@hicaru/pi-rlm 0.2.0 → 0.2.2

package/src/core/limits.ts

@@ -1,7 +1,9 @@
 /**
- * LimitGuard — wall-clock, token, cost, and consecutive-error caps for a headless RLM run
+ * LimitGuard — wall-clock, token, and consecutive-error caps for a headless RLM run
  * (ported from rlm/core/rlm.py `_check_timeout` / `_check_iteration_limits`). Any breach throws
  * a LimitError; the engine catches it and returns the best partial answer it has.
+ *
+ * Cost is tracked for reporting only — there is no USD spend ceiling.
  */
 
 import type { Usage } from "@earendil-works/pi-ai";
@@ -9,14 +11,12 @@ import type { Usage } from "@earendil-works/pi-ai";
 export interface Limits {
   readonly maxTimeoutMs?: number;
   readonly maxTokens?: number;
-  readonly maxBudgetUsd?: number;
   readonly maxErrors?: number;
 }
 
 /** Pick the limit caps out of a config (`RlmConfig` satisfies this structurally). */
 export function limitsFromConfig(config: Limits): Limits {
   return {
-    maxBudgetUsd: config.maxBudgetUsd,
     maxTimeoutMs: config.maxTimeoutMs,
     maxTokens: config.maxTokens,
     maxErrors: config.maxErrors,
@@ -33,7 +33,7 @@ export interface UsageSnapshot {
 
 export class LimitError extends Error {
   constructor(
-    public readonly kind: "timeout" | "tokens" | "budget" | "errors",
+    public readonly kind: "timeout" | "tokens" | "errors",
     message: string,
   ) {
     super(message);
@@ -77,16 +77,13 @@ export class LimitGuard {
   /** Call after each turn with whether the turn's REPL produced an error. */
   observe(hadError: boolean): void {
     this.consecutiveErrors = hadError ? this.consecutiveErrors + 1 : 0;
-    const { maxErrors, maxTokens, maxBudgetUsd } = this.limits;
+    const { maxErrors, maxTokens } = this.limits;
     if (maxErrors && this.consecutiveErrors >= maxErrors) {
       throw new LimitError("errors", `${this.consecutiveErrors} consecutive errors (limit ${maxErrors})`);
     }
     if (maxTokens && this.inputTokens + this.outputTokens > maxTokens) {
       throw new LimitError("tokens", `${this.inputTokens + this.outputTokens} tokens (limit ${maxTokens})`);
     }
-    if (maxBudgetUsd && this.costUsd > maxBudgetUsd) {
-      throw new LimitError("budget", `$${this.costUsd.toFixed(4)} spent (limit $${maxBudgetUsd})`);
-    }
   }
 
   usage(): UsageSnapshot {
@@ -98,10 +95,6 @@ export class LimitGuard {
     };
   }
 
-  remainingBudgetUsd(): number | undefined {
-    return this.limits.maxBudgetUsd === undefined ? undefined : this.limits.maxBudgetUsd - this.costUsd;
-  }
-
   remainingTimeoutMs(): number | undefined {
     return this.limits.maxTimeoutMs === undefined ? undefined : this.limits.maxTimeoutMs - (Date.now() - this.start);
   }

package/src/core/resource-limits.ts

@@ -3,12 +3,10 @@
 import { formatError } from "../util/errors.ts";
 
 export interface RemainingResources {
-  readonly budgetUsd?: number;
   readonly timeoutMs?: number;
 }
 
 export function checkResourceLimits(resources: RemainingResources): string | undefined {
-  if (resources.budgetUsd !== undefined && resources.budgetUsd <= 0) return formatError("budget exhausted");
   if (resources.timeoutMs !== undefined && resources.timeoutMs <= 0) return formatError("timeout exhausted");
   return undefined;
 }

package/src/core/types.ts

@@ -1,8 +1,6 @@
 /** Shared configuration + runtime types for the RLM engine. */
 
 import type { ThinkingLevel } from "@earendil-works/pi-ai";
-import type { AskAnswer, AskQuestion } from "../sandbox/protocol.ts";
-import type { ReconstructResult } from "../state/resume.ts";
 
 export interface Sampling {
   readonly maxTokens?: number;
@@ -10,17 +8,6 @@ export interface Sampling {
   readonly reasoning?: ThinkingLevel;
 }
 
-export interface RunLogConfig {
-  /** Default: true — always-on, opt-out. */
-  readonly enabled?: boolean;
-  /** Default: ".rlm/runs". Directory under cwd for run artifacts. */
-  readonly dir?: string;
-  /** Default: true — whether to write sandbox.pkl snapshots. */
-  readonly snapshot?: boolean;
-  /** Default: 50 — prune oldest runs beyond this count on each new run. */
-  readonly maxRuns?: number;
-}
-
 export interface RlmConfig {
   /** Persistent editor-routing mode; when enabled, plain interactive prompts use RLM. */
   readonly enabled: boolean;
@@ -34,10 +21,11 @@ export interface RlmConfig {
   readonly requestTimeoutMs: number;
   /** Concurrency pool for *_batched sub-calls. */
   readonly maxConcurrentSubcalls: number;
+  /** Concurrent recursive child engines admitted per depth. Lower than maxConcurrentSubcalls:
+   *  each child is a Python subprocess holding its own copy of the inherited context. */
+  readonly maxConcurrentChildren: number;
   /** Reject sub-LLM prompts larger than this many chars. */
   readonly maxPromptChars: number;
-  /** Max USD spend across the whole tree before the engine stops (undefined = no cap). */
-  readonly maxBudgetUsd?: number;
   /** Max wall-clock ms across the whole tree before the engine stops (undefined = no cap). */
   readonly maxTimeoutMs?: number;
   /** Max total input+output tokens across the whole tree before the engine stops (undefined = no cap). */
@@ -46,10 +34,6 @@ export interface RlmConfig {
   readonly maxErrors?: number;
   /** Append the orchestrator addendum to the system prompt. */
   readonly orchestrator: boolean;
-  /** Enable the phase pipeline (advance_phase + stall nags) at depth 0. */
-  readonly pipeline: boolean;
-  /** Max validate→blueprint corrective re-entries when validation reports blockers (default 2). */
-  readonly maxBackwardJumps: number;
   /** Summarize the trajectory when it grows past the threshold (keeps the root window small). */
   readonly compaction: boolean;
   /** Compact when estimated history tokens reach this fraction of the model's context window. */
@@ -58,10 +42,6 @@ export interface RlmConfig {
   readonly python: string;
   /** Worker startup wait before treating sandbox init as failed (ms). */
   readonly sandboxInitTimeoutMs: number;
-  /** Allow ask_user_question() calls from the root REPL. */
-  readonly askUserQuestion: boolean;
-  /** Allow todo() calls from the REPL. */
-  readonly todo: boolean;
   /** Enable the load_library() REPL scaffold (external dirs/files/git repos as extra context slots). */
   readonly libraryLoader: boolean;
   /** ThinkingLevel for the root smart model (set via /rlm-config). */
@@ -76,8 +56,6 @@ export interface RlmConfig {
   readonly subSystemPrompt?: string;
   /** Sampling for sub-LLM (worker) calls. */
   readonly subSampling: Readonly<Sampling>;
-  /** Optional run-state persistence configuration. Enabled by default. */
-  readonly runLog?: RunLogConfig;
 }
 
 /** Input to a (headless) RLM run. */
@@ -92,12 +70,8 @@ export interface RlmInput {
   readonly parentNodeId?: string;
   /** "provider/id" — overrides the root model for this run (set by recursive rlm_query). */
   readonly modelOverride?: string;
-  /** Remaining budget for this subtree (set by parent from its LimitGuard). */
-  readonly remainingBudgetUsd?: number;
   /** Remaining timeout for this subtree (set by parent from its LimitGuard). */
   readonly remainingTimeoutMs?: number;
-  /** Depth-0 resume payload — controller rebuilds this from the trail's `reconstructRlmState()`. */
-  readonly resume?: ReconstructResult & { readonly ok: true };
 }
 
 /** Result of a completed RLM run. */
@@ -111,11 +85,4 @@ export interface RlmResult {
 }
 
 /** A function that runs an RLM to completion — used to wire recursion (rlm_query). */
-export interface InteractiveDeps {
-  /** Called when the sandbox issues ask_user_question; undefined = feature disabled. */
-  readonly onAskUserQuestion?: (questions: readonly AskQuestion[]) => Promise<AskAnswer[]>;
-  /** Called when the sandbox issues todo; undefined = feature disabled. */
-  readonly onTodo?: (action: string, params: Record<string, unknown>) => Promise<string>;
-}
-
 export type RunRlm = (input: RlmInput) => Promise<RlmResult>;

package/src/index.ts

@@ -7,17 +7,22 @@ import { registerRlmConfigCommand } from "./commands/rlm-config.ts";
 import { createRlmTool } from "./tool/rlm-tool.ts";
 import { createReplTool } from "./tool/repl-tool.ts";
 import { loadSettings, mergeConfig, resolveModelId } from "./config/settings.ts";
-import { RlmController, cheapestModel } from "./mode/rlm-mode.ts";
+import { RlmController } from "./mode/rlm-mode.ts";
+import { cheapestModel } from "./mode/llm-model.ts";
 import { postRlmGuide } from "./ui/intro.ts";
 import { setRlmModeStatus } from "./ui/status.ts";
 import { markdownTheme } from "./ui/theme-adapter.ts";
 import { SandboxManager } from "./sandbox/sandbox-manager.ts";
+import { createSubcallGates } from "./util/concurrency.ts";
+import { BackgroundTasks } from "./tool/background-tasks.ts";
 import { packRepository, formatForLLM, serializeForSandbox } from "./context/repomix-context.ts";
-import { buildNativeSystemPrompt, NATIVE_TURN_REMINDER } from "./prompts/system.ts";
+import { buildNativeSystemPrompt, NATIVE_TURN_REMINDER } from "./prompts/native.ts";
 import { bashCommandFromInput, isFileReadingCommand, capToolResultText, BASH_BLOCK_REASON } from "./mode/native-guards.ts";
 import { errorMessage } from "./util/errors.ts";
 
 const BLOCKED_NATIVE_TOOLS = Object.freeze(new Set(["read", "grep"]));
+/** How often to keep the parent sandbox's request watchdog alive during detached work. */
+const WATCHDOG_HEARTBEAT_MS = 30_000;
 const CAPPED_RESULT_TOOLS = Object.freeze(new Set(["bash", "find", "ls"]));
 
 export default function rlmExtension(pi: ExtensionAPI): void {
@@ -31,8 +36,27 @@ export default function rlmExtension(pi: ExtensionAPI): void {
     python: config.python,
     sandboxInitTimeoutMs: config.sandboxInitTimeoutMs,
     maxPromptChars: config.maxPromptChars,
+    // Same wall budget as the parent request watchdog — a stalled sub-call should surface
+    // inside the cell rather than hang the session forever.
+    awaitTimeoutS: Math.round(config.requestTimeoutMs / 1000),
     onSandboxDiscarded: () => { onSandboxDiscardExtra?.(); },
   });
+  // One admission gate for the whole session: spawn() lets the sandbox put many requests on
+  // the wire at once, so nothing smaller than session scope actually bounds fan-out.
+  const gates = createSubcallGates(config.maxConcurrentSubcalls, config.maxConcurrentChildren);
+  const background = new BackgroundTasks({
+    maxTimeoutMs: config.maxTimeoutMs,
+    maxTokens: config.maxTokens,
+    maxErrors: config.maxErrors,
+  });
+  // A detached child works in its OWN sandbox, so this one sees no frames and its request
+  // watchdog would fire mid-await and SIGKILL a healthy worker, taking the REPL namespace
+  // with it. Keep it alive while detached work is genuinely in flight.
+  const watchdogHeartbeat = setInterval(() => {
+    if (background.pending > 0) sandboxManager.refreshWatchdog();
+  }, WATCHDOG_HEARTBEAT_MS);
+  watchdogHeartbeat.unref();
+
   let packedContextText: string | undefined;
   let contextPackPromise: Promise<string | undefined> | undefined;
   const ensureRepositoryContext = async (cwd: string): Promise<string | undefined> => {
@@ -55,7 +79,7 @@ export default function rlmExtension(pi: ExtensionAPI): void {
   const settingsReady = loadSettings()
     .then((persisted) => {
       controller.config = mergeConfig(persisted.config);
-      controller.savedWorkerRef = persisted.worker;
+      controller.savedLlmRef = persisted.llm;
     })
     .catch((err) => {
       console.warn(`[rlm] settings load failed: ${errorMessage(err)}`);
@@ -96,24 +120,37 @@ export default function rlmExtension(pi: ExtensionAPI): void {
     const flag = pi.getFlag("rlm");
     if (typeof flag === "boolean") controller.setConfig(Object.freeze({ ...controller.config, enabled: flag }));
 
-    if (controller.savedWorkerRef) {
-      const resolved = resolveModelId(ctx.modelRegistry, controller.savedWorkerRef);
-      if (resolved) controller.workerModel = resolved;
+    // Reload the catalog before the worker-model pick below reads it. Newer pi builds make
+    // `getAvailable()` an async-populated snapshot that starts empty, and picking from an empty
+    // catalog silently falls back to the root model. Called with no arguments and awaited so it
+    // is valid whether `refresh` returns void (current) or a promise (newer); fail-soft, because
+    // a refresh error must not abort session start.
+    try {
+      await ctx.modelRegistry.refresh();
+    } catch (err) {
+      console.warn(`[rlm] model registry refresh failed: ${errorMessage(err)}`);
+    }
+
+    if (controller.savedLlmRef) {
+      const resolved = resolveModelId(ctx.modelRegistry, controller.savedLlmRef);
+      if (resolved) controller.llmModel = resolved;
     }
 
     // Re-register repl tool each session to pick up model provider changes
-    const workerModel = controller.workerModel ?? cheapestModel(ctx.modelRegistry) ?? ctx.model;
+    const llmModel = controller.llmModel ?? cheapestModel(ctx.modelRegistry) ?? ctx.model;
     const model = ctx.model;
-    if (workerModel && model) {
+    if (llmModel && model) {
       try {
         pi.registerTool(createReplTool({
           sandboxManager,
           model,
-          workerModel,
+          llmModel,
           getModel: () => controller.resolveModels(ctx)?.model,
-          getWorkerModel: () => controller.resolveModels(ctx)?.worker,
+          getLlmModel: () => controller.resolveModels(ctx)?.llm,
           registry: ctx.modelRegistry,
           getConfig: () => controller.config,
+          gates,
+          background,
           registerDiscardHook: (reset) => { onSandboxDiscardExtra = reset; },
           ensureContext: async () => {
             const contextText = await ensureRepositoryContext(ctx.cwd ?? process.cwd());
@@ -224,6 +261,8 @@ export default function rlmExtension(pi: ExtensionAPI): void {
   // ── Session shutdown: cleanup ──
   pi.on("session_shutdown", async () => {
     controller.abort();
+    clearInterval(watchdogHeartbeat);
+    background.dispose();
     await sandboxManager.dispose();
     contextInjected = false;
     packedContextText = undefined;

package/src/mode/llm-model.ts

@@ -0,0 +1,54 @@
+/**
+ * Sub-LLM model ranking — "cheapest available", with free models winning outright.
+ *
+ * Pi's `ModelCost` is non-nullable (`packages/ai/src/types.ts`), so a free model is a literal 0,
+ * not a null. That makes plain price sorting ambiguous rather than wrong: subscription and
+ * token-plan providers also publish 0, and a stable sort would hand back whichever 0-cost entry
+ * happened to be first in catalog order. The tie-breaks below are what actually pick a usable
+ * free model — and what make the pick identical across sessions and catalog reorderings.
+ */
+
+import type { Api, Model } from "@earendil-works/pi-ai";
+import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
+
+/** $/Mtok, input-weighted 3:1 — a sub-call sends a file body and gets back a sentence. */
+function priceOf(model: Model<Api>): number {
+  const { input, output, cacheRead } = model.cost;
+  return input * 3 + output + cacheRead;
+}
+
+/** True when the model costs nothing to call on any axis. */
+export function isFreeModel(model: Model<Api>): boolean {
+  return priceOf(model) === 0;
+}
+
+/**
+ * Negative when `a` is the better sub-LLM.
+ *
+ * Window before maxTokens before id: a free model with a 4K context is useless for bulk reading,
+ * so price alone must not decide. The final id comparison exists only to make the result
+ * deterministic — without it the pick drifts whenever a provider reorders its catalog.
+ */
+export function compareLlm(a: Model<Api>, b: Model<Api>): number {
+  return (priceOf(a) - priceOf(b))
+    || (b.contextWindow - a.contextWindow)
+    || (b.maxTokens - a.maxTokens)
+    || `${a.provider}/${a.id}`.localeCompare(`${b.provider}/${b.id}`);
+}
+
+/**
+ * Best sub-LLM among the models whose provider has configured auth.
+ *
+ * Single pass rather than `[...models].sort()[0]`: the copy and the sort both allocate for a
+ * result that is one element.
+ */
+export function cheapestModel(registry: ModelRegistry): Model<Api> | undefined {
+  const models = registry.getAvailable();
+  let best: Model<Api> | undefined;
+  for (let i = 0; i < models.length; i++) {
+    const model = models[i];
+    if (model === undefined) continue;
+    if (best === undefined || compareLlm(model, best) < 0) best = model;
+  }
+  return best;
+}

package/src/mode/rlm-mode.ts

@@ -1,42 +1,35 @@
 /**
  * RlmController — holds RLM config + chosen models.
  *
- * The engine drives the root model turn-by-turn over ```repl``` blocks with full budget/token/
+ * The engine drives the root model turn-by-turn over ```repl``` blocks with full token/
  * timeout/error guards, compaction, and a finalize fallback. `start()` returns a RunHandle with
  * the completion promise.
  */
 
 import type { Api, Model } from "@earendil-works/pi-ai";
-import type { ExtensionContext, ModelRegistry } from "@earendil-works/pi-coding-agent";
-import { DEFAULT_RUN_DIR } from "../config/defaults.ts";
+import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { modelRef, resolveModelId, saveSettings } from "../config/settings.ts";
 import { createEngine } from "../core/engine.ts";
 import { limitsFromConfig } from "../core/limits.ts";
-import type { InteractiveDeps, RlmConfig, RlmInput, RlmResult } from "../core/types.ts";
-import type { ReconstructResult } from "../state/resume.ts";
+import type { RlmConfig, RlmResult } from "../core/types.ts";
 import { packRepository, serializeForSandbox } from "../context/repomix-context.ts";
 import { RlmEmitter } from "../tool/rlm-events.ts";
 import { formatError } from "../util/errors.ts";
-
-export function cheapestModel(registry: ModelRegistry): Model<Api> | undefined {
-  const models = registry.getAvailable();
-  if (models.length === 0) return undefined;
-  return [...models].sort((a, b) => a.cost.input + a.cost.output - (b.cost.input + b.cost.output))[0];
-}
+import { cheapestModel } from "./llm-model.ts";
 
 export interface RunHandle {
   readonly abort: () => void;
   readonly done: Promise<RlmResult>;
 }
 
-/** B5+SA: discriminated union removes non-null `!` assertions and the `context: ""` hack. */
-export type StartInput =
-  | { readonly kind: "fresh"; readonly rootPrompt: string; readonly context: unknown }
-  | { readonly kind: "resume"; readonly resume: ReconstructResult & { ok: true }; readonly context: unknown };
+export interface StartInput {
+  readonly rootPrompt: string;
+  readonly context: unknown;
+}
 
 export class RlmController {
-  workerModel: Model<Api> | undefined;
-  savedWorkerRef: string | undefined;
+  llmModel: Model<Api> | undefined;
+  savedLlmRef: string | undefined;
   private active: AbortController | null = null;
 
   constructor(public config: RlmConfig) {}
@@ -65,7 +58,7 @@ export class RlmController {
   async persist(): Promise<boolean> {
     return await saveSettings({
       config: this.config,
-      worker: modelRef(this.workerModel) ?? this.savedWorkerRef,
+      llm: modelRef(this.llmModel) ?? this.savedLlmRef,
     });
   }
 
@@ -77,15 +70,15 @@ export class RlmController {
     this.active?.abort();
   }
 
-  resolveModels(ctx: ExtensionContext): { model: Model<Api>; worker: Model<Api> } | undefined {
-    if (!this.workerModel && this.savedWorkerRef) this.workerModel = resolveModelId(ctx.modelRegistry, this.savedWorkerRef);
+  resolveModels(ctx: ExtensionContext): { model: Model<Api>; llm: Model<Api> } | undefined {
+    if (!this.llmModel && this.savedLlmRef) this.llmModel = resolveModelId(ctx.modelRegistry, this.savedLlmRef);
     const model = ctx.model ?? cheapestModel(ctx.modelRegistry);
     if (!model) return undefined;
-    const worker = this.workerModel ?? cheapestModel(ctx.modelRegistry) ?? model;
-    return { model, worker };
+    const llm = this.llmModel ?? cheapestModel(ctx.modelRegistry) ?? model;
+    return { model, llm };
   }
 
-  start(ctx: ExtensionContext, input: StartInput, emitter?: RlmEmitter, interactive?: InteractiveDeps): RunHandle {
+  start(ctx: ExtensionContext, input: StartInput, emitter?: RlmEmitter): RunHandle {
     const models = this.resolveModels(ctx);
     if (!models) throw new Error("no model with configured auth is available");
     if (this.active) throw new Error("RLM run already in progress"); // QC: mutual-exclusion guard
@@ -93,50 +86,26 @@ export class RlmController {
     const abortController = new AbortController();
     this.active = abortController;
 
-    const runState = this.config.runLog?.enabled !== false
-      ? { cwd: ctx.cwd ?? process.cwd(), dir: this.config.runLog?.dir ?? DEFAULT_RUN_DIR, snapshot: this.config.runLog?.snapshot !== false }
-      : undefined;
-
     const done = (async () => {
-      let engineInput: RlmInput;
-      if (input.kind === "fresh") {
-        // Auto-pack empty/undefined context via repomix; pass explicit context through.
-        let contextValue: unknown = input.context;
-        if (contextValue === undefined || contextValue === "" || (typeof contextValue === "string" && contextValue.trim() === "")) {
-          const cwd = ctx.cwd ?? process.cwd();
-          const result = await packRepository(cwd, abortController.signal);
-          if (result.ok) {
-            contextValue = serializeForSandbox(result.value);
-          } else {
-            contextValue = formatError(`failed to pack repository — ${result.error}`);
-          }
-        }
-        engineInput = {
-          rootPrompt: input.rootPrompt,
-          context: contextValue,
-          depth: 0,
-        };
-      } else {
-        engineInput = {
-          rootPrompt: input.resume.header.rootPrompt,
-          context: input.context, // B5: load the actual context from the sidecar, not ""
-          depth: 0,
-          resume: input.resume,
-        };
+      // Auto-pack empty/undefined context via repomix; pass explicit context through.
+      let contextValue: unknown = input.context;
+      if (contextValue === undefined || (typeof contextValue === "string" && contextValue.trim() === "")) {
+        const cwd = ctx.cwd ?? process.cwd();
+        const result = await packRepository(cwd, abortController.signal);
+        contextValue = result.ok
+          ? serializeForSandbox(result.value)
+          : formatError(`failed to pack repository — ${result.error}`);
       }
       const engine = createEngine({
         model: models.model,
-        workerModel: models.worker,
+        llmModel: models.llm,
         registry: ctx.modelRegistry,
         config: this.config,
         signal: abortController.signal,
         emitter: emitter ?? new RlmEmitter(),
-        runState,
-        onAskUserQuestion: interactive?.onAskUserQuestion,
-        onTodo: interactive?.onTodo,
         limits: limitsFromConfig(this.config),
       });
-      return await engine(engineInput);
+      return await engine({ rootPrompt: input.rootPrompt, context: contextValue, depth: 0 });
     })().finally(() => {
       if (this.active === abortController) this.active = null;
     });