@hicaru/pi-rlm 0.3.20 → 0.3.22

package/README.md

@@ -5,7 +5,7 @@
 <p align="center">
   <a href="https://www.npmjs.com/package/@hicaru/pi-rlm"><img src="https://img.shields.io/npm/v/@hicaru/pi-rlm?color=cb3837&logo=npm" alt="npm version"></a>
   <a href="https://github.com/openzebra/rlm.pi/blob/master/pi-plugin/rlm/LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT license"></a>
-  <a href="https://github.com/earendil-works/pi"><img src="https://img.shields.io/badge/for-Pi-7c3aed" alt="Built for Pi"></a>
+  <a href="https://github.com/earendil-works/pi"><img src="https://img.shields.io/badge/for-Pi%20%C2%B7%20omp-7c3aed" alt="Built for Pi and oh-my-pi"></a>
 </p>
 
 <p align="center">
@@ -17,17 +17,17 @@
 ## Install
 
 ```bash
-pi install npm:@hicaru/pi-rlm
+pi install npm:@hicaru/pi-rlm        # Pi
+omp plugin install @hicaru/pi-rlm    # oh-my-pi
 ```
 
-Run `/reload` in Pi — `/rlm`, `/rlm-config`, `/rlm-stop` appear under **[Extensions]**.
-Toggle with `Ctrl+Shift+R` or `/rlm`.
+Then `/reload` (or restart) — `/rlm`, `/rlm-config`, `/rlm-stop` appear under
+**[Extensions]**. Toggle with `Ctrl+Shift+R` or `/rlm`.
 
-To remove it later:
-
-```bash
-pi uninstall npm:@hicaru/pi-rlm
-```
+|              | install                             | upgrade                                     | remove                                |
+| ------------ | ----------------------------------- | ------------------------------------------- | ------------------------------------- |
+| **Pi**       | `pi install npm:@hicaru/pi-rlm`     | `pi install npm:@hicaru/pi-rlm --force`     | `pi uninstall npm:@hicaru/pi-rlm`     |
+| **oh-my-pi** | `omp plugin install @hicaru/pi-rlm` | `omp plugin install @hicaru/pi-rlm --force` | `omp plugin uninstall @hicaru/pi-rlm` |
 
 <p align="center">
   <img src="https://github.com/openzebra/rlm.pi/blob/master/assets/hero.png?raw=true" width="100%" alt="rlm.pi — OOLONG benchmark results">
@@ -46,14 +46,17 @@ models, recursively. Same Pi session, same tools, same keys: `/rlm` and go. Read
 **OOLONG (oolong-synth)** — paper-tier long-context suite (the only suite); latest
 journal per model, cost per task from real `costUsd`:
 
-| Model | Score | Avg. cost/task |
-|-------|-------|----------------|
-| `zai/glm-4.7` | **83%** | $0.0000 * |
-| `qwen/qwen3.8-27b` | 49% | $0.1038 |
-| `inception/mercury-2.5` | 38% | $0.0052 |
+| Model                              | Score             | Avg. tokens/task | Avg. cost/task |
+| ---------------------------------- | ----------------- | ---------------- | -------------- |
+| `zai/glm-4.7`                      | **91.7%** (22/24) | ~36k             | $0.0000 *      |
+| `openrouter/qwen/qwen3.8-27b`      | 49%               | ~36k             | $0.1038        |
+| `openrouter/inception/mercury-2.5` | 38%               | ~36k             | $0.0052        |
 
 \* glm-4.7 runs on Z.ai's coding-plan endpoint — subscription billing, `costUsd` stays $0.
 
+In the [RLM paper](https://arxiv.org/abs/2512.24601), GPT-5-mini driven as an RLM outperforms
+GPT-o3 on OOLONG — recursion beats raw context, at a fraction of the price.
+
 Raw per-task rows (correct, recall, latency, tokens, cost) live in
 `bench/runs/*.jsonl` — one JSONL row per task, committed as history.
 
@@ -99,17 +102,17 @@ One suite (`oolong`). Regenerate the hero chart:
 
 Drop ANY of these into `add_context()` — they auto-convert to Markdown and land in `context`:
 
-| Category | Formats |
-|----------|---------|
-| **Word** | `.docx` |
-| **PDF** | `.pdf` |
-| **PowerPoint** | `.pptx` |
-| **Excel** | `.xlsx` |
-| **EPUB** | `.epub` |
-| **Rich Text** | `.rtf` |
-| **OpenDocument** | `.odt` |
-| **CSV / TSV** | `.csv` `.tsv` |
-| **HTML / XML** | `.html` `.htm` `.xml` `.rss` `.atom` |
+| Category              | Formats                                                                    |
+| --------------------- | -------------------------------------------------------------------------- |
+| **Word**              | `.docx`                                                                    |
+| **PDF**               | `.pdf`                                                                     |
+| **PowerPoint**        | `.pptx`                                                                    |
+| **Excel**             | `.xlsx`                                                                    |
+| **EPUB**              | `.epub`                                                                    |
+| **Rich Text**         | `.rtf`                                                                     |
+| **OpenDocument**      | `.odt`                                                                     |
+| **CSV / TSV**         | `.csv` `.tsv`                                                              |
+| **HTML / XML**        | `.html` `.htm` `.xml` `.rss` `.atom`                                       |
 | **+ Pandoc fallback** | `.doc` `.ppt` `.xls` `.pptm` `.xlsm` `.xlsb` `.ppsm` `.docm` `.odp` `.ods` |
 
 ```python
@@ -126,24 +129,33 @@ A **Recursive Language Model (RLM)** replaces `llm.completion(prompt)` with
 launch sub-LLM and sub-RLM calls as ordinary Python functions — decomposing,
 delegating, and synthesizing across a tree of models, not a single context window.
 
+Each step of a run is one validated state transition (paper §3.2, §5.7):
+
+```
+A_t = (P, Σ_t, O_t)          fixed prompt + execution state Σ_t + tool surface
+ΔΣ_t = μ(A_t)                the model emits a repl() patch, not prose
+V(ΔΣ_t, Σ_t)                 deterministic validator — no crash paths
+Σ_{t+1} = Σ_t ⊕ ΔΣ_t         deep merge, null = delete
+```
+
 ## Commands
 
-| Command | Shortcut | What it does |
-|---------|----------|--------------|
-| `/rlm` | `Ctrl+Shift+R` | Toggle RLM mode on/off |
-| `/rlm-stop` | | Abort current run |
-| `/rlm-config` | | Pick models, tune limits |
+| Command       | Shortcut       | What it does             |
+| ------------- | -------------- | ------------------------ |
+| `/rlm`        | `Ctrl+Shift+R` | Toggle RLM mode on/off   |
+| `/rlm-stop`   |                | Abort current run        |
+| `/rlm-config` |                | Pick models, tune limits |
 
 ## Settings (`/rlm-config`)
 
-| Setting | Default | Why you'd change it |
-|---------|---------|---------------------|
-| Smart model | Pi's active | Use your best model as orchestrator |
-| Worker model | cheapest available | Free/cheap model for leaf `llm_query` calls |
-| Max recursion depth | `4` | Deeper trees for harder problems |
-| Max iterations | `30` | Longer runs for complex tasks |
-| REPL timeout | `120`s | Bump for slow computations |
-| Max concurrent subs | `16` | More parallelism (costs RAM) |
+| Setting             | Default            | Why you'd change it                         |
+| ------------------- | ------------------ | ------------------------------------------- |
+| Smart model         | Pi's active        | Use your best model as orchestrator         |
+| Worker model        | cheapest available | Free/cheap model for leaf `llm_query` calls |
+| Max recursion depth | `4`                | Deeper trees for harder problems            |
+| Max iterations      | `30`               | Longer runs for complex tasks               |
+| REPL timeout        | `120`s             | Bump for slow computations                  |
+| Max concurrent subs | `16`               | More parallelism (costs RAM)                |
 
 ## Sampling & reproducibility
 
@@ -152,14 +164,14 @@ temperature 0 took OOLONG from 71% pooled / 40% flips to 91.7–100% all-stable
 total. Those knobs are first-class in `rlm.json` (`~/.pi/agent/rlm.json`) and on the
 `/rlm-config` panel:
 
-| Field | Where | Default | What it governs |
-|-------|-------|---------|-----------------|
-| `rootSampling.maxTokens` | rlm.json, panel | `16384` | Output cap per root-model turn (finalize included) |
-| `rootSampling.temperature` | rlm.json, panel | provider default | Root + finalize sampling temperature; `0` = deterministic |
-| `smartReasoning` | rlm.json, panel | none | Thinking effort for the root model |
-| `subSampling.maxTokens` | rlm.json, panel | `8192` | Output cap per leaf sub-call (`llm_query`, `llm_batch`, `map_files`) |
-| `subSampling.temperature` | rlm.json, panel | provider default | Leaf sampling temperature |
-| `enableVerificationNudge` | rlm.json | off | One coached redo when the root finalizes early with a bare number / short label |
+| Field                      | Where           | Default          | What it governs                                                                 |
+| -------------------------- | --------------- | ---------------- | ------------------------------------------------------------------------------- |
+| `rootSampling.maxTokens`   | rlm.json, panel | `16384`          | Output cap per root-model turn (finalize included)                              |
+| `rootSampling.temperature` | rlm.json, panel | provider default | Root + finalize sampling temperature; `0` = deterministic                       |
+| `smartReasoning`           | rlm.json, panel | none             | Thinking effort for the root model                                              |
+| `subSampling.maxTokens`    | rlm.json, panel | `8192`           | Output cap per leaf sub-call (`llm_query`, `llm_batch`, `map_files`)            |
+| `subSampling.temperature`  | rlm.json, panel | provider default | Leaf sampling temperature                                                       |
+| `enableVerificationNudge`  | rlm.json        | off              | One coached redo when the root finalizes early with a bare number / short label |
 
 **Reproducibility recipe (validated by r3):**
 

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@hicaru/pi-rlm",
-	"version": "0.3.20",
+	"version": "0.3.22",
 	"author": "hicaru",
 	"repository": {
 		"type": "git",

package/src/commands/rlm.ts

@@ -4,7 +4,12 @@ import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import type { RlmController } from "../mode/rlm-mode.ts";
 import { setRlmModeStatus } from "../ui/status.ts";
 
-export function registerRlmCommand(pi: ExtensionAPI, controller: RlmController): void {
+/**
+ * `stopNative` aborts the native-mode session work (repl cells, child engines, detached
+ * spawn() tasks) and reports whether any was in flight. Optional so tests can register the
+ * bare command; production wires the closure accessors from src/index.ts.
+ */
+export function registerRlmCommand(pi: ExtensionAPI, controller: RlmController, stopNative?: () => boolean): void {
   pi.registerCommand("rlm", {
     description: "Toggle persistent RLM mode (route plain prompts through the RLM engine).",
     handler: async (_args, ctx) => {
@@ -15,14 +20,16 @@ export function registerRlmCommand(pi: ExtensionAPI, controller: RlmController):
   });
 
   pi.registerCommand("rlm-stop", {
-    description: "Abort the in-progress RLM run.",
+    description: "Abort in-progress RLM work: RLM runs, native repl cells, background tasks.",
     handler: async (_args, ctx) => {
-      if (!controller.isBusy()) {
-        ctx.ui.notify("No RLM run in progress.", "info");
-        return;
+      const rlmBusy = controller.isBusy();
+      if (rlmBusy) controller.abort();
+      const nativeBusy = stopNative?.() ?? false;
+      if (rlmBusy || nativeBusy) {
+        ctx.ui.notify("RLM work aborted — runs, repl cells and background tasks stopped.", "info");
+      } else {
+        ctx.ui.notify("No RLM work in progress.", "info");
       }
-      controller.abort();
-      ctx.ui.notify("RLM run aborted.", "info");
     },
   });
 

package/src/config/defaults.ts

@@ -1,10 +1,18 @@
 import type { RlmConfig } from "../core/types.ts";
 
-/** Frozen default sub-LLM system prompt — avoids re-allocation on every llm_query call. */
+/**
+ * Frozen default sub-LLM system prompt — avoids re-allocation on every llm_query call.
+ * Anthropic prompting canon applied: a role ("precise extraction assistant") and the
+ * hallucination out ("reply exactly NOT_FOUND when the material lacks the answer") — the
+ * convention the glossary documents so roots can branch on a leaf's NOT_FOUND instead of
+ * retrying identical prompts against slices that cannot answer.
+ */
 const DEFAULT_SUB_SYSTEM_PROMPT =
-  "Answer directly and concisely. Return only the requested information. " +
-  "No preamble, no meta-commentary, no explanation of your approach. " +
-  "If listing items, use compact bullet form.";
+  "You are a precise extraction and analysis assistant. " +
+  "Answer directly and concisely from the material provided in the prompt. " +
+  "Return only the requested information — no preamble, no meta-commentary, no explanation of your approach. " +
+  "If listing items, use compact bullet form. " +
+  "If the material does not contain the answer, reply exactly: NOT_FOUND";
 
 export const DEFAULT_CONFIG: Readonly<RlmConfig> = Object.freeze({
   enabled: true,
@@ -64,9 +72,11 @@ export const DEFAULT_CONFIG: Readonly<RlmConfig> = Object.freeze({
   // v5 TaskLedger blackboard
   enableLedger: true,
   rlmBudget: 8,
-  // Verification-discipline nudge — deliberately OFF (plan guardrail): when on, an early
-  // bare-number finalize gets one coached redo instead of being accepted. Opt-in via rlm.json.
-  enableVerificationNudge: false,
+  // Verification-discipline nudge — default ON (knowlange rec: 28/33 bench failures were
+  // early confident wrong answers; the RLM paper ships the same discipline). An early bare
+  // finalize — or one from a run that never inspected its context — gets ONE coached redo
+  // before the answer is accepted. Opt out via rlm.json (`"enableVerificationNudge": false`).
+  enableVerificationNudge: true,
   // SKILL.state integration: Σ_t execution state + cross-session distilled knowledge.
   // Paradigm flags are ENFORCED (R0, /tmp/ROOT_FULL_SKILLSTATE_PLAN.md) — validateEnforcedOn
   // forces true whatever rlm.json carries; only calibrations are tunable.
@@ -78,7 +88,12 @@ export const DEFAULT_CONFIG: Readonly<RlmConfig> = Object.freeze({
   enableSkillStateDistill: true, // ENFORCED — see /tmp/ROOT_FULL_SKILLSTATE_PLAN.md R0
   skillStateMaxTokens: 1_200,
   skillStateLeafTokens: 200,
-  skillStateMinScore: 4.0,
+  // 2.5 (was 4.0 — recall W3): at 4.0 grounding silently no-oped on most prompts; 2.5 keeps
+  // the byte-identical-below-threshold contract while letting genuinely-relevant facts land.
+  skillStateMinScore: 2.5,
+  // Ξ block floor (recall W3): was Number.MIN_VALUE — every positively-scored stale note
+  // rode every root prompt. 2.0 admits relevant notes without the cross-session noise.
+  skillStateXiMinScore: 2.0,
   skillStateNotesPerProject: 128,
   // Root Σ integration (WS-2..WS-4): every LLM call assembles A_t = (P, Σ_t, O_t) — discard
   // semantics on stale payloads + exactly one Σ snapshot splice, and model-proposed ΔΣ_t
@@ -89,8 +104,16 @@ export const DEFAULT_CONFIG: Readonly<RlmConfig> = Object.freeze({
   rootDigestKeepRecentChars: 12_000,
   rootDigestMaxChars: 8_000,
   enableRootContextTransform: true, // ENFORCED — see /tmp/ROOT_FULL_SKILLSTATE_PLAN.md R0 (was soak-OFF pre-v2)
-  rootContextKeepTurns: 2,
-  rootContextElideChars: 1_500,
+  // Recall W4 calibration: 4 verbatim turns (was 2). SKILL.state's budget-matched ablation
+  // (Table 5/11) shows truncated windows collapse recall (0.18 vs structured 0.94); a 2-turn
+  // window sat dangerously close to that shape. Still O(1) per call; the session archive
+  // (rootArchiveMaxChars) makes anything older dereferenceable instead of gone.
+  rootContextKeepTurns: 4,
+  rootContextElideChars: 3_000,
+  // Recall W1: elided turns archive into the sandbox (ctx/session-log/*) so search()/
+  // grep_context() recall them — elision becomes dereferenceable, and the stubs stay honest.
+  // 0 disables the archive (stubs degrade to the plain Σ line).
+  rootArchiveMaxChars: 2_000_000,
   rootContextSnapshot: true,
   enableRootStateFences: true, // ENFORCED — see /tmp/ROOT_FULL_SKILLSTATE_PLAN.md R0 (was soak-OFF pre-v2)
 });

package/src/config/settings.ts

@@ -167,6 +167,8 @@ export function validateConfig(raw: unknown): Partial<RlmConfig> {
   if (skillStateLeafTokens !== undefined) out.skillStateLeafTokens = skillStateLeafTokens;
   const skillStateMinScore = validateNumber(r.skillStateMinScore, 0);
   if (skillStateMinScore !== undefined) out.skillStateMinScore = skillStateMinScore;
+  const skillStateXiMinScore = validateNumber(r.skillStateXiMinScore, 0);
+  if (skillStateXiMinScore !== undefined) out.skillStateXiMinScore = skillStateXiMinScore;
   const skillStateNotesPerProject = validateNumber(r.skillStateNotesPerProject, 1);
   if (skillStateNotesPerProject !== undefined) out.skillStateNotesPerProject = skillStateNotesPerProject;
   // Root Σ integration (WS-2..WS-4) — paradigm flags ENFORCED (R0); the window/byte knobs
@@ -181,6 +183,10 @@ export function validateConfig(raw: unknown): Partial<RlmConfig> {
   if (rootContextKeepTurns !== undefined) out.rootContextKeepTurns = rootContextKeepTurns;
   const rootContextElideChars = validateNumber(r.rootContextElideChars, 100);
   if (rootContextElideChars !== undefined) out.rootContextElideChars = rootContextElideChars;
+  // Recall W1 archive calibration: 0 legitimately disables the archive (plain stubs), so the
+  // floor differs from the other char knobs.
+  const rootArchiveMaxChars = validateNumber(r.rootArchiveMaxChars, 0);
+  if (rootArchiveMaxChars !== undefined) out.rootArchiveMaxChars = rootArchiveMaxChars;
   const rootContextSnapshot = validateBoolean(r.rootContextSnapshot);
   if (rootContextSnapshot !== undefined) out.rootContextSnapshot = rootContextSnapshot;
   out.enableRootStateFences = validateEnforcedOn(r.enableRootStateFences, "enableRootStateFences");