v0.21.0

57.9% faster than shell and 37.4% faster than upstream

Code Cortex reduced complete-task time by 57.9% versus shell tools across 22 eligible pairs.
It reduced time by 37.4% versus upstream across 14 eligible pairs.

Comparison	Overall	10K lines	1M lines	100M lines	One-shot	Multi-step
Code Cortex vs shell	57.9% (22/24)	57.7% (8/8)	56.1% (6/8)	59.4% (8/8)	33.3% (12/12)	75.7% (10/12)
Code Cortex vs upstream	37.4% (14/16)	37.8% (8/8)	36.9% (6/8)	Unavailable	33.3% (8/8)	42.6% (6/8)

Each result includes only terminal, source-correct, exact pairs from a controlled Linux run on September 10, 2026.
The run used fresh gpt-5.6-sol sessions, medium reasoning effort, and warm synthetic C++ indexes.
Upstream exceeded both the 4 GiB and 32 GiB memory budgets at 100M lines.
These results cover one-shot lookups and multi-step three-hop chains.

Read the full results.
Use the reproduction kit to repeat the controlled comparison.

What changed

• Bounded result storage reduces task-time memory pressure on large repositories.
• Targeted trace_path searches bound graph work between known endpoints.
• inspect_symbol returns source, callers, tests, cross-language links, and trust signals in one call.
• CBM_UPDATE_CHECK=0 disables the background release check for controlled runs.
• The public harness compares Code Cortex, shell tools, and pinned upstream with source-correct grading.

Install or update

curl -fsSL https://raw.githubusercontent.com/tigercosmos/code-cortex-mcp/main/install.sh | bash

Restart the coding agent after the update.