GTST-20 feat: add compact TOON-inspired output format for agent token efficiency

[MonteYin]

Summary

• Add self-implemented compact output format inspired by TOON — no external dependencies
• Default format is now toon (compact), json available via --format json
• Remove human format and Auto detection — gitsift is agent-only
• Remove similar crate dependency (was only used for human output)
• ~40% token savings vs JSON on real-world diffs (38KB → 23KB)

Key optimizations

• Tabular rows for hunk lines: {tag,content,old,new}: header + CSV-like rows
• Context lines (unchanged) stripped — agents only need change lines for staging
• Redundant file_path removed from hunks (already on parent file entry)

Test plan

• cargo fmt --check passes
• cargo clippy --all-targets — zero warnings
• cargo test — all 96 tests pass
• cargo build --release — binary works: diff, stage, status, protocol
• Manual stage test on real repo verified
• Size comparison: JSON 38,187 bytes vs toon 22,558 bytes (40.9% reduction)

[MonteYin]

Review -- Opus 4.6

Review Result

Overall Verdict: HAS_ISSUES

---

1. Correctness

Verdict: HAS_ISSUES

Issues

Important

• FileStatus casing inconsistency between toon and JSON formats (src/toon.rs:28): The toon format uses {:?} (Rust Debug trait) to render FileStatus, producing PascalCase values like Modified, Added, Deleted, Renamed. The JSON format uses #[serde(rename_all = "snake_case")] producing modified, added, deleted, renamed. An agent that parses both formats will see different status values depending on --format. This will break any agent code that checks status == "modified" against toon output.

Minor

• Unused Response wrapper in format_stage_result (src/toon.rs:76): Response::success(result) is constructed only to read resp.version, which is always PROTOCOL_VERSION. The other two formatters (format_diff, format_status) use crate::models::PROTOCOL_VERSION directly. This allocates an unnecessary Response struct. Not a bug, but inconsistent with the sibling functions.

Fix Instructions

• For the FileStatus casing: Replace {:?} with an explicit match or implement Display for FileStatus that matches serde's snake_case output. Simplest fix at src/toon.rs:28:

  let status_str = match file.status {
      FileStatus::Modified => "modified",
      FileStatus::Added => "added",
      FileStatus::Deleted => "deleted",
      FileStatus::Renamed => "renamed",
  };
  writeln!(buf, "    status: {}", status_str).unwrap();

• For the unused Response wrapper: Replace line 76-79 with writeln!(buf, "version: {}", crate::models::PROTOCOL_VERSION).unwrap(); and remove Response from the import if no longer needed.

---

2. Architecture

Verdict: HAS_ISSUES

Issues

Important

• CLAUDE.md not updated (CLAUDE.md:28,36,46,51,56): CLAUDE.md is the primary guidance file for AI agents working on this codebase. It still references GlobalArgs, human format, Auto detection, IsTerminal, and --format json|human|auto. Since this PR removes all of those, CLAUDE.md will actively mislead agents into using nonexistent CLI options. Specific stale lines:
  • Line 28: GlobalArgs (struct no longer exists, was renamed/removed)
  • Line 36: output.rs described as "JSON (default when piped) vs human-readable"
  • Line 46: "auto-detect terminal via IsTerminal; JSON by default when piped"
  • Line 51: "list unstaged hunks (JSON or human)"
  • Line 56: --format json|human|auto
• Skill doc not updated (skills/gitsift-staging/SKILL.md:32-35): The skill doc still recommends gitsift diff --format human which will now fail with a clap parse error. This will cause agent tool invocations to fail.

Minor

• No toon format error response: The toon formatter only handles success cases (ok: true). Error responses from the CLI (e.g., invalid repo path) go through anyhow and print to stderr via main(), so this isn't a bug -- but it's worth noting that the toon format has no ok: false representation. If error formatting through toon is ever needed (e.g., for stage partial failures), the current format_stage_result does handle the errors array, so the main gap is only for top-level command failures, which correctly use stderr + exit code.

Fix Instructions

• Update CLAUDE.md to reflect the new format options: --format toon|json, default is toon, no auto-detection, no human format, toon.rs in the architecture diagram, and remove IsTerminal from design decisions.
• Update skills/gitsift-staging/SKILL.md lines 32-35: Remove the --format human section entirely, or replace with gitsift diff (default compact format). Also consider mentioning the toon format in the "See what changed" section since it's now the default.

---

3. Security

Verdict: PASS

No security concerns. The toon format is output-only, no new input parsing is introduced. The escape function properly handles the relevant special characters for the CSV-like row format.

---

4. Performance

Verdict: PASS

The implementation is straightforward String building with write! macros. Filtering context lines with iter().filter().collect() is fine for the expected hunk sizes. The removal of the similar crate dependency is a positive change -- less code to compile, smaller binary.

---

5. Simplicity

Verdict: PASS

The new code is clean and well-organized. The toon.rs module is self-contained with clear formatting functions. The removal of the Auto variant and resolve_format() indirection simplifies output.rs significantly. Test coverage is thorough with both unit tests in toon.rs and integration tests in tests/cli.rs.

The escape function is minimal and correct for the use case. The tabular row format is a reasonable design choice for token efficiency.

---

Summary

The core toon formatter implementation is solid -- clean code, good test coverage, and a meaningful token reduction. However, there are two issues that should be addressed before merging. First, FileStatus renders as PascalCase in toon output (Modified) but snake_case in JSON (modified) due to using Rust's Debug trait instead of matching serde's rename_all convention -- this will trip up agents that parse both formats. Second, CLAUDE.md and the skill doc (skills/gitsift-staging/SKILL.md) still reference the now-removed --format human and Auto detection, which will actively mislead AI agents working on or using this tool. Both are straightforward fixes.