[yuanhao]

Follow-up to #347's review (finding 5): since 0.25.9 sends every text block of a tool result, tool_output_max_lines — applied per block — let an N-block result (an MCP server's) reach the model as N × the budget.

Change (context::block_budgets, used by truncate_tool_output_keyed):

• If the text blocks fit the budget together, nothing changes.
• Otherwise blocks shorter than an even share stay whole and the long ones split what is left (remainder to the earliest), each cut head + tail with its own marker — so per-block stash keys (-b{i}) and the loop's stash code are unchanged.
• A cut block keeps at least 5 lines (head, marker, tail): a result with more long blocks than the budget can mark goes over it by that much rather than losing lines without a marker.
• One-block results are unchanged. Idempotent (a cut block is exactly its budget; the floor case reassigns the same budgets), so truncated history is never rewritten — prefix cache intact.

Tests (src/context.rs): 20 × 200-line blocks fit 200 lines with a marker each; blocks that fit together are untouched; a short block stays whole between long ones, an image keeps its place and the keyed marker names k-b3; ten long blocks under a 20-line budget each keep 5 lines and a marker; every case truncated twice is unchanged.

Docs: docs/concepts/context-management.md, CHANGELOG ## Unreleased.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG

[yuanhao]

Review (code-reviewer, with a brute-force idempotence check over max_lines 0–12/20/25/MAX, 1–6 blocks, \r\n / trailing-newline / blank-line text, images between): no code issues; keyed and unkeyed passes mark the same blocks. Fixed in 133a8f1: the CHANGELOG now says multi-block results stored by earlier versions are re-cut once at their first level-1 compaction (one prefix-cache rewrite); floor wording limited to budgets of 5 or more; block_budgets doc says a cut block is at most its budget.

[yuanhao]

Test and comment reviews addressed in 75b8e6b: tests for the remainder rule ([67, 67, 66]), freed slack ([10, 60, 80]), an override on a many-block result, short blocks past the floor, and line counts on the loop's append path ([10, 10]); docs now state the overshoot exactly (even share under 5 lines, short blocks included; no sharing below a budget of 5) and stale per-block wording is updated.