jeo-code 0.4.6 → 0.4.8

package/README.ja.md

@@ -150,11 +150,11 @@ CI は `.github/workflows/npm-publish.yml` で公開します — GitHub リリ
 ## 変更履歴 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.4.8]** (2026-06-14) — Live-frame stability: constant-height live turn, renderer self-heal off-by-one fix, and frame-safe child-stdout sanitizing — no more duplicate model bar or torn escapes.
+- **[0.4.7]** (2026-06-14) — Detached subagents + `subagent` control tool, live shaded in-flight output, registry-driven providers, fuller `read` budget, styled italics in the final report, and `gjc` retired.
 - **[0.4.6]** (2026-06-14) — Width-correct forge cards for CJK/emoji, red borders on failed tool cards, aligned `ooo ralph` monitor HUD, and a per-theme user-card palette.
 - **[0.4.5]** (2026-06-14) — First-class filesystem make/remove tools.
 - **[0.4.4]** (2026-06-13) — Live subagent status mirroring, always-useful Ctrl+O activity tail, read lineRange crash guard.
-- **[0.4.3]** (2026-06-13) — Readability pass for autopilot, subagent activity, and worked-history review.
-- **[0.4.2]** (2026-06-13) — Thinking-loop termination guarantees (cycle guard + turn wall-clock budget), unboxed live status without step counters, self-contained `.jeo` namespace, live next-prompt input card, role-targeted model/thinking picker.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.ko.md

@@ -150,11 +150,11 @@ CI는 `.github/workflows/npm-publish.yml`로 배포합니다 — GitHub 릴리
 ## 변경 이력 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.4.8]** (2026-06-14) — Live-frame stability: constant-height live turn, renderer self-heal off-by-one fix, and frame-safe child-stdout sanitizing — no more duplicate model bar or torn escapes.
+- **[0.4.7]** (2026-06-14) — Detached subagents + `subagent` control tool, live shaded in-flight output, registry-driven providers, fuller `read` budget, styled italics in the final report, and `gjc` retired.
 - **[0.4.6]** (2026-06-14) — Width-correct forge cards for CJK/emoji, red borders on failed tool cards, aligned `ooo ralph` monitor HUD, and a per-theme user-card palette.
 - **[0.4.5]** (2026-06-14) — First-class filesystem make/remove tools.
 - **[0.4.4]** (2026-06-13) — Live subagent status mirroring, always-useful Ctrl+O activity tail, read lineRange crash guard.
-- **[0.4.3]** (2026-06-13) — Readability pass for autopilot, subagent activity, and worked-history review.
-- **[0.4.2]** (2026-06-13) — Thinking-loop termination guarantees (cycle guard + turn wall-clock budget), unboxed live status without step counters, self-contained `.jeo` namespace, live next-prompt input card, role-targeted model/thinking picker.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.md

@@ -150,11 +150,11 @@ Required npm token permissions (repository secret `NPM_TOKEN`):
 ## Changelog
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.4.8]** (2026-06-14) — Live-frame stability: constant-height live turn, renderer self-heal off-by-one fix, and frame-safe child-stdout sanitizing — no more duplicate model bar or torn escapes.
+- **[0.4.7]** (2026-06-14) — Detached subagents + `subagent` control tool, live shaded in-flight output, registry-driven providers, fuller `read` budget, styled italics in the final report, and `gjc` retired.
 - **[0.4.6]** (2026-06-14) — Width-correct forge cards for CJK/emoji, red borders on failed tool cards, aligned `ooo ralph` monitor HUD, and a per-theme user-card palette.
 - **[0.4.5]** (2026-06-14) — First-class filesystem make/remove tools.
 - **[0.4.4]** (2026-06-13) — Live subagent status mirroring, always-useful Ctrl+O activity tail, read lineRange crash guard.
-- **[0.4.3]** (2026-06-13) — Readability pass for autopilot, subagent activity, and worked-history review.
-- **[0.4.2]** (2026-06-13) — Thinking-loop termination guarantees (cycle guard + turn wall-clock budget), unboxed live status without step counters, self-contained `.jeo` namespace, live next-prompt input card, role-targeted model/thinking picker.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.zh.md

@@ -150,11 +150,11 @@ CI 通过 `.github/workflows/npm-publish.yml` 发布 — GitHub 发布 release
 ## 更新日志 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.4.8]** (2026-06-14) — Live-frame stability: constant-height live turn, renderer self-heal off-by-one fix, and frame-safe child-stdout sanitizing — no more duplicate model bar or torn escapes.
+- **[0.4.7]** (2026-06-14) — Detached subagents + `subagent` control tool, live shaded in-flight output, registry-driven providers, fuller `read` budget, styled italics in the final report, and `gjc` retired.
 - **[0.4.6]** (2026-06-14) — Width-correct forge cards for CJK/emoji, red borders on failed tool cards, aligned `ooo ralph` monitor HUD, and a per-theme user-card palette.
 - **[0.4.5]** (2026-06-14) — First-class filesystem make/remove tools.
 - **[0.4.4]** (2026-06-13) — Live subagent status mirroring, always-useful Ctrl+O activity tail, read lineRange crash guard.
-- **[0.4.3]** (2026-06-13) — Readability pass for autopilot, subagent activity, and worked-history review.
-- **[0.4.2]** (2026-06-13) — Thinking-loop termination guarantees (cycle guard + turn wall-clock budget), unboxed live status without step counters, self-contained `.jeo` namespace, live next-prompt input card, role-targeted model/thinking picker.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jeo-code",
-  "version": "0.4.6",
+  "version": "0.4.8",
   "description": "Clean, highly optimized AI coding agent using spec-first loop",
   "type": "module",
   "main": "src/cli.ts",

package/src/agent/dev/evolution-bridge.ts

@@ -1,7 +1,40 @@
-import { runGjcCommand } from "../../commands/gjc";
+import { runAgentLoop, executorSystemPrompt, DEFAULT_TOOLS } from "../engine";
+import { readGlobalConfig } from "../state";
+import { runPostImplementationHooks } from "../hooks";
 import { runAdvancedAnalysis } from "./advanced-analyzer";
 import { logEvolution } from "./evolution-logger";
 
+async function runEvolutionLoop(intent: string, cwd: string): Promise<void> {
+  const config = await readGlobalConfig();
+  const model = config.defaultModel || "fast";
+  const systemPrompt = executorSystemPrompt();
+
+  await runAgentLoop([{ role: "user", content: intent }], {
+    cwd,
+    systemPrompt,
+    model,
+    tools: DEFAULT_TOOLS,
+    maxSteps: 50,
+  });
+
+  console.log("\n[jeo] Verifying implementation...");
+  const verify = await runPostImplementationHooks(cwd, intent);
+  
+  if (!verify.success) {
+    console.error("\n[jeo] Verification FAILED. Auto-repairing...");
+    const repairTask = `Previous implementation failed verification.\nErrors:\n${verify.output}\n\nPlease fix.`;
+    await runAgentLoop([{ role: "user", content: repairTask }], {
+      cwd,
+      systemPrompt,
+      model,
+      tools: DEFAULT_TOOLS,
+      maxSteps: 30,
+    });
+  } else {
+    console.log("\n[jeo] Verification SUCCESSFUL.");
+  }
+}
+
 export async function consultGjcForAdvancedEvolution(cwd: string) {
   const report = await runAdvancedAnalysis(cwd);
   const timestamp = new Date().toISOString();
@@ -27,7 +60,7 @@ As my implementation guide (gjc), please:
   `;
 
   try {
-    await runGjcCommand([request]);
+    await runEvolutionLoop(request, cwd);
     
     await logEvolution({
       timestamp: new Date().toISOString(),
@@ -62,7 +95,7 @@ export async function consultGjcForEvolution(cwd: string) {
 
   console.log();
   try {
-    await runGjcCommand([report]);
+    await runEvolutionLoop(report, cwd);
     await logEvolution({
       timestamp: new Date().toISOString(),
       target: "src/agent/engine.ts",

package/src/agent/dev/self-analysis.ts

@@ -22,7 +22,12 @@ export async function runSelfAnalysis(cwd: string): Promise<string> {
   const content = await fs.readFile(targetPath, "utf-8");
   
   const lineCount = content.split("\n").length;
-  const hasTooManyResponsibilities = content.includes("runAgentLoop") && content.includes("truncateToolOutput") && content.includes("spillToolResult");
+  // Ownership-accurate SRP check: the loop drives steps, while output shaping
+  // (truncate/spill) lives in tool-output.ts. Flag only when those are DEFINED
+  // here again, not merely imported or re-exported for backward compatibility.
+  const definesOutputShaping =
+    /\bfunction\s+truncateToolOutput\b/.test(content) && /\bfunction\s+spillToolResult\b/.test(content);
+  const hasTooManyResponsibilities = content.includes("runAgentLoop") && definesOutputShaping;
 
   let report = "Analysis of src/agent/engine.ts:\n";
   report += "- File length: " + lineCount + " lines.\n";

package/src/agent/engine.ts

@@ -16,7 +16,8 @@ import { webSearchTool, setWebSearchActiveModel } from "./web-search";
 import { friendlyProviderError, isContextOverflowError, isRefusalError } from "../util/provider-error";
 import { isRateLimitError } from "../util/retry";
 import { runPreToolHooks, runPostTurnHooks } from "./hooks";
-import { minimizeToolOutput } from "./output-minimizer";
+import { truncateToolOutput, formatToolResultBody } from "./tool-output";
+export { TOOL_OUTPUT_MAX, READ_OUTPUT_MAX, TOOL_SPILL_THRESHOLD, MAX_TOOL_ARTIFACTS, truncateToolOutput, spillToolResult } from "./tool-output";
 import { StepBudget, dynamicStepBudgetConfig, resolveStepBudgetConfig, hashSignature, type StepBudgetConfig } from "./step-budget";
 import { historyTokens, trimToolResultsInPlace } from "./compaction";
 import { jeoEnv } from "../util/env";
@@ -30,6 +31,7 @@ async function invokeCallLlm(history: Message[], options: {
   onUsage?: (u: { inputTokens?: number; outputTokens?: number }) => void;
   onRetry?: (attempt: number, err: unknown, delayMs: number) => void;
   onToken?: (delta: string) => void;
+  onReasoning?: (delta: string) => void;
 }): Promise<string> {
   const mod = await import("./loop");
   return mod.callLlm(history, options);
@@ -39,14 +41,14 @@ export interface ToolInvocation {
   arguments?: Record<string, any>;
 }
 
-export type ToolHandler = (args: Record<string, any>, cwd: string) => Promise<ToolResult>;
+export type ToolHandler = (args: Record<string, any>, cwd: string, onProgress?: (partialOutput: string) => void) => Promise<ToolResult>;
 
-/** The default executor toolset (read / write / edit / bash / find / search). */
+/** The default executor toolset (read / write / edit / bash / find / search / ls / mkdir / delete / web_search). */
 export const DEFAULT_TOOLS: Record<string, ToolHandler> = {
   read: (a, cwd) => readTool(a.filePath ?? a.path, a.lineRange ?? a.range, cwd, !!a.raw),
   write: (a, cwd) => writeTool(a.filePath ?? a.path, a.content ?? "", cwd),
   edit: (a, cwd) => editTool(a.filePath ?? a.path, a.editBlock ?? a.edit ?? "", cwd),
-  bash: (a, cwd) => bashTool(a.command ?? a.cmd, cwd, typeof a.timeoutMs === "number" ? a.timeoutMs : undefined, typeof a.cwd === "string" ? a.cwd : (typeof a.subdir === "string" ? a.subdir : undefined), a.env && typeof a.env === "object" ? a.env : undefined),
+  bash: (a, cwd, onProgress) => bashTool(a.command ?? a.cmd, cwd, typeof a.timeoutMs === "number" ? a.timeoutMs : undefined, typeof a.cwd === "string" ? a.cwd : (typeof a.subdir === "string" ? a.subdir : undefined), a.env && typeof a.env === "object" ? a.env : undefined, onProgress),
   find: (a, cwd) => findTool(a.globPattern ?? a.pattern, cwd),
   search: (a, cwd) => searchTool(a.pattern, a.globPattern ?? "*", cwd, !!(a.ignoreCase ?? a.i), { before: a.before, after: a.after, context: a.context, maxMatches: a.maxMatches }),
   ls: (a, cwd) => lsTool(a.dirPath ?? a.path ?? a.dir ?? ".", cwd),
@@ -134,6 +136,10 @@ export interface AgentLoopEvents {
   onStep?(step: number): void | Promise<void>;
   onAssistant?(raw: string, invocation: ToolInvocation | null): void;
   onToolResult?(tool: string, success: boolean, output: string): void;
+  /** Streaming partial output of the currently-running tool (e.g. bash stdout as it
+   *  arrives) — drives a live DIMMED output view that the final formatted result
+   *  replaces on onToolResult. Only bash emits today; other tools are unaffected. */
+  onToolProgress?(tool: string, partial: string): void;
   /** Transient progress notice (e.g. "rate limited — retrying in Ns"); NOT a terminal error. */
   onNotice?(message: string): void;
   /** Cumulative token usage after each LLM call — drives live usage meters. */
@@ -141,6 +147,9 @@ export interface AgentLoopEvents {
   /** Accumulated streamed model response so far — drives the live reasoning view. Only
    *  requested when a consumer sets it (the engine streams solely for the TUI). */
   onModelStream?(textSoFar: string): void;
+  /** Accumulated native reasoning/thinking text so far — drives a transient dimmed
+   *  "thinking" view. Only requested when a consumer (TUI) attaches. */
+  onReasoningStream?(textSoFar: string): void;
   /** Step-budget change (gjc-style retry flow): the limit was extended because the
    *  turn is making progress. `limit` is the new max; `reason` is display-ready. */
   onBudget?(limit: number, reason: string): void;
@@ -192,14 +201,6 @@ export interface AgentLoopResult {
   usage?: { inputTokens: number; outputTokens: number };
 }
 
-/** Env-tunable output budget (plan/gjc-inheritance.md B10, gjc settings-driven
- *  output handling 계승): JEO_TOOL_OUTPUT_MAX caps the model-visible tool result;
- *  the spill threshold tracks it so anything truncated stays artifact-recoverable. */
-function envOutputMax(): number {
-  const raw = Number(jeoEnv("TOOL_OUTPUT_MAX") ?? "");
-  return Number.isFinite(raw) && raw >= 500 && raw <= 200_000 ? Math.trunc(raw) : 4_000;
-}
-export const TOOL_OUTPUT_MAX = envOutputMax();
 
 /** Wall-clock budget for ONE agent turn (ms). JEO_TURN_MAX_MS overrides; 0 disables.
  *  Default 30 minutes: long autonomous runs stay alive, while a turn that spins in
@@ -214,55 +215,6 @@ export function turnMaxMs(env: Record<string, string | undefined> = process.env)
   return 30 * 60 * 1000;
 }
 
-/**
- * Cap a tool result fed back to the model, keeping both ends: the head holds the
- * start (e.g. a file's top / a command's invocation) and the tail holds what's
- * usually decisive (test summaries, the final error). A pure head-cut loses that.
- */
-export function truncateToolOutput(s: string, max = TOOL_OUTPUT_MAX): string {
-  if (s.length <= max) return s;
-  const head = Math.floor(max * 0.6);
-  const tail = max - head;
-  return `${s.slice(0, head)}\n…(${s.length - max} chars truncated)…\n${s.slice(s.length - tail)}`;
-}
-
-/** Tool output larger than this is spilled to a recoverable artifact file. Aligned
- *  with `truncateToolOutput`'s cap so that whenever the model-visible result drops
- *  content, the full output is recoverable via the artifact. */
-export const TOOL_SPILL_THRESHOLD = TOOL_OUTPUT_MAX;
-
-/**
- * Write an oversized tool result verbatim under `.jeo/artifacts/tool-results/` and
- * return the workspace-relative path (for the model to `read`). Best-effort: throws
- * are caught by the caller, which simply omits the artifact note.
- */
-/** Most recent tool-result artifacts to keep; older ones are pruned on each spill. */
-export const MAX_TOOL_ARTIFACTS = 50;
-
-/** Best-effort retention: keep the newest `MAX_TOOL_ARTIFACTS` files in `dir`, delete the rest. */
-async function pruneToolArtifacts(dir: string): Promise<void> {
-  const files = await fs.readdir(dir).catch(() => [] as string[]);
-  if (files.length <= MAX_TOOL_ARTIFACTS) return;
-  const stamped = await Promise.all(
-    files.map(async f => ({ f, m: (await fs.stat(path.join(dir, f)).catch(() => null))?.mtimeMs ?? 0 })),
-  );
-  stamped.sort((a, b) => b.m - a.m); // newest first
-  for (const { f } of stamped.slice(MAX_TOOL_ARTIFACTS)) {
-    await fs.rm(path.join(dir, f), { force: true }).catch(() => {});
-  }
-}
-
-export async function spillToolResult(tool: string, output: string, cwd: string): Promise<string> {
-  const dir = path.join(cwd, ".jeo", "artifacts", "tool-results");
-  await fs.mkdir(dir, { recursive: true });
-  const safeTool = tool.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 32) || "tool";
-  const stamp = `${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
-  const rel = path.join(".jeo", "artifacts", "tool-results", `${stamp}-${safeTool}.txt`);
-  await fs.writeFile(path.join(cwd, rel), output, "utf-8");
-  // Retention so a long session can't grow the artifact dir without bound.
-  await pruneToolArtifacts(dir);
-  return rel;
-}
 
 /** Levenshtein distance (small inputs: tool/command names). */
 function editDistance(a: string, b: string): number {
@@ -453,6 +405,10 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
     const onToken = ev.onModelStream
       ? (delta: string) => { streamBuf += delta; ev.onModelStream!(streamBuf); }
       : undefined;
+    let reasonBuf = "";
+    const onReasoning = ev.onReasoningStream
+      ? (delta: string) => { reasonBuf += delta; ev.onReasoningStream!(reasonBuf); }
+      : undefined;
     let responseText: string;
     try {
       responseText = await invokeCallLlm(history, {
@@ -462,6 +418,7 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
               signal: opts.signal,
               onUsage: u => { acc.inputTokens += u.inputTokens ?? 0; acc.outputTokens += u.outputTokens ?? 0; sawUsage = true; },
               onToken,
+              onReasoning,
               // Make provider auto-retry visible: previously a rate-limited call sat in a
               // silent backoff wait, then surfaced "auto-retry was exhausted" with no trace
               // of the retries that DID happen.
@@ -783,7 +740,8 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
             output = preHookResult.error + (preHookResult.output ? `\n${preHookResult.output}` : "");
           } else {
             try {
-              const res = await handler(args ?? {}, cwd);
+              const onProgress = ev.onToolProgress ? (partial: string) => ev.onToolProgress!(tool, partial) : undefined;
+              const res = await handler(args ?? {}, cwd, onProgress);
               success = res.success;
               output = res.success ? res.output : (res.error ? (res.output ? `${res.error}\n${res.output}` : res.error) : res.output);
             } catch (err: any) {
@@ -875,15 +833,7 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
 
         ev.onToolResult?.(call.tool, res.success, res.output);
 
-        const minimized = minimizeToolOutput(res.output, call.tool);
-        const visible = minimized.text;
-        let resultBody = truncateToolOutput(visible);
-        if (res.output.length > TOOL_SPILL_THRESHOLD) {
-          const artifact = await spillToolResult(call.tool, res.output, cwd).catch(() => null);
-          if (artifact) {
-            resultBody += `\n[full output (${res.output.length} chars) saved to ${artifact} — read it for the elided middle]`;
-          }
-        }
+        const resultBody = await formatToolResultBody(call.tool, res.output, cwd);
 
         const { diags: hookDiags, ran: hooksRan } = await runPostTurnHooks(
           cwd,

package/src/agent/loop.ts

@@ -19,6 +19,8 @@ export interface ChatOptions {
    *  delivered here (concatenation equals the returned string). Absent ⇒ a single
    *  non-streaming `call()` (unchanged behavior for non-interactive/test callers). */
   onToken?: (delta: string) => void;
+  /** Streaming sink for native reasoning/thinking deltas (drives the dimmed live view). */
+  onReasoning?: (delta: string) => void;
 }
 
 const manager = createModelManager();

package/src/agent/subagent-registry.ts

@@ -0,0 +1,131 @@
+/**
+ * In-process detached-subagent registry (gjc `subagent`/`job` parity, scoped down
+ * to one runtime). The synchronous `task` tool blocks the parent until a subagent
+ * finishes; a DETACHED launch registers the run here and returns immediately, so
+ * the parent can keep working and later list / inspect / await / cancel it via the
+ * `subagent` control tool. Concurrency is real (JS event loop): a detached run's
+ * awaits interleave with the parent's between steps.
+ *
+ * Lifecycle is bounded to the turn that created the registry — `cancelAll()` on
+ * turn teardown guarantees no background promise leaks into the next turn.
+ */
+import type { ToolResult } from "./tools";
+
+export type SubagentStatus = "running" | "completed" | "failed" | "cancelled";
+
+export interface SubagentRecord {
+  /** Stable id, e.g. "executor-1". */
+  id: string;
+  role: string;
+  /** The assignment text (trimmed for display). */
+  task: string;
+  status: SubagentStatus;
+  startedAt: number;
+  finishedAt?: number;
+  /** Whether the finished run reported success (contract satisfied). */
+  success?: boolean;
+  /** Final subagent report/output, set once the run settles. */
+  result?: string;
+}
+
+interface Entry {
+  record: SubagentRecord;
+  promise: Promise<void>;
+  abort: AbortController;
+}
+
+/** A detached run: receives its own AbortSignal and resolves to the subagent's
+ *  final ToolResult. The runner is responsible for streaming live events itself. */
+export type DetachedRunner = (signal: AbortSignal) => Promise<ToolResult>;
+
+export class SubagentRegistry {
+  private readonly entries = new Map<string, Entry>();
+  private readonly seq = new Map<string, number>();
+
+  /** Register and START a detached run; returns the (running) record immediately. */
+  launch(role: string, task: string, runner: DetachedRunner): SubagentRecord {
+    const n = (this.seq.get(role) ?? 0) + 1;
+    this.seq.set(role, n);
+    const id = `${role}-${n}`;
+    const abort = new AbortController();
+    const record: SubagentRecord = {
+      id,
+      role,
+      task: task.length > 200 ? task.slice(0, 197) + "…" : task,
+      status: "running",
+      startedAt: Date.now(),
+    };
+    const promise = (async () => {
+      try {
+        const res = await runner(abort.signal);
+        // A cancel that already fired wins — don't overwrite the terminal state.
+        if (record.status === "cancelled") return;
+        record.status = res.success ? "completed" : "failed";
+        record.success = res.success;
+        record.result = res.output || res.error || "";
+      } catch (err) {
+        if (record.status === "cancelled") return;
+        record.status = "failed";
+        record.result = err instanceof Error ? err.message : String(err);
+      } finally {
+        if (record.finishedAt === undefined) record.finishedAt = Date.now();
+      }
+    })();
+    this.entries.set(id, { record, promise, abort });
+    return record;
+  }
+
+  list(): SubagentRecord[] {
+    return [...this.entries.values()].map(e => e.record);
+  }
+
+  get(id: string): SubagentRecord | undefined {
+    return this.entries.get(id)?.record;
+  }
+
+  running(): SubagentRecord[] {
+    return this.list().filter(r => r.status === "running");
+  }
+
+  /** Wait for the given ids (or all running, when empty). With `timeoutMs` the wait
+   *  is bounded — unfinished runs simply stay "running" in the returned snapshot. */
+  async awaitIds(ids: string[], timeoutMs?: number): Promise<SubagentRecord[]> {
+    const targets = ids
+      .map(id => this.entries.get(id))
+      .filter((e): e is Entry => e !== undefined);
+    const all = Promise.all(targets.map(e => e.promise)).then(() => {});
+    if (timeoutMs !== undefined && timeoutMs > 0) {
+      let handle: ReturnType<typeof setTimeout> | undefined;
+      const timer = new Promise<void>(resolve => {
+        handle = setTimeout(resolve, timeoutMs);
+      });
+      await Promise.race([all, timer]);
+      if (handle !== undefined) clearTimeout(handle);
+    } else {
+      await all;
+    }
+    return targets.map(e => e.record);
+  }
+
+  /** Cancel the given ids (or all running, when empty): aborts the run and marks the
+   *  record cancelled. Already-terminal records are returned unchanged. */
+  cancel(ids: string[]): SubagentRecord[] {
+    const out: SubagentRecord[] = [];
+    for (const id of ids) {
+      const e = this.entries.get(id);
+      if (!e) continue;
+      if (e.record.status === "running") {
+        e.record.status = "cancelled";
+        e.record.finishedAt = Date.now();
+        e.abort.abort();
+      }
+      out.push(e.record);
+    }
+    return out;
+  }
+
+  /** Abort every still-running subagent (turn teardown / Ctrl-C). */
+  cancelAll(): SubagentRecord[] {
+    return this.cancel(this.running().map(r => r.id));
+  }
+}

package/src/agent/subagent-tool.ts

@@ -0,0 +1,89 @@
+/**
+ * `subagent` control tool (#9) — the parent's handle on DETACHED subagents launched
+ * via `task {detached:true}`. Mirrors gjc's `subagent`/`job` control surface, scoped
+ * to an in-process registry: list, inspect, await (optionally bounded), and cancel.
+ *
+ * Out of scope here (separate subsystems, not stubbed): live peer messaging (IRC)
+ * and pause/resume — a step-budget loop has no safe mid-step checkpoint to resume
+ * from, so those are intentionally absent rather than faked.
+ */
+import type { ToolHandler } from "./engine";
+import type { ToolResult } from "./tools";
+import type { SubagentRegistry, SubagentRecord } from "./subagent-registry";
+
+/** One-line protocol description appended to the launch system prompt. */
+export const SUBAGENT_TOOL_PROTOCOL_LINE =
+  `subagent {action:"list"|"inspect"|"await"|"cancel", ids?, timeoutMs?} — control DETACHED ` +
+  `subagents started with task{detached:true}. 'await' blocks (optionally up to timeoutMs ms) and ` +
+  `returns their reports; 'inspect' shows status + result; 'cancel' aborts them. Omit ids to target all running.`;
+
+function elapsed(rec: SubagentRecord): string {
+  const end = rec.finishedAt ?? Date.now();
+  return `${Math.max(0, Math.round((end - rec.startedAt) / 1000))}s`;
+}
+
+function rowLine(rec: SubagentRecord): string {
+  return `- ${rec.id} [${rec.status.toUpperCase()}] ${elapsed(rec)} · ${rec.task}`;
+}
+
+function detailBlock(rec: SubagentRecord): string {
+  const head = rowLine(rec);
+  if (rec.status === "running" || !rec.result) return head;
+  return `${head}\n${rec.result}`;
+}
+
+function idsOf(args: Record<string, any>): string[] {
+  if (Array.isArray(args.ids)) return args.ids.map((x: unknown) => String(x));
+  if (args.id !== undefined) return [String(args.id)];
+  return [];
+}
+
+export function createSubagentTool(registry: SubagentRegistry): ToolHandler {
+  return async (args: Record<string, any>, _cwd: string): Promise<ToolResult> => {
+    const action = String(args.action ?? "list").trim().toLowerCase();
+    const ids = idsOf(args);
+
+    if (action === "list") {
+      const rows = registry.list();
+      if (rows.length === 0) {
+        return { success: true, output: "No detached subagents this turn. Launch one with task {detached:true}." };
+      }
+      const running = rows.filter(r => r.status === "running").length;
+      return { success: true, output: `${rows.length} subagent(s), ${running} running:\n${rows.map(rowLine).join("\n")}` };
+    }
+
+    if (action === "inspect") {
+      const targets = (ids.length ? ids.map(id => registry.get(id)) : registry.list())
+        .filter((r): r is SubagentRecord => r !== undefined);
+      if (targets.length === 0) {
+        return { success: false, output: "", error: ids.length ? `No subagent matches ${ids.join(", ")}.` : "No detached subagents this turn." };
+      }
+      return { success: true, output: targets.map(detailBlock).join("\n\n") };
+    }
+
+    if (action === "await") {
+      const targets = ids.length ? ids : registry.running().map(r => r.id);
+      if (targets.length === 0) {
+        return { success: true, output: "No running subagents to await." };
+      }
+      const timeoutMs = typeof args.timeoutMs === "number" && args.timeoutMs > 0 ? args.timeoutMs : undefined;
+      const recs = await registry.awaitIds(targets, timeoutMs);
+      const stillRunning = recs.filter(r => r.status === "running").length;
+      const head = stillRunning > 0
+        ? `Awaited ${recs.length} subagent(s); ${stillRunning} still running after the ${timeoutMs}ms timeout — await again or cancel.`
+        : `Awaited ${recs.length} subagent(s); all settled.`;
+      return { success: stillRunning === 0, output: `${head}\n\n${recs.map(detailBlock).join("\n\n")}` };
+    }
+
+    if (action === "cancel") {
+      const targets = ids.length ? ids : registry.running().map(r => r.id);
+      if (targets.length === 0) {
+        return { success: true, output: "No running subagents to cancel." };
+      }
+      const recs = registry.cancel(targets);
+      return { success: true, output: `Cancelled ${recs.length} subagent(s):\n${recs.map(rowLine).join("\n")}` };
+    }
+
+    return { success: false, output: "", error: `Unknown subagent action '${action}'. Use list | inspect | await | cancel.` };
+  };
+}

package/src/agent/subagents.ts

@@ -178,20 +178,39 @@ function renderRolePrompt(template: string, role: SubagentRole): string {
     .trim();
 }
 
+/** True when `marker` is present in `text` AND the span between it and the next
+ *  required marker (or end of text) carries non-whitespace content. A label-only
+ *  section ("Summary:" with an empty body) is not a real report, so it fails. */
+function markerHasContent(text: string, marker: string, allMarkers: string[]): boolean {
+  const start = text.indexOf(marker);
+  if (start < 0) return false;
+  const after = start + marker.length;
+  let end = text.length;
+  for (const other of allMarkers) {
+    if (other === marker) continue;
+    const j = text.indexOf(other, after);
+    if (j >= 0 && j < end) end = j;
+  }
+  return text.slice(after, end).trim().length > 0;
+}
+
 export function validateSubagentDoneReason(role: SubagentRole, reason: string | undefined): { ok: boolean; missing?: string[] } {
   const trimmed = (reason ?? "").trim();
   if (!trimmed) return { ok: false, missing: ["done.reason"] };
+  const markers = role.requiredDoneMarkers ?? [];
+  // Each required section must be PRESENT and carry non-empty content — a report of
+  // bare labels (no prose) is rejected, which the substring-presence check let pass.
+  const sectionMissing = markers.filter(m => !markerHasContent(trimmed, m, markers));
   if (role.id === "critic") {
     const verdicts = ["[OKAY]", "[ITERATE]", "[REJECT]"];
     const hasVerdict = verdicts.some(marker => trimmed.startsWith(marker));
     const missing = [
       ...(hasVerdict ? [] : ["[OKAY]|[ITERATE]|[REJECT]"]),
-      ...((role.requiredDoneMarkers ?? []).filter(marker => !trimmed.includes(marker))),
+      ...sectionMissing,
     ];
     return { ok: missing.length === 0, missing };
   }
-  const missing = (role.requiredDoneMarkers ?? []).filter(marker => !trimmed.includes(marker));
-  return { ok: missing.length === 0, missing };
+  return { ok: sectionMissing.length === 0, missing: sectionMissing };
 }
 
 /** Build a role-specific system prompt from its dedicated template. */