How to stretch your Claude Code and Codex usage limits

Know where the limit goes

Every agent turn costs tokens twice. Once for what the model reads: the whole conversation from the start, files, command output. Once more for what it writes: reasoning, code, the answer to you.

That gives three main sources of waste:

• a long conversation: on message one hundred the model rereads the ninety-nine before it;
• wordy answers: greetings, retelling what was done, lists of “what I could also do”;
• a strong model on simple work: renaming a file or fixing a typo does not need the top model.

Habits that help in any agent: one task per conversation, full file paths instead of “it is somewhere in there”, a new conversation for a new topic.

Ask the agent to answer briefly

Extra words in an answer cost tokens, and then cost again when the model rereads them on the next turn. A “be brief” rule pays for itself quickly.

Hedoq does this for you. Every agent gets a starter set of rules: answer briefly, no greetings or praise, do the work first, then write a short result (what was done, how it was checked, what is needed from you). If the agent does not know or did not check, it says so.

The set is on by default. To turn it off, open the settings, the Agents block, and switch off “Brief agent answers”. The change applies from the next new session.

Clear the context without losing the thread

/clear in Claude Code or /new in Codex wipes the conversation. That is cheap, but the agent forgets everything: the project, your rules, what is already done. The next ten messages go to explaining it all again.

Hedoq has “Refresh session” for this:

1. Under the input field, the context line shows how much of the model window is used.
2. At 70 % and 90 %, a “Refresh session” hint appears above the field.
3. After you confirm, the agent writes a note: the project, your rules, what is done, where it stopped, what is next, which files to read.
4. The context is cleared, the new conversation gets the note and tells you briefly where things stand.

With Claude, the note is written by an invisible copy of the conversation, so nothing extra shows up in your session. If the note fails, the context is not cleared. The old conversation stays in the feed above a line, so you can reread it. It works for Claude Code and Codex when the agent is not in the middle of a turn. More on long work: long tasks without losing the thread.

Pick the model for the task

The strongest model earns its cost on architecture, hard debugging and big rewrites. For small edits, code search or a first draft, a lighter model does the job and uses the limit more slowly.

In Hedoq you switch the model inside the session, with the model chip in the input field:

• Claude Code offers Fable 5.1, Opus, Sonnet and Haiku; the switch goes through /model when the agent is free;
• Codex takes its list from Codex itself; the session restarts with the chosen model and the same conversation.

A sensible order: start a hard task on a strong model, then switch to a lighter one once the plan is clear and only routine work is left.

Keep the limits in sight

Finding out about a limit from an error in the middle of a task hurts. In Hedoq the limits of every connected subscription sit by the camera notch on your Mac: hover the notch.

• Each subscription shows what is left as a bar, like a battery: the five-hour window and the week.
• When less than 10 % is left in a window, you get a notification.
• The same panel lists sessions with their model and state, so you see who is spending the limit right now.

With two subscriptions, say Claude and ChatGPT, you see which one has more room and can give the next task to it. How to keep both agents side by side: Claude Code or Codex.

Watch the modes that multiply the cost

Some techniques spend the limit of several agents on purpose. “Model council” gives one task to several models, “Split the task into parts” starts several sessions. Worth it on hard questions, expensive on easy ones.

In Hedoq the lead session of a council warns you before a large task that the mode will spend the limit of every model, and suggests narrowing the question. Council members only read and reply; they do not edit files.

A short daily checklist:

1. Brief answers are on.
2. New topic, new session.
3. At 70 % context, run “Refresh session”.
4. Give routine work to a lighter model.
5. Before a big task, glance at the notch.