@aexhq/agentloop-pi 1.6.0 → 2.0.0

package/README.md

@@ -2,28 +2,31 @@
 
 A pi-style agent loop for Brain: a semantic port of the pi coding agent's loop,
 pinned against [earendil-works/pi](https://github.com/earendil-works/pi) tag
-`v0.84.4` (`@earendil-works/pi-agent-core@0.84.4`). The loop drives each turn
-through Brain's services from inside its WebAssembly component and reproduces
+`v0.84.4` (`@earendil-works/pi-agent-core@0.84.4`). The package ships a precompiled
+WebAssembly Component that drives each turn through Brain's Agentloop host imports and reproduces
 pi's per-turn contract:
 
-- tool calls are issued as one **parallel batch**, and results return in
-  assistant source order;
-- a `length`-stopped response that carries tool calls **fails the whole batch
-  without executing it** and re-asks the model;
+- tool calls are issued as one **parallel batch**, and results return in assistant source order;
+- a `length`-stopped response that carries tool calls **fails the whole batch without executing
+  it** and re-asks the model;
 - **automatic compaction**: when the estimated context exceeds
-  `contextWindow - reserveTokens` (default 16384), history older than
-  ~`keepRecentTokens` (default 20000) is summarized into pi's structured
-  context checkpoint (`## Goal` … `## Critical Context`) that replaces it.
+  `contextWindow - reserveTokens` (default 16384), history older than ~`keepRecentTokens`
+  (default 20000) is summarized into pi's structured context checkpoint.
 
-pi's steering and follow-up queues and per-tool `executionMode` are host-app
-seams with no Brain equivalent and are not ported.
+pi's steering and follow-up queues and per-tool `executionMode` are host-app seams with no Brain
+equivalent and are not ported.
 
 ```ts
+import { brainWasm } from "@aexhq/brain";
 import { pi } from "@aexhq/agentloop-pi";
 
+const loopRuntime = brainWasm();
 const session = await brain.sessions.create({
-  agentloop: pi({ contextWindow: 200_000 }),
+  agentloop: pi({ env: loopRuntime, contextWindow: 200_000 }),
   model,
   tools: [read({ env: workspace })],
 });
 ```
+
+The component is built by this package's publisher. Brain consumes the resulting Component and
+does not compile its JavaScript source.

package/dist/index.mjs

@@ -1,5 +1,18 @@
-import { installExtensionIdentity } from "@aexhq/brain";
-import * as definitions from "./source.mjs";
-
-export const pi = definitions.pi;
-installExtensionIdentity(pi, "pi", new URL("./pi.brain.json", import.meta.url));
+// src/index.mjs
+import { agentloop, component } from "@aexhq/brain";
+import { z } from "zod";
+var options = z.object({
+  contextWindow: z.number().int().positive().default(2e5),
+  // pi defaults (compaction.ts): compact when context exceeds
+  // contextWindow - reserveTokens, keep ~keepRecentTokens of recent messages.
+  reserveTokens: z.number().int().positive().default(16384),
+  keepRecentTokens: z.number().int().positive().default(2e4),
+  compaction: z.boolean().default(true)
+}).strict();
+var pi = agentloop({
+  options,
+  implementation: component(new URL("./loop.component.wasm", import.meta.url))
+});
+export {
+  pi
+};

package/index.d.ts

@@ -1,6 +1,7 @@
-import type { Agentloop } from "@aexhq/brain";
+import type { AgentloopBinding, Environment } from "@aexhq/brain";
 
 export interface PiOptions {
+  readonly env: Environment;
   /** Model context window in tokens the compaction budget is measured against. Default 200000. */
   readonly contextWindow?: number;
   /** Compact when the estimated context exceeds contextWindow - reserveTokens. pi default 16384. */
@@ -11,4 +12,4 @@ export interface PiOptions {
   readonly compaction?: boolean;
 }
 
-export declare const pi: (options?: PiOptions) => Agentloop;
+export declare const pi: (options: PiOptions) => AgentloopBinding;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aexhq/agentloop-pi",
-  "version": "1.6.0",
+  "version": "2.0.0",
   "description": "Pi-style parallel-Tool Brain extension",
   "license": "MIT",
   "repository": {
@@ -15,7 +15,7 @@
   },
   "publishConfig": {
     "access": "public",
-    "provenance": true,
+    "provenance": false,
     "tag": "next"
   },
   "exports": {
@@ -29,12 +29,12 @@
     "README.md"
   ],
   "scripts": {
-    "build": "brain build src/index.mjs --out dist",
+    "build": "node ../../tools/build-agentloop.mjs src/component.mjs src/index.mjs dist",
     "test": "npm run build && node --test",
     "prepack": "npm run build"
   },
   "dependencies": {
-    "@aexhq/brain": "0.15.1",
+    "@aexhq/brain": "0.16.0",
     "zod": "4.4.3"
   },
   "devDependencies": {}

package/src/component.mjs

@@ -0,0 +1,47 @@
+import * as host from "brain:agentloop/host@0.1.0";
+
+import { runPi } from "./logic.mjs";
+
+function hosted(call) {
+  try {
+    return call();
+  } catch (error) {
+    const payload = error !== null && typeof error === "object" && "payload" in error ? error.payload : undefined;
+    const failure = new Error(payload?.message ?? String(error?.message ?? error));
+    failure.code = payload?.code ?? "host_error";
+    failure.retryable = payload?.retryable ?? false;
+    throw failure;
+  }
+}
+
+export async function turn(input) {
+  try {
+    const output = await runPi({
+      input: JSON.parse(input.inputJson),
+      transcript: JSON.parse(input.transcriptJson),
+      slots: JSON.parse(input.slotsJson),
+      events: JSON.parse(input.eventsJson),
+      configuration: JSON.parse(input.configurationJson),
+      system: input.system,
+      tools: JSON.parse(input.toolsJson),
+    }, {
+      model: (request) => JSON.parse(hosted(() => host.model(JSON.stringify(request)))),
+      dispatch: (calls) => JSON.parse(hosted(() => host.dispatch(JSON.stringify(calls)))),
+      emit: (kind, data) => hosted(() => host.emit(kind, JSON.stringify(data ?? null))),
+      telemetry: (record) => host.telemetry(JSON.stringify(record ?? null)),
+    });
+    return {
+      transcriptJson: JSON.stringify(output.transcript),
+      slotsJson: JSON.stringify(output.slots),
+      resultJson: output.result === undefined ? undefined : JSON.stringify(output.result),
+    };
+  } catch (error) {
+    const failure = new Error(String(error?.message ?? error) || "Agentloop turn failed");
+    failure.payload = {
+      code: typeof error?.code === "string" && error.code.length > 0 ? error.code : "agentloop_failed",
+      message: failure.message,
+      retryable: Boolean(error?.retryable),
+    };
+    throw failure;
+  }
+}

package/src/index.mjs

@@ -1,189 +1,15 @@
-import { agentloop } from "@aexhq/brain";
+import { agentloop, component } from "@aexhq/brain";
 import { z } from "zod";
-
-// Semantic port of the pi coding agent's loop, pinned against
-// earendil-works/pi tag v0.84.4 (@earendil-works/pi-agent-core@0.84.4,
-// packages/agent/src/agent-loop.ts + packages/coding-agent/src/core/compaction/compaction.ts).
-//
-// The loop drives one whole turn through Brain's services: it calls the model,
-// hands tool calls to Brain, and edits the transcript Brain persists. What it
-// reproduces of pi:
-// - tool calls are issued as one parallel batch, results return as one
-//   user message of tool_result blocks in assistant source order
-// - a length-stop that carries tool calls fails the whole batch without
-//   executing it, and the loop re-asks the model
-// - automatic compaction: when the estimated context exceeds
-//   contextWindow - reserveTokens, everything older than ~keepRecentTokens
-//   is summarized into a structured context checkpoint that replaces it
-//   (cut points never split a tool_result away from its tool_use)
-// Host-app seams of pi (steering and follow-up queues, per-tool
-// executionMode) have no Brain equivalent and are not ported.
-
-const optionsSchema = z.object({
+const options = z.object({
   contextWindow: z.number().int().positive().default(200_000),
   // pi defaults (compaction.ts): compact when context exceeds
   // contextWindow - reserveTokens, keep ~keepRecentTokens of recent messages.
   reserveTokens: z.number().int().positive().default(16_384),
   keepRecentTokens: z.number().int().positive().default(20_000),
   compaction: z.boolean().default(true),
-}).default({});
-
-// The checkpoint is the transcript's first message once a compaction has
-// happened; the slot remembers its text so the next compaction can update it.
-const checkpointSchema = z.object({ summary: z.union([z.string(), z.null()]) });
-
-const CHECKPOINT_PREFIX = "Context checkpoint from earlier in this conversation:\n\n";
-
-const SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.
-
-Use this EXACT format:
-
-## Goal
-
-## Constraints & Preferences
-
-## Progress
-
-### Done
-
-### In Progress
-
-### Blocked
-
-## Key Decisions
-
-## Next Steps
-
-## Critical Context
-
-Keep each section concise. Preserve exact file paths, function names, and error messages.`;
-
-const UPDATE_RULES = `Update the existing structured summary with new information.
-
-RULES:
-- PRESERVE all existing information from the previous summary
-- ADD new progress, decisions, and context from the new messages
-- Move items from In Progress to Done as they complete
-
-`;
-
-const TRUNCATED_CALL_MESSAGE =
-  "Tool call was not executed: the model response was cut off by the output token limit, so the arguments may be truncated. Re-issue the tool call.";
-
-const text = (message) =>
-  message.content
-    .filter((block) => block.type === "text")
-    .map((block) => block.text)
-    .join("");
-
-const blockText = (block) => {
-  if (block.type === "text") return block.text;
-  if (block.type === "tool_use") return `[tool_use ${block.name}] ${JSON.stringify(block.input)}`;
-  const output = typeof block.content === "string" ? block.content : JSON.stringify(block.content);
-  return `[tool_result${block.is_error ? " (error)" : ""}] ${output.length > 2000 ? `${output.slice(0, 2000)}…` : output}`;
-};
-
-const serializeConversation = (messages) =>
-  messages.map((message) => `${message.role}:\n${message.content.map(blockText).join("\n")}`).join("\n\n");
-
-export const pi = agentloop({ options: optionsSchema }, (author) => {
-  const options = author.options;
-  const checkpoint = author.slot("checkpoint", checkpointSchema, () => ({ summary: null }));
-
-  // The conversation proper: the transcript minus the checkpoint message.
-  const body = (transcript) => (checkpoint.summary === null ? transcript : transcript.slice(1));
-
-  const shouldCompact = (transcript) =>
-    options.compaction && author.context.estimateTokens(transcript) > options.contextWindow - options.reserveTokens;
-
-  // Earliest index the recent-token budget allows, then forward to the next
-  // boundary that does not split a tool_result away from its tool_use.
-  const cutPoint = (messages) => {
-    let kept = 0;
-    let cut = 0;
-    for (let index = messages.length - 1; index >= 0; index -= 1) {
-      kept += author.context.estimateTokens([messages[index]]);
-      if (kept > options.keepRecentTokens) {
-        cut = index + 1;
-        break;
-      }
-    }
-    while (
-      cut < messages.length &&
-      messages[cut].role === "user" &&
-      messages[cut].content.some((block) => block.type === "tool_result")
-    ) {
-      cut += 1;
-    }
-    return cut;
-  };
-
-  const compact = async (turn) => {
-    const messages = body(turn.transcript);
-    const cut = cutPoint(messages);
-    if (cut === 0) return;
-    const previous = checkpoint.summary === null ? "" : `Previous summary:\n\n${checkpoint.summary}\n\n`;
-    const prompt = checkpoint.summary === null ? SUMMARIZATION_PROMPT : `${UPDATE_RULES}${SUMMARIZATION_PROMPT}`;
-    const { message } = await turn.model({
-      messages: [
-        {
-          role: "user",
-          content: [{ type: "text", text: `${previous}${serializeConversation(messages.slice(0, cut))}\n\n${prompt}` }],
-        },
-      ],
-    });
-    checkpoint.summary = text(message);
-    turn.transcript.splice(
-      0,
-      turn.transcript.length,
-      { role: "user", content: [{ type: "text", text: `${CHECKPOINT_PREFIX}${checkpoint.summary}` }] },
-      ...messages.slice(cut),
-    );
-  };
+}).strict();
 
-  author.turn(async (turn) => {
-    turn.transcript.push({ role: "user", content: [{ type: "text", text: turn.input.message }] });
-    for (;;) {
-      if (shouldCompact(turn.transcript)) await compact(turn);
-      const { message, stop_reason } = await turn.model({ messages: turn.transcript });
-      turn.transcript.push(message);
-      const calls = message.content
-        .filter((block) => block.type === "tool_use")
-        .map((block) => ({ callId: block.id, name: block.name, input: block.input }));
-      if (calls.length === 0) {
-        await turn.reply(text(message));
-        return turn.done();
-      }
-      if (stop_reason === "max_tokens") {
-        // pi fails the whole batch without executing it when the response was
-        // cut off: the arguments cannot be trusted.
-        turn.transcript.push({
-          role: "user",
-          content: calls.map((call) => ({
-            type: "tool_result",
-            tool_use_id: call.callId,
-            content: TRUNCATED_CALL_MESSAGE,
-            is_error: true,
-          })),
-        });
-        continue;
-      }
-      // One dispatch: Brain runs the batch in parallel and reports back once.
-      const results = await turn.dispatch(calls);
-      const byCall = new Map(results.map((result) => [result.callId, result]));
-      turn.transcript.push({
-        role: "user",
-        // Assistant source order, the order pi appends result messages in.
-        content: calls.map(({ callId }) => {
-          const result = byCall.get(callId);
-          return {
-            type: "tool_result",
-            tool_use_id: callId,
-            content: result === undefined ? "Tool produced no result." : result.output,
-            is_error: result === undefined ? true : result.isError,
-          };
-        }),
-      });
-    }
-  });
+export const pi = agentloop({
+  options,
+  implementation: component(new URL("./loop.component.wasm", import.meta.url)),
 });

package/src/logic.mjs

@@ -0,0 +1,130 @@
+const CHECKPOINT_PREFIX = "Context checkpoint from earlier in this conversation:\n\n";
+
+const SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.
+
+Use this EXACT format:
+
+## Goal
+
+## Constraints & Preferences
+
+## Progress
+
+### Done
+
+### In Progress
+
+### Blocked
+
+## Key Decisions
+
+## Next Steps
+
+## Critical Context
+
+Keep each section concise. Preserve exact file paths, function names, and error messages.`;
+
+const UPDATE_RULES = `Update the existing structured summary with new information.
+
+RULES:
+- PRESERVE all existing information from the previous summary
+- ADD new progress, decisions, and context from the new messages
+- Move items from In Progress to Done as they complete
+
+`;
+
+const TRUNCATED_CALL_MESSAGE =
+  "Tool call was not executed: the model response was cut off by the output token limit, so the arguments may be truncated. Re-issue the tool call.";
+
+const estimateTokens = (messages) => Math.ceil(JSON.stringify(messages).length / 4);
+const text = (message) => message.content.filter((block) => block.type === "text").map((block) => block.text).join("");
+
+const blockText = (block) => {
+  if (block.type === "text") return block.text;
+  if (block.type === "tool_use") return `[tool_use ${block.name}] ${JSON.stringify(block.input)}`;
+  const output = typeof block.content === "string" ? block.content : JSON.stringify(block.content);
+  return `[tool_result${block.is_error ? " (error)" : ""}] ${output.length > 2000 ? `${output.slice(0, 2000)}…` : output}`;
+};
+
+const serializeConversation = (messages) =>
+  messages.map((message) => `${message.role}:\n${message.content.map(blockText).join("\n")}`).join("\n\n");
+
+export async function runPi(input, context) {
+  const options = {
+    contextWindow: 200_000,
+    reserveTokens: 16_384,
+    keepRecentTokens: 20_000,
+    compaction: true,
+    ...input.configuration,
+  };
+  const transcript = cloneJson(input.transcript);
+  const saved = input.slots.checkpoint;
+  const checkpoint = saved === undefined ? { summary: null } : cloneJson(saved);
+  const body = () => checkpoint.summary === null ? transcript : transcript.slice(1);
+  const shouldCompact = () =>
+    options.compaction && estimateTokens(transcript) > options.contextWindow - options.reserveTokens;
+  const cutPoint = (messages) => {
+    let kept = 0;
+    let cut = 0;
+    for (let index = messages.length - 1; index >= 0; index -= 1) {
+      kept += estimateTokens([messages[index]]);
+      if (kept > options.keepRecentTokens) {
+        cut = index + 1;
+        break;
+      }
+    }
+    while (cut < messages.length && messages[cut].role === "user" && messages[cut].content.some((block) => block.type === "tool_result")) cut += 1;
+    return cut;
+  };
+  const compact = async () => {
+    const messages = body();
+    const cut = cutPoint(messages);
+    if (cut === 0) return;
+    const previous = checkpoint.summary === null ? "" : `Previous summary:\n\n${checkpoint.summary}\n\n`;
+    const prompt = checkpoint.summary === null ? SUMMARIZATION_PROMPT : `${UPDATE_RULES}${SUMMARIZATION_PROMPT}`;
+    const { message } = await context.model({
+      messages: [{ role: "user", content: [{ type: "text", text: `${previous}${serializeConversation(messages.slice(0, cut))}\n\n${prompt}` }] }],
+    });
+    checkpoint.summary = text(message);
+    transcript.splice(0, transcript.length,
+      { role: "user", content: [{ type: "text", text: `${CHECKPOINT_PREFIX}${checkpoint.summary}` }] },
+      ...messages.slice(cut));
+  };
+
+  transcript.push({ role: "user", content: [{ type: "text", text: input.input.message }] });
+  for (;;) {
+    if (shouldCompact()) await compact();
+    const { message, stop_reason } = await context.model({ messages: transcript });
+    transcript.push(message);
+    const calls = message.content
+      .filter((block) => block.type === "tool_use")
+      .map((block) => ({ call_id: block.id, name: block.name, input: block.input }));
+    if (calls.length === 0) {
+      await context.emit("output_emitted", { type: "assistant_message", message: text(message) });
+      return { transcript, slots: { checkpoint } };
+    }
+    if (stop_reason === "max_tokens") {
+      transcript.push({
+        role: "user",
+        content: calls.map((call) => ({ type: "tool_result", tool_use_id: call.call_id, content: TRUNCATED_CALL_MESSAGE, is_error: true })),
+      });
+      continue;
+    }
+    const results = await context.dispatch(calls);
+    const byCall = new Map(results.map((result) => [result.call_id, result]));
+    transcript.push({
+      role: "user",
+      content: calls.map(({ call_id }) => {
+        const result = byCall.get(call_id);
+        return {
+          type: "tool_result",
+          tool_use_id: call_id,
+          content: result === undefined ? "Tool produced no result." : result.output,
+          is_error: result === undefined ? true : result.is_error,
+        };
+      }),
+    });
+  }
+}
+
+const cloneJson = (value) => JSON.parse(JSON.stringify(value));