agentv 2.18.2 → 2.18.4

package/dist/{chunk-PJBBVLLB.js → chunk-RMUVJ44Z.js}

@@ -11,7 +11,7 @@ import {
   validateEvalFile,
   validateFileReferences,
   validateTargetsFile
-} from "./chunk-ZCUOH72A.js";
+} from "./chunk-KSUL3F3R.js";
 import {
   assembleLlmJudgePrompt,
   buildPromptInputs,
@@ -27,7 +27,7 @@ import {
   toCamelCaseDeep,
   toSnakeCaseDeep as toSnakeCaseDeep2,
   trimBaselineResult
-} from "./chunk-ID5SDIYE.js";
+} from "./chunk-YTHTGLMT.js";
 import {
   __commonJS,
   __esm,
@@ -3771,6 +3771,129 @@ async function processEvaluator(config, evalCase, candidate, promptInputs) {
 }
 
 // src/commands/eval/commands/prompt/overview.ts
+function getEvalMode() {
+  const mode = process.env.AGENTV_PROMPT_EVAL_MODE ?? "agent";
+  if (mode !== "agent" && mode !== "cli") {
+    throw new Error(`Invalid AGENTV_PROMPT_EVAL_MODE="${mode}". Valid values: agent, cli`);
+  }
+  return mode;
+}
+async function generateOverviewPrompt(evalPaths) {
+  const cwd = process.cwd();
+  const resolvedPaths = await resolveEvalPaths(evalPaths, cwd);
+  const repoRoot = await findRepoRoot(cwd);
+  const mode = getEvalMode();
+  const fileEntries = [];
+  for (const evalPath of resolvedPaths) {
+    const tests = await loadTests(evalPath, repoRoot);
+    fileEntries.push({ path: evalPath, tests });
+  }
+  const totalCases = fileEntries.reduce((sum, e) => sum + e.tests.length, 0);
+  if (mode === "cli") {
+    return generateCliModePrompt(fileEntries, totalCases);
+  }
+  return generateAgentModePrompt(fileEntries, totalCases);
+}
+function generateAgentModePrompt(fileEntries, totalCases) {
+  const timestamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-").slice(0, -1);
+  const lines = [
+    "# AgentV Eval Orchestration",
+    "",
+    "**Mode: agent** \u2014 You orchestrate the evaluation using agents. No API keys needed.",
+    "",
+    `You are orchestrating ${totalCases} evaluation case${totalCases === 1 ? "" : "s"}.`,
+    "",
+    "## Setup",
+    "",
+    `- **Results file:** \`.agentv/results/eval_${timestamp}.jsonl\``,
+    "- **Temp answers:** `.agentv/tmp/`",
+    "",
+    "Ensure both directories exist before starting.",
+    "",
+    "## For each test case",
+    "",
+    "Run these two agents **sequentially**:",
+    "",
+    "### 1. Dispatch `eval-candidate` agent",
+    "",
+    "Parameters:",
+    "- `eval-path`: Path to the eval YAML file",
+    "- `test-id`: The test case ID",
+    "- `answer-file`: `.agentv/tmp/eval_<test-id>.txt`",
+    "",
+    "The agent retrieves the task input, acts as the candidate LLM, and saves its response.",
+    "",
+    "### 2. Dispatch `eval-judge` agent (after candidate completes)",
+    "",
+    "Parameters:",
+    "- `eval-path`: Path to the eval YAML file",
+    "- `test-id`: The test case ID",
+    "- `answer-file`: `.agentv/tmp/eval_<test-id>.txt`",
+    `- \`results-file\`: \`.agentv/results/eval_${timestamp}.jsonl\``,
+    "",
+    "The agent runs evaluators, scores the response, and appends results to the JSONL file.",
+    ""
+  ];
+  for (const { path: evalPath, tests } of fileEntries) {
+    lines.push(`## ${evalPath}`);
+    lines.push("");
+    for (const evalCase of tests) {
+      const evaluatorSummary = describeEvaluators(evalCase);
+      lines.push(`### ${evalCase.id}`);
+      lines.push(`Criteria: ${evalCase.criteria}`);
+      if (evaluatorSummary) {
+        lines.push(`Evaluators: ${evaluatorSummary}`);
+      }
+      lines.push("");
+      lines.push("**1. Dispatch `eval-candidate` agent:**");
+      lines.push(`- eval-path: \`${evalPath}\``);
+      lines.push(`- test-id: \`${evalCase.id}\``);
+      lines.push(`- answer-file: \`.agentv/tmp/eval_${evalCase.id}.txt\``);
+      lines.push("");
+      lines.push("**2. Dispatch `eval-judge` agent** (after candidate completes):");
+      lines.push(`- eval-path: \`${evalPath}\``);
+      lines.push(`- test-id: \`${evalCase.id}\``);
+      lines.push(`- answer-file: \`.agentv/tmp/eval_${evalCase.id}.txt\``);
+      lines.push(`- results-file: \`.agentv/results/eval_${timestamp}.jsonl\``);
+      lines.push("");
+    }
+  }
+  return lines.join("\n");
+}
+function generateCliModePrompt(fileEntries, totalCases) {
+  const evalPathArgs = fileEntries.map((e) => e.path).join(" ");
+  const lines = [
+    "# AgentV Eval Orchestration",
+    "",
+    "**Mode: cli** \u2014 Run the evaluation end-to-end using the CLI.",
+    "",
+    `You are orchestrating ${totalCases} evaluation case${totalCases === 1 ? "" : "s"}.`,
+    "",
+    "## Run the evaluation",
+    "",
+    "```bash",
+    `agentv eval ${evalPathArgs}`,
+    "```",
+    "",
+    "Results are written to `.agentv/results/`. The output path is printed in the CLI output.",
+    "Parse the JSONL file for per-test scores, hits, and misses.",
+    ""
+  ];
+  for (const { path: evalPath, tests } of fileEntries) {
+    lines.push(`## ${evalPath}`);
+    lines.push("");
+    for (const evalCase of tests) {
+      const evaluatorSummary = describeEvaluators(evalCase);
+      lines.push(`### ${evalCase.id}`);
+      lines.push(`Criteria: ${evalCase.criteria}`);
+      if (evaluatorSummary) {
+        lines.push(`Evaluators: ${evaluatorSummary}`);
+      }
+      lines.push("");
+    }
+  }
+  return lines.join("\n");
+}
 var evalPromptOverviewCommand = command({
   name: "overview",
   description: "Output orchestration prompt for host agent to run evals",
@@ -3782,65 +3905,8 @@ var evalPromptOverviewCommand = command({
     })
   },
   handler: async (args) => {
-    const cwd = process.cwd();
-    const resolvedPaths = await resolveEvalPaths(args.evalPaths, cwd);
-    const repoRoot = await findRepoRoot(cwd);
-    const fileEntries = [];
-    for (const evalPath of resolvedPaths) {
-      const tests = await loadTests(evalPath, repoRoot);
-      fileEntries.push({ path: evalPath, tests });
-    }
-    const totalCases = fileEntries.reduce((sum, e) => sum + e.tests.length, 0);
-    const lines = [
-      "# AgentV Eval Orchestration",
-      "",
-      `You are orchestrating ${totalCases} evaluation case${totalCases === 1 ? "" : "s"}. For each case: get the task input, execute it, then judge the result.`,
-      "",
-      "## Step 1: Get Task Input",
-      "",
-      "Run `agentv prompt eval input <path> --test-id <id>` to get the task as JSON.",
-      "",
-      "The output contains:",
-      '- `input` \u2014 `[{role, content}]` array. Content segments are either `{type: "text", value: "..."}` or `{type: "file", path: "/absolute/path"}`. Read file segments from the filesystem.',
-      "- `guideline_paths` \u2014 files containing additional instructions to prepend to the system message (may be empty). Read these from the filesystem.",
-      "- `criteria` \u2014 what a good answer should accomplish (for your reference, do not leak to the agent being tested)",
-      "",
-      "## Step 2: Execute the Task",
-      "",
-      "Send the prompt to the agent/LLM being evaluated. Save the complete response text to a file.",
-      "",
-      "## Step 3: Judge the Result",
-      "",
-      "Run `agentv prompt eval judge <path> --test-id <id> --answer-file <response-file>`.",
-      "",
-      "The output contains an `evaluators` array. Each evaluator has a `status`:",
-      "",
-      '- **`"completed"`** \u2014 Score is final (code-judge ran deterministically). Read `result.score` (0.0\u20131.0).',
-      '- **`"prompt_ready"`** \u2014 LLM grading required. Send `prompt.system_prompt` as system and',
-      "  `prompt.user_prompt` as user to your LLM. Parse the JSON response to get `score`, `hits`, `misses`.",
-      ""
-    ];
-    for (const { path: evalPath, tests } of fileEntries) {
-      lines.push(`## ${evalPath}`);
-      lines.push("");
-      for (const evalCase of tests) {
-        const evaluatorSummary = describeEvaluators(evalCase);
-        lines.push(`### ${evalCase.id}`);
-        lines.push(`Criteria: ${evalCase.criteria}`);
-        if (evaluatorSummary) {
-          lines.push(`Evaluators: ${evaluatorSummary}`);
-        }
-        lines.push("");
-        lines.push("```bash");
-        lines.push(`agentv prompt eval input ${evalPath} --test-id ${evalCase.id}`);
-        lines.push(
-          `agentv prompt eval judge ${evalPath} --test-id ${evalCase.id} --answer-file <response-file>`
-        );
-        lines.push("```");
-        lines.push("");
-      }
-    }
-    process.stdout.write(lines.join("\n"));
+    const output = await generateOverviewPrompt(args.evalPaths);
+    process.stdout.write(output);
   }
 });
 function describeEvaluators(evalCase) {
@@ -3938,7 +4004,7 @@ var evalRunCommand = command({
     agentTimeout: option({
       type: optional(number),
       long: "agent-timeout",
-      description: "Timeout in seconds for provider responses (default: 120)"
+      description: "Optional top-level evaluation timeout in seconds. Unset by default."
     }),
     maxRetries: option({
       type: optional(number),
@@ -4006,7 +4072,7 @@ var evalRunCommand = command({
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-RJTBLMVF.js");
+      const { launchInteractiveWizard } = await import("./interactive-J4IBXJF7.js");
       await launchInteractiveWizard();
       return;
     }
@@ -5885,4 +5951,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-PJBBVLLB.js.map
+//# sourceMappingURL=chunk-RMUVJ44Z.js.map