+      try {
+        writeFileSync(schemaPath, `${JSON.stringify(evaluatorOutputSchema(), null, 2)}\n`, "utf8");
+        const prompt = buildEvaluatorPrompt(input, candidate);
+        const args = [
+          "exec", "--sandbox", "read-only", "--ephemeral", "--ignore-user-config",
+          "--output-schema", schemaPath, "--output-last-message", outputPath, "--color", "never",
+        ];
+        if (options.model) args.push("--model", options.model);
+        args.push(prompt);
+        guardedExecFileSync(codexBin, args, {
+          cwd: options.cwd,
+          encoding: "utf8",
+          stdio: ["ignore", "pipe", "pipe"],
+          timeout: options.timeoutMs ?? 120000,
+        });
+        return validateEvaluation(JSON.parse(readFileSync(outputPath, "utf8")), candidate.id);
+      } finally {
+        rmSync(temp, { recursive: true, force: true });
+      }
+    },
+  };
+}