fix(pi-devin): parse Devin CLI 1M cost_summary correctly

Devin catalog emits '$4 / 1M Input · $0.2 / 1M Cached input · $20 / 1M Output'
but the vendored parseCost only matched '/MTok In/Out', leaving output and
cached prices at 0 so session cost totals for devin models were wrong.
Accept 1M Input/Output/Cached input labels and use the real cached-input
price for cacheRead instead of the 10% input heuristic.

.github/workflows/test.yml

@@ -1,22 +0,0 @@
-name: Tests
-on:
-  push:
-  pull_request:
-permissions:
-  contents: read
-jobs:
-  test:
-    strategy:
-      fail-fast: false
-      matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
-        node: [22, 24]
-    runs-on: ${{ matrix.os }}
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-node@v4
-        with:
-          node-version: ${{ matrix.node }}
-      - run: npm install --ignore-scripts --no-audit --no-fund
-      - run: npm run typecheck
-      - run: npm test

.gitignore

@@ -1,7 +0,0 @@
-node_modules/
-dist/
-package-lock.json
-.DS_Store
-*.log
-.env
-.reference/

README.md

@@ -16,9 +16,9 @@ Those models are available through the local Devin CLI. This package uses that C
 
 ## Requirements
 
-- Pi Coding Agent 0.86+
+- Pi Coding Agent 0.86+ (managed fork validated against 0.87.1)
 - A signed-in [Devin CLI](https://docs.devin.ai/cli) (`devin auth status`)
-- Node 22.19+ (required by Pi 0.86)
+- Node 22.19+
 
 The CLI binary is resolved in this order:
 
@@ -27,8 +27,6 @@ The CLI binary is resolved in this order:
 3. Devin.app's bundled `devin` binary
 4. `which devin`
 
-On Windows, discovery checks the CLI installer and Devin Desktop locations, then uses `where.exe` to search PATH. Native executables and `.cmd`/`.bat` wrappers are supported, including paths with spaces. `DEVIN_CLI` takes precedence on every platform.
-
 ## Install
 
 From git:
@@ -37,16 +35,12 @@ From git:
 pi install git:github.com/kashyab12/pi-devin
 ```
 
-From npm:
+After npm publish:
 
 ```bash
 pi install npm:pi-devin
 ```
 
-Version 0.2.0 requires Pi 0.86+ and Node 22.19+. It restores system instructions and tools on the current Pi transcript format and sends system instructions through Devin's dedicated prompt field. Version 0.1.2 predates Pi 0.86 support.
-
-`pi-devin-local` is a separate npm package maintained in the [mizorewww/pi-devin fork](https://github.com/mizorewww/pi-devin). Changes merged here do not update that package. Install only one: both packages register the `devin` provider, so their registrations can overwrite each other.
-
 Local checkout:
 
 ```bash
@@ -71,8 +65,6 @@ Commands:
 - `/devin-status` — CLI path, version, auth
 - `/devin-refresh` — reload `devin models list --format json`
 
-The model catalog is cached for six hours in `$XDG_CACHE_HOME/pi-devin/models.json` (or `~/.cache/pi-devin/models.json`). A fresh cache avoids the CLI call at startup; an older cache remains available while it refreshes in the background. Set `PI_OFFLINE=1` to skip automatic catalog refreshes. `/devin-refresh` still requests a refresh explicitly.
-
 ## What this is / is not
 
 | This package | Not this package |
@@ -89,3 +81,7 @@ This is a standard Pi package (`keywords: ["pi-package"]` + `pi.extensions`). Af
 ## License
 
 MIT. Unofficial. Not affiliated with Cognition.
+
+## Managed fork
+
+Based on npm pi-devin 0.2.1. Retains native signature/response replay, corrected wire fields, bounded frames, context clamping, multi-window catalog variants and abort cleanup. Uses the upstream six-hour XDG catalog cache. Tests and typecheck are maintained in the parent app tests directory.

package.json

@@ -28,10 +28,6 @@
     "AGENTS.md",
     "LICENSE"
   ],
-  "scripts": {
-    "test": "node test-runner.mjs",
-    "typecheck": "tsc --noEmit"
-  },
   "peerDependencies": {
     "@earendil-works/pi-ai": ">=0.86.0",
     "@earendil-works/pi-coding-agent": ">=0.86.0"

src/context-estimate.ts

@@ -0,0 +1,118 @@
+import { getSystemMessageText, type Api, type TranscriptContext, type Message, type Model, type Usage } from "@earendil-works/pi-ai";
+
+/**
+ * Local port of pi-ai 0.87.1's `utils/estimate.js` + `clampMaxTokensToContext`.
+ * Extensions can only resolve the pi-ai compat entry, so the upstream helpers
+ * are reproduced here to keep the maxTokens context clamp identical.
+ */
+const CHARS_PER_TOKEN = 4;
+const ESTIMATED_IMAGE_CHARS = 4800;
+const CONTEXT_SAFETY_TOKENS = 4096;
+const MIN_MAX_TOKENS = 1;
+
+function calculateContextTokens(usage: Usage): number {
+  return usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;
+}
+
+function safeJsonStringify(value: unknown): string {
+  try {
+    return JSON.stringify(value) ?? "undefined";
+  } catch {
+    return "[unserializable]";
+  }
+}
+
+function estimateTextAndImageContentChars(content: Message["content"]): number {
+  if (typeof content === "string") return content.length;
+  let chars = 0;
+  for (const block of content) chars += block.type === "text" ? block.text.length : ESTIMATED_IMAGE_CHARS;
+  return chars;
+}
+
+function estimateTextTokens(text: string): number {
+  return Math.ceil(text.length / CHARS_PER_TOKEN);
+}
+
+function estimateTextAndImageContentTokens(content: Message["content"]): number {
+  return Math.ceil(estimateTextAndImageContentChars(content) / CHARS_PER_TOKEN);
+}
+
+function estimateMessageTokens(message: Message): number {
+  if (message.role === "system") {
+    return estimateTextTokens(getSystemMessageText(message)) +
+      estimateToolsTokens(message.toolsAdded) + estimateToolsTokens(message.toolsRemoved);
+  }
+  if (message.role === "user" || message.role === "toolResult") {
+    return estimateTextAndImageContentTokens(message.content);
+  }
+  let chars = 0;
+  for (const block of message.content) {
+    if (block.type === "text") chars += block.text.length;
+    else if (block.type === "thinking") chars += block.thinking.length;
+    else chars += block.name.length + safeJsonStringify(block.arguments).length;
+  }
+  return Math.ceil(chars / CHARS_PER_TOKEN);
+}
+
+interface UsageInfo {
+  usage: Usage;
+  index: number;
+}
+
+function getLastAssistantUsageInfo(messages: Message[]): UsageInfo | undefined {
+  let latestPrefixTimestamp = Number.NEGATIVE_INFINITY;
+  let usageInfo: UsageInfo | undefined;
+  for (let i = 0; i < messages.length; i++) {
+    const message = messages[i];
+    if (message.role === "assistant") {
+      const usageAppliesToPrefix = message.timestamp >= latestPrefixTimestamp;
+      if (
+        usageAppliesToPrefix &&
+        message.stopReason !== "aborted" &&
+        message.stopReason !== "error" &&
+        calculateContextTokens(message.usage) > 0
+      ) {
+        usageInfo = { usage: message.usage, index: i };
+      }
+    }
+    latestPrefixTimestamp = Math.max(latestPrefixTimestamp, message.timestamp);
+  }
+  return usageInfo;
+}
+
+interface ContextEstimate {
+  tokens: number;
+  usageTokens: number;
+  trailingTokens: number;
+  lastUsageIndex: number | null;
+}
+
+function estimateMessages(messages: Message[]): ContextEstimate {
+  const usageInfo = getLastAssistantUsageInfo(messages);
+  if (usageInfo) {
+    const usageTokens = calculateContextTokens(usageInfo.usage);
+    let trailingTokens = 0;
+    for (let i = usageInfo.index + 1; i < messages.length; i++) {
+      trailingTokens += estimateMessageTokens(messages[i]);
+    }
+    return { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex: usageInfo.index };
+  }
+  let tokens = 0;
+  for (const message of messages) tokens += estimateMessageTokens(message);
+  return { tokens, usageTokens: 0, trailingTokens: tokens, lastUsageIndex: null };
+}
+
+function estimateToolsTokens(tools: readonly unknown[] | undefined): number {
+  if (!tools || tools.length === 0) return 0;
+  return estimateTextTokens(safeJsonStringify(tools));
+}
+
+export function estimateContextTokens(context: TranscriptContext): ContextEstimate {
+  return estimateMessages(context.messages);
+}
+
+export function clampMaxTokensToContext(model: Model<Api>, context: TranscriptContext, maxTokens: number): number {
+  if (model.contextWindow <= 0) return Math.max(MIN_MAX_TOKENS, maxTokens);
+  const available = model.contextWindow - estimateContextTokens(context).tokens - CONTEXT_SAFETY_TOKENS;
+  return Math.min(maxTokens, Math.max(MIN_MAX_TOKENS, available));
+}

src/context-map.ts

@@ -1,11 +1,6 @@
 import {
-  collapseSystemMessages,
-  getCurrentSystemPrompt,
-  getCurrentTools,
-  withoutInitialSystemMessage,
-  type Message,
-  type Tool,
-  type TranscriptContext,
+  collapseSystemMessages, getCurrentSystemPrompt, getCurrentTools, withoutInitialSystemMessage,
+  type Api, type TranscriptContext, type Message, type Model, type Tool,
 } from "@earendil-works/pi-ai";
 
 export interface ContentPart {
@@ -17,9 +12,14 @@ export interface ContentPart {
 
 export interface ChatHistoryItem {
   role: "user" | "assistant" | "tool";
-  content: string | ContentPart[];
+  content: ContentPart[];
   tool_call_id?: string;
   tool_calls?: Array<{ id: string; name: string; arguments: string }>;
+  tool_result_is_error?: boolean;
+  thinking?: string;
+  signature?: string;
+  thinking_redacted?: boolean;
+  message_id?: string;
 }
 
 export interface ToolDef {
@@ -29,15 +29,14 @@ export interface ToolDef {
 }
 
 export interface MappedChat {
-  /** Sent in GetChatMessageRequest.prompt, separately from conversation history. */
   systemPrompt?: string;
   messages: ChatHistoryItem[];
   tools: ToolDef[];
 }
 
-function userContent(content: Message["content"]): string | ContentPart[] {
-  if (typeof content === "string") return content;
-  if (!Array.isArray(content)) return "";
+function contentParts(content: Message["content"] | string): ContentPart[] {
+  if (typeof content === "string") return content ? [{ type: "text", text: content }] : [];
+  if (!Array.isArray(content)) return [];
   const parts: ContentPart[] = [];
   for (const part of content) {
     if (!part || typeof part !== "object") continue;
@@ -49,22 +48,39 @@ function userContent(content: Message["content"]): string | ContentPart[] {
   return parts;
 }
 
-export function mapContextToChat(context: TranscriptContext): MappedChat {
-  // Fold later system updates into the single prompt sent separately to Devin.
+export function mapContextToChat(context: TranscriptContext, model: Model<Api>): MappedChat {
   const transcript = collapseSystemMessages(context);
-  const systemPrompt = getCurrentSystemPrompt(transcript.messages);
   const messages: ChatHistoryItem[] = [];
+
   for (const message of withoutInitialSystemMessage(transcript.messages)) {
     if (message.role === "user") {
-      messages.push({ role: "user", content: userContent(message.content) });
+      const content = contentParts(message.content);
+      if (content.length === 0) continue;
+      messages.push({ role: "user", content });
       continue;
     }
     if (message.role === "assistant") {
-      const texts: string[] = [];
+      const content: ContentPart[] = [];
       const toolCalls: Array<{ id: string; name: string; arguments: string }> = [];
+      let thinking = "";
+      let signature: string | undefined;
+      let thinkingRedacted = false;
+      // Only messages produced by this provider/model carry wire-replayable
+      // thinking signatures and response ids.
+      const native =
+        message.api === model.api &&
+        message.provider === model.provider &&
+        message.model === model.id;
       for (const part of message.content) {
-        if (part.type === "text") texts.push(part.text);
-        if (part.type === "toolCall") {
+        if (part.type === "text") {
+          content.push({ type: "text", text: part.text });
+        } else if (part.type === "thinking") {
+          thinking += part.thinking;
+          if (part.redacted) thinkingRedacted = true;
+          if (native && !signature && part.thinkingSignature) {
+            signature = part.thinkingSignature;
+          }
+        } else if (part.type === "toolCall") {
           toolCalls.push({
             id: part.id,
             name: part.name,
@@ -72,25 +88,24 @@ export function mapContextToChat(context: TranscriptContext): MappedChat {
           });
         }
       }
+      if (content.length === 0 && !thinking && !signature && toolCalls.length === 0) continue;
       messages.push({
         role: "assistant",
-        content: texts.join("\n"),
+        content,
         tool_calls: toolCalls.length > 0 ? toolCalls : undefined,
+        thinking: thinking || undefined,
+        signature,
+        thinking_redacted: thinkingRedacted || undefined,
+        message_id: native ? message.responseId : undefined,
       });
       continue;
     }
     if (message.role === "toolResult") {
-      const text =
-        typeof message.content === "string"
-          ? message.content
-          : message.content
-              .filter((part) => part.type === "text")
-              .map((part) => part.text)
-              .join("\n");
       messages.push({
         role: "tool",
-        content: text,
+        content: contentParts(message.content),
         tool_call_id: message.toolCallId,
+        tool_result_is_error: message.isError || undefined,
       });
     }
   }
@@ -101,5 +116,7 @@ export function mapContextToChat(context: TranscriptContext): MappedChat {
     parameters: tool.parameters,
   }));
 
-  return { systemPrompt: systemPrompt || undefined, messages, tools };
+  const systemPrompt = getCurrentSystemPrompt(transcript.messages) || undefined;
+
+  return { systemPrompt, messages, tools };
 }

src/jwt.ts

@@ -7,7 +7,7 @@ export interface MintedUserJwt {
   expiresAt: number;
 }
 
-function anySignal(signals: AbortSignal[]): AbortSignal {
+export function anySignal(signals: AbortSignal[]): AbortSignal {
   const builtin = (AbortSignal as unknown as { any?: (s: AbortSignal[]) => AbortSignal }).any;
   if (typeof builtin === "function") return builtin(signals);
   const controller = new AbortController();

src/metadata.ts

@@ -11,7 +11,9 @@ import {
  * Luna) by client ide name. With ide="windsurf" GetChatMessage rejects them
  * with "This model is only in Devin Local."; with ide="devin-desktop" the
  * server serves them and the response header echoes the exact model
- * (verified: "GPT-5.6 Sol High Thinking" for gpt-5-6-sol-high, 2026-08-29).
+ * (verified: gpt-5-6-sol-high, 2026-09-15). The version string is still
+ * gated against a known Windsurf/Devin Desktop release: prefer the installed
+ * Devin.app product.json, then a current desktop release.
  */
 const FALLBACK_WINDSURF_VERSION = "3.6.27";
 const PRODUCT_JSON =

src/models.ts

@@ -30,14 +30,18 @@ const THINKING_ORDER = ["off", "minimal", "low", "medium", "high", "xhigh", "max
 function parseCost(summary?: string): ProviderModelConfig["cost"] {
   const empty = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
   if (!summary) return empty;
-  const input = summary.match(/\$([0-9.]+)\s*\/\s*MTok In/i);
-  const output = summary.match(/\$([0-9.]+)\s*\/\s*MTok Out/i);
+  // Devin CLI emits e.g. "$10 / 1M Input · $0.25 / 1M Cached input · $50 / 1M Output";
+  // older catalogs used "/MTok In" style labels. Match the unit token loosely.
+  const input = summary.match(/\$([0-9.]+)\s*\/\s*1?M(?:Tok)?\s+In(?:put)?/i);
+  const output = summary.match(/\$([0-9.]+)\s*\/\s*1?M(?:Tok)?\s+Out(?:put)?/i);
+  const cached = summary.match(/\$([0-9.]+)\s*\/\s*1?M(?:Tok)?\s+Cached\s+in(?:put)?/i);
   const inCost = input ? Number(input[1]) : 0;
   const outCost = output ? Number(output[1]) : 0;
+  const cacheRead = cached ? Number(cached[1]) : Number((inCost * 0.1).toFixed(4));
   return {
     input: inCost,
     output: outCost,
-    cacheRead: Number((inCost * 0.1).toFixed(4)),
+    cacheRead,
     cacheWrite: Number((inCost * 1.25).toFixed(4)),
   };
 }
@@ -65,8 +69,13 @@ function variantKey(uid: string): string | null {
     "max",
     "minimal",
   ];
+  // Catalog uids are not uniform: some families use `MODEL_GPT_5_2_HIGH`-style
+  // uppercase/underscore ids and long-context twins append `-1m` after the
+  // thinking suffix. Normalize case, separators, and the context suffix before
+  // matching so every variant lands on its real thinking level.
+  const normalized = uid.toLowerCase().replace(/_/g, "-").replace(/-1m$/, "");
   for (const suffix of suffixes) {
-    if (uid === suffix || uid.endsWith(`-${suffix}`)) return suffix;
+    if (normalized === suffix || normalized.endsWith(`-${suffix}`)) return suffix;
   }
   return null;
 }
@@ -92,39 +101,82 @@ function preferredDefault(map: ThinkingLevelMap): string | undefined {
   return undefined;
 }
 
+function contextLabel(tokens: number | undefined): string {
+  if (!tokens || tokens <= 0) return "";
+  if (tokens % 1_000_000 === 0) return `${tokens / 1_000_000}M`;
+  return `${Math.round(tokens / 1_000)}K`;
+}
+
+function variantsToModel(
+  family: DevinFamily,
+  variants: DevinVariant[],
+  name?: string,
+): ProviderModelConfig | null {
+  // Seed every level with null so Pi treats levels the catalog lacks as
+  // explicitly unsupported instead of selectable-but-silently-remapped.
+  const thinkingLevelMap: ThinkingLevelMap = {
+    off: null,
+    minimal: null,
+    low: null,
+    medium: null,
+    high: null,
+    xhigh: null,
+    max: null,
+  };
+  for (const variant of variants) {
+    const level = thinkingFromSuffix(variantKey(variant.model_uid));
+    if (level && thinkingLevelMap[level] == null) {
+      thinkingLevelMap[level] = variant.model_uid;
+    }
+  }
+
+  const defaultUid = preferredDefault(thinkingLevelMap) ?? variants[0]?.model_uid ?? family.family_uid;
+  const sample = variants.find((variant) => variant.model_uid === defaultUid) ?? variants[0];
+  if (!sample) return null;
+
+  const mappedLevels = THINKING_ORDER.filter((level) => typeof thinkingLevelMap[level] === "string");
+  const reasoning = mappedLevels.length > 1;
+
+  return {
+    id: defaultUid,
+    name: name ?? (family.family_label || family.slug || defaultUid),
+    reasoning,
+    thinkingLevelMap: reasoning ? thinkingLevelMap : undefined,
+    input: ["text", "image"],
+    cost: parseCost(sample.cost_summary),
+    contextWindow: sample.max_context_tokens ?? 256_000,
+    maxTokens: sample.max_output_tokens ?? 128_000,
+  };
+}
+
 function familyToModels(family: DevinFamily): ProviderModelConfig[] {
   const usable = family.variants.filter((variant) => {
     const key = variantKey(variant.model_uid);
     return !key || (!key.includes("priority") && !key.includes("fast") && key !== "thinking-1m");
   });
   const source = usable.length > 0 ? usable : family.variants;
-  const thinkingLevelMap: ThinkingLevelMap = {};
+  // Group by advertised context window so long-context twins (`*-1m`) remain
+  // selectable instead of collapsing into the primary variant group. The
+  // smallest-context group keeps the plain family label for stable model ids.
+  const groups = new Map<number, DevinVariant[]>();
   for (const variant of source) {
-    const level = thinkingFromSuffix(variantKey(variant.model_uid));
-    if (level && thinkingLevelMap[level] === undefined) {
-      thinkingLevelMap[level] = variant.model_uid;
-    }
+    const context = variant.max_context_tokens ?? 0;
+    const group = groups.get(context) ?? [];
+    group.push(variant);
+    groups.set(context, group);
   }
-
-  const defaultUid = preferredDefault(thinkingLevelMap) ?? source[0]?.model_uid ?? family.family_uid;
-  const sample = source.find((variant) => variant.model_uid === defaultUid) ?? source[0];
-  if (!sample) return [];
-
-  const mappedLevels = THINKING_ORDER.filter((level) => typeof thinkingLevelMap[level] === "string");
-  const reasoning = mappedLevels.length > 1;
-
-  return [
-    {
-      id: defaultUid,
-      name: family.family_label || family.slug || defaultUid,
-      reasoning,
-      thinkingLevelMap: reasoning ? thinkingLevelMap : undefined,
-      input: ["text", "image"],
-      cost: parseCost(sample.cost_summary),
-      contextWindow: sample.max_context_tokens ?? 256_000,
-      maxTokens: sample.max_output_tokens ?? 128_000,
-    },
-  ];
+  const ordered = [...groups.values()].sort(
+    (a, b) => (a[0]?.max_context_tokens ?? 0) - (b[0]?.max_context_tokens ?? 0),
+  );
+  const models: ProviderModelConfig[] = [];
+  ordered.forEach((variants, index) => {
+    const context = variants[0]?.max_context_tokens;
+    const name =
+      index === 0 ? undefined : `${family.family_label || family.slug} (${contextLabel(context)})`;
+    const model = variantsToModel(family, variants, name);
+    if (model) models.push(model);
+  });
+  return models;
 }
 
 export const FALLBACK_MODELS: ProviderModelConfig[] = [
@@ -133,6 +185,8 @@ export const FALLBACK_MODELS: ProviderModelConfig[] = [
     name: "Claude Opus 5",
     reasoning: true,
     thinkingLevelMap: {
+      off: null,
+      minimal: null,
       low: "claude-opus-5-low",
       medium: "claude-opus-5-medium",
       high: "claude-opus-5-high",
@@ -149,6 +203,8 @@ export const FALLBACK_MODELS: ProviderModelConfig[] = [
     name: "Claude Fable 5",
     reasoning: true,
     thinkingLevelMap: {
+      off: null,
+      minimal: null,
       low: "claude-5-fable-low",
       medium: "claude-5-fable-medium",
       high: "claude-5-fable-high",
@@ -166,6 +222,7 @@ export const FALLBACK_MODELS: ProviderModelConfig[] = [
     reasoning: true,
     thinkingLevelMap: {
       off: "gpt-5-6-sol-none",
+      minimal: null,
       low: "gpt-5-6-sol-low",
       medium: "gpt-5-6-sol-medium",
       high: "gpt-5-6-sol-high",
@@ -181,6 +238,15 @@ export const FALLBACK_MODELS: ProviderModelConfig[] = [
     id: "swe-1-7",
     name: "SWE-1.7",
     reasoning: true,
+    thinkingLevelMap: {
+      off: null,
+      minimal: null,
+      low: null,
+      medium: "swe-1-7",
+      high: "swe-1-7",
+      xhigh: null,
+      max: null,
+    },
     input: ["text", "image"],
     cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
     contextWindow: 262_000,
@@ -211,11 +277,12 @@ export function resolveModelUid(
   thinkingLevelMap: ThinkingLevelMap | undefined,
   reasoning?: string,
 ): string {
-  if (reasoning && thinkingLevelMap) {
-    const mapped = thinkingLevelMap[reasoning as keyof ThinkingLevelMap];
-    if (typeof mapped === "string") return mapped;
-  }
   if (thinkingLevelMap) {
+    // Pi encodes thinkingLevel "off" as an absent reasoning option; resolve it
+    // against the explicit "off" mapping (the catalog's *-none variant) before
+    // falling back to the family's preferred default.
+    const mapped = thinkingLevelMap[(reasoning ?? "off") as keyof ThinkingLevelMap];
+    if (typeof mapped === "string") return mapped;
     const fallback = preferredDefault(thinkingLevelMap);
     if (fallback) return fallback;
   }

src/stream.ts

@@ -4,15 +4,15 @@ import {
   type Api,
   type AssistantMessage,
   type AssistantMessageEventStream,
+  type TranscriptContext,
   type Model,
   type SimpleStreamOptions,
-  type TranscriptContext,
   calculateCost,
   createAssistantMessageEventStream,
-  parseStreamingJson,
 } from "@earendil-works/pi-ai";
-import { mapContextToChat, type ChatHistoryItem, type ContentPart, type ToolDef } from "./context-map.js";
-import { getCachedUserJwt } from "./jwt.js";
+import { clampMaxTokensToContext } from "./context-estimate.js";
+import { mapContextToChat, type ChatHistoryItem, type ToolDef } from "./context-map.js";
+import { anySignal, getCachedUserJwt } from "./jwt.js";
 import { buildMetadata } from "./metadata.js";
 import { resolveModelUid } from "./models.js";
 import {
@@ -24,32 +24,40 @@ import {
   iterFields,
 } from "./wire.js";
 
-const SOURCE_BY_ROLE: Record<ChatHistoryItem["role"], number> = {
+const SOURCE_BY_ROLE: Record<string, number> = {
   user: 1,
   assistant: 2,
   tool: 4,
 };
 
+/** Connect frame payloads larger than this indicate a corrupt length prefix. */
+const MAX_CONNECT_FRAME_PAYLOAD = 16 * 1024 * 1024;
+/**
+ * Recovery heuristic for opaque Devin `invalid_argument`/`internal error`
+ * trailers: compactable message history this large is likely the cause, so the
+ * failure is reported in a shape Pi recognizes as context overflow.
+ */
+const LARGE_HISTORY_RECOVERY_BYTES = 512 * 1024;
+
 export type CloudChatEvent =
   | { kind: "text"; text: string }
   | { kind: "reasoning"; text: string }
+  | { kind: "signature"; text: string }
+  | { kind: "thinking_redacted" }
+  | { kind: "message_id"; id: string }
+  | { kind: "actual_model"; uid: string }
   | { kind: "tool_call_start"; id: string; name: string }
   | { kind: "tool_call_args"; argsDelta: string; id?: string }
-  | { kind: "finish"; reason: "stop" | "tool_calls" | "length" | "content_filter" }
+  | { kind: "finish"; reason: "stop" | "tool_calls" | "length" | "content_filter" | "error" }
   | {
       kind: "usage";
-      promptTokens?: number;
-      completionTokens?: number;
-      totalTokens?: number;
-      cachedInputTokens?: number;
-      cacheCreationInputTokens?: number;
+      input?: number;
+      output?: number;
+      cacheRead?: number;
+      cacheWrite?: number;
+      modelUid?: string;
     };
 
-function normalizeContent(content: string | ContentPart[]): ContentPart[] {
-  if (typeof content === "string") return [{ type: "text", text: content }];
-  return content;
-}
-
 function encodeImageData(img: { mimeType?: string; base64Data?: string }): Buffer {
   return Buffer.concat([
     encodeString(1, img.base64Data ?? ""),
@@ -61,39 +69,41 @@ function encodeChatToolCall(tc: { id: string; name: string; arguments: string })
   return Buffer.concat([encodeString(1, tc.id), encodeString(2, tc.name), encodeString(3, tc.arguments)]);
 }
 
-function encodeChatMessagePrompt(
-  content: ContentPart[],
-  source: number,
-  opts?: { toolCallId?: string; toolCalls?: Array<{ id: string; name: string; arguments: string }> },
-): Buffer {
-  const text = content
+export function encodeChatMessagePrompt(item: ChatHistoryItem, messageId: string): Buffer {
+  const text = item.content
     .filter((part) => part.type === "text")
     .map((part) => part.text ?? "")
     .join("\n");
   const parts: Buffer[] = [
-    encodeVarintField(2, source),
+    encodeString(1, messageId),
+    encodeVarintField(2, SOURCE_BY_ROLE[item.role] ?? 1),
     encodeString(3, text),
-    encodeVarintField(4, Math.max(1, Math.floor(text.length / 4))),
-    encodeVarintField(5, 1),
   ];
-  if (opts?.toolCallId) parts.push(encodeString(7, opts.toolCallId));
-  for (const tc of opts?.toolCalls ?? []) parts.push(encodeMessage(6, encodeChatToolCall(tc)));
-  for (const img of content.filter((part) => part.type === "image")) {
+  for (const tc of item.tool_calls ?? []) parts.push(encodeMessage(6, encodeChatToolCall(tc)));
+  if (item.tool_call_id) parts.push(encodeString(7, item.tool_call_id));
+  if (item.tool_result_is_error) parts.push(encodeVarintField(9, 1));
+  for (const img of item.content.filter((part) => part.type === "image")) {
     parts.push(encodeMessage(10, encodeImageData(img)));
   }
+  if (item.thinking) parts.push(encodeString(11, item.thinking));
+  if (item.signature) parts.push(encodeString(12, item.signature));
+  if (item.thinking_redacted) parts.push(encodeVarintField(13, 1));
   return Buffer.concat(parts);
 }
 
-function encodeCompletionConfiguration(maxOutputTokens?: number): Buffer {
+function encodeCompletionConfiguration(args: {
+  maxTokens: number;
+  temperature: number;
+  topP: number;
+  topK: number;
+}): Buffer {
   return Buffer.concat([
     encodeVarintField(1, 1),
-    encodeVarintField(2, 64_000),
-    encodeVarintField(3, maxOutputTokens ?? 128_000),
-    encodeFixed64Field(5, 0.7),
-    encodeFixed64Field(6, 0.95),
-    encodeVarintField(7, 50),
-    encodeFixed64Field(8, 1.0),
-    encodeFixed64Field(11, 1.0),
+    encodeVarintField(2, args.maxTokens),
+    encodeVarintField(3, 400),
+    encodeFixed64Field(5, args.temperature),
+    encodeVarintField(7, args.topK),
+    encodeFixed64Field(8, args.topP),
   ]);
 }
 
@@ -106,7 +116,7 @@ function encodeToolDef(tool: ToolDef): Buffer {
   ]);
 }
 
-function buildGetChatMessageRequest(args: {
+export interface WireRequestArgs {
   apiKey: string;
   userJwt: string;
   modelUid: string;
@@ -118,8 +128,13 @@ function buildGetChatMessageRequest(args: {
   sessionId: string;
   requestId: bigint;
   triggerId: string;
-  maxOutputTokens?: number;
-}): Buffer {
+  maxTokens: number;
+  temperature: number;
+  topP: number;
+  topK: number;
+}
+
+export function buildGetChatMessageRequest(args: WireRequestArgs): { body: Buffer; historyBytes: number } {
   const metadata = buildMetadata({
     apiKey: args.apiKey,
     userJwt: args.userJwt,
@@ -127,36 +142,46 @@ function buildGetChatMessageRequest(args: {
     requestId: args.requestId,
     triggerId: args.triggerId,
   });
-  const prompts = args.messages.map((message) =>
-    encodeMessage(
-      3,
-      encodeChatMessagePrompt(normalizeContent(message.content), SOURCE_BY_ROLE[message.role], {
-        toolCallId: message.role === "tool" ? message.tool_call_id : undefined,
-        toolCalls: message.role === "assistant" ? message.tool_calls : undefined,
-      }),
-    ),
-  );
-  return Buffer.concat([
+  const promptFields: Buffer[] = [];
+  args.messages.forEach((message, index) => {
+    const messageId = message.message_id ?? `${args.cascadeId}-${index}`;
+    promptFields.push(encodeMessage(3, encodeChatMessagePrompt(message, messageId)));
+  });
+  const history = Buffer.concat(promptFields);
+  const body = Buffer.concat([
     encodeMessage(1, metadata),
-    ...(args.systemPrompt ? [encodeString(2, args.systemPrompt)] : []),
-    ...prompts,
+    args.systemPrompt ? encodeString(2, args.systemPrompt) : Buffer.alloc(0),
+    history,
     encodeVarintField(7, 5),
-    encodeMessage(8, encodeCompletionConfiguration(args.maxOutputTokens)),
+    encodeMessage(8, encodeCompletionConfiguration(args)),
     ...(args.tools ?? []).map((tool) => encodeMessage(10, encodeToolDef(tool))),
     encodeString(16, args.cascadeId),
+    encodeString(17, args.promptId),
+    encodeVarintField(20, 1),
     encodeString(21, args.modelUid),
-    encodeString(22, args.promptId),
   ]);
+  return { body, historyBytes: history.length };
 }
 
-function* decodeChatFrame(proto: Buffer): Generator<CloudChatEvent> {
+export function* decodeChatFrame(proto: Buffer): Generator<CloudChatEvent> {
   for (const field of iterFields(proto)) {
-    if (field.num === 3 && field.wire === 2 && Buffer.isBuffer(field.value)) {
+    if (field.num === 1 && field.wire === 2 && Buffer.isBuffer(field.value)) {
+      const id = field.value.toString("utf8");
+      if (id) yield { kind: "message_id", id };
+    } else if (field.num === 3 && field.wire === 2 && Buffer.isBuffer(field.value)) {
       const text = field.value.toString("utf8");
       if (text) yield { kind: "text", text };
     } else if (field.num === 9 && field.wire === 2 && Buffer.isBuffer(field.value)) {
       const text = field.value.toString("utf8");
       if (text) yield { kind: "reasoning", text };
+    } else if (field.num === 10 && field.wire === 2 && Buffer.isBuffer(field.value)) {
+      const signature = field.value.toString("utf8");
+      if (signature) yield { kind: "signature", text: signature };
+    } else if (field.num === 11 && field.wire === 0 && field.value) {
+      yield { kind: "thinking_redacted" };
+    } else if (field.num === 23 && field.wire === 2 && Buffer.isBuffer(field.value)) {
+      const uid = field.value.toString("utf8");
+      if (uid) yield { kind: "actual_model", uid };
     } else if (field.num === 6 && field.wire === 2 && Buffer.isBuffer(field.value)) {
       let id: string | undefined;
       let name: string | undefined;
@@ -173,59 +198,77 @@ function* decodeChatFrame(proto: Buffer): Generator<CloudChatEvent> {
       if (argsDelta !== undefined) yield { kind: "tool_call_args", argsDelta, ...(id ? { id } : {}) };
     } else if (field.num === 5 && field.wire === 0) {
       const value = Number(field.value);
-      let reason: Extract<CloudChatEvent, { kind: "finish" }>['reason'] = "stop";
+      let reason: Extract<CloudChatEvent, { kind: "finish" }>["reason"] = "stop";
       if (value === 10) reason = "tool_calls";
+      else if (value === 1 || value === 3 || value === 5) reason = "length";
       else if (value === 11) reason = "content_filter";
-      else if (value === 1 || value === 3) reason = "length";
+      else if (value === 12 || value === 13) reason = "error";
       yield { kind: "finish", reason };
-    } else if (field.num === 28 && field.wire === 2 && Buffer.isBuffer(field.value)) {
-      const usage = decodeUsage(field.value);
+    } else if (field.num === 7 && field.wire === 2 && Buffer.isBuffer(field.value)) {
+      const usage = decodeUsageStats(field.value);
       if (usage) yield usage;
     }
   }
 }
 
-function decodeUsage(buf: Buffer): CloudChatEvent | null {
-  let promptTokens: number | undefined;
-  let completionTokens: number | undefined;
-  let cachedInputTokens: number | undefined;
-  let cacheCreationInputTokens: number | undefined;
+function decodeUsageStats(buf: Buffer): CloudChatEvent | null {
+  let input: number | undefined;
+  let output: number | undefined;
+  let cacheWrite: number | undefined;
+  let cacheRead: number | undefined;
+  let modelUid: string | undefined;
   for (const field of iterFields(buf)) {
-    if (field.num !== 2 || field.wire !== 2 || !Buffer.isBuffer(field.value)) continue;
-    let metric: string | undefined;
-    let value: number | undefined;
-    for (const inner of iterFields(field.value)) {
-      if (inner.num === 5 && inner.wire === 2 && Buffer.isBuffer(inner.value)) {
-        metric = inner.value.toString("utf8");
-      } else if (inner.num === 4 && inner.wire === 2 && Buffer.isBuffer(inner.value)) {
-        for (const dim of iterFields(inner.value)) {
-          if (dim.num === 2 && dim.wire === 5 && Buffer.isBuffer(dim.value)) {
-            value = dim.value.readFloatLE(0);
-          }
-        }
-      }
+    if (field.wire === 0) {
+      const n = Number(field.value);
+      if (field.num === 2) input = n;
+      else if (field.num === 3) output = n;
+      else if (field.num === 4) cacheWrite = n;
+      else if (field.num === 5) cacheRead = n;
+    } else if (field.num === 9 && field.wire === 2 && Buffer.isBuffer(field.value)) {
+      modelUid = field.value.toString("utf8");
     }
-    if (!metric || value === undefined || !Number.isFinite(value)) continue;
-    const n = Math.round(value);
-    if (metric === "input_tokens") promptTokens = n;
-    else if (metric === "output_tokens") completionTokens = n;
-    else if (metric.includes("cached") || metric.includes("cache_read")) cachedInputTokens = n;
-    else if (metric.includes("cache_creation")) cacheCreationInputTokens = n;
   }
-  if ([promptTokens, completionTokens, cachedInputTokens, cacheCreationInputTokens].every((n) => n === undefined)) return null;
-  return {
-    kind: "usage",
-    promptTokens,
-    completionTokens,
-    totalTokens: (promptTokens ?? 0) + (completionTokens ?? 0) + (cachedInputTokens ?? 0) + (cacheCreationInputTokens ?? 0),
-    cachedInputTokens,
-    cacheCreationInputTokens,
-  };
+  if (input === undefined && output === undefined && cacheRead === undefined && cacheWrite === undefined) {
+    return null;
+  }
+  return { kind: "usage", input, output, cacheRead, cacheWrite, modelUid };
+}
+
+export interface TrailerError {
+  code: string;
+  message: string;
+}
+
+export function readTrailerError(text: string): TrailerError | null {
+  if (!text.includes('"error"')) return null;
+  try {
+    const parsed = JSON.parse(text) as { error?: { code?: unknown; message?: unknown } };
+    if (!parsed.error || typeof parsed.error !== "object") return null;
+    return {
+      code: typeof parsed.error.code === "string" ? parsed.error.code : "",
+      message: typeof parsed.error.message === "string" ? parsed.error.message : "",
+    };
+  } catch {
+    return { code: "", message: text.slice(0, 512) };
+  }
 }
 
 const sessionCache = new Map<string, { sessionId: string; cascadeId: string }>();
 
-function sessionIds(apiKey: string, host: string) {
+export function sessionIds(
+  apiKey: string,
+  host: string,
+  options?: { sessionId?: string; cacheRetention?: string },
+): { sessionId: string; cascadeId: string } {
+  // Compaction and other no-cache callers must not share or reuse a cascade.
+  if (options?.cacheRetention === "none") {
+    return { sessionId: randomUUID(), cascadeId: randomUUID() };
+  }
+  // Pi threads its session id through so each session tree keeps its own
+  // Devin cascade instead of sharing one process-wide cascade.
+  if (options?.sessionId) {
+    return { sessionId: options.sessionId, cascadeId: options.sessionId };
+  }
   const key = `${host}\x1f${apiKey}`;
   let ids = sessionCache.get(key);
   if (!ids) {
@@ -235,45 +278,41 @@ function sessionIds(apiKey: string, host: string) {
   return ids;
 }
 
+function mergeHeaders(extra: Record<string, string | null> | undefined): Record<string, string> {
+  const headers = new Map<string, string>([
+    ["content-type", "application/connect+proto"],
+    ["connect-protocol-version", "1"],
+    ["connect-accept-encoding", "gzip"],
+    ["user-agent", "connect-go/1.18.1 (go1.26.3)"],
+  ]);
+  for (const [name, value] of Object.entries(extra ?? {})) {
+    if (value == null) headers.delete(name.toLowerCase());
+    else headers.set(name.toLowerCase(), value);
+  }
+  return Object.fromEntries(headers);
+}
+
 async function* streamChatEvents(args: {
-  apiKey: string;
   host: string;
-  modelUid: string;
-  systemPrompt?: string;
-  messages: ChatHistoryItem[];
-  tools?: ToolDef[];
-  maxOutputTokens?: number;
+  request: { body: Buffer; historyBytes: number };
+  headers?: Record<string, string | null>;
   signal?: AbortSignal;
+  fetchFn?: typeof fetch;
+  onResponse?: (response: { status: number; headers: Record<string, string> }) => void | Promise<void>;
 }): AsyncGenerator<CloudChatEvent> {
   const host = args.host.replace(/\/$/, "");
-  const userJwt = await getCachedUserJwt(args.apiKey, host, args.signal);
-  const ids = sessionIds(args.apiKey, host);
-  const proto = buildGetChatMessageRequest({
-    apiKey: args.apiKey,
-    userJwt,
-    modelUid: args.modelUid,
-    systemPrompt: args.systemPrompt,
-    messages: args.messages,
-    tools: args.tools,
-    cascadeId: ids.cascadeId,
-    promptId: randomUUID(),
-    sessionId: ids.sessionId,
-    requestId: BigInt(Date.now()),
-    triggerId: randomUUID(),
-    maxOutputTokens: args.maxOutputTokens,
-  });
-
-  const resp = await fetch(`${host}/exa.api_server_pb.ApiServerService/GetChatMessage`, {
-    method: "POST",
-    headers: {
-      "Content-Type": "application/connect+proto",
-      "Connect-Protocol-Version": "1",
-      "Connect-Content-Encoding": "gzip",
-      "Connect-Accept-Encoding": "gzip",
+  const resp = await (args.fetchFn ?? fetch)(
+    `${host}/exa.api_server_pb.ApiServerService/GetChatMessage`,
+    {
+      method: "POST",
+      headers: mergeHeaders(args.headers),
+      body: new Uint8Array(frameConnectStream(args.request.body, false)),
+      signal: args.signal,
     },
-    body: new Uint8Array(frameConnectStream(proto, true)),
-    signal: args.signal,
-  });
+  );
+  if (args.onResponse) {
+    await args.onResponse({ status: resp.status, headers: Object.fromEntries(resp.headers.entries()) });
+  }
   if (!resp.ok) {
     throw new Error(`GetChatMessage HTTP ${resp.status}: ${(await resp.text()).slice(0, 300)}`);
   }
@@ -283,7 +322,8 @@ async function* streamChatEvents(args: {
   const queue: Buffer[] = [];
   let queued = 0;
   let sawEos = false;
-  let trailerError: string | null = null;
+  let producedOutput = false;
+  let trailerError: TrailerError | null = null;
 
   const peek = (n: number): Buffer | null => {
     if (queued < n) return null;
@@ -331,6 +371,9 @@ async function* streamChatEvents(args: {
         if (!header) break;
         const flags = header[0];
         const len = header.readUInt32BE(1);
+        if (len > MAX_CONNECT_FRAME_PAYLOAD) {
+          throw new Error(`Devin stream frame too large: ${len} bytes`);
+        }
         if (queued < 5 + len) break;
         drop(5);
         const raw = peek(len) ?? Buffer.alloc(0);
@@ -339,18 +382,15 @@ async function* streamChatEvents(args: {
         if (flags & 0x01) payload = zlib.gunzipSync(raw);
         if (flags & 0x02) {
           sawEos = true;
-          const text = payload.toString("utf8");
-          if (text.includes('"error"')) {
-            try {
-              const parsed = JSON.parse(text) as { error?: { message?: string } };
-              trailerError = parsed.error?.message ?? text;
-            } catch {
-              trailerError = text;
-            }
-          }
+          trailerError = readTrailerError(payload.toString("utf8").trim());
           continue;
         }
-        yield* decodeChatFrame(payload);
+        for (const event of decodeChatFrame(payload)) {
+          if (event.kind === "text" || event.kind === "reasoning" || event.kind === "tool_call_start") {
+            producedOutput = true;
+          }
+          yield event;
+        }
       }
     }
   } finally {
@@ -360,13 +400,30 @@ async function* streamChatEvents(args: {
       // ignore
     }
     try {
+      // Await teardown so a rejection from an errored stream is observed.
       await resp.body?.cancel();
     } catch {
       // ignore
     }
   }
 
-  if (trailerError) throw new Error(trailerError);
+  if (trailerError) {
+    if (
+      !producedOutput &&
+      trailerError.code.toLowerCase() === "invalid_argument" &&
+      /\binternal error\b/i.test(trailerError.message) &&
+      args.request.historyBytes >= LARGE_HISTORY_RECOVERY_BYTES
+    ) {
+      throw new Error(
+        `Devin stream rejected: ${trailerError.message} (message history of ${args.request.historyBytes} bytes likely exceeds the context window)`,
+      );
+    }
+    throw new Error(
+      trailerError.code
+        ? `Devin stream rejected (${trailerError.code}): ${trailerError.message}`
+        : `Devin stream rejected: ${trailerError.message}`,
+    );
+  }
   if (!sawEos) throw new Error("Devin stream ended without an EOS trailer");
 }
 
@@ -398,10 +455,9 @@ export function streamDevin(
 
     let textOpen = false;
     let thinkingOpen = false;
-    let toolIndex = -1;
-    let partialJson = "";
-    let toolId = "";
-    let toolName = "";
+    const openTools = new Map<string, number>();
+    const toolJson = new Map<string, string>();
+    let lastToolId: string | undefined;
 
     const closeText = () => {
       if (!textOpen) return;
@@ -421,38 +477,92 @@ export function streamDevin(
       }
       thinkingOpen = false;
     };
-    const closeTool = () => {
-      if (toolIndex < 0) return;
-      const block = output.content[toolIndex];
+    const closeTool = (id: string) => {
+      const idx = openTools.get(id);
+      if (idx === undefined) return;
+      openTools.delete(id);
+      const block = output.content[idx];
       if (block.type === "toolCall") {
-        block.arguments = parseStreamingJson(partialJson);
+        const json = toolJson.get(id) ?? "";
+        try {
+          block.arguments = JSON.parse(json);
+        } catch {
+          // keep last parsed object
+        }
         stream.push({
           type: "toolcall_end",
-          contentIndex: toolIndex,
-          toolCall: { type: "toolCall", id: toolId, name: toolName, arguments: block.arguments },
+          contentIndex: idx,
+          toolCall: { type: "toolCall", id, name: block.name, arguments: block.arguments },
           partial: output,
         });
       }
-      toolIndex = -1;
+      toolJson.delete(id);
+      if (lastToolId === id) lastToolId = undefined;
+    };
+    const closeAllTools = () => {
+      for (const id of [...openTools.keys()]) closeTool(id);
+    };
+    const currentThinkingBlock = () => {
+      const block = output.content[output.content.length - 1];
+      return block?.type === "thinking" ? block : undefined;
     };
 
     try {
       const apiKey = options?.apiKey;
       if (!apiKey) throw new Error("No Devin credentials. Run /login devin (uses the local Devin CLI).");
+      if (options?.deferred) throw new Error("The devin provider does not support deferred responses");
       const host = (options?.env?.DEVIN_API_SERVER_URL || "https://server.codeium.com").replace(/\/$/, "");
       const modelUid = resolveModelUid(model.id, model.thinkingLevelMap, options?.reasoning);
-      const mapped = mapContextToChat(context);
-      stream.push({ type: "start", partial: output });
+      output.providerThinkingLevel = modelUid;
+      output.responseModel = modelUid;
+      const mapped = mapContextToChat(context, model);
+      const timeoutMs = options?.timeoutMs;
+      const signal =
+        timeoutMs && timeoutMs > 0
+          ? anySignal(
+              [options?.signal, AbortSignal.timeout(timeoutMs)].filter((s): s is AbortSignal => Boolean(s)),
+            )
+          : options?.signal;
+      const ids = sessionIds(apiKey, host, options);
+      const userJwt = await getCachedUserJwt(apiKey, host, signal);
 
-      for await (const event of streamChatEvents({
+      const sampling =
+        model.samplingParams || options?.samplingParams
+          ? { ...model.samplingParams, ...options?.samplingParams }
+          : {};
+      const wireArgs = {
         apiKey,
-        host,
+        userJwt,
         modelUid,
         systemPrompt: mapped.systemPrompt,
         messages: mapped.messages,
         tools: mapped.tools.length > 0 ? mapped.tools : undefined,
-        maxOutputTokens: options?.maxTokens,
-        signal: options?.signal,
+        cascadeId: ids.cascadeId,
+        promptId: randomUUID(),
+        sessionId: ids.sessionId,
+        requestId: BigInt(Date.now()),
+        triggerId: randomUUID(),
+        maxTokens: clampMaxTokensToContext(model, context, options?.maxTokens ?? model.maxTokens),
+        temperature: options?.temperature ?? 1.0,
+        topP: typeof sampling.topP === "number" ? sampling.topP : 0.95,
+        topK: typeof sampling.topK === "number" ? sampling.topK : 40,
+      };
+      const overridden = await options?.onPayload?.(wireArgs, model);
+      const finalArgs =
+        overridden && typeof overridden === "object" ? { ...wireArgs, ...overridden } : wireArgs;
+      const request = buildGetChatMessageRequest(finalArgs);
+
+      stream.push({ type: "start", partial: output });
+
+      for await (const event of streamChatEvents({
+        host,
+        request,
+        headers: options?.headers,
+        signal,
+        fetchFn: options?.fetch,
+        onResponse: options?.onResponse
+          ? (response) => options?.onResponse?.(response, model)
+          : undefined,
       })) {
         if (event.kind === "text") {
           closeThinking();
@@ -480,46 +590,66 @@ export function streamDevin(
             block.thinking += event.text;
             stream.push({ type: "thinking_delta", contentIndex: idx, delta: event.text, partial: output });
           }
+        } else if (event.kind === "signature") {
+          const block = currentThinkingBlock();
+          if (block) block.thinkingSignature = event.text;
+        } else if (event.kind === "thinking_redacted") {
+          const block = currentThinkingBlock();
+          if (block) block.redacted = true;
+        } else if (event.kind === "message_id") {
+          if (!output.responseId) output.responseId = event.id;
+        } else if (event.kind === "actual_model") {
+          output.responseModel = event.uid;
         } else if (event.kind === "tool_call_start") {
           closeText();
           closeThinking();
-          closeTool();
-          toolId = event.id;
-          toolName = event.name;
-          partialJson = "";
           output.content.push({ type: "toolCall", id: event.id, name: event.name, arguments: {} });
-          toolIndex = output.content.length - 1;
-          stream.push({ type: "toolcall_start", contentIndex: toolIndex, partial: output });
+          openTools.set(event.id, output.content.length - 1);
+          toolJson.set(event.id, "");
+          lastToolId = event.id;
+          stream.push({ type: "toolcall_start", contentIndex: output.content.length - 1, partial: output });
         } else if (event.kind === "tool_call_args") {
-          if (toolIndex < 0) continue;
-          partialJson += event.argsDelta;
-          const block = output.content[toolIndex];
+          const id = event.id ?? lastToolId;
+          const idx = id === undefined ? undefined : openTools.get(id);
+          if (id === undefined || idx === undefined) continue;
+          const json = (toolJson.get(id) ?? "") + event.argsDelta;
+          toolJson.set(id, json);
+          const block = output.content[idx];
           if (block.type === "toolCall") {
-            block.arguments = parseStreamingJson(partialJson);
+            try {
+              block.arguments = JSON.parse(json);
+            } catch {
+              // incomplete json
+            }
           }
-          stream.push({ type: "toolcall_delta", contentIndex: toolIndex, delta: event.argsDelta, partial: output });
+          stream.push({ type: "toolcall_delta", contentIndex: idx, delta: event.argsDelta, partial: output });
         } else if (event.kind === "finish") {
           closeText();
           closeThinking();
-          closeTool();
+          closeAllTools();
+          if (event.reason === "content_filter") {
+            throw new Error("Devin stop_reason: content_filter");
+          }
+          if (event.reason === "error") {
+            throw new Error("Devin stop_reason: error");
+          }
           output.stopReason =
             event.reason === "tool_calls" ? "toolUse" : event.reason === "length" ? "length" : "stop";
         } else if (event.kind === "usage") {
-          // Empty trailers must not erase usage already received. Metrics are
-          // snapshots, and a frame may contain only some of the components.
-          if (!event.totalTokens) continue;
-          output.usage.input = event.promptTokens ?? output.usage.input;
-          output.usage.output = event.completionTokens ?? output.usage.output;
-          output.usage.cacheRead = event.cachedInputTokens ?? output.usage.cacheRead;
-          output.usage.cacheWrite = event.cacheCreationInputTokens ?? output.usage.cacheWrite;
-          output.usage.totalTokens = output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
+          output.usage.input = event.input ?? 0;
+          output.usage.output = event.output ?? 0;
+          output.usage.cacheRead = event.cacheRead ?? 0;
+          output.usage.cacheWrite = event.cacheWrite ?? 0;
+          output.usage.totalTokens =
+            output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
+          if (!output.responseModel && event.modelUid) output.responseModel = event.modelUid;
           calculateCost(model, output.usage);
         }
       }
 
       closeText();
       closeThinking();
-      closeTool();
+      closeAllTools();
       stream.push({ type: "done", reason: output.stopReason as "stop" | "length" | "toolUse", message: output });
       stream.end();
     } catch (error) {

test-runner.mjs

@@ -1,8 +0,0 @@
-import { readdirSync } from "node:fs";
-import { spawnSync } from "node:child_process";
-import { join } from "node:path";
-
-const files = readdirSync("test").filter((file) => file.endsWith(".test.mjs")).sort().map((file) => join("test", file));
-const result = spawnSync(process.execPath, ["--import", "tsx", "--test", ...files], { stdio: "inherit" });
-if (result.error) throw result.error;
-process.exit(result.status ?? 1);

test/catalog-cache.test.mjs

@@ -1,65 +0,0 @@
-import assert from "node:assert/strict";
-import { mkdtempSync, rmSync, readFileSync, writeFileSync } from "node:fs";
-import { tmpdir } from "node:os";
-import { join } from "node:path";
-import test from "node:test";
-import {
-  catalogCachePath,
-  isUsableCatalog,
-  isCatalogCacheFresh,
-  readCatalogCache,
-  writeCatalogCache,
-} from "../src/catalog-cache.ts";
-
-const catalog = {
-  families: [{
-    family_label: "SWE-2",
-    family_uid: "swe-2",
-    slug: "swe-2",
-    variants: [{ model_uid: "swe-2-high", label: "SWE-2 High" }],
-  }],
-};
-
-test("writes and reads a validated catalog atomically", (t) => {
-  const directory = mkdtempSync(join(tmpdir(), "pi-devin-cache-"));
-  t.after(() => rmSync(directory, { recursive: true, force: true }));
-  const path = join(directory, "nested", "models.json");
-  writeCatalogCache(catalog, path, 1_000);
-  assert.deepEqual(readCatalogCache(path), { catalog, fetchedAt: 1_000 });
-  assert.equal(JSON.parse(readFileSync(path, "utf8")).version, 1);
-});
-
-test("rejects malformed cache content", (t) => {
-  const directory = mkdtempSync(join(tmpdir(), "pi-devin-cache-"));
-  t.after(() => rmSync(directory, { recursive: true, force: true }));
-  const path = join(directory, "models.json");
-  writeFileSync(path, JSON.stringify({ version: 1, fetchedAt: 1_000, catalog: { families: [{}] } }));
-  assert.throws(() => readCatalogCache(path), /Invalid Devin model catalog cache/);
-});
-
-test("rejects catalogs without usable model families", () => {
-  assert.equal(isUsableCatalog({ families: [] }), false);
-  assert.equal(isUsableCatalog({
-    families: [{ family_label: "Empty", family_uid: "empty", slug: "empty", variants: [] }],
-  }), false);
-});
-
-test("checks cache age and future timestamps", () => {
-  const cached = { catalog, fetchedAt: 1_000 };
-  assert.equal(isCatalogCacheFresh(cached, 1_500, 500), true);
-  assert.equal(isCatalogCacheFresh(cached, 1_501, 500), false);
-  assert.equal(isCatalogCacheFresh(cached, 999, 500), false);
-});
-
-test("uses XDG cache directory when configured", () => {
-  assert.equal(catalogCachePath({ XDG_CACHE_HOME: "/cache" }, "/home/test"), join("/cache", "pi-devin", "models.json"));
-  assert.equal(catalogCachePath({}, "/home/test"), join("/home/test", ".cache", "pi-devin", "models.json"));
-});
-
-for (const [field, value] of Object.entries({ cost_summary: {}, max_context_tokens: "1000", max_output_tokens: -1, model_uid: "", is_beta: "yes" })) {
-  test(`rejects invalid ${field} before model conversion`, () => {
-    const malformed = structuredClone(catalog);
-    malformed.families[0].variants[0][field] = value;
-    assert.equal(isUsableCatalog(malformed), false);
-  });
-}

test/catalog-startup.test.mjs

@@ -1,101 +0,0 @@
-import assert from "node:assert/strict";
-import test, { after } from "node:test";
-import fs from "node:fs";
-import { tmpdir } from "node:os";
-import { join } from "node:path";
-import { syncBuiltinESMExports } from "node:module";
-import { catalogCachePath, writeCatalogCache } from "../src/catalog-cache.ts";
-
-const directory = fs.mkdtempSync(join(tmpdir(), "pi-devin-catalog-test-"));
-after(() => fs.rmSync(directory, { recursive: true, force: true }));
-const cli = join(directory, process.platform === "win32" ? "devin.cmd" : "devin");
-const counter = join(directory, "calls");
-const catalog = { families: [{ family_label: "Test", family_uid: "test", slug: "test", variants: [{ model_uid: "test-high", label: "Test high", cost_summary: "$1/MTok In, $2/MTok Out" }] }] };
-const script = join(directory, "catalog.cjs");
-fs.writeFileSync(script, `const fs = require('node:fs');\nfs.appendFileSync(${JSON.stringify(counter)}, 'call\\n');\nsetTimeout(() => { console.log(process.env.TEST_CATALOG); process.exit(Number(process.env.TEST_CLI_EXIT || 0)); }, 30);\n`);
-fs.writeFileSync(cli, process.platform === "win32" ? `@echo off\r\n"${process.execPath}" "${script}"\r\n` : `#!/bin/sh\nexec "${process.execPath}" "${script}"\n`, { mode: 0o700 });
-const originalEnv = { ...process.env };
-process.env.DEVIN_CLI = cli;
-const { default: extension } = await import("../extensions/index.ts");
-const { clearDevinBinCache } = await import("../src/cli.ts");
-after(() => { if (originalEnv.DEVIN_CLI === undefined) delete process.env.DEVIN_CLI; else process.env.DEVIN_CLI = originalEnv.DEVIN_CLI; });
-
-function setup(t) {
-  clearDevinBinCache();
-  const root = fs.mkdtempSync(join(directory, "cache-"));
-  const previous = { ...process.env };
-  process.env.XDG_CACHE_HOME = root;
-  process.env.TEST_CATALOG = JSON.stringify(catalog);
-  delete process.env.PI_OFFLINE;
-  delete process.env.TEST_CLI_EXIT;
-  fs.writeFileSync(counter, "");
-  const read = fs.readFileSync, exists = fs.existsSync;
-  const isCredentials = (path) => String(path).replaceAll("\\", "/").endsWith("devin/credentials.toml");
-  t.mock.method(fs, "readFileSync", (path, ...args) => isCredentials(path) ? 'api_key = "synthetic-test-key"\n' : read(path, ...args));
-  t.mock.method(fs, "existsSync", (path) => isCredentials(path) || exists(path));
-  t.mock.method(console, "warn", () => {});
-  syncBuiltinESMExports();
-  const registrations = [], hooks = {}, commands = {};
-  const pi = { registerProvider: (_id, config) => registrations.push(config), on: (name, fn) => hooks[name] = fn, registerCommand: (name, command) => commands[name] = command };
-  t.after(async () => {
-    await hooks.session_shutdown?.();
-    t.mock.restoreAll(); syncBuiltinESMExports();
-    for (const key of ["XDG_CACHE_HOME", "PI_OFFLINE", "TEST_CATALOG", "TEST_CLI_EXIT"]) {
-      if (previous[key] === undefined) delete process.env[key]; else process.env[key] = previous[key];
-    }
-  });
-  return { registrations, hooks, commands, pi, path: catalogCachePath(), calls: () => fs.readFileSync(counter, "utf8").split("\n").filter(Boolean).length };
-}
-
-test("fresh cache registers once without invoking the CLI", async (t) => {
-  const x = setup(t); writeCatalogCache(catalog);
-  await extension(x.pi);
-  assert.equal(x.calls(), 0); assert.equal(x.registrations.length, 1);
-  assert.equal(x.registrations[0].models[0].id, "test-high");
-});
-
-test("malformed cached cost data cannot prevent provider and command registration", async (t) => {
-  const x = setup(t); process.env.PI_OFFLINE = "1";
-  const malformed = structuredClone(catalog); malformed.families[0].variants[0].cost_summary = {};
-  fs.mkdirSync(join(process.env.XDG_CACHE_HOME, "pi-devin"));
-  fs.writeFileSync(x.path, JSON.stringify({ version: 1, fetchedAt: Date.now(), catalog: malformed }));
-  await extension(x.pi);
-  assert.equal(x.registrations.length, 1); assert.ok(x.registrations[0].models.length);
-  assert.ok(x.commands["devin-refresh"]); assert.equal(x.calls(), 0);
-});
-
-test("cold startup loads and caches the CLI catalog", async (t) => {
-  const x = setup(t); await extension(x.pi);
-  assert.equal(x.calls(), 1); assert.equal(x.registrations.at(-1).models[0].id, "test-high");
-  assert.deepEqual(JSON.parse(fs.readFileSync(x.path, "utf8")).catalog, catalog);
-});
-
-test("stale cache shares its background request with manual refresh", async (t) => {
-  const x = setup(t); writeCatalogCache(catalog, x.path, 1);
-  await extension(x.pi);
-  const notifications = [];
-  await x.commands["devin-refresh"].handler("", { ui: { notify: (...args) => notifications.push(args) } });
-  assert.equal(x.calls(), 1); assert.equal(notifications.at(-1)[1], "info");
-  assert.equal(x.registrations.at(-1).models[0].id, "test-high");
-});
-
-test("failed refresh retains cached models and does not overwrite the cache", async (t) => {
-  const x = setup(t); writeCatalogCache(catalog, x.path, 1); process.env.TEST_CATALOG = '{"families":[{}]}';
-  await extension(x.pi);
-  const notifications = [];
-  await x.commands["devin-refresh"].handler("", { ui: { notify: (...args) => notifications.push(args) } });
-  assert.equal(notifications.at(-1)[1], "error"); assert.equal(x.registrations.length, 1);
-  assert.deepEqual(JSON.parse(fs.readFileSync(x.path, "utf8")).catalog, catalog);
-});
-
-test("offline cold startup uses fallback without running the CLI", async (t) => {
-  const x = setup(t); process.env.PI_OFFLINE = "true";
-  await extension(x.pi); assert.equal(x.calls(), 0); assert.ok(x.registrations[0].models.length);
-});
-
-test("shutdown prevents a late background result from registering a provider", async (t) => {
-  const x = setup(t); writeCatalogCache(catalog, x.path, 1);
-  await extension(x.pi); await x.hooks.session_shutdown();
-  await x.commands["devin-refresh"].handler("", { ui: { notify() {} } });
-  assert.equal(x.registrations.length, 1);
-});

test/cli.test.mjs

@@ -1,73 +0,0 @@
-import assert from "node:assert/strict";
-import test from "node:test";
-import { mkdtempSync, mkdirSync, writeFileSync, rmSync, realpathSync } from "node:fs";
-import { tmpdir } from "node:os";
-import { dirname, join } from "node:path";
-import { spawnSync } from "node:child_process";
-import { clearDevinBinCache, findDevinBin, whichDevin, runDevin } from "../src/cli.ts";
-
-function fixture(t) {
-  const directory = mkdtempSync(join(tmpdir(), "pi devin cli "));
-  // Windows environment lookups ignore case; a spread object does not.
-  const keys = ["DEVIN_CLI", "LOCALAPPDATA", "ProgramFiles", "PATH"];
-  const original = Object.fromEntries(keys.map((key) => [key, process.env[key]]));
-  clearDevinBinCache();
-  t.after(() => {
-    for (const key of keys) {
-      if (original[key] === undefined) delete process.env[key]; else process.env[key] = original[key];
-    }
-    clearDevinBinCache(); rmSync(directory, { recursive: true, force: true });
-  });
-  return directory;
-}
-
-test("honors an override set after the discovery cache has already been populated", async (t) => {
-  const directory = fixture(t);
-  const first = join(directory, "first"), second = join(directory, "second");
-  writeFileSync(first, ""); writeFileSync(second, "");
-  process.env.DEVIN_CLI = first; assert.equal(findDevinBin(), first);
-  process.env.DEVIN_CLI = second; assert.equal(await whichDevin(), second);
-});
-
-test("runs a native executable directly with argument boundaries intact", async (t) => {
-  fixture(t); process.env.DEVIN_CLI = process.execPath;
-  const value = 'spaces & punctuation "stay literal"';
-  const result = await runDevin(["-e", "process.stdout.write(process.argv[1])", value]);
-  assert.equal(result.code, 0); assert.equal(result.stdout, value);
-});
-
-test("finds the Windows installer binary and ignores a directory override", { skip: process.platform !== "win32" }, (t) => {
-  const directory = fixture(t); process.env.LOCALAPPDATA = directory;
-  process.env.DEVIN_CLI = directory;
-  const binary = join(directory, "devin", "cli", "bin", "devin.exe");
-  mkdirSync(dirname(binary), { recursive: true }); writeFileSync(binary, "");
-  assert.equal(findDevinBin(), binary);
-});
-
-test("Windows PATH lookup uses where.exe and reads multiple results", { skip: process.platform !== "win32" }, async (t) => {
-  const directory = fixture(t); delete process.env.DEVIN_CLI;
-  process.env.LOCALAPPDATA = directory; process.env.ProgramFiles = directory;
-  const bin = join(directory, "path bin"); mkdirSync(bin);
-  const binary = join(bin, "devin.exe"); writeFileSync(binary, "");
-  const otherBin = join(directory, "other bin"); mkdirSync(otherBin);
-  writeFileSync(join(otherBin, "devin.exe"), "");
-  process.env.PATH = `${bin};${otherBin};${process.env.PATH}`;
-  const located = await whichDevin();
-  const probe = spawnSync("where.exe", ["devin.exe"], { encoding: "utf8" });
-  assert.ok(located, JSON.stringify({ status: probe.status, stdout: probe.stdout, stderr: probe.stderr, error: probe.error?.message }));
-  // where.exe expands Windows short paths such as RUNNER~1.
-  assert.equal(realpathSync.native(located), realpathSync.native(binary));
-  assert.ok(probe.stdout.trim().split(/\r?\n/).length >= 2);
-});
-
-test("Windows batch wrappers work in paths with spaces, including inherited stdio", { skip: process.platform !== "win32" }, async (t) => {
-  const directory = fixture(t), script = join(directory, "echo.cjs"), wrapper = join(directory, "devin.cmd");
-  writeFileSync(script, 'process.stdout.write(JSON.stringify(process.argv.slice(2)));');
-  writeFileSync(wrapper, `@echo off\r\n"${process.execPath}" "${script}" %*\r\n`);
-  process.env.DEVIN_CLI = wrapper;
-  const result = await runDevin(["models", "list", "--format", "json", "a b"]);
-  assert.equal(result.code, 0, result.stderr);
-  assert.deepEqual(JSON.parse(result.stdout), ["models", "list", "--format", "json", "a b"]);
-  assert.equal((await runDevin(["version"], { inheritStdio: true })).code, 0);
-  await assert.rejects(runDevin(["%PATH%"]), /Unsupported shell characters/);
-});

test/context-map.test.mjs

@@ -1,53 +0,0 @@
-import assert from "node:assert/strict";
-import test from "node:test";
-import { normalizeContext } from "@earendil-works/pi-ai";
-import { mapContextToChat } from "../src/context-map.ts";
-
-const tool = (name) => ({
-  name,
-  description: name,
-  parameters: { type: "object", properties: {} },
-});
-
-// Regression for #7: Pi 0.86 moves prompt and tools into system messages.
-test("preserves the prompt and tools from a normalized transcript", () => {
-  const read = tool("read");
-  const context = normalizeContext({
-    systemPrompt: "Inspect the repository.",
-    tools: [read],
-    messages: [{ role: "user", content: "Read README", timestamp: 1 }],
-  });
-  assert.deepEqual(mapContextToChat(context), {
-    systemPrompt: "Inspect the repository.",
-    messages: [
-      { role: "user", content: "Read README" },
-    ],
-    tools: [read],
-  });
-});
-
-test("applies later instruction, section, and tool changes", () => {
-  const bash = tool("bash");
-  const context = {
-    messages: [
-      { role: "system", content: "Inspect the repository.", sections: { policy: "Old policy", obsolete: "Remove me" }, toolsAdded: [tool("read")], timestamp: 0 },
-      { role: "user", content: "Continue", timestamp: 1 },
-      { role: "system", content: "Use bash.", sections: { policy: "Do not edit.", obsolete: null }, toolsRemoved: [{ name: "read" }], toolsAdded: [bash], timestamp: 2 },
-    ],
-  };
-  assert.deepEqual(mapContextToChat(context), {
-    systemPrompt: "Inspect the repository.\n\nUse bash.\n\nDo not edit.",
-    messages: [
-      { role: "user", content: "Continue" },
-    ],
-    tools: [bash],
-  });
-});
-
-test("accepts an empty transcript", () => {
-  assert.deepEqual(mapContextToChat(normalizeContext({ messages: [] })), {
-    systemPrompt: undefined,
-    messages: [],
-    tools: [],
-  });
-});

test/stream-abort.test.mjs

@@ -1,104 +0,0 @@
-import assert from "node:assert/strict";
-import test from "node:test";
-import { normalizeContext } from "@earendil-works/pi-ai";
-import { clearCachedUserJwt } from "../src/jwt.ts";
-import { streamDevin } from "../src/stream.ts";
-import { encodeMessage } from "../src/wire.ts";
-
-const model = {
-  id: "test-model",
-  name: "Test model",
-  api: "devin-local",
-  provider: "devin",
-  baseUrl: "https://devin.invalid",
-  reasoning: false,
-  input: ["text", "image"],
-  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-  contextWindow: 262_000,
-  maxTokens: 1_000,
-};
-const user = (content) => ({ role: "user", content, timestamp: 1 });
-
-test("aborting an in-flight stream reports aborted without leaking a rejection", async (t) => {
-  clearCachedUserJwt();
-  t.after(clearCachedUserJwt);
-
-  const rejections = [];
-  const onRejection = (reason) => rejections.push(reason);
-  process.on("unhandledRejection", onRejection);
-  t.after(() => process.off("unhandledRejection", onRejection));
-
-  const ac = new AbortController();
-  t.mock.method(globalThis, "fetch", async (url, options) => {
-    if (url === "https://devin.invalid/exa.auth_pb.AuthService/GetUserJwt") {
-      return new Response(encodeMessage(1, Buffer.from("eyJtest.jwt")));
-    }
-    assert.equal(url, "https://devin.invalid/exa.api_server_pb.ApiServerService/GetChatMessage");
-    // Mirror undici: an aborted fetch errors the response body with signal.reason.
-    const body = new ReadableStream({
-      start(controller) {
-        controller.enqueue(new TextEncoder().encode("partial"));
-        const onAbort = () => controller.error(options.signal.reason);
-        if (options.signal?.aborted) onAbort();
-        else options.signal?.addEventListener("abort", onAbort, { once: true });
-      },
-    });
-    return new Response(body);
-  });
-
-  const stream = streamDevin(model, normalizeContext({ messages: [user("hello")] }), {
-    apiKey: "synthetic-test-key",
-    env: { DEVIN_API_SERVER_URL: "https://devin.invalid" },
-    signal: ac.signal,
-  });
-
-  const events = [];
-  const consume = (async () => {
-    for await (const event of stream) events.push(event);
-  })();
-  ac.abort();
-  await consume;
-  // Give any floating rejection a chance to surface.
-  await new Promise((resolve) => setTimeout(resolve, 50));
-
-  const last = events.at(-1);
-  assert.equal(last?.type, "error");
-  assert.equal(last?.reason, "aborted");
-  assert.equal((await stream.result()).stopReason, "aborted");
-  assert.deepEqual(rejections, []);
-});
-
-test("real fetch cancellation after a response chunk leaves the process usable", { timeout: 10_000 }, async (t) => {
-  const { createServer } = await import("node:http");
-  const { once } = await import("node:events");
-  const { encodeString, encodeVarintField, frameConnectStream } = await import("../src/wire.ts");
-  clearCachedUserJwt(); t.after(clearCachedUserJwt);
-  let chats = 0;
-  const server = createServer((req, res) => {
-    if (req.url.endsWith("/GetUserJwt")) {
-      res.end(encodeString(1, "eyJtest.jwt")); return;
-    }
-    chats++;
-    res.writeHead(200, { "Content-Type": "application/connect+proto" });
-    res.write(frameConnectStream(encodeString(3, "started"), false));
-    if (chats > 1) {
-      res.write(frameConnectStream(encodeVarintField(5, 0), false));
-      res.end(Buffer.from([2, 0, 0, 0, 2, 123, 125]));
-    }
-  });
-  t.after(() => { server.closeAllConnections(); server.close(); });
-  server.listen(0, "127.0.0.1"); await once(server, "listening");
-  const ac = new AbortController();
-  const options = { apiKey: "synthetic-test-key", env: { DEVIN_API_SERVER_URL: `http://127.0.0.1:${server.address().port}` } };
-  const stream = streamDevin(model, normalizeContext({ messages: [user("hello")] }), { ...options, signal: ac.signal });
-  let sawText = false;
-  for await (const event of stream) {
-    if (event.type === "text_delta") { sawText = true; ac.abort(); }
-  }
-  assert.equal(sawText, true);
-  assert.equal((await stream.result()).stopReason, "aborted");
-  const next = streamDevin(model, normalizeContext({ messages: [user("again")] }), options);
-  for await (const event of next) assert.notEqual(event.type, "error");
-  assert.equal((await next.result()).stopReason, "stop");
-  assert.equal(chats, 2);
-});

test/stream-deltas.test.mjs

@@ -1,86 +0,0 @@
-import assert from "node:assert/strict";
-import test from "node:test";
-import { setImmediate } from "node:timers/promises";
-import { streamDevin } from "../src/stream.ts";
-import { clearCachedUserJwt } from "../src/jwt.ts";
-import { encodeMessage, encodeString, encodeTag, encodeVarintField, frameConnectStream } from "../src/wire.ts";
-
-const model = {
-  id: "test-model", name: "Test", api: "devin-local", provider: "devin",
-  baseUrl: "https://devin.invalid", reasoning: false, input: ["text"],
-  cost: { input: 1, output: 2, cacheRead: 0.1, cacheWrite: 1.25 },
-  contextWindow: 1_000_000, maxTokens: 1000,
-};
-
-async function run(t, frames) {
-  clearCachedUserJwt();
-  t.after(clearCachedUserJwt);
-  t.mock.method(globalThis, "fetch", async (url) => {
-    if (url === "https://devin.invalid/exa.auth_pb.AuthService/GetUserJwt") {
-      return new Response(encodeString(1, "eyJtest.jwt"));
-    }
-    assert.equal(url, "https://devin.invalid/exa.api_server_pb.ApiServerService/GetChatMessage");
-    const chunks = [...frames.map((frame) => frameConnectStream(frame, false)), Buffer.from([2, 0, 0, 0, 2, 123, 125])];
-    return new Response(new ReadableStream({
-      async pull(controller) {
-        await setImmediate();
-        if (chunks.length) controller.enqueue(chunks.shift());
-        else controller.close();
-      },
-    }));
-  });
-  const stream = streamDevin(model, { messages: [] }, {
-    apiKey: "synthetic-test-key", env: { DEVIN_API_SERVER_URL: "https://devin.invalid" },
-  });
-  const events = [];
-  for await (const event of stream) events.push(structuredClone(event));
-  const result = await stream.result();
-  assert.notEqual(result.stopReason, "error", result.errorMessage);
-  return { events, result };
-}
-
-function usage(metrics) {
-  return encodeMessage(28, Buffer.concat(Object.entries(metrics).map(([name, value]) => {
-    const float = Buffer.alloc(4);
-    float.writeFloatLE(value);
-    return encodeMessage(2, Buffer.concat([
-      encodeString(5, name), encodeMessage(4, Buffer.concat([encodeTag(2, 5), float])),
-    ]));
-  })));
-}
-
-const metrics = { input_tokens: 461, output_tokens: 262, cached_input_tokens: 228_162, cache_creation_input_tokens: 17 };
-
-test("counts all four usage components and preserves them after an empty frame (#3, #4)", async (t) => {
-  const { result } = await run(t, [usage(metrics), usage(Object.fromEntries(Object.keys(metrics).map((key) => [key, 0])))]);
-  assert.deepEqual({ ...result.usage, cost: undefined }, {
-    input: 461, output: 262, cacheRead: 228_162, cacheWrite: 17, totalTokens: 228_902, cost: undefined,
-  });
-  assert.equal(result.usage.cost.input, 461 / 1_000_000);
-  assert.equal(result.usage.cost.cacheRead, 228_162 * 0.1 / 1_000_000);
-});
-
-test("accepts cache-only usage and combines separate component snapshots without double counting", async (t) => {
-  const { result } = await run(t, [
-    usage({ cached_input_tokens: 58_905 }), usage({ input_tokens: 280, output_tokens: 118 }),
-    usage({ input_tokens: 280, output_tokens: 118 }), usage({ cache_creation_input_tokens: 17 }),
-  ]);
-  assert.equal(result.usage.totalTokens, 59_320);
-  assert.equal(result.usage.cacheRead, 58_905);
-  assert.equal(result.usage.cacheWrite, 17);
-});
-
-test("keeps growing tool arguments across split keys, strings, and Unicode escapes (#2)", async (t) => {
-  const fragments = ['{"path":"file.txt","cont', 'ent":"hello ', '\\u26', '3a"}'];
-  const frames = fragments.map((part, index) => encodeMessage(6, Buffer.concat([
-    ...(index === 0 ? [encodeString(1, "call-1"), encodeString(2, "write")] : []), encodeString(3, part),
-  ])));
-  frames.push(encodeVarintField(5, 10));
-  const { events, result } = await run(t, frames);
-  const deltas = events.filter((event) => event.type === "toolcall_delta");
-  assert.equal(deltas.length, 4);
-  for (const event of deltas) assert.equal(event.partial.content[0].arguments.path, "file.txt");
-  assert.equal(deltas[1].partial.content[0].arguments.content, "hello");
-  assert.deepEqual(result.content[0].arguments, { path: "file.txt", content: "hello ☺" });
-  assert.equal(result.stopReason, "toolUse");
-});

test/stream-request.test.mjs

@@ -1,164 +0,0 @@
-import assert from "node:assert/strict";
-import test from "node:test";
-import { gunzipSync } from "node:zlib";
-import { normalizeContext } from "@earendil-works/pi-ai";
-import { clearCachedUserJwt } from "../src/jwt.ts";
-import { streamDevin } from "../src/stream.ts";
-import { encodeMessage, encodeString, encodeVarintField, frameConnectStream, iterFields } from "../src/wire.ts";
-
-const model = {
-  id: "test-model",
-  name: "Test model",
-  api: "devin-local",
-  provider: "devin",
-  baseUrl: "https://devin.invalid",
-  reasoning: false,
-  input: ["text", "image"],
-  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-  contextWindow: 262_000,
-  maxTokens: 1_000,
-};
-const readTool = {
-  name: "read",
-  description: "Read a file",
-  parameters: { type: "object", properties: { path: { type: "string" } }, required: ["path"] },
-};
-const user = (content) => ({ role: "user", content, timestamp: 1 });
-const fields = (buffer) => [...iterFields(buffer)];
-const messages = (request) => request.filter((field) => field.num === 3).map((field) => fields(field.value));
-const stringField = (fields, number) => fields.find((field) => field.num === number)?.value.toString("utf8");
-
-function response(body) {
-  // Connect's end-of-stream frame contains the JSON trailer {}.
-  return new Response(Buffer.concat([frameConnectStream(body, false), Buffer.from([2, 0, 0, 0, 2, 123, 125])]));
-}
-
-function mockDevin(t, replies = [Buffer.concat([encodeString(3, "OK"), encodeVarintField(5, 0)])]) {
-  const requests = [];
-  clearCachedUserJwt();
-  t.after(clearCachedUserJwt);
-  t.mock.method(globalThis, "fetch", async (url, options) => {
-    if (url === "https://devin.invalid/exa.auth_pb.AuthService/GetUserJwt") {
-      return new Response(encodeMessage(1, Buffer.from("eyJtest.jwt")));
-    }
-    assert.equal(url, "https://devin.invalid/exa.api_server_pb.ApiServerService/GetChatMessage");
-    const frame = Buffer.from(options.body);
-    assert.equal(frame[0], 1, "request uses Connect gzip compression");
-    assert.equal(frame.readUInt32BE(1), frame.length - 5);
-    requests.push(fields(gunzipSync(frame.subarray(5))));
-    const reply = replies[requests.length - 1];
-    assert.ok(reply, "unexpected extra chat request");
-    return response(reply);
-  });
-  return requests;
-}
-
-async function complete(context) {
-  const stream = streamDevin(model, context, {
-    apiKey: "synthetic-test-key",
-    env: { DEVIN_API_SERVER_URL: "https://devin.invalid" },
-  });
-  for await (const event of stream) {
-    assert.notEqual(event.type, "error", event.error?.errorMessage);
-  }
-  return stream.result();
-}
-
-// #7: verify the normalized transcript survives the complete request-encoding path.
-test("encodes the current system prompt once in field 2, with unchanged user text and images", async (t) => {
-  const requests = mockDevin(t);
-  const text = "  Inspect this: <system>literal user text</system>\nこんにちは  ";
-  const image = { type: "image", mimeType: "image/png", data: "aW1hZ2UtZml4dHVyZQ==" };
-  const context = normalizeContext({
-    systemPrompt: "Initial instructions.",
-    tools: [readTool],
-    messages: [user([{ type: "text", text }, image])],
-  });
-  const tools = Array.from({ length: 90 }, (_, i) => ({ ...readTool, name: `read_${i}` }));
-  context.messages.push({
-    role: "system",
-    content: "Updated instructions.",
-    sections: { policy: "Do not edit." },
-    toolsRemoved: [{ name: "read" }],
-    toolsAdded: tools,
-    timestamp: 2,
-  });
-  await complete(context);
-  assert.equal(requests.length, 1);
-  const request = requests[0];
-  const promptFields = request.filter((field) => field.num === 2);
-  assert.equal(promptFields.length, 1);
-  assert.equal(promptFields[0].wire, 2);
-  assert.equal(promptFields[0].value.toString(), "Initial instructions.\n\nUpdated instructions.\n\nDo not edit.");
-  const history = messages(request);
-  assert.equal(history.length, 1);
-  assert.equal(history[0].find((field) => field.num === 2).value, 1n);
-  assert.equal(stringField(history[0], 3), text);
-  const images = history[0].filter((field) => field.num === 10);
-  assert.equal(images.length, 1);
-  assert.equal(stringField(fields(images[0].value), 1), image.data);
-  assert.equal(stringField(fields(images[0].value), 2), image.mimeType);
-  assert.deepEqual(request.filter((field) => field.num === 10).map((field) => {
-    const definition = fields(field.value);
-    return { name: stringField(definition, 1), description: stringField(definition, 2), parameters: JSON.parse(stringField(definition, 3)) };
-  }), tools);
-});
-
-test("omits an empty system prompt while preserving tools and user content", async (t) => {
-  const requests = mockDevin(t);
-  await complete(normalizeContext({ systemPrompt: "", tools: [readTool], messages: [user("Read probe.txt")] }));
-  assert.equal(requests[0].filter((field) => field.num === 2).length, 0);
-  assert.equal(requests[0].filter((field) => field.num === 10).length, 1);
-  assert.deepEqual(messages(requests[0]).map((message) => stringField(message, 3)), ["Read probe.txt"]);
-});
-
-test("sends a system-only transcript without inventing a user message", async (t) => {
-  const requests = mockDevin(t);
-  await complete(normalizeContext({ systemPrompt: "System-only instructions.", messages: [] }));
-  assert.equal(stringField(requests[0], 2), "System-only instructions.");
-  assert.deepEqual(messages(requests[0]), []);
-});
-
-test("encodes an empty transcript without a prompt or history", async (t) => {
-  const requests = mockDevin(t);
-  await complete(normalizeContext({ messages: [] }));
-  assert.equal(stringField(requests[0], 2), undefined);
-  assert.deepEqual(messages(requests[0]), []);
-});
-
-test("retains the system prompt and replays assistant tool calls and results on the next request", async (t) => {
-  const callId = "call-read";
-  const args = { path: "probe.txt" };
-  const requests = mockDevin(t, [
-    Buffer.concat([
-      encodeMessage(6, Buffer.concat([encodeString(1, callId), encodeString(2, "read"), encodeString(3, JSON.stringify(args))])),
-      encodeVarintField(5, 10),
-    ]),
-    Buffer.concat([encodeString(3, "file-value"), encodeVarintField(5, 0)]),
-  ]);
-  const context = normalizeContext({ systemPrompt: "Use tools; never guess.", tools: [readTool], messages: [user("Read probe.txt")] });
-  const first = await complete(context);
-  assert.equal(first.stopReason, "toolUse");
-  assert.deepEqual(first.content, [{ type: "toolCall", id: callId, name: "read", arguments: args }]);
-  context.messages.push(first, {
-    role: "toolResult", toolCallId: callId, toolName: "read",
-    content: [{ type: "text", text: "file-value" }], isError: false, timestamp: 3,
-  });
-  const second = await complete(context);
-  assert.equal(second.stopReason, "stop");
-  assert.equal(second.content[0].text, "file-value");
-  assert.equal(requests.length, 2);
-  for (const request of requests) {
-    assert.equal(request.filter((field) => field.num === 2).length, 1);
-    assert.equal(stringField(request, 2), "Use tools; never guess.");
-  }
-  const history = messages(requests[1]);
-  assert.deepEqual(history.map((message) => message.find((field) => field.num === 2).value), [1n, 2n, 4n]);
-  assert.equal(stringField(history[0], 3), "Read probe.txt");
-  const call = fields(history[1].find((field) => field.num === 6).value);
-  assert.equal(stringField(call, 1), callId);
-  assert.equal(stringField(call, 2), "read");
-  assert.deepEqual(JSON.parse(stringField(call, 3)), args);
-  assert.equal(stringField(history[2], 7), callId);
-  assert.equal(stringField(history[2], 3), "file-value");
-});

tsconfig.json

@@ -1,12 +0,0 @@
-{
-  "compilerOptions": {
-    "target": "ES2022",
-    "module": "NodeNext",
-    "moduleResolution": "NodeNext",
-    "strict": true,
-    "skipLibCheck": true,
-    "noEmit": true,
-    "types": ["node"]
-  },
-  "include": ["extensions/**/*.ts", "src/**/*.ts"]
-}