[BUG] Hidden `isMeta` system-reminder causes 200k+ cache_creation burst mid-session (Opus 4.6)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Every Opus 4.6 turn that satisfies a specific internal gate pushes a hidden meta-message onto the messages array. On certain turns this push correlates with the prompt cache being invalidated, so cache_creation_input_tokens spikes from the normal ~1–3k range to 200k–282k per turn. The user prompt that triggers a burst is often a single token ("ok", "1"); the cost is hidden in the transcript and invisible in the UI.

In a 183-turn session on Opus 4.6, 9 such burst turns accounted for ~56% of total cache_creation (1.5M of 2.66M). The reminder content is never shown to the user and the injection site is the model's messages array, not the system-prompt block — so even when the string itself is stable, the breakpoint position moves per turn and invalidates the cached prefix.

Binary evidence

// constant
am7 = "<system-reminder>Respond with just the action or changes and without a thinking block, unless this is a redesign or requires fresh reasoning.</system-reminder>"

// dynamic boundary marker used elsewhere to split system prompt into
// cached (global) vs. non-cached (null) blocks
WMH = "__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__"

// injection site (paraphrased from minified source)
mainLoopModel && $?.some(U => U.type === "assistant") &&
  F.messages.push(n_({ content: am7, isMeta: true }));
return F;

Injection gate (traced from binary)

All of the following must hold for the push to occur:

• input mode is prompt
• no external-loading path
• no customSystemPrompt set
• thinking is not disabled
• Pz8(mainLoopModel) returns true
• at least one prior assistant message exists

Pz8 returns true only when:

• the model belongs to the opus-4-6 family, and
• D_().clientDataCache?.loud_sugary_rock === "true" (remote experiment gate)

What Should Happen?

The reminder string is a constant per product version; it should either:

• Live inside the cacheable global system-prompt block (before __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__), so it cannot invalidate the prefix on subsequent turns, or
• Be placed after the last cache_control breakpoint so its turn-to-turn presence doesn't move the cached prefix, or
• Be controllable with an environment variable (e.g., CLAUDE_CODE_DISABLE_ACTION_REMINDER=1) so quota-sensitive users can opt out.

Short trivial user prompts should never cost 200k+ cache_creation tokens.

Error Messages/Logs

There is no error message. The only visible signal is the `usage.cache_creation_input_tokens` field in the JSONL transcript at `~/.claude/projects/<project>/<sid>.jsonl`.

### Observed real session (Opus 4.6, 183 turns)

| Metric | Value |
|---|---|
| Total assistant turns | 183 |
| Normal turn `cache_creation` | 1k – 3k |
| Burst turn `cache_creation` | 200k – 282k |
| Burst turns | 9 |
| Share of total `cache_creation` from burst turns | ~56% |
| Total `cache_creation` | 2.66M |
| Total `cache_read` (normal caching working) | 24.6M |

Bursts were not user-initiated (no `/clear`, no `/compact`, no large tool results). They appeared at internal state transitions.

### Controlled repro (see "Steps to Reproduce")

Switching output style while in an Opus 4.6 session produces a one-shot burst within 2 turns of the switch, matching the pattern observed in normal sessions:

| Step | outputStyle | User prompt | `cache_creation` |
|---|---|---|---|
| baseline (normal work) | Explanatory | — | 932 – 1,417 |
| `/config` → default | default | `ok` | 589 – 824 |
| next turn | default | `네` | **335,411** |
| next turn | default | `1` | 278 – 354 |
| `/config` → Explanatory | Explanatory | `2` | 365 |
| next turn | Explanatory | `3` | 432 |

The `"네"` turn is two bytes; the 335k spike is consistent with the cached prefix up to the last breakpoint being re-created.

Steps to Reproduce

1. Start a fresh session on claude-opus-4-6 with thinking enabled. Do not set --system-prompt.
2. Run a few normal turns to ensure at least one prior assistant message exists.
3. Issue /config and switch output style from the current value to default (or from default to any named style). Send a short prompt such as ok.
4. Send another short prompt such as 1 or 네.
5. Open the session JSONL at ~/.claude/projects/<project>/<sid>.jsonl. For each type:"assistant" entry, inspect message.usage.cache_creation_input_tokens.
6. Expect the second turn after the style change to show a cache_creation spike of ~200k+ with cache_read_input_tokens = 0, while surrounding turns stay under 2k.

Alternative (non-deliberate) repro: running any long Opus 4.6 session will exhibit the same pattern at internal state transitions (e.g., when the loud_sugary_rock gate flips or on the first turn that gains a prior assistant message).

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.109

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Workarounds verified

• --thinking disabled short-circuits one of the gate conditions. Empirically reduces burst frequency in the same session; needs broader confirmation.
• Switching the model family away from opus-4-6 appears to fail Pz8, avoiding the push.
• --system-prompt bypasses the gate but is too heavy-handed for normal use.
• --append-system-prompt does not help (verified).
• --exclude-dynamic-system-prompt-sections does not help (verified).
• No settings.json key appears to control this path.

Suggested fix

Preferred: move the reminder into the static cacheable system-prompt block (Option A above). The reminder content is constant per version and does not need to be a per-turn dynamic isMeta push.

Backup: expose an env-var opt-out. Users on quota-sensitive plans can mitigate immediately while the caching fix is being designed.

Additional notes

• The reminder content itself is reasonable guidance; the concern is the insertion mechanism, not the text.
• The binary already has a Xz8 / WMH two-tier (global vs. null) cache-scope mechanism; routing am7 through the global tier would resolve this cleanly.
• Happy to provide redacted transcript snippets, binary offsets, and additional JSONL samples if useful — kept out of the public body to avoid leaking private content.