refactor(client): relocate turn_meta to tail of user message for prefix stability

[HUQIANTAO]

Summary

Move the <turn_meta> content block from position 0 to the last position in each user message's content array. This ensures the leading bytes of every user message stay stable across date boundaries, model route changes, and working set mutations, which directly improves DeepSeek's KV prefix cache hit rate.

Background

The <turn_meta> block contains volatile per-turn data:

• Current local date (chrono::Local::now().format("%Y-%m-%d"))
• Auto model route (when auto_model is enabled)
• Auto reasoning effort (when auto-selected)
• Working set summary (file paths observed this session)

When this block sits at content position 0, the leading bytes of every user message change at every date boundary. DeepSeek's KV prefix cache matches byte sequences from the start of each message; a change at position 0 invalidates the cache for the entire user message prefix, forcing a full re-prefill on the next turn.

What changed

The reorder is a single swap in user_text_message_with_turn_metadata_for_route:

Before:

content: vec![
    turn_metadata_block(...),  // volatile: date, model route, working set
    ContentBlock::Text { text, .. },  // stable: user input
]

After:

content: vec![
    ContentBlock::Text { text, .. },  // stable: user input (now at position 0)
    turn_metadata_block(...),  // volatile: date, model route, working set (now at tail)
]

Why this is safe

• The model does not depend on the position of metadata within a single user message. V4 processes all content blocks in a user message sequentially regardless of order.
• The messages_with_turn_metadata() function returns stored messages unchanged (it clones self.session.messages). The metadata is baked into messages at append time, not injected at request time. The reorder only affects the initial storage layout.
• Tool result messages are unaffected; they carry a single ToolResult content block and are never mixed with turn_meta.

Testing

• Updated 4 existing tests that previously asserted content.first() was the turn_meta block to use content.last() instead.
• Added new test user_message_turn_meta_is_appended_not_prepended that explicitly verifies the user text is at position 0 and turn_meta is at position 1.
• All 8 turn_metadata-related tests pass.

Cache impact

For sessions that span midnight or change model routes mid-conversation, the user message prefix bytes now remain stable. This means:

• Cross-date sessions no longer bust the user message prefix cache at the date boundary.
• Model route changes (auto_model switching between v4-pro and v4-flash) no longer invalidate the user message prefix.
• Working set mutations (new files observed) only affect the trailing metadata block, not the leading user input.

[greptile-apps]

HUQIANTAO has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[gemini-code-assist]

Code Review

This pull request swaps the order of content blocks in user messages so that the user text is placed first and the turn metadata block is placed last. This change prevents DeepSeek's KV prefix cache from being invalidated by volatile metadata changes (like dates). While the tests in the PR were updated to reflect this change, a review comment correctly identifies that an existing test (post_edit_hook_injects_diagnostics_message_before_next_request) was missed and will now fail because it still expects the metadata block to be at index 0.

[gemini-code-assist]

An existing test post_edit_hook_injects_diagnostics_message_before_next_request (around line 2956-2960 in crates/tui/src/core/engine/tests.rs) is not modified in this PR but will now fail because of this change.\n\nIn that test, it asserts that the first content block of the appended message starts with <turn_meta>\\n:\n```rust\n let last = engine.session.messages.last().expect("message appended");\n assert_eq!(last.role, "user");\n let meta = match &last.content[0] {\n crate::models::ContentBlock::Text { text, .. } => text.clone(),\n other => panic!("expected text block, got {other:?}"),\n };\n assert!(meta.starts_with("<turn_meta>\n"));\n```\nSince this PR swaps the order of the content blocks, last.content[0] is now the diagnostics text block, and last.content[1] (or last.content.last()) is the turn_meta block. This assertion will fail.\n\nPlease update that test to check last.content.last() instead of last.content[0] for the turn_meta block.

[carloslindhstrom-code]

An existing test post_edit_hook_injects_diagnostics_message_before_next_request (around line 2956-2960 in crates/tui/src/core/engine/tests.rs) is not modified in this PR but will now fail because of this change.\n\nIn that test, it asserts that the first content block of the appended message starts with <turn_meta>\\n:\n```rust\n let last = engine.session.messages.last().expect("message appended");\n assert_eq!(last.role, "user");\n let meta = match &last.content[0] {\n crate::models::ContentBlock::Text { text, .. } => text.clone(),\n other => panic!("expected text block, got {other:?}"),\n };\n assert!(meta.starts_with("<turn_meta>\n"));\n```\nSince this PR swaps the order of the content blocks, last.content[0] is now the diagnostics text block, and last.content[1] (or last.content.last()) is the turn_meta block. This assertion will fail.\n\nPlease update that test to check last.content.last() instead of last.content[0] for the turn_meta block.

[carloslindhstrom-code]

😃

[greptile-apps]

HUQIANTAO has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.