菜单
Interpret session token statistics and recommend the next move — /compact, delegate to a subagent, switch to a smaller model, or stop and persist learnings. Use after `/sustain:status`, when the session feels long, or when context-window pressure is showing up in degraded responses.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Walk through stale or conflicting memory entries and routing-table drift. Use when the user runs /sustain:audit, asks "is my memory store healthy?", or before a major release of personal rules.
Search persistent memory across mempalace MCP, claude-mem MCP, and the structured filesystem fallback. Use when the user asks "how did we solve X before?", "what did we decide about Y?", or anything that implies recalling prior-conversation context.
Decide whether and where to persist a learning. Use at phase boundaries when something seems worth long-term memory — a rule, a non-obvious decision, an external resource, or a project-state change. Picks the right wing (feedback / project / reference / user) and routes through the user's active memory backend.
Run the 6-question phase-end checklist when a logical phase completes (after a feature, before a context switch, or when the Stop hook prompts you). Use to surface skipped Iron Rules, missed memory writes, and overdue /compact opportunities.
Establish Iron Rules (3+1) and R1–R5 token-saving rules at session start or whenever the rules feel forgotten. Use when user asks about token efficiency, when rules need re-grounding, or proactively at the start of complex multi-turn work.
| name | token-budget-coach |
| description | Interpret session token statistics and recommend the next move — /compact, delegate to a subagent, switch to a smaller model, or stop and persist learnings. Use after `/sustain:status`, when the session feels long, or when context-window pressure is showing up in degraded responses. |
Use this skill when there's a token-usage signal worth acting on. Sources of signal:
/sustain:status output (session token tally + cache hit rate).tokens: line.The session JSONL records four token streams:
| Field | What it is | What to optimize |
|---|---|---|
input_tokens | Fresh prompt tokens (not cached). | Fewer is better. Driven by file reads and tool-result echoes. |
cache_creation_input_tokens | New cache writes. | Spikes early; should plateau. |
cache_read_input_tokens | Cached prompt reuse. | Higher is better — the prompt cache is paying off. |
output_tokens | Generated tokens. | Bound by R4 (concise responses); large output usually means a long search/summary should have been a subagent. |
Cache hit rate = cache_read / (cache_read + cache_creation + input). Healthy long sessions: > 70%.
Is cache hit rate < 50% AND session is long?
├─ Probably context churn (lots of new file reads / re-searches).
│ → Fix R5.1: re-read what's already in context instead of re-searching.
│ → Apply R1.1/R1.3: offset+limit reads, not full files.
Is output_tokens of any single turn > ~5K?
├─ You echoed a file or wrote a long explanation.
│ → R4.1/R4.3 — be concise. If it was a search summary, delegate to a subagent next time.
Is input_tokens climbing turn over turn (no plateau)?
├─ Context is filling with raw tool output.
│ → Time to /compact at the next phase boundary.
│ → Or dispatch the remaining work to a subagent (R3.1) so the bulk lives in *its* context.
Is the user's task long-running and we've just finished a phase?
├─ → Apply phase-self-check (the 6 questions).
│ → Especially Q5 (compact?) and Q6 (anything worth saving as memory?).
subagent_type: "Explore" for the next codebase question. Bulk discovery happens in their context, not yours.model: "haiku" for a simple lookup (R3.3)./compact at the next natural phase boundary (don't compact mid-task — Iron-1 friction).memory-write-router then /clear and resume in a fresh session with the search-bridge./compact reflexively every N turns. Compaction has a cost (it summarises, dropping detail). Use it when the data shows pressure, not on a clock.memory-write-router).