Merge PR #4882: fix(catalog): collapse Devin GLM-5.2 variants so free 200K model works when quota is exhausted (@oldschoola)

# Conflicts: # packages/catalog/src/models.json
2026-07-20 22:50:16 +02:00
parent b3cb3e6463 55d922fabf
commit 024497995c
6 changed files with 173 additions and 8 deletions
@@ -174,6 +174,13 @@

 - Updated cost and token configurations for various models across providers
 - Renamed several models for consistency (e.g., MiniMax M3, Gemma 4 31B, Qwen variants)
+### Added
+
+- Added static fallback seed for Devin's `swe-1-7` model so it is bundled even when catalog generation runs without a Devin session token.
+
+### Fixed
+
+- Collapsed Devin's six GLM-5.2 variants into two logical entries (`glm-5-2` for 200K free, `glm-5-2-1m` for 1M paid). The 200K entry routes every thinking effort to the free `glm-5-2` wire UID — never to the quota-gated `glm-5-2-max` or `glm-5-2-none` — so GLM-5.2 works even when the weekly usage quota is exhausted.

 ## [16.3.12] - 2026-07-08

@@ -17,6 +17,7 @@ import { getGitLabDuoModels } from "@oh-my-pi/pi-ai/providers/gitlab-duo";
 import { $env } from "@oh-my-pi/pi-utils";
 import { ANTIGRAVITY_PRIMARY_ENDPOINT, fetchAntigravityDiscoveryModels } from "../src/discovery/antigravity";
 import { fetchCodexModels } from "../src/discovery/codex";
+import { DEVIN_STATIC_FALLBACK_MODELS } from "../src/discovery/devin";
 import { buildGitLabDuoWorkflowFallbackModel } from "../src/discovery/gitlab-duo-workflow";
 import { createModelManager } from "../src/model-manager";
 import prevModelsJson from "../src/models.json" with { type: "json" };
@@ -541,6 +542,13 @@ async function generateModels() {
 	if (!authoritativeCatalogProviders.has("gitlab-duo-agent")) {
 		allModels.push(buildGitLabDuoWorkflowFallbackModel());
 	}
+	// Seed Devin fallback models so newly released free models (e.g. `swe-1-7`)
+	// are bundled even when catalog generation runs without a Devin session
+	// token. Devin is `dynamicModelsAuthoritative: true`, so live discovery
+	// replaces these at runtime when a key is present.
+	if (!authoritativeCatalogProviders.has("devin")) {
+		allModels.push(...DEVIN_STATIC_FALLBACK_MODELS);
+	}
 	// Seed Fireworks "Fast" serving-path variants (`<id>-fast`). Fast routers are
 	// not enumerated by the serverless control-plane list, so discovery never
 	// surfaces them; the seed projects each base entry into a fast variant.
@@ -149,3 +149,27 @@ function normalizeDevinModels(
 	}
 	return [...byId.values()].sort((a, b) => a.id.localeCompare(b.id));
 }
+
+/**
+ * Static fallback Devin models for catalog generation without a live API key.
+ *
+ * Devin discovery requires an authenticated session token; when catalog
+ * generation runs without one, these seeds ensure new free models (like
+ * `swe-1-7`) are still bundled. Live discovery is authoritative — when it
+ * succeeds, it replaces these seeds entirely (stale entries are pruned).
+ */
+export const DEVIN_STATIC_FALLBACK_MODELS: readonly ModelSpec<"devin-agent">[] = [
+	{
+		id: "swe-1-7",
+		name: "SWE-1.7",
+		api: "devin-agent",
+		provider: "devin",
+		baseUrl: DEVIN_DEFAULT_BASE_URL,
+		reasoning: true,
+		input: ["text"],
+		supportsTools: true,
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+		contextWindow: 262_000,
+		maxTokens: DEFAULT_MAX_TOKENS,
+	},
+];
@@ -532,6 +532,39 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Minimal, Effort.Low, Effort.Medium, Effort.High],
 		),
+		// GLM-5.2 200K — only the base wire UID `glm-5-2` is free on Devin's
+		// Coding Plan (verified via streamDevin: `glm-5-2-none` and `glm-5-2-max`
+		// both return "weekly usage quota exhausted" while `glm-5-2` streams
+		// successfully).  Route every effort to `glm-5-2` so the collapsed entry
+		// is always free; include the paid 200K variants as members so they are
+		// hidden from the model list.  The 1M-context variants stay as separate
+		// paid entries (collapsed below).
+		{
+			id: "glm-5-2",
+			name: "GLM-5.2",
+			members: ["glm-5-2", "glm-5-2-none", "glm-5-2-max"],
+			routing: {
+				[Effort.High]: "glm-5-2",
+				[Effort.XHigh]: "glm-5-2",
+			},
+			thinking: {
+				mode: "effort",
+				efforts: [Effort.High, Effort.XHigh],
+				requiresEffort: true,
+			},
+		},
+		// GLM-5.2 1M — paid variants that consume weekly quota.  Collapse the
+		// three 1M-context variants into one entry with proper effort routing.
+		devinTierFamily(
+			"glm-5-2-1m",
+			"GLM-5.2 1M",
+			{
+				off: "glm-5-2-none-1m",
+				high: "glm-5-2-1m",
+				xhigh: "glm-5-2-max-1m",
+			},
+			[Effort.High, Effort.XHigh],
+		),
 	],
 };

@@ -791,3 +791,75 @@ describe("antigravity discovery collapsing", () => {
 		expect(models?.[0]?.baseUrl).toBe(ANTIGRAVITY_PRIMARY_ENDPOINT);
 	});
 });
+
+describe("Devin GLM-5.2 collapse", () => {
+	function devinMemberSpec(id: string, overrides: Partial<ModelSpec<"devin-agent">> = {}): ModelSpec<"devin-agent"> {
+		return {
+			id,
+			name: id,
+			api: "devin-agent",
+			provider: "devin",
+			baseUrl: "https://server.codeium.com",
+			reasoning: true,
+			input: ["text"],
+			supportsTools: true,
+			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+			contextWindow: 200_000,
+			maxTokens: 64_000,
+			...overrides,
+		};
+	}
+
+	it("collapses the three 200K GLM-5.2 variants into one logical entry routing all efforts to the free glm-5-2 wire UID", () => {
+		const out = collapseEffortVariants(
+			[
+				devinMemberSpec("glm-5-2"),
+				devinMemberSpec("glm-5-2-max"),
+				devinMemberSpec("glm-5-2-none", { reasoning: false }),
+			],
+			DEVIN_VARIANT_COLLAPSE_TABLE,
+		);
+
+		expect(out).toHaveLength(1);
+		const spec = out[0];
+		expect(spec?.id).toBe("glm-5-2");
+		expect(spec?.thinking?.effortRouting).toEqual({
+			high: "glm-5-2",
+			xhigh: "glm-5-2",
+		});
+	});
+
+	it("routes every effort to glm-5-2 (never to the quota-gated glm-5-2-max or glm-5-2-none)", () => {
+		const out = collapseEffortVariants(
+			[devinMemberSpec("glm-5-2"), devinMemberSpec("glm-5-2-max")],
+			DEVIN_VARIANT_COLLAPSE_TABLE,
+		);
+
+		const spec = out[0];
+		const routing = spec?.thinking?.effortRouting ?? {};
+		for (const wire of Object.values(routing)) {
+			expect(wire).toBe("glm-5-2");
+		}
+	});
+
+	it("collapses the three 1M GLM-5.2 variants into one paid entry with proper effort routing", () => {
+		const out = collapseEffortVariants(
+			[
+				devinMemberSpec("glm-5-2-1m", { contextWindow: 1_000_000 }),
+				devinMemberSpec("glm-5-2-max-1m", { contextWindow: 1_000_000 }),
+				devinMemberSpec("glm-5-2-none-1m", { contextWindow: 1_000_000, reasoning: false }),
+			],
+			DEVIN_VARIANT_COLLAPSE_TABLE,
+		);
+
+		expect(out).toHaveLength(1);
+		const spec = out[0];
+		expect(spec?.id).toBe("glm-5-2-1m");
+		expect(spec?.contextWindow).toBe(1_000_000);
+		expect(spec?.thinking?.effortRouting).toEqual({
+			off: "glm-5-2-none-1m",
+			high: "glm-5-2-1m",
+			xhigh: "glm-5-2-max-1m",
+		});
+	});
+});
@@ -34,16 +34,37 @@ describe("AgentSession advisor descriptor thinking level", () => {
 		sharedDir = TempDir.createSync("@pi-advisor-devin-thinking-shared-");
 		authStorage = await AuthStorage.create(path.join(sharedDir.path(), "testauth.db"));
 		authStorage.setRuntimeApiKey("anthropic", "test-key");
-		// Seeding a runtime API key exposes the bundled Devin catalog for
-		// `resolveAdvisorRoleSelection` / `getAvailable()` without any live
-		// network discovery.
-		authStorage.setRuntimeApiKey("devin", "test-key");
 		modelRegistry = new ModelRegistry(authStorage);
 		const anthropic = getBundledModel("anthropic", "claude-sonnet-4-5");
-		const devin = getBundledModel("devin", "glm-5-2");
 		if (!anthropic) throw new Error("Expected bundled anthropic/claude-sonnet-4-5 to exist");
-		if (!devin) throw new Error("Expected bundled devin/glm-5-2 to exist");
 		anthropicModel = anthropic;
+
+		// Register a synthetic `devin-agent` provider with a reasoning model
+		// that has NO `thinking` metadata. This is the exact catalog shape that
+		// triggered #4579: `reasoning: true` with no controllable effort surface.
+		// Using a synthetic model avoids brittleness from upstream catalog drift
+		// (e.g. variant-collapse adding `thinking.effortRouting` to bundled Devin
+		// models).
+		modelRegistry.registerProvider("devin-advisor-test", {
+			api: "devin-agent",
+			apiKey: "test-key",
+			baseUrl: "https://test.example.com",
+			models: [
+				{
+					id: "no-thinking",
+					name: "Test No-Thinking",
+					api: "devin-agent",
+					reasoning: true,
+					input: ["text"],
+					supportsTools: true,
+					cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+					contextWindow: 200_000,
+					maxTokens: 64_000,
+				},
+			],
+		});
+		const devin = modelRegistry.find("devin-advisor-test", "no-thinking");
+		if (!devin) throw new Error("Expected synthetic devin-advisor-test/no-thinking to register");
 		devinModel = devin;
 	});

@@ -88,8 +109,8 @@ describe("AgentSession advisor descriptor thinking level", () => {

 	it("Devin advisor with no configured thinking suffix boots without an unsupported-effort throw", () => {
 		// Confirm the catalog shape that triggered the bug: `reasoning: true` with
-		// no controllable `thinking.efforts`. If this drifts upstream the
-		// regression's assumptions no longer hold.
+		// no controllable `thinking.efforts`. The synthetic model in `beforeAll`
+		// guarantees this shape regardless of upstream catalog drift.
 		expect(devinModel.reasoning).toBe(true);
 		expect(devinModel.thinking).toBeUndefined();