Mark the request's end on its last block, so a 1h block TTL does not clash with a top-level 5m mark

[0xTomDaniel]

With a proxy that gives Anthropic block cache marks a 1h TTL (ours is a self-hosted CLIProxyAPI in front of Anthropic), every request without a status line fails, which is every compactor call:

400 Top-level cache_control has ttl='5m' but the target block already has cache_control with ttl='1h'. When both are specified on the same block, they must have matching TTLs.

cachePayload ends such a request with top-level automatic caching (payload.cache_control), whose TTL defaults to 5m and which lands on the last block. Once the proxy has turned the block marks into 1h marks, Anthropic refuses the pair. Requests with a status line already avoid this, because their turn mark is a block mark.

Fix: in src/cache.ts, the end-of-request mark goes on the last block itself, the same way the turn mark does: the last block of the last non-system message, skipping thinking blocks. Nothing is written at the top level any more. It is still four marks (three on the view, one at the end), Anthropic's limit, and a cached prefix ends in the same place, so caching works as before. When a proxy rewrites TTLs, all four now carry the same one.

Tests (test/memory.test.ts):

• a compactor-shaped request (view, instruction, an assistant reply with thinking, a string follow-up, Pi's empty settings message) gets no top-level mark and an end mark on the last user block; with every mark rewritten to 1h, nothing is left at a default 5m
• a request ending in a thinking block marks the text block before it
• the existing view-marks test now expects the end mark on the last block instead of payload.cache_control

Checked live through the proxy with claude-sonnet-5-5 and max_tokens: 1: the old shape (top-level plus a block mark) returns the 400 above; the payload cachePayload now builds returns 200, with cache_creation.ephemeral_1h_input_tokens showing the proxy's 1h TTL applied to the block marks.

npm run check passes; npm test: 289 pass, 1 skipped (the Windows-only lock test).

[Copilot]

🟢 Approval recommended

The focused change has regression coverage for the affected request shapes, with no unresolved issues identified.

0 open findings

What changed in this PR

Moves end-of-request caching from the top level to an eligible message block, avoiding TTL conflicts when a proxy applies 1-hour block caching.

Changes:

• Preserves turn-mark behavior while adding block-level end marks.
• Adds regression coverage for compactor-shaped requests and trailing thinking blocks.

File	Description
test/​memory.test.ts	Verifies block-level marks and absence of top-level caching.
src/​cache.ts	Places the end mark on the last eligible non-system message block.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.