orizu 0.3.2 → 0.3.3

package/README.md

@@ -101,6 +101,17 @@ orizu logout
     <td>Create a project</td>
     <td><code>orizu projects create --name "Support QA" --team ops-eval</code></td>
   </tr>
+  <tr>
+    <td colspan="2"><strong>Prompts</strong></td>
+  </tr>
+  <tr>
+    <td>List prompts</td>
+    <td><code>orizu prompts list --project ops-eval/support-qa</code></td>
+  </tr>
+  <tr>
+    <td>Show prompt comments</td>
+    <td><code>orizu prompts comments &lt;promptIdOrName&gt; --project ops-eval/support-qa</code></td>
+  </tr>
   <tr>
     <td colspan="2"><strong>Datasets</strong></td>
   </tr>

package/dist/index.js

@@ -30,10 +30,10 @@ function printVersion() {
     printLine(`orizu ${getCliVersion()}`);
 }
 function printUsage() {
-    printLine(`orizu global options:\n\n  --local                 Use http://localhost:3000\n  --server <url>          Use a specific server origin (for example: https://preview.example.com)\n  --version, -v           Print the orizu CLI version\n\norizu commands:\n\n  orizu login [--no-prompt-if-logged-in]\n  orizu logout\n  orizu whoami\n  orizu env [--project <team/project>] [--project-id <projectId>]\n  orizu log <event_type> --run-id <id> --sequence <n> --payload @event.json\n  orizu teams list\n  orizu teams create [--name <name>]\n  orizu teams members list [--team <teamSlug>]\n  orizu teams members add --email <email> [--team <teamSlug>]\n  orizu teams members remove --email <email> [--team <teamSlug>]\n  orizu teams members role --team <teamSlug> --email <email> --role <admin|member>\n  orizu projects list [--team <teamSlug>]\n  orizu projects create --name <name> [--team <teamSlug>]\n  orizu prompts list --project <team/project>\n  orizu prompts pull <prompt-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu prompts push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu prompts labels set <prompt-name> <label> --version <version-id> [--project <team/project>] [--json]\n  orizu prompts scorers set-headline <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu prompts scorers add <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu scorers list --project <team/project>\n  orizu scorers register --project <team/project> --name <name> --manifest <manifest.json> [--prompt-version <id>] [--runner-version <id>] [--label <label>] [--json]\n  orizu scorers detail <scorer-id-or-name> --project <team/project> [--json]\n  orizu scorers labels set <scorer-name> <label> --version <scorer-version-id> [--project <team/project>] [--json]\n  orizu scores submit <results.jsonl|results.json> --scorer-version <id> --subject-version <prompt-version-id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu judges list --project <team/project>\n  orizu judges pull <judge-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu judges push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu runners push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runners exec (--prompt <prompt-version-id> | --prompt-version <id> --runner-version <id> | --scorer-version <id>) --dataset-version <id> --split-set <id-or-name> --split <name> [--runner-dir <dir>] --out <results.jsonl|results.jsonl.gz>\n  orizu optimizers push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runs submit <results.jsonl|results.jsonl.gz> --prompt-version <id> --runner-version <id> --dataset-version <id> --split-set <id> --split <name> [--project <team/project>]\n  orizu apps list [--project <team/project>]\n  orizu apps create --project <team/project> --name <name> --dataset <datasetId> --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps update [--app <appId>] [--project <team/project>] --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps link-dataset --dataset <datasetId> [--app <appId>] [--project <team/project>] [--version <n>]\n  orizu apps detail --app <appId> [--project <team/project>] [--json]\n  orizu tasks list [--project <team/project>]\n  orizu tasks create --project <team/project> --dataset <datasetId> --app <appId> --title <title> --assignees <userIdOrEmail1,userIdOrEmail2> [--version <n>] [--instructions <text>] [--labels-per-item <n>] [--json]\n  orizu tasks assign --task <taskId> --assignees <userId1,userId2>\n  orizu tasks status --task <taskId> [--json]\n  orizu tasks pause --task <taskId>\n  orizu tasks unpause --task <taskId>\n  orizu datasets upload --file <path> [--project <team/project>] [--name <name>]\n  orizu datasets push <path> [--project <team/project>] [--name <name>] [--json]\n  orizu datasets versions create <datasetId|dataset-name> [--project <team/project>] [--label <label>] [--json]\n  orizu datasets splits create <datasetVersionId> [--from-file <split.json>] [--json]\n  orizu datasets download [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--format <csv|json|jsonl>] [--out <path>]\n  orizu datasets append [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets edit-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets delete-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --row-ids <id1,id2>\n  orizu datasets delete [--dataset <datasetId|datasetUrl>] [--project <team/project>]\n  orizu datasets lock [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--reason <text>]\n  orizu datasets clone [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--name <name>]\n  orizu tasks export [--task <taskId>] [--format <csv|json|jsonl>] [--out <path>]`);
+    printLine(`orizu global options:\n\n  --local                 Use http://localhost:3000\n  --server <url>          Use a specific server origin (for example: https://preview.example.com)\n  --version, -v           Print the orizu CLI version\n\norizu commands:\n\n  orizu login [--no-prompt-if-logged-in]\n  orizu logout\n  orizu whoami\n  orizu env [--project <team/project>] [--project-id <projectId>]\n  orizu log <event_type> --run-id <id> --sequence <n> --payload @event.json\n  orizu teams list\n  orizu teams create [--name <name>]\n  orizu teams members list [--team <teamSlug>]\n  orizu teams members add --email <email> [--team <teamSlug>]\n  orizu teams members remove --email <email> [--team <teamSlug>]\n  orizu teams members role --team <teamSlug> --email <email> --role <admin|member>\n  orizu projects list [--team <teamSlug>]\n  orizu projects create --name <name> [--team <teamSlug>]\n  orizu prompts list --project <team/project>\n  orizu prompts comments <prompt-id-or-name> --project <team/project> [--label <label> | --version <id>] [--json]\n  orizu prompts pull <prompt-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu prompts push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu prompts labels set <prompt-name> <label> --version <version-id> [--project <team/project>] [--json]\n  orizu prompts scorers set-headline <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu prompts scorers add <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu scorers list --project <team/project>\n  orizu scorers register --project <team/project> --name <name> --manifest <manifest.json> [--prompt-version <id>] [--runner-version <id>] [--label <label>] [--json]\n  orizu scorers detail <scorer-id-or-name> --project <team/project> [--json]\n  orizu scorers labels set <scorer-name> <label> --version <scorer-version-id> [--project <team/project>] [--json]\n  orizu scores submit <results.jsonl|results.json> --scorer-version <id> --subject-version <prompt-version-id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu judges list --project <team/project>\n  orizu judges pull <judge-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu judges push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu runners push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runners exec (--prompt <prompt-version-id> | --prompt-version <id> --runner-version <id> | --scorer-version <id>) --dataset-version <id> --split-set <id-or-name> --split <name> [--runner-dir <dir>] --out <results.jsonl|results.jsonl.gz>\n  orizu optimizers push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runs submit <results.jsonl|results.jsonl.gz> --prompt-version <id> --runner-version <id> --dataset-version <id> --split-set <id> --split <name> [--project <team/project>]\n  orizu apps list [--project <team/project>]\n  orizu apps create --project <team/project> --name <name> --dataset <datasetId> --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps update [--app <appId>] [--project <team/project>] --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps link-dataset --dataset <datasetId> [--app <appId>] [--project <team/project>] [--version <n>]\n  orizu apps detail --app <appId> [--project <team/project>] [--json]\n  orizu tasks list [--project <team/project>]\n  orizu tasks create --project <team/project> --dataset <datasetId> --app <appId> --title <title> --assignees <userIdOrEmail1,userIdOrEmail2> [--version <n>] [--instructions <text>] [--labels-per-item <n>] [--json]\n  orizu tasks assign --task <taskId> --assignees <userId1,userId2>\n  orizu tasks status --task <taskId> [--json]\n  orizu tasks pause --task <taskId>\n  orizu tasks unpause --task <taskId>\n  orizu datasets upload --file <path> [--project <team/project>] [--name <name>]\n  orizu datasets push <path> [--project <team/project>] [--name <name>] [--json]\n  orizu datasets versions create <datasetId|dataset-name> [--project <team/project>] [--label <label>] [--json]\n  orizu datasets splits create <datasetVersionId> [--from-file <split.json>] [--json]\n  orizu datasets download [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--format <csv|json|jsonl>] [--out <path>]\n  orizu datasets append [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets edit-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets delete-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --row-ids <id1,id2>\n  orizu datasets delete [--dataset <datasetId|datasetUrl>] [--project <team/project>]\n  orizu datasets lock [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--reason <text>]\n  orizu datasets clone [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--name <name>]\n  orizu tasks export [--task <taskId>] [--format <csv|json|jsonl>] [--out <path>]`);
 }
 function printOptimizationUsage() {
-    printLine(`\nOptimization lifecycle commands:\n\n  orizu optimizations start --project <team/project> --optimizer-version <id> --prompt-version <id[,id]> --selection-scorer <id> [--reflection-scorer <id>] [--pareto-scorer <id>] [--best-scorer <id>] --dataset-version <id> --split-set <id> [--train-split <name>] [--validation-split <name>] [--metadata <json|@file>] [--json]\n  orizu optimizations run-gepa --project <team/project> --optimizer-version-id <id> --candidate-version-id <id> --runner-version-id <id> --candidate-runner-dir <dir> --scorer-version-id <id> --scorer-runner-version-id <id> --scorer-runner-dir <dir> --dataset-version-id <id> --split-set-id <id> [--train-split train] [--val-split validation]\n  orizu optimizations pause <run-id> [--reason <text>] [--json]\n  orizu optimizations resume <run-id> [--json]\n  orizu optimizations finish <run-id> [--best-score <n>] [--best-candidate <id>] [--result-prompt-version <id>] [--metadata <json|@file>] [--json]\n  orizu optimizations fail <run-id> [--reason <text>] [--metadata <json|@file>] [--json]\n  orizu optimizations cancel <run-id> [--reason <text>] [--json]`);
+    printLine(`\nOptimization lifecycle commands:\n\n  orizu optimizations start --project <team/project> --optimizer-version <id> --prompt-version <id[,id]> --selection-scorer <id> [--reflection-scorer <id>] [--pareto-scorer <id>] [--best-scorer <id>] --dataset-version <id> --split-set <id> [--train-split <name>] [--validation-split <name>] [--metadata <json|@file>] [--json]\n  orizu optimizations run-gepa --project <team/project> --optimizer-version-id <id> --candidate-version-id <id> --runner-version-id <id> --candidate-runner-dir <dir> --scorer-version-id <id> --scorer-runner-version-id <id> --scorer-runner-dir <dir> --dataset-version-id <id> --split-set-id <id> [--train-split train] [--val-split validation] [--log-dir logs]\n  orizu optimizations export <run-id> [--out <path>] [--json]\n  orizu optimizations pause <run-id> [--reason <text>] [--json]\n  orizu optimizations resume <run-id> [--json]\n  orizu optimizations finish <run-id> [--best-score <n>] [--best-candidate <id>] [--result-prompt-version <id>] [--metadata <json|@file>] [--json]\n  orizu optimizations fail <run-id> [--reason <text>] [--metadata <json|@file>] [--json]\n  orizu optimizations cancel <run-id> [--reason <text>] [--json]`);
 }
 let cliArgs = process.argv.slice(2);
 function getArg(name) {
@@ -480,6 +480,63 @@ function printTaskStatusSummary(data) {
         });
     }
 }
+function compactTerminalText(value, maxLength = 160) {
+    const compact = sanitizeTerminalText(value).replace(/\s+/g, ' ').trim();
+    if (compact.length <= maxLength) {
+        return compact;
+    }
+    return `${compact.slice(0, Math.max(0, maxLength - 1)).trim()}…`;
+}
+function printIndentedBody(value, indent) {
+    const lines = sanitizeTerminalText(value || '').split(/\r?\n/);
+    for (const line of lines) {
+        printLine(`${indent}${line}`);
+    }
+}
+function formatPromptCommentAnchor(anchor) {
+    if (!anchor) {
+        return null;
+    }
+    const lineLabel = anchor.startLine && anchor.endLine
+        ? anchor.startLine === anchor.endLine
+            ? `line ${anchor.startLine}`
+            : `lines ${anchor.startLine}-${anchor.endLine}`
+        : null;
+    const selectedText = anchor.text ? `"${compactTerminalText(anchor.text, 180)}"` : null;
+    if (lineLabel && selectedText) {
+        return `${lineLabel}: ${selectedText}`;
+    }
+    return lineLabel || selectedText;
+}
+function printPromptComments(data) {
+    const versionLabel = data.version.versionLabel || (data.version.versionNumber ? `v${data.version.versionNumber}` : data.version.id);
+    printLine(`Prompt: ${sanitizeTerminalText(data.prompt.name)} (${sanitizeTerminalText(data.prompt.id)})`);
+    printLine(`Version: ${sanitizeTerminalText(String(versionLabel))} (${sanitizeTerminalText(data.version.id)})`);
+    printLine(`Comments: ${data.summary.threadCount} thread${data.summary.threadCount === 1 ? '' : 's'} ` +
+        `(${data.summary.openThreadCount} open, ${data.summary.resolvedThreadCount} resolved, ${data.summary.replyCount} replies)`);
+    if (data.comments.length === 0) {
+        printLine('\nNo comments found for this prompt version.');
+        return;
+    }
+    for (const [index, thread] of data.comments.entries()) {
+        const authorName = thread.author?.name || 'Unknown';
+        const anchor = formatPromptCommentAnchor(thread.anchor);
+        printLine('');
+        printLine(`${index + 1}. [${thread.status}] ${sanitizeTerminalText(authorName)} · ${sanitizeTerminalText(thread.createdAt)}`);
+        if (anchor) {
+            printLine(`   Selection: ${anchor}`);
+        }
+        printIndentedBody(thread.body, '   ');
+        if (thread.replies.length > 0) {
+            printLine(`   Replies (${thread.replies.length})`);
+            for (const reply of thread.replies) {
+                const replyAuthor = reply.author?.name || 'Unknown';
+                printLine(`   - ${sanitizeTerminalText(replyAuthor)} · ${sanitizeTerminalText(reply.createdAt)}`);
+                printIndentedBody(reply.body, '     ');
+            }
+        }
+    }
+}
 const DEFAULT_AUTH_CALLBACK_PORT = 43123;
 function resolveAuthCallbackPort() {
     const envPort = process.env.ORIZU_AUTH_PORT;
@@ -788,6 +845,33 @@ async function listJudges() {
     const data = await parseJsonResponse(response, 'Judges list');
     printPromptSummaries(data.judges, 'No judges found.');
 }
+async function listPromptComments() {
+    const promptRef = getPositionalArg(2);
+    const project = getArg('--project') || await resolveProjectSlug(null);
+    const label = getArg('--label');
+    const version = getArg('--version');
+    if (!promptRef) {
+        throw new Error('Usage: orizu prompts comments <prompt-id-or-name> --project <team/project> [--label <label> | --version <version-id>] [--json]');
+    }
+    if (label && version) {
+        throw new Error('Use either --label or --version, not both');
+    }
+    const params = new URLSearchParams({ project });
+    if (label)
+        params.set('label', label);
+    if (version)
+        params.set('version', version);
+    const response = await authedFetch(`/api/cli/prompts/${encodeURIComponent(promptRef)}/comments?${params.toString()}`);
+    if (!response.ok) {
+        throw new Error(`Failed to fetch prompt comments: ${await response.text()}`);
+    }
+    const data = await parseJsonResponse(response, 'Prompt comments');
+    if (hasJsonFlag()) {
+        printJson(data);
+        return;
+    }
+    printPromptComments(data);
+}
 async function listScorers() {
     const project = getArg('--project') || await resolveProjectSlug(null);
     const response = await authedFetch(`/api/cli/scorers?project=${encodeURIComponent(project)}`);
@@ -1473,6 +1557,27 @@ async function updateOptimizationRunLifecycle(action) {
         printLine(`Cancelled optimization run ${id}`);
     }
 }
+async function exportOptimizationRun() {
+    const runId = getPositionalArg(2) || getArg('--run-id');
+    const outPathArg = getArg('--out');
+    if (!runId) {
+        throw new Error('Usage: orizu optimizations export <run-id> [--out <path>] [--json]');
+    }
+    const response = await authedFetch(`/api/cli/optimization-runs/${encodeURIComponent(runId)}/export`);
+    if (!response.ok) {
+        throw new Error(`Failed to export optimization run: ${await response.text()}`);
+    }
+    const data = await parseJsonResponse(response, 'Optimization export');
+    if (hasJsonFlag() && !outPathArg) {
+        printJson(data);
+        return;
+    }
+    const filename = outPathArg
+        ? expandHomePath(outPathArg)
+        : `${runId}.optimization.json`;
+    writeTextFileEnsuringDir(filename, `${JSON.stringify(data, null, 2)}\n`);
+    printLine(`Saved optimization export to ${sanitizeTerminalText(filename)}`);
+}
 function removeFlagWithValue(args, flag) {
     const filtered = [];
     for (let index = 0; index < args.length; index += 1) {
@@ -2937,6 +3042,10 @@ export async function main(rawArgs = process.argv.slice(2)) {
         await listPrompts();
         return;
     }
+    if (command === 'prompts' && subcommand === 'comments') {
+        await listPromptComments();
+        return;
+    }
     if (command === 'prompts' && subcommand === 'pull') {
         await pullPromptArtifact('prompt');
         return;
@@ -3013,6 +3122,10 @@ export async function main(rawArgs = process.argv.slice(2)) {
         await runGepaOptimization();
         return;
     }
+    if (command === 'optimizations' && subcommand === 'export') {
+        await exportOptimizationRun();
+        return;
+    }
     if (command === 'optimizations' &&
         (subcommand === 'pause' ||
             subcommand === 'resume' ||

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "orizu",
-  "version": "0.3.2",
+  "version": "0.3.3",
   "private": false,
   "type": "module",
   "bin": {

package/src/index.ts

@@ -99,6 +99,62 @@ interface PromptSummary {
   description?: string | null
 }
 
+interface PromptCommentAuthor {
+  name: string
+  initials?: string
+  color?: string
+}
+
+interface PromptCommentAnchor {
+  text: string | null
+  startLine: number | null
+  endLine: number | null
+}
+
+interface PromptCommentReply {
+  id: string
+  body: string
+  author: PromptCommentAuthor
+  createdAt: string
+  updatedAt: string
+}
+
+interface PromptCommentThread {
+  id: string
+  status: 'open' | 'resolved'
+  body: string
+  anchor: PromptCommentAnchor | null
+  author: PromptCommentAuthor
+  createdAt: string
+  updatedAt: string
+  resolvedAt: string | null
+  resolvedByUserId: string | null
+  replyCount: number
+  replies: PromptCommentReply[]
+}
+
+interface PromptCommentsPayload {
+  prompt: {
+    id: string
+    name: string
+    role: string
+    description?: string | null
+  }
+  version: {
+    id: string
+    versionNumber?: number
+    versionLabel?: string | null
+    status?: string | null
+  }
+  summary: {
+    threadCount: number
+    openThreadCount: number
+    resolvedThreadCount: number
+    replyCount: number
+  }
+  comments: PromptCommentThread[]
+}
+
 interface ScorerSummary {
   id: string
   name: string
@@ -200,11 +256,11 @@ function printVersion() {
 }
 
 function printUsage() {
-  printLine(`orizu global options:\n\n  --local                 Use http://localhost:3000\n  --server <url>          Use a specific server origin (for example: https://preview.example.com)\n  --version, -v           Print the orizu CLI version\n\norizu commands:\n\n  orizu login [--no-prompt-if-logged-in]\n  orizu logout\n  orizu whoami\n  orizu env [--project <team/project>] [--project-id <projectId>]\n  orizu log <event_type> --run-id <id> --sequence <n> --payload @event.json\n  orizu teams list\n  orizu teams create [--name <name>]\n  orizu teams members list [--team <teamSlug>]\n  orizu teams members add --email <email> [--team <teamSlug>]\n  orizu teams members remove --email <email> [--team <teamSlug>]\n  orizu teams members role --team <teamSlug> --email <email> --role <admin|member>\n  orizu projects list [--team <teamSlug>]\n  orizu projects create --name <name> [--team <teamSlug>]\n  orizu prompts list --project <team/project>\n  orizu prompts pull <prompt-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu prompts push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu prompts labels set <prompt-name> <label> --version <version-id> [--project <team/project>] [--json]\n  orizu prompts scorers set-headline <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu prompts scorers add <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu scorers list --project <team/project>\n  orizu scorers register --project <team/project> --name <name> --manifest <manifest.json> [--prompt-version <id>] [--runner-version <id>] [--label <label>] [--json]\n  orizu scorers detail <scorer-id-or-name> --project <team/project> [--json]\n  orizu scorers labels set <scorer-name> <label> --version <scorer-version-id> [--project <team/project>] [--json]\n  orizu scores submit <results.jsonl|results.json> --scorer-version <id> --subject-version <prompt-version-id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu judges list --project <team/project>\n  orizu judges pull <judge-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu judges push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu runners push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runners exec (--prompt <prompt-version-id> | --prompt-version <id> --runner-version <id> | --scorer-version <id>) --dataset-version <id> --split-set <id-or-name> --split <name> [--runner-dir <dir>] --out <results.jsonl|results.jsonl.gz>\n  orizu optimizers push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runs submit <results.jsonl|results.jsonl.gz> --prompt-version <id> --runner-version <id> --dataset-version <id> --split-set <id> --split <name> [--project <team/project>]\n  orizu apps list [--project <team/project>]\n  orizu apps create --project <team/project> --name <name> --dataset <datasetId> --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps update [--app <appId>] [--project <team/project>] --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps link-dataset --dataset <datasetId> [--app <appId>] [--project <team/project>] [--version <n>]\n  orizu apps detail --app <appId> [--project <team/project>] [--json]\n  orizu tasks list [--project <team/project>]\n  orizu tasks create --project <team/project> --dataset <datasetId> --app <appId> --title <title> --assignees <userIdOrEmail1,userIdOrEmail2> [--version <n>] [--instructions <text>] [--labels-per-item <n>] [--json]\n  orizu tasks assign --task <taskId> --assignees <userId1,userId2>\n  orizu tasks status --task <taskId> [--json]\n  orizu tasks pause --task <taskId>\n  orizu tasks unpause --task <taskId>\n  orizu datasets upload --file <path> [--project <team/project>] [--name <name>]\n  orizu datasets push <path> [--project <team/project>] [--name <name>] [--json]\n  orizu datasets versions create <datasetId|dataset-name> [--project <team/project>] [--label <label>] [--json]\n  orizu datasets splits create <datasetVersionId> [--from-file <split.json>] [--json]\n  orizu datasets download [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--format <csv|json|jsonl>] [--out <path>]\n  orizu datasets append [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets edit-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets delete-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --row-ids <id1,id2>\n  orizu datasets delete [--dataset <datasetId|datasetUrl>] [--project <team/project>]\n  orizu datasets lock [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--reason <text>]\n  orizu datasets clone [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--name <name>]\n  orizu tasks export [--task <taskId>] [--format <csv|json|jsonl>] [--out <path>]`)
+  printLine(`orizu global options:\n\n  --local                 Use http://localhost:3000\n  --server <url>          Use a specific server origin (for example: https://preview.example.com)\n  --version, -v           Print the orizu CLI version\n\norizu commands:\n\n  orizu login [--no-prompt-if-logged-in]\n  orizu logout\n  orizu whoami\n  orizu env [--project <team/project>] [--project-id <projectId>]\n  orizu log <event_type> --run-id <id> --sequence <n> --payload @event.json\n  orizu teams list\n  orizu teams create [--name <name>]\n  orizu teams members list [--team <teamSlug>]\n  orizu teams members add --email <email> [--team <teamSlug>]\n  orizu teams members remove --email <email> [--team <teamSlug>]\n  orizu teams members role --team <teamSlug> --email <email> --role <admin|member>\n  orizu projects list [--team <teamSlug>]\n  orizu projects create --name <name> [--team <teamSlug>]\n  orizu prompts list --project <team/project>\n  orizu prompts comments <prompt-id-or-name> --project <team/project> [--label <label> | --version <id>] [--json]\n  orizu prompts pull <prompt-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu prompts push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu prompts labels set <prompt-name> <label> --version <version-id> [--project <team/project>] [--json]\n  orizu prompts scorers set-headline <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu prompts scorers add <prompt-id> --scorer-version <id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu scorers list --project <team/project>\n  orizu scorers register --project <team/project> --name <name> --manifest <manifest.json> [--prompt-version <id>] [--runner-version <id>] [--label <label>] [--json]\n  orizu scorers detail <scorer-id-or-name> --project <team/project> [--json]\n  orizu scorers labels set <scorer-name> <label> --version <scorer-version-id> [--project <team/project>] [--json]\n  orizu scores submit <results.jsonl|results.json> --scorer-version <id> --subject-version <prompt-version-id> [--dataset-version <id> --split-set <id> --split <name>] [--project <team/project>] [--json]\n  orizu judges list --project <team/project>\n  orizu judges pull <judge-id-or-name> --project <team/project> --out <dir> [--label <label> | --version <id>] [--json]\n  orizu judges push <dir> [--runner-version <id>] [--project <team/project>] [--parent <version-id>] [--json]\n  orizu runners push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runners exec (--prompt <prompt-version-id> | --prompt-version <id> --runner-version <id> | --scorer-version <id>) --dataset-version <id> --split-set <id-or-name> --split <name> [--runner-dir <dir>] --out <results.jsonl|results.jsonl.gz>\n  orizu optimizers push <dir> [--project <team/project>] [--name <name>] [--label <label>] [--json]\n  orizu runs submit <results.jsonl|results.jsonl.gz> --prompt-version <id> --runner-version <id> --dataset-version <id> --split-set <id> --split <name> [--project <team/project>]\n  orizu apps list [--project <team/project>]\n  orizu apps create --project <team/project> --name <name> --dataset <datasetId> --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps update [--app <appId>] [--project <team/project>] --file <path> --input-schema <json-path> --output-schema <json-path> [--component <name>]\n  orizu apps link-dataset --dataset <datasetId> [--app <appId>] [--project <team/project>] [--version <n>]\n  orizu apps detail --app <appId> [--project <team/project>] [--json]\n  orizu tasks list [--project <team/project>]\n  orizu tasks create --project <team/project> --dataset <datasetId> --app <appId> --title <title> --assignees <userIdOrEmail1,userIdOrEmail2> [--version <n>] [--instructions <text>] [--labels-per-item <n>] [--json]\n  orizu tasks assign --task <taskId> --assignees <userId1,userId2>\n  orizu tasks status --task <taskId> [--json]\n  orizu tasks pause --task <taskId>\n  orizu tasks unpause --task <taskId>\n  orizu datasets upload --file <path> [--project <team/project>] [--name <name>]\n  orizu datasets push <path> [--project <team/project>] [--name <name>] [--json]\n  orizu datasets versions create <datasetId|dataset-name> [--project <team/project>] [--label <label>] [--json]\n  orizu datasets splits create <datasetVersionId> [--from-file <split.json>] [--json]\n  orizu datasets download [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--format <csv|json|jsonl>] [--out <path>]\n  orizu datasets append [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets edit-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --file <path>\n  orizu datasets delete-rows [--dataset <datasetId|datasetUrl>] [--project <team/project>] --row-ids <id1,id2>\n  orizu datasets delete [--dataset <datasetId|datasetUrl>] [--project <team/project>]\n  orizu datasets lock [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--reason <text>]\n  orizu datasets clone [--dataset <datasetId|datasetUrl>] [--project <team/project>] [--name <name>]\n  orizu tasks export [--task <taskId>] [--format <csv|json|jsonl>] [--out <path>]`)
 }
 
 function printOptimizationUsage() {
-  printLine(`\nOptimization lifecycle commands:\n\n  orizu optimizations start --project <team/project> --optimizer-version <id> --prompt-version <id[,id]> --selection-scorer <id> [--reflection-scorer <id>] [--pareto-scorer <id>] [--best-scorer <id>] --dataset-version <id> --split-set <id> [--train-split <name>] [--validation-split <name>] [--metadata <json|@file>] [--json]\n  orizu optimizations run-gepa --project <team/project> --optimizer-version-id <id> --candidate-version-id <id> --runner-version-id <id> --candidate-runner-dir <dir> --scorer-version-id <id> --scorer-runner-version-id <id> --scorer-runner-dir <dir> --dataset-version-id <id> --split-set-id <id> [--train-split train] [--val-split validation]\n  orizu optimizations pause <run-id> [--reason <text>] [--json]\n  orizu optimizations resume <run-id> [--json]\n  orizu optimizations finish <run-id> [--best-score <n>] [--best-candidate <id>] [--result-prompt-version <id>] [--metadata <json|@file>] [--json]\n  orizu optimizations fail <run-id> [--reason <text>] [--metadata <json|@file>] [--json]\n  orizu optimizations cancel <run-id> [--reason <text>] [--json]`)
+  printLine(`\nOptimization lifecycle commands:\n\n  orizu optimizations start --project <team/project> --optimizer-version <id> --prompt-version <id[,id]> --selection-scorer <id> [--reflection-scorer <id>] [--pareto-scorer <id>] [--best-scorer <id>] --dataset-version <id> --split-set <id> [--train-split <name>] [--validation-split <name>] [--metadata <json|@file>] [--json]\n  orizu optimizations run-gepa --project <team/project> --optimizer-version-id <id> --candidate-version-id <id> --runner-version-id <id> --candidate-runner-dir <dir> --scorer-version-id <id> --scorer-runner-version-id <id> --scorer-runner-dir <dir> --dataset-version-id <id> --split-set-id <id> [--train-split train] [--val-split validation] [--log-dir logs]\n  orizu optimizations export <run-id> [--out <path>] [--json]\n  orizu optimizations pause <run-id> [--reason <text>] [--json]\n  orizu optimizations resume <run-id> [--json]\n  orizu optimizations finish <run-id> [--best-score <n>] [--best-candidate <id>] [--result-prompt-version <id>] [--metadata <json|@file>] [--json]\n  orizu optimizations fail <run-id> [--reason <text>] [--metadata <json|@file>] [--json]\n  orizu optimizations cancel <run-id> [--reason <text>] [--json]`)
 }
 
 let cliArgs = process.argv.slice(2)
@@ -835,6 +891,78 @@ function printTaskStatusSummary(data: TaskStatusPayload) {
   }
 }
 
+function compactTerminalText(value: string, maxLength = 160): string {
+  const compact = sanitizeTerminalText(value).replace(/\s+/g, ' ').trim()
+  if (compact.length <= maxLength) {
+    return compact
+  }
+
+  return `${compact.slice(0, Math.max(0, maxLength - 1)).trim()}…`
+}
+
+function printIndentedBody(value: string, indent: string) {
+  const lines = sanitizeTerminalText(value || '').split(/\r?\n/)
+  for (const line of lines) {
+    printLine(`${indent}${line}`)
+  }
+}
+
+function formatPromptCommentAnchor(anchor: PromptCommentAnchor | null): string | null {
+  if (!anchor) {
+    return null
+  }
+
+  const lineLabel = anchor.startLine && anchor.endLine
+    ? anchor.startLine === anchor.endLine
+      ? `line ${anchor.startLine}`
+      : `lines ${anchor.startLine}-${anchor.endLine}`
+    : null
+  const selectedText = anchor.text ? `"${compactTerminalText(anchor.text, 180)}"` : null
+
+  if (lineLabel && selectedText) {
+    return `${lineLabel}: ${selectedText}`
+  }
+  return lineLabel || selectedText
+}
+
+function printPromptComments(data: PromptCommentsPayload) {
+  const versionLabel = data.version.versionLabel || (
+    data.version.versionNumber ? `v${data.version.versionNumber}` : data.version.id
+  )
+
+  printLine(`Prompt: ${sanitizeTerminalText(data.prompt.name)} (${sanitizeTerminalText(data.prompt.id)})`)
+  printLine(`Version: ${sanitizeTerminalText(String(versionLabel))} (${sanitizeTerminalText(data.version.id)})`)
+  printLine(
+    `Comments: ${data.summary.threadCount} thread${data.summary.threadCount === 1 ? '' : 's'} ` +
+    `(${data.summary.openThreadCount} open, ${data.summary.resolvedThreadCount} resolved, ${data.summary.replyCount} replies)`
+  )
+
+  if (data.comments.length === 0) {
+    printLine('\nNo comments found for this prompt version.')
+    return
+  }
+
+  for (const [index, thread] of data.comments.entries()) {
+    const authorName = thread.author?.name || 'Unknown'
+    const anchor = formatPromptCommentAnchor(thread.anchor)
+    printLine('')
+    printLine(`${index + 1}. [${thread.status}] ${sanitizeTerminalText(authorName)} · ${sanitizeTerminalText(thread.createdAt)}`)
+    if (anchor) {
+      printLine(`   Selection: ${anchor}`)
+    }
+    printIndentedBody(thread.body, '   ')
+
+    if (thread.replies.length > 0) {
+      printLine(`   Replies (${thread.replies.length})`)
+      for (const reply of thread.replies) {
+        const replyAuthor = reply.author?.name || 'Unknown'
+        printLine(`   - ${sanitizeTerminalText(replyAuthor)} · ${sanitizeTerminalText(reply.createdAt)}`)
+        printIndentedBody(reply.body, '     ')
+      }
+    }
+  }
+}
+
 const DEFAULT_AUTH_CALLBACK_PORT = 43123
 
 function resolveAuthCallbackPort(): number {
@@ -1185,6 +1313,37 @@ async function listJudges() {
   printPromptSummaries(data.judges, 'No judges found.')
 }
 
+async function listPromptComments() {
+  const promptRef = getPositionalArg(2)
+  const project = getArg('--project') || await resolveProjectSlug(null)
+  const label = getArg('--label')
+  const version = getArg('--version')
+
+  if (!promptRef) {
+    throw new Error('Usage: orizu prompts comments <prompt-id-or-name> --project <team/project> [--label <label> | --version <version-id>] [--json]')
+  }
+  if (label && version) {
+    throw new Error('Use either --label or --version, not both')
+  }
+
+  const params = new URLSearchParams({ project })
+  if (label) params.set('label', label)
+  if (version) params.set('version', version)
+
+  const response = await authedFetch(`/api/cli/prompts/${encodeURIComponent(promptRef)}/comments?${params.toString()}`)
+  if (!response.ok) {
+    throw new Error(`Failed to fetch prompt comments: ${await response.text()}`)
+  }
+
+  const data = await parseJsonResponse<PromptCommentsPayload>(response, 'Prompt comments')
+  if (hasJsonFlag()) {
+    printJson(data as unknown as Record<string, unknown>)
+    return
+  }
+
+  printPromptComments(data)
+}
+
 async function listScorers() {
   const project = getArg('--project') || await resolveProjectSlug(null)
   const response = await authedFetch(`/api/cli/scorers?project=${encodeURIComponent(project)}`)
@@ -2070,6 +2229,32 @@ async function updateOptimizationRunLifecycle(action: OptimizationLifecycleActio
   }
 }
 
+async function exportOptimizationRun() {
+  const runId = getPositionalArg(2) || getArg('--run-id')
+  const outPathArg = getArg('--out')
+
+  if (!runId) {
+    throw new Error('Usage: orizu optimizations export <run-id> [--out <path>] [--json]')
+  }
+
+  const response = await authedFetch(`/api/cli/optimization-runs/${encodeURIComponent(runId)}/export`)
+  if (!response.ok) {
+    throw new Error(`Failed to export optimization run: ${await response.text()}`)
+  }
+
+  const data = await parseJsonResponse<Record<string, unknown>>(response, 'Optimization export')
+  if (hasJsonFlag() && !outPathArg) {
+    printJson(data)
+    return
+  }
+
+  const filename = outPathArg
+    ? expandHomePath(outPathArg)
+    : `${runId}.optimization.json`
+  writeTextFileEnsuringDir(filename, `${JSON.stringify(data, null, 2)}\n`)
+  printLine(`Saved optimization export to ${sanitizeTerminalText(filename)}`)
+}
+
 function removeFlagWithValue(args: string[], flag: string): string[] {
   const filtered: string[] = []
   for (let index = 0; index < args.length; index += 1) {
@@ -3930,6 +4115,11 @@ export async function main(rawArgs = process.argv.slice(2)) {
     return
   }
 
+  if (command === 'prompts' && subcommand === 'comments') {
+    await listPromptComments()
+    return
+  }
+
   if (command === 'prompts' && subcommand === 'pull') {
     await pullPromptArtifact('prompt')
     return
@@ -4025,6 +4215,11 @@ export async function main(rawArgs = process.argv.slice(2)) {
     return
   }
 
+  if (command === 'optimizations' && subcommand === 'export') {
+    await exportOptimizationRun()
+    return
+  }
+
   if (
     command === 'optimizations' &&
     (

package/vendor/orizu-gepa-python/src/orizu_gepa/cli.py

@@ -6,6 +6,7 @@ from pathlib import Path
 from typing import Any
 
 from .client import OrizuClient, OrizuEventSink
+from .local_log import LocalOptimizationLogger
 from .optimizer import TextGepaConfig, optimize_loaded_text_candidate
 from .reflection import reflect_with_provider
 from .runner import make_candidate_runner, make_scorer_runner
@@ -74,6 +75,8 @@ def main() -> None:
     parser.add_argument("--auto-promote", action="store_true")
     parser.add_argument("--promotion-label")
     parser.add_argument("--log-row-snapshots", action="store_true")
+    parser.add_argument("--log-dir", default="logs")
+    parser.add_argument("--no-local-log", action="store_true")
     parser.add_argument("--metadata", default="{}")
     args = parser.parse_args()
 
@@ -123,6 +126,19 @@ def main() -> None:
         split=args.val_split,
     )
 
+    run_metadata = {
+        **metadata,
+        "optimizer_package": "orizu-gepa-python",
+        "optimizer_family": "gepa",
+        "mode": "text-candidate",
+        "inference_lm": prompt_context.provider_settings.get("model"),
+        "reflection_lm": config.reflection_model,
+        "scorer_lm": scorer_context.provider_settings.get("model"),
+        "dataset_size": len(trainset) + len(valset),
+        "train_count": len(trainset),
+        "validation_count": len(valset),
+    }
+
     run_id = client.start_run(
         project=args.project,
         optimizer_version_id=args.optimizer_version_id,
@@ -132,21 +148,29 @@ def main() -> None:
         split_set_id=args.split_set_id,
         train_split=args.train_split,
         validation_split=args.val_split,
-        metadata={
-            **metadata,
-            "optimizer_package": "orizu-gepa-python",
-            "optimizer_family": "gepa",
-            "mode": "text-candidate",
-            "inference_lm": prompt_context.provider_settings.get("model"),
-            "reflection_lm": config.reflection_model,
-            "scorer_lm": scorer_context.provider_settings.get("model"),
-            "dataset_size": len(trainset) + len(valset),
-            "train_count": len(trainset),
-            "validation_count": len(valset),
-        },
+        metadata=run_metadata,
     )
     print(f"[orizu-gepa] started optimization run {run_id}", flush=True)
 
+    local_logger = None
+    if not args.no_local_log:
+        local_logger = LocalOptimizationLogger.create(args.log_dir, run_id)
+        local_logger.write_context(
+            project=args.project,
+            run_id=run_id,
+            args={
+                key: value
+                for key, value in vars(args).items()
+                if key not in {"metadata"}
+            },
+            prompt_context=prompt_context,
+            scorer_context=scorer_context,
+            trainset=trainset,
+            valset=valset,
+            metadata=run_metadata,
+        )
+        print(f"[orizu-gepa] local log: {local_logger.path}", flush=True)
+
     result = optimize_loaded_text_candidate(
         run_id=run_id,
         prompt_context=prompt_context,
@@ -156,9 +180,17 @@ def main() -> None:
         candidate_runner=make_candidate_runner(Path(args.candidate_runner_dir), run_id),
         scorer_runner=make_scorer_runner(Path(args.scorer_runner_dir), run_id),
         reflector=reflect_with_provider,
-        event_sink=OrizuEventSink(client, run_id, fail_on_log_error=config.fail_on_log_error),
+        event_sink=OrizuEventSink(
+            client,
+            run_id,
+            fail_on_log_error=config.fail_on_log_error,
+            local_logger=local_logger,
+        ),
         config=config,
+        local_logger=local_logger,
     )
+    if local_logger is not None:
+        local_logger.write_result(result)
     print(json.dumps({
         "optimization_run_id": result.run_id,
         "best_candidate_id": result.best_candidate_id,
@@ -166,6 +198,7 @@ def main() -> None:
         "seed_score": result.seed_score,
         "promoted_prompt_version_id": result.promoted_prompt_version_id,
         "budget": result.budget.to_payload(),
+        "local_log_dir": local_logger.path if local_logger is not None else None,
     }))
 
 

package/vendor/orizu-gepa-python/src/orizu_gepa/client.py

@@ -255,12 +255,14 @@ class OrizuEventSink:
         *,
         fail_on_log_error: bool = True,
         max_log_retries: int = 2,
+        local_logger: Any | None = None,
     ):
         self.client = client
         self.run_id = run_id
         self.sequence = 0
         self.fail_on_log_error = fail_on_log_error
         self.max_log_retries = max(0, max_log_retries)
+        self.local_logger = local_logger
 
     def log_event(
         self,
@@ -275,6 +277,20 @@ class OrizuEventSink:
         child_candidate_id: str | None = None,
     ) -> None:
         self.sequence += 1
+        local_event = {
+            "run_id": self.run_id,
+            "sequence": self.sequence,
+            "event_type": event_type,
+            "event_layer": event_layer,
+            "optimizer_family": optimizer_family,
+            "iteration": iteration,
+            "candidate_id": candidate_id,
+            "parent_candidate_id": parent_candidate_id,
+            "child_candidate_id": child_candidate_id,
+            "payload": payload or {},
+        }
+        if self.local_logger is not None:
+            self.local_logger.append_event(local_event)
         last_error: Exception | None = None
         for attempt in range(self.max_log_retries + 1):
             try:
@@ -308,7 +324,37 @@ class OrizuEventSink:
         prompt_version_id = self.client.promote_candidate(self.run_id, **kwargs)
         # Promotion writes a system event at max(sequence)+1 in Postgres.
         self.sequence += 1
+        if self.local_logger is not None:
+            self.local_logger.append_event({
+                "run_id": self.run_id,
+                "sequence": self.sequence,
+                "event_type": "candidate_promoted",
+                "event_layer": "system",
+                "optimizer_family": "gepa",
+                "iteration": None,
+                "candidate_id": kwargs.get("candidate_id"),
+                "parent_candidate_id": None,
+                "child_candidate_id": None,
+                "payload": {
+                    **kwargs,
+                    "prompt_version_id": prompt_version_id,
+                },
+            })
         return prompt_version_id
 
     def finish_run(self, **kwargs: Any) -> None:
+        if self.local_logger is not None:
+            self.sequence += 1
+            self.local_logger.append_event({
+                "run_id": self.run_id,
+                "sequence": self.sequence,
+                "event_type": "run_status_updated",
+                "event_layer": "system",
+                "optimizer_family": "gepa",
+                "iteration": None,
+                "candidate_id": kwargs.get("best_candidate_id"),
+                "parent_candidate_id": None,
+                "child_candidate_id": None,
+                "payload": kwargs,
+            })
         self.client.update_run(self.run_id, **kwargs)

package/vendor/orizu-gepa-python/src/orizu_gepa/local_log.py

@@ -0,0 +1,123 @@
+from __future__ import annotations
+
+import dataclasses
+import json
+from pathlib import Path
+from typing import Any
+
+from .optimizer import DatasetRow, PromptContext, RowEvaluation, TextGepaResult
+
+
+def _json_default(value: Any) -> Any:
+    if dataclasses.is_dataclass(value):
+        return dataclasses.asdict(value)
+    return str(value)
+
+
+def _write_json(path: Path, payload: dict[str, Any]) -> None:
+    path.write_text(
+        json.dumps(payload, indent=2, ensure_ascii=False, default=_json_default) + "\n",
+        encoding="utf-8",
+    )
+
+
+def _append_jsonl(path: Path, payload: dict[str, Any]) -> None:
+    with path.open("a", encoding="utf-8") as handle:
+        handle.write(json.dumps(payload, ensure_ascii=False, default=_json_default) + "\n")
+
+
+class LocalOptimizationLogger:
+    def __init__(self, directory: Path):
+        self.directory = directory
+        self.directory.mkdir(parents=True, exist_ok=True)
+
+    @classmethod
+    def create(cls, root: str | Path, run_id: str) -> "LocalOptimizationLogger":
+        return cls(Path(root) / run_id)
+
+    @property
+    def path(self) -> str:
+        return str(self.directory)
+
+    def write_context(
+        self,
+        *,
+        project: str,
+        run_id: str,
+        args: dict[str, Any],
+        prompt_context: PromptContext,
+        scorer_context: PromptContext,
+        trainset: list[DatasetRow],
+        valset: list[DatasetRow],
+        metadata: dict[str, Any],
+    ) -> None:
+        _write_json(self.directory / "run.json", {
+            "schema_version": "orizu.optimization-local-log.v1",
+            "optimization_run_id": run_id,
+            "project": project,
+            "args": args,
+            "metadata": metadata,
+        })
+        _write_json(self.directory / "prompt_context.json", dataclasses.asdict(prompt_context))
+        _write_json(self.directory / "scorer_context.json", dataclasses.asdict(scorer_context))
+        _write_json(self.directory / "trainset.json", {
+            "rows": [dataclasses.asdict(row) for row in trainset],
+        })
+        _write_json(self.directory / "valset.json", {
+            "rows": [dataclasses.asdict(row) for row in valset],
+        })
+
+    def append_event(self, event: dict[str, Any]) -> None:
+        _append_jsonl(self.directory / "events.jsonl", event)
+
+    def append_evaluations(
+        self,
+        *,
+        stage: str,
+        split: str,
+        iteration: int | None,
+        candidate_id: str,
+        parent_candidate_id: str | None = None,
+        results: list[RowEvaluation],
+    ) -> None:
+        for result in results:
+            _append_jsonl(self.directory / "evaluations.jsonl", {
+                "stage": stage,
+                "split": split,
+                "iteration": iteration,
+                "candidate_id": candidate_id,
+                "parent_candidate_id": parent_candidate_id,
+                **result.to_payload(),
+            })
+
+    def append_reflection(
+        self,
+        *,
+        iteration: int,
+        parent_candidate_id: str,
+        child_candidate_id: str,
+        row_ids: list[str],
+        prompt: str,
+        response: str,
+        candidate_text: str,
+    ) -> None:
+        _append_jsonl(self.directory / "reflections.jsonl", {
+            "iteration": iteration,
+            "parent_candidate_id": parent_candidate_id,
+            "child_candidate_id": child_candidate_id,
+            "row_ids": row_ids,
+            "prompt": prompt,
+            "response": response,
+            "candidate_text": candidate_text,
+        })
+
+    def write_result(self, result: TextGepaResult) -> None:
+        _write_json(self.directory / "result.json", {
+            "optimization_run_id": result.run_id,
+            "best_candidate_id": result.best_candidate_id,
+            "best_candidate_text": result.best_candidate_text,
+            "best_score": result.best_score,
+            "seed_score": result.seed_score,
+            "promoted_prompt_version_id": result.promoted_prompt_version_id,
+            "budget": result.budget.to_payload(),
+        })

package/vendor/orizu-gepa-python/src/orizu_gepa/optimizer.py

@@ -658,6 +658,7 @@ def optimize_loaded_text_candidate(
     reflector: Reflector,
     event_sink: EventSink,
     config: TextGepaConfig,
+    local_logger: Any | None = None,
 ) -> TextGepaResult:
     seed_text = prompt_context.body or ""
     budget = Budget.from_config(config)
@@ -709,6 +710,14 @@ def optimize_loaded_text_candidate(
             budget=budget,
             evaluation_cache=evaluation_cache,
         )
+        if local_logger is not None:
+            local_logger.append_evaluations(
+                stage="seed_val_set",
+                split="validation",
+                iteration=None,
+                candidate_id="seed",
+                results=seed_results,
+            )
         if not _completed_all_rows(seed_results, valset):
             raise RuntimeError("Budget exhausted before seed validation completed")
         budget.used_full_evals += 1
@@ -804,6 +813,14 @@ def optimize_loaded_text_candidate(
                 budget=budget,
                 evaluation_cache=evaluation_cache,
             )
+            if local_logger is not None:
+                local_logger.append_evaluations(
+                    stage="parent_minibatch",
+                    split="train",
+                    iteration=iteration,
+                    candidate_id=parent_candidate_id,
+                    results=parent_results,
+                )
             if not _completed_all_rows(parent_results, minibatch):
                 event_sink.log_event(
                     "budget_exhausted",
@@ -852,11 +869,21 @@ def optimize_loaded_text_candidate(
             child_text = reflection.candidate_text
             candidate_text_by_id[child_id] = child_text
             budget.used_candidate_proposals += 1
+            if local_logger is not None:
+                local_logger.append_reflection(
+                    iteration=iteration,
+                    parent_candidate_id=parent_candidate_id,
+                    child_candidate_id=child_id,
+                    row_ids=[row.id for row in minibatch],
+                    prompt=reflection.prompt,
+                    response=reflection.response,
+                    candidate_text=reflection.candidate_text,
+                )
             event_sink.log_event(
                 "reflection_completed",
                 {
                     **_text_log_fields("prompt", reflection.prompt, include_text=config.log_row_snapshots),
-                    **_text_log_fields("response", reflection.response, include_text=config.log_row_snapshots),
+                    "response": reflection.response,
                     "child_candidate_id": child_id,
                     "candidate_text": child_text,
                 },
@@ -908,6 +935,15 @@ def optimize_loaded_text_candidate(
                 budget=budget,
                 evaluation_cache=evaluation_cache,
             )
+            if local_logger is not None:
+                local_logger.append_evaluations(
+                    stage="child_minibatch",
+                    split="train",
+                    iteration=iteration,
+                    candidate_id=child_id,
+                    parent_candidate_id=parent_candidate_id,
+                    results=child_results,
+                )
             if not _completed_all_rows(child_results, minibatch):
                 event_sink.log_event(
                     "budget_exhausted",
@@ -981,6 +1017,15 @@ def optimize_loaded_text_candidate(
                     budget=budget,
                     evaluation_cache=evaluation_cache,
                 )
+                if local_logger is not None:
+                    local_logger.append_evaluations(
+                        stage="child_val_set",
+                        split="validation",
+                        iteration=iteration,
+                        candidate_id=child_id,
+                        parent_candidate_id=parent_candidate_id,
+                        results=child_val_results,
+                    )
                 if not _completed_all_rows(child_val_results, valset):
                     event_sink.log_event(
                         "budget_exhausted",