C31-compound

Multilingual Triggers

Language	Trigger phrases
EN	solved, fixed, working now, root cause, that worked, compound, document this, 解决了, 搞定了, 原来是因为, C一下
ZH	复利, 记录知识, 总结
JA	知識を記録する, コンパウンド, まとめる

Output language: Respond automatically in the user's conversation language.

C31-compound — Knowledge Compounding

Capture a recently solved problem as structured documentation so future occurrences take minutes instead of research cycles.

Preconditions

• Problem has been solved (not in-progress)
• Solution has been verified working
• Non-trivial problem (compound_score ≥ 15, or user insists)

Structural Routing (pstack/reflect, merged 2026-08-19)

Before writing the solution doc, route the learning:

1. Structural enforcement check: if this lesson is enforced more reliably by a lint rule, script, metadata flag, runtime check, or test than by a doc, propose the structural artifact instead (doc becomes secondary). Lessons living only in text re-decay; lessons in structure fire themselves.
2. Skill-edit branch: if the learning generalizes to future agent behavior (a skill's wording caused the dead end, or a missing rule would have prevented it), the primary artifact is a skill edit proposal — present it to the user and apply only after explicit approval. Skill changes affect every future session; never auto-apply.
3. One-off filter: skip when the episode is trivial, off-topic, or already covered by an existing skill the agent simply failed to follow. One-offs are not learnings; recording them dilutes the library.

Auto-Draft Mode (Default)

When triggered by auto-invoke phrases (not manual command), enter lightweight draft mode first.

Behavior:

1. Skip blocking questions
2. Extract from recent conversation context:
   • Problem: 1-2 sentence description
   • Symptoms: Observable symptoms
   • What Didn't Work: Failed attempts
   • Solution: Actual fix
3. Duplicate detection:
   • Use memory_search to query key words from the Problem description + core tool/filename from Solution
   • If a similar solution with score > 0.7 is found:
     • Emit a brief note: ⚠️ Similar record detected: {filename} (date)
     • Append options: [Overwrite] update existing file / [Append] add new case study / [Skip] do not save
     • If user says "Overwrite" → update existing file (append last_updated: YYYY-MM-DD, add new observations at end)
     • If user says "Append" → append a new case study block at end of existing file
     • If user says "Skip" or no response (30s) → discard current draft, do not save
     • Interrupt current Auto-Draft flow; next steps are user-driven
   • If no match or score ≤ 0.7 → continue to step 4
4. Auto-calculate compound_score = Severity × Frequency × Uniqueness × Scope
   • Severity: 1-5 (how bad was the problem)
   • Frequency: 1-5 (how likely to recur)
   • Uniqueness: 1-5 (how non-obvious was the solution)
   • Scope: 1-5 (impact range: single file → cross-skill → systemic)
     • 1 = within a single file / single skill
     • 2-3 = across files or across skills
     • 4-5 = systemic / framework-level / affects multiple workflows
5. Tiered auto-save:
   • compound_score ≥ 50 → trigger Full Mode async execution
     • Spawn background subagent to run Phases 0.5–4 (parallel research + full assembly)
     • Main conversation only emits: ✍️ High-value problem detected ({score} pts) — generating full compound doc in background...
     • Completes silently; no further interruption
   • 15 ≤ compound_score < 50 → auto-save lightweight version to memory/solutions/, silently
     • Emit brief note: ✍️ Auto-recorded — {filename} (compound_score: {score})
   • compound_score < 15 → skip silently. No notification.