fix(ai,coding-agent): correct Copilot context window discovery and await it during startup

packages/ai/CHANGELOG.md

@@ -5,6 +5,8 @@
 ### Fixed
 
 - Fixed CI type-check failures from test files referencing model IDs no longer returned by upstream `/models` endpoints. `models.generated.ts` is regenerated each CI run, so tests hardcoding stale IDs (`gpt-4o` on github-copilot, `qwen-3-235b-a22b-instruct-2507` on cerebras, `google/gemini-2.0-flash-001` on openrouter) failed `getModel`'s typed lookup. Swapped to currently-served siblings (`gpt-4.1`, `llama3.1-8b`, `google/gemini-2.5-flash`).
+- Fixed GitHub Copilot capability discovery silently failing on `api.individual.githubcopilot.com` (HTTP 421 Misdirected Request for some accounts): discovery now falls back through the canonical Copilot hosts (`api.githubcopilot.com`, `api.business.githubcopilot.com`, `api.enterprise.githubcopilot.com`) and uses the host that responded for the registered model's `baseUrl`, so 1M-context variants are actually surfaced.
+- Fixed Copilot Anthropic model `contextWindow` being set to `max_context_window_tokens` (total input+output budget) instead of `max_prompt_tokens` (real input cap), causing the compaction threshold to never trigger for 1M-context variants until the provider rejected the request. Both the capability cache and the registered model entry now use `max_prompt_tokens`, so compaction fires at the real input ceiling.
 - Fixed Anthropic adaptive-thinking rejection on `claude-opus-4-7` (and the `opus-4-6` family more generally): the request builder now consults a per-(provider, model) capability table and emits the adaptive `thinking.type=adaptive` + `output_config.effort` shape for any model that requires it, instead of the legacy `thinking.type=enabled` + `budget_tokens` shape ([#11](https://github.com/JuliusBrussee/caveman-code/issues/11))
 - Fixed Anthropic 1M context tier not being opted into for `claude-opus-4-5` on the direct Anthropic API and Bedrock: the provider now sends `anthropic-beta: context-1m-2025-08-07` for that family and the model registry overrides its `contextWindow` to 1_000_000 so the modeline / picker / compaction logic see the real ceiling. The opt-in is provider-scoped: the GitHub Copilot Anthropic relay rejects this beta header and exposes 1M-context Claude models as separate model ids instead, so the beta is not sent on Copilot. Discovery of those Copilot-only model ids is out of scope for this PR; a follow-up runtime-discovery layer will surface them ([#12](https://github.com/JuliusBrussee/caveman-code/issues/12))
 - Fixed bare `readline` import to use `node:readline` prefix for Deno compatibility ([#2885](https://github.com/badlogic/pi-mono/issues/2885) by [@milosv-vtool](https://github.com/milosv-vtool))

packages/ai/src/providers/anthropic-discovery.ts

@@ -129,19 +129,43 @@ interface CopilotModelEntry {
 	};
 }
 
+// Copilot's /models endpoint is served by a different proxy than chat. Some
+// accounts (individual/business/enterprise) get HTTP 421 "Misdirected
+// Request" on the host that serves chat. Try the configured host first,
+// then fall back to canonical hosts so discovery succeeds regardless of
+// which Copilot tier the user is on.
+const COPILOT_DISCOVERY_FALLBACK_HOSTS = [
+	"https://api.githubcopilot.com",
+	"https://api.individual.githubcopilot.com",
+	"https://api.business.githubcopilot.com",
+	"https://api.enterprise.githubcopilot.com",
+];
+
 async function discoverCopilot(baseUrl: string, apiKey: string, extraHeaders?: Record<string, string>): Promise<void> {
-	const res = await fetch(`${baseUrl}/models`, {
-		method: "GET",
-		headers: {
-			Authorization: `Bearer ${apiKey}`,
-			"User-Agent": "GitHubCopilotChat/0.35.0",
-			"Editor-Version": "vscode/1.107.0",
-			"Editor-Plugin-Version": "copilot-chat/0.35.0",
-			"Copilot-Integration-Id": "vscode-chat",
-			...extraHeaders,
-		},
-	});
-	if (!res.ok) return;
+	const headers = {
+		Authorization: `Bearer ${apiKey}`,
+		"User-Agent": "GitHubCopilotChat/0.35.0",
+		"Editor-Version": "vscode/1.107.0",
+		"Editor-Plugin-Version": "copilot-chat/0.35.0",
+		"Copilot-Integration-Id": "vscode-chat",
+		...extraHeaders,
+	};
+	const hosts = [baseUrl, ...COPILOT_DISCOVERY_FALLBACK_HOSTS.filter((h) => h !== baseUrl)];
+	let res: Response | undefined;
+	let workingHost = baseUrl;
+	for (const host of hosts) {
+		try {
+			const r = await fetch(`${host}/models`, { method: "GET", headers });
+			if (r.ok) {
+				res = r;
+				workingHost = host;
+				break;
+			}
+		} catch {
+			// try next host
+		}
+	}
+	if (!res) return;
 	const body = (await res.json()) as { data?: CopilotModelEntry[] };
 	const entries = body?.data ?? [];
 
@@ -159,10 +183,16 @@ async function discoverCopilot(baseUrl: string, apiKey: string, extraHeaders?: R
 		const efforts = sup.reasoning_effort ?? [];
 		const xhighEffort = efforts.includes("max") || efforts.includes("xhigh");
 
+		// Use max_prompt_tokens (real input cap) for contextWindow, not
+		// max_context_window_tokens (input+output+cache total). Compaction
+		// compares against the input budget and the provider rejects
+		// requests over max_prompt_tokens.
+		const promptCap = lim.max_prompt_tokens ?? lim.max_context_window_tokens;
+
 		setDiscoveredCapabilities("github-copilot", entry.id, {
 			thinkingSchema: adaptive ? "adaptive" : "legacy",
 			xhighEffort: xhighEffort || undefined,
-			contextWindow: lim.max_context_window_tokens,
+			contextWindow: promptCap,
 		});
 
 		// Build a registry entry so previously-unknown ids (e.g.
@@ -172,7 +202,7 @@ async function discoverCopilot(baseUrl: string, apiKey: string, extraHeaders?: R
 			name: entry.name || entry.id,
 			api: "anthropic-messages",
 			provider: "github-copilot",
-			baseUrl,
+			baseUrl: workingHost,
 			headers: {
 				"User-Agent": "GitHubCopilotChat/0.35.0",
 				"Editor-Version": "vscode/1.107.0",
@@ -182,7 +212,7 @@ async function discoverCopilot(baseUrl: string, apiKey: string, extraHeaders?: R
 			reasoning: adaptive,
 			input: sup.vision ? ["text", "image"] : ["text"],
 			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-			contextWindow: lim.max_context_window_tokens ?? 128000,
+			contextWindow: promptCap ?? 128000,
 			maxTokens: lim.max_output_tokens ?? 8192,
 		});
 	}

packages/coding-agent/CHANGELOG.md

@@ -4,6 +4,7 @@
 
 ### Fixed
 
+- Fixed initial model resolution falling back to a different model when the user's saved default isn't in the static registry (e.g. discovered-only Copilot ids). `createAgentSession` and the session-services factory now await Anthropic capability discovery before resolving the initial model if the saved default is missing from the snapshot, so the correct `contextWindow` is used from the first turn instead of silently degrading to a smaller-window sibling and over-compacting.
 - Added missing `@sinclair/typebox` runtime dependency. Fixes `ERR_MODULE_NOT_FOUND` on `caveman-code` startup when installed globally ([#6](https://github.com/JuliusBrussee/caveman-code/issues/6)).
 - Fixed broken extensions-migration links in the hooks-folder warning — now point to `JuliusBrussee/caveman-code` instead of the renamed-away `caveman-cli` repo ([#8](https://github.com/JuliusBrussee/caveman-code/issues/8)).
 - Self-update check now queries the correct GitHub repo (`caveman-code`) for releases.

packages/coding-agent/src/core/agent-session-services.ts

@@ -134,8 +134,15 @@ export async function createAgentSessionServices(
 	const authStorage = options.authStorage ?? AuthStorage.create(join(agentDir, "auth.json"));
 	const settingsManager = options.settingsManager ?? SettingsManager.create(cwd, agentDir);
 	const modelRegistry = options.modelRegistry ?? ModelRegistry.create(authStorage, join(agentDir, "models.json"));
-	// See sdk.ts for rationale; fire-and-forget Anthropic capability discovery.
-	void modelRegistry.discoverAnthropicCapabilities().catch(() => {});
+	// Anthropic capability discovery. If the user's saved default model isn't
+	// in the static registry, await discovery so resolution sees the freshly
+	// merged ids (e.g. github-copilot's claude-opus-4.7-1m-internal).
+	const discoveryPromise = modelRegistry.discoverAnthropicCapabilities().catch(() => {});
+	const savedDefaultProvider = settingsManager.getDefaultProvider();
+	const savedDefaultModelId = settingsManager.getDefaultModel();
+	if (savedDefaultProvider && savedDefaultModelId && !modelRegistry.find(savedDefaultProvider, savedDefaultModelId)) {
+		await discoveryPromise;
+	}
 	const resourceLoader = new DefaultResourceLoader({
 		...(options.resourceLoaderOptions ?? {}),
 		cwd,

packages/coding-agent/src/core/sdk.ts

@@ -179,13 +179,21 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 	const authStorage = options.authStorage ?? AuthStorage.create(authPath);
 	const modelRegistry = options.modelRegistry ?? ModelRegistry.create(authStorage, modelsPath);
 
+	const settingsManager = options.settingsManager ?? SettingsManager.create(cwd, agentDir);
+
 	// Kick off per-account capability discovery for Anthropic-API providers
 	// (e.g. GitHub Copilot's /models endpoint surfaces 1M variants like
-	// claude-opus-4.6-1m that aren't in the static registry). Fire-and-forget;
-	// onModelRegistryChange refreshes the local snapshot as results arrive.
-	void modelRegistry.discoverAnthropicCapabilities().catch(() => {});
+	// claude-opus-4.6-1m that aren't in the static registry).
+	// If the user's saved default model isn't in the static registry, we MUST
+	// await discovery before resolving — otherwise resolution falls back to a
+	// different model with a smaller context window. Otherwise fire-and-forget.
+	const discoveryPromise = modelRegistry.discoverAnthropicCapabilities().catch(() => {});
+	const savedDefaultProvider = settingsManager.getDefaultProvider();
+	const savedDefaultModelId = settingsManager.getDefaultModel();
+	if (savedDefaultProvider && savedDefaultModelId && !modelRegistry.find(savedDefaultProvider, savedDefaultModelId)) {
+		await discoveryPromise;
+	}
 
-	const settingsManager = options.settingsManager ?? SettingsManager.create(cwd, agentDir);
 	const sessionManager = options.sessionManager ?? SessionManager.create(cwd, getDefaultSessionDir(cwd, agentDir));
 
 	if (!resourceLoader) {