Native Token Compression: Why Harness CLI Does Not Install RTK or Caveman

Token cost is not only a billing problem. It is a reliability problem.

Long AI coding sessions fail when the model sees too much noise: full logs, repeated stack traces, page boilerplate, and verbose status updates. The obvious shortcut is to install a token-saving competitor tool. We chose not to do that.

Harness CLI references the useful ideas from RTK-style input filtering and Caveman-style output brevity, then implements the workflow inside AIOS.

What Changed

Harness CLI now treats token reduction as two native layers:

1. Input compression: reduce command, browser, and ContextDB payloads before they enter the model.
2. Output compression: make the agent answer compactly without losing exact commands, paths, errors, or risk warnings.

No RTK install. No Caveman install. No global shell hook that rewrites user commands.

Native Input Compression

ContextDB already has a self-contained token strategy engine:

cd mcp-server
npm run contextdb -- context:pack \
  --session <session_id> \
  --limit 60 \
  --token-budget 1200 \
  --token-strategy balanced \
  --out memory/context-db/exports/<session_id>-context.md

The strategy is conservative by default:

• compress repeated lines and stack-run noise;
• preserve critical errors, file paths, commands, and latest state;
• drop low-priority events before truncating protected events;
• emit telemetry for strategy, rawTokenUsed, compressed, dropped, and truncated.

For browser work, the new aios-browser-compress skill pushes agents toward compact reads:

1. page.semantic_snapshot
2. targeted page.extract_text
3. full page.extract_text
4. page.get_html
5. screenshot only when visual evidence is needed

For CLI work, Harness CLI prefers scoped output over full dumps: rg, git diff --stat, sed -n, head, tail, and focused test selectors.

Native Output Compression

The new aios-compress skill defines three response levels:

Level	Use case	Behavior
tight	normal coding work	short technical answers, no filler
ultra	harness logs and checkpoints	one-line evidence + next action
precise	browser operations, safety, irreversible steps	full explicit wording

The key rule: compression must never hide risk. Errors, commands, paths, selectors, dates, and verification gaps stay exact.

Why Not Install Competitor Tools?

Installing another token tool looks fast, but it creates hidden coupling:

• command behavior can change outside Harness CLI's control;
• cross-client behavior differs between Codex, Claude, Gemini, and opencode;
• docs become harder to verify;
• users inherit another dependency, update path, and failure mode.

Native implementation keeps the behavior auditable. The code, skills, and docs live in the repo.

How To Use It

For ContextDB packets:

cd mcp-server
npm run contextdb -- context:pack --session <session_id> --token-budget 1200 --token-strategy balanced

For agent output:

/compress tight
/compress ultra
/compress precise
stop compress

For browser automation, prefer semantic snapshots and targeted extraction before full page text.

Bottom Line

Harness CLI token savings are now native: input compression in ContextDB and browser workflows, output compression in AIOS skills, and no competitor installation step.

That keeps long-running agent work cheaper, quieter, and easier to verify.