jeo-code 0.4.6 → 0.4.8

package/src/agent/task-tool.ts

@@ -26,6 +26,7 @@ import {
   validateSubagentDoneReason,
 } from "./subagents";
 import { thinkingMaxTokens } from "../ai/model-manager";
+import type { SubagentRegistry } from "./subagent-registry";
 
 /** Lifecycle event emitted while a delegated subagent runs. */
 export interface TaskSubEvent {
@@ -41,6 +42,12 @@ export interface TaskSubEvent {
   summary?: string;
   /** Model selected for this subagent run. */
   model?: string;
+  /** 1-based task position within a fan-out batch (omitted for single-task runs). */
+  index?: number;
+  /** Total tasks in the fan-out batch (omitted for single-task runs). */
+  total?: number;
+  /** Provider token usage for the finished subagent (done events only). */
+  tokens?: { input: number; output: number };
 }
 
 export interface TaskToolOptions {
@@ -51,18 +58,46 @@ export interface TaskToolOptions {
   signal?: AbortSignal;
   /** Optional live sink (e.g. plain-stream rendering of nested progress). */
   onEvent?: (ev: TaskSubEvent) => void;
-  /** Mid-turn steering drain (gjc parity). Forwarded to a SINGLE running subagent so
-   *  an additional user query typed while the subagent works reaches it live. While
-   *  the subagent runs the parent loop is blocked inside this tool call, so the
-   *  subagent is the only active drainer — the message is not double-consumed.
-   *  Fan-out batches do NOT forward it (parallel drains would deliver to one arbitrary
-   *  subagent); pending steering stays for the parent after the batch returns. */
+  /** Mid-turn steering drain (gjc parity): an additional user query typed while a
+   *  subagent works is forwarded live. Single-task runs and the SERIAL executor
+   *  batch (concurrency 1) forward to the one active subagent. A parallel read-only
+   *  batch routes through a broadcast hub (createSteerHub) so every running worker
+   *  sees each message exactly once. Unconsumed messages stay for the parent. */
   steer?: () => string[];
+  /** When present, a `task` call with `detached: true` registers a background run
+   *  here and returns immediately; the parent controls it via the `subagent` tool. */
+  registry?: SubagentRegistry;
 }
 
 /** Max concurrent read-only subagents in a fan-out batch. */
 const MAX_FANOUT = 4;
 
+/** Hard cap on a SERIAL (mutating executor) fan-out batch: it runs one task at a
+ *  time inside one blocking tool call, so an unbounded queue would monopolize the
+ *  parent turn. Split larger efforts into sequential task calls. */
+const MAX_SERIAL_EXECUTOR = 6;
+
+/** Broadcast steering hub for a fan-out batch. Each concurrent worker registers
+ *  ONCE and then sees every parent steer message exactly once (append-only log +
+ *  per-worker cursor), so a mid-batch redirect reaches all running subagents
+ *  without the double-consume hazard of several workers draining one inbox. */
+function createSteerHub(drain?: () => string[]) {
+  const log: string[] = [];
+  return {
+    worker(): (() => string[]) | undefined {
+      if (!drain) return undefined;
+      let cursor = 0;
+      return () => {
+        const fresh = drain();
+        if (fresh.length) log.push(...fresh);
+        const out = log.slice(cursor);
+        cursor = log.length;
+        return out;
+      };
+    },
+  };
+}
+
 /** One-line protocol description appended to the launch system prompt. Pass a
  *  config so CONFIG-DECLARED custom roles are advertised to the model too. */
 export function taskToolProtocolLine(config?: Pick<Config, "subagents">): string {
@@ -141,14 +176,26 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
     taskText: string,
     context: string,
     cwd: string,
-    steer?: () => string[],
+    extra: {
+      steer?: () => string[];
+      slot?: { index: number; total: number };
+      projectContext?: Awaited<ReturnType<typeof loadProjectContext>>;
+      /** Overrides opts.signal — a detached run uses its own registry signal so it
+       *  is cancellable independently of the parent turn. */
+      signal?: AbortSignal;
+    } = {},
   ): Promise<ToolResult> => {
+    const { steer, slot, projectContext: preloadedContext, signal: signalOverride } = extra;
+    // Tag every live event with its fan-out slot so a parent monitor can tell
+    // task 1 from task 3 when several same-role subagents stream concurrently.
+    const emit = (ev: TaskSubEvent) =>
+      opts.onEvent?.(slot ? { ...ev, index: slot.index, total: slot.total } : ev);
     const model = resolveSubagentModel(role.id, opts.config);
     const maxSteps = resolveSubagentMaxSteps(role.id, opts.config);
     // gjc parity: a role may pin its own reasoning budget; absent = inherit the
     // session/global thinking level (the "(inherit)" row in the picker).
     const thinking = resolveSubagentThinking(role.id, opts.config) ?? opts.config.thinkingLevel;
-    const projectContext = await loadProjectContext(cwd);
+    const projectContext = preloadedContext ?? await loadProjectContext(cwd);
     const history: Message[] = [
       { role: "system", content: withProjectContext(subagentSystemPrompt(role), projectContext) },
       { role: "user", content: `${taskText}${context}` },
@@ -157,10 +204,13 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
     let lastTarget = "";
     let currentStep = 0;
     // Round-8 (architect ref 7-Round7Workflow): count the subagent's SUCCESSFUL
-    // mutating calls so the parent can audit a "Changed Files:" claim against
-    // observed reality instead of trusting the report's substring markers.
-    let mutationsOk = 0;
-    opts.onEvent?.({ role: role.id, kind: "start", detail: taskText, maxSteps, model });
+    // calls so the parent can audit a "Changed Files:" claim against observed
+    // reality. File-writing tools (write/edit/mkdir/delete) are tracked apart from
+    // bash: read-only bash (e.g. `bun test`) MUST NOT count as edit evidence, but
+    // bash CAN mutate, so the audit message distinguishes the two cases.
+    let fileMutations = 0;
+    let bashRuns = 0;
+    emit({ role: role.id, kind: "start", detail: taskText, maxSteps, model });
     const result = await runAgentLoop(history, {
       cwd,
       model,
@@ -169,7 +219,7 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
       // Bounded delegation: a subagent's step contract stays exact — the parent
       // owns any retry/extension decision, so the gjc retry flow is disabled here.
       budget: { maxExtensions: 0 },
-      signal: opts.signal,
+      signal: signalOverride ?? opts.signal,
       steer,
       tools: subagentToolset(role),
       events: {
@@ -178,38 +228,45 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
           if (invocation && invocation.tool && invocation.tool !== "done") {
             lastTarget = toolTarget(invocation.tool, invocation.arguments);
             trace.push(`  step ${currentStep}/${maxSteps}: ${lastTarget}`);
-            opts.onEvent?.({ role: role.id, kind: "step", detail: lastTarget, step: currentStep, maxSteps, model });
+            emit({ role: role.id, kind: "step", detail: lastTarget, step: currentStep, maxSteps, model });
           }
         },
         onToolResult: (tool, success, output) => {
-          if (success && (tool === "write" || tool === "edit" || tool === "bash")) mutationsOk++;
+          if (success) {
+            if (tool === "write" || tool === "edit" || tool === "mkdir" || tool === "delete") fileMutations++;
+            else if (tool === "bash") bashRuns++;
+          }
           const label = lastTarget || tool;
           const summary = firstUsefulLine(output);
           const suffix = summary ? ` — ${summary}` : "";
           trace.push(`  ${success ? "✓" : "✗"} ${label}${suffix}`);
-          opts.onEvent?.({ role: role.id, kind: "tool", detail: label, success, summary, step: currentStep, maxSteps, model });
+          emit({ role: role.id, kind: "tool", detail: label, success, summary, step: currentStep, maxSteps, model });
           lastTarget = "";
         },
         // Retry notices (rate-limit backoff etc.) surface as live "step" beats so the
         // parent's monitor shows WHY a subagent is pausing instead of going silent.
-        onNotice: msg => opts.onEvent?.({ role: role.id, kind: "step", detail: msg, step: currentStep, maxSteps, model }),
+        onNotice: msg => emit({ role: role.id, kind: "step", detail: msg, step: currentStep, maxSteps, model }),
         // Mid-turn steering reached this subagent: surface it as a live beat so the
         // parent's monitor shows the redirect instead of an unexplained behavior change.
-        onSteer: text => opts.onEvent?.({ role: role.id, kind: "step", detail: `↳ steer: ${text}`, step: currentStep, maxSteps, model }),
+        onSteer: text => emit({ role: role.id, kind: "step", detail: `↳ steer: ${text}`, step: currentStep, maxSteps, model }),
       },
     });
     const reason = result.doneReason?.trim() || `(subagent reached the ${result.steps}-step limit without signaling done)`;
     const validation = validateSubagentDoneReason(role, reason);
     const complete = result.done && validation.ok;
     const detail = validation.ok ? reason : `${reason}\n\n[contract incomplete: missing ${validation.missing?.join(", ")}]`;
-    opts.onEvent?.({ role: role.id, kind: "done", detail, success: complete, step: result.steps, maxSteps, model });
-    const header = `[${role.title} subagent] ${complete ? "completed" : "stopped"} in ${result.steps} step(s) on ${model}.`;
+    emit({ role: role.id, kind: "done", detail, success: complete, step: result.steps, maxSteps, model, tokens: result.usage ? { input: result.usage.inputTokens, output: result.usage.outputTokens } : undefined });
+    const tokNote = result.usage ? `, ${result.usage.inputTokens + result.usage.outputTokens} tok` : "";
+    const header = `[${role.title} subagent] ${complete ? "completed" : "stopped"} in ${result.steps} step(s) on ${model}${tokNote}.`;
     const body = trace.length ? `\nSteps:\n${trace.join("\n")}` : "";
-    // Parent-side audit: a mutating role that "completed" without ONE successful
-    // write/edit/bash cannot have changed anything — flag the claim as unverified
-    // (the report's markers prove formatting, not work).
-    const audit = complete && !role.readOnly && mutationsOk === 0
-      ? `\n[parent audit] No successful write/edit/bash was observed in this run — treat any "Changed Files:" claims above as UNVERIFIED.`
+    // Parent-side audit: a mutating role that "completed" without a successful file
+    // mutation (write/edit/mkdir/delete) likely changed nothing — flag the claim.
+    // bash is tracked separately: it CAN mutate, so an only-bash run downgrades to
+    // "verify independently" instead of the stronger UNVERIFIED.
+    const audit = complete && !role.readOnly && fileMutations === 0
+      ? bashRuns === 0
+        ? `\n[parent audit] No successful write/edit/bash was observed in this run — treat any "Changed Files:" claims above as UNVERIFIED.`
+        : `\n[parent audit] No successful write/edit was observed (only bash ran); bash may or may not have mutated files — verify any "Changed Files:" claims above independently.`
       : "";
     return { success: complete, output: `${header}${body}\n\nResult:\n${fenceSubagentReport(detail)}${audit}` };
   };
@@ -237,6 +294,18 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
       if (items.length === 0) {
         return { success: false, output: "", error: "task fan-out requires a non-empty 'tasks' array of assignments." };
       }
+      // #5: the mutating executor fan-out is SERIAL (concurrency 1) and blocks the
+      // turn; cap it regardless of justification so a huge queue can't monopolize
+      // the parent. Split larger efforts into sequential task calls.
+      if (!role.readOnly && items.length > MAX_SERIAL_EXECUTOR) {
+        return {
+          success: false,
+          output: "",
+          error:
+            `Executor fan-out of ${items.length} exceeds the serial cap of ${MAX_SERIAL_EXECUTOR}. ` +
+            `The mutating executor runs one task at a time and blocks the turn — split into ≤${MAX_SERIAL_EXECUTOR}-task batches or sequential task calls.`,
+        };
+      }
       // Spawn-gate lite (plan/gjc-inheritance.md B9, gjc spawn-gate 계승): a batch
       // wider than MAX_FANOUT is refused BEFORE any subagent launches unless the
       // model justifies the parallelism — silent capping hid the cost decision.
@@ -257,13 +326,22 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
       // Read-only roles fan out concurrently (bounded). The mutating executor is serialized
       // (concurrency 1) so parallel subagents can't race on the same files.
       const limit = role.readOnly ? Math.min(items.length, MAX_FANOUT) : 1;
+      // Load project context ONCE per batch instead of re-scanning AGENTS.md for
+      // every fan-out task (redundant IO + duplicated tokens).
+      const batchContext = await loadProjectContext(cwd);
       const results: ToolResult[] = new Array(items.length);
       let next = 0;
+      // #7: broadcast steering hub — each concurrent worker sees every parent
+      // steer message exactly once (safe even for parallel read-only fan-out).
+      const steerHub = createSteerHub(opts.steer);
       const worker = async () => {
+        // One steer cursor per concurrent worker (not per item) so a worker that
+        // processes several items sees each parent message once across them all.
+        const workerSteer = steerHub.worker();
         while (true) {
           const i = next++;
           if (i >= items.length) return;
-          results[i] = await runOne(role, items[i]!.task, items[i]!.context, cwd);
+          results[i] = await runOne(role, items[i]!.task, items[i]!.context, cwd, { slot: { index: i + 1, total: items.length }, projectContext: batchContext, steer: workerSteer });
         }
       };
       await Promise.all(Array.from({ length: limit }, () => worker()));
@@ -279,6 +357,20 @@ export function createTaskTool(opts: TaskToolOptions): ToolHandler {
     if (!taskText) {
       return { success: false, output: "", error: `task tool requires a non-empty 'task' (or a 'tasks' array). Valid roles: ${subagentRoleIds(opts.config).join(", ")}.` };
     }
-    return runOne(role, taskText, ctx(args.context), cwd, opts.steer);
+    // Detached form (#9): register a background run and return immediately so the
+    // parent can keep working, then list/inspect/await/cancel via the `subagent`
+    // tool. Steering is not forwarded to a detached run (no single active drainer).
+    if (args.detached === true && opts.registry) {
+      const rec = opts.registry.launch(role.id, taskText, signal =>
+        runOne(role, taskText, ctx(args.context), cwd, { signal }),
+      );
+      return {
+        success: true,
+        output:
+          `[detached] launched ${role.title} subagent '${rec.id}'. It runs in the background — ` +
+          `keep working, then use the 'subagent' tool ({action:"await"|"list"|"inspect"|"cancel", ids?}) to collect its result.`,
+      };
+    }
+    return runOne(role, taskText, ctx(args.context), cwd, { steer: opts.steer });
   };
 }

package/src/agent/tool-output.ts

@@ -0,0 +1,115 @@
+/**
+ * Tool-result output handling — the model-visible output budget, both-ends
+ * truncation, recoverable artifact spilling, and the minimize→truncate→spill
+ * orchestration the agent loop applies to every tool result.
+ *
+ * Extracted from `engine.ts` (single-responsibility: the loop drives steps; this
+ * module owns how a tool's raw output is shaped before it re-enters context).
+ * `engine.ts` re-exports the public surface for backward compatibility.
+ */
+import * as fs from "node:fs/promises";
+import * as path from "node:path";
+import { jeoEnv } from "../util/env";
+import { minimizeToolOutput } from "./output-minimizer";
+
+/** 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();
+
+/** Read results are deliberate, contiguous file slices the model explicitly asked
+ *  for (via lineRange), already line-capped by the read tool and recoverable via
+ *  spill. They get a much larger model-visible budget than the generic
+ *  noise-control cap, so a 500-line read is not silently re-shrunk to ~100 lines.
+ *  JEO_READ_OUTPUT_MAX overrides (1k..200k). */
+function envReadOutputMax(): number {
+  const raw = Number(jeoEnv("READ_OUTPUT_MAX") ?? "");
+  return Number.isFinite(raw) && raw >= 1_000 && raw <= 200_000 ? Math.trunc(raw) : 32_000;
+}
+export const READ_OUTPUT_MAX = envReadOutputMax();
+
+/**
+ * Cap a tool result fed back to the model. Default mode keeps both ends: the head
+ * holds the start (e.g. a command's invocation) and the tail holds what's usually
+ * decisive (test summaries, the final error). A pure head-cut loses that.
+ *
+ * `headOnly` truncates from the front only — for `read` results, which are a
+ * contiguous file slice the model explicitly requested; head/tail splitting would
+ * mangle the code into two non-adjacent fragments.
+ */
+export function truncateToolOutput(s: string, max = TOOL_OUTPUT_MAX, headOnly = false): string {
+  if (s.length <= max) return s;
+  if (headOnly) {
+    return `${s.slice(0, max)}\n…(${s.length - max} chars truncated; narrow the lineRange or read the spilled artifact)…`;
+  }
+  const head = Math.floor(max * 0.6);
+  const tail = max - head;
+  return `${s.slice(0, head)}\n…(${s.length - max} chars truncated)…\n${s.slice(s.length - tail)}`;
+}
+
+/** Non-read tool output larger than this is spilled to a recoverable artifact file.
+ *  Aligned with `truncateToolOutput`'s generic cap so that whenever the model-visible
+ *  result drops content, the full output is recoverable via the artifact. (`read`
+ *  spills against the larger READ_OUTPUT_MAX in the result loop.) */
+export const TOOL_SPILL_THRESHOLD = TOOL_OUTPUT_MAX;
+
+/** Most recent tool-result artifacts to keep; older ones are pruned on each spill. */
+export const MAX_TOOL_ARTIFACTS = 50;
+
+/** Best-effort retention: keep the newest `MAX_TOOL_ARTIFACTS` files in `dir`, delete the rest. */
+async function pruneToolArtifacts(dir: string): Promise<void> {
+  const files = await fs.readdir(dir).catch(() => [] as string[]);
+  if (files.length <= MAX_TOOL_ARTIFACTS) return;
+  const stamped = await Promise.all(
+    files.map(async f => ({ f, m: (await fs.stat(path.join(dir, f)).catch(() => null))?.mtimeMs ?? 0 })),
+  );
+  stamped.sort((a, b) => b.m - a.m); // newest first
+  for (const { f } of stamped.slice(MAX_TOOL_ARTIFACTS)) {
+    await fs.rm(path.join(dir, f), { force: true }).catch(() => {});
+  }
+}
+
+/**
+ * Write an oversized tool result verbatim under `.jeo/artifacts/tool-results/` and
+ * return the workspace-relative path (for the model to `read`). Best-effort: throws
+ * are caught by the caller, which simply omits the artifact note.
+ */
+export async function spillToolResult(tool: string, output: string, cwd: string): Promise<string> {
+  const dir = path.join(cwd, ".jeo", "artifacts", "tool-results");
+  await fs.mkdir(dir, { recursive: true });
+  const safeTool = tool.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 32) || "tool";
+  const stamp = `${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+  const rel = path.join(".jeo", "artifacts", "tool-results", `${stamp}-${safeTool}.txt`);
+  await fs.writeFile(path.join(cwd, rel), output, "utf-8");
+  // Retention so a long session can't grow the artifact dir without bound.
+  await pruneToolArtifacts(dir);
+  return rel;
+}
+
+/**
+ * Shape one tool's raw output into the model-visible result body: strip runner
+ * noise (minimize), cap to the per-tool budget (`read` gets the larger read budget
+ * and a head-only cut), and spill the full output to a recoverable artifact when it
+ * exceeds the budget. Behavior-identical to the inline logic it replaces in
+ * `runAgentLoop`.
+ */
+export async function formatToolResultBody(tool: string, rawOutput: string, cwd: string): Promise<string> {
+  const visible = minimizeToolOutput(rawOutput, tool).text;
+  // `read` is a deliberate, contiguous file slice: give it the larger read budget
+  // and truncate head-only (head/tail splitting mangles code). Other tools keep the
+  // generic noise-control cap + both-ends truncation.
+  const isReadResult = tool === "read";
+  const outputBudget = isReadResult ? READ_OUTPUT_MAX : TOOL_OUTPUT_MAX;
+  let body = truncateToolOutput(visible, outputBudget, isReadResult);
+  if (rawOutput.length > outputBudget) {
+    const artifact = await spillToolResult(tool, rawOutput, cwd).catch(() => null);
+    if (artifact) {
+      body += `\n[full output (${rawOutput.length} chars) saved to ${artifact} — read it for the truncated remainder]`;
+    }
+  }
+  return body;
+}

package/src/agent/tools.ts

@@ -3,6 +3,7 @@ import * as fs from "node:fs/promises";
 import * as path from "node:path";
 import { readWorkflowState, readWorkflowStateStrict, type WorkflowState } from "./state";
 import { jeoEnv } from "../util/env";
+import { READ_OUTPUT_MAX } from "./tool-output";
 
 /** Read the deep-interview lock; on corrupt state fail CLOSED (treat as active lock). */
 async function readMutationLock(cwd: string): Promise<WorkflowState | null> {
@@ -291,10 +292,27 @@ export async function readTool(
       return { success: true, output: out.join("\n") };
     }
 
-    const MAX_LINES = 500;
-    const annotated = lines.slice(0, MAX_LINES).map((l, i) => `${i + 1}${lineAnchor(l)}|${l}`).join("\n");
-    if (lines.length > MAX_LINES) {
-      const notice = `\n…(showing lines 1-${MAX_LINES} of ${lines.length}; pass lineRange "${MAX_LINES + 1}-" to read the rest)`;
+    // Default (no lineRange): fill the model-visible read budget with WHOLE lines
+    // instead of a fixed 500-line cap that left half the 32k budget unused and forced
+    // needless pagination (the read tool's biggest "reads too little per call" pain).
+    // READ_OUTPUT_MAX is the real cap; a hard line ceiling (JEO_READ_MAX_LINES) guards
+    // pathological files, and a small reserve keeps the pagination notice inside the
+    // budget so it is never trimmed by the downstream head-only truncation.
+    const HARD_LINE_CEILING = Math.max(500, Number(jeoEnv("READ_MAX_LINES") ?? "") || 5000);
+    const charBudget = Math.max(1_000, READ_OUTPUT_MAX - 256);
+    const shownLines: string[] = [];
+    let usedChars = 0;
+    for (let i = 0; i < lines.length && shownLines.length < HARD_LINE_CEILING; i++) {
+      const annotatedLine = `${i + 1}${lineAnchor(lines[i]!)}|${lines[i]}`;
+      const cost = annotatedLine.length + 1; // + newline
+      if (shownLines.length > 0 && usedChars + cost > charBudget) break; // always emit ≥1 line
+      shownLines.push(annotatedLine);
+      usedChars += cost;
+    }
+    const annotated = shownLines.join("\n");
+    if (shownLines.length < lines.length) {
+      const shown = shownLines.length;
+      const notice = `\n…(showing lines 1-${shown} of ${lines.length}; pass lineRange "${shown + 1}-" to read the rest)`;
       return { success: true, output: annotated + notice };
     }
     return { success: true, output: annotated };
@@ -574,7 +592,8 @@ export async function bashTool(
   cwd: string = process.cwd(),
   timeoutMs: number = 120_000,
   subdir?: string,
-  env?: Record<string, string>
+  env?: Record<string, string>,
+  onProgress?: (partialOutput: string) => void,
 ): Promise<ToolResult> {
   if (jeoEnv("BASH_FIXUPS") === "1") {
     const fx = applyBashFixups(command);
@@ -608,12 +627,27 @@ export async function bashTool(
       killTimer = setTimeout(() => { try { proc.kill(9); } catch {} }, 3_000);
     }, TIMEOUT_MS);
 
+    // Stream stdout incrementally when a progress sink is attached (drives the live
+    // DIMMED bash output view); read stderr fully in parallel. Without a sink, fall
+    // back to a single post-exit read (identical content, no streaming overhead).
+    const stderrPromise = new Response(proc.stderr).text();
+    let stdout = "";
+    if (onProgress) {
+      const decoder = new TextDecoder();
+      let lastEmit = 0;
+      for await (const chunk of proc.stdout as unknown as AsyncIterable<Uint8Array>) {
+        stdout += decoder.decode(chunk, { stream: true });
+        const now = Date.now();
+        if (now - lastEmit >= 80) { lastEmit = now; onProgress(stdout); }
+      }
+      stdout += decoder.decode();
+      onProgress(stdout);
+    }
     await proc.exited;
     clearTimeout(timer);
     if (killTimer) clearTimeout(killTimer);
-
-    const stdout = await new Response(proc.stdout).text();
-    const stderr = await new Response(proc.stderr).text();
+    if (!onProgress) stdout = await new Response(proc.stdout).text();
+    const stderr = await stderrPromise;
 
     let output = [stdout, stderr].filter(Boolean).join("\n");
     const MAX_OUTPUT = 100_000;

package/src/ai/model-manager.ts

@@ -2,11 +2,7 @@ import { providerRegistry } from "./provider-registry";
 import { OAUTH_FLOW_REGISTRY } from "../auth/flows";
 import { readGlobalConfig } from "../agent/state";
 import { resolveCredential, type AuthProvider, type Credential } from "../auth";
-import { anthropicAdapter } from "./providers/anthropic";
-import { openaiAdapter } from "./providers/openai";
-import { geminiAdapter } from "./providers/gemini";
-import { ollamaAdapter } from "./providers/ollama";
-import { antigravityAdapter } from "./providers/antigravity";
+import "./register-providers"; // side-effect: registers built-in adapters into providerRegistry
 import type { CallOptions, Message, ProviderAdapter, ProviderName } from "./types";
 import { expandAlias, resolveModelId, effectiveAliasesFor } from "./model-registry";
 import { findCatalogEntry, type ModelCatalogEntry } from "./model-catalog-compat";
@@ -16,12 +12,6 @@ import { jeoEnv } from "../util/env";
 import type { Config } from "../agent/state";
 
 
-// Initialize Provider Registry
-providerRegistry.register("anthropic", anthropicAdapter);
-providerRegistry.register("openai", openaiAdapter);
-providerRegistry.register("gemini", geminiAdapter);
-providerRegistry.register("antigravity", antigravityAdapter);
-providerRegistry.register("ollama", ollamaAdapter);
 
 
 export function resolveProvider(model: string): ProviderName {
@@ -315,6 +305,7 @@ async function resolveCall(options: Partial<CallOptions>, kind: "request" | "str
     onUsage: options.onUsage,
     signal: options.signal,
     reasoningEffort: options.reasoningEffort ?? thinkingToReasoningEffort(config.thinkingLevel),
+    onReasoning: options.onReasoning,
   };
   // Caller-supplied retry sink rides on the config-derived retry budget so the
   // engine/TUI can surface "rate limited — retrying in Ns" instead of a silent wait.

package/src/ai/providers/antigravity.ts

@@ -160,13 +160,18 @@ export function antigravityRequest(messages: Message[], options: CallOptions, cr
 type CcaUsage = { promptTokenCount?: number; candidatesTokenCount?: number; thoughtsTokenCount?: number };
 interface CcaChunk {
   response?: {
-    candidates?: { content?: { parts?: { text?: string }[] }; finishReason?: string }[];
+    candidates?: { content?: { parts?: { text?: string; thought?: boolean }[] }; finishReason?: string }[];
     usageMetadata?: CcaUsage;
   };
 }
 
 function textOf(chunk: CcaChunk): string {
-  return chunk.response?.candidates?.[0]?.content?.parts?.map(p => p.text ?? "").join("") ?? "";
+  return chunk.response?.candidates?.[0]?.content?.parts?.filter(p => !p.thought).map(p => p.text ?? "").join("") ?? "";
+}
+
+/** Native thinking text (`thought` parts) — kept separate so it never pollutes the JSON tool call. */
+function thoughtOf(chunk: CcaChunk): string {
+  return chunk.response?.candidates?.[0]?.content?.parts?.filter(p => p.thought).map(p => p.text ?? "").join("") ?? "";
 }
 
 async function fetchAntigravity(messages: Message[], options: CallOptions, credential: Credential): Promise<Response> {
@@ -194,6 +199,8 @@ export const antigravityAdapter: ProviderAdapter = {
     for await (const data of readSse(response.body)) {
       let chunk: CcaChunk;
       try { chunk = JSON.parse(data); } catch { continue; }
+      const thought = thoughtOf(chunk);
+      if (thought) options.onReasoning?.(thought);
       out += textOf(chunk);
       if (chunk.response?.usageMetadata) usage = chunk.response.usageMetadata;
     }
@@ -209,6 +216,8 @@ export const antigravityAdapter: ProviderAdapter = {
     for await (const data of readSse(response.body)) {
       let chunk: CcaChunk;
       try { chunk = JSON.parse(data); } catch { continue; }
+      const thought = thoughtOf(chunk);
+      if (thought) options.onReasoning?.(thought);
       const delta = textOf(chunk);
       if (delta) { yielded = true; yield delta; }
       if (chunk.response?.usageMetadata) usage = chunk.response.usageMetadata;

package/src/ai/providers/gemini.ts

@@ -119,7 +119,7 @@ export function geminiCliRequest(messages: Message[], options: CallOptions, acce
 }
 
 interface GeminiChunk {
-  candidates?: { content?: { parts?: { text?: string }[] }; finishReason?: string }[];
+  candidates?: { content?: { parts?: { text?: string; thought?: boolean }[] }; finishReason?: string }[];
   promptFeedback?: { blockReason?: string };
   usageMetadata?: { promptTokenCount?: number; candidatesTokenCount?: number; thoughtsTokenCount?: number };
 }
@@ -130,7 +130,13 @@ interface CcaChunk {
 }
 
 function textOf(chunk: GeminiChunk): string {
-  return chunk.candidates?.[0]?.content?.parts?.map(p => p.text ?? "").join("") ?? "";
+  return chunk.candidates?.[0]?.content?.parts?.filter(p => !p.thought).map(p => p.text ?? "").join("") ?? "";
+}
+
+/** Native thinking text (`thought` parts), present only when the model emits thought
+ *  summaries. Kept SEPARATE from textOf so thoughts never pollute the JSON tool call. */
+function thoughtOf(chunk: GeminiChunk): string {
+  return chunk.candidates?.[0]?.content?.parts?.filter(p => p.thought).map(p => p.text ?? "").join("") ?? "";
 }
 
 /** When Gemini returns HTTP 200 with no text, surface the real cause (safety block /
@@ -176,6 +182,8 @@ async function* ccaTurn(messages: Message[], options: CallOptions, credential: C
     }
     const inner = chunk.response;
     if (!inner) continue;
+    const thought = thoughtOf(inner);
+    if (thought) options.onReasoning?.(thought);
     const delta = textOf(inner);
     if (delta) {
       yieldedAny = true;
@@ -239,6 +247,8 @@ export const geminiAdapter: ProviderAdapter = {
       } catch {
         continue;
       }
+      const thought = thoughtOf(chunk);
+      if (thought) options.onReasoning?.(thought);
       const delta = textOf(chunk);
       if (delta) {
         yieldedAny = true;

package/src/ai/register-providers.ts

@@ -0,0 +1,21 @@
+/**
+ * Built-in provider registration (the registry bootstrap).
+ *
+ * Importing this module for its side effect registers every bundled LLM adapter
+ * into the shared `providerRegistry`. `model-manager` then resolves adapters
+ * through the registry alone — it no longer imports, or even names, concrete
+ * providers. To add a new built-in provider, register it HERE only; nothing in
+ * `model-manager` changes.
+ */
+import { providerRegistry } from "./provider-registry";
+import { anthropicAdapter } from "./providers/anthropic";
+import { openaiAdapter } from "./providers/openai";
+import { geminiAdapter } from "./providers/gemini";
+import { ollamaAdapter } from "./providers/ollama";
+import { antigravityAdapter } from "./providers/antigravity";
+
+providerRegistry.register("anthropic", anthropicAdapter);
+providerRegistry.register("openai", openaiAdapter);
+providerRegistry.register("gemini", geminiAdapter);
+providerRegistry.register("antigravity", antigravityAdapter);
+providerRegistry.register("ollama", ollamaAdapter);

package/src/ai/types.ts

@@ -43,6 +43,10 @@ export interface CallOptions {
   /** Notified before each auto-retry backoff wait (rate limits / transient errors).
    *  NOT forwarded to provider adapters — consumed by the manager's retry layer. */
   onRetry?: (attempt: number, err: unknown, delayMs: number) => void;
+  /** Streaming sink for native model reasoning/thinking text deltas (separate from the
+   *  answer text). Surfaced as a transient dimmed view; absent for models that emit no
+   *  thought text. */
+  onReasoning?: (delta: string) => void;
 }
 
 export interface ProviderAdapter {

package/src/cli/runner.ts

@@ -172,15 +172,6 @@ export const COMMANDS: readonly CommandSpec[] = [
       return args => m.runUpdateCommand(args);
     },
   },
-  {
-    name: "gjc",
-    summary: "Run the gjc workflow skill as an autonomous build loop (plan → implement → verify).",
-    usage: "gjc \"<intent>\"",
-    loader: async () => {
-      const m = await import("../commands/gjc");
-      return args => m.runGjcCommand(args);
-    },
-  },
   {
     name: "ooo-seed",
     summary: "Generate an immutable ooo seed from a specification (spec-first automation).",