Is Opus 4.6 extended thinking and effort levels in conflict with Ralph restarting?

think fix for this is here https://github.com/snarktank/ralph/pull/93 Check if all stories in prd.json have passes: true and not relying on the output of claude 🤷

think fix for this is here https://github.com/snarktank/ralph/pull/93 Check if all stories in prd.json have passes: true and not relying on the output of claude 🤷

think fix for this is here https://github.com/snarktank/ralph/pull/93 Check if all stories in prd.json have passes: true and not relying on the output of claude 🤷

The linked issue is more about ensuring a robust loop as far as I can see?

I think the slow / no progress is due to thinking and effort

The linked issue is more about ensuring a robust loop as far as I can see?

I think the slow / no progress is due to thinking and effort

The linked issue is more about ensuring a robust loop as far as I can see?

I think the slow / no progress is due to thinking and effort

The linked issue is more about ensuring a robust loop as far as I can see?

I think the slow / no progress is due to thinking and effort

I've noticed the loop is hanging when using opus 4.6 between iterations? Why would that be?

I've noticed the loop is hanging when using opus 4.6 between iterations? Why would that be?

I've noticed the loop is hanging when using opus 4.6 between iterations? Why would that be?

I'm guessing that opus 4.6 is reacting differently, maybe asking follow up questions.

I didn't manage to figure it out. Maybe you could try lowering effort and disable thinking

I'm guessing that opus 4.6 is reacting differently, maybe asking follow up questions.

I didn't manage to figure it out. Maybe you could try lowering effort and disable thinking

I'm guessing that opus 4.6 is reacting differently, maybe asking follow up questions.

I didn't manage to figure it out. Maybe you could try lowering effort and disable thinking

Yes, there is a real tension here.

If the expensive reasoning only lives inside the live model session, then a per-story restart throws away exactly the thing you just paid for. The loop gets a clean boundary, but it also loses any deep intermediate reasoning that was never externalized.

That usually means one of two approaches has to win:

• keep longer-lived sessions and accept more context drift risk
• restart aggressively, but force the run to externalize durable artifacts between stories

The second path tends to age better operationally, but only if the artifacts are actually useful: checkpointed decisions, verifier evidence, failure reasons, and compact context worth carrying forward.

Otherwise you get the worst of both worlds: expensive thinking inside the story, then a restart that forgets it and makes the next story pay again.