fix: L2 policy induction crystallizes conversational acts into dead skills

[smoryan]

Pre-submission checklist

• I have searched existing issues and this hasn't been mentioned before
• I have read the project documentation and confirmed this issue doesn't already exist
• This issue is specific to MemOS and not a general software issue

Bug Description

The L2 policy induction prompt (apps/memos-local-plugin/core/llm/prompts/l2-induction.ts, v2) has no guidance against conversational acts as policy actions. When a trace cluster's core behaviour is dialogue — asking a clarifying question, requesting confirmation, notifying, or reporting status — the induced policy captures that dialogue act as its action.

Such policies later crystallize into skills that no agent can meaningfully invoke: they are not procedures, just one-off dialogue behaviours. In a 2026-08 audit of our local deployment's skill pool, these conversational-act skills were a recurring failure mode — always retrieved as candidates, never usable.

How to Reproduce

1. Run an agent session where a trace cluster's dominant behaviour is asking the user for confirmation before acting
2. Let L2 induction run on that cluster (prompt l2.induction v2)
3. Inspect the induced policy: action is e.g. "ask the user to confirm the skill name before viewing it" — a conversational act, not a reusable procedure
4. The policy later qualifies for skill crystallization, producing a dead skill

Environment

• memos-local-plugin: 2.0.18 (apps/memos-local-plugin)
• Node.js: 26
• OS: macOS

Willingness to Implement

• I'm willing to implement this myself

[Memtensor-AI]

🤖 AutoDev has picked up this issue and started working on it.

Task ID: 8aa945e8b611eae0
Working branch: bugfix/autodev-2318-20260902004642055
Target branch: latest dev* branch
Workflow: opsp (analysis → coding → testing → PR)

I will post the PR link here once done. If I need more information, I will ask in the comments.

[Memtensor-AI]

✅ AutoDev task 8aa945e8b611eae0 completed.

Summary: Fixes #2318 — L2 induction prompt was silent on conversational-act "actions", so trace clusters whose dominant behaviour was dialogue (ask / confirm / notify / report status) produced policies whose action was that dialogue act, later crystallising into dead skills that no agent could invoke.

Change: extended L2_INDUCTION_PROMPT (apps/memos-local-plugin/core/llm/prompts/l2-induction.ts) with a second boundary subsection symmetric to the existing L3-drift guard. It enumerates the dialogue-verb classes to reject as ACTION templates, shows a wrong/right contrast on the audit's canonical skill-viewing scenario, and offers an explicit {"abstain":true,"reason":"..."} escape hatch for dialogue-only clusters. The existing induce.ts validator already treats a missing-title response as a soft failure — the correct outcome for an abstain. Prompt version bumped 2→3; the 11 op-key fixtures across 4 vitest test files were migrated l2.l2.induction.v2 → l2.l2.induction.v3 in the same commit.

Verification: TDD-first (new red test in tests/unit/llm/prompts.test.ts pins version=3 + the guardrail language + abstain contract + still-present L3 guard, went green after the prompt edit). Full plugin vitest suite: 183 files, 1558 passed / 2 skipped / 0 failed. Type-check (tsc --noEmit) clean. No runtime code path changes — the fix is prompt-only, cutting the failure at its origin rather than mopping up downstream.

Scope: 6 files touched (1 prompt + 5 test files). .ai-tasks/ and openspec/changes/ remain excluded from the PR per the workflow contract; both were archived to the sibling memos-autodev-specs repo (2026-09-02-2318-pre-submission-checklist/).

Base branch: main
Branch: bugfix/autodev-2318-20260902004642055
Commit: ad220ef
PR: #2323
Assigned to: @whipser030
Reviewers: @whipser030, @hijzy