@letterstory/cli 0.123.0 → 0.125.0

package/lib/cli.mjs

@@ -61,13 +61,15 @@ import {
 } from "./commands.mjs";
 
 // Keep in sync with cli/package.json.
-export const VERSION = "0.123.0";
+export const VERSION = "0.125.0";
 
 // Flags that never take a value. Listing them explicitly means `deploy get --json <id>`
 // can't accidentally swallow the id as --json's value. Every `type: "bool"` flag (and its
 // `negate` spelling) in a command's ARGS map must be listed; boolean-flags.test.ts enforces it.
 export const BOOLEAN_FLAGS = new Set([
 	"json",
+	// admin models runs: only the starred ones.
+	"starred",
 	// shredder file / arena file: save the version as a candidate beside live.
 	"candidate",
 	// site tree: include children a fold-by-default family would hide; items over the whole subtree.
@@ -843,6 +845,8 @@ Admin (STAFF key — --staff-key / LETTERSTORY_STAFF_API_KEY; acts before an org
                                                        session (a trial per arm), linked to its
                                                        ledger entry. Written by \`npm run lab --
                                                        publish\` in the kernels repo.
+  admin models runs [--model M] [--starred] | run <id> | star|unstar <id> | scoreboard --model M
+                                                       The model hub's saved runs and blind arena (models.letter.dev)
   admin spend summary [--range day|week|month|year | --from D --to D] [--grain day|week|month]
           | reconcile [--from D --to D] | pnl [--range day|week|month|year | --from D --to D]
           | reference add --vendor V [--key K] --metric billed_usd --value N --from D [--to D] --ref "…"
@@ -1332,6 +1336,8 @@ Repo changes (Claude Code edits your repo in a sandbox; the Letterstory GitHub A
   repo list [--status running|succeeded|no_changes|failed|cancelled] [--limit n] [--audit <id>]
              [--source site_audit_issue|site_quick_check|letterprove_install]
   repo publish <run-id>                                Make PR on a preview in review
+  repo revise <run-id> --instructions <text>|--instructions-file <path|->
+                                                       Ask for more on a preview; it takes over the preview's link
   repo cancel <run-id>                                 Stop it and delete the sandbox; no PR (Discard a preview)
 
 Feedback (tell us what's broken, missing, unclear, or wanted):

package/lib/commands/admin.mjs

@@ -19,6 +19,7 @@
 //   admin brand-detection <org-id>                                                   (fleet:read)
 //   admin brand-detection <org-id> [--term t]... [--exact t]... [--exclude p]... [--note "…"] | --clear [--note "…"]   (fleet:instrument)
 //   admin recipes list|show|vocab|export|import|save|delete …                        (recipes:read / recipes:write)
+//   admin models runs|run|star|unstar|scoreboard …                                    (models:read / models:write)
 //   admin replays [--org <id>] [--target <id>] [--status …] [--limit n]                (fleet:read)
 //   admin replay --org <id> [<response-id> | --target <id>] [--surface …]           (fleet:replay)
 //   admin create-org --domain acme.com [--name "Acme"] [--test] [--no-key] [--key-name N]
@@ -3962,6 +3963,83 @@ function spendEntryFromFlags(flags) {
 	});
 }
 
+const MODELS = "/api/superadmin/models";
+const MODELS_USAGE =
+	"Usage: admin models runs [--model ballpoint-1] [--starred] [--limit N] | run <id> | star <id> | unstar <id> | scoreboard --model ballpoint-1";
+
+/** The model hub's saved runs and blind arena (models.letter.dev), same functions as its pages. */
+async function cmdAdminModels(ctx) {
+	const { flags, positionals, io } = ctx;
+	const sub = positionals[0];
+	switch (sub) {
+		case "runs": {
+			const params = new URLSearchParams(
+				compact({
+					model: flagStr(flags.model),
+					starred: flagBool(flags.starred) ? "true" : undefined,
+					limit: flagNum(flags.limit)?.toString(),
+				})
+			);
+			const result = await staffRequest(ctx, "GET", `${MODELS}/runs${params.size ? `?${params}` : ""}`);
+			printResult(io, flags, result, (r) =>
+				(r.runs ?? [])
+					.map(
+						(x) =>
+							`${x.created_at.slice(0, 16)}  ${x.starred ? "★" : " "} ${x.model.padEnd(14)} ${String(x.outputs).padStart(3)} outputs ${String(x.votes).padStart(3)} votes  ${x.contenders.join(", ")}  ${x.id}`
+					)
+					.join("\n")
+			);
+			return 0;
+		}
+		case "run": {
+			const id = positionals[1];
+			if (!id) throw new CliError("Usage: admin models run <id>");
+			const result = await staffRequest(ctx, "GET", `${MODELS}/runs/${id}`);
+			printResult(io, flags, result, (r) =>
+				[
+					`${r.model} · ${r.created_at} · ${JSON.stringify(r.input)}`,
+					...(r.outputs ?? []).map(
+						(o) =>
+							`  [${o.slot}] ${o.contender}${o.ours ? " (ours)" : ""}: ${o.error ? `error: ${o.error}` : Object.values(o.output).join(" — ")}`
+					),
+				].join("\n")
+			);
+			return 0;
+		}
+		case "star":
+		case "unstar": {
+			const id = positionals[1];
+			if (!id) throw new CliError(`Usage: admin models ${sub} <id>`);
+			const result = await staffRequest(ctx, "PATCH", `${MODELS}/runs/${id}`, { starred: sub === "star" });
+			ok(ctx, `${sub === "star" ? "Starred" : "Unstarred"} ${result.id}`);
+			printResult(io, flags, result, () => "");
+			return 0;
+		}
+		case "scoreboard": {
+			const model = flagStr(flags.model);
+			if (!model) throw new CliError("Usage: admin models scoreboard --model ballpoint-1");
+			const result = await staffRequest(
+				ctx,
+				"GET",
+				`${MODELS}/arena/scoreboard?model=${encodeURIComponent(model)}`
+			);
+			printResult(io, flags, result, (r) =>
+				(r.opponents ?? []).length
+					? r.opponents
+							.map(
+								(o) =>
+									`${r.model} vs ${o.opponent}: ${o.wins}-${o.losses} (${o.ties} ties, ${o.both_bad} both bad) · ${o.win_rate == null ? "—" : `${Math.round(o.win_rate * 100)}% [${Math.round(o.low * 100)}–${Math.round(o.high * 100)}]`} · ${o.voters} voters`
+							)
+							.join("\n")
+					: `No votes on ${r.model} yet.`
+			);
+			return 0;
+		}
+		default:
+			throw new CliError(MODELS_USAGE);
+	}
+}
+
 async function cmdAdminSpend(ctx) {
 	const { flags, positionals, io } = ctx;
 	const sub = positionals[0];
@@ -4767,6 +4845,7 @@ export async function cmdAdmin(ctx) {
 		return cmdAdminReplays({ ...ctx, positionals: ctx.positionals.slice(1) }, sub);
 	if (sub === "crm") return cmdAdminCrm({ ...ctx, positionals: ctx.positionals.slice(1) });
 	if (sub === "spend") return cmdAdminSpend({ ...ctx, positionals: ctx.positionals.slice(1) });
+	if (sub === "models") return cmdAdminModels({ ...ctx, positionals: ctx.positionals.slice(1) });
 	if (sub === "create-org") return cmdAdminCreateOrg({ ...ctx, positionals: ctx.positionals.slice(1) });
 	if (sub === "ask") return cmdAdminAsk({ ...ctx, positionals: ctx.positionals.slice(1) });
 	if (sub === "patterns") return cmdAdminPatterns({ ...ctx, positionals: ctx.positionals.slice(1) });

package/lib/commands/repo.mjs

@@ -4,6 +4,7 @@
 //   repo activity → get_repo_change_activity (what the sandbox is doing; --follow polls)
 //   repo list    → list_repo_changes     repo cancel → cancel_repo_change
 //   repo publish → publish_repo_change (Make PR on a preview)
+//   repo revise  → revise_repo_change (ask for more on a preview, before the PR)
 //   repo repos   → list_github_app_repos  repo branches → list_repo_branches
 //   repo prs     → list_open_pull_requests (every open PR, whoever opened it)
 //   repo install-letterprove → letterprove_open_install_pr
@@ -167,6 +168,10 @@ const ARGS = {
 	get_repo_change_activity: { run_id: { positional: 0, required: true, label: "run-id" } },
 	cancel_repo_change: { run_id: { positional: 0, required: true, label: "run-id" } },
 	publish_repo_change: { run_id: { positional: 0, required: true, label: "run-id" } },
+	revise_repo_change: {
+		run_id: { positional: 0, required: true, label: "run-id" },
+		instructions: { value: (flags) => readTextInput(flags, "instructions", "instructions-file") },
+	},
 	list_repo_changes: {
 		status: { flag: "status", enum: ["running", "succeeded", "no_changes", "failed", "cancelled", "sent"] },
 		limit: { flag: "limit", type: "num" },
@@ -195,8 +200,14 @@ function describe(run) {
 		`${run.id}  ${run.repo}@${run.baseBranch}  ${run.status} (${run.phase})  $${run.costUsd.toFixed(2)}`,
 	];
 	if (run.preview?.url) lines.push(`  Preview: ${run.preview.url}`);
+	for (const [i, change] of (run.preview?.changes ?? []).entries())
+		lines.push(`    ${i + 1}. ${change.path}  ${change.note}${change.url ? `  ${change.url}` : ""}`);
 	if (run.preview?.state === "failed") lines.push(`  Preview failed: ${run.preview.detail ?? ""}`);
-	if (run.phase === "review") lines.push(`  Make PR: letterstory repo publish ${run.id}`);
+	if (run.preview?.revisionRunId) lines.push(`  Follow-up being built: ${run.preview.revisionRunId}`);
+	else if (run.phase === "review") {
+		lines.push(`  Make PR: letterstory repo publish ${run.id}`);
+		lines.push(`  Ask for more: letterstory repo revise ${run.id} --instructions "…"`);
+	}
 	if (run.prUrl) lines.push(`  PR: ${run.prUrl} (${run.prState ?? "state unknown"})`);
 	if (run.errorCode) lines.push(`  Error (${run.errorCode}): ${run.errorDetail ?? ""}`);
 	if (run.summary) lines.push(...run.summary.split("\n").map((l) => `  ${l}`));
@@ -235,6 +246,8 @@ export async function cmdRepo(ctx) {
 			return repoCancel(rest);
 		case "publish":
 			return repoPublish(rest);
+		case "revise":
+			return repoRevise(rest);
 		case "repos":
 			return repoRepos(rest);
 		case "branches":
@@ -549,6 +562,17 @@ async function repoPublish(ctx) {
 	return 0;
 }
 
+/** `repo revise <run-id> --instructions <text>`: a follow-up on a preview, built on top of it. */
+async function repoRevise(ctx) {
+	const { client, flags, io } = ctx;
+	const args = argsFromFlags(ARGS.revise_repo_change, flags, ctx.positionals);
+	if (!args.instructions) throw new CliError("Missing --instructions (or --instructions-file <path|->)");
+	const result = await client.callTool("revise_repo_change", args);
+	ok(ctx, result.message);
+	printResult(io, flags, result.run, describe);
+	return 0;
+}
+
 async function repoRepos(ctx) {
 	const { client, flags, io } = ctx;
 	const result = await client.callTool("list_github_app_repos", argsFromFlags(ARGS.list_github_app_repos, flags));

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@letterstory/cli",
-	"version": "0.123.0",
+	"version": "0.125.0",
 	"description": "Spin up and manage Letterstory phantom blogs from your terminal.",
 	"type": "module",
 	"license": "UNLICENSED",