Decide when to continue or start fresh

Continue when the session is making coherent progress on one task. Start a new session when unrelated work or repeated rejected approaches are obscuring the goal, and carry forward what you learned.

A restart is useful when it improves the input or isolates a new task. It is not a universal remedy after a fixed number of messages.

Keep the product-specific guidance distinct

Claude Code's best-practices guide offers a practical heuristic: repeated corrections on the same issue can justify clearing the conversation and restarting with a better prompt. This is Claude guidance to evaluate, not a cross-provider law.

Codex guidance also supports sustained work on a coherent task, with maintained context and verification. Repeatedly restarting such a task can discard useful investigation and decisions. See Codex planning and verification.

Use the available session and compaction controls for the product you are running. A prompt asking for a “fresh mind” does not establish a new session or an independent review.

Distinguish missing evidence from clutter

Suppose the agent keeps changing retry arithmetic, but the real ambiguity is whether the caller starts counting at zero or one. Another wording correction does not settle that contract. Inspect the caller or ask the person who owns the requirement.

If the session now contains several abandoned fixes, write down what the evidence established and why those fixes failed. That record can support either continued work or a restart.

Prepare a restart packet

Task and acceptance criteria
Current source revision and actual working-tree changes
Confirmed evidence and checks already run
Rejected approaches and the evidence against them
Open questions
Next useful investigation

Do not treat an old check result as evidence for a changed revision. Do not infer that a new tool inherits the previous session's permissions, credentials, or decisions.

For an independent review, provide the source, diff, and acceptance criteria before the implementer's conclusions. Keep enough context to review accurately without steering the reviewer toward agreement.

Try the choice on a real failure

Choose a task that has repeated the same unsuccessful approach. First establish the missing facts and record the current state. Then try a fresh session with that improved packet.

Compare the checked result and additional work with the previous attempt. If restarting helps repeatedly on that task class, adopt it as a local heuristic. If the task continues to progress well in one session, retain that continuity.

What goes wrong

Restarting with the same ambiguous prompt repeats the information problem. Restarting without the current diff can undo useful work. Continuing through unrelated tasks can make old assumptions hard to distinguish from current ones. Treating one improved answer as proof of a universal turn limit overstates the evidence.

How to check

The new or continued session should identify the correct revision, preserved constraints, and next useful action. Confirm these against the task record. Finish with the same acceptance checks, regardless of the session strategy.

The artifact and approval guide shows how a pipeline retains decisions independently of a model's conversation.