feat: GPT-6 Sol and GPT-6 Luna - catalog rows, family prompt preset, recommended default

.agents/skills/senpi-qa/scripts/gpt-6-family-preset-mock-loop.mjs

@@ -0,0 +1,133 @@
+/**
+ * Channel 3 proof: GPT-6 Sol and GPT-6 Luna reach the wire on the GPT-6 preset.
+ *
+ * Boots the real senpi CLI in --print mode against a fake OpenAI Responses
+ * server that serves the built-in catalog rows for `gpt-6-sol` and `gpt-6-luna`,
+ * then inspects each captured request: the model id and the requested reasoning
+ * effort must be the ones the CLI was asked for, and the developer/system message
+ * must carry the GPT-6 core (its "Asynchronous Work" and "Instructions From Files"
+ * sections) and none of the GPT-5.6 core's own sections.
+ *
+ * Runs under bun (`bun gpt-6-family-preset-mock-loop.mjs`): the CLI is spawned
+ * from TypeScript source with the current runtime, so no tsx shim is needed.
+ */
+import { spawn } from "node:child_process";
+import { writeFileSync } from "node:fs";
+import { join } from "node:path";
+import { cliEntry, evidenceDir, guardRealAuth, installCleanupHooks, makeSandbox, repoRoot, track } from "./lib/common.mjs";
+import { startFakeModelServer } from "./lib/fake-model-server.mjs";
+import { hermeticEnv, writeMockModelsJson } from "./lib/mock-loop-support.mjs";
+
+const EVIDENCE_SLUG = "gpt-6-family-preset-mock-loop";
+const FINAL_MARKER = "SENPI-QA-GPT6-FAMILY-PRESET-APPLIED-7d2a";
+const GPT6_ONLY_SECTIONS = ["## Asynchronous Work", "## Instructions From Files"];
+const GPT56_ONLY_SECTIONS = ["## Manual QA Gate", "## Pragmatism & Scope"];
+const SCENARIOS = [
+	{ id: "gpt-6-sol", name: "GPT-6 Sol", thinking: "medium", expectedEffort: "medium" },
+	{ id: "gpt-6-luna", name: "GPT-6 Luna", thinking: "off", expectedEffort: "none" },
+];
+
+const checks = [];
+function check(label, condition) {
+	checks.push({ label, pass: !!condition });
+	console.log(`[${condition ? "PASS" : "FAIL"}] ${label}`);
+}
+
+function runCliWithCurrentRuntime(args, { env, cwd, timeoutMs }) {
+	return new Promise((resolve) => {
+		const root = repoRoot();
+		const child = track(spawn(process.execPath, [cliEntry(root), ...args], { cwd, env, stdio: ["pipe", "pipe", "pipe"] }));
+		let stdout = "";
+		let stderr = "";
+		child.stdout.on("data", (chunk) => {
+			stdout += chunk.toString();
+		});
+		child.stderr.on("data", (chunk) => {
+			stderr += chunk.toString();
+		});
+		const timer = setTimeout(() => {
+			child.kill("SIGKILL");
+			resolve({ code: null, stdout, stderr, timedOut: true });
+		}, timeoutMs);
+		child.on("close", (code) => {
+			clearTimeout(timer);
+			resolve({ code, stdout, stderr, timedOut: false });
+		});
+		child.stdin.end();
+	});
+}
+
+function systemTextOf(body) {
+	if (typeof body?.instructions === "string") return body.instructions;
+	const input = Array.isArray(body?.input) ? body.input : [];
+	const systemItems = input.filter((item) => item?.role === "developer" || item?.role === "system");
+	return systemItems
+		.map((item) => (typeof item.content === "string" ? item.content : (item.content ?? []).map((part) => part?.text ?? "").join("\n")))
+		.join("\n");
+}
+
+async function runScenario(scenario, evidence, authGuard) {
+	const sandbox = makeSandbox(`senpi-qa-${scenario.id}`);
+	const env = hermeticEnv(sandbox.env);
+	const server = await startFakeModelServer({ turns: [{ text: FINAL_MARKER }] });
+	try {
+		writeMockModelsJson(sandbox.agentDir, server, "openai-responses", { id: scenario.id, name: scenario.name, reasoning: true });
+
+		const result = await runCliWithCurrentRuntime(
+			["--print", "--provider", "openai", "--model", scenario.id, "--thinking", scenario.thinking, "Say hello"],
+			{ env, cwd: sandbox.cwd, timeoutMs: 60000 },
+		);
+
+		check(`${scenario.id}: --print exits 0`, result.code === 0);
+		check(`${scenario.id}: --print output contains the final marker`, result.stdout.includes(FINAL_MARKER));
+		check(`${scenario.id}: fake server captured exactly 1 request`, server.requests.length === 1);
+
+		const request = server.requests[0];
+		if (request) {
+			check(`${scenario.id}: request names ${scenario.id}`, request.body?.model === scenario.id);
+			check(
+				`${scenario.id}: request carries reasoning.effort=${scenario.expectedEffort}`,
+				request.body?.reasoning?.effort === scenario.expectedEffort,
+			);
+			const systemText = systemTextOf(request.body);
+			check(`${scenario.id}: request carries a developer/system message`, systemText.length > 0);
+			for (const section of GPT6_ONLY_SECTIONS) {
+				check(`${scenario.id}: system prompt carries the GPT-6 section ${section}`, systemText.includes(section));
+			}
+			for (const section of GPT56_ONLY_SECTIONS) {
+				check(`${scenario.id}: system prompt does NOT carry the GPT-5.6 section ${section}`, !systemText.includes(section));
+			}
+			check(`${scenario.id}: system prompt does not name Astra`, !/\bAstra\b/.test(systemText));
+			writeFileSync(join(evidence, `${scenario.id}-system-prompt.txt`), systemText);
+			writeFileSync(join(evidence, `${scenario.id}-request.json`), JSON.stringify(request.body, null, 2));
+		}
+
+		writeFileSync(
+			join(evidence, `${scenario.id}-stdout.txt`),
+			`exit=${result.code}\n---STDOUT---\n${result.stdout}\n---STDERR---\n${result.stderr}\n`,
+		);
+		check(`${scenario.id}: real auth store unchanged`, authGuard.assertUnchanged());
+	} finally {
+		await server.stop();
+		sandbox.cleanup();
+	}
+}
+
+async function main() {
+	const evidence = evidenceDir(EVIDENCE_SLUG);
+	installCleanupHooks();
+	const authGuard = guardRealAuth();
+
+	for (const scenario of SCENARIOS) {
+		await runScenario(scenario, evidence, authGuard);
+	}
+
+	const passed = checks.filter((entry) => entry.pass).length;
+	console.log(`\n${EVIDENCE_SLUG}: ${passed}/${checks.length} passed (evidence: ${evidence})`);
+	process.exitCode = passed === checks.length ? 0 : 1;
+}
+
+main().catch((error) => {
+	console.error(error);
+	process.exitCode = 1;
+});

packages/ai/CHANGELOG.md

@@ -6,10 +6,14 @@
 
 ### Added
 
+- GPT-6 Sol (`gpt-6-sol`) and GPT-6 Luna (`gpt-6-luna`) join the catalog on OpenAI, ChatGPT Subscription, Azure OpenAI, OpenCode Zen, OpenRouter, Venice and Vercel AI Gateway, with `-fast` Priority-tier variants on OpenAI and ChatGPT Subscription. Both carry their published prices (Sol \$2/\$10 per 1M tokens with \$0.20 cache reads, Luna \$0.10/\$0.50 with \$0.01 cache reads, both doubling input and 1.5x output past 272k), 128k output, text and image input, tool search and additional-tools support, and the documented effort ladder `none`/`low`/`medium`/`high`/`xhigh`/`max`. Project prompt budgets: Luna ships the full 922k input cap, Sol ships 400k, on every provider that lists the model.
+
 ### Changed
 
 ### Fixed
 
+- OpenRouter's passthrough rows for `openai/gpt-6-sol` and `openai/gpt-6-luna` (and their `-pro` / `:batch` siblings) shipped in 2026.9.22-4 with no effort ladder at all, so `xhigh` and `max` were not selectable there and the rows sat at the raw 922k window instead of the tier budget. They now carry the same GPT-6 ladder and budget as the first-party rows.
+
 ### Removed
 
 ## [2026.9.22-4] - 2026-09-22

packages/ai/changes.md

@@ -1,3 +1,24 @@
+## 2026-09-23 - GPT-6 Sol and GPT-6 Luna catalog rows
+
+### What changed
+
+- `packages/ai/scripts/generate-models.ts`: `gpt-6-sol` and `gpt-6-luna` join every OpenAI id table that already carried `gpt-6-astra` (tool search + additional tools on `openai` and `chatgpt-subscription`, short-context cap, long-context pricing tiers, `none` reasoning on the direct provider, Priority `-fast` variants on both first-party providers). Hand-added rows for both tiers land in the `openai` fallback list (used only when models.dev lacks the id) and in the `chatgpt-subscription` list (models.dev never carries the Codex backend), priced from the new `OPENAI_GPT_6_STANDARD_COSTS` table (Sol 2/10/0.2/2.5, Luna 0.1/0.5/0.01/0.125 per MTok) through `withOpenAiLongContextPricing`; `OPENAI_STANDARD_COSTS` merges the 5.6 and 6 tables for the direct-provider and Cloudflare price overrides. `supportsOpenAiXhigh` / `supportsOpenAiMax` and the former Astra-only final pass now key on `isGpt6FamilyId` (astra | sol | luna markers, prefixed or suffixed ids): `applyGpt6ContextWindow` stamps the tier budget on every provider row (Astra 600,000 unchanged, Sol 400,000, Luna 922,000 = the documented input cap) and `applyGpt6ThinkingLevels` stamps the documented ladder (`minimal: null`, low..max) and forces `off: null` for Astra only, since Sol and Luna document `none`.
+- `packages/ai/src/providers/data/` regenerated with `--strict`: +4 rows each on `openai.json` and `chatgpt-subscription.json` (base + `-fast`), +2 on `azure-openai-responses.json`, `opencode.json` (plus incidental upstream additions `claude-opus-5-5`, `grok-4.7`), `venice.json`, +4 on `vercel-ai-gateway.json`; the eight pre-existing `openrouter.json` GPT-6 Sol/Luna rows gain the ladder and the Sol budget. Incidental upstream drift: OpenRouter pricing/context refreshes (aion, deepseek, hy3, glm, `~latest` aliases), Vercel gemini metadata. No model id was removed.
+- Tests: `test/gpt-6-family-catalog.test.ts` (first-party rows, pricing tiers, ladder incl. `off`, tool metadata, `-fast` variants, map-less id inference, family-wide budget across every catalog); `test/openai-input-cap-catalog.test.ts` exempts the deliberate `gpt-6-sol @ 400,000` pairing from the documented-total check (a 1,050,000 Sol row would still fail) and pins the new direct-provider defaults.
+
+### Why
+
+OpenAI's GPT-6 family is Astra, Sol and Luna (developers.openai.com/api/docs/guides/latest-model). 2026.9.22-4 shipped Astra rows only; models.dev had since added Sol/Luna rows for openai, opencode, azure, openrouter and vercel, and 2026.9.22-4's regeneration had already pulled the OpenRouter passthrough rows without any effort ladder, so `xhigh` / `max` were unreachable there and the Codex backend could not select either tier at all. Budgets follow the user-stated defaults (Luna full window, Sol 400k) rather than the 272k short-context tier.
+
+### Why an extension could not handle it
+
+The catalog shards ship inside this package; nothing loaded at runtime can add a first-party row or change what the generator wrote.
+
+### Expected merge conflict zones
+
+- `packages/ai/scripts/generate-models.ts`: the OpenAI id tables near the top, `supportsOpenAiXhigh` / `supportsOpenAiMax`, the hand-added `openai` and `chatgpt-subscription` row lists, and the final metadata pass.
+- `packages/ai/src/providers/data/*.json` + `.manifest.json`: regenerate rather than merge.
+
 ## 2026-09-22 - Claude Opus 5.5 catalog rows and request compat
 
 ### What changed