@tranhoangnguyen0310/pi-flow-external 2.1.0-external.0 → 2.1.0-external.2

package/CHANGELOG.md

@@ -4,6 +4,20 @@ All notable changes to pi-flow external are documented here.
 
 ## Unreleased
 
+## [2.1.0-external.2] - 2026-09-18
+
+### Fixed
+
+- Preserved and exposed available partial assistant output on failure, cancellation, and timeout across all backend adapters (`codex`, `claude`, `agy`, and in-process `pi`). Unsuccessful terminal receipts retain the failure status and error diagnostic while appending a bounded `Interrupted output:` preview.
+- Recovered interrupted assistant output in run evidence inspection (`external_runs(action: "inspect", view: "output")`) from completed non-done summary documents, including Pi SDK child runs.
+- `ChildRunError` preserves `partialOutput` and `assistantOutput` for workflow scripts to inspect, and uncaught child errors surface the interrupted output preview in workflow failure receipts.
+- Hardened CI release workflow polling loop against npm registry replication latency (increased timeout to 150s).
+
+### Fixed
+
+- Coordinator guidance now names all four harnesses (`Agent` description, `buildCoordinatorPrompt` body): registered Pi harnesses are first-class alongside Claude Code/Codex CLI/Antigravity, not "external CLIs only" leftovers of the pre-2.1.0 copy. `test/agent-contract`/`agent-rendering` pin the copy.
+- Parent guidance states the blocking default explicitly (`background:true` is opt-in, use `external_runs` on the returned handle) and how same-harness parallel work actually shares the one `maxConcurrentSubagents` limiter: workflow `parallel([() => agent(...), ...])` or separate background `Agent` calls in one turn; sequential awaits stay serial. Includes a three-`agy` workflow example plus fore/background meta examples in `external_help`. Pinned by a new `test/external-help` regression.
+
 ## [2.1.0-external.0] - 2026-09-15
 
 Named Pi harness configurations (`pi-*`): a fourth delegation surface running in-process through Pi's own SDK, selectable exactly like `agy`/`claude`/`codex`. Design: `docs/plans/pi-named-configurations-design.md`; plan: `docs/plans/pi-named-configurations-implementation-plan.md`.

package/CONTEXT.md

@@ -32,6 +32,8 @@ The split is global and intentional to avoid tool ambiguity across projects. The
 
 User operations are namespaced under `/external`: status, Doctor, settings, profiles, profile creation, workflows, interactive run navigation, durable run summary/pruning, and help. `/external runs` uses the same registry/readers as `external_runs`, including every cursor-paged workflow child and output/diagnostic page. Extension-owned concurrency, timeout, and global default-harness settings live in `$PI_CODING_AGENT_DIR/pi-flow-external/settings.json`; profiles remain authoritative for downstream backend/model/thinking metadata. Named Pi harness configurations live in the sibling `pi-flow-external/harnesses.json` registry. A trusted project may override only `defaultHarness`, which may itself name a registered `pi-*` harness.
 
+> Next agent: full implementation map is `docs/ARCHITECTURE_SNAPSHOT.md` (verified 2026-09-16). Keep this file and `AGENTS.md` as the persistent breadcrumbs.
+
 ## Design stance
 
 Guardrails exist to keep lanes from bleeding into each other (a reviewer that edits, a worker that fixes), not to constrain how a model works. Role bodies stay short: define the job, state the boundary, then get out of the way and trust the model's judgment on approach, depth, and method. The generic `worker` role covers non-coding tasks with no method constraints at all. The shipped default roster is five code-oriented roles (explorer, planner, implementer, reviewer, qa) plus `worker`; other roles are user-created through `/external profile create` and still receive the name-based execution permission floor. External agents must not be arbitrarily handcuffed just to satisfy a theoretical safety checklist: what transpires on the ground is what matters. When an external agent needs heightened permissions to inspect, run tests, and execute tools, the system defaults to providing that authority rather than blocking execution.

package/README.md

@@ -253,7 +253,7 @@ Workflows show access once at the workflow level, retain done/active/queued/fail
 
 ## Background runs and supervision
 
-`Agent` and `workflow` are blocking by default. Add `background: true` to return a stable `run_...` or `wf_...` handle after validation and registration while the originating Pi session continues to own the work:
+`Agent` and `workflow` are blocking by default — an ordinary call waits for its child and returns the result, run everything this way unless the parent has other work to do first. Add `background: true` only when the parent can productively proceed before completion; it returns a stable `run_...` or `wf_...` handle after validation and registration while the originating Pi session continues to own the work:
 
 ```ts
 Agent({
@@ -275,7 +275,19 @@ Interrupting a blocking `Agent`/`workflow` call cancels its work. Interrupting `
 
 ## Workflow usage
 
-The `workflow` tool runs trusted JavaScript that calls one or more external roles and returns a JSON-serializable result. Its first statement must be the current declaration `export const meta = { apiVersion: 1, name, description }`; missing or unsupported versions fail before any child launches. Every `agent()` child uses the same `role`/optional `harness` resolution as direct `Agent` calls, with legacy exact `subagent_type` also supported.
+The `workflow` tool runs trusted JavaScript that calls one or more external roles and returns a JSON-serializable result. Workflows are for explicit fan-out — and for the case this extension was built to make explicit: running several children on the **same** harness at once. Example:
+
+```js
+export const meta = { apiVersion: 1, name: "triage", description: "Three parallel agy reviewers" };
+const [a, b, c] = await parallel([
+  () => agent("Review slice A read-only: /absolute/repo", { label: "rev-a", role: "reviewer", harness: "agy" }),
+  () => agent("Review slice B read-only: /absolute/repo", { label: "rev-b", role: "reviewer", harness: "agy" }),
+  () => agent("Review slice C read-only: /absolute/repo", { label: "rev-c", role: "reviewer", harness: "agy" }),
+]);
+return { a, b, c };
+```
+
+Sequential `await agent(...)` stays serial by construction; `parallel([...])` is how parallel children share the single `maxConcurrentSubagents` limiter at once. Top-level direct `Agent` calls issued together also run together (`executionMode: "parallel"`, default under one limiter). `external_runs wait` and `external_runs inspect` cover both surfaces. Its first statement must be the current declaration `export const meta = { apiVersion: 1, name, description }`; missing or unsupported versions fail before any child launches. Every `agent()` child uses the same `role`/optional `harness` resolution as direct `Agent` calls, with legacy exact `subagent_type` also supported.
 
 Saved workflows are discovered on demand through `external_help({ topic: "workflow" })`; project `.pi/workflows` entries are included only when Pi reports the project trusted.
 
@@ -285,7 +297,7 @@ Example request:
 Use the workflow tool to ask role "explorer" on harness "claude" for an architecture map and role "reviewer" on harness "codex" for a risk review, then synthesize their findings.
 ```
 
-Direct `Agent` calls and workflow children share the same concurrency and timeout controls. Workflow `agent(prompt, { role, context: { mode: "recent", turns: 5 } })` accepts the same context modes. Every current-version `agent()` returns its value or throws a catchable `ChildRunError` with `runId`, `outcome` (`failed`, `cancelled`, or `timed_out`), `message`, and output/diagnostic references. Catch an optional failure explicitly; an uncaught child error fails the workflow and drains active siblings. `parallel` and `pipeline` preserve this contract and never convert failure to `null`.
+Direct `Agent` calls and workflow children share one concurrency limiter (`maxConcurrentSubagents`, default 12) and the same timeout controls. There is no per-harness cap: three `agy` children run together when started together — `parallel([() => agent(..., { harness: "agy" }), ...])` in a workflow, or three `background: true` `Agent` calls in one turn followed by `external_runs wait`. Sequentially `await`ed children stay serial by construction. Direct calls default `executionMode: "parallel"` so top-level `Agent` calls issued together run together. Workflow `agent(prompt, { role, context: { mode: "recent", turns: 5 } })` accepts the same context modes. Every current-version `agent()` returns its value or throws a catchable `ChildRunError` with `runId`, `outcome` (`failed`, `cancelled`, or `timed_out`), `message`, and output/diagnostic references. Catch an optional failure explicitly; an uncaught child error fails the workflow and drains active siblings. `parallel` and `pipeline` preserve this contract and never convert failure to `null`.
 
 Every child selects from one parent snapshot and effective settings frozen at workflow invocation; earlier child results must still be passed explicitly. `resumeFromRunId` is explicit replay for a persisted `scriptPath`: it reuses only the longest unchanged prefix of successful child calls. The API version, transferred context, prompt, selection, and relevant options participate in fingerprints. The first changed, failed, cancelled, or timed-out call and everything after it executes again. Recomposition is cheap, but rerun children may cost money or repeat side effects; the runtime never retries or replays a repaired script automatically.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tranhoangnguyen0310/pi-flow-external",
-  "version": "2.1.0-external.0",
+  "version": "2.1.0-external.2",
   "description": "External Claude Code, Codex CLI, and Antigravity delegation for pi.",
   "type": "module",
   "main": "./index.ts",

package/src/core/agy.ts

@@ -1,6 +1,6 @@
 import { spawn, type ChildProcess } from "node:child_process";
 import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
-import { assistantOutput, createProgressEmitter, textResult, type AgentToolResult } from "./progress.ts";
+import { assistantOutput, createProgressEmitter, formatInterruptedOutputPreview, textResult, type AgentToolResult } from "./progress.ts";
 import {
   createBoundedBuffer,
   MAX_STDERR_CHARS,
@@ -486,7 +486,8 @@ export async function spawnAgySubagent(params: {
     if (child) abortChildTree(child);
     const message = error instanceof Error ? error.message : String(error);
     const status = params.signal?.aborted ? "aborted" : "error";
-    const output = assistantOutput(assistantMessages, "interrupted");
+    const candidate = terminalResult ? textFromAgyResult(terminalResult)?.trim() : undefined;
+    const output = assistantOutput(assistantMessages, "interrupted", candidate);
     params.onUsage(latestUsage);
     if (progress) {
       progress.status = status;
@@ -495,7 +496,8 @@ export async function spawnAgySubagent(params: {
       progress.assistantOutput = output;
       progress.endedAt = Date.now();
     }
-    return textResult(`Subagent "${params.description}" (${subagentType}) ${status === "aborted" ? "aborted" : "failed"}: ${message}`, {
+    const preview = formatInterruptedOutputPreview(output);
+    return textResult(`Subagent "${params.description}" (${subagentType}) ${status === "aborted" ? "aborted" : "failed"}: ${message}${preview}`, {
       description: params.description,
       subagentType,
       backend: params.profile.backend,

package/src/core/claude.ts

@@ -3,6 +3,7 @@ import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
 import {
   assistantOutput,
   createProgressEmitter,
+  formatInterruptedOutputPreview,
   textResult,
   type AgentToolResult,
 } from "./progress.ts";
@@ -570,7 +571,7 @@ export async function spawnClaudeSubagent(params: {
     if (child) abortChildTree(child);
     const message = error instanceof Error ? error.message : String(error);
     const status = params.signal?.aborted ? "aborted" : "error";
-    const output = assistantOutput(assistantMessages, "interrupted");
+    const output = assistantOutput(assistantMessages, "interrupted", resultText);
     params.onUsage(latestUsage);
     if (progress) {
       progress.status = status;
@@ -580,7 +581,8 @@ export async function spawnClaudeSubagent(params: {
       progress.endedAt = Date.now();
     }
     const verb = status === "aborted" ? "aborted" : "failed";
-    return textResult(`Subagent "${params.description}" (${subagentType}) ${verb}: ${message}`, {
+    const preview = formatInterruptedOutputPreview(output);
+    return textResult(`Subagent "${params.description}" (${subagentType}) ${verb}: ${message}${preview}`, {
       description: params.description,
       subagentType,
       backend: params.profile.backend,

package/src/core/codex.ts

@@ -6,6 +6,7 @@ import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
 import {
   assistantOutput,
   createProgressEmitter,
+  formatInterruptedOutputPreview,
   textResult,
   type AgentToolResult,
 } from "./progress.ts";
@@ -533,7 +534,7 @@ export async function spawnCodexSubagent(params: {
     if (child) abortChildTree(child);
     const message = error instanceof Error ? error.message : String(error);
     const status = params.signal?.aborted ? "aborted" : "error";
-    const output = assistantOutput(assistantMessages, "interrupted");
+    const output = assistantOutput(assistantMessages, "interrupted", resultText);
     params.onUsage(latestUsage);
     if (progress) {
       progress.status = status;
@@ -543,7 +544,8 @@ export async function spawnCodexSubagent(params: {
       progress.endedAt = Date.now();
     }
     const verb = status === "aborted" ? "aborted" : "failed";
-    return textResult(`Subagent "${params.description}" (${subagentType}) ${verb}: ${message}`, {
+    const preview = formatInterruptedOutputPreview(output);
+    return textResult(`Subagent "${params.description}" (${subagentType}) ${verb}: ${message}${preview}`, {
       description: params.description,
       subagentType,
       backend: params.profile.backend,

package/src/core/progress.ts

@@ -51,6 +51,12 @@ export function assistantOutput(
   return output.length ? { status, messages: output } : undefined;
 }
 
+export function formatInterruptedOutputPreview(output: SubagentAssistantOutput | undefined): string {
+  if (!output?.messages?.length) return "";
+  const text = output.messages.map((m) => m.text.trim()).filter(Boolean).join("\n\n");
+  return text ? `\n\nInterrupted output:\n${text.slice(-4_000)}` : "";
+}
+
 export function extractAssistantMessages(messages: readonly unknown[]): Array<{ id?: string; text: string }> {
   return messages.flatMap((value) => {
     const message = value as { id?: unknown; role?: unknown; content?: unknown };
@@ -88,7 +94,7 @@ function getFirstTextLine(text: string): string {
   return text.split("\n").find((line) => line.trim()) ?? text;
 }
 
-function extractTextContent(content: unknown): string {
+export function extractTextContent(content: unknown): string {
   if (typeof content === "string") {
     return content;
   }

package/src/core/run-inspection.ts

@@ -5,6 +5,7 @@ import { join } from "node:path";
 import { agyActivityFromEvent } from "./agy.ts";
 import { claudeActivityFromEvent, extractClaudeFinalText } from "./claude.ts";
 import { codexActivityFromEvent, extractCodexFinalText } from "./codex.ts";
+import { extractTextContent } from "./progress.ts";
 
 const CURSOR_VERSION = 1;
 const DEFAULT_PAGE_BYTES = 32 * 1024;
@@ -117,7 +118,7 @@ export async function inspectRun({
   }
 
   if (decoded?.kind === "inspect" && decoded.source === "summary") throw new Error("Cursor source does not match inspection view");
-  if (view === "output" && summary.document && terminalStatus === "done") {
+  if (view === "output" && summary.document && (terminalStatus === "done" || assistantItems(terminal).length > 0)) {
     if (decoded?.kind === "inspect" && decoded.source === "events") throw staleCursorError();
     const observation = await readObservation(eventsPath, runId);
     const items = terminalOutputItems(terminal);
@@ -347,7 +348,7 @@ function summaryProjection(runId: string, summary: SummaryState, observation: Ru
   };
 }
 
-function outputFromEvent(event: EvidenceEvent): ({ kind: "claude" | "codex" | "agy"; id?: string; text: string }) | undefined {
+function outputFromEvent(event: EvidenceEvent): ({ kind: "claude" | "codex" | "agy" | "pi"; id?: string; text: string }) | undefined {
   if (event.type !== "backend_event") return undefined;
   const envelope = asRecord(event.data);
   const backendEvent = asRecord(envelope?.event);
@@ -368,6 +369,14 @@ function outputFromEvent(event: EvidenceEvent): ({ kind: "claude" | "codex" | "a
     const id = asString(update?.step_id);
     return text ? { kind: "agy", ...(id ? { id } : {}), text } : undefined;
   }
+  if (envelope?.backend === "pi" && backendEvent.type === "message_end") {
+    const message = asRecord(backendEvent.message);
+    if (message?.role === "assistant") {
+      const text = extractTextContent(message.content);
+      const id = asString(message.id);
+      return text ? { kind: "pi", ...(id ? { id } : {}), text } : undefined;
+    }
+  }
   return undefined;
 }
 

package/src/core/spawn.ts

@@ -13,6 +13,7 @@ import {
   extractFinalAssistantText,
   extractAssistantMessages,
   assistantOutput,
+  formatInterruptedOutputPreview,
   getFinalAssistantFailure,
   getSubagentUsage,
   textResult,
@@ -223,8 +224,7 @@ function rewriteTimeoutResult(
 ): AgentToolResult {
   const details = markSubagentTimedOut(result.details as SubagentToolDetails, params.timeoutMs);
   const message = details.error;
-  const partial = details.assistantOutput?.messages.at(-1)?.text;
-  const preview = partial ? `\n\nInterrupted output:\n${partial.slice(-4_000)}` : "";
+  const preview = formatInterruptedOutputPreview(details.assistantOutput);
   return textResult(`Subagent "${params.description}" (${params.profile.name}) aborted: ${message}${preview}`, {
     ...details,
     description: params.description,
@@ -345,11 +345,13 @@ export async function spawnSubagent(params: SpawnSubagentParams): Promise<AgentT
       const details = result.details as SubagentToolDetails;
       if (details.status === "aborted") {
         const reason = params.signal.reason instanceof Error ? params.signal.reason.message : String(params.signal.reason);
-        const previous = details.error;
         details.error = reason;
         if (details.progress) details.progress.error = reason;
         const first = result.content[0];
-        if (first?.type === "text") first.text = previous ? first.text.replace(previous, reason) : `${first.text}: ${reason}`;
+        if (first?.type === "text") {
+          const preview = formatInterruptedOutputPreview(details.assistantOutput);
+          first.text = `Subagent "${params.description}" (${params.profile.name}) aborted: ${reason}${preview}`;
+        }
       }
     }
     if (params.context) {
@@ -829,7 +831,8 @@ async function spawnSubagentRuntime(params: SpawnSubagentRuntimeParams): Promise
       if (thinkingClamped) progress.thinkingClamped = thinkingClamped;
     }
     const verb = status === "aborted" ? "aborted" : "failed";
-    return textResult(`Subagent "${description}" (${subagentType}) ${verb}: ${message}`, {
+    const preview = formatInterruptedOutputPreview(output);
+    return textResult(`Subagent "${description}" (${subagentType}) ${verb}: ${message}${preview}`, {
       description,
       subagentType,
       backend: profile.backend,

package/src/external-help.ts

@@ -94,11 +94,14 @@ Provide exactly one workflow source: name, scriptPath, or script. Every script s
 
 APIs: agent(prompt, { label, role, harness, subagent_type, permission, max_budget_usd, resume, context, schema, phase }); parallel(thunks); pipeline(items, ...stages); phase(title); log(message). Globals: args and cwd. Agent options other than the profile selector are optional. Choose role/harness or legacy exact subagent_type, never both. Use unique labels and clear task prompts. An agent() succeeds with its value or throws ChildRunError { runId, outcome, message, outputRef, diagnosticsRef }; catch optional failures explicitly. Uncaught failures terminate the workflow and drain siblings. Helpers never convert failures to null. Context options: {mode:"none"} (default), {mode:"recent",turns:N} (positive integer, includes current user turn), or {mode:"full"} (available context after compaction). All children share invocation-time context/settings; earlier child results must still be passed explicitly. Context excludes thinking/system instructions and pending calls; images and snapshots over 1 MiB fail without truncation. Context sharing cannot be combined with resume. Use schema for results that control branching or aggregation. Imports, filesystem globals, Date APIs, and Math.random() are unavailable.
 
-Background and supervision: set background:true on Agent or workflow to return a stable run ID while session-owned work continues. external_runs actions are list (optional workflowRunId/cursor/workflowCursor/limit), inspect (runId, view summary|output|diagnostics, optional opaque cursor/limitBytes), wait (runId or runIds, mode any|all), and cancel (runId, optional reason). Follow nextCursor/nextWorkflowCursor for complete results. Wait returns selected terminal outcomes and pending IDs, never cancels pending work, and returns an unsuccessful workflow early; interrupting wait stops only the wait. Cancelling a workflow stops active children, while cancelling one child is a catchable workflow error. Blocking-call interruption cancels that call; background work survives tool return but is cancelled on orderly owning-session shutdown. It is not a daemon: crashes leave unfinished evidence interrupted/uncertain, restart does not adopt work, and live steering is unavailable.
+Background and supervision: blocking by default — an ordinary call waits for its child and returns the result, run everything this way unless the parent has other work to do first. Set background:true only when the parent can proceed before completion, then use external_runs on the returned stable run ID while session-owned work continues. Parallel same-harness work (e.g. three agy workers) runs concurrently under the shared maxConcurrentSubagents cap: call parallel([() => agent(...), ...]) in a workflow, or issue separate Agent calls with background:true in one turn, then external_runs wait/inspect/cancel. Awaiting agent() calls sequentially stays serial by construction. external_runs actions are list (optional workflowRunId/cursor/workflowCursor/limit), inspect (runId, view summary|output|diagnostics, optional opaque cursor/limitBytes), wait (runId or runIds, mode any|all), and cancel (runId, optional reason). Follow nextCursor/nextWorkflowCursor for complete results. Wait returns selected terminal outcomes and pending IDs, never cancels pending work, and returns an unsuccessful workflow early; interrupting wait stops only the wait. Cancelling a workflow stops active children, while cancelling one child is a catchable workflow error. Blocking-call interruption cancels that call; background work survives tool return but is cancelled on orderly owning-session shutdown. It is not a daemon: crashes leave unfinished evidence interrupted/uncertain, restart does not adopt work, and live steering is unavailable.
 
 Replay: resumeFromRunId works only with persisted scriptPath and starts an explicit new attempt. It reuses the longest unchanged prefix of successful child calls; the first changed/failed/cancelled/timed-out call and its suffix run again. Script recomposition is cheap, but child reruns can cost money or repeat side effects. There is no automatic repaired-script replay.
 
-Example:
+Background example:
+export const meta = { apiVersion: 1, name: "review", description: "Map and review a repository", background: true };
+
+Foreground example:
 export const meta = { apiVersion: 1, name: "review", description: "Map and review a repository" };
 const [map, review] = await parallel([
   () => agent("Map /absolute/repo read-only.", { label: "map", role: "explorer" }),

package/src/pi-subagent.ts

@@ -352,7 +352,7 @@ function createAgentTool(
   return defineTool({
     name: "Agent",
     label: "Agent",
-    description: "Delegate one task to an external Claude Code, Codex CLI, or Antigravity role.",
+    description: "Delegate one task to an external Claude Code, Codex CLI, Antigravity, or registered Pi harness role.",
     promptSnippet: AGENT_PROMPT_SNIPPET,
     parameters: agentToolParameters,
     executionMode: "parallel",

package/src/prompts.ts

@@ -3,7 +3,7 @@ import { EXTERNAL_HARNESSES, type SubagentProfile } from "./types.ts";
 import { externalProfileRole, externalRoleAvailability } from "./profiles.ts";
 
 export const AGENT_PROMPT_SNIPPET =
-  "Delegate one task to an external Claude Code, Codex CLI, or Antigravity role.";
+  "Delegate one task to an external Claude Code, Codex CLI, Antigravity, or registered Pi harness role.";
 
 export const WORKFLOW_PROMPT_SNIPPET =
   "Run requested multi-agent orchestration with external roles; use external_help for syntax and saved workflows.";
@@ -92,5 +92,5 @@ export function buildCoordinatorPrompt(
 
 ${formatExternalRoleCatalog(profiles, defaultHarness, configuredHarnessNames)}
 
-Catalog availability reflects configured profiles, not CLI installation or authentication. External delegation tools use external CLIs only; use native Pi subagents for Pi-backed work. Give each child a clear task, absolute paths, and read-only/edit intent; nested agents may start elsewhere. Share parent context explicitly when a child needs it: context: {mode:"recent", turns:N} for the last N user turns (including the current turn), {mode:"full"} for available post-compaction conversation, or omit it for independent tasks. Prefer the smallest sufficient snapshot; snapshots exclude thinking/system instructions and pending calls, fail on images or over 1 MiB, and may contain sensitive data sent to the external harness. Use resume to follow up on an existing child; never combine it with sharing. This is text transfer, not a native session clone or guaranteed cache reuse. Set background:true only when the parent can proceed before completion; use external_runs to list, inspect, wait for, or cancel the returned run ID. External CLIs have host access, and agy always runs unsandboxed (readonly/edit are advisory). Role selection never falls back. Do not retry failed runs: agy alone may make one disclosed infrastructure retry. Use external_help for role descriptions, permission details, workflow syntax, supervision syntax, and saved workflows.`;
+Catalog availability reflects configured profiles, not CLI installation or authentication. External delegation tools use external CLIs and registered Pi harnesses; use native Pi subagents for Pi-backed work. Give each child a clear task, absolute paths, and read-only/edit intent; nested agents may start elsewhere. Share parent context explicitly when a child needs it: context: {mode:"recent", turns:N} for the last N user turns (including the current turn), {mode:"full"} for available post-compaction conversation, or omit it for independent tasks. Prefer the smallest sufficient snapshot; snapshots exclude thinking/system instructions and pending calls, fail on images or over 1 MiB, and may contain sensitive data sent to the external harness. Use resume to follow up on an existing child; never combine it with sharing. This is text transfer, not a native session clone or guaranteed cache reuse. Blocking by default; set background:true only when the parent can proceed before completion, then use external_runs on the returned run ID. Parallel same-harness work shares one maxConcurrentSubagents cap: use parallel([() => agent(...), ...]) or separate background Agent calls in one turn; sequential awaits stay serial. External CLIs have host access, and agy always runs unsandboxed (readonly/edit are advisory). Role selection never falls back. Do not retry failed runs: agy alone may make one disclosed infrastructure retry. Use external_help for role descriptions, permission details, workflow syntax, supervision syntax, and saved workflows.`;
 }

package/src/workflow/script-worker.ts

@@ -63,6 +63,8 @@ class ChildRunError extends Error {
     this.outcome = error.outcome;
     this.outputRef = error.outputRef;
     this.diagnosticsRef = error.diagnosticsRef;
+    this.assistantOutput = error.assistantOutput;
+    this.partialOutput = error.partialOutput;
   }
 }
 
@@ -116,6 +118,8 @@ function postError(error) {
         message: error.message,
         ...(error.outputRef ? { outputRef: error.outputRef } : {}),
         ...(error.diagnosticsRef ? { diagnosticsRef: error.diagnosticsRef } : {}),
+        ...(error.assistantOutput ? { assistantOutput: error.assistantOutput } : {}),
+        ...(error.partialOutput ? { partialOutput: error.partialOutput } : {}),
       } } : {}),
     });
   } finally {

package/src/workflow/tool.ts

@@ -306,12 +306,15 @@ export function createWorkflowTool(
         }
         if (resultDetails.status !== "done") {
           const runId = resultDetails.runId ?? call.runRecord?.runId ?? childId;
+          const partialOutput = resultDetails.assistantOutput?.messages.map((m) => m.text.trim()).filter(Boolean).join("\n\n");
           throw new ChildRunError({
             runId,
             outcome: childOutcome(resultDetails),
             message: resultDetails.error ?? "subagent failed",
             outputRef: { runId, view: "output" },
             diagnosticsRef: { runId, view: "diagnostics" },
+            ...(resultDetails.assistantOutput ? { assistantOutput: resultDetails.assistantOutput } : {}),
+            ...(partialOutput ? { partialOutput } : {}),
           });
         }
         if (externalOutputSchema) {
@@ -528,8 +531,11 @@ export function createWorkflowTool(
             agent.endedAt = Date.now();
           }
         }
+        const partial = error instanceof ChildRunError && error.partialOutput
+          ? `\n\nInterrupted child output (${error.runId}):\n${error.partialOutput.slice(-4_000)}`
+          : "";
         return workflowResult(
-          `Workflow "${metaName}" ${aborted ? "aborted" : "failed"}: ${message}${formatRecentLogs(snapshot.logs)}`,
+          `Workflow "${metaName}" ${aborted ? "aborted" : "failed"}: ${message}${partial}${formatRecentLogs(snapshot.logs)}`,
           cloneSnapshot(snapshot),
         );
       } finally {

package/src/workflow/types.ts

@@ -1,6 +1,7 @@
 import type { ParentContextMessages, ParentContextReceipt } from "../core/parent-context.ts";
 import type { ConcurrencyLimiter } from "../core/concurrency.ts";
 import type { RunRecord } from "../core/run-record.ts";
+import type { SubagentAssistantOutput } from "../types.ts";
 
 export type ChildRunOutcome = "failed" | "cancelled" | "timed_out";
 
@@ -15,6 +16,8 @@ export interface SerializedChildRunError {
   message: string;
   outputRef?: ChildRunReference;
   diagnosticsRef?: ChildRunReference;
+  assistantOutput?: SubagentAssistantOutput;
+  partialOutput?: string;
 }
 
 export class ChildRunError extends Error {
@@ -22,6 +25,8 @@ export class ChildRunError extends Error {
   readonly outcome: ChildRunOutcome;
   readonly outputRef?: ChildRunReference;
   readonly diagnosticsRef?: ChildRunReference;
+  readonly assistantOutput?: SubagentAssistantOutput;
+  readonly partialOutput?: string;
 
   constructor(error: SerializedChildRunError) {
     super(error.message);
@@ -30,6 +35,8 @@ export class ChildRunError extends Error {
     this.outcome = error.outcome;
     this.outputRef = error.outputRef;
     this.diagnosticsRef = error.diagnosticsRef;
+    this.assistantOutput = error.assistantOutput;
+    this.partialOutput = error.partialOutput;
   }
 }
 
@@ -40,6 +47,8 @@ export function childRunErrorData(error: ChildRunError): SerializedChildRunError
     message: error.message,
     ...(error.outputRef ? { outputRef: error.outputRef } : {}),
     ...(error.diagnosticsRef ? { diagnosticsRef: error.diagnosticsRef } : {}),
+    ...(error.assistantOutput ? { assistantOutput: error.assistantOutput } : {}),
+    ...(error.partialOutput ? { partialOutput: error.partialOutput } : {}),
   };
 }