@hicaru/pi-rlm 0.2.1 → 0.2.2

package/src/tool/repl-tool.ts

@@ -16,16 +16,15 @@
  */
 
 import { Type } from "typebox";
-import type { Theme, ToolDefinition } from "@earendil-works/pi-coding-agent";
-import { Container, Spacer, Text } from "@earendil-works/pi-tui";
+import type { ToolDefinition } from "@earendil-works/pi-coding-agent";
+import { Text } from "@earendil-works/pi-tui";
 import type { Model, Usage, Api } from "@earendil-works/pi-ai";
 import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
-import { buildInteractiveHandlers } from "../bridge/interactive.ts";
 import { buildLibraryHandler } from "../bridge/library.ts";
-import { createPiInteractiveDeps } from "../bridge/pi-interactive.ts";
+import { libraryPrefixesIn } from "../context/library-context.ts";
 import type { SubcallGates } from "../util/concurrency.ts";
 import { LimitGuard, limitsFromConfig } from "../core/limits.ts";
-import type { InteractiveDeps, RlmConfig, RlmInput, RlmResult } from "../core/types.ts";
+import type { RlmConfig, RlmInput, RlmResult } from "../core/types.ts";
 import { SandboxManager } from "../sandbox/sandbox-manager.ts";
 import type { SubcallOpts } from "../sandbox/sandbox.ts";
 import { createSubcallHandlers, type Invocation } from "../bridge/subcall-handlers.ts";
@@ -39,26 +38,13 @@ import { createEngine } from "../core/engine.ts";
 import { spinnerFrame } from "../ui/theme.ts";
 import { previewText } from "../text/preview.ts";
 import { errorMessage } from "../util/errors.ts";
-import {
-  cardHeader,
-  cardStatsLine,
-  renderCollapsedCard,
-  renderExpandedSubcallTree,
-} from "./subcall-render.ts";
 import { createProgressNotifier, validateToolParams } from "./tool-utils.ts";
-import { capReplResultText, replDelegationNudge } from "../mode/native-guards.ts";
+import { buildReplResultText, collectReplWarnings } from "./repl-result.ts";
+import { renderReplCollapsed, renderReplExpanded } from "./repl-render.ts";
 import { attachTracer, trace, traceEnabled } from "../util/trace.ts";
 
-/** Chars of code shown on the tool call line, and of stdout in the expanded view. */
+/** Chars of code shown on the tool call line. */
 const CALL_PREVIEW_CHARS = 80;
-const EXPANDED_STDOUT_CHARS = 2_000;
-const EXPANDED_STDERR_CHARS = 500;
-
-// ── Parameter schema ──
-
-export const ReplToolParams = Object.freeze(Type.Object({
-  code: Type.String({ description: "Python code to execute in the persistent REPL sandbox" }),
-}));
 
 /** Last non-empty line of a Python traceback — the `TypeError: …` line, not the frames. */
 function lastLine(text: string): string {
@@ -70,67 +56,11 @@ function lastLine(text: string): string {
   return "";
 }
 
-/** Model-visible text assembled from a repl() result. */
-export interface ReplResultText {
-  readonly text: string;
-}
-
-/**
- * Assemble the model-visible text for a repl() result: cap stdout, append a zero-subcall
- * delegation nudge when a bulk read went undelegated, and report tasks still running.
- *
- * The pending line is the model's only signal that `spawn()`ed work is outstanding — without
- * it a model that spawned and moved on has no way to know it should still collect.
- *
- * `varNames` covers the opposite failure: a block that stores its results in `answers` and
- * prints nothing reads as a bare "(no output)", so the model concludes the block did nothing
- * and re-runs it — paying twice for the same sub-calls. The headless engine already answers
- * this with the same hint (core/answer.ts); native mode was the only path missing it.
- */
-export function buildReplResultText(
-  stdout: string,
-  finalAnswer: string | undefined,
-  subcalls: readonly RlmSubcall[],
-  backgroundPending = 0,
-  varNames: readonly string[] = [],
-): ReplResultText {
-  const answerSubmitted = finalAnswer !== undefined;
-  const noOutput = !answerSubmitted && !stdout;
-  const varsHint = noOutput && varNames.length > 0
-    ? ` — the block ran fine and these REPL vars are defined: ${varNames.join(", ")}. `
-      + "Do NOT re-run it; read them in the next block."
-    : "";
-  const rawText = answerSubmitted
-    ? `ANSWER_SUBMITTED (${finalAnswer.length} chars) — delivered to user. Do not restate it.`
-    : stdout || `(no output)${varsHint}`;
-  // Model-visible text is capped; the caller keeps full stdout in `details` for the TUI.
-  const cappedText = capReplResultText(rawText) ?? rawText;
-  const delegated = subcalls.some((s) => s.kind === "llm" || s.kind === "batch" || s.kind === "rlm");
-  const nudge = answerSubmitted ? undefined : replDelegationNudge(rawText.length, delegated);
-  const failedBg = subcalls.filter((s) => s.id.startsWith("bg") && s.status === "error").length;
-  const pendingLine = backgroundPending > 0
-    ? `\n\n[rlm] ${backgroundPending} background task(s) still running — rlm_await_all(tasks) to collect.`
-    : "";
-  const failedLine = failedBg > 0
-    ? `\n[rlm] ${failedBg} background sub-call(s) FAILED — their rlm_await value is an "Error: …" string, not data.`
-    : "";
-  return { text: cappedText + (nudge ?? "") + pendingLine + failedLine };
-}
+// ── Parameter schema ──
 
-/** Advisory diagnostics derived from a completed invocation's sub-calls. */
-export function collectReplWarnings(subcalls: readonly RlmSubcall[]): readonly string[] | undefined {
-  let failed = 0;
-  let total = 0;
-  for (let i = 0; i < subcalls.length; i++) {
-    const call = subcalls[i];
-    if (call.status !== "error") continue;
-    // A batch subcall stands for many prompts; a single call stands for one.
-    failed += call.failedCount ?? 1;
-    total += call.totalCount ?? 1;
-  }
-  if (failed === 0) return undefined;
-  return Object.freeze([`${failed}/${total} sub-call(s) failed — results may be incomplete`]);
-}
+export const ReplToolParams = Object.freeze(Type.Object({
+  code: Type.String({ description: "Python code to execute in the persistent REPL sandbox" }),
+}));
 
 // ── Mutable bridge state (handler indirection) ──
 
@@ -141,21 +71,18 @@ export function collectReplWarnings(subcalls: readonly RlmSubcall[]): readonly s
  * calls rather than rebuilding handlers (which would lose REPL variable state). Handlers
  * capture the Invocation synchronously at interrupt entry and never re-read it — with
  * spawn() a sub-call can outlive its exec, and a later read would attribute it to whichever
- * turn happened to be current when it resumed.
+ * turn happened to be current when it settled.
  *
  * Detached work resolves to the session-scoped background Invocation instead, whose emitter
  * and LimitGuard are not torn down at the end of a turn.
  */
 class NativeBridgeState {
   private current: Invocation | null = null;
-  /** Interactive callbacks for the turn in progress; child engines inherit them. */
-  interactive: InteractiveDeps | null = null;
 
   constructor(private readonly background: BackgroundTasks) {}
 
-  swap(inv: Invocation, interactive: InteractiveDeps): void {
+  swap(inv: Invocation): void {
     this.current = Object.freeze({ ...inv });
-    this.interactive = interactive;
   }
 
   /** Detached ⇒ session registry; otherwise the turn that is currently executing. */
@@ -175,9 +102,9 @@ class NativeBridgeState {
 export interface ReplToolDeps {
   readonly sandboxManager: SandboxManager;
   readonly model: Model<Api>;
-  readonly workerModel: Model<Api>;
+  readonly llmModel: Model<Api>;
   readonly getModel?: () => Model<Api> | undefined;
-  readonly getWorkerModel?: () => Model<Api> | undefined;
+  readonly getLlmModel?: () => Model<Api> | undefined;
   readonly registry: ModelRegistry;
   /** Live accessor — `/rlm-config` replaces the config object, so never capture the value. */
   readonly getConfig: () => RlmConfig;
@@ -193,13 +120,13 @@ export interface ReplToolDeps {
 }
 
 export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplToolParams, ReplDetails> {
-  const { sandboxManager, workerModel, registry, getConfig, signal, onUsage, background } = deps;
+  const { sandboxManager, llmModel, registry, getConfig, signal, onUsage, background } = deps;
   const bridgeState = new NativeBridgeState(background);
 
   // Late-bound cwd — getOrCreate installs handlers only at spawn; never rebuild the closure.
   let sessionCwd = process.cwd();
 
-  const getWorkerModel = (): Model<Api> => deps.getWorkerModel?.() ?? workerModel;
+  const getLlmModel = (): Model<Api> => deps.getLlmModel?.() ?? llmModel;
   const getModel = (): Model<Api> => deps.getModel?.() ?? deps.model;
 
   // Each rlm_query spawns a child RLM with its own sandbox and turn loop, not a flat
@@ -207,7 +134,7 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
   // progress and cost deltas land on the emitter the parent invocation is using.
   const runChild = (input: RlmInput, inv: Invocation): Promise<RlmResult> => createEngine({
     model: getModel(),
-    workerModel: getWorkerModel(),
+    llmModel: getLlmModel(),
     registry,
     config: getConfig(),
     signal,
@@ -218,8 +145,6 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
     // the child's own root turns — so fold both roles into "sub" rather than casting.
     onUsage: onUsage === undefined ? undefined : (usage: Usage) => onUsage(usage, "sub"),
     limits: limitsFromConfig(getConfig()),
-    onTodo: bridgeState.interactive?.onTodo,
-    onAskUserQuestion: bridgeState.interactive?.onAskUserQuestion,
   })(input);
 
   // Built once: the same closures stay correct across repl() calls because everything
@@ -228,12 +153,16 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
     resolve: (opts) => bridgeState.resolve(opts),
     gates: deps.gates,
     registry,
-    getWorkerModel,
+    getLlmModel,
     getModel,
     getConfig,
     signal,
     onUsage,
     runChild,
+    // The session sandbox's context is the child's world. Read lazily so a load_library from an
+    // earlier repl() reaches a child spawned in a later one. Populated before any interrupt can
+    // fire: execute() awaits ensureContext() before getOrCreate().
+    getChildContext: () => sandboxManager.contextPayload ?? undefined,
     trackDetached: (task) => background.track(task),
   });
 
@@ -241,12 +170,22 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
     ? buildLibraryHandler({
         getCwd: () => sessionCwd,
         getEmitter: () => bridgeState.currentEmitter,
+        // Refuse pre-flight whatever the worker would reject, so host idempotency is never
+        // committed for an append that did not happen.
+        getContext: () => sandboxManager.contextPayload,
         parentId: undefined,
         signal,
-        startIndex: 1,
+        // Keep the manager's replay copy in step with the worker's live `context`, and with it
+        // whatever a child spawned after this load will inherit.
+        onLoaded: (payload) => { sandboxManager.appendLibrary(payload); },
       })
     : undefined;
-  if (libraryBundle) deps.registerDiscardHook?.(libraryBundle.reset);
+  if (libraryBundle) {
+    // Re-derive the loaded-prefix cache from the payload that will actually be replayed —
+    // clearing it outright would make the host re-clone a library the recreated worker already has.
+    const bundle = libraryBundle;
+    deps.registerDiscardHook?.(() => bundle.reset(libraryPrefixesIn(sandboxManager.contextPayload)));
+  }
 
   return {
     name: "repl",
@@ -258,7 +197,7 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
       "semantic reading to map_files / llm_query / llm_query_batched / llm_query_chunked " +
       "(rlm_query for iterative sub-tasks) — stdout returned to you is hard-capped at 4K chars, " +
       "so printing file bodies is useless. Variables, imports, and the `answers`/`plan` memo " +
-      "persist across calls. Also supports todo, ask_user_question, and load_library.",
+      "persist across calls. Also supports load_library.",
     promptSnippet:
       "repl: run Python in a persistent sandbox holding the whole repository in `context`; " +
       "search/grep_context/outline to locate, map_files/llm_query* to read.",
@@ -327,24 +266,11 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
       }
 
       try {
-        // Build interactive handlers (session-stable callbacks)
-        const interactive = createPiInteractiveDeps(ctx);
-        const interactiveHandlers = buildInteractiveHandlers({
-          onAskUserQuestion: getConfig().askUserQuestion ? interactive.onAskUserQuestion : undefined,
-          onTodo: interactive.onTodo,
-          onTodoRow: undefined,
-          emitter,
-          depth: 0,
-          parentId: undefined,
-        });
-
         sessionCwd = ctx.cwd ?? process.cwd();
 
         await deps.ensureContext?.();
         await sandboxManager.getOrCreate({
           ...subcallHandlers,
-          askUserQuestion: interactiveHandlers.askUserQuestion,
-          todo: interactiveHandlers.todo,
           ...(libraryBundle?.handlers ?? {}),
         });
 
@@ -366,7 +292,7 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
           // Wire per-invocation mutable state only after the serialized exec slot
           // is active. Swapping earlier would let queued repl() calls overwrite
           // emitter/limits for the currently running REPL execution.
-          bridgeState.swap({ emitter, parentId: undefined, depth: 0, limits }, interactive);
+          bridgeState.swap({ emitter, parentId: undefined, depth: 0, limits });
         }, execSignal);
         const elapsed = Date.now() - start;
         capturedStdout = result.stdout;
@@ -480,51 +406,3 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
     },
   };
 }
-
-// ── Collapsed view ──
-
-function replStats(details: ReplDetails, theme: Theme): string {
-  const elapsed = details.executionTimeMs > 0 ? `${details.executionTimeMs}ms` : undefined;
-  return cardStatsLine(details.totals, theme, elapsed, details.backgroundPending);
-}
-
-function renderReplCollapsed(details: ReplDetails, theme: Theme): Text {
-  return renderCollapsedCard("REPL", details.status, replStats(details, theme), details.subcalls, theme);
-}
-
-// ── Expanded view ──
-
-function renderReplExpanded(details: ReplDetails, theme: Theme): Container {
-  const container = new Container();
-
-  container.addChild(new Text(cardHeader("REPL", details.status, replStats(details, theme), theme), 0, 0));
-
-  // Output
-  if (details.output) {
-    container.addChild(new Spacer(1));
-    const out = details.output.length > EXPANDED_STDOUT_CHARS
-      ? `${details.output.slice(0, EXPANDED_STDOUT_CHARS)}…`
-      : details.output;
-    container.addChild(new Text(out, 0, 0));
-  }
-
-  if (details.warnings && details.warnings.length > 0) {
-    container.addChild(new Spacer(1));
-    container.addChild(new Text(theme.fg("muted", details.warnings.join("\n")), 0, 0));
-  }
-
-  // Stderr
-  if (details.stderr) {
-    container.addChild(new Spacer(1));
-    container.addChild(new Text(theme.fg("error", details.stderr.slice(0, EXPANDED_STDERR_CHARS)), 0, 0));
-  }
-
-  // Sub-call tree
-  if (details.subcalls.length > 0) {
-    container.addChild(new Spacer(1));
-    container.addChild(new Text(theme.fg("muted", "─── Sub-calls ───"), 0, 0));
-    container.addChild(renderExpandedSubcallTree(details.subcalls, theme));
-  }
-
-  return container;
-}

package/src/tool/rlm-aggregator.ts

@@ -6,14 +6,14 @@
  * getState(): RlmDetails for direct access (spinner loop, final return).
  *
  * Subcall storage and totals are delegated to SubcallStore. Root-level state
- * (status, prompt, turns, answer, warnings) is kept in the aggregator.
+ * (status, prompt, turns, answer) is kept in the aggregator.
  *
  * Replaces RlmToolBridge's internal state accumulation. The emitter is pure
  * dispatch; the aggregator is pure state. Separated for independent testing.
  */
 
 import type { AgentToolUpdateCallback } from "@earendil-works/pi-agent-core";
-import type { RlmEmitter, TurnEvent, RootUsageEvent, AnswerEvent, StatusEvent, RootPromptEvent, WarningsEvent } from "./rlm-events.ts";
+import type { RlmEmitter, TurnEvent, RootUsageEvent, AnswerEvent, StatusEvent, RootPromptEvent } from "./rlm-events.ts";
 import type { RlmDetails, RlmRunStatus } from "./rlm-details.ts";
 import { EmitterListener } from "./emitter-listener.ts";
 import { SubcallStore } from "./subcall-store.ts";
@@ -27,7 +27,6 @@ export class RlmEventAggregator extends EmitterListener {
   private turnCurrent = 0;
   private turnMax = 0;
   private answer?: string;
-  private warnings?: readonly string[];
 
   constructor(
     emitter: RlmEmitter,
@@ -40,7 +39,6 @@ export class RlmEventAggregator extends EmitterListener {
       emitter.onTurn((e) => this.handleTurn(e)),
       emitter.onRootUsage((e) => this.handleRootUsage(e)),
       emitter.onAnswer((e) => this.handleAnswer(e)),
-      emitter.onWarnings((e) => this.handleWarnings(e)),
       emitter.onStatus((e) => this.handleStatus(e)),
       emitter.onRootPrompt((e) => this.handleRootPrompt(e)),
     ]);
@@ -64,11 +62,6 @@ export class RlmEventAggregator extends EmitterListener {
     this.notify();
   }
 
-  private handleWarnings(event: WarningsEvent): void {
-    this.warnings = event.warnings;
-    this.notify();
-  }
-
   private handleStatus(event: StatusEvent): void {
     this.rootStatus = event.status;
     this.notify();
@@ -90,7 +83,6 @@ export class RlmEventAggregator extends EmitterListener {
       subcalls: this.store.getSubcalls(),
       totals: this.store.getTotals(),
       answer: this.answer,
-      warnings: this.warnings,
     };
   }
 

package/src/tool/rlm-details.ts

@@ -40,7 +40,5 @@ export interface RlmDetails {
   readonly subcalls: readonly RlmSubcall[];
   readonly totals: { readonly costUsd: number; readonly tokens: number };
   readonly answer?: string;
-  /** Advisory diagnostics — surfaced to the user, never a failure. */
-  readonly warnings?: readonly string[];
 }
 

package/src/tool/rlm-events.ts

@@ -45,10 +45,6 @@ export interface SubcallUpdatedEvent {
   readonly totalCount?: number;
 }
 
-export interface WarningsEvent {
-  readonly warnings: readonly string[];
-}
-
 export interface TurnEvent {
   readonly current: number;
   readonly max: number;
@@ -121,11 +117,6 @@ export class RlmEmitter {
     this.ee.emit("answer", { text } satisfies AnswerEvent);
   }
 
-  /** Set advisory warnings (root-only; never a failure). */
-  emitWarnings(warnings: readonly string[]): void {
-    this.ee.emit("warnings", { warnings } satisfies WarningsEvent);
-  }
-
   /** Set the root run status (done/error/aborted). */
   emitStatus(status: RlmRunStatus): void {
     this.ee.emit("status", { status } satisfies StatusEvent);
@@ -163,11 +154,6 @@ export class RlmEmitter {
     return () => { this.ee.off("answer", handler); };
   }
 
-  onWarnings(handler: (event: WarningsEvent) => void): () => void {
-    this.ee.on("warnings", handler);
-    return () => { this.ee.off("warnings", handler); };
-  }
-
   onStatus(handler: (event: StatusEvent) => void): () => void {
     this.ee.on("status", handler);
     return () => { this.ee.off("status", handler); };

package/src/tool/rlm-tool.ts

@@ -8,7 +8,6 @@
 import { type Theme, type ToolDefinition } from "@earendil-works/pi-coding-agent";
 import { Container, Markdown, Spacer, Text, type Component } from "@earendil-works/pi-tui";
 import { Type } from "typebox";
-import { createPiInteractiveDeps } from "../bridge/pi-interactive.ts";
 import type { RlmController, StartInput } from "../mode/rlm-mode.ts";
 import { spinnerFrame } from "../ui/theme.ts";
 import { markdownTheme } from "../ui/theme-adapter.ts";
@@ -82,15 +81,10 @@ export function createRlmTool(controller: RlmController): ToolDefinition<typeof
 
       try {
         const input: StartInput = {
-          kind: "fresh",
           rootPrompt: params.prompt,
           context: params.context ?? undefined,
         };
-        const interactive = createPiInteractiveDeps(ctx);
-        const { done } = controller.start(ctx, input, emitter, {
-          onAskUserQuestion: controller.config.askUserQuestion ? interactive.onAskUserQuestion : undefined,
-          onTodo: controller.config.todo ? interactive.onTodo : undefined,
-        });
+        const { done } = controller.start(ctx, input, emitter);
         const result = await done;
 
         emitter.emitAnswer(result.answer);
@@ -152,11 +146,6 @@ function renderExpanded(details: RlmDetails, theme: Theme): Component {
     container.addChild(new Markdown(details.answer, 0, 0, markdownTheme(theme)));
   }
 
-  if (details.warnings && details.warnings.length > 0) {
-    container.addChild(new Spacer(1));
-    container.addChild(new Text(theme.fg("muted", details.warnings.join("\n")), 0, 0));
-  }
-
   return container;
 }
 

package/src/ui/config-panel.ts

@@ -9,21 +9,17 @@ const CHOICES = Object.freeze({
   maxDepth: Object.freeze(["1", "2", "3", "4"]),
   maxIterations: Object.freeze(["10", "20", "30", "50"]),
   execTimeoutS: Object.freeze(["30", "60", "120", "300"]),
-  maxConcurrentSubcalls: Object.freeze(["2", "4", "8", "16"]),
-  maxBudgetUsd: Object.freeze(["none", "0.50", "1", "5"]),
+  maxConcurrentSubcalls: Object.freeze(["2", "4", "8", "16", "32"]),
+  maxConcurrentChildren: Object.freeze(["1", "2", "3", "4", "6", "8"]),
   maxTimeoutMs: Object.freeze(["none", "60", "120", "300"]),
   maxTokens: Object.freeze(["none", "10000", "50000", "100000"]),
   maxErrors: Object.freeze(["3", "5", "10", "none"]),
   orchestrator: Object.freeze(["on", "off"]),
-  pipeline: Object.freeze(["on", "off"]),
-  maxBackwardJumps: Object.freeze(["0", "1", "2", "3"]),
   compaction: Object.freeze(["on", "off"]),
   compactionThresholdPct: Object.freeze(["50", "65", "80", "90"]),
   rootSamplingMaxTokens: Object.freeze(["4096", "8192", "16384", "32768"]),
   sandboxInitTimeoutMs: Object.freeze(["10000", "30000", "60000", "120000"]),
   requestTimeoutMs: Object.freeze(["2", "5", "10", "20"]),
-  askUserQuestion: Object.freeze(["on", "off"]),
-  todo: Object.freeze(["on", "off"]),
   libraryLoader: Object.freeze(["on", "off"]),
 });
 
@@ -43,20 +39,16 @@ export async function showConfigPanel(ctx: ExtensionContext, config: RlmConfig):
     item("maxIterations", "Max iterations", String(config.maxIterations), CHOICES.maxIterations, "Maximum root REPL turns before RLM asks the model for a final answer."),
     item("execTimeoutS", "REPL block timeout (s)", String(config.execTimeoutS), CHOICES.execTimeoutS, "Wall-clock limit for one model-authored Python REPL block."),
     item("maxConcurrentSubcalls", "Max concurrent sub-calls", String(config.maxConcurrentSubcalls), CHOICES.maxConcurrentSubcalls, "Concurrency pool size for llm_query_batched and rlm_query_batched."),
-    item("maxBudgetUsd", "Budget ceiling (USD)", config.maxBudgetUsd != null ? String(config.maxBudgetUsd) : "none", CHOICES.maxBudgetUsd, "Total spend cap for the whole recursive tree; none disables the cap."),
+    item("maxConcurrentChildren", "Max concurrent children", String(config.maxConcurrentChildren), CHOICES.maxConcurrentChildren, "Concurrent rlm_query child engines per depth. Each is a Python process holding its own copy of the inherited context."),
     item("maxTimeoutMs", "Wall-clock ceiling (min)", config.maxTimeoutMs != null ? String(Math.round(config.maxTimeoutMs / 60_000)) : "none", CHOICES.maxTimeoutMs, "Total runtime cap for the whole recursive tree; none disables the cap."),
     item("maxTokens", "Token ceiling", config.maxTokens != null ? String(config.maxTokens) : "none", CHOICES.maxTokens, "Total input+output token cap for the whole recursive tree."),
     item("maxErrors", "Max consecutive errors", config.maxErrors != null ? String(config.maxErrors) : "none", CHOICES.maxErrors, "Stop after this many consecutive failing turns; none disables the guard."),
     item("orchestrator", "Orchestrator addendum", config.orchestrator ? "on" : "off", CHOICES.orchestrator, "Append extra divide-and-conquer guidance to the root model system prompt."),
-    item("pipeline", "Phase pipeline", config.pipeline ? "on" : "off", CHOICES.pipeline, "Enable artifact-gated phases: clarify→research→blueprint→validate (read-only plan pipeline; clarify needs Ask user on)."),
-    item("maxBackwardJumps", "Max validate→blueprint loops", String(config.maxBackwardJumps), CHOICES.maxBackwardJumps, "Bounded corrective re-entries when validation reports blockers_count > 0."),
     item("compaction", "Trajectory compaction", config.compaction ? "on" : "off", CHOICES.compaction, "Summarize old turns when history approaches the model context window."),
     item("compactionThresholdPct", "Compaction threshold (%)", String(Math.round(config.compactionThresholdPct * 100)), CHOICES.compactionThresholdPct, "Compact once estimated history tokens reach this share of the root model's context window."),
     item("rootSamplingMaxTokens", "Root model output cap (tok)", String(config.rootSampling?.maxTokens ?? 16384), CHOICES.rootSamplingMaxTokens, "Max output tokens per root-model turn. Lower values keep each turn lean."),
     item("sandboxInitTimeoutMs", "Sandbox init timeout", String(config.sandboxInitTimeoutMs), CHOICES.sandboxInitTimeoutMs, "How long to wait for the Python worker to start."),
     item("requestTimeoutMs", "Sandbox request timeout (min)", String(Math.round(config.requestTimeoutMs / 60_000)), CHOICES.requestTimeoutMs, "Parent-side watchdog per sandbox request; on breach the Python worker is killed."),
-    item("askUserQuestion", "[Interactive] Ask user", config.askUserQuestion ? "on" : "off", CHOICES.askUserQuestion, "Allow root REPL code to present structured ask_user_question dialogs."),
-    item("todo", "[Interactive] Todo", config.todo ? "on" : "off", CHOICES.todo, "Allow REPL code to manage a visible todo task list."),
     item("libraryLoader", "Library loader", config.libraryLoader ? "on" : "off", CHOICES.libraryLoader,
       "Allow load_library() to pull an external dir, file, or git repo into the shared context list."),
     item("__save__", "Save & close", "↵", ["↵"], "Save these settings and close (Esc also saves)."),
@@ -101,21 +93,17 @@ export function applySetting(config: RlmConfig, id: string, value: string): RlmC
     case "maxIterations": return Object.freeze({ ...config, maxIterations: Number(value) });
     case "execTimeoutS": return Object.freeze({ ...config, execTimeoutS: Number(value) });
     case "maxConcurrentSubcalls": return Object.freeze({ ...config, maxConcurrentSubcalls: Number(value) });
-    case "maxBudgetUsd": return Object.freeze({ ...config, maxBudgetUsd: optionalNumber(value) });
+    case "maxConcurrentChildren": return Object.freeze({ ...config, maxConcurrentChildren: Number(value) });
     case "maxTimeoutMs": return Object.freeze({ ...config, maxTimeoutMs: optionalNumber(value, 60_000) });
     case "maxTokens": return Object.freeze({ ...config, maxTokens: optionalNumber(value) });
     case "maxErrors": return Object.freeze({ ...config, maxErrors: optionalNumber(value) });
     case "orchestrator": return Object.freeze({ ...config, orchestrator: value === "on" });
-    case "pipeline": return Object.freeze({ ...config, pipeline: value === "on" });
-    case "maxBackwardJumps": return Object.freeze({ ...config, maxBackwardJumps: Number(value) });
     case "compaction": return Object.freeze({ ...config, compaction: value === "on" });
     case "compactionThresholdPct": return Object.freeze({ ...config, compactionThresholdPct: Number(value) / 100 });
     case "rootSamplingMaxTokens":
       return Object.freeze({ ...config, rootSampling: Object.freeze({ ...config.rootSampling, maxTokens: Number(value) }) });
     case "sandboxInitTimeoutMs": return Object.freeze({ ...config, sandboxInitTimeoutMs: Number(value) });
     case "requestTimeoutMs": return Object.freeze({ ...config, requestTimeoutMs: Number(value) * 60_000 });
-    case "askUserQuestion": return Object.freeze({ ...config, askUserQuestion: value === "on" });
-    case "todo": return Object.freeze({ ...config, todo: value === "on" });
     case "libraryLoader": return Object.freeze({ ...config, libraryLoader: value === "on" });
     default: return config;
   }

package/src/ui/intro.ts

@@ -11,9 +11,8 @@ export const RLM_GUIDE = `# RLM mode
 ## Commands
 
 - \`/rlm\` — toggle RLM mode (shortcut: Ctrl+Shift+R). Turning it OFF also stops a running query.
-- \`/rlm-config\` — choose models, reasoning, and budget limits
+- \`/rlm-config\` — choose models, reasoning, and run limits
 - \`/rlm-stop\` — cancel the current run but stay in RLM mode (use /rlm or Ctrl+Shift+R to leave)
-- \`/rlm-help\` — show this guide again
 
 When RLM mode is ON, \`read\`/\`grep\` are disabled and the agent reads the repository through the
 \`repl\` tool, delegating bulk analysis to sub-LLMs. The footer/status line shows the current state.`;

package/src/ui/model-picker.ts

@@ -1,10 +1,11 @@
 /** Model picker TUI — choose a model and, when supported, a thinking level. */
 
-import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
+import type { ExtensionContext, ModelRegistry } from "@earendil-works/pi-coding-agent";
 import { DynamicBorder } from "@earendil-works/pi-coding-agent";
 import type { Api, Model, ThinkingLevel } from "@earendil-works/pi-ai";
 import { Container, type Component, type SelectItem, SelectList, Text, truncateToWidth } from "@earendil-works/pi-tui";
 import { formatCost } from "./theme.ts";
+import { compareLlm } from "../mode/llm-model.ts";
 
 export interface ModelSelection {
   readonly model: Model<Api>;
@@ -16,15 +17,38 @@ type SelectableThinkingLevel = (typeof LEVELS)[number];
 
 const CHEAPEST_VALUE = "__rlm_cheapest__";
 
-function items(models: Model<Api>[], includeCheapest = false): SelectItem[] {
-  const modelItems = models.map((m) => ({
-    value: `${m.provider}/${m.id}`,
-    label: `${m.provider}/${m.id}`,
-    description: `in ${formatCost(m.cost.input)}/Mtok · out ${formatCost(m.cost.output)}/Mtok${m.reasoning ? " · reasoning" : ""}`,
-  }));
+/**
+ * Models Pi itself would offer for this session, cheapest-first.
+ *
+ * Mirrors the built-in model switcher:
+ *   - if the session has scoped models (`--models` / enabledModels) → those only
+ *   - else → `getAvailable()` (providers with configured auth)
+ *
+ * Deliberately NOT `getAll()`: the full catalog dumps every provider's catalog entry and is
+ * not what the user sees in Pi natively. See Pi extension docs on `ctx.scopedModels`.
+ */
+export function pickableModels(
+  registry: ModelRegistry,
+  scoped?: readonly { readonly model: Model<Api> }[],
+): readonly Model<Api>[] {
+  const source = scoped !== undefined && scoped.length > 0
+    ? scoped.map((s) => s.model)
+    : registry.getAvailable();
+  return [...source].sort(compareLlm);
+}
+
+function items(models: readonly Model<Api>[], includeCheapest: boolean): SelectItem[] {
+  const modelItems = models.map((m) => {
+    const price = `in ${formatCost(m.cost.input)}/Mtok · out ${formatCost(m.cost.output)}/Mtok`;
+    return {
+      value: `${m.provider}/${m.id}`,
+      label: `${m.provider}/${m.id}`,
+      description: `${price}${m.reasoning ? " · reasoning" : ""}`,
+    };
+  });
   if (!includeCheapest) return modelItems;
   return [
-    { value: CHEAPEST_VALUE, label: "⟳ cheapest (auto)", description: "Always use the cheapest available model" },
+    { value: CHEAPEST_VALUE, label: "⟳ cheapest (auto)", description: "Always use the cheapest model with a configured key" },
     ...modelItems,
   ];
 }
@@ -79,12 +103,12 @@ async function selectThinkingLevel(
 export async function selectModel(
   ctx: ExtensionContext,
   title: string,
-  models: Model<Api>[],
+  models: readonly Model<Api>[],
   current?: Model<Api>,
   currentThinking?: ThinkingLevel,
 ): Promise<ModelSelection | null | undefined> {
   if (models.length === 0) {
-    ctx.ui.notify("RLM: no models with configured auth", "warning");
+    ctx.ui.notify("RLM: no models available (add a provider key in Pi, or widen --models / enabledModels)", "warning");
     return undefined;
   }
   if (ctx.mode !== "tui") {

package/src/ui/status.ts

@@ -12,18 +12,14 @@ export function modelLabel(model: Model<Api> | undefined, fallback: string): str
 
 export function formatRlmStateLine(controller: RlmController, contextUsage?: ContextUsage): string {
   if (!controller.enabled) return "○ RLM OFF";
-  const worker = modelLabel(controller.workerModel, controller.savedWorkerRef ?? "cheapest");
-  const workerSuffix = controller.config.subSampling.reasoning ? `:${controller.config.subSampling.reasoning}` : "";
+  const llm = modelLabel(controller.llmModel, controller.savedLlmRef ?? "cheapest");
+  const llmSuffix = controller.config.subSampling.reasoning ? `:${controller.config.subSampling.reasoning}` : "";
   // `percent` is null right after a compaction, before the next assistant response reports usage.
   const percent = contextUsage?.percent;
   const ctxSuffix = percent === null || percent === undefined ? "" : ` · ctx ${Math.round(percent)}%`;
-  return `● RLM ON · worker=${worker}${workerSuffix}${ctxSuffix}`;
+  return `● RLM ON · llm=${llm}${llmSuffix}${ctxSuffix}`;
 }
 
 export function setRlmModeStatus(ui: ExtensionUIContext, controller: RlmController, contextUsage?: ContextUsage): void {
   ui.setStatus(KEY, formatRlmStateLine(controller, contextUsage));
 }
-
-export function clearRlmStatus(ui: ExtensionUIContext): void {
-  ui.setStatus(KEY, undefined);
-}