feat(chat): add summary budget discipline for compaction

[Cstewart-HC]

Summary

Enforce hard budget on LLM-generated compaction summaries via new compress_summary() function in crates/chat/src/lib.rs.

Without this, auto-compact can produce arbitrarily long summaries that overflow the context window on the next turn — defeating the purpose of compaction.

Changes

• crates/chat/src/lib.rs (+269 lines) — new compress_summary() function:
  • Max 1,200 characters total
  • Max 24 lines
  • Max 160 characters per line
  • Case-insensitive line deduplication
  • Priority-based line dropping: headers > bullets > plain text
  • Omission notice appended when lines are dropped
• Applied to both compact() method and standalone compact_session() function
• 8 unit tests covering all budget enforcement paths

Testing

running 8 tests
test tests::compress_summary_empty_input ... ok
test tests::compress_summary_long_line_truncation ... ok
test tests::compress_summary_deduplication ... ok
test tests::compress_summary_single_very_long_line ... ok
test tests::compress_summary_under_budget_returns_unchanged ... ok
test tests::compress_summary_over_char_limit ... ok
test tests::compress_summary_header_preservation ... ok
test tests::compress_summary_over_line_count ... ok

Also verified: cargo check, cargo clippy, just format-check — all clean.

Checklist

• Code follows project Rust style (no .unwrap()/.expect(), iterators over loops, #[must_use] on pure fns)
• Tests included with high coverage
• Single-file, minimal scope change
• Commit message follows conventional commits format

[greptile-apps]

Greptile Summary

This PR adds compress_summary() to enforce a hard budget (1,200 chars, 24 lines, 160 chars/line) on LLM-generated compaction summaries, applying it in both compact() and compact_session(). The logic is well-structured and the 8 unit tests cover the main paths, but two minor inconsistencies were found beyond the already-flagged off-by-one and notice-undercount issues: (1) the SUMMARY_MAX_LINES doc comment says "excluding omission notice" while the drop loop actually counts the notice toward the limit, and (2) the edge-case char_budget is pre-computed with a potentially shorter notice string than what gets appended, allowing up to 2 bytes of overage in a degenerate all-header input.

Confidence Score: 4/5

Safe to merge after addressing the previously flagged off-by-one and notice-undercount bugs; the two new findings here are low-impact.

The three issues from prior review threads (off-by-one in fold budget, omission notice undercount for dropped headers, blank-line dedup) remain open and affect correctness of the budget enforcement in real inputs. The two new findings (doc comment vs implementation mismatch and stale notice estimate) are P2 but the doc mismatch weakens the test contract. Score is 4 pending resolution of at least the previously flagged P1-level issues.

crates/chat/src/lib.rs — the compress_summary function, specifically the fold-based total_len calculations and edge-case char_budget path

Important Files Changed

Filename

Overview

crates/chat/src/lib.rs

Adds compress_summary() with character/line/dedup budget enforcement and applies it to both compact() and compact_session(); two minor inconsistencies remain: the SUMMARY_MAX_LINES doc comment contradicts the drop-loop implementation (notice is included in the count, not excluded), and the edge-case char_budget calculation uses a stale notice length estimate that can allow 1–2 bytes of overage in rare inputs.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A([compress_summary input]) --> B{text.trim is empty?}
    B -- yes --> Z([return empty string])
    B -- no --> C[Step 1: deduplicate lines case-insensitive, keep first]
    C --> D[Step 2: truncate lines > 160 chars at char boundary]
    D --> E{len <= 24 lines AND <= 1200 chars?}
    E -- yes --> F([return joined as-is])
    E -- no --> G[Partition into headers / bullets / others]
    G --> H[candidates = bullets ++ others, headers always kept]
    H --> I{all candidates fit with headers?}
    I -- yes --> J([return headers ++ candidates])
    I -- no --> K[Drop loop: drop from tail of candidates until line_count + notice <= 24 AND total chars <= 1200]
    K -- fits --> L([return headers ++ kept ++ notice])
    K -- never fits --> M[Edge case: headers alone too long, keep headers until char/line budget, append notice with full drop count]
    M --> N([return kept headers ++ notice])

Loading

_{Reviews (2): Last reviewed commit: "fix(chat): correct budget accounting bug..." | Re-trigger Greptile}

[Cstewart-HC]

@greptileai new commit landed to address issues.