@nanhara/hara 0.106.0 → 0.107.0

package/CHANGELOG.md

@@ -5,6 +5,18 @@ All notable changes to `@nanhara/hara`.
 > Versioning (pre-1.0, SemVer-style): the **minor** (middle) number bumps for a **new feature**; the
 > **patch** (last) number bumps for **optimizations/fixes of existing features**.
 
+## 0.107.0 — interjection triage: the model is the scheduler, the todo list is the queue
+
+- **Mid-task messages get triaged, not blindly folded in.** Typing while hara works always reached
+  the model between tool calls (type-ahead steering) — but nothing told it HOW to handle the
+  interjection. Now every mid-task message carries a triage contract, backed by a standing policy in
+  the system prompt: a **refinement** folds into the current task immediately; a **new independent
+  task** goes onto the todo queue (`todo_write`, one-line acknowledgment, current work continues);
+  something **urgent** — a bug, "stop", "this first" — finishes the current step safely (no half-done
+  edits), re-plans the queue, and switches immediately. Same architecture codex and Claude Code
+  landed on: no engine-level priority scheduler — classification is exactly what the model is best
+  at, and the todo list (with its live panel + attention refresh) is the task queue.
+
 ## 0.106.0 — gateway session hygiene
 
 - **Idle chats auto-rotate to a fresh thread.** A WeChat/Feishu chat is one endless surface — days-old

package/dist/agent/loop.js

@@ -59,7 +59,9 @@ you can do. When analyzing a project, start wide in ONE batch — manifest (pack
 pyproject.toml / go.mod), README, build/CI config — then chase only what the task needs with narrow
 grep/glob; don't read whole large files when a targeted search answers the question. For broad,
 open-ended exploration (more than ~3 searches), spawn \`agent\` sub-agents — several in one response for
-independent questions (role "explore") — each returns conclusions, not dumps. For a multi-step task, call \`todo_write\` to plan a short checklist and keep it updated as
+independent questions (role "explore") — each returns conclusions, not dumps. Messages the user sends
+mid-task arrive marked as interjections — triage them (refine current / queue as todo / urgent-switch)
+instead of blindly folding everything into the current task; the todo list is your task queue. For a multi-step task, call \`todo_write\` to plan a short checklist and keep it updated as
 you go (one item in_progress at a time) — skip it for trivial one-step tasks. You have a persistent
 memory: use memory_search before answering about prior decisions,
 conventions, or the user's preferences, and memory_write to proactively save durable facts you learn.

package/dist/agent/reminders.js

@@ -30,6 +30,15 @@ export function wrapReminders(items) {
 /** How many tool rounds a checklist may sit untouched (with unfinished items) before the model gets an
  *  attention refresh. Reset on every todo_write; re-arms after firing so it nags at most once per N. */
 export const TODO_STALE_ROUNDS = 5;
+/** Prefix for a message the user sent MID-TASK (type-ahead steering). Carries the triage contract
+ *  inline (self-contained even for role-overridden runs): the model — not the engine — is the
+ *  scheduler, and the todo list is the task queue (codex/Claude-Code's model too: neither ships an
+ *  engine-level priority scheduler; classification is exactly what the LLM is best at). */
+export const INTERJECT_PREFIX = "[Sent while you were working on the above — TRIAGE before continuing: " +
+    "a refinement/correction of the current task → fold it in now; " +
+    "a NEW independent task → todo_write it onto the queue, acknowledge in one line, continue the current task; " +
+    "URGENT (a bug, \"stop\", \"this first\") → finish the current step safely (no half-done edits), " +
+    "todo_write the re-plan (current task → pending, this → in_progress), and switch to it immediately.]";
 /** Parallel fan-outs at/above this size get a synthesis nudge (CC's KN5 synthesizer, hara-shaped:
  *  instead of a dedicated merger agent, the MAIN model is reminded to merge before acting). */
 export const SYNTHESIS_MIN_AGENTS = 3;

package/dist/index.js

@@ -26,6 +26,7 @@ import { loadPermissionRules, scaffoldPermissions, globalPermissionsPath, projec
 import { routingProvider } from "./agent/route.js";
 import { shouldAutoCompact, COMPACT_SYSTEM, buildFileRestore } from "./agent/compact.js";
 import { recentTouched } from "./agent/touched.js";
+import { INTERJECT_PREFIX } from "./agent/reminders.js";
 import { checkForUpdate } from "./update-check.js";
 import { formatContextReport } from "./agent/context-report.js";
 import { userTurnPreviews, rewindTo } from "./agent/rewind.js";
@@ -3054,7 +3055,7 @@ program.action(async (opts) => {
                         const attach = !r2.skip && r2.attach?.length ? r2.attach : undefined;
                         if (!body.trim() && !attach)
                             continue; // image-only message whose image was skipped → nothing to add
-                        out.push({ role: "user", content: `[I sent this while you were working on the above]\n\n${body}`, ...(attach ? { images: attach } : {}) });
+                        out.push({ role: "user", content: `${INTERJECT_PREFIX}\n\n${body}`, ...(attach ? { images: attach } : {}) });
                     }
                     return out;
                 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nanhara/hara",
-  "version": "0.106.0",
+  "version": "0.107.0",
   "description": "hara — a coding agent CLI that runs like an engineering org.",
   "bin": {
     "hara": "dist/index.js"