zfinix/aster/context-economy

Context economy

1. Big outputs go to files, then get queried.

git diff main > /tmp/change.diff then grep -n "fn " /tmp/change.diff; gh pr diff 12 > /tmp/pr12.diff; a build log to /tmp/build.log then grep -m5 error /tmp/build.log. Never print thousands of lines to look at twenty.

1. Summaries before bodies. git diff --stat before any full diff,

wc -l before a full read, ls -la before catting files. Decide what to open from the summary.

1. Delegate bulk reading. Reading ten files to answer one question fills

the conversation with nine files of noise. Send the question to the explorer agent and get back only the conclusion.

1. Ranged reads, always. You almost never need a whole file: read the

function, not the module. Ask for the specific line range the search hit pointed at.

1. Do not re-read or re-run what is above. A file read earlier and

unchanged is still in the conversation; an identical command gives an identical answer. Scroll back instead.

1. Minimal-output probes for external checks. Checking that a URL exists

needs a status code, not the page. Checking a fact in a doc needs one grep, not the document.