poe-code 3.0.129 → 3.0.130

package/dist/index.js

@@ -8154,7 +8154,7 @@ function interpolate(template, values) {
   }
   return output;
 }
-async function resolveFileIncludes(template, cwd, readFile14) {
+async function resolveFileIncludes(template, cwd, readFile15) {
   const matches = [...template.matchAll(FILE_INCLUDE_PATTERN)];
   if (matches.length === 0) {
     return template;
@@ -8162,7 +8162,7 @@ async function resolveFileIncludes(template, cwd, readFile14) {
   let result = template;
   for (const match of matches) {
     const absolutePath = path14.resolve(cwd, match[1]);
-    const content = await readFile14(absolutePath, "utf8");
+    const content = await readFile15(absolutePath, "utf8");
     result = result.replace(match[0], content);
   }
   return result;
@@ -8288,10 +8288,10 @@ function createDefaultFs3() {
 function isAbortError(error2) {
   return error2 instanceof Error && error2.name === "AbortError";
 }
-async function resolveVars(vars, cwd, readFile14) {
+async function resolveVars(vars, cwd, readFile15) {
   const resolved = {};
   for (const [key, value] of Object.entries(vars)) {
-    resolved[key] = await resolveFileIncludes(value, cwd, readFile14);
+    resolved[key] = await resolveFileIncludes(value, cwd, readFile15);
   }
   return resolved;
 }
@@ -10640,12 +10640,6 @@ function parseExperimentFrontmatter(content) {
     body: parsed.content
   };
 }
-async function writeExperimentFrontmatter(docPath, frontmatter, body, fs3) {
-  await fs3.mkdir(dirname(docPath), { recursive: true });
-  const serialized = matter.stringify(body, serializeFrontmatter(frontmatter));
-  const content = body.endsWith("\n") || !serialized.endsWith("\n") ? serialized : serialized.slice(0, -1);
-  await fs3.writeFile(docPath, content);
-}
 function parseFrontmatterData2(value) {
   const parsed = isRecord5(value) ? value : void 0;
   const agent2 = parseAgent2(parsed?.agent);
@@ -10657,22 +10651,7 @@ function parseFrontmatterData2(value) {
     ...metric !== void 0 ? { metric } : {},
     baseline: parseBaseline(parsed?.baseline),
     ...maxExperiments !== void 0 ? { maxExperiments } : {},
-    ...metricTimeout !== void 0 ? { metricTimeout } : {},
-    status: parseStatus(parsed?.status)
-  };
-}
-function serializeFrontmatter(frontmatter) {
-  return {
-    ...frontmatter.agent !== void 0 ? { agent: frontmatter.agent } : {},
-    ...frontmatter.metric !== void 0 ? { metric: frontmatter.metric } : {},
-    baseline: frontmatter.baseline,
-    ...frontmatter.maxExperiments !== void 0 ? { maxExperiments: frontmatter.maxExperiments } : {},
-    ...frontmatter.metricTimeout !== void 0 ? { metricTimeout: frontmatter.metricTimeout } : {},
-    status: {
-      state: frontmatter.status.state,
-      experiment: frontmatter.status.experiment,
-      kept: frontmatter.status.kept
-    }
+    ...metricTimeout !== void 0 ? { metricTimeout } : {}
   };
 }
 function parseMetric(value) {
@@ -10716,14 +10695,6 @@ function parseBaseline(value) {
   }).filter((entry) => entry !== void 0);
   return baselineEntries.length === Object.keys(value).length ? Object.fromEntries(baselineEntries) : null;
 }
-function parseStatus(value) {
-  const parsed = isRecord5(value) ? value : void 0;
-  return {
-    state: parseString(parsed?.state) ?? DEFAULT_STATUS2.state,
-    experiment: parseNonNegativeInteger2(parsed?.experiment) ?? DEFAULT_STATUS2.experiment,
-    kept: parseNonNegativeInteger2(parsed?.kept) ?? DEFAULT_STATUS2.kept
-  };
-}
 function parseAgent2(value) {
   if (typeof value === "string") {
     return parseString(value);
@@ -10754,20 +10725,23 @@ function parseNonNegativeInteger2(value) {
 function isRecord5(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
-var DEFAULT_STATUS2;
 var init_frontmatter2 = __esm({
   "packages/experiment-loop/src/frontmatter/frontmatter.ts"() {
     "use strict";
-    DEFAULT_STATUS2 = {
-      state: "open",
-      experiment: 0,
-      kept: 0
-    };
   }
 });
 
 // packages/experiment-loop/src/journal/journal.ts
 import { dirname as dirname2 } from "node:path";
+function baselineFromEntry(metrics, entry) {
+  if (entry.scores) {
+    return entry.scores;
+  }
+  if (entry.score !== null && metrics.length === 1) {
+    return { [metrics[0].name]: entry.score };
+  }
+  return null;
+}
 function parseLine(line) {
   try {
     return [JSON.parse(line)];
@@ -10833,6 +10807,20 @@ var init_journal = __esm({
         }
         return content.split("\n").filter((line) => line.length > 0).flatMap((line) => parseLine(line));
       }
+      async updateLast(updates) {
+        const entries = await this.readAll();
+        if (entries.length === 0) {
+          return null;
+        }
+        const last = entries[entries.length - 1];
+        const updated = { ...last, ...updates };
+        entries[entries.length - 1] = updated;
+        await this.fs.writeFile(
+          this.journalPath,
+          entries.map((e) => JSON.stringify(e)).join("\n") + "\n"
+        );
+        return updated;
+      }
       async format() {
         const entries = await this.readAll();
         return [
@@ -10919,20 +10907,6 @@ async function readCurrentHash(exec3, cwd) {
 }
 function createDefaultGit(exec3) {
   return {
-    async commitAll(message2, cwd) {
-      await runOrThrow(exec3, `git add -A -- . ':!${EXPERIMENT_DOCS_PATH}'`, cwd);
-      const diffResult = await exec3("git diff --cached --quiet", { cwd });
-      if (diffResult.exitCode === 0) {
-        return readCurrentHash(exec3, cwd);
-      }
-      if (diffResult.exitCode !== 1) {
-        throw new Error(
-          diffResult.stderr || diffResult.stdout || "Failed to detect staged git changes"
-        );
-      }
-      await runOrThrow(exec3, `git commit -m ${shellEscape(message2)}`, cwd);
-      return readCurrentHash(exec3, cwd);
-    },
     async reset(commitHash, cwd) {
       const stashResult = await exec3(
         `git stash push -q --include-untracked -- ${EXPERIMENT_DOCS_PATH}`,
@@ -10959,6 +10933,8 @@ var init_git = __esm({
 
 // packages/experiment-loop/src/config/loader.ts
 import path23 from "node:path";
+import { readFile as readFile5 } from "node:fs/promises";
+import { fileURLToPath as fileURLToPath2 } from "node:url";
 import { parse as parse7 } from "yaml";
 function isRecord6(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
@@ -10996,8 +10972,20 @@ function parseRunConfigDocument(filePath, content) {
   }
   return { prompt };
 }
-function defaultRunConfig() {
-  return { prompt: DEFAULT_PROMPT };
+async function readBundledFile(name) {
+  const filePath = fileURLToPath2(new URL(`./${name}`, import.meta.url));
+  return readFile5(filePath, "utf8");
+}
+async function readDefaultRunConfig() {
+  const content = await readBundledFile("default-run.yaml");
+  const config = parseRunConfigDocument("default-run.yaml", content);
+  if (!config) {
+    throw new Error("default-run.yaml is empty or invalid.");
+  }
+  return config;
+}
+async function loadInstructions() {
+  return readBundledFile("default-instructions.md");
 }
 async function loadRunConfig(options) {
   const projectPath = path23.join(options.cwd, ".poe-code", "experiments", "run.yaml");
@@ -11016,26 +11004,11 @@ async function loadRunConfig(options) {
       return config;
     }
   }
-  return defaultRunConfig();
+  return readDefaultRunConfig();
 }
-var DEFAULT_PROMPT;
 var init_loader2 = __esm({
   "packages/experiment-loop/src/config/loader.ts"() {
     "use strict";
-    DEFAULT_PROMPT = `{{body}}
-
-## Metrics
-
-{{metrics}}
-
-## Journal
-
-{{journal}}
-
-{{crash_output}}
-
-You are autonomous, do not stop or ask for input.
-Do not write to the journal file or commit changes. Both are managed automatically after your run completes.`;
   }
 });
 
@@ -11135,9 +11108,6 @@ function createAbortError5() {
 function isAbortError3(error2) {
   return error2 instanceof Error && error2.name === "AbortError";
 }
-function combineOutput(stdout, stderr) {
-  return `${stdout}${stderr}`;
-}
 function interpolate2(template, values) {
   let output = template;
   for (const [key, value] of Object.entries(values)) {
@@ -11160,41 +11130,25 @@ function formatMetrics(metrics, baseline) {
   }).join("\n");
 }
 function buildPrompt(options) {
-  return interpolate2(options.runConfig.prompt, {
+  const vars = {
     body: options.body.trim(),
     journal: options.journal,
     metrics: formatMetrics(options.metrics, options.baseline),
-    crash_output: options.lastCrashOutput ? `## Last crash output
-
-${options.lastCrashOutput.trim()}` : "",
     experiment_index: String(options.experimentIndex),
-    baseline: options.baseline ? JSON.stringify(options.baseline) : ""
-  }).trim();
+    baseline: options.baseline ? JSON.stringify(options.baseline) : "",
+    doc_path: options.docPath,
+    doc_name: options.docName
+  };
+  const context = interpolate2(options.runConfig.prompt, vars).trim();
+  const instructions = interpolate2(options.instructions, vars).trim();
+  return `${context}
+
+${instructions}`;
 }
 function allMetricsPassed(metrics, results) {
   return results.length === metrics.length && results.every((result) => result.passed);
 }
-function allScoresImproved(metrics, results, baseline) {
-  return metrics.every((metric, index) => {
-    const score = results[index]?.score;
-    if (score === null || score === void 0) {
-      return false;
-    }
-    const baselineScore = baseline?.[metric.name];
-    if (baselineScore === void 0) {
-      return true;
-    }
-    const delta = metric.delta;
-    if (metric.direction === "stable") {
-      return delta !== void 0 ? Math.abs(score - baselineScore) <= delta : score === baselineScore;
-    }
-    if (metric.direction === "maximize") {
-      return delta !== void 0 ? score >= baselineScore - delta : score > baselineScore;
-    }
-    return delta !== void 0 ? score <= baselineScore + delta : score < baselineScore;
-  });
-}
-function updateBaseline(metrics, results) {
+function baselineFromResults(metrics, results) {
   return Object.fromEntries(
     metrics.map((metric, index) => {
       const score = results[index]?.score;
@@ -11205,53 +11159,16 @@ function updateBaseline(metrics, results) {
     })
   );
 }
-function formatEvaluationOutput(metrics, results) {
-  return metrics.map((metric, index) => {
-    const result = results[index];
-    if (!result) {
-      return `${metric.name}: not_run`;
-    }
-    return [
-      `${metric.name}: score=${result.score === null ? "null" : result.score}, passed=${String(result.passed)}`,
-      result.output.trim()
-    ].filter((entry) => entry.length > 0).join("\n");
-  }).join("\n\n");
-}
-function selectJournalScore(metrics, results) {
-  if (metrics.length !== 1) {
-    return null;
-  }
-  return results[0]?.score ?? null;
-}
-function createEntry(options) {
+function deriveStateFromJournal(entries, metrics) {
+  const keepEntries = entries.filter((e) => e.status === "keep");
+  const lastKeep = keepEntries[keepEntries.length - 1];
   return {
-    commit: options.commit,
-    status: options.status,
-    score: options.score,
-    output: options.output,
-    agentOutput: options.agentOutput,
-    durationMs: options.durationMs,
-    timestamp: (/* @__PURE__ */ new Date()).toISOString()
+    experimentsCompleted: entries.length,
+    experimentsKept: keepEntries.length,
+    baseline: lastKeep ? baselineFromEntry(metrics, lastKeep) : null,
+    baselineHash: lastKeep?.commit
   };
 }
-async function persistDoc(options) {
-  const rawContent = await options.fs.readFile(options.docPath, "utf8");
-  const { frontmatter, body } = parseExperimentFrontmatter(rawContent);
-  await writeExperimentFrontmatter(
-    options.docPath,
-    {
-      ...frontmatter,
-      baseline: options.baseline,
-      status: {
-        ...frontmatter.status,
-        experiment: options.experimentsCompleted,
-        kept: options.experimentsKept
-      }
-    },
-    body,
-    options.fs
-  );
-}
 async function runExperimentLoop(options) {
   const fs3 = options.fs ?? createDefaultFs6();
   const exec3 = options.exec ?? createDefaultExec();
@@ -11263,35 +11180,35 @@ async function runExperimentLoop(options) {
   const absoluteDocPath = resolveAbsoluteDocPath2(options.docPath, options.cwd, options.homeDir);
   const journal = new ExperimentJournal(resolveJournalPath(absoluteDocPath), fs3);
   await journal.init();
-  const runConfig = await loadRunConfig({ cwd: options.cwd, homeDir: options.homeDir, fs: fs3 });
+  const [runConfig, instructions] = await Promise.all([
+    loadRunConfig({ cwd: options.cwd, homeDir: options.homeDir, fs: fs3 }),
+    loadInstructions()
+  ]);
   const startTime = Date.now();
   async function readDoc() {
     const rawContent = await fs3.readFile(absoluteDocPath, "utf8");
     return parseExperimentFrontmatter(rawContent);
   }
-  let { frontmatter, body } = await readDoc();
-  let experimentsCompleted = 0;
-  let experimentsKept = 0;
-  let lastCrashOutput;
-  let baselineHash;
-  if (frontmatter.baseline === null) {
-    const metrics = normalizeMetrics(frontmatter.metric);
-    const metricTimeoutMs = frontmatter.metricTimeout ? frontmatter.metricTimeout * 1e3 : void 0;
-    const baselineResults = await evaluateChain(metrics, options.cwd, exec3, options.onMetricResult, metricTimeoutMs);
-    if (allMetricsPassed(metrics, baselineResults)) {
-      const baseline = updateBaseline(metrics, baselineResults);
+  const { frontmatter: initialFrontmatter } = await readDoc();
+  const initialMetrics = normalizeMetrics(initialFrontmatter.metric);
+  const journalEntries = await journal.readAll();
+  const journalState = deriveStateFromJournal(journalEntries, initialMetrics);
+  let experimentsCompleted = journalState.experimentsCompleted;
+  let experimentsKept = journalState.experimentsKept;
+  let baselineHash = journalState.baselineHash;
+  let baseline = journalState.baseline ?? initialFrontmatter.baseline;
+  if (baseline === null) {
+    const metricTimeoutMs = initialFrontmatter.metricTimeout ? initialFrontmatter.metricTimeout * 1e3 : void 0;
+    const baselineResults = await evaluateChain(
+      initialMetrics,
+      options.cwd,
+      exec3,
+      options.onMetricResult,
+      metricTimeoutMs
+    );
+    if (allMetricsPassed(initialMetrics, baselineResults)) {
+      baseline = baselineFromResults(initialMetrics, baselineResults);
       options.onBaselineCollected?.(baseline);
-      frontmatter = {
-        ...frontmatter,
-        baseline
-      };
-      await persistDoc({
-        fs: fs3,
-        docPath: absoluteDocPath,
-        baseline: frontmatter.baseline,
-        experimentsCompleted,
-        experimentsKept
-      });
     }
   }
   async function finalize(stopReason) {
@@ -11307,8 +11224,7 @@ async function runExperimentLoop(options) {
     while (true) {
       assertNotAborted3(options.signal);
       const doc = await readDoc();
-      body = doc.body;
-      frontmatter = doc.frontmatter;
+      const { body, frontmatter } = doc;
       const maxExperiments = validateMaxExperiments(
         options.maxExperiments ?? frontmatter.maxExperiments
       );
@@ -11317,20 +11233,21 @@ async function runExperimentLoop(options) {
       }
       const metrics = normalizeMetrics(frontmatter.metric);
       const agents = normalizeAgents2(options.agent ?? frontmatter.agent);
-      const metricTimeoutMs = frontmatter.metricTimeout ? frontmatter.metricTimeout * 1e3 : void 0;
       const experimentIndex = experimentsCompleted + 1;
+      baselineHash ??= await git.currentHash(options.cwd);
+      const preExperimentHash = baselineHash;
+      const journalLengthBefore = (await journal.readAll()).length;
       const prompt = buildPrompt({
         runConfig,
+        instructions,
         body,
         journal: await journal.format(),
         metrics,
         experimentIndex,
-        baseline: frontmatter.baseline,
-        ...lastCrashOutput ? { lastCrashOutput } : {}
+        baseline,
+        docPath: options.docPath,
+        docName: path24.basename(absoluteDocPath, path24.extname(absoluteDocPath))
       });
-      baselineHash ??= await git.currentHash(options.cwd);
-      const preExperimentHash = baselineHash;
-      const experimentStart = Date.now();
       const currentSpecifier = agents[(experimentIndex - 1) % agents.length];
       const model = currentSpecifier.model;
       options.onExperimentStart?.(experimentIndex, currentSpecifier.agent);
@@ -11349,108 +11266,39 @@ async function runExperimentLoop(options) {
         }
         throw error2;
       }
-      const agentOutput = combineOutput(agentResult.stdout, agentResult.stderr);
-      if (agentResult.exitCode !== 0) {
-        const entry2 = createEntry({
-          commit: preExperimentHash,
-          status: "crash",
-          score: null,
-          output: agentOutput,
-          agentOutput,
-          durationMs: Date.now() - experimentStart
-        });
-        experimentsCompleted += 1;
-        lastCrashOutput = entry2.output;
-        await journal.log(entry2);
-        await git.reset(preExperimentHash, options.cwd);
-        options.onReset?.(preExperimentHash);
-        options.onExperimentComplete?.(experimentIndex, entry2);
-        await persistDoc({
-          fs: fs3,
-          docPath: absoluteDocPath,
-          baseline: frontmatter.baseline,
-          experimentsCompleted,
-          experimentsKept
-        });
-        continue;
-      }
-      const commitMessage = `experiment-loop: ${path24.basename(absoluteDocPath, path24.extname(absoluteDocPath))} #${experimentIndex}`;
-      let commitHash;
-      try {
-        commitHash = await git.commitAll(commitMessage, options.cwd);
-      } catch (commitError) {
-        const errorMessage = commitError instanceof Error ? commitError.message : String(commitError);
-        options.onRecoveryAttempt?.(errorMessage);
-        const recoveryResult = await runAgent({
-          agent: currentSpecifier.agent,
-          prompt: `A git command failed. Fix the issue so the commit can succeed.
-
-Error:
-${errorMessage}`,
-          cwd: options.cwd,
-          ...model ? { model } : {},
-          ...options.signal ? { signal: options.signal } : {}
-        });
-        if (recoveryResult.exitCode !== 0) {
-          const entry2 = createEntry({
-            commit: preExperimentHash,
-            status: "crash",
-            score: null,
-            output: errorMessage,
-            agentOutput,
-            durationMs: Date.now() - experimentStart
-          });
-          experimentsCompleted += 1;
-          lastCrashOutput = entry2.output;
-          await journal.log(entry2);
+      const journalAfter = await journal.readAll();
+      let newEntry = journalAfter.length > journalLengthBefore ? journalAfter[journalAfter.length - 1] : null;
+      experimentsCompleted += 1;
+      if (newEntry && newEntry.score === null && !newEntry.scores && metrics.length > 0) {
+        const metricTimeoutMs = frontmatter.metricTimeout ? frontmatter.metricTimeout * 1e3 : void 0;
+        const results = await evaluateChain(
+          metrics,
+          options.cwd,
+          exec3,
+          options.onMetricResult,
+          metricTimeoutMs
+        );
+        if (allMetricsPassed(metrics, results)) {
+          const scores = baselineFromResults(metrics, results);
+          const score = metrics.length === 1 ? results[0].score : null;
+          newEntry = await journal.updateLast({ score, scores }) ?? newEntry;
+        }
+      }
+      if (newEntry) {
+        if (newEntry.status === "keep") {
+          experimentsKept += 1;
+          baselineHash = newEntry.commit;
+          baseline = baselineFromEntry(metrics, newEntry) ?? baseline;
+          options.onCommit?.(newEntry.commit);
+        } else {
           await git.reset(preExperimentHash, options.cwd);
           options.onReset?.(preExperimentHash);
-          options.onExperimentComplete?.(experimentIndex, entry2);
-          await persistDoc({
-            fs: fs3,
-            docPath: absoluteDocPath,
-            baseline: frontmatter.baseline,
-            experimentsCompleted,
-            experimentsKept
-          });
-          continue;
         }
-        commitHash = await git.commitAll(commitMessage, options.cwd);
-      }
-      options.onCommit?.(commitHash);
-      const evaluationResults = await evaluateChain(metrics, options.cwd, exec3, options.onMetricResult, metricTimeoutMs);
-      const keep = allMetricsPassed(metrics, evaluationResults) && allScoresImproved(metrics, evaluationResults, frontmatter.baseline);
-      const entry = createEntry({
-        commit: commitHash,
-        status: keep ? "keep" : "discard",
-        score: selectJournalScore(metrics, evaluationResults),
-        output: formatEvaluationOutput(metrics, evaluationResults),
-        agentOutput,
-        durationMs: Date.now() - experimentStart
-      });
-      experimentsCompleted += 1;
-      await journal.log(entry);
-      if (keep) {
-        experimentsKept += 1;
-        frontmatter = {
-          ...frontmatter,
-          baseline: updateBaseline(metrics, evaluationResults)
-        };
-        baselineHash = commitHash;
-        lastCrashOutput = void 0;
+        options.onExperimentComplete?.(experimentIndex, newEntry);
       } else {
         await git.reset(preExperimentHash, options.cwd);
         options.onReset?.(preExperimentHash);
-        lastCrashOutput = void 0;
       }
-      options.onExperimentComplete?.(experimentIndex, entry);
-      await persistDoc({
-        fs: fs3,
-        docPath: absoluteDocPath,
-        baseline: frontmatter.baseline,
-        experimentsCompleted,
-        experimentsKept
-      });
     }
   } catch (error2) {
     if (isAbortError3(error2)) {
@@ -11545,6 +11393,13 @@ async function readExperimentJournal(options) {
   const journal = new ExperimentJournal(resolveJournalPath2(absoluteDocPath), fs3);
   return await journal.readAll();
 }
+async function appendExperimentJournalEntry(options) {
+  const fs3 = options.fs ?? createDefaultFs7();
+  const absoluteDocPath = resolveAbsoluteDocPath3(options.docPath, options.cwd, options.homeDir);
+  const journal = new ExperimentJournal(resolveJournalPath2(absoluteDocPath), fs3);
+  await journal.init();
+  await journal.log(options.entry);
+}
 var init_experiment = __esm({
   async "src/sdk/experiment.ts"() {
     "use strict";
@@ -11804,7 +11659,7 @@ var init_frontmatter3 = __esm({
 });
 
 // packages/github-workflows/src/discover.ts
-import { readdir as readdir9, readFile as readFile7 } from "node:fs/promises";
+import { readdir as readdir9, readFile as readFile8 } from "node:fs/promises";
 import { join } from "node:path";
 async function discoverAutomations(builtInDir, projectDir) {
   const automationsByName = /* @__PURE__ */ new Map();
@@ -11845,7 +11700,7 @@ async function listMarkdownFiles(dir) {
   }
 }
 async function readAutomation(dir, fileName) {
-  const content = await readFile7(join(dir, fileName), "utf8");
+  const content = await readFile8(join(dir, fileName), "utf8");
   const { frontmatter, body } = parseFrontmatter2(content);
   const name = fileName.slice(0, -3);
   return {
@@ -12042,7 +11897,7 @@ var init_src13 = __esm({
 });
 
 // packages/cmdkit/src/index.ts
-import { fileURLToPath as fileURLToPath2 } from "node:url";
+import { fileURLToPath as fileURLToPath3 } from "node:url";
 function cloneScope(scope) {
   return scope === void 0 ? void 0 : [...scope];
 }
@@ -12074,7 +11929,7 @@ function cloneRequires(requires) {
 function parseStackPath(candidate) {
   if (candidate.startsWith("file://")) {
     try {
-      return fileURLToPath2(candidate);
+      return fileURLToPath3(candidate);
     } catch {
       return void 0;
     }
@@ -12470,9 +12325,9 @@ var init_require_comment_prefix = __esm({
 });
 
 // packages/github-workflows/src/commands.ts
-import { access, mkdir as mkdir8, readFile as readFile8, unlink as unlink2, writeFile as writeFile6 } from "node:fs/promises";
+import { access, mkdir as mkdir8, readFile as readFile9, unlink as unlink2, writeFile as writeFile6 } from "node:fs/promises";
 import path26 from "node:path";
-import { fileURLToPath as fileURLToPath3 } from "node:url";
+import { fileURLToPath as fileURLToPath4 } from "node:url";
 import Mustache3 from "mustache";
 function formatLabel(name) {
   const capitalize = (s) => s.charAt(0).toUpperCase() + s.slice(1);
@@ -12490,7 +12345,7 @@ async function loadNamedAutomation(name, cwd) {
 }
 async function readBuiltInPromptFile(name) {
   try {
-    return await readFile8(path26.join(await resolveBuiltInPromptsDir(), `${name}.md`), "utf8");
+    return await readFile9(path26.join(await resolveBuiltInPromptsDir(), `${name}.md`), "utf8");
   } catch (error2) {
     if (isMissingPathError2(error2)) {
       throw new UserError(`Automation "${name}" was not found.`);
@@ -12501,7 +12356,7 @@ async function readBuiltInPromptFile(name) {
 async function readBuiltInWorkflowTemplate(name, variant) {
   const templatePath = path26.join(await resolveBuiltInWorkflowTemplatesDir(), `${name}.${variant}.yml`);
   try {
-    const content = await readFile8(templatePath, "utf8");
+    const content = await readFile9(templatePath, "utf8");
     const header = variant === "ejected" ? `# Auto-generated by: poe-code github-workflows install ${name}
 # Edit .poe-code/github-workflows/poe-code-${name}.md to customize the prompt.
 ` : `# Auto-generated by: poe-code github-workflows install ${name}
@@ -12718,12 +12573,12 @@ var init_commands = __esm({
     init_require_comment_prefix();
     UPSTREAM_REPO = "poe-platform/poe-code";
     builtInPromptsDirCandidates = [
-      fileURLToPath3(new URL("./prompts", import.meta.url)),
-      fileURLToPath3(new URL("../src/prompts", import.meta.url))
+      fileURLToPath4(new URL("./prompts", import.meta.url)),
+      fileURLToPath4(new URL("../src/prompts", import.meta.url))
     ];
     builtInWorkflowTemplatesDirCandidates = [
-      fileURLToPath3(new URL("./workflow-templates", import.meta.url)),
-      fileURLToPath3(new URL("../src/workflow-templates", import.meta.url))
+      fileURLToPath4(new URL("./workflow-templates", import.meta.url)),
+      fileURLToPath4(new URL("../src/workflow-templates", import.meta.url))
     ];
     originalMustacheEscape = Mustache3.escape;
     installableAutomations = [
@@ -13001,7 +12856,7 @@ var init_src15 = __esm({
 
 // src/utils/execution-context.ts
 import { basename as basename2, dirname as dirname3 } from "node:path";
-import { fileURLToPath as fileURLToPath4 } from "node:url";
+import { fileURLToPath as fileURLToPath5 } from "node:url";
 function detectExecutionContext(input) {
   const { argv, env, moduleUrl } = input;
   if (isDevelopmentMode(argv, env)) {
@@ -13060,7 +12915,7 @@ function detectNpxVersion(env) {
   return "default";
 }
 function createDevelopmentContext(moduleUrl) {
-  const modulePath = fileURLToPath4(moduleUrl);
+  const modulePath = fileURLToPath5(moduleUrl);
   const srcIndex = modulePath.lastIndexOf("/src/");
   const projectRoot = srcIndex !== -1 ? modulePath.substring(0, srcIndex) : dirname3(dirname3(modulePath));
   return {
@@ -15920,8 +15775,8 @@ var init_poe_agent_plugin_shell = __esm({
 
 // packages/poe-agent/src/system-prompt.ts
 import { readFileSync } from "node:fs";
-import { readFile as readFile9 } from "node:fs/promises";
-import { fileURLToPath as fileURLToPath5 } from "node:url";
+import { readFile as readFile10 } from "node:fs/promises";
+import { fileURLToPath as fileURLToPath6 } from "node:url";
 function loadSystemPromptSync() {
   return readFileSync(SYSTEM_PROMPT_PATH, "utf8");
 }
@@ -15929,7 +15784,7 @@ var SYSTEM_PROMPT_PATH;
 var init_system_prompt = __esm({
   "packages/poe-agent/src/system-prompt.ts"() {
     "use strict";
-    SYSTEM_PROMPT_PATH = fileURLToPath5(new URL("./SYSTEM_PROMPT.md", import.meta.url));
+    SYSTEM_PROMPT_PATH = fileURLToPath6(new URL("./SYSTEM_PROMPT.md", import.meta.url));
   }
 });
 
@@ -19241,7 +19096,7 @@ var init_renderer2 = __esm({
 });
 
 // packages/cmdkit/src/cli.ts
-import { access as access2, readFile as readFile10, writeFile as writeFile7 } from "node:fs/promises";
+import { access as access2, readFile as readFile11, writeFile as writeFile7 } from "node:fs/promises";
 import path32 from "node:path";
 import { Command as CommanderCommand, CommanderError, InvalidArgumentError, Option } from "commander";
 function unwrapOptional(schema) {
@@ -19817,7 +19672,7 @@ async function withOutputFormat2(output, fn) {
 }
 function createFs() {
   return {
-    readFile: async (path42, encoding = "utf8") => readFile10(path42, { encoding }),
+    readFile: async (path42, encoding = "utf8") => readFile11(path42, { encoding }),
     writeFile: async (path42, contents) => {
       await writeFile7(path42, contents);
     },
@@ -19909,7 +19764,7 @@ function validatePresetFieldValue(value, field, presetPath) {
 async function loadPresetValues(fields, presetPath) {
   let rawPreset;
   try {
-    rawPreset = await readFile10(presetPath, {
+    rawPreset = await readFile11(presetPath, {
       encoding: "utf8"
     });
   } catch (error2) {
@@ -20158,7 +20013,7 @@ async function loadFixtureScenario(command, selector) {
   const fixturePath = resolveFixturePath(commandPath);
   let rawFixture;
   try {
-    rawFixture = await readFile10(fixturePath, {
+    rawFixture = await readFile11(fixturePath, {
       encoding: "utf8"
     });
   } catch {
@@ -24564,7 +24419,7 @@ var init_configs3 = __esm({
 });
 
 // packages/agent-skill-config/src/templates.ts
-import { readFile as readFile11 } from "node:fs/promises";
+import { readFile as readFile12 } from "node:fs/promises";
 async function getTemplates() {
   if (templatesCache) {
     return templatesCache;
@@ -24573,7 +24428,7 @@ async function getTemplates() {
     "./templates/poe-generate.md",
     import.meta.url
   );
-  const poeGenerateTemplate = await readFile11(poeGenerateTemplateUrl, "utf8");
+  const poeGenerateTemplate = await readFile12(poeGenerateTemplateUrl, "utf8");
   templatesCache = {
     "poe-generate.md": poeGenerateTemplate
   };
@@ -25615,8 +25470,8 @@ var init_models2 = __esm({
 
 // src/cli/commands/pipeline.ts
 import path39 from "node:path";
-import { readFile as readFile12, stat as stat9 } from "node:fs/promises";
-import { fileURLToPath as fileURLToPath6 } from "node:url";
+import { readFile as readFile13, stat as stat9 } from "node:fs/promises";
+import { fileURLToPath as fileURLToPath7 } from "node:url";
 async function resolvePipelinePlanDirectory(container) {
   const configDoc = await readMergedDocument(
     container.fs,
@@ -25675,7 +25530,7 @@ async function loadPipelineTemplates() {
   if (pipelineTemplatesCache) {
     return pipelineTemplatesCache;
   }
-  const packageRoot = await findPackageRoot(fileURLToPath6(import.meta.url));
+  const packageRoot = await findPackageRoot(fileURLToPath7(import.meta.url));
   const templateRoots = [
     path39.join(packageRoot, "src", "templates", "pipeline"),
     path39.join(packageRoot, "dist", "templates", "pipeline")
@@ -25685,8 +25540,8 @@ async function loadPipelineTemplates() {
       continue;
     }
     const [skillPlan, steps] = await Promise.all([
-      readFile12(path39.join(templateRoot, "SKILL_plan.md"), "utf8"),
-      readFile12(path39.join(templateRoot, "steps.yaml.hbs"), "utf8")
+      readFile13(path39.join(templateRoot, "SKILL_plan.md"), "utf8"),
+      readFile13(path39.join(templateRoot, "steps.yaml.hbs"), "utf8")
     ]);
     pipelineTemplatesCache = { skillPlan, steps };
     return pipelineTemplatesCache;
@@ -25897,16 +25752,16 @@ function registerPipelineCommand(program, container) {
       resources.logger.success("Plan is valid.");
       const opts = this.opts();
       if (opts.preview) {
-        const readFile14 = container.fs.readFile.bind(container.fs);
+        const readFile15 = container.fs.readFile.bind(container.fs);
         const resolvedVars = {};
         for (const [key, value] of Object.entries(plan.vars ?? {})) {
-          resolvedVars[key] = await resolveFileIncludes(value, container.env.cwd, readFile14);
+          resolvedVars[key] = await resolveFileIncludes(value, container.env.cwd, readFile15);
         }
         const resolvedSetup = plan.setup === null ? void 0 : plan.setup ?? steps.setup;
         const resolvedTeardown = plan.teardown === null ? void 0 : plan.teardown ?? steps.teardown;
         if (resolvedSetup) {
           const raw = Object.keys(resolvedVars).length > 0 ? interpolate(resolvedSetup.prompt, resolvedVars) : resolvedSetup.prompt;
-          const expanded = await resolveFileIncludes(raw, container.env.cwd, readFile14);
+          const expanded = await resolveFileIncludes(raw, container.env.cwd, readFile15);
           resources.logger.resolved("setup", expanded);
         }
         for (const task of plan.tasks) {
@@ -25914,7 +25769,7 @@ function registerPipelineCommand(program, container) {
             const expanded = await resolveFileIncludes(
               buildExecutionPrompt({ selection: { kind: "run", task }, steps: steps.steps, planPath: file, vars: resolvedVars }),
               container.env.cwd,
-              readFile14
+              readFile15
             );
             resources.logger.resolved(`task: ${task.id} \u2014 ${task.title}`, expanded);
           } else {
@@ -25922,7 +25777,7 @@ function registerPipelineCommand(program, container) {
               const expanded = await resolveFileIncludes(
                 buildExecutionPrompt({ selection: { kind: "run", task, stepName }, steps: steps.steps, planPath: file, vars: resolvedVars }),
                 container.env.cwd,
-                readFile14
+                readFile15
               );
               resources.logger.resolved(`task: ${task.id} / ${stepName}`, expanded);
             }
@@ -25930,7 +25785,7 @@ function registerPipelineCommand(program, container) {
         }
         if (resolvedTeardown) {
           const raw = Object.keys(resolvedVars).length > 0 ? interpolate(resolvedTeardown.prompt, resolvedVars) : resolvedTeardown.prompt;
-          const expanded = await resolveFileIncludes(raw, container.env.cwd, readFile14);
+          const expanded = await resolveFileIncludes(raw, container.env.cwd, readFile15);
           resources.logger.resolved("teardown", expanded);
         }
       }
@@ -26498,8 +26353,8 @@ var init_ralph3 = __esm({
 
 // src/cli/commands/experiment.ts
 import path41 from "node:path";
-import { readFile as readFile13, stat as stat10 } from "node:fs/promises";
-import { fileURLToPath as fileURLToPath7 } from "node:url";
+import { readFile as readFile14, stat as stat10 } from "node:fs/promises";
+import { fileURLToPath as fileURLToPath8 } from "node:url";
 function resolveExperimentPaths(scope, cwd, homeDir) {
   const rootPath = scope === "global" ? path41.join(homeDir, ".poe-code", "experiments") : path41.join(cwd, ".poe-code", "experiments");
   const displayRoot = scope === "global" ? "~/.poe-code/experiments" : ".poe-code/experiments";
@@ -26536,7 +26391,7 @@ async function loadExperimentTemplates() {
   if (experimentTemplatesCache) {
     return experimentTemplatesCache;
   }
-  const packageRoot = await findPackageRoot2(fileURLToPath7(import.meta.url));
+  const packageRoot = await findPackageRoot2(fileURLToPath8(import.meta.url));
   const templateRoots = [
     path41.join(packageRoot, "src", "templates", "experiment"),
     path41.join(packageRoot, "dist", "templates", "experiment")
@@ -26546,8 +26401,8 @@ async function loadExperimentTemplates() {
       continue;
     }
     const [skillPlan, runYaml] = await Promise.all([
-      readFile13(path41.join(templateRoot, "SKILL_experiment.md"), "utf8"),
-      readFile13(path41.join(templateRoot, "run.yaml.hbs"), "utf8")
+      readFile14(path41.join(templateRoot, "SKILL_experiment.md"), "utf8"),
+      readFile14(path41.join(templateRoot, "run.yaml.hbs"), "utf8")
     ]);
     experimentTemplatesCache = { skillPlan, runYaml };
     return experimentTemplatesCache;
@@ -26588,11 +26443,6 @@ function validateExperimentDoc(frontmatter) {
       }
     }
   }
-  if (frontmatter.status.kept > frontmatter.status.experiment) {
-    errors.push(
-      `Status inconsistency: kept (${frontmatter.status.kept}) exceeds experiment count (${frontmatter.status.experiment})`
-    );
-  }
   return errors;
 }
 function formatDuration3(ms) {
@@ -26821,7 +26671,7 @@ function registerExperimentCommand(program, container) {
       resources.context.finalize();
     }
   });
-  experiment.command("journal").description("Display the experiment journal as a table.").argument("[doc]", "Experiment doc path").action(async function(docArg) {
+  const journalCommand = experiment.command("journal").description("Display the experiment journal as a table.").argument("[doc]", "Experiment doc path").action(async function(docArg) {
     const flags = resolveCommandFlags(program);
     const resources = createExecutionResources(container, flags, "experiment:journal");
     resources.logger.intro("experiment journal");
@@ -26872,6 +26722,60 @@ function registerExperimentCommand(program, container) {
       resources.context.finalize();
     }
   });
+  journalCommand.command("log").description("Append an entry to the experiment journal.").argument("[doc]", "Experiment doc path").requiredOption("--status <status>", "Entry status: keep or discard").requiredOption("--commit <hash>", "Git commit hash").option("--score <number>", "Metric score (for single-metric display)").option("--scores <json>", `Multi-metric scores as JSON object, e.g. '{"tests":2}'`).option("--output <text>", "Metric output text", "").option("--duration-ms <number>", "Duration in milliseconds", "0").action(async function(docArg) {
+    const flags = resolveCommandFlags(program);
+    const resources = createExecutionResources(container, flags, "experiment:journal:log");
+    const opts = this.opts();
+    try {
+      const docPath = await resolveDocPath2({
+        container,
+        program,
+        providedDoc: docArg,
+        selectMessage: "Select the experiment doc to log to:",
+        cancelMessage: "Journal log cancelled."
+      });
+      if (!docPath) {
+        return;
+      }
+      const status = opts.status;
+      if (status !== "keep" && status !== "discard") {
+        throw new ValidationError(`Invalid status "${opts.status}". Must be keep or discard.`);
+      }
+      const score = opts.score !== void 0 ? Number.parseFloat(opts.score) : null;
+      let scores;
+      if (opts.scores) {
+        try {
+          scores = JSON.parse(opts.scores);
+        } catch {
+          throw new ValidationError(`Invalid --scores JSON: ${opts.scores}`);
+        }
+      }
+      const entry = {
+        commit: opts.commit,
+        status,
+        score: score !== null && Number.isFinite(score) ? score : null,
+        ...scores ? { scores } : {},
+        output: opts.output,
+        agentOutput: "",
+        durationMs: Number.parseInt(opts.durationMs, 10) || 0,
+        timestamp: (/* @__PURE__ */ new Date()).toISOString()
+      };
+      if (!flags.dryRun) {
+        await appendExperimentJournalEntry({
+          cwd: container.env.cwd,
+          homeDir: container.env.homeDir,
+          docPath,
+          entry
+        });
+      }
+      resources.context.complete({
+        success: `Logged ${status} entry (commit: ${opts.commit})`,
+        dry: `Would log ${status} entry (commit: ${opts.commit})`
+      });
+    } finally {
+      resources.context.finalize();
+    }
+  });
   experiment.command("validate").description("Validate an experiment doc without running it.").argument("[doc]", "Experiment doc path").action(async function(docArg) {
     const flags = resolveCommandFlags(program);
     const resources = createExecutionResources(
@@ -27475,7 +27379,7 @@ var init_package = __esm({
   "package.json"() {
     package_default = {
       name: "poe-code",
-      version: "3.0.129",
+      version: "3.0.130",
       description: "CLI tool to configure Poe API for developer workflows.",
       type: "module",
       main: "./dist/index.js",