fix(chat): bump pi-ai and validate gateway model ids at startup (#244)

## Summary

Fixes the production crash `Error: Unknown AI Gateway model id:
openai/gpt-5.4-mini`
([JUNIOR-1Q](https://sentry.sentry.io/issues/JUNIOR-1Q)) and adds the
type/runtime guardrails that would have caught it before it shipped.

**Root cause.** #237 switched the default `AI_FAST_MODEL` to
`openai/gpt-5.4-mini`, but `@mariozechner/pi-ai@0.59.0` was pinned and
its generated `vercel-ai-gateway` registry predates that entry. Every
turn where `AI_FAST_MODEL` was unset hit `resolveGatewayModel` in
`packages/junior/src/chat/pi/client.ts`, found no match, and threw.
`openai/gpt-5.4-mini` is the correct gateway id per [Vercel's
catalog](https://vercel.com/ai-gateway/models/gpt-5.4-mini) — 0.68.1 is
just the first pi-ai release that knows about it.

**Fix:**

- Bump `@mariozechner/pi-ai` and `@mariozechner/pi-agent-core` from
`0.59.0` → `0.68.1`. Catalog now includes `openai/gpt-5.4-mini`,
`gpt-5.4-nano`, `gpt-5.3-chat`, etc.
- `Agent.replaceMessages()` was removed in pi-agent-core 0.68; assign
`agent.state.messages` directly instead (the setter copies the top-level
array). One call site in `respond.ts`.

**Guardrails so this doesn't regress silently again** (answering
@dcramer's review question on whether types could have caught this):

- **Compile-time check on default literals.** `config.ts` wraps
`DEFAULT_MODEL_ID` / `DEFAULT_FAST_MODEL_ID` in
`getModel("vercel-ai-gateway", ...)` calls. `getModel`'s second generic
is `TModelId extends keyof (typeof MODELS)[TProvider]`, so a stale
literal becomes a `tsc` error at the call site. Verified by temporarily
swapping in a bogus id — `tsc` emitted `TS2345: Argument of type
'"openai/gpt-5.4-mini-bogus"' is not assignable to parameter of type
'"openai/gpt-5.4" | ... | "zai/glm-5v-turbo"'`.
- **Runtime check on env overrides.** New `toGatewayModelId()` validates
every `AI_MODEL` / `AI_FAST_MODEL` / `AI_VISION_MODEL` against pi-ai's
registry at `readBotConfig` time. A typo in env now throws "Unknown AI
Gateway model id: …" at startup, not mid-turn. New regression test in
`chat-config.test.ts` pins this behavior; existing tests moved from
placeholder ids (`anthropic/custom-model`) to real catalog ids
(`anthropic/claude-opus-4.6`, `anthropic/claude-haiku-4.5`).

We can't go further (i.e. type `BotConfig.fastModelId: GatewayModelId`)
because pi-ai doesn't export `MODELS` from its package entry — the
literal union is not importable as a nameable type. Defaults are
compile-checked and env is runtime-checked at startup, which covers both
observed regression paths.

Fixes JUNIOR-1Q.

## Review & Testing Checklist for Human

- [ ] Skim the pi-ai / pi-agent-core 0.60 → 0.68 changelog for any other
API breaks we haven't hit yet — I verified via full typecheck + `pnpm
test` but didn't walk every intermediate release.
- [ ] After merge, confirm JUNIOR-1Q stops firing in production with the
new release.
- [ ] If any deployed env has a typo in `AI_MODEL` / `AI_FAST_MODEL` /
`AI_VISION_MODEL`, the process will now fail to boot instead of failing
on the first turn. Worth eyeballing current prod env before merge.

### Notes

- `pnpm typecheck`, `pnpm lint`, and `pnpm test` (707 tests) all pass
locally.
- No `BotConfig` signature changes —
`modelId`/`fastModelId`/`visionModelId` stay typed as `string` since we
can't name a `GatewayModelId` union without pi-ai re-exporting `MODELS`.

Link to Devin session:
https://app.devin.ai/sessions/bf10e4f407dd4265a27a7d7f463eb1c3

---------

Co-authored-by: Devin AI <devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: GPT-5 <devin@cognition.ai>

Author: devin-ai-integration[bot]

packages/junior/package.json

@@ -39,8 +39,8 @@
     "@chat-adapter/state-memory": "4.26.0",
     "@chat-adapter/state-redis": "4.26.0",
     "@logtape/logtape": "^2.0.5",
-    "@mariozechner/pi-agent-core": "0.59.0",
-    "@mariozechner/pi-ai": "0.59.0",
+    "@mariozechner/pi-agent-core": "0.68.1",
+    "@mariozechner/pi-ai": "0.68.1",
     "@modelcontextprotocol/sdk": "1.29.0",
     "@sinclair/typebox": "^0.34.49",
     "@slack/web-api": "^7.15.1",

packages/junior/src/chat/config.ts

@@ -1,4 +1,6 @@
+import { getModel } from "@mariozechner/pi-ai";
 import { toOptionalTrimmed } from "@/chat/optional-string";
+import { resolveGatewayModel } from "@/chat/pi/client";
 
 const MIN_AGENT_TURN_TIMEOUT_MS = 10 * 1000;
 const DEFAULT_AGENT_TURN_TIMEOUT_MS = 12 * 60 * 1000;
@@ -101,16 +103,33 @@ function parseLoadingMessages(rawValue: string | undefined): string[] {
   });
 }
 
+// Compile-time assertion: `getModel`'s second generic is constrained to
+// `keyof (typeof MODELS)[TProvider]`, so a stale default becomes a tsc error.
+const DEFAULT_MODEL_ID = getModel("vercel-ai-gateway", "openai/gpt-5.4").id;
+const DEFAULT_FAST_MODEL_ID = getModel(
+  "vercel-ai-gateway",
+  "openai/gpt-5.4-mini",
+).id;
+
+function validateGatewayModelId(raw: string | undefined): string | undefined {
+  const trimmed = toOptionalTrimmed(raw);
+  if (trimmed === undefined) return undefined;
+  resolveGatewayModel(trimmed);
+  return trimmed;
+}
+
 function readBotConfig(env: NodeJS.ProcessEnv): BotConfig {
   const functionMaxDurationSeconds = resolveFunctionMaxDurationSeconds(env);
   const maxTurnTimeoutMs = resolveMaxTurnTimeoutMs(functionMaxDurationSeconds);
 
   return {
     userName: env.JUNIOR_BOT_NAME ?? "junior",
-    modelId: env.AI_MODEL ?? "openai/gpt-5.4",
-    fastModelId: env.AI_FAST_MODEL ?? env.AI_MODEL ?? "openai/gpt-5.4-mini",
+    modelId: validateGatewayModelId(env.AI_MODEL) ?? DEFAULT_MODEL_ID,
+    fastModelId:
+      validateGatewayModelId(env.AI_FAST_MODEL ?? env.AI_MODEL) ??
+      DEFAULT_FAST_MODEL_ID,
     loadingMessages: parseLoadingMessages(env.JUNIOR_LOADING_MESSAGES),
-    visionModelId: toOptionalTrimmed(env.AI_VISION_MODEL),
+    visionModelId: validateGatewayModelId(env.AI_VISION_MODEL),
     turnTimeoutMs: parseAgentTurnTimeoutMs(
       env.AGENT_TURN_TIMEOUT_MS,
       maxTurnTimeoutMs,

packages/junior/src/chat/pi/client.ts

@@ -129,9 +129,15 @@ function parseJsonCandidate(text: string): unknown {
   }
 }
 
+/**
+ * Look up a gateway model by id. Throws `Unknown AI Gateway model id: …` if
+ * the id is not in pi-ai's registry — callers at the config boundary can use
+ * this to fail fast at startup instead of mid-turn.
+ */
 export function resolveGatewayModel(modelId: string): Model<any> {
-  const models = getModels(GATEWAY_PROVIDER);
-  const matched = models.find((model: Model<any>) => model.id === modelId);
+  const matched = getModels(GATEWAY_PROVIDER).find(
+    (model: Model<any>) => model.id === modelId,
+  );
   if (!matched) {
     throw new Error(`Unknown AI Gateway model id: ${modelId}`);
   }

packages/junior/src/chat/respond.ts

@@ -885,7 +885,7 @@ export async function generateAssistantReply(
 
     try {
       if (resumedFromCheckpoint) {
-        agent.replaceMessages(existingCheckpoint!.piMessages);
+        agent.state.messages = existingCheckpoint!.piMessages;
       }
       beforeMessageCount = agent.state.messages.length;
 

packages/junior/tests/unit/config/chat-config.test.ts

@@ -14,20 +14,20 @@ describe("chat config", () => {
   });
 
   it("uses AI_MODEL for fastModelId when AI_FAST_MODEL is unset", async () => {
-    process.env.AI_MODEL = "anthropic/custom-model";
+    process.env.AI_MODEL = "anthropic/claude-opus-4.6";
     delete process.env.AI_FAST_MODEL;
 
     const { botConfig } = await loadConfig();
-    expect(botConfig.modelId).toBe("anthropic/custom-model");
-    expect(botConfig.fastModelId).toBe("anthropic/custom-model");
+    expect(botConfig.modelId).toBe("anthropic/claude-opus-4.6");
+    expect(botConfig.fastModelId).toBe("anthropic/claude-opus-4.6");
   });
 
   it("prefers AI_FAST_MODEL over AI_MODEL for fastModelId", async () => {
-    process.env.AI_MODEL = "anthropic/custom-model";
-    process.env.AI_FAST_MODEL = "anthropic/custom-fast-model";
+    process.env.AI_MODEL = "anthropic/claude-opus-4.6";
+    process.env.AI_FAST_MODEL = "anthropic/claude-haiku-4.5";
 
     const { botConfig } = await loadConfig();
-    expect(botConfig.fastModelId).toBe("anthropic/custom-fast-model");
+    expect(botConfig.fastModelId).toBe("anthropic/claude-haiku-4.5");
   });
 
   it("uses the default fast model when AI_MODEL and AI_FAST_MODEL are unset", async () => {
@@ -46,31 +46,37 @@ describe("chat config", () => {
   });
 
   it("ignores AI_LIGHT_MODEL and keeps using AI_FAST_MODEL", async () => {
-    process.env.AI_MODEL = "anthropic/custom-model";
-    process.env.AI_FAST_MODEL = "anthropic/custom-fast-model";
+    process.env.AI_MODEL = "anthropic/claude-opus-4.6";
+    process.env.AI_FAST_MODEL = "anthropic/claude-haiku-4.5";
     process.env.AI_LIGHT_MODEL = "openai/gpt-5.4-mini";
 
     const { botConfig } = await loadConfig();
-    expect(botConfig.fastModelId).toBe("anthropic/custom-fast-model");
+    expect(botConfig.fastModelId).toBe("anthropic/claude-haiku-4.5");
   });
 
   it("leaves visionModelId unset when AI_VISION_MODEL is absent", async () => {
-    process.env.AI_MODEL = "anthropic/custom-model";
+    process.env.AI_MODEL = "anthropic/claude-opus-4.6";
     delete process.env.AI_VISION_MODEL;
 
     const { botConfig } = await loadConfig();
     expect(botConfig.visionModelId).toBeUndefined();
   });
 
   it("uses AI_VISION_MODEL without falling back to AI_MODEL", async () => {
-    process.env.AI_MODEL = "anthropic/custom-model";
+    process.env.AI_MODEL = "anthropic/claude-opus-4.6";
     process.env.AI_VISION_MODEL = "openai/gpt-5.4";
 
     const { botConfig } = await loadConfig();
-    expect(botConfig.modelId).toBe("anthropic/custom-model");
+    expect(botConfig.modelId).toBe("anthropic/claude-opus-4.6");
     expect(botConfig.visionModelId).toBe("openai/gpt-5.4");
   });
 
+  it("throws at config load when AI_MODEL is not a registered gateway model id", async () => {
+    process.env.AI_MODEL = "openai/gpt-definitely-not-real";
+
+    await expect(loadConfig()).rejects.toThrow(/Unknown AI Gateway model id/);
+  });
+
   it("uses the default assistant loading messages when unset", async () => {
     delete process.env.JUNIOR_LOADING_MESSAGES;
     const { botConfig } = await loadConfig();