fix(compaction): respect Model.contextWindow; size Codebase Auto correctly

User report: CLI compacting at 80–100k tokens on Codebase Auto routes
that have 200k of headroom. Two stacked bugs:

1. CompactionEngine ignored the resolved Model's contextWindow field
   entirely. It read modelId (a string) and regex-matched against a
   built-in table — IDs like "MiniMax-M2.7" or proxy-routed names
   missed every pattern and silently fell back to 128k.
2. buildProxiedConfig synthesized the Codebase Auto model by spreading
   a Groq llama-3.3-70b template, which carries contextWindow: 128_000.
   Even after fix #1, the synthesized model was lying about its true
   window.

Now: CompactionEngine takes an explicit contextWindow option that
agent.ts passes from model.contextWindow. buildProxiedConfig sets
contextWindow: 200_000 on the Codebase Auto path and uses a
guessContextWindow helper for explicit-provider synthesis so Gemini
routes get 1M, GPT-5 gets 400k, Claude/proxy-default gets 200k.

Threshold math: old = 128_000 * 0.75 = 96_000 (matches the report).
New on Codebase Auto = 200_000 * 0.75 = 150_000.

Author: halfaipg

package-lock.json

@@ -1,6 +1,6 @@
 {
 	"name": "codebase-cli",
-	"version": "2.0.0-pre.57",
+	"version": "2.0.0-pre.58",
 	"description": "Codebase CLI — a TypeScript coding agent on the pi-mono runtime. OAuth-aware, any LLM provider, single install.",
 	"keywords": [
 		"ai",

package.json

@@ -175,7 +175,11 @@ export function createAgent(opts: CreateAgentOptions = {}): AgentBundle {
 		smartModel: glueModels.smart,
 		getApiKey,
 	});
-	const compaction = new CompactionEngine({ glue, modelId: model.id });
+	// Pass model.contextWindow explicitly so proxy-synthesized models
+	// (Codebase Auto, custom in-house IDs) get the real window instead of
+	// the regex-based fallback in tokens.ts, which would otherwise lock
+	// them at 128k and trigger compaction at ~96k on a 200k-context route.
+	const compaction = new CompactionEngine({ glue, modelId: model.id, contextWindow: model.contextWindow });
 	const compactionMonitor = new CompactionMonitor();
 	const sessions = new SessionStore({ cwd });
 	const resumed = opts.resume ? sessions.load(model.id) : null;

src/agent/agent.ts

@@ -202,6 +202,13 @@ function buildProxiedConfig(
 			// `model.id` only, so this cast is safe.
 			provider: "codebase" as Model<string>["provider"],
 			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+			// Codebase Auto routes to large-context models on the backend
+			// (Claude Sonnet 4 default, open-weight alternates also 128k+).
+			// The Groq llama template's 128k contextWindow was leaking
+			// through and triggering compaction at ~96k tokens on routes
+			// that have 200k of headroom. Set explicitly so the compaction
+			// engine reads the right value.
+			contextWindow: 200_000,
 		};
 		return { model, apiKey: accessToken, source: "proxy" };
 	}
@@ -230,10 +237,31 @@ function buildProxiedConfig(
 		baseUrl: proxyBase,
 		provider: explicitProvider as Model<string>["provider"],
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+		// Same reasoning as Codebase Auto: synthesized proxy models point
+		// at large-context backends. Without overriding contextWindow the
+		// Groq llama template's 128k leaks through and the compaction
+		// engine triggers ~30% earlier than it should.
+		contextWindow: guessContextWindow(modelId, 200_000),
 	};
 	return { model: synthesized, apiKey: accessToken, source: "proxy" };
 }
 
+/**
+ * Best-effort context-window guess for synthesized proxy models whose
+ * IDs pi-ai doesn't know natively. Pattern-matches a few common families
+ * so we don't gimp a 1M-context Gemini model to 200k; everything else
+ * defaults to the supplied fallback (200k, matching Claude Sonnet 4 /
+ * GPT-5 / most open-weight large models the proxy routes to).
+ */
+function guessContextWindow(modelId: string, fallback: number): number {
+	const id = modelId.toLowerCase();
+	if (id.startsWith("gemini-")) return 1_000_000;
+	if (id.startsWith("gpt-5")) return 400_000;
+	if (id.startsWith("claude-")) return 200_000;
+	if (id.startsWith("llama-3-3-70b") || id.startsWith("llama-3.3-70b")) return 128_000;
+	return fallback;
+}
+
 /**
  * OpenAI Chat-Completions compatible custom endpoint. Used by
  * MiniMax in-house, Qwen in-house, Groq custom URLs, Ollama, vLLM,

src/agent/config.ts

@@ -25,25 +25,37 @@ Be concrete. Reference paths, function names, commit subjects when you have them
 export interface CompactionEngineOptions {
 	glue: GlueClient;
 	modelId: string;
+	/**
+	 * Authoritative context-window size from the resolved model. When set,
+	 * this is used verbatim and the modelId-based regex fallback is
+	 * bypassed entirely. Codebase Auto + proxy-synthesized models need
+	 * this — their IDs (e.g. "MiniMax-M2.7") don't match any built-in
+	 * regex and would otherwise fall back to 128k, triggering compaction
+	 * at ~96k against a model that has 200k of headroom.
+	 */
+	contextWindow?: number;
 	thresholdRatio?: number;
 	keepRecent?: number;
 }
 
 export class CompactionEngine {
 	private readonly glue: GlueClient;
 	private readonly modelId: string;
+	private readonly explicitContextWindow: number | undefined;
 	private readonly thresholdRatio: number;
 	private readonly keepRecent: number;
 
 	constructor(options: CompactionEngineOptions) {
 		this.glue = options.glue;
 		this.modelId = options.modelId;
+		this.explicitContextWindow = options.contextWindow;
 		this.thresholdRatio = options.thresholdRatio ?? DEFAULT_THRESHOLD;
 		this.keepRecent = options.keepRecent ?? DEFAULT_KEEP_RECENT;
 	}
 
 	threshold(): number {
-		return Math.floor(contextWindow(this.modelId) * this.thresholdRatio);
+		const window = this.explicitContextWindow ?? contextWindow(this.modelId);
+		return Math.floor(window * this.thresholdRatio);
 	}
 
 	needsCompaction(messages: AgentMessage[]): boolean {