Codex can compact a long conversation so the agent can continue after the context window fills up. OpenAI describes compaction as replacing the conversation input with a smaller representation that lets the agent continue with an understanding of what happened. That is useful. It is not the same thing as a verified handoff.

Radar read: compaction preserves a conversation path. A recovery packet preserves the work: changed files, failed attempts, verification state, decisions, and the next smallest action.

Why this matters.

A compacted task may remember the broad story while losing precise edge details. It may know that a deploy happened, but not whether the live URL was checked. It may remember that a script edited files, but not which files were backed up. It may preserve intent, while the source of truth still lives on disk.

Codex usage also depends on context, prompt size, model choice, tool use, and active limits. When a task is getting long enough to compact, it is already the right time to create a smaller recovery packet.

Before a compacted session continues.

  1. Check the working tree. Changed and untracked files matter more than the summary.
  2. Record failed commands. A failure that disappeared from context may be the key to the next safe step.
  3. Verify claims. Mark live checks, tests, deploys, and backups as done only when evidence exists.
  4. Write the next smallest action. Keep the continuation prompt narrow.
  5. Archive risky assumptions. Label guesses so the next agent does not treat them as facts.

What ShardStitch does here.

ShardStitch creates a continuation packet outside the tired chat: disk facts, changed files, failed attempts, verified checks, inferred claims, and next action. Then Codex can continue, or the work can move to Claude, Cursor, Gemini, Roo Code (archived May 2026), or another tool without relying on compaction alone.

Source trail

Related pages