Repository navigation
ContextLean v0.2.0
ContextLean v0.2.0
ContextLean gives coding agents a compact, maintained map of a repository so they can find relevant code without repeatedly exploring everything.
Highlights
- Bootstrap Repository
- Audit Context Locality
- Lean Change Review
- Benchmark Context Usage
- Codex and Claude Code support
- Compact AGENTS.md guidance with conditional detailed references
- Project-map-driven navigation
- Semantic transfer and safe-removal contracts
- Original bootstrap intent preserved
ContextLean remains dependency-free and skills-only, with no MCP server, hooks, telemetry or automatic benchmark runs.
Compact context
Representative automatic AGENTS.md guidance was reduced from 7,956 bytes → approximately 5.5 KB (5,476 bytes), while maintaining 74 PASS / 0 PARTIAL / 0 MISSING across the semantic coverage gate. Two mechanisms are intentionally replaced; all 71/71 permanent facets are preserved.
Preliminary validation
The release-aligned batch used five existing tasks and 10 fresh runs, with gpt-5.6-terra, low reasoning, and one Vanilla and one ContextLean run per task. All ten solutions passed acceptance, regression and submitted tests, including independent offline regrading.
- Total tokens: 482,024 → 341,676 (-29.12%)
- Wall time: 189.97s → 134.52s (-29.19%)
- Commands: 18 → 16 (-11.11%)
- Task success: 5/5 → 5/5
Task token differences:
- Navigation: -8.18%
- Bug Fix: -11.35%
- Feature: -17.97%
- Refactor: -24.02%
- Documentation/config: -58.77%
These results are preliminary. There was one run per condition per task. They do not establish statistical confidence, causality, or universal savings. Results apply to the tested configuration.
Unfavorable submetrics remain disclosed: Bug Fix used 35 more output tokens (+3.35%) and two more commands (3→5, +66.67%); Navigation had a higher dated credit-equivalent (+59.93%) because cache composition differed. Raw diagnostics, recoveries and previous datasets are preserved separately.
Product tested: 86043499d5a375cc6a3b295bc7c03b0999eb458c.
Full committed evidence.
Verification
114 offline tests, Ruff lint/format, all four Skill validators, Codex/Claude package validators, evidence checksums, relative links and hygiene pass. No model benchmark was rerun for release.