self-improvement beforeResetNote injects visible message to user on /new — should use prependContext instead

[eightHundreds]

Problem

When selfImprovement.beforeResetNote is enabled (default: true), the command:new / command:reset hook injects a /note self-improvement (before reset): ... string via event.messages.push() (index.ts:3041-3050).

This message is visible to the end user in the conversation, and triggers an Unknown skill: note error because note is not a registered skill — the raw parameter text is then exposed verbatim:

Unknown skill: note
Args from unknown skill: note self-improvement (before reset):
  If anything was learned/corrected, log it now:
  .learnings/LEARNINGS.md (corrections/best practices)
  .learnings/ERRORS.md (failures/root causes)
  Distill reusable rules to AGENTS.md / SOUL.md / TOOLS.md.
  If reusable across tasks, extract a new skill from the learning.
  Then proceed with the new session.

Root Cause

event.messages.push() in command:new / command:reset hooks injects content as a user-visible conversation message, not as silent system context.

Suggested Fix

The same codebase already has a better pattern — the before_prompt_build event handler supports a prependContext return value that injects content into the system context without exposing it to the user (index.ts:3157-3163):

api.on("before_prompt_build", async (_event, ctx) => {
  return {
    prependContext: [
      "<self-improvement-reminder>",
      "If anything was learned/corrected in the previous session, log it now:",
      "  - .learnings/LEARNINGS.md (corrections/best practices)",
      "  - .learnings/ERRORS.md (failures/root causes)",
      "Distill reusable rules to AGENTS.md / SOUL.md / TOOLS.md.",
      "</self-improvement-reminder>",
    ].join("\n"),
  };
});

Alternatively, the agent:bootstrap hook with bootstrapFiles.push() (already used at index.ts:2988) could also work.

Available injection points in the codebase

Method	Code location	Visible to user?
api.registerHook("command:new") + event.messages.push()	index.ts:3060	Yes (current, problematic)
api.registerHook("agent:bootstrap") + bootstrapFiles.push()	index.ts:2959	No
api.on("before_prompt_build") + return { prependContext }	index.ts:3143, 3170	No

Since agent:bootstrap already injects SELF_IMPROVEMENT_REMINDER.md as a virtual bootstrap file, the command:new event.messages.push() injection may be redundant and could simply be removed.

Environment

• memory-lancedb-pro v1.1.0-beta.10

[eightHundreds]

Additional analysis from codebase

After deeper investigation, the self-improvement feature actually has two injection points, not just one:

1. agent:bootstrap hook (index.ts:2959) — silent, works correctly

Injects SELF_IMPROVEMENT_REMINDER.md as a virtual bootstrap file via bootstrapFiles.push(). This is not visible to the user. ✅

2. command:new / command:reset hook (index.ts:3060-3064) — visible, problematic

Injects /note self-improvement ... string via event.messages.push(). This is visible to the user and triggers Unknown skill: note. ❌

Available alternatives in the same codebase

The codebase already uses before_prompt_build + prependContext for memory-reflection (index.ts:3143-3163, 3170-3218). This pattern injects content into system context silently.

Suggested fix: Replace the command:new event.messages.push() with a before_prompt_build handler that returns { prependContext } containing the self-improvement reminder. This way the reminder enters the LLM context on the first turn after /new without being visible to the user.

Alternatively, since agent:bootstrap already handles the bootstrap case, the command:new injection may be redundant and could simply be removed.