@hicaru/pi-rlm 0.1.0 → 0.1.1

package/README.md

@@ -155,16 +155,12 @@ These functions are injected into the model's Python namespace inside the REPL:
 > defaults (depth 4, conc 4) that's 4³ = 64 in the pathological case. Budget and error
 > caps (above) bound total spend regardless of fan-out.
 
-## Telemetry & run logs
+## Run logs
 
 - **Run logs** (`runLog`): always-on by default. Each run writes a JSONL trail to `.rlm/runs/`
   (default), capped at `maxRuns` (50). Supports **snapshots** (`sandbox.pkl`) and **resume**
   of interrupted runs via `/rlm-resume`. Snapshots are protected by a per-session `nonce`
   to prevent cross-session replay.
-- **MLflow tracing** (`telemetry`): optional. Set `MLFLOW_TRACKING_URI` or configure
-  `trackingUri` / `experimentId` in `/rlm-config`. The root run is tagged as an MLflow span
-  for trace correlation on resume. The Bearer token comes from the `MLFLOW_TRACKING_TOKEN`
-  env var and is **never persisted** to `rlm.json`.
 
 ## Security
 
@@ -191,11 +187,10 @@ src/
   core/       engine.ts (the loop) · iteration · limits · answer · compaction · pipeline · types
   prompts/    system + per-turn prompts (ported from the Python reference)
   text/       parsing (repl blocks) · tokens · preview · edits
-  state/      agent-tree · events · reads/writes · resume · paths · rows
+  state/      reads/writes · resume · paths · rows
   tool/       repl-tool · rlm-events · aggregator · propose-edits · emitter-listener
   config/     defaults · settings (rlm.json persistence + validation)
   context/    repomix-based repository packing + caching
-  telemetry/  MLflow sink · dispatcher · mlflow-config
   ui/         tree-widget · status · model-picker · config-panel · intro · theme
   commands/   rlm · rlm-config
   mode/       rlm-mode (controller) · input-router

package/README.ru.md

@@ -133,10 +133,9 @@ rm -rf ~/.pi/agent/extensions/rlm
 
 > **Примечание по параллелизму:** каждый дочерний `rlm_query` запускает собственного worker `python3` (~50–150 мс «холодного старта»). В худшем случае количество параллельных интерпретаторов ≈ `maxConcurrentSubcalls`^(depth−1); при настройках по умолчанию (глубина 4, параллелизм 4) это 4³ = 64 в патологическом случае. Лимиты бюджета и ошибок (см. выше) ограничивают общие затраты независимо от степени разветвления.
 
-## Телеметрия и логи запусков
+## Логи запусков
 
 - **Логи запусков** (`runLog`): включены по умолчанию. Каждый запуск записывает след в формате JSONL в `.rlm/runs/` (по умолчанию) с ограничением `maxRuns` (50). Поддерживает **снимки** (`sandbox.pkl`) и **возобновление** прерванных запусков через `/rlm-resume`. Снимки защищены сессионным `nonce` для предотвращения повторов между сессиями.
-- **Трассировка MLflow** (`telemetry`): опционально. Установите `MLFLOW_TRACKING_URI` или настройте `trackingUri` / `experimentId` в `/rlm-config`. Корневой запуск помечается как span MLflow для корреляции трасс при возобновлении. Bearer-токен берется из переменной окружения `MLFLOW_TRACKING_TOKEN` и **никогда не сохраняется** в `rlm.json`.
 
 ## Безопасность
 
@@ -155,11 +154,10 @@ src/
   core/       engine.ts (цикл) · iteration · limits · answer · compaction · pipeline · types
   prompts/    системные промпты и промпты для каждого шага (перенесены из Python-референса)
   text/       парсинг (repl-блоки) · токены · превью · правки
-  state/      дерево-агентов · события · чтения/записи · возобновление · пути · строки
+  state/      чтения/записи · возобновление · пути · строки
   tool/       repl-tool · rlm-events · агрегатор · предложение-правок · emitter-listener
   config/     значения по умолчанию · настройки (сохранение и валидация rlm.json)
   context/    упаковка репозитория на базе repomix + кеширование
-  telemetry/  MLflow sink · диспетчер · mlflow-config
   ui/         виджет-дерева · статус · выбор-модели · панель-конфигурации · вступление · тема
   commands/   rlm · rlm-config
   mode/       rlm-mode (контроллер) · маршрутизатор-ввода

package/README.zh-CN.md

@@ -143,15 +143,11 @@ rm -rf ~/.pi/agent/extensions/rlm
 > 默认设置下 (深度 4, 并发 4)，极端情况下为 4³ = 64。预算和错误
 > 上限 (见上文) 无论扇出 (fan-out) 如何都会限制总支出。
 
-## 遥测与运行日志
+## 运行日志
 
 - **运行日志** (`runLog`)：默认始终开启。每次运行将 JSONL 轨迹写入 `.rlm/runs/`
   (默认)，上限为 `maxRuns` (50)。支持通过 `/rlm-resume` 进行**快照** (`sandbox.pkl`) 和**恢复**
   被中断的任务。快照受每个会话的 `nonce` 保护，以防止跨会话重放。
-- **MLflow 追踪** (`telemetry`)：可选。设置 `MLFLOW_TRACKING_URI` 或在
-  `/rlm-config` 中配置 `trackingUri` / `experimentId`。根运行被标记为 MLflow span
-  以便在恢复时进行追踪关联。Bearer 令牌来自 `MLFLOW_TRACKING_TOKEN`
-  环境变量，且**绝不会**持久化到 `rlm.json`。
 
 ## 安全性
 
@@ -178,11 +174,10 @@ src/
   core/       engine.ts (the loop) · iteration · limits · answer · compaction · pipeline · types
   prompts/    system + per-turn prompts (ported from the Python reference)
   text/       parsing (repl blocks) · tokens · preview · edits
-  state/      agent-tree · events · reads/writes · resume · paths · rows
+  state/      reads/writes · resume · paths · rows
   tool/       repl-tool · rlm-events · aggregator · propose-edits · emitter-listener
   config/     defaults · settings (rlm.json persistence + validation)
   context/    repomix-based repository packing + caching
-  telemetry/  MLflow sink · dispatcher · mlflow-config
   ui/         tree-widget · status · model-picker · config-panel · intro · theme
   commands/   rlm · rlm-config
   mode/       rlm-mode (controller) · input-router

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hicaru/pi-rlm",
-  "version": "0.1.0",
+  "version": "0.1.1",
   "type": "module",
   "description": "Recursive Language Model (RLM) for the Pi coding agent — native, server-less.",
   "license": "MIT",
@@ -13,17 +13,17 @@
   "bugs": {
     "url": "https://github.com/hicaru/rlm.pi/issues"
   },
-  "exports": {
-    ".": "./src/index.ts"
-  },
   "files": [
-    "src",
+    "src/",
     "README.md",
     "LICENSE"
   ],
   "keywords": [
     "pi-package",
-    "pi-extension"
+    "pi-extension",
+    "rlm",
+    "recursive",
+    "ai-agent"
   ],
   "scripts": {
     "check": "tsc --noEmit",
@@ -38,14 +38,12 @@
     "access": "public"
   },
   "peerDependencies": {
-    "@earendil-works/pi-ai": ">=0.79.0",
-    "@earendil-works/pi-coding-agent": ">=0.79.0",
-    "@earendil-works/pi-tui": ">=0.79.0",
+    "@earendil-works/pi-ai": "*",
+    "@earendil-works/pi-coding-agent": "*",
+    "@earendil-works/pi-tui": "*",
     "typebox": "*"
   },
   "dependencies": {
-    "@mlflow/core": "~0.2.0",
-    "diff": "^9.0.0",
     "repomix": "^1.15.0"
   },
   "devDependencies": {

package/src/bridge/pi-interactive.ts

@@ -4,29 +4,6 @@ import type { AskAnswer, AskQuestion } from "../sandbox/protocol.ts";
 import { formatError } from "../util/errors.ts";
 import { createTodoFallback } from "./fallback-todo.ts";
 
-interface ToolInvoker {
-  readonly callTool?: (name: string, params: unknown) => Promise<unknown>;
-}
-
-function hasAnswers(value: unknown): value is { readonly answers: readonly unknown[] } {
-  return typeof value === "object" && value !== null && Array.isArray((value as { readonly answers?: unknown }).answers);
-}
-
-function isAskAnswer(value: unknown): value is AskAnswer {
-  if (typeof value !== "object" || value === null) return false;
-  const candidate = value as { readonly question?: unknown; readonly selected?: unknown; readonly custom?: unknown };
-  return typeof candidate.question === "string"
-    && Array.isArray(candidate.selected)
-    && candidate.selected.every((item) => typeof item === "string")
-    && (candidate.custom === undefined || typeof candidate.custom === "string");
-}
-
-function normalizeAnswers(result: unknown): AskAnswer[] | undefined {
-  if (!hasAnswers(result)) return undefined;
-  const answers = result.answers;
-  return answers.every(isAskAnswer) ? Array.from(answers) : undefined;
-}
-
 async function askViaUi(ctx: ExtensionContext, questions: readonly AskQuestion[]): Promise<AskAnswer[]> {
   if (!ctx.hasUI) throw new Error("ask_user_question requires UI");
   const answers = new Array<AskAnswer>(questions.length);
@@ -57,30 +34,8 @@ async function askViaUi(ctx: ExtensionContext, questions: readonly AskQuestion[]
 export function createPiInteractiveDeps(ctx: ExtensionContext): InteractiveDeps {
   const fallbackTodo = createTodoFallback();
   return Object.freeze({
-    onAskUserQuestion: async (questions: readonly AskQuestion[]): Promise<AskAnswer[]> => {
-      const callTool = (ctx as unknown as ToolInvoker).callTool;
-      if (typeof callTool === "function") {
-        try {
-          const result = await callTool.call(ctx, "ask_user_question", { questions });
-          const answers = normalizeAnswers(result);
-          if (answers) return answers;
-        } catch {
-          // Fall through to native UI fallback when the extension tool is not registered or fails.
-        }
-      }
-      return askViaUi(ctx, questions);
-    },
-    onTodo: async (action: string, params: Record<string, unknown>): Promise<string> => {
-      const callTool = (ctx as unknown as ToolInvoker).callTool;
-      if (typeof callTool === "function") {
-        try {
-          const result = await callTool.call(ctx, "todo", { action, ...params });
-          return typeof result === "string" ? result : JSON.stringify(result);
-        } catch {
-          // Fall through to in-process task store when the extension tool is not registered or fails.
-        }
-      }
-      return fallbackTodo(action, params);
-    },
+    onAskUserQuestion: (questions: readonly AskQuestion[]): Promise<AskAnswer[]> => askViaUi(ctx, questions),
+    onTodo: (action: string, params: Record<string, unknown>): Promise<string> =>
+      Promise.resolve(fallbackTodo(action, params)),
   });
 }

package/src/commands/rlm-config.ts

@@ -11,15 +11,22 @@ export async function runRlmConfig(controller: RlmController, ctx: ExtensionCont
   const models = ctx.modelRegistry.getAvailable();
 
   const worker = await selectModel(ctx, "Worker model (sub-LLM / llm_query)", models, controller.workerModel, controller.config.subSampling.reasoning);
-  if (worker) {
+  if (worker === null) {
+    controller.workerModel = undefined;
+    controller.config.subSampling.reasoning = undefined;
+  } else if (worker) {
     controller.workerModel = worker.model;
     controller.config.subSampling.reasoning = worker.thinkingLevel;
   }
 
   await showConfigPanel(ctx, controller.config);
 
-  const effectiveWorker = controller.workerModel ?? cheapestModel(ctx.modelRegistry);
-  controller.savedWorkerRef = modelRef(controller.workerModel) ?? modelRef(effectiveWorker);
+  if (worker === null) {
+    controller.savedWorkerRef = undefined;
+  } else {
+    const effectiveWorker = controller.workerModel ?? cheapestModel(ctx.modelRegistry);
+    controller.savedWorkerRef = modelRef(controller.workerModel) ?? modelRef(effectiveWorker);
+  }
   const persisted = await controller.persist();
   if (!persisted) ctx.ui.notify("RLM: failed to save settings to ~/.pi/agent/rlm.json", "error");
   setRlmModeStatus(ctx.ui, controller);

package/src/commands/rlm.ts

@@ -13,9 +13,6 @@ import type { RunHeader } from "../state/rows.ts";
 import { buildRlmSystemPrompt } from "../prompts/system.ts";
 import { RlmEmitter } from "../tool/rlm-events.ts";
 import { RlmEventAggregator } from "../tool/rlm-aggregator.ts";
-import { createTelemetrySink } from "../telemetry/index.ts";
-import { applyEdits } from "../patch/index.ts";
-import { tryExtractDiff } from "../core/answer.ts";
 
 export function registerRlmCommand(pi: ExtensionAPI, controller: RlmController): void {
   pi.registerCommand("rlm", {
@@ -112,14 +109,10 @@ async function executeRlmRunWithResume(
   context: unknown,
 ): Promise<void> {
   let handle: RunHandle | undefined;
-  let sink: Awaited<ReturnType<typeof createTelemetrySink>>;
   let emitter: RlmEmitter | undefined;
-  let detachSink: (() => void) | undefined;
   let aggregator: RlmEventAggregator | undefined;
   try {
-    sink = await createTelemetrySink(controller.config.telemetry);
     emitter = new RlmEmitter();
-    if (sink) detachSink = emitter.attachSink(sink);
     aggregator = new RlmEventAggregator(emitter, (partial) => {
       const d = partial.details;
       if (!d) return;
@@ -146,20 +139,12 @@ async function executeRlmRunWithResume(
   try {
     const result = await done;
     pi.sendMessage({ customType: "rlm-answer", content: result.answer, display: true });
-    const proposedDiffs = result.diffs?.length ? result.diffs : tryExtractDiff(result.answer);
-    await applyEdits(
-      result.edits ?? [],
-      proposedDiffs,
-      ctx,
-    );
   } catch (e) {
     ctx.ui.notify(`RLM resume failed: ${e instanceof Error ? e.message : String(e)}`, "error");
   } finally {
     clearRlmStatus(ctx.ui);
     ctx.ui.setWidget?.("rlm-status", undefined);
-    detachSink?.();
     aggregator?.dispose();
     emitter?.shutdown();
-    try { await sink?.shutdown(); } catch { /* best-effort */ }
   }
 }

package/src/config/settings.ts

@@ -4,7 +4,7 @@ import { mkdir, readFile, writeFile } from "node:fs/promises";
 import { dirname, join } from "node:path";
 import { getAgentDir, type ModelRegistry } from "@earendil-works/pi-coding-agent";
 import type { Api, Model, ThinkingLevel } from "@earendil-works/pi-ai";
-import type { RlmConfig, RunLogConfig, TelemetryConfig } from "../core/types.ts";
+import type { RlmConfig, RunLogConfig } from "../core/types.ts";
 import { DEFAULT_CONFIG } from "./defaults.ts";
 
 export interface PersistedSettings {
@@ -30,26 +30,6 @@ function validateString(v: unknown): string | undefined {
   return typeof v === "string" && v.trim() ? v : undefined;
 }
 
-function validateTelemetry(raw: unknown): TelemetryConfig | undefined {
-  if (typeof raw !== "object" || raw === null) return undefined;
-  const r = raw as Record<string, unknown>;
-  const out: {
-    enabled?: boolean;
-    trackingUri?: string;
-    experimentId?: string;
-    maxQueueSize?: number;
-  } = {};
-  const enabled = validateBoolean(r.enabled);
-  if (enabled !== undefined) out.enabled = enabled;
-  const trackingUri = validateString(r.trackingUri);
-  if (trackingUri !== undefined) out.trackingUri = trackingUri;
-  const experimentId = validateString(r.experimentId);
-  if (experimentId !== undefined) out.experimentId = experimentId;
-  const maxQueueSize = validateNumber(r.maxQueueSize, 1);
-  if (maxQueueSize !== undefined) out.maxQueueSize = maxQueueSize;
-  return Object.freeze(out);
-}
-
 function validateRunLog(raw: unknown): Partial<RunLogConfig> | undefined {
   if (typeof raw !== "object" || raw === null) return undefined;
   const r = raw as Record<string, unknown>;
@@ -102,8 +82,6 @@ function validateConfig(raw: unknown): Partial<RlmConfig> {
   if (typeof r.smartReasoning === "string") out.smartReasoning = r.smartReasoning as ThinkingLevel;
   const subSystemPrompt = validateString(r.subSystemPrompt);
   if (subSystemPrompt !== undefined) out.subSystemPrompt = subSystemPrompt;
-  const telemetry = validateTelemetry(r.telemetry);
-  if (telemetry) out.telemetry = telemetry;
   const runLog = validateRunLog(r.runLog);
   if (runLog) out.runLog = runLog;
   const sandboxInitTimeoutMs = validateNumber(r.sandboxInitTimeoutMs, 100);
@@ -167,7 +145,6 @@ export function mergeConfig(partial: Partial<RlmConfig>): RlmConfig {
     ...partial,
     subSampling: { ...DEFAULT_CONFIG.subSampling, ...partial.subSampling },
     rootSampling: Object.freeze({ ...DEFAULT_CONFIG.rootSampling, ...partial.rootSampling }),
-    ...(partial.telemetry ? { telemetry: Object.freeze({ ...partial.telemetry }) } : {}),
     ...(partial.runLog ? { runLog: Object.freeze({ ...DEFAULT_CONFIG.runLog, ...partial.runLog }) } : {}),
   };
 }

package/src/core/answer.ts

@@ -1,6 +1,6 @@
 /** Helpers for detecting and formatting the RLM final answer from a turn's REPL results. */
 
-import type { ProposedDiffEdit, ProposedEdit, ReplResult } from "../sandbox/protocol.ts";
+import type { ProposedEdit, ReplResult } from "../sandbox/protocol.ts";
 import { truncateOutput } from "../text/parsing.ts";
 
 /** First non-null final answer across a turn's executed blocks, or null. */
@@ -27,22 +27,6 @@ export function collectEdits(results: readonly ReplResult[]): ProposedEdit[] {
   return [];
 }
 
-/** Last cumulative diff proposal set reported by a turn. */
-export function collectDiffs(results: readonly ReplResult[]): ProposedDiffEdit[] {
-  for (let i = results.length - 1; i >= 0; i--) {
-    const diffs = results[i]?.diffs;
-    if (diffs && diffs.length > 0) return [...diffs];
-  }
-  return [];
-}
-
-/** Parses a ```diff fence from answer text; returns [] if none found. */
-export function tryExtractDiff(answer: string): ProposedDiffEdit[] {
-  const match = /```diff\n([\s\S]*?)```/.exec(answer);
-  if (!match) return [];
-  return [{ diff: match[1].trim() }];
-}
-
 /** True if any block in the turn raised an exception. Plain stderr does not count. */
 export function turnHadError(results: readonly ReplResult[]): boolean {
   return results.some((r) => r.raised);

package/src/core/engine.ts

@@ -21,7 +21,7 @@ import { PythonSandbox } from "../sandbox/sandbox.ts";
 import { advancePhase as validatePhaseTransition, phaseGatePrompt, type PhaseState } from "./pipeline.ts";
 import { previewStdout, previewText } from "../text/preview.ts";
 import { contextLength, contextSizeStats, contextTypeLabel } from "../text/tokens.ts";
-import { collectDiffs, collectEdits, finalAnswerOf, formatReplOutputs, latestAnswerContentOf, turnHadError } from "./answer.ts";
+import { collectEdits, finalAnswerOf, formatReplOutputs, latestAnswerContentOf, turnHadError } from "./answer.ts";
 import { compactHistory, shouldCompact } from "./compaction.ts";
 import { appendUserMessage } from "./history.ts";
 import { runTurn } from "./iteration.ts";
@@ -32,7 +32,7 @@ import { appendRow, appendTodoRow, generateRunId, pruneRuns, snapshotPath, write
 import { STATE_SCHEMA_VERSION } from "../state/rows.ts";
 import type { PhaseRow, RunHeader } from "../state/rows.ts";
 import { serializeForSandbox, type ContextBundle } from "../context/repomix-context.ts";
-import type { ProposedDiffEdit, ProposedEdit } from "../sandbox/protocol.ts";
+import type { ProposedEdit } from "../sandbox/protocol.ts";
 import { formatError } from "../util/errors.ts";
 
 
@@ -57,7 +57,7 @@ export function createEngine(deps: EngineDeps): RunRlm {
   const run: RunRlm = async (input: RlmInput): Promise<RlmResult> => {
     const nowIso = (): string => new Date().toISOString(); // local helper — 4 call sites below
     const persist = input.depth === 0 && deps.runState !== undefined;
-    // Compute runId early so it can tag the MLflow root span (Ops: trace correlation on resume).
+    // Compute runId early for run-state correlation on resume.
     const runId = persist
       ? (input.resume ? input.resume.header.runId : generateRunId())
       : undefined;
@@ -136,7 +136,6 @@ export function createEngine(deps: EngineDeps): RunRlm {
     let compactions = 0;
     let completedTurns = 0;
     let editsAcc: ProposedEdit[] = [];
-    let diffsAcc: ProposedDiffEdit[] = [];
     let phaseState: PhaseState | undefined;
     let nodeStatus: "done" | "error" = "done";
     let persistOn = persist;
@@ -228,7 +227,6 @@ export function createEngine(deps: EngineDeps): RunRlm {
         limits.addRaw(input.resume.usageSeed.costUsd, input.resume.usageSeed.inputTokens, input.resume.usageSeed.outputTokens);
         best = input.resume.best;
         editsAcc = [];
-        diffsAcc = [];
         compactions = input.resume.compactions;
         completedTurns = input.resume.completedTurns;
         if (input.resume.phase) {
@@ -312,12 +310,9 @@ export function createEngine(deps: EngineDeps): RunRlm {
         completedTurns = i + 1;
         const proposedEdits = collectEdits(turn.results);
         if (proposedEdits.length > 0) editsAcc = proposedEdits;
-        const proposedDiffs = collectDiffs(turn.results);
-        if (proposedDiffs.length > 0) diffsAcc = proposedDiffs;
-
         const final = finalAnswerOf(turn.results);
         if (final != null) {
-          const done = result(final, i + 1, limits, editsAcc, diffsAcc);
+          const done = result(final, i + 1, limits, editsAcc);
           await recordTerminal("completed", done);
           lastAnswer = done.answer;
           return done;
@@ -348,21 +343,21 @@ export function createEngine(deps: EngineDeps): RunRlm {
         }
       }
       if (pendingReplOutputs) appendUserMessage(history, pendingReplOutputs);
-      const finalized = result(await finalize(history, deps, limits), deps.config.maxIterations, limits, editsAcc, diffsAcc);
+      const finalized = result(await finalize(history, deps, limits), deps.config.maxIterations, limits, editsAcc);
       await recordTerminal("finalized", finalized);
       lastAnswer = finalized.answer;
       return finalized;
     } catch (err) {
       // Abort is a user action — resolve with the best partial, not an error.
       if (deps.signal?.aborted) {
-        const aborted = result(best.trim() || "(aborted)", completedTurns, limits, editsAcc, diffsAcc);
+        const aborted = result(best.trim() || "(aborted)", completedTurns, limits, editsAcc);
         await recordTerminal("aborted", aborted);
         lastAnswer = aborted.answer;
         return aborted;
       }
       if (err instanceof LimitError) {
         nodeStatus = "error";
-        const stopped = result(best.trim() || `(stopped: ${err.message})`, completedTurns, limits, editsAcc, diffsAcc);
+        const stopped = result(best.trim() || `(stopped: ${err.message})`, completedTurns, limits, editsAcc);
         await recordTerminal("stopped", stopped);
         lastAnswer = stopped.answer;
         return stopped;
@@ -388,9 +383,9 @@ export function createEngine(deps: EngineDeps): RunRlm {
   return run;
 }
 
-function result(answer: string, iterations: number, limits: LimitGuard, edits: ProposedEdit[] = [], diffs: ProposedDiffEdit[] = []): RlmResult {
+function result(answer: string, iterations: number, limits: LimitGuard, edits: ProposedEdit[] = []): RlmResult {
   const u = limits.usage();
-  return { answer, edits, diffs, iterations, costUsd: u.costUsd, inputTokens: u.inputTokens, outputTokens: u.outputTokens, durationMs: u.durationMs };
+  return { answer, edits, iterations, costUsd: u.costUsd, inputTokens: u.inputTokens, outputTokens: u.outputTokens, durationMs: u.durationMs };
 }
 
 /** Out of turns: ask the model for its best final answer (plain text). */

package/src/core/types.ts

@@ -1,7 +1,7 @@
 /** Shared configuration + runtime types for the RLM engine. */
 
 import type { ThinkingLevel } from "@earendil-works/pi-ai";
-import type { AskAnswer, AskQuestion, ProposedDiffEdit, ProposedEdit } from "../sandbox/protocol.ts";
+import type { AskAnswer, AskQuestion, ProposedEdit } from "../sandbox/protocol.ts";
 import type { ReconstructResult } from "../state/resume.ts";
 
 export interface Sampling {
@@ -12,15 +12,6 @@ export interface Sampling {
 
 type MutableSampling = { -readonly [Key in keyof Sampling]?: Sampling[Key] };
 
-export interface TelemetryConfig {
-  /** Default: enabled iff a tracking URI resolves from config or MLFLOW_TRACKING_URI. */
-  readonly enabled?: boolean;
-  readonly trackingUri?: string;
-  readonly experimentId?: string;
-  /** Bearer token is env-only via MLFLOW_TRACKING_TOKEN; never persisted in rlm.json. */
-  readonly maxQueueSize?: number;
-}
-
 export interface RunLogConfig {
   /** Default: true — always-on, opt-out. */
   readonly enabled?: boolean;
@@ -81,8 +72,6 @@ export interface RlmConfig {
   subSystemPrompt?: string;
   /** Sampling for sub-LLM (worker) calls. */
   subSampling: MutableSampling;
-  /** Optional MLflow telemetry export configuration. Omitted by default. */
-  readonly telemetry?: TelemetryConfig;
   /** Optional run-state persistence configuration. Enabled by default. */
   readonly runLog?: RunLogConfig;
 }
@@ -112,7 +101,6 @@ export interface RlmResult {
   readonly answer: string;
   /** Legacy anchor edits retained for compatibility while older run-state rows exist. */
   readonly edits?: readonly ProposedEdit[];
-  readonly diffs?: readonly ProposedDiffEdit[];
   readonly iterations: number;
   readonly costUsd: number;
   readonly inputTokens: number;

package/src/index.ts

@@ -6,7 +6,6 @@ import { Markdown } from "@earendil-works/pi-tui";
 import { registerRlmCommand } from "./commands/rlm.ts";
 import { registerRlmConfigCommand } from "./commands/rlm-config.ts";
 import { createRlmTool } from "./tool/rlm-tool.ts";
-import { createApplyDiffTool } from "./tool/apply-diff-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";
@@ -17,21 +16,12 @@ import { packRepository, formatForLLM, serializeForSandbox } from "./context/rep
 import { buildNativeSystemPrompt } from "./prompts/system.ts";
 import { errorMessage } from "./util/errors.ts";
 
-const BLOCKED_NATIVE_TOOLS = Object.freeze(new Set(["read", "grep", "bash", "write", "edit"]));
+const BLOCKED_NATIVE_TOOLS = Object.freeze(new Set(["read", "grep", "bash"]));
 
 export default function rlmExtension(pi: ExtensionAPI): void {
-  void setupRlmExtension(pi).catch((error) => {
-    console.warn(`[rlm] extension setup failed: ${errorMessage(error)}`);
-  });
-}
-
-async function setupRlmExtension(pi: ExtensionAPI): Promise<void> {
-  const persisted = await loadSettings();
-  const config = mergeConfig(persisted.config);
+  // Init synchronously with defaults — ensures commands/tools/handlers register before session_start
+  const config = mergeConfig({});
   const controller = new RlmController(config);
-  controller.savedWorkerRef = persisted.worker;
-
-  // ── SandboxManager — persistent singleton for native-mode repl() ──
   const sandboxManager = new SandboxManager({
     execTimeoutS: config.execTimeoutS,
     requestTimeoutMs: config.requestTimeoutMs,
@@ -39,6 +29,16 @@ async function setupRlmExtension(pi: ExtensionAPI): Promise<void> {
     sandboxInitTimeoutMs: config.sandboxInitTimeoutMs,
   });
 
+  // Load persisted settings async — applied before session_start handler reads controller state
+  const settingsReady = loadSettings()
+    .then((persisted) => {
+      controller.config = mergeConfig(persisted.config);
+      controller.savedWorkerRef = persisted.worker;
+    })
+    .catch((err) => {
+      console.warn(`[rlm] settings load failed: ${errorMessage(err)}`);
+    });
+
   // ── Message renderers ──
   pi.registerMessageRenderer(
     "rlm-answer",
@@ -56,23 +56,19 @@ async function setupRlmExtension(pi: ExtensionAPI): Promise<void> {
   registerRlmConfigCommand(pi, controller);
 
   // ── Tool registration ──
-  // Existing rlm tool (stays for backward compat with /rlm mode)
   pi.registerTool(createRlmTool(controller));
-  pi.registerTool(createApplyDiffTool());
-
-  // Native repl tool — re-registered each session to pick up model provider changes
   let guidePosted = false;
 
   pi.on("session_start", async (_event, ctx) => {
-    // Restore saved worker ref — retry lazily if registry/auth is not ready yet.
-    // session_start can fire before the provider/auth registry is fully loaded, so a miss
-    // here does NOT mean the ref is stale: RlmController.resolveModels() retries later.
+    // Wait for persisted settings before reading controller state
+    await settingsReady;
+
     if (controller.savedWorkerRef) {
       const resolved = resolveModelId(ctx.modelRegistry, controller.savedWorkerRef);
       if (resolved) controller.workerModel = resolved;
     }
 
-    // Register repl tool with current models (re-registers each session for provider changes)
+    // Re-register repl tool each session to pick up model provider changes
     const workerModel = controller.workerModel ?? cheapestModel(ctx.modelRegistry) ?? ctx.model;
     const model = ctx.model;
     if (workerModel && model) {
@@ -84,7 +80,7 @@ async function setupRlmExtension(pi: ExtensionAPI): Promise<void> {
           getModel: () => controller.resolveModels(ctx)?.model,
           getWorkerModel: () => controller.resolveModels(ctx)?.worker,
           registry: ctx.modelRegistry,
-          config,
+          config: controller.config,
         }));
       } catch { /* re-registration on provider change — ignore if already registered */ }
     }
@@ -139,19 +135,20 @@ async function setupRlmExtension(pi: ExtensionAPI): Promise<void> {
   });
 
   // ── Input routing: native mode — agent decides whether to use repl() or other tools ──
-  // The old black-box rlm() routing is removed; the main agent receives messages normally
-  // and chooses natively when to call repl(), read, grep, zebra-mcp, etc.
   pi.on("input", async (_event, _ctx) => {
     return { action: "continue" };
   });
 
-  // ── Tool restriction: block read/grep/bash when RLM is ON ──
+  // ── Tool restriction: block analysis tools when RLM is ON ──
+  // `edit`/`write` stay unblocked so the agent modifies files through Pi's native
+  // tool flow (visible to all plugins, +/- diff preview). Only read/grep/bash are
+  // blocked — the repository is pre-loaded in the REPL `context` variable.
   pi.on("tool_call", async (event) => {
     if (!controller.enabled) return;
     if (BLOCKED_NATIVE_TOOLS.has(event.toolName)) {
       return {
         block: true,
-        reason: "RLM mode active. Use repl({code}) to read files and apply_diff({diff}) to modify them. All files are pre-loaded in the REPL. If sub-LLM credits are exhausted, report to the user.",
+        reason: "RLM mode active. Use repl({code}) to read files and search the repository — all files are pre-loaded in the REPL `context` variable. Use `edit`/`write` for file changes. If sub-LLM credits are exhausted, report to the user.",
       };
     }
   });