pi-continual-harness 0.3.0 → 0.5.0

package/README.md

@@ -16,8 +16,9 @@ with:
 The durable markdown file (`.pi/harness-state.md`) is the composition seam,
 and it is **two-way**: `/refine --commit` and `/harness export` write it;
 `/harness import` parses it back and merges into the live store (offline edits
-win on conflict), so refinements pi-reflect makes flow back online. pi-mem
-ingestion is a future/manual step.
+win on conflict), so refinements pi-reflect makes flow back online. `/harness push-mem`
+pushes active items into pi-mem's semantic store (see
+[Composing with pi-mem](#composing-with-pi-mem)).
 
 ## Why
 
@@ -40,7 +41,7 @@ leaving offline and storage to the packages that already do them well.
 ## Install
 
 ```
-pi install git:github.com/<owner>/pi-continual-harness
+pi install npm:pi-continual-harness
 ```
 
 Or drop `src/index.ts` into `~/.pi/agent/extensions/`.
@@ -51,6 +52,7 @@ Or drop `src/index.ts` into `~/.pi/agent/extensions/`.
 /refine              # review last 25 turns, propose deltas
 /refine 50           # review last 50 turns
 /refine 25 --commit  # also export durable state to ~/.pi/agent/harness-state.md
+/refine --proposer dedupe  # run the rule-based dedupe proposer instead of steering
 ```
 
 Durable I/O (round-trip with pi-reflect):
@@ -62,6 +64,7 @@ Durable I/O (round-trip with pi-reflect):
 /harness prune [--decay <days>]  # drop items below the importance floor
 /harness keep <id>               # nudge importance up (+0.1)
 /harness drop <id>               # nudge importance down (−0.1)
+/harness push-mem [--all|--kind <kind>]  # persist active items to pi-mem (save_memory)
 ```
 
 `import` reconciles the file into the live store: items whose id matches an
@@ -83,6 +86,24 @@ The model-facing tools:
 Active items are injected into the system prompt each turn as a structured
 block, appended to (never replacing) the base prompt.
 
+## Composing with pi-mem
+
+`/harness push-mem` copies active harness items into
+[pi-mem](https://github.com/georgebashi/pi-mem)'s semantic memory store, so they
+become searchable across sessions. It works by **steering** the agent to call
+pi-mem's `save_memory` tool — there is **no dependency** on pi-mem: it is a
+soft-fail composition. If pi-mem (or any memory tool) is not installed, the
+agent tells you so rather than fabricating one.
+
+```
+pi install npm:pi-mem        # optional companion
+```
+
+By default only **memory**-kind items are pushed (the clean 1:1 mapping); use
+`--all` for every active item or `--kind prompt|skill|subagent` for a specific
+kind. The harness store itself is unchanged by a push — pi-mem gets a separate
+copy.
+
 ## Configuration
 
 Optional config at `~/.pi/agent/harness.json` (missing or malformed → defaults):
@@ -90,8 +111,10 @@ Optional config at `~/.pi/agent/harness.json` (missing or malformed → defaults
 ```json
 {
   "durableScope": "global",
+  "proposer": "steering",
   "remindRefine": { "enabled": false, "everyTurns": 50 },
-  "autoRefine": { "enabled": false, "everyTurns": 100, "commit": false }
+  "autoRefine": { "enabled": false, "everyTurns": 100, "commit": false },
+  "outcomeImportance": { "enabled": false, "bump": 0.03 }
 }
 ```
 
@@ -99,29 +122,84 @@ Optional config at `~/.pi/agent/harness.json` (missing or malformed → defaults
   `~/.pi/agent/harness-state.md`; `"project"` writes to
   `~/.pi/agent/harness-state/<slug>.md` (slug derived from `cwd`) so each
   project keeps separate state for pi-reflect.
+- **`proposer`** — which delta proposer `/refine` and auto-refine use. Defaults
+  to `steering` (the agent reasons via a steering message). `dedupe` applies a
+  rule-based dedupe directly. See [Proposers](#proposers).
 - **`remindRefine`** — opt-in `turn_end` nudge. `{ "enabled": true,
   "everyTurns": 50 }` notifies you to run `/refine` on a cadence. It is
   informational only — it never mutates state.
 - **`autoRefine`** — opt-in autonomous self-improvement (**off by default**).
   When `enabled`, the agent runs `/refine` itself every `everyTurns` turns
-  (default 100). This is the package's one autonomous-self-mutation path: it
-  reuses the exact `/refine` routine (audited `REFINE_ENTRY` tagged
-  `source: "auto"`, branch-local, `/tree` rollback) and notifies before firing.
-  `commit: true` also flushes durable state on each run.
+  (default 100). It is one of the package's opt-in autonomous paths: it reuses
+  the exact `/refine` routine (audited `REFINE_ENTRY` tagged `source: "auto"`,
+  branch-local, `/tree` rollback) and notifies before firing. `commit: true`
+  also flushes durable state on each run.
+- **`outcomeImportance`** — opt-in autonomous **promotion** loop (**off by
+  default**). When `enabled`, a `turn_end` hook bumps (+`bump`, default 0.03)
+  the importance of any active item the agent cited by its `[h_xxxx]` tag in the
+  turn's output. Referenced items gain importance and get their `updatedAt`
+  touched (so they survive time-based decay); ignored items keep decaying. This
+  is the package's second opt-in autonomous path — promotion only, never
+  deletes, persisted/branchable like `harness_mutate`. Autonomous demotion from
+  outcomes is intentionally NOT done (high false-positive); use `/harness drop`,
+  `prune --decay`, or the `dedupe` proposer for that.
 
 ## How it works
 
 1. `/refine` gathers recent trajectory evidence from the current session branch.
-2. It sends a steering user message asking the agent to propose evidence-backed
-   CRUD deltas via `harness_mutate`.
-3. The agent calls the tools; each accepted delta updates the in-memory state,
-   which is snapshotted to the session via `appendEntry("harness-state", ...)`.
+2. A **proposer** decides what to do. The default (`steering`) sends a steering
+   user message asking the agent to propose evidence-backed CRUD deltas via
+   `harness_mutate`; rule-based proposers (e.g. `dedupe`) return deltas the
+   harness applies directly. See [Proposers](#proposers).
+3. The agent calls the tools (steering path); each accepted delta updates the
+   in-memory state, which is snapshotted to the session via
+   `appendEntry("harness-state", ...)`. Direct-apply proposers snapshot the
+   same way.
 4. Because pi's session tree branches at any entry, `/tree` navigation gives
    rollback to any pre-refinement point for free — no bespoke snapshot system.
 
 This reuses the existing agent loop (no nested/hidden model calls), is
 model-agnostic, and keeps every delta visible and reviewable in the transcript.
 
+## Proposers
+
+`/refine` is split into two stages: **propose** (given evidence + state, decide
+what deltas to pursue) and **apply** (send a steering message, or apply returned
+deltas directly). The propose stage is pluggable via a registry
+(`src/proposer.ts`).
+
+| Name | What it does |
+|---|---|
+| `steering` (default) | Delegates reasoning to the agent via a steering message — reuses the agent loop, model-agnostic, fully visible. |
+| `dedupe` | Rule-based: drops near-duplicate active items (token-overlap ≥ 0.6), keeping the higher-importance one. No model call. |
+
+Select a proposer per run with `/refine --proposer <name>`, or set the default
+for auto-refine via `proposer` in the [config](#configuration). Both paths are
+audited: the `harness-refinement` entry records which proposer ran and how many
+deltas it applied directly.
+
+A dedicated-model proposer (a hidden LLM call) is the obvious next alternate
+the interface supports, but is intentionally not shipped — non-visible model
+spend is a tradeoff kept as a separate decision (see `docs/ROADMAP.md`).
+
+Register your own proposer from another extension (the registry is re-exported
+from the package entry):
+
+```ts
+import { registerProposer } from "pi-continual-harness";
+
+registerProposer({
+  name: "my-proposer",
+  async propose({ evidence, state }) {
+    /* inspect evidence + state, return deltas (and/or a steering message) */
+    return { deltas: [/* { delta, rationale } */] };
+  },
+});
+```
+
+Then `"my-proposer"` is selectable via `/refine --proposer my-proposer` or
+`"proposer": "my-proposer"` in the config.
+
 ## Scope and non-goals
 
 - **In scope:** unified state store, online `/refine`, structured deltas,
@@ -132,22 +210,30 @@ model-agnostic, and keeps every delta visible and reviewable in the transcript.
 
 ## Status
 
-0.2.x. Implemented:
+0.5.x. Implemented:
 
 - Unified harness-state store with branch-local snapshots (`/tree` rollback).
 - Online `/refine` + `harness_mutate` / `harness_list` tools.
 - Two-way durable round-trip with pi-reflect (`/harness import|export|status`).
 - Importance hygiene: `/harness prune [--decay <days>]` and `/harness keep|drop
   <id>`.
+- **pi-mem composition**: `/harness push-mem [--all|--kind]` steers the agent
+  to persist active items into pi-mem (soft-fail; no dependency).
 - Optional config (`~/.pi/agent/harness.json`): project-local durable scope, an
-  opt-in `turn_end` reminder, and opt-in `turn_end` auto-refine (the one
-  autonomous-self-mutation path; off by default).
+  opt-in `turn_end` reminder, opt-in `turn_end` auto-refine, and an opt-in
+  `turn_end` outcome-importance loop — the package's two opt-in autonomous
+  paths (both off by default).
+- **Pluggable delta proposers** with a registry: `steering` (default) and
+  `dedupe` (rule-based) shipped; `registerProposer()` for custom ones.
 
 Open extension points (see `docs/ROADMAP.md`):
 
-- A pluggable delta proposer (currently the agent itself, via steering).
-- pi-mem ingestion of the durable export.
-- Outcome-driven importance (currently model-set + manual nudges).
+- A dedicated-model proposer (interface ready; the hidden-model-spend tradeoff
+  is a separate decision).
+- Correction-side outcome signals (promotion is shipped; autonomous demotion
+  is high-false-positive, so it is served by `/harness drop`, `prune --decay`,
+  and the `dedupe` proposer — a fuzzy `corrections` proposer is the natural
+  future extension).
 
 ## License
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-continual-harness",
-  "version": "0.3.0",
+  "version": "0.5.0",
   "description": "Online self-improvement layer for pi. A unified, in-trajectory harness-state store (prompt notes, memory, skill descriptions, sub-agent specs) with a manual /refine that proposes evidence-backed structured CRUD deltas. Composes with pi-reflect (offline) and pi-mem (storage); owns only the online optimizer layer. Inspired by Continual Harness (arXiv 2605.09998) and ACE (arXiv 2510.04618).",
   "keywords": [
     "pi-package",

package/src/auto-refine.ts

@@ -53,7 +53,12 @@ export function registerAutoRefine(pi: ExtensionAPI): void {
     const every = config.autoRefine?.everyTurns ?? DEFAULT_AUTO_EVERY_TURNS;
     ctx.ui.notify(`Auto-refine: running /refine (every ${every} turns, opt-in).`, "info");
     try {
-      await runRefine(pi, ctx, { commit: config.autoRefine?.commit ?? false }, "auto");
+      await runRefine(
+        pi,
+        ctx,
+        { commit: config.autoRefine?.commit ?? false, ...(config.proposer ? { proposer: config.proposer } : {}) },
+        "auto",
+      );
     } catch (err) {
       ctx.ui.notify(`Auto-refine failed: ${(err as Error).message}`, "error");
     }

package/src/config.ts

@@ -17,6 +17,8 @@ import { DEFAULT_DURABLE_PATH } from "./store.js";
 
 export const DEFAULT_EVERY_TURNS = 50;
 export const DEFAULT_AUTO_EVERY_TURNS = 100;
+/** Default per-reference importance bump for the opt-in outcome loop. */
+export const DEFAULT_REF_BUMP = 0.03;
 
 export interface HarnessConfig {
   durableScope?: "global" | "project";
@@ -29,12 +31,22 @@ export interface HarnessConfig {
     everyTurns?: number;
     commit?: boolean;
   };
+  /** Delta proposer name (see proposer.ts registry). Defaults to "steering". */
+  proposer?: string;
+  /** Opt-in turn_end outcome loop: promote importance of items the agent
+   *  references by their [h_xxxx] tag. Off by default (autonomous mutation). */
+  outcomeImportance?: {
+    enabled?: boolean;
+    bump?: number;
+  };
 }
 
 export const DEFAULT_CONFIG: HarnessConfig = {
   durableScope: "global",
   remindRefine: { enabled: false, everyTurns: DEFAULT_EVERY_TURNS },
   autoRefine: { enabled: false, everyTurns: DEFAULT_AUTO_EVERY_TURNS, commit: false },
+  proposer: "steering",
+  outcomeImportance: { enabled: false, bump: DEFAULT_REF_BUMP },
 };
 
 export const CONFIG_PATH = join(homedir(), ".pi", "agent", "harness.json");
@@ -54,6 +66,11 @@ function mergeConfig(over: Partial<HarnessConfig>): HarnessConfig {
       everyTurns: over.autoRefine?.everyTurns ?? DEFAULT_AUTO_EVERY_TURNS,
       commit: over.autoRefine?.commit ?? false,
     },
+    proposer: over.proposer ?? "steering",
+    outcomeImportance: {
+      enabled: over.outcomeImportance?.enabled ?? false,
+      bump: over.outcomeImportance?.bump ?? DEFAULT_REF_BUMP,
+    },
   };
 }
 

package/src/harness.ts

@@ -22,14 +22,15 @@ import {
   reconstructFromDurable,
 } from "./store.js";
 import { loadConfig, resolveDurablePath } from "./config.js";
-import type { ComponentKind } from "./types.js";
+import type { ComponentKind, HarnessItem } from "./types.js";
 
 export function registerHarness(pi: ExtensionAPI): void {
   pi.registerCommand("harness", {
     description:
       "Durable harness-state I/O + importance hygiene. Subcommands: " +
       "import [--prune] [path] · export [path] · status [path] · " +
-      "prune [--decay <days>] · keep <id> · drop <id>.",
+      "prune [--decay <days>] · keep <id> · drop <id> · " +
+      "push-mem [--all|--kind <kind>] (persist active items to pi-mem).",
     handler: async (args, ctx) => {
       const parts = args.trim().split(/\s+/).filter(Boolean);
       const sub = (parts[0] ?? "status").toLowerCase();
@@ -50,6 +51,9 @@ export function registerHarness(pi: ExtensionAPI): void {
         case "drop":
           await handleBump(pi, ctx, rest, -0.1, "drop");
           return;
+        case "push-mem":
+          await handlePushMem(pi, ctx, rest);
+          return;
         case "status":
         default:
           await handleStatus(ctx, rest);
@@ -160,6 +164,61 @@ async function handleBump(
   ctx.ui.notify(`${label}: "${preview}" → importance ${item.importance.toFixed(2)}`, "info");
 }
 
+/** Compose a steering message that asks the agent to persist the given active
+ *  items to long-term memory via a memory tool (pi-mem's save_memory).
+ *  Tool-agnostic: if no memory tool is present the agent says so; we never
+ *  fabricate one. No dependency on pi-mem — soft-fail composition. */
+function buildPushMemPrompt(items: HarnessItem[], scope: string): string {
+  const lines = [
+    `/harness push-mem — persist ${items.length} Continual Harness ${scope} to long-term memory`,
+    "",
+    "For EACH item below, call your memory tool (pi-mem exposes `save_memory`) once, mapping it as shown. This copies harness state into the semantic memory store so it is searchable across sessions.",
+    "",
+    "If you do NOT have a memory tool, do not fabricate one: tell the user to install pi-mem (`pi install npm:pi-mem`) and stop.",
+    "",
+  ];
+  items.forEach((i, n) => {
+    const title = `${i.id} ${i.content.slice(0, 48)}`;
+    const text = `${i.content} (evidence: ${i.evidence})`;
+    lines.push(
+      `${n + 1}. [${i.id}] ${i.content}`,
+      `   evidence: ${i.evidence}`,
+      `   → save_memory({ title: ${JSON.stringify(title)}, text: ${JSON.stringify(text)}, concepts: ["${i.kind}", "continual-harness"] })`,
+      "",
+    );
+  });
+  return lines.join("\n");
+}
+
+async function handlePushMem(
+  pi: ExtensionAPI,
+  ctx: ExtensionCommandContext,
+  rest: string[],
+): Promise<void> {
+  const all = rest.includes("--all");
+  const kindIdx = rest.indexOf("--kind");
+  let kind: ComponentKind | undefined;
+  if (kindIdx >= 0 && kindIdx + 1 < rest.length) {
+    const k = rest[kindIdx + 1]!;
+    if (k === "prompt" || k === "memory" || k === "skill" || k === "subagent") kind = k;
+  }
+  // Default: memory kind (the clean 1:1 mapping). --all or --kind override.
+  const items = getState().items.filter((i) =>
+    i.active && (all || (kind ? i.kind === kind : i.kind === "memory")),
+  );
+  if (items.length === 0) {
+    ctx.ui.notify(
+      "No active items to push (default: memory kind; use --all or --kind <kind>).",
+      "warning",
+    );
+    return;
+  }
+  const scope = all ? "item(s)" : `${kind ?? "memory"} item(s)`;
+  const msg = buildPushMemPrompt(items, scope);
+  pi.sendUserMessage(msg);
+  ctx.ui.notify(`Steering agent to persist ${items.length} ${scope} to pi-mem.`, "info");
+}
+
 async function handleStatus(ctx: ExtensionCommandContext, rest: string[]): Promise<void> {
   const path = await resolvePath(rest, ctx);
   const items = getState().items;

package/src/index.ts

@@ -13,6 +13,7 @@ import { registerRefine } from "./refine.js";
 import { registerTools } from "./tools.js";
 import { registerReminder } from "./remind.js";
 import { registerAutoRefine } from "./auto-refine.js";
+import { registerOutcome } from "./outcome.js";
 import { getState, reconstruct, STATE_ENTRY } from "./store.js";
 
 export default function continualHarness(pi: ExtensionAPI): void {
@@ -32,7 +33,22 @@ export default function continualHarness(pi: ExtensionAPI): void {
   registerRefine(pi);
   registerHarness(pi);
   registerReminder(pi);
+  // Register outcome BEFORE auto-refine: the test fake overwrites turn_end on
+  // each registration (Map.set), and the auto-refine integration tests assume
+  // auto-refine is the surviving handler. In production pi, all three run.
+  registerOutcome(pi);
   registerAutoRefine(pi);
 }
 
 export { STATE_ENTRY };
+
+// Public extension API: other extensions can register their own delta proposer
+// (see src/proposer.ts) and it becomes selectable via /refine --proposer <name>
+// or the `proposer` config key.
+export {
+  listProposers,
+  registerProposer,
+  type DeltaProposer,
+  type ProposeInput,
+  type ProposeResult,
+} from "./proposer.js";

package/src/outcome.ts

@@ -0,0 +1,104 @@
+// turn_end outcome loop (Phase 5 / limit B-phase-2). The package's SECOND
+// autonomous-mutation path (after auto-refine), so — like auto-refine — it is
+// locked behind an explicit opt-in:
+//
+//   ~/.pi/agent/harness.json
+//   { "outcomeImportance": { "enabled": true, "bump": 0.03 } }
+//
+// What it does: when the agent's own output cites an injected item by its
+// [h_xxxx] tag, that is unambiguous POSITIVE evidence the item was useful, so
+// we nudge its importance up (+bump). This closes the outcome half of the
+// fitness loop that manual /harness keep|drop (B-phase-1) and /harness prune
+// --decay (A) started: useful items gain importance AND get their updatedAt
+// touched (so they survive time-based decay); ignored items keep decaying.
+//
+// Hard scope cut — CORRECTION/demotion from outcomes is genuinely fuzzy and
+// high-false-positive, so it is intentionally NOT autonomized here. Demotion is
+// served by the existing audited, reviewable primitives: /harness drop,
+// /harness prune --decay, and the dedupe proposer. (See ROADMAP.md.)
+//
+// Safety properties (promotion only, never deletes): clamped to [0,1]; persisted
+// via the same harness-state entries as harness_mutate (branch-local, /tree
+// rollback); visible (notify per turn). Pure reference extraction is factored
+// into findReferences() so it is unit-testable without a live pi runtime.
+
+import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
+import { DEFAULT_REF_BUMP, type HarnessConfig, loadConfig } from "./config.js";
+import { bumpImportance, getState } from "./store.js";
+
+// Minimal entry shape; kept loose to avoid coupling to pi's internal types.
+type AnyEntry = {
+  type?: string;
+  message?: {
+    role?: string;
+    content?: Array<{ type?: string; text?: string }>;
+  };
+};
+
+// Matches [h_xxxx] citation tags (the same form inject.ts emits).
+const REF_RE = /\[(h_[a-z0-9_]+)\]/gi;
+
+/**
+ * Extract distinct harness-item ids the agent cited by their [h_xxxx] tag,
+ * keeping only those that are currently known (active or not). Pure.
+ */
+export function findReferences(text: string, knownIds: Set<string>): string[] {
+  const found = new Set<string>();
+  for (const m of text.matchAll(REF_RE)) {
+    const id = m[1]!.toLowerCase();
+    if (knownIds.has(id)) found.add(id);
+  }
+  return [...found];
+}
+
+// Number of assistant message-entries already consumed. -1 = unseen; the first
+// observed turn_end seeds the cursor (so enabling mid-session never bumps
+// retroactively for the whole history).
+let scannedMessages = -1;
+
+/** Test hook: reset internal cursor. */
+export function resetOutcome(): void {
+  scannedMessages = -1;
+}
+
+/** Subscribe to turn_end and promote importance of referenced items (opt-in). */
+export function registerOutcome(pi: ExtensionAPI): void {
+  pi.on("turn_end", async (_event, ctx) => {
+    const config = await loadConfig();
+    if (!config.outcomeImportance?.enabled) return;
+    const bump = config.outcomeImportance.bump ?? DEFAULT_REF_BUMP;
+    if (bump <= 0) return;
+
+    const entries = ctx.sessionManager.getBranch() as AnyEntry[];
+    const asst = entries.filter((e) => e.type === "message" && e.message?.role === "assistant");
+    if (scannedMessages < 0) {
+      scannedMessages = asst.length; // seed: no retroactive bump
+      return;
+    }
+    const fresh = asst.slice(scannedMessages);
+    scannedMessages = asst.length;
+
+    const text = fresh
+      .map((e) => (e.message?.content ?? []).map((c) => c.text ?? "").join(" "))
+      .join("\n");
+    if (!text.trim()) return;
+
+    const activeIds = new Set(getState().items.filter((i) => i.active).map((i) => i.id));
+    const refs = findReferences(text, activeIds);
+    if (refs.length === 0) return;
+
+    let bumped = 0;
+    for (const id of refs) {
+      const item = bumpImportance(id, bump, (snapshot, ver) => {
+        pi.appendEntry("harness-state", { state: snapshot, version: ver });
+      });
+      if (item) bumped += 1;
+    }
+    if (bumped > 0) {
+      ctx.ui.notify(
+        `Outcome: +${bump} importance for ${bumped} referenced item(s): ${refs.join(", ")}`,
+        "info",
+      );
+    }
+  });
+}

package/src/proposer.ts

@@ -0,0 +1,174 @@
+// Pluggable delta proposers (Phase 4 of the roadmap).
+//
+// /refine is split into two stages:
+//   1. PROPOSE — given trajectory evidence + current state, decide what deltas
+//      to apply (or what to ask the agent to do).
+//   2. APPLY   — either send a steering message (the agent reasons + calls
+//      harness_mutate) or apply returned deltas directly.
+//
+// This file owns stage 1. The default `steeringProposer` preserves the
+// original behavior exactly (delegates to the agent via a steering message,
+// reusing the agent loop — model-agnostic, fully visible). Alternate proposers
+// produce deltas directly:
+//   - `dedupeProposer` (rule-based): drops near-duplicate active items by token
+//     overlap, keeping the higher-importance one. Pure function of state —
+//     deterministic, no model call, fully testable.
+//
+// A dedicated-model proposer (a hidden LLM call) is the obvious next alternate
+// the interface supports, but is intentionally NOT shipped here: it is the one
+// tradeoff (non-visible model spend) the roadmap flags as a separate decision.
+//
+// Extension is open: call registerProposer() from your own extension to add a
+// named proposer, then select it via /refine --proposer <name> or config.
+
+import type { Delta, HarnessItem, HarnessState } from "./types.js";
+
+/** Inputs to a proposer. */
+export interface ProposeInput {
+  /** Recent trajectory evidence (the same string /refine gathers). */
+  evidence: string;
+  /** Current harness state as a defensive copy; mutations are ignored — use
+   *  the returned deltas to change state. */
+  state: HarnessState;
+  /** Lookback window in turns. */
+  lookback: number;
+}
+
+/** A single delta plus a human-readable reason for the audit trail. */
+export interface ProposedDelta {
+  delta: Delta;
+  rationale: string;
+}
+
+export interface ProposeResult {
+  /** Deltas the proposer produced directly (rule-based / model proposers). */
+  deltas?: ProposedDelta[];
+  /**
+   * A steering message for the agent, when the proposer delegates reasoning to
+   * the agent loop (the default). runRefine sends this via sendUserMessage.
+   * A proposer MAY return both — deltas are applied first, then steering sent.
+   */
+  steeringMessage?: string;
+}
+
+/** A strategy for turning evidence + state into (or toward) harness deltas. */
+export interface DeltaProposer {
+  readonly name: string;
+  propose(input: ProposeInput): Promise<ProposeResult>;
+}
+
+// ---- default: steering (delegates reasoning to the agent) ------------------
+
+function buildSteeringPrompt(evidence: string, lookback: number): string {
+  return [
+    `/refine (online self-improvement, last ${lookback} turns as evidence)`,
+    "",
+    "Review the trajectory evidence below. Identify DURABLE, REUSABLE corrections — things that will help in future turns, not one-off fixes. Then update the Continual Harness state.",
+    "",
+    "Steps:",
+    "1. Call harness_list to see the current state.",
+    "2. Call harness_mutate with small, surgical CRUD deltas. Every create MUST include concrete evidence from the trajectory. Prefer updating existing items over creating duplicates; delete items that are wrong, stale, or contradicted.",
+    "3. Keep prompt notes terse and behavioral. Memory facts should be specific. Skill/sub-agent entries should describe reusable patterns, not one task.",
+    "",
+    "Constraints: never rewrite the whole store; ground every change in evidence; if nothing durable is worth recording, say so and do nothing.",
+    "",
+    "## Trajectory evidence",
+    evidence,
+  ].join("\n");
+}
+
+/** Default proposer: delegates reasoning to the agent via a steering message. */
+export const steeringProposer: DeltaProposer = {
+  name: "steering",
+  async propose({ evidence, lookback }): Promise<ProposeResult> {
+    return { steeringMessage: buildSteeringPrompt(evidence, lookback) };
+  },
+};
+
+// ---- rule-based alternate: dedupe -----------------------------------------
+
+export const DEDUPE_THRESHOLD = 0.6;
+
+/** Tokenize for overlap comparison: lowercase alnum tokens. */
+export function tokenize(text: string): string[] {
+  return text.toLowerCase().replace(/[^a-z0-9\s]/g, " ").split(/\s+/).filter(Boolean);
+}
+
+/** Jaccard token overlap in [0,1]. 0 if either side has no tokens. */
+export function tokenOverlap(a: string, b: string): number {
+  const sa = new Set(tokenize(a));
+  const sb = new Set(tokenize(b));
+  if (sa.size === 0 || sb.size === 0) return 0;
+  let inter = 0;
+  for (const t of sa) if (sb.has(t)) inter++;
+  return inter / (sa.size + sb.size - inter);
+}
+
+function truncate(s: string, n = 48): string {
+  return s.length > n ? `${s.slice(0, n - 1)}…` : s;
+}
+
+/**
+ * Rule-based proposer: drops near-duplicate ACTIVE items (same kind, Jaccard
+ * token overlap >= DEDUPE_THRESHOLD), keeping the higher-importance one.
+ *
+ * Greedy and contradiction-free: items are considered in importance-descending
+ * order; the first item is always a keeper, and each later item is compared
+ * only against keepers, so a keeper is never subsequently dropped.
+ */
+export const dedupeProposer: DeltaProposer = {
+  name: "dedupe",
+  async propose({ state }): Promise<ProposeResult> {
+    const ordered = state.items
+      .filter((i) => i.active)
+      .sort((a, b) => b.importance - a.importance);
+    const keepers: HarnessItem[] = [];
+    const proposals: ProposedDelta[] = [];
+    for (const cand of ordered) {
+      let dup: HarnessItem | undefined;
+      let best = 0;
+      for (const k of keepers) {
+        if (k.kind !== cand.kind) continue;
+        const sim = tokenOverlap(k.content, cand.content);
+        if (sim >= DEDUPE_THRESHOLD && sim > best) {
+          best = sim;
+          dup = k;
+        }
+      }
+      if (dup) {
+        proposals.push({
+          delta: {
+            op: "delete",
+            id: cand.id,
+            reason: `near-duplicate of ${dup.id} (overlap ${best.toFixed(2)})`,
+          },
+          rationale: `dedupe: "${truncate(cand.content)}" ≈ keeper "${truncate(dup.content)}" (Jaccard ${best.toFixed(2)}); kept higher-importance ${dup.id}.`,
+        });
+      } else {
+        keepers.push(cand);
+      }
+    }
+    return proposals.length ? { deltas: proposals } : {};
+  },
+};
+
+// ---- registry -------------------------------------------------------------
+
+const registry = new Map<string, DeltaProposer>([
+  ["steering", steeringProposer],
+  ["dedupe", dedupeProposer],
+]);
+
+/** Register (or replace) a named proposer. For external extensions. */
+export function registerProposer(p: DeltaProposer): void {
+  registry.set(p.name, p);
+}
+
+/** Look up a proposer by name; falls back to the steering default. */
+export function getProposer(name: string | undefined): DeltaProposer {
+  return registry.get(name ?? "steering") ?? steeringProposer;
+}
+
+export function listProposers(): string[] {
+  return [...registry.keys()];
+}

package/src/refine.ts

@@ -1,19 +1,25 @@
-// /refine — the online optimizer. Manual only (no turn_end auto-trigger).
+// /refine — the online optimizer. Manual by default; opt-in turn_end
+// auto-refine lives in auto-refine.ts and reuses runRefine().
 //
 // Flow:
 //  1. Gather recent trajectory evidence from the current branch.
-//  2. Send a steering user message that asks the agent to propose evidence-
-//     backed CRUD deltas via harness_mutate (and to inspect state with
-//     harness_list first).
-//  3. The agent does the reasoning and calls the tools. Each accepted delta is
-//     recorded as a session entry, so /tree gives rollback for free.
+//  2. Hand evidence + current state to a pluggable DeltaProposer (proposer.ts).
+//     The default `steering` proposer returns a steering user message; the
+//     agent then reasons and calls harness_mutate. Rule-based/model proposers
+//     return deltas directly, which runRefine applies itself.
+//  3. Every mutation — whether from the agent (harness_mutate) or a direct-apply
+//     proposer — is recorded as a session entry, so /tree gives rollback for
+//     free. Which proposer ran is captured in the audit entry.
 //
-// Why a steering message instead of a nested LLM call: it reuses the existing
-// agent loop, is model-agnostic, and keeps every delta visible and reviewable
-// in the transcript. No hidden model calls.
+// Why a steering message as the DEFAULT instead of a nested LLM call: it reuses
+// the existing agent loop, is model-agnostic, and keeps every delta visible and
+// reviewable in the transcript. Alternate proposers can opt out via the
+// registry; the dedicated-model variant is intentionally not shipped (hidden
+// model spend is a tradeoff the roadmap keeps as a separate decision).
 
 import type { ExtensionAPI, ExtensionCommandContext, ExtensionContext } from "@earendil-works/pi-coding-agent";
-import { exportDurable, REFINE_ENTRY } from "./store.js";
+import { applyDeltas, exportDurable, REFINE_ENTRY, snapshotState } from "./store.js";
+import { getProposer } from "./proposer.js";
 
 const DEFAULT_EVIDENCE_BYTES = 16000;
 export const DEFAULT_LOOKBACK_TURNS = 25;
@@ -32,10 +38,10 @@ type AnyEntry = {
 export function registerRefine(pi: ExtensionAPI): void {
   pi.registerCommand("refine", {
     description:
-      "Online self-improvement: review recent trajectory and propose evidence-backed CRUD deltas to the harness state. Usage: /refine [lookback-turns] [--commit]",
+      "Online self-improvement: review recent trajectory and propose evidence-backed CRUD deltas to the harness state. Usage: /refine [lookback-turns] [--commit] [--proposer steering|dedupe]",
     handler: async (args, ctx) => {
-      const { lookback, commit } = parseArgs(args);
-      await runRefine(pi, ctx, { lookback, commit });
+      const { lookback, commit, proposer } = parseArgs(args);
+      await runRefine(pi, ctx, { lookback, commit, ...(proposer ? { proposer } : {}) });
     },
   });
 }
@@ -43,6 +49,8 @@ export function registerRefine(pi: ExtensionAPI): void {
 export interface RefineOptions {
   lookback?: number;
   commit?: boolean;
+  /** Proposer name (see proposer.ts registry). Defaults to "steering". */
+  proposer?: string;
 }
 
 export interface RefineResult {
@@ -50,13 +58,18 @@ export interface RefineResult {
   lookback: number;
   commit: boolean;
   source: "manual" | "auto";
+  /** Which proposer ran. */
+  proposer: string;
+  /** Deltas the proposer applied directly (0 for the steering path). */
+  applied: number;
 }
 
-/** Core refine routine shared by /refine (manual) and turn_end auto-refine.
- *  Gathers trajectory evidence, records an audit entry, optionally flushes the
- *  durable state, then sends the steering user message that drives the agent to
- *  propose evidence-backed CRUD deltas. `source` tags the audit entry so
- *  autonomous runs are distinguishable in the transcript. */
+/**
+ * Core refine routine shared by /refine (manual) and turn_end auto-refine.
+ * Resolves a proposer, gathers evidence, lets the proposer decide what to do,
+ * then either applies returned deltas directly or sends a steering message.
+ * `source` tags the audit entry so autonomous runs are distinguishable.
+ */
 export async function runRefine(
   pi: ExtensionAPI,
   ctx: ExtensionContext,
@@ -65,12 +78,40 @@ export async function runRefine(
 ): Promise<RefineResult> {
   const lookback = options.lookback ?? DEFAULT_LOOKBACK_TURNS;
   const commit = options.commit ?? false;
-  ctx.ui.setStatus("harness", `Refining (last ${lookback} turns)…`);
+  const proposer = getProposer(options.proposer);
+  ctx.ui.setStatus("harness", `Refining (last ${lookback} turns)… [${proposer.name}]`);
+
   const evidence = gatherEvidence(ctx, lookback);
-  const prompt = buildSteeringPrompt(evidence, lookback);
+  // Pass a defensive copy: proposers are a public extension point and must not
+  // be able to mutate the live store outside applyDeltas (no audit / no rollback).
+  const result = await proposer.propose({ evidence, state: snapshotState(), lookback });
+  const proposedDeltas = result.deltas ?? [];
+
+  let applied = 0;
+  if (proposedDeltas.length > 0) {
+    // Direct-apply path (rule-based / model proposers): persist each batch the
+    // same way harness_mutate does, so /tree rollback covers these too.
+    const appliedDeltas = applyDeltas(
+      proposedDeltas.map((d) => d.delta),
+      (snapshot, ver) => pi.appendEntry("harness-state", { state: snapshot, version: ver }),
+    );
+    applied = appliedDeltas.length;
+    ctx.ui.notify(`Proposer "${proposer.name}" applied ${applied} delta(s) directly.`, "info");
+  }
+
   // Audit trail (branchable via /tree). `source` distinguishes manual /refine
-  // from opt-in auto-refine so autonomous runs are visible in the transcript.
-  pi.appendEntry(REFINE_ENTRY, { lookback, commit, startedAt: Date.now(), source });
+  // from opt-in auto-refine; `proposer` records WHICH strategy ran; the
+  // rationales make rule-based decisions visible in the transcript.
+  pi.appendEntry(REFINE_ENTRY, {
+    lookback,
+    commit,
+    startedAt: Date.now(),
+    source,
+    proposer: proposer.name,
+    applied,
+    rationales: proposedDeltas.map((d) => d.rationale),
+  });
+
   if (commit) {
     // Flush current state to the durable file first so the agent refines
     // against the same view pi-reflect / pi-mem would see.
@@ -81,20 +122,38 @@ export async function runRefine(
       ctx.ui.notify(`Durable export failed: ${(err as Error).message}`, "warning");
     }
   }
-  pi.sendUserMessage(prompt);
+
+  if (result.steeringMessage) {
+    pi.sendUserMessage(result.steeringMessage);
+  }
   ctx.ui.setStatus("harness", undefined);
-  return { evidenceBytes: evidence.length, lookback, commit, source };
+  return {
+    evidenceBytes: evidence.length,
+    lookback,
+    commit,
+    source,
+    proposer: proposer.name,
+    applied,
+  };
 }
 
-function parseArgs(args: string): { lookback: number; commit: boolean } {
+function parseArgs(args: string): {
+  lookback: number;
+  commit: boolean;
+  proposer: string | undefined;
+} {
   const parts = args.trim().split(/\s+/).filter(Boolean);
   let lookback = DEFAULT_LOOKBACK_TURNS;
   let commit = false;
-  for (const p of parts) {
+  let proposer: string | undefined;
+  for (let i = 0; i < parts.length; i++) {
+    const p = parts[i]!;
     if (p === "--commit") commit = true;
+    else if (p === "--proposer" && i + 1 < parts.length) proposer = parts[++i];
+    else if (p.startsWith("--proposer=")) proposer = p.slice("--proposer=".length);
     else if (/^\d+$/.test(p)) lookback = Math.max(1, Math.min(200, Number(p)));
   }
-  return { lookback, commit };
+  return { lookback, commit, proposer };
 }
 
 function gatherEvidence(ctx: ExtensionContext, lookback: number): string {
@@ -121,21 +180,3 @@ function gatherEvidence(ctx: ExtensionContext, lookback: number): string {
   }
   return lines.join("\n") || "(no recent trajectory evidence found)";
 }
-
-function buildSteeringPrompt(evidence: string, lookback: number): string {
-  return [
-    `/refine (online self-improvement, last ${lookback} turns as evidence)`,
-    "",
-    "Review the trajectory evidence below. Identify DURABLE, REUSABLE corrections — things that will help in future turns, not one-off fixes. Then update the Continual Harness state.",
-    "",
-    "Steps:",
-    "1. Call harness_list to see the current state.",
-    "2. Call harness_mutate with small, surgical CRUD deltas. Every create MUST include concrete evidence from the trajectory. Prefer updating existing items over creating duplicates; delete items that are wrong, stale, or contradicted.",
-    "3. Keep prompt notes terse and behavioral. Memory facts should be specific. Skill/sub-agent entries should describe reusable patterns, not one task.",
-    "",
-    "Constraints: never rewrite the whole store; ground every change in evidence; if nothing durable is worth recording, say so and do nothing.",
-    "",
-    "## Trajectory evidence",
-    evidence,
-  ].join("\n");
-}

package/src/store.ts

@@ -32,6 +32,13 @@ export function getState(): HarnessState {
   return state;
 }
 
+/** A defensive deep-enough copy of current state (items are flat). Use when
+ *  handing state to untrusted/external code (e.g. a DeltaProposer) so it cannot
+ *  mutate the live store outside applyDeltas. */
+export function snapshotState(): HarnessState {
+  return { items: state.items.map((i) => ({ ...i })) };
+}
+
 export function listItems(kind?: ComponentKind): HarnessItem[] {
   return kind ? state.items.filter((i) => i.kind === kind) : state.items;
 }