Configurable RepoMap (--map-tokens) + faster, thread-safe runs

[cryptekbits]

Summary

• Adds --map-tokens to control RepoMap token budget in benchmarks (set 0 to disable).
• Initializes a per-test git repo and enables use_git=True so RepoMap is available during runs.
• Uses an in-memory RepoMap cache for threaded runs to avoid contention and speed up builds.
• Scopes RepoMap cache to each test directory and reduces tree size for faster benchmarks.
• Records and prints map_tokens in saved results and summaries.

Key Changes

• CLI: Add --map-tokens option and wire through to the Coder.
  • benchmark.py
  • README.md
• Repo setup for tests: Initialize a local git repo if missing, set identity, and run with use_git=True (no auto/dirty commits).
  • benchmark.py
• RepoMap tuning for benchmarks:
  • Per-test cache dir: map_cache_dir=str(testdir)
  • In-memory tags cache when threads > 1: repomap_in_memory=True
  • Smaller map for speed: map_mul_no_files=4
  • benchmark.py
• Reporting: Persist and display map_tokens in results and summarize_results.
  • benchmark.py
• Library support (used by benchmark):
  • Coder exposes repomap_in_memory and passes repo_root and use_memory_cache to RepoMap.
  • RepoMap accepts repo_root and use_memory_cache; Coder filters .meta/.docs from map inputs.
  • base_coder.py, repomap.py

Usage

# Enable RepoMap with a budget
./benchmark/benchmark.py --map-tokens 1024 …

# Disable RepoMap
./benchmark/benchmark.py --map-tokens 0 …

Notes

• Threaded runs default to in-memory RepoMap caching; single-threaded runs use the standard on-disk cache.
• Defaults are backward compatible; if --map-tokens is omitted, behavior remains unchanged for typical runs.

Files Touched

• benchmark/benchmark.py
• benchmark/README.md
• aider/coders/base_coder.py
• aider/repomap.py