[Bug] Prompt cache drops to 7,085-token system floor across runs of 43–58 calls

Bug Description
Title: [BUG] Prompt cache collapses to the system floor (7,085 tokens) on every request for 43 and 58 consecutive calls; no images, no compaction, 1h TTL (2.1.285, macOS) ## Environment - Claude Code 2.1.285 on macOS. The session started on 2.1.284 and auto-updated mid-session. - Model claude-opus-5-5, effort medium - Subscription (OAuth) login, promptCacheTtl 1h: every write is ephemeral_1h_input_tokens - One interactive CLI session, not resumed. There were 0 compactions and 7 background general-purpose subagents. - Several plugins add PreToolUse/PostToolUse hooks. Their hook_additional_context is appended as attachments. ## What's wrong Across 359 main-thread API calls, 110 are full cache misses. On each of them, cache_read_input_tokens falls to exactly 7,085, which is the system-prompt floor. cache_creation_input_tokens then re-writes the whole conversation (up to ~500k tokens). Most of the misses do not stand alone. They come in two long runs, and every call in a run misses, even when the gap between calls is 8–60 seconds: | Run | Calls | Time (UTC, 2026-09-30) | Re-written per call | |---|---|---|---| | A | 43 consecutive | 02:52:54 – 04:45:23 | ~285k–330k | | B | 58 consecutive | 05:24:21 – 06:51:47 | ~438k–500k | Impact: 42.4M of the session's 42.9M cache-write tokens (99%) are these re-writes. If the cache had held, the same session would have written ~0.4M. ## Evidence that the client mutates the prefix Let promptTotal = input_tokens + cache_read_input_tokens + cache_creation_input_tokens. For an append-only conversation, this value can only grow. Inside the runs it shrinks between consecutive calls: | Call | promptTotal before → after | Delta | |---|---|---| | #162 → #163 | 309,114 → 292,224 | −16,890 | | #178 → #179 | 328,336 → 310,625 | −17,711 | | #191 → #192 | 349,562 → 330,714 | −18,848 | | #203 → #204 | 340,994 → 334,409 | −6,585 | Content of roughly 17k tokens appears to be added to and removed from the request between calls. Nothing is removed from the transcript at those points. ## Causes ruled out - TTL: the 1h tier is in use, and the gaps inside the runs are mostly seconds. - Compaction: none happened. - Images / image eviction (#90716): only 2 images entered the context in the whole session. - Persisting cd (#91706): every Bash cd result carries "Shell cwd was reset", and the transcript cwd never changes. - Resume: the session was never resumed. ## Isolated misses that do have a visible trigger (for completeness) - Mid-session auto-update from 2.1.284 to 2.1.285: one miss. - First ToolSearch that loaded Playwright MCP tools: one miss, same signature as #91514. - Usage-limit wait followed by /login, which brought MCP deltas: one miss. ## What started and ended the runs - Run A: starts on a plain user text message with no tool change visible in the transcript. It ends ~2h later, again with no visible event. - Run B: starts mid tool loop, right after a Bash tool_result. It ends only after the usage-limit wait and /login. ## Request IDs Good call → first miss: - req_011CfYmX31zV9TtuGTAg87BC → req_011CfYmgPERCQrhmkMkhJMhw (01:39Z) - req_011CfZ4nCoFzxptRHv2Cnkij → start of run B (05:24:21Z) Misses inside the runs: - req_011CfYsFGapPKoqFX33ueHzt: first call of run A - req_011CfYxFADfUxVZXJzgYTmkU, req_011CfYxFq7VY1NsdWgxfPctL: consecutive misses in run A with the promptTotal drop above - req_011CfYph88wibV25nToCZbtV: the ToolSearch-triggered miss ## Related #91514 (collapse to the system+tools floor after ToolSearch / Skill / tool_result), #90716, #91706, #91151. None of them explains a run of consecutive misses with no images and no cd. I can share the per-call usage table (timestamps, read/write/input) if useful. I will not share the raw transcript, because it contains private code.

Environment Info

• Platform: darwin
• Terminal: Orca
• Version: 2.1.286
• Feedback ID: 847e96d8-3d13-474f-9963-9101757808ed

Errors

[]