thinker

A knowledge cache for your coding agents Pre-seeds repo architecture & conventions before each turn. Verified against code changes.

Docs →

$ thinker benchmark report

PR change evaluation

14% faster

Wall clock time

2.7m → 2.3m

18–21% less

Input tokens

stops context re-reads

17–22% fewer

Tool calls

cuts blind exploration

+33% solved

Correctness

Claude Fable pass rate

Measured on real PR tasks across Claude Fable, Gemini Flash & GPT-6 Astra View full report →

Claude Code • Codex CLI • Cursor • Gemini CLI

Dimension	Metric	Claude Fable (Claude Code)	Gemini 3.8 Flash (Antigravity CLI)	GPT-6 Astra (Codex CLI)
Timing	Wall clock time	14% faster	8.5% faster	7.0% faster
Token Usage	Input context	18% less	21% less	19% less
Token Usage	Output tokens	11% less	22% less	11% less
Tool Efficiency	Tool calls	17% fewer	22% fewer	18% fewer
Tool Efficiency	File reads	no diff	27% fewer	14% fewer
Cost	Cost per task	10% cheaper	16% cheaper	no diff
Correctness	Criteria accuracy	10% improved	no diff	no diff
Correctness	Fully solved	33% more solved	no diff	no diff