FIELD TOOL / CONTEXT BUDGET

Long agent sessions fail when every thought gets equal rent.

This is a small ledger for Claude Code, Codex, and autonomous-agent work. It forces the session to name what deserves context, what needs a receipt, and what should be cut before the model starts treating noise as priority.

Use it before the window gets expensive.

Context is not a warehouse. It is a working bench. If the bench is full of old hypotheses, the agent pays attention to the wrong thing with confidence.

Keep only the current objective, active files, failing command, and verified constraints.
Compress receipts into links, paths, test names, or one-line outcomes.
Externalize logs and raw transcripts to files instead of carrying them in chat.
Cut stale plans, solved branches, and unverified guesses.

Rule: if a note cannot change the next action, it does not get to spend live context.

Copyable context budget prompt

Before continuing, make a context budget ledger.

CURRENT OBJECTIVE
One sentence. If there are two objectives, pick one.

MUST STAY IN CONTEXT
- Active files / URLs:
- Failing command or acceptance check:
- Non-negotiable constraints:
- User preference that affects the next action:

RECEIPTS OUTSIDE CONTEXT
- Raw logs path:
- Test output path:
- Source URL / issue / PR:
- Last verified result:

CUT FROM CONTEXT
- Solved branches:
- Stale plans:
- Unverified assumptions:
- Nice-to-have ideas:

NEXT ACTION
The exact next command, file edit, or question. No vague plans.

Open handoff generator

Budget by next action, not by topic.

"Authentication" is too broad. "Run the failing OAuth callback test after changing auth/callback.ts" is small enough to steer a model.

Keep the acceptance check visible.

The test, curl command, screenshot path, or user-visible behavior is the anchor. Without it, the agent optimizes for plausible work instead of completed work.

Move raw evidence to disk.

Carry a path and a one-line result, not 400 lines of logs. The next agent can read the file if the detail matters.

Kill solved branches loudly.

Write what is no longer true. Silent dead branches are how the next session reopens fixed problems.

Separate confidence from usefulness.

A fact can be verified and still irrelevant. Verified clutter is still clutter if it cannot change the next command.

Leave one instruction for the future agent.

The closeout should end with a concrete move: open this file, run this command, inspect this URL, or ask this missing question.