fix(channel): refuse attachments that can't fit the chat's model, and explain context overflows

.lane/memory/packages/channel/src/channel.ts/01M4XQKND0900E94CG3BXBEJP7-a-run-s-end-is-read-from-the.md

@@ -4,9 +4,9 @@ anchor: class Channel
 created: 2026-10-14T17:33:22Z
 norm: '1'
 sig: c905677fa4be33c1
-body_hash: 3b6eafcbe8faee26
-raw_hash: 67f0e882288fa8ce
-vouched: 2026-10-16T22:27:53Z
+body_hash: 3d7816658a739ed7
+raw_hash: 85f3d4da91b05dae
+vouched: 2026-10-16T23:12:54Z
 lines: 114-792
 ---
 

.lane/memory/packages/channel/src/channel.ts/01M4YFFTK7TWGKENJAD1W10HJJ-the-desktop-default-changes.md

@@ -4,9 +4,9 @@ anchor: class Channel
 created: 2026-10-15T00:30:42Z
 norm: '1'
 sig: c905677fa4be33c1
-body_hash: 3b6eafcbe8faee26
-raw_hash: 67f0e882288fa8ce
-vouched: 2026-10-16T22:27:53Z
+body_hash: 3d7816658a739ed7
+raw_hash: 85f3d4da91b05dae
+vouched: 2026-10-16T23:12:54Z
 lines: 114-792
 ---
 

.lane/memory/packages/channel/src/channel.ts/01M4YPJQB6KKEW1YTRD08YCYJY-only-gateways-admit-clients.md

@@ -4,9 +4,9 @@ anchor: class Channel
 created: 2026-10-15T02:34:37Z
 norm: '1'
 sig: c905677fa4be33c1
-body_hash: 3b6eafcbe8faee26
-raw_hash: 67f0e882288fa8ce
-vouched: 2026-10-16T22:27:53Z
+body_hash: 3d7816658a739ed7
+raw_hash: 85f3d4da91b05dae
+vouched: 2026-10-16T23:12:54Z
 lines: 114-794
 ---
 

.lane/memory/packages/channel/src/channel.ts/01M4ZH31DXZK0Y57C97DP84ZKX-handoffs-learns-queued-hando.md

@@ -4,9 +4,9 @@ anchor: class Channel
 created: 2026-10-15T10:17:55Z
 norm: '1'
 sig: c905677fa4be33c1
-body_hash: 3b6eafcbe8faee26
-raw_hash: 67f0e882288fa8ce
-vouched: 2026-10-16T22:27:53Z
+body_hash: 3d7816658a739ed7
+raw_hash: 85f3d4da91b05dae
+vouched: 2026-10-16T23:12:54Z
 lines: 121-817
 ---
 

.lane/memory/packages/channel/src/channel.ts/01M53CAGW2RVRKCYF5BCJRTZX2-a-fork-s-starting-lane-is-re.md

@@ -4,9 +4,9 @@ anchor: class Channel
 created: 2026-10-16T22:11:35Z
 norm: '1'
 sig: c905677fa4be33c1
-body_hash: 3b6eafcbe8faee26
-raw_hash: 67f0e882288fa8ce
-vouched: 2026-10-16T22:27:53Z
+body_hash: 3d7816658a739ed7
+raw_hash: 85f3d4da91b05dae
+vouched: 2026-10-16T23:12:54Z
 lines: 127-873
 ---
 

.lane/memory/packages/channel/src/channel.ts/01M53FTPSETKJ5KVVERXS3KAHK-overflow-errors-are-explaine.md

@@ -0,0 +1,12 @@
+---
+id: 01M53FTPSETKJ5KVVERXS3KAHK
+anchor: function failed
+created: 2026-10-16T23:12:51Z
+norm: '1'
+sig: ef0729315ff1212a
+body_hash: d35c24211c9f9acd
+raw_hash: a835bc42c69d89ed
+lines: 1033-1039
+---
+
+overflow errors are explained in the projection, not stored, so replies already stored by older channels get the recovery hint too

.lane/memory/packages/channel/src/room.ts/01M53FTPRW0WRSS93WVMFYZBKE-estimate-from-utf-8-bytes-no.md

@@ -0,0 +1,12 @@
+---
+id: 01M53FTPRW0WRSS93WVMFYZBKE
+anchor: function checkContext
+created: 2026-10-16T23:12:51Z
+norm: '1'
+sig: 9bf4c87f920895b7
+body_hash: 393fd86707463e3c
+raw_hash: 338145fbeb8c2ee8
+lines: 100-116
+---
+
+estimate from UTF-8 bytes, not chars; pi-ai's chars/4 undercounts non-Latin text by ~2x, which is how multilingual PDFs overflowed

packages/channel/src/channel.ts

@@ -1,4 +1,10 @@
-import { getSupportedThinkingLevels, type Model, type Models } from "@earendil-works/pi-ai";
+import {
+	type AssistantMessage,
+	getSupportedThinkingLevels,
+	isContextOverflow,
+	type Model,
+	type Models,
+} from "@earendil-works/pi-ai";
 import { BACKGROUND_CONTEXT as context } from "@earendil-works/chord/context";
 import {
 	AgentDoc,
@@ -589,6 +595,10 @@ export class Channel {
 		this.#author(request.author);
 		const conversation = await this.#conversation(request.chat);
 		room.checkAttachments(request);
+		// The chat's next run reads this message, so it must fit the model the chat has now.
+		const { model } = await conversation.agent(context);
+		const selected = model && this.#options.models.getModel(model.provider, model.modelId);
+		if (selected) room.checkContext(selected, request);
 		const entry = room.draft(request.author, request.text, Date.now(), request);
 		const submission = await conversation.submit({
 			type: "write",
@@ -616,7 +626,7 @@ export class Channel {
 		room.checkAttachments(request);
 		await this.#mayInvoke(request.author);
 		const conversation = await this.#conversation(request.chat);
-		await this.#select(conversation, request.model, request.effort);
+		await this.#select(conversation, request, request.model, request.effort);
 		const submission = await conversation.submit({
 			type: "input",
 			content: room.content(request.author, request.text, request),
@@ -637,7 +647,12 @@ export class Channel {
 	}
 
 	/** Model and effort apply to a whole run, never just its later turns. */
-	async #select(conversation: Conversation, requested: ModelRef | undefined, effort?: Effort) {
+	async #select(
+		conversation: Conversation,
+		attached: room.Attached,
+		requested: ModelRef | undefined,
+		effort?: Effort,
+	) {
 		const agent = await conversation.agent(context);
 		const current = agent.model;
 		// A chat's inherited model can lose its provider key; only an explicit request must keep it.
@@ -651,6 +666,8 @@ export class Channel {
 		if (!selected) {
 			throw new Error(`Unknown model ${model.provider}/${model.modelId}`);
 		}
+		// Before configuring, so a refused message leaves the chat's model as it was.
+		room.checkContext(selected, attached);
 		const levels = getSupportedThinkingLevels(selected);
 		if (effort !== undefined && effort !== "off" && !levels.includes(effort)) {
 			throw new Error(
@@ -1009,6 +1026,18 @@ function readable(error: string | undefined): string {
 	}
 }
 
+/**
+ * The transcript keeps whatever overflowed, so every later run in the chat fails the same way until
+ * a person removes it; channels from before attachments were checked against the model can hold one.
+ */
+function failed(message: AssistantMessage): string {
+	const error = readable(message.errorMessage);
+	if (!isContextOverflow(message)) return error;
+	const hint = "This chat no longer fits the model's context, so every run fails this way until "
+		+ "you delete the message with the large attachment or pick a model with a larger context.";
+	return `${error.replace(/\.?$/, ".")} ${hint}`;
+}
+
 function live(chat: ChatId, event: AgentEvent): Event[] {
 	if (event.type === "message_end" || event.type === "entry_appended") return events(event.entry);
 	if (event.type === "message_update") {
@@ -1047,7 +1076,7 @@ function events(entry: EntryRecord): Event[] {
 				at,
 				model,
 				text,
-				...(error ? { error: readable(message.errorMessage) } : {}),
+				...(error ? { error: failed(message) } : {}),
 				...(stopped ? { stopped } : {}),
 			});
 		}

packages/channel/src/room.ts

@@ -1,4 +1,11 @@
-import type { ImageContent, Message, TextContent, UserMessage } from "@earendil-works/pi-ai";
+import type {
+	Api,
+	ImageContent,
+	Message,
+	Model,
+	TextContent,
+	UserMessage,
+} from "@earendil-works/pi-ai";
 import {
 	defineEntry,
 	type EntryRecord,
@@ -80,6 +87,36 @@ export function checkAttachments({ images = [], files = [] }: Attached) {
 	if (size > MAX_BYTES) throw new Error("Attachments are too large; send smaller or fewer files");
 }
 
+/**
+ * Bytes, not characters: non-Latin text costs more tokens per character, and Keynote's
+ * multilingual license PDF measured about 2.8 bytes per token on Claude.
+ */
+const BYTES_PER_TOKEN = 3;
+
+/**
+ * A message whose files can't fit the model's context fails its run and every later one, since
+ * the transcript keeps it. The rest of the window holds the prompt, tools, history, and answer.
+ */
+export function checkContext(model: Model<Api>, { files = [] }: Attached) {
+	if (!files.length) return;
+	const encoder = new TextEncoder();
+	const tokens = files.reduce(
+		(sum, file) => sum + Math.ceil(encoder.encode(file.text).length / BYTES_PER_TOKEN),
+		0,
+	);
+	const limit = Math.floor(model.contextWindow * 0.75);
+	if (tokens <= limit) return;
+	const names = files.map((file) => file.name).join(", ");
+	const verb = files.length === 1 ? "is" : "are";
+	throw new Error(
+		`${names} ${verb} about ${thousands(tokens)} tokens; ${model.name} reads at most about `
+			+ `${thousands(limit)} tokens of attachments in one message. Attach a smaller file or an `
+			+ "excerpt, or pick a model with a larger context.",
+	);
+}
+
+const thousands = (tokens: number) => `${Math.round(tokens / 1000)}K`;
+
 export function draft(author: string, text: string, timestamp: number, attached?: Attached) {
 	return {
 		kind: MessageEntry.kind,