@hicaru/pi-rlm 0.1.2 → 0.1.5

package/src/tool/repl-tool.ts

@@ -31,6 +31,7 @@ import type { ProposedEdit, ReplResult } from "../sandbox/protocol.ts";
 import { RlmEmitter } from "./rlm-events.ts";
 import { SubcallStore } from "./subcall-store.ts";
 import type { ReplDetails } from "./repl-details.ts";
+import type { RlmSubcall } from "./rlm-details.ts";
 import { createEngine } from "../core/engine.ts";
 import { formatCost, formatTokens, spinnerFrame } from "../ui/theme.ts";
 import { errorMessage, formatError, isErrorText } from "../util/errors.ts";
@@ -40,6 +41,7 @@ import {
   renderExpandedSubcallTree,
 } from "./subcall-render.ts";
 import { createProgressNotifier, validateToolParams } from "./tool-utils.ts";
+import { capReplResultText, replDelegationNudge } from "../mode/native-guards.ts";
 
 // ── Parameter schema ──
 
@@ -51,6 +53,37 @@ export function surfaceReplEdits(edits: readonly ProposedEdit[], raised: boolean
   return edits.length > 0 && !raised ? edits : undefined;
 }
 
+/** Model-visible text assembled from a repl() result, plus the surfaced edits for `details`. */
+export interface ReplResultText {
+  readonly text: string;
+  readonly surfacedEdits: readonly ProposedEdit[] | undefined;
+}
+
+/**
+ * Assemble the model-visible text for a repl() result: cap stdout, append a zero-subcall
+ * delegation nudge (suppressed when edits were staged), and append the STAGED_EDITS block
+ * AFTER capping so edit JSON is never truncated. Extracted as a pure function so the
+ * capping/ordering invariants are testable independently of the sandbox.
+ */
+export function buildReplResultText(
+  stdout: string,
+  answerContent: string | undefined,
+  edits: readonly ProposedEdit[],
+  raised: boolean,
+  subcalls: readonly RlmSubcall[],
+): ReplResultText {
+  const rawText = stdout || answerContent || "(no output)";
+  const surfacedEdits = surfaceReplEdits(edits, raised);
+  const editsBlock = surfacedEdits
+    ? `\n\nSTAGED_EDITS:\n${JSON.stringify(surfacedEdits)}`
+    : "";
+  // Model-visible text is capped; the caller keeps full stdout in `details.output` for the TUI.
+  const cappedText = capReplResultText(rawText) ?? rawText;
+  const delegated = subcalls.some((s) => s.kind === "llm" || s.kind === "batch" || s.kind === "rlm");
+  const nudge = surfacedEdits ? undefined : replDelegationNudge(rawText.length, delegated);
+  return { text: cappedText + (nudge ?? "") + editsBlock, surfacedEdits };
+}
+
 // ── Mutable bridge state (handler indirection) ──
 
 /**
@@ -91,6 +124,11 @@ class NativeBridgeState {
       modelRef(model ? (resolveModelId(deps.registry, model) ?? workerModel()) : workerModel()) ?? workerModel().id;
 
     async function complete1(prompt: string, model: string | null, track: (u: Usage) => void): Promise<string> {
+      const limits = state.currentLimits;
+      if (limits) {
+        const limitError = checkResourceLimits({ budgetUsd: limits.remainingBudgetUsd(), timeoutMs: limits.remainingTimeoutMs() });
+        if (limitError !== undefined) return limitError;
+      }
       if (prompt.length > deps.maxPromptChars) {
         return formatError(`sub-LLM prompt exceeded size limit (${prompt.length.toLocaleString()} chars > ${deps.maxPromptChars.toLocaleString()})`);
       }
@@ -107,6 +145,7 @@ class NativeBridgeState {
           reasoning: deps.sampling?.reasoning,
           signal: deps.signal,
         });
+        limits?.addUsage(res.usage);
         track(res.usage);
         return res.text;
       } catch (err) {
@@ -132,7 +171,6 @@ class NativeBridgeState {
           costUsd: cost, tokens, resultPreview: previewText(out),
           detail: isErrorText(out) ? out : undefined,
         });
-        state.currentLimits?.addRaw(cost, 0, tokens);
         return out;
       },
 
@@ -155,7 +193,6 @@ class NativeBridgeState {
           status: error ? "error" : "done", costUsd: cost, tokens,
           resultPreview: previewText(out[0] ?? ""), detail: error,
         });
-        state.currentLimits?.addRaw(cost, 0, tokens);
         return out;
       },
     };
@@ -317,7 +354,13 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
   return {
     name: "repl",
     label: "REPL",
-    description: "Execute Python code in a persistent REPL sandbox with the full repository context pre-loaded. Variables, imports, and state persist across calls. Supports llm_query, rlm_query, todo, and ask_user_question inside the sandbox.",
+    description:
+      "PRIMARY tool for ALL repository reading and analysis (read/grep are disabled in RLM mode). " +
+      "Persistent Python sandbox with every file pre-loaded in `context`. You are an orchestrator: " +
+      "chunk `context` and delegate semantic work to llm_query / llm_query_batched / " +
+      "llm_query_chunked / rlm_query — stdout returned to you is hard-capped at 4K chars, so " +
+      "printing file bodies is useless. Variables, imports, and state persist across calls. " +
+      "Also supports todo and ask_user_question inside the sandbox.",
     parameters: ReplToolParams,
 
     async execute(_toolCallId, rawParams, _execSignal, onUpdate, ctx) {
@@ -421,11 +464,13 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
 
         if (queuedId) emitter.emitSubcallUpdated({ id: queuedId, status: "done" });
 
-        const baseText = result.stdout || result.answerContent || "(no output)";
-        const surfacedEdits = surfaceReplEdits(result.edits, result.raised);
-        const editsBlock = surfacedEdits
-          ? `\n\nSTAGED_EDITS:\n${JSON.stringify(surfacedEdits)}`
-          : "";
+        const { text: resultText, surfacedEdits } = buildReplResultText(
+          result.stdout,
+          result.answerContent,
+          result.edits,
+          result.raised,
+          store.getSubcalls(),
+        );
 
         const details: ReplDetails = {
           status: "done",
@@ -438,7 +483,7 @@ export function createReplTool(deps: ReplToolDeps): ToolDefinition<typeof ReplTo
         };
         // Final progressive update
         onUpdate?.({ content: [{ type: "text", text: result.stdout.slice(0, 500) || "(no output)" }], details });
-        return { content: [{ type: "text", text: baseText + editsBlock }], details };
+        return { content: [{ type: "text", text: resultText }], details };
       } catch (e) {
         progressStatus = "error";
         const msg = errorMessage(e);

package/src/ui/config-panel.ts

@@ -15,6 +15,7 @@ const CHOICES = Object.freeze({
   maxTokens: Object.freeze(["none", "10000", "50000", "100000"]),
   maxErrors: Object.freeze(["3", "5", "10", "none"]),
   orchestrator: Object.freeze(["on", "off"]),
+  pipeline: Object.freeze(["on", "off"]),
   compaction: Object.freeze(["on", "off"]),
   rootSamplingMaxTokens: Object.freeze(["4096", "8192", "16384", "32768"]),
   sandboxInitTimeoutMs: Object.freeze(["10000", "30000", "60000", "120000"]),
@@ -38,6 +39,7 @@ export async function showConfigPanel(ctx: ExtensionContext, config: RlmConfig):
     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 advance_phase plus phase-stall reminders at root depth."),
     item("compaction", "Trajectory compaction", config.compaction ? "on" : "off", CHOICES.compaction, "Summarize old turns when history approaches the model 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."),
@@ -83,6 +85,7 @@ function applySetting(config: RlmConfig, id: string, value: string): void {
     case "maxTokens": config.maxTokens = value === "none" ? undefined : Number(value); break;
     case "maxErrors": config.maxErrors = value === "none" ? undefined : Number(value); break;
     case "orchestrator": config.orchestrator = value === "on"; break;
+    case "pipeline": config.pipeline = value === "on"; break;
     case "compaction": config.compaction = value === "on"; break;
     case "rootSamplingMaxTokens": config.rootSampling = Object.freeze({ ...config.rootSampling, maxTokens: Number(value) }); break;
     case "sandboxInitTimeoutMs": config.sandboxInitTimeoutMs = Number(value); break;