Multilingual Triggers

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.

   No checkpoint. No user selection. The skill runs and completes without interruption.

If user explicitly says "compound this" / "document this" / "保存" / "记录": → Proceed directly to Full Mode (Phase 0.5 onwards), bypassing Auto-Draft tiering.

If user explicitly says "skip" / "不要记" / "不用记录": → Discard silently, no save, no notification.

Full Mode (Manual or Selected)

When user chooses "Full" or invoked with "compound this" / "document this":

Phase 0.5: Auto Memory Scan

Run memory_search with query matching the problem topic. If relevant entries found, prepare labeled excerpt block for Phase 1 subagents.

Phase 1: Parallel Research

Critical: subagents return TEXT DATA only. Orchestrator writes files.

🛑 STOP: Based on the problem description, make a preliminary determination of track type:

• Unexpected failure / error / abnormal behavior → bug
• Working pattern / best practice / design decision → knowledge
• Cannot determine → mark as unknown; let the Context Analyzer subagent make the final call

After recording the preliminary track determination, launch the following subagents:

Launch these subagents in parallel:

1. Context Analyzer

• Determines track: bug (unexpected failure) or knowledge (pattern/guidance)
• Identifies: problem_type, component, category
• Suggests filename: YYYY-MM-DD-[sanitized-slug].md
• Returns: YAML frontmatter skeleton, category, track

2. Solution Extractor

• Adapts output structure based on track:

Bug track sections:

• Problem: 1-2 sentence description
• Symptoms: Observable symptoms
• What Didn't Work: Failed attempts and why
• Solution: Actual fix with examples (before/after)
• Why This Works: Root cause explanation
• Prevention: Concrete strategies, test cases, lint rules

Knowledge track sections:

• Context: Situation that prompted guidance
• Guidance: Practice/pattern with examples
• Why This Matters: Rationale and impact
• When to Apply: Conditions where this applies
• Examples: Concrete before/after or usage

3. Related Docs Finder

• Searches memory/solutions/ for related documentation
• Uses grep/ripgrep for pre-filtering on title:, tags:, problem_type:
• Assesses overlap: High (same problem, update existing) / Moderate (related) / Low (distinct)
• Returns: links, relationships, refresh candidates

4. Session History (optional)

Run memory_search with tight query (default: 7 days). Structure findings as:

• What was tried before
• What didn't work
• Key decisions

Phase 2: Assembly & Write

WAIT for all Phase 1 subagents to complete.

1. Collect all Phase 1 text results
2. Check overlap assessment:

   Overlap	Action
   High	Update existing doc. Add last_updated: YYYY-MM-DD.
   Moderate/Low/None	Create new doc.

3. Assemble complete markdown
4. Validate YAML frontmatter
5. Write to memory/solutions/YYYY-MM-DD-[slug].md

Phase 3: Discoverability Check

Check whether AGENTS.md / MEMORY.md references memory/solutions/:

1. Read existing instruction files
2. Assess whether an agent would learn:
   • That a searchable knowledge store exists
   • Its structure (YAML frontmatter, category organization)
   • When to search it
3. If not met → draft smallest addition to communicate these three things

Interactive mode: Show proposed change, ask consent before editing. Headless mode: Apply edit silently, surface in report.

Phase 4: Project Integration (if .planning/ exists)

If memory/.planning/ exists:

• Append a summary entry to .planning/debug/knowledge-base.md
• Format: - YYYY-MM-DD: [title] (compound_score: X, file: solutions/...)
• This links compound knowledge to active project tracking

YAML Frontmatter Template

---
compound_score: [1-125]
date: YYYY-MM-DD
category: [tooling | workflow | database | api | frontend | backend | security | performance | unknown]
problem_type: [brief-classification]
track: [bug | knowledge]
tags: [tag1, tag2]
status: [active | resolved | deprecated]
---

Success Output

✓ Documentation complete

File: memory/solutions/YYYY-MM-DD-[slug].md  (created | updated)
Track: <bug | knowledge>
Category: <category>
compound_score: <score>
Overlap: <none | low | moderate | high — existing doc updated>
Knowledge-base: <not updated | appended to .planning/debug/knowledge-base.md>

The Compounding Philosophy

Build → Test → Find Issue → Research → Improve → Document → Validate → Deploy
    ↑                                                                      ↓
    └──────────────────────────────────────────────────────────────────────┘

Each unit of work should make subsequent units easier — not harder.

Error Handling

This skill follows the AGENTS/error-handling.md standard.

Common Error Codes

Escalation Conditions

• Same error_code fails 2 times in a row → force escalate on the 3rd
• retry_count reaches max_retries → escalate
• Non-recoverable errors (PERMISSION_DENIED, SECURITY_BLOCKED, etc.) → escalate immediately

Fallback Chain

If-Then Failure Mode Fallbacks

Anti-Pattern Blacklist

The following anti-patterns are strictly forbidden during C31-compound execution:

1. ❌ Never record unverified solutions
   • User says "should be fine" but hasn't actually tested it → do not record; mark as pending_verification
   • Must see "solved" / "verified" or equivalent confirmation before entering draft flow
2. ❌ Never skip user confirmation in Auto-Draft mode
   • Even a draft must be shown to the user with [1]-[4] options
   • Silent saving is forbidden; bypassing pre-select recommendation logic is forbidden
3. ❌ Never modify the original SKILL.md file content
   • compound only writes to memory/solutions/; it does not modify skills/C31-compound/SKILL.md
   • If a bug is found in the skill itself → go through the normal fix process; do not patch it via compound
4. ❌ Never record "solutions" from pure casual chat / emotional expressions
   • "I understand" / "I see" / "Thanks" → do not trigger compound
   • There must be a reusable technical / process / cognitive output
5. ❌ Never forcefully save when user explicitly says "don't record / skip"
   • User selects [skip] → immediately discard draft; do not write to any file
   • User says "stop recording" / "no need to record" → equivalent to [skip]; do not ask again or try to retain

Blacklist Violation Consequences

Auto-Invoke

<auto_invoke> <trigger_phrases>

• "that worked"
• "it's fixed"
• "working now"
• "problem solved"
• "solved"
• "fixed"
• "root cause"
• "解决了"
• "搞定了"
• "原来是因为"
• "根因是"
• "终于找到原因"
• "原来要这样"
• "没想到是"
• "错在" </trigger_phrases> <manual_override> Invoke with: "compound this" or "document this solution" or "记录" </manual_override> </auto_invoke>