pi-continual-harness 0.1.1 → 0.2.0

package/README.md

@@ -13,8 +13,11 @@ with:
   → behavioral-file refinement (the "deep" path).
 - **pi-mem / pi-memory** — durable memory storage.
 
-The durable markdown export (`.pi/harness-state.md`) is the composition seam:
-pi-reflect can read and refine it, pi-mem can ingest it.
+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.
 
 ## Why
 
@@ -50,6 +53,24 @@ Or drop `src/index.ts` into `~/.pi/agent/extensions/`.
 /refine 25 --commit  # also export durable state to ~/.pi/agent/harness-state.md
 ```
 
+Durable I/O (round-trip with pi-reflect):
+
+```
+/harness status                  # counts + durable file presence/mtime
+/harness export [path]           # write active items to a markdown file
+/harness import [--prune] [path] # parse it back and merge (offline edits win)
+```
+
+`import` reconciles the file into the live store: items whose id matches an
+existing entry are updated (offline edits win on content/evidence/importance),
+new entries are created. By default nothing is deleted — `--prune` also drops
+active items whose id is no longer in the file (inactive items are always
+preserved). Point pi-reflect at the same file to refine it offline:
+
+```
+/reflect ~/.pi/agent/harness-state.md
+```
+
 The model-facing tools:
 
 - `harness_list [kind]` — read current state (optionally filtered by kind).
@@ -82,13 +103,15 @@ model-agnostic, and keeps every delta visible and reviewable in the transcript.
 
 ## Status
 
-Scaffold (0.1.0). Extension points to grow into:
+0.1.x. The durable round-trip with pi-reflect is implemented
+(`/harness import|export|status`). Open extension points:
 
 - A pluggable delta proposer (currently the agent itself, via steering).
 - Project-local vs global durable paths.
 - Optional `turn_end` auto-trigger behind a setting (intentionally off by
   default — autonomous self-mutation is a sharp edge).
-- Importance decay/prune policy.
+- Importance decay/prune policy — `decayAndPrune()` exists but is not yet wired.
+- pi-mem ingestion of the durable export (manual/future).
 
 ## License
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-continual-harness",
-  "version": "0.1.1",
+  "version": "0.2.0",
   "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/harness.ts

@@ -0,0 +1,110 @@
+// /harness — durable I/O. The two-way round-trip seam with pi-reflect.
+//
+//   /harness status [path]               counts + durable file presence/mtime
+//   /harness export [path]               write active items to a markdown file
+//   /harness import [--prune] [path]     parse it back and merge (durable wins)
+//
+// export writes the active items to ~/.pi/agent/harness-state.md (best-effort);
+// import parses that file and merges into the live store. Because pi-reflect
+// edits markdown files and git-commits, pointing it at the same file closes the
+// loop: offline refinement flows back into the online store.
+//
+// Manual only (no session_start auto-import): importing is an explicit,
+// reviewable action, matching the package's "no autonomous mutation" stance.
+
+import type { ExtensionAPI, ExtensionCommandContext } from "@earendil-works/pi-coding-agent";
+import { stat } from "node:fs/promises";
+import { DEFAULT_DURABLE_PATH, exportDurable, getState, reconstructFromDurable } from "./store.js";
+import type { ComponentKind } from "./types.js";
+
+export function registerHarness(pi: ExtensionAPI): void {
+  pi.registerCommand("harness", {
+    description:
+      "Durable harness-state I/O (round-trip seam with pi-reflect). " +
+      "Subcommands: import [--prune] [path] · export [path] · status [path].",
+    handler: async (args, ctx) => {
+      const parts = args.trim().split(/\s+/).filter(Boolean);
+      const sub = (parts[0] ?? "status").toLowerCase();
+      const rest = parts.slice(1);
+      switch (sub) {
+        case "import":
+          await handleImport(pi, ctx, rest);
+          return;
+        case "export":
+          await handleExport(ctx, rest);
+          return;
+        case "status":
+        default:
+          await handleStatus(ctx, rest);
+          return;
+      }
+    },
+  });
+}
+
+function resolvePath(rest: string[]): string {
+  return rest.find((a) => !a.startsWith("-")) ?? DEFAULT_DURABLE_PATH;
+}
+
+async function handleImport(
+  pi: ExtensionAPI,
+  ctx: ExtensionCommandContext,
+  rest: string[],
+): Promise<void> {
+  const prune = rest.includes("--prune");
+  const path = resolvePath(rest);
+  ctx.ui.setStatus("harness", `Importing durable state${prune ? " (prune)" : ""}…`);
+  try {
+    const res = await reconstructFromDurable(path, { prune }, (snapshot, ver) => {
+      pi.appendEntry("harness-state", { state: snapshot, version: ver });
+    });
+    if (res.missingFile) {
+      ctx.ui.notify(
+        `No durable file at ${path}. Run /refine --commit or /harness export first.`,
+        "warning",
+      );
+      return;
+    }
+    const bits = [`${res.created} created`, `${res.updated} updated`];
+    if (prune) bits.push(`${res.pruned} pruned`);
+    ctx.ui.notify(`Imported ${res.imported} item(s) from ${path} (${bits.join(", ")}).`, "info");
+  } catch (err) {
+    ctx.ui.notify(`Harness import failed: ${(err as Error).message}`, "error");
+  }
+  ctx.ui.setStatus("harness", undefined);
+}
+
+async function handleExport(ctx: ExtensionCommandContext, rest: string[]): Promise<void> {
+  const path = resolvePath(rest);
+  ctx.ui.setStatus("harness", "Exporting durable state…");
+  try {
+    const written = await exportDurable(path);
+    const n = getState().items.filter((i) => i.active).length;
+    ctx.ui.notify(`Exported ${n} active item(s) to ${written}`, "info");
+  } catch (err) {
+    ctx.ui.notify(`Harness export failed: ${(err as Error).message}`, "error");
+  }
+  ctx.ui.setStatus("harness", undefined);
+}
+
+async function handleStatus(ctx: ExtensionCommandContext, rest: string[]): Promise<void> {
+  const path = resolvePath(rest);
+  const items = getState().items;
+  const active = items.filter((i) => i.active);
+  const counts: Record<ComponentKind, number> = { prompt: 0, memory: 0, skill: 0, subagent: 0 };
+  for (const i of active) counts[i.kind] += 1;
+  let fileState: string;
+  try {
+    const st = await stat(path);
+    fileState = `${path} (modified ${st.mtime.toISOString()})`;
+  } catch {
+    fileState = `none at ${path}`;
+  }
+  ctx.ui.setStatus("harness", undefined);
+  ctx.ui.notify(
+    `Harness: ${active.length} active / ${items.length} total — ` +
+      `prompt ${counts.prompt}, memory ${counts.memory}, skill ${counts.skill}, subagent ${counts.subagent}. ` +
+      `Durable: ${fileState}.`,
+    "info",
+  );
+}

package/src/index.ts

@@ -7,6 +7,7 @@
 // See README for design rationale and the research it is grounded in.
 
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
+import { registerHarness } from "./harness.js";
 import { registerInjection } from "./inject.js";
 import { registerRefine } from "./refine.js";
 import { registerTools } from "./tools.js";
@@ -27,6 +28,7 @@ export default function continualHarness(pi: ExtensionAPI): void {
   registerInjection(pi);
   registerTools(pi);
   registerRefine(pi);
+  registerHarness(pi);
 }
 
 export { STATE_ENTRY };

package/src/store.ts

@@ -10,7 +10,7 @@
 //     to a markdown file pi-reflect can read and refine offline, and pi-mem can
 //     ingest. Best-effort; the package works without it.
 
-import { mkdir, writeFile } from "node:fs/promises";
+import { mkdir, readFile, writeFile } from "node:fs/promises";
 import { homedir } from "node:os";
 import { dirname, join } from "node:path";
 import type { AppliedDelta, ComponentKind, Delta, HarnessItem, HarnessState } from "./types.js";
@@ -18,7 +18,7 @@ import type { AppliedDelta, ComponentKind, Delta, HarnessItem, HarnessState } fr
 const STATE_ENTRY = "harness-state";
 const REFINE_ENTRY = "harness-refinement";
 
-const DEFAULT_DURABLE_PATH = join(homedir(), ".pi", "agent", "harness-state.md");
+export const DEFAULT_DURABLE_PATH = join(homedir(), ".pi", "agent", "harness-state.md");
 
 const IMPORTANCE_FLOOR = 0.3;
 
@@ -167,3 +167,190 @@ function titleFor(kind: ComponentKind): string {
       return "Sub-agent specs";
   }
 }
+
+// ---- Durable import (round-trip seam with pi-reflect) ---------------------
+//
+// exportDurable() is write-only by design. reconstructFromDurable() closes the
+// loop: it parses the markdown back into items and merges them into the live
+// store, so offline edits pi-reflect makes to harness-state.md flow back in.
+//
+// Merge semantics (predictable, loss-free by default):
+//   - parsed item whose id matches an existing item → UPDATE in place
+//     (durable wins on content/evidence/importance; reactivated; createdAt kept).
+//   - parsed item with a new/foreign id → CREATE.
+//   - items in the store but absent from the file → KEPT by default.
+//     Pass { prune: true } to also drop active items whose id is not in the
+//     file (inactive items are always preserved — the durable export never
+//     contains them, so they cannot have been "deleted" by pi-reflect).
+
+export interface DurableImportResult {
+  /** Items successfully parsed from the file. */
+  imported: number;
+  created: number;
+  updated: number;
+  /** Only non-zero when { prune: true }. */
+  pruned: number;
+  /** True if the file did not exist (no-op). */
+  missingFile: boolean;
+}
+
+interface ParsedItem {
+  id?: string;
+  kind: ComponentKind;
+  importance: number;
+  content: string;
+  evidence: string;
+}
+
+// Section title → kind. Exact export titles first, then tolerant keyword
+// fallbacks so pi-reflect's edits to headings still parse.
+const TITLE_TO_KIND: Array<[RegExp, ComponentKind]> = [
+  [/supplemental prompt notes/i, "prompt"],
+  [/memory facts/i, "memory"],
+  [/skill descriptions/i, "skill"],
+  [/sub-?agent specs/i, "subagent"],
+  [/\bprompt\b/i, "prompt"],
+  [/\bmemory\b/i, "memory"],
+  [/\bskills?\b/i, "skill"],
+  [/\bsub-?agents?\b/i, "subagent"],
+];
+
+function titleToKind(title: string): ComponentKind | undefined {
+  for (const [re, kind] of TITLE_TO_KIND) if (re.test(title)) return kind;
+  return undefined;
+}
+
+const RE_H2 = /^##\s+(.*)$/;
+const RE_ID_BULLET = /^-\s+\*\*\[([^\]]+)\]\*\*\s*\(importance\s+([\d.]+)\)\s*(.*)$/;
+const RE_PLAIN_BULLET = /^-\s+(.+)$/;
+const RE_EVIDENCE = /^\s+-\s+evidence:\s*(.*)$/i;
+
+/** Parse a durable markdown export into items. Tolerant of pi-reflect's edits. */
+export function parseDurable(text: string): ParsedItem[] {
+  const out: ParsedItem[] = [];
+  let kind: ComponentKind | undefined;
+  let pending: ParsedItem | null = null;
+  const flush = (): void => {
+    if (pending && pending.content) out.push(pending);
+    pending = null;
+  };
+
+  for (const raw of text.split(/\r?\n/)) {
+    const line = raw.replace(/\s+$/, "");
+    const h2 = line.match(RE_H2);
+    if (h2) {
+      flush();
+      kind = titleToKind(h2[1]!.trim());
+      continue;
+    }
+    if (/^#/.test(line)) {
+      flush(); // any other header (h1, h3, …)
+      continue;
+    }
+    if (!kind) continue; // ignore bullets outside a known section
+
+    const ev = line.match(RE_EVIDENCE);
+    if (ev) {
+      if (pending) pending.evidence = ev[1]!.trim();
+      continue;
+    }
+
+    const idm = line.match(RE_ID_BULLET);
+    if (idm) {
+      flush();
+      const imp = Number(idm[2]!);
+      pending = {
+        id: idm[1]!.trim(),
+        kind,
+        importance: Number.isNaN(imp) ? 0.5 : imp,
+        content: idm[3]!.trim(),
+        evidence: "",
+      };
+      continue;
+    }
+
+    const pm = line.match(RE_PLAIN_BULLET);
+    if (pm) {
+      flush();
+      const body = pm[1]!.trim();
+      const idInBody = body.match(/^\*\*\[([^\]]+)\]\*\*/);
+      pending = {
+        kind,
+        importance: 0.5,
+        content: body.replace(/^\*\*\[([^\]]+)\]\*\*\s*/, "").trim(),
+        evidence: "",
+      };
+      if (idInBody) pending.id = idInBody[1]!.trim();
+      continue;
+    }
+  }
+  flush();
+  return out.filter((p) => p.content && !/^\(?no active items\)?$/i.test(p.content));
+}
+
+/**
+ * Parse the durable file and merge it into the live store, then persist a
+ * snapshot. See the file-level comment above for merge semantics.
+ */
+export async function reconstructFromDurable(
+  path: string,
+  options: { prune?: boolean },
+  persist: (snapshot: HarnessState, version: number) => void,
+): Promise<DurableImportResult> {
+  let text: string;
+  try {
+    text = await readFile(path, "utf8");
+  } catch {
+    return { imported: 0, created: 0, updated: 0, pruned: 0, missingFile: true };
+  }
+
+  const parsed = parseDurable(text);
+  const fileIds = new Set(parsed.map((p) => p.id).filter((id): id is string => Boolean(id)));
+  const existingById = new Map(state.items.map((i) => [i.id, i] as const));
+  const preActiveIds = new Set(state.items.filter((i) => i.active).map((i) => i.id));
+
+  let created = 0;
+  let updated = 0;
+  const now = Date.now();
+  for (const p of parsed) {
+    const existing = p.id ? existingById.get(p.id) : undefined;
+    if (existing) {
+      // Durable wins; reactivate; keep createdAt.
+      existing.content = p.content;
+      existing.evidence = p.evidence;
+      existing.importance = clamp(p.importance);
+      existing.active = true;
+      existing.updatedAt = now;
+      updated += 1;
+    } else {
+      const id = p.id && /^h_/.test(p.id) ? p.id : genId();
+      state.items.push({
+        id,
+        kind: p.kind,
+        content: p.content,
+        evidence: p.evidence,
+        importance: clamp(p.importance),
+        active: true,
+        createdAt: now,
+        updatedAt: now,
+      });
+      created += 1;
+    }
+  }
+
+  let pruned = 0;
+  if (options.prune) {
+    const before = state.items.length;
+    state.items = state.items.filter((i) => {
+      if (!i.active) return true; // inactive items are never touched by durable I/O
+      if (!preActiveIds.has(i.id)) return true; // created during this import
+      if (fileIds.has(i.id)) return true; // present in the file
+      return false; // was active before, absent from the file → drop
+    });
+    pruned = before - state.items.length;
+  }
+
+  version += 1;
+  persist(state, version);
+  return { imported: parsed.length, created, updated, pruned, missingFile: false };
+}