Cache

/ar:cache — Cache Miss / Compaction Protection Skill

A PreToolUse gate that blocks commands when the prompt cache is effectively cold — saving tokens by preventing full-rate re-reads of context that has just fallen out of the 5-minute prompt-cache TTL. Off by default.

When to use this skill

Invoke when the user asks about any of:

• setting a cache-miss / cache-hit-ratio threshold
• protecting against compaction-triggered context loss
• /ar:cache, /ar:cache on, /ar:cache ok 5m, /ar:cache set …
• "how do I stop Claude from burning tokens after idle time"
• "what is the cache gate", "cache protection"

Harness note: this is a skill, while /ar:cache is a runtime autorun command. Claude Code can usually pass the slash form to hooks. Codex may intercept unknown slash commands before hooks see them, so use ar:cache or ar cache ... in Codex prompts when you need the runtime command.

1. Philosophy

• Off by default. No block fires unless the user runs /ar:cache on.
• Fail-open. If usage data is unavailable (old Claude Code version, Gemini CLI without cache-token fields, malformed JSONL), the gate allows.
• Reuse, don't reinvent. The gate uses the same ScopedAllow grants, parse_scope_args grammar, session_state filelock backend, and check_blocked_commands block UX that /ar:ok / /ar:no already use.
• Stay out of the statusline. The gate does not install or modify the user's statusLine setting. An optional one-line tap exists for users who want to feed their statusline's rich JSON into the gate themselves.

2. Commands

/ar:cache                       # show status
/ar:cache on [5m|1h|perm]       # enable
/ar:cache off [5m|1h|perm]      # disable (temporarily or until re-enabled)

/ar:cache set ratio 0.3         # block when cache hit ratio < 30%
/ar:cache set read 50k          # block when cache_read_tokens < 50,000
/ar:cache set age 10m           # block when last cache hit older than 10 min
/ar:cache set full 0.9          # block when total_input / window > 90%

/ar:cache ok 5m                 # override for 5 minutes
/ar:cache ok 3                  # override next 3 tool uses
/ar:cache ok perm               # override until axis clears / session end
/ar:cache no                    # clear outstanding overrides

Units: tokens accept 50000, 50,000, 50_000, 50k, .5M, 1.5M. Percent: 85% or bare decimal 0.85. Durations: 5m, 1h, 2h30m, perm.

3. Grammar decision: /ar:cache ok not /ar:ok cache

The two grammars are semantically distinct. /ar:ok <pattern> allows a command-string pattern; /ar:cache ok <scope> overrides the feature gate. Overloading /ar:ok cache would silently collide with any user command named cache. Decision in plan §6.5.1 is: no /ar:ok cache alias in v1. A sigil-based alias (e.g. /ar:ok @cache) is backwards-compatible to add later if users ask.

4. How the decision flows (PreToolUse)

 tool use
     │
     ▼
 FeatureToggle("cache").is_enabled()  ──► false → ALLOW (default)
     │ true
     ▼
 CacheGuard._read_usage:
   1) cache/last_usage memo (2 s TTL) — reuse on burst calls
   2) cache/statusline_snapshot       — opt-in fast path
   3) JSONL tail (64 KB reverse scan) — find last assistant message.usage
     │   none available → ALLOW (fail-open)
     ▼
 CacheThreshold axes:
   - cache_hit_ratio_min     (cache_read / (input + cache_read + cache_creation))
   - cache_read_tokens_min   (raw cache_read_input_tokens floor)
   - cache_age_max_seconds   (age of last assistant timestamp)
   - compaction_used_max     (total_input / context_window)
   - rate_limit_5h_max       (from statusline snapshot; rate-limit wording)
   - rate_limit_7d_max       (from statusline snapshot; rate-limit wording)
     │ none tripped → ALLOW
     ▼
 ScopedAllow override active? (/ar:cache ok …)
   yes → consume + ALLOW
   no  → BLOCK with copy-paste override hints

5. Cross-CLI matrix

Capability	Claude Code	Gemini CLI
hook stdin transcript_path	yes	yes
message.usage.cache_read_* in JSONL	yes	unknown schema; probe
Compaction event name	PreCompact / PostCompact / SessionStart(compact)	PreCompress (advisory) / SessionStart
cache_hit_ratio_min axis	active	fail-open (no data)
cache_read_tokens_min axis	active	fail-open
cache_age_max_seconds axis	active	active
compaction_used_max axis	active (with snapshot)	active when total tokens available
rate_limit_*_max axes	only with statusline tap	unavailable

/ar:cache status prints which axes are inactive for the current CLI.

6. Multiprocess safety

The gate uses session_state() — a single shared JSON file with prefixed keys and filelock + atomic tempfile rename. All writes are serialised; reads re-load inside the lock. Parallel hook invocations (e.g. rtk-spawned) inherit the same _PARALLEL_GRACE_SECONDS = 1.0 s window that ScopedAllow already uses. No new state file, no new lock.

7. Files

Path	Role
plugins/autorun/src/autorun/cache_guard.py	Single-file feature (~500 LOC).
plugins/autorun/tests/test_cache_guard.py	Full test matrix (parser → multiprocess).
plugins/autorun/commands/cache.md	/ar:cache command definition.
plugins/autorun/skills/cache/SKILL.md	This file.

8. Troubleshooting

• "Gate doesn't fire." Run /ar:cache status — is it enabled? Are any axes configured? Set at least one axis (/ar:cache set ratio 0.5) after enabling.
• "Blocked when I didn't want it." /ar:cache ok 5m or /ar:cache off.
• "Gemini says axes unavailable." Correct — Gemini CLI hooks do not surface cache-token fields. Use age axis instead; it works on both CLIs.
• "I want the rich statusline numbers." Add one line to your own statusline script: printf '%s' "$INPUT" | autorun --cache-snapshot >/dev/null 2>&1 &.

9. Plan reference

Full design: ~/.claude/plans/make-a-plan-to-sunny-sparkle.md. Key sections: §6.1 (reused magic classes), §6.5 (usage detection), §6.5.1 (grammar decision), §6.7 (block message), §6.8 (multiprocess), §7 (ASCII diagrams).