Smaller context¶
Output filters¶
keelyard run <command> runs a command and prints a shorter version of its output. Filters exist for
git status/log/diff, pytest, cargo, go and jest tests, grep and rg, find and ls, and ruff, eslint and mypy.
Every failure, error and file:line is kept; what is dropped is repetition and progress noise. The filtered
output is never longer than the original.
In Claude Code the same filters run on Bash results through PostToolUse. The command itself is not
changed, only what the model reads. Each shortened output ends with an id, and keelyard recall <id>
(or the MCP recall tool) returns the full text.
keelyard run pytest -q
keelyard run -- git log --stat -n 50
pytest -v 2>&1 | keelyard filter --for "pytest -v"
On 8 real commands, the filters cut 56,522 tokens to 5,901 and kept 100% of the facts we extracted (method).
Deduplicated reads¶
When the agent reads a file it has already read in this session, and the file has not changed, keelyard
answers with a one-line notice instead of the file. If the file changed, the MCP read_file tool returns
only the diff. When the agent's context is compacted, this memory is cleared so it can read again.
Structure first¶
project_map: files per directory, with approximate token sizesoutline path: classes and signatures with line numbers (Python is parsed properly; other languages use patterns)find_symbol name: where something is defined
keelyard code map
keelyard code outline src/cart.py
keelyard code find Cart