TokenTrim copilot-instructions.md

Terminal output compression with tt (TokenTrim)

To reduce token usage, prefix heavy terminal commands with tt when you run them. tt compresses noisy command output (60-90% fewer tokens) while keeping everything you need: it preserves exit codes and, if a command fails, it saves the full raw output to disk and prints the path.

Prefix these commands with tt (e.g. tt git diff, tt terraform plan):

• Version control: git status, git diff, git log, git show, git branch, git add/commit/push/pull
• Tests / build: pytest, jest, vitest, cargo test, go test, npm test, make, mvn, gradle
• Containers / K8s: docker ps, docker logs, docker images, kubectl get, kubectl logs, kubectl describe, helm, oc, podman
• Cloud / IaC: terraform plan, terraform apply, terraform validate, az ..., aws ..., gcloud ...
• Packages / lint: npm install, pnpm install, pip list, eslint, tsc, ruff, mypy
• System: journalctl, systemctl status, curl
• Files: ls, grep, find, cat

You can also compress arbitrary text before quoting it back by piping it through tt trim (auto-detects JSON, diffs, stack traces, ML training logs, pandas df.info(), and sklearn classification_report). For an ML training run, use tt train <command> to get the metric curve, best epoch, and early-stop signal.

At the START of a task, run tt map once: it prints a compact overview of the repository (directories + function/class signatures) so you do not need many exploratory ls/cat/grep calls.

Batch work into one step (CodeAct)

When a task needs SEVERAL independent operations (read multiple files, search, then inspect), do them in ONE tt code call instead of many separate tool calls. Fewer turns means the whole conversation is re-sent to the model fewer times = large token savings (this compounds when MCP servers are loaded).

Inside tt code these helpers are preloaded and already return compressed text:

• glob(pattern) - list files (recursive, skips .git/node_modules/...)
• view(path, sig=False) - read a file (sig=True: signatures only)
• grep(pattern, path=".") - grouped matches (pure Python, no grep needed)
• sh("cmd ...") - run a shell command, return compressed output
• run("cmd1","cmd2",...) - run several, return combined compressed output