omo runs on senpi 2026.10.10-9, which stops automatic turns from running without bound. After one message from you, stream-rule nudges, goal continuations and other automatic follow-ups pause after 150 turns, or once 12 automatic turns in a minute do no work, and each stream rule corrects a message at most once. When it pauses, the session says so, and your next message picks it back up. Full engine notes: senpi 2026.10.10-9. (senpi#2967, senpi#2969, #9772)
A required question can't be read as permission. senpi's ask_user_question accepts required: true for a question that gates an action, such as an approval before an irreversible step: if it times out, is dismissed or can't be shown, the model is told not to take that action. (senpi#2949, senpi#2959)
The configuration guide documents the Anthropic 1-hour prompt cache and the cache keep-alive. (#9770, #9771)
Fixed
Memory locks stop churning on machines with many sessions. A session waiting on a memory lock no longer republishes a recovery lock on every retry: it reads first, and lock polls back off with jitter instead of a fixed 25 ms. Locks and candidate files left by processes that died are reclaimed, so the lock directory stops growing. (#9749, #9750)
Single-turn answers no longer carry the progress-ledger block. Hephaestus and Sisyphus keep the Now/Next ledger for multi-step work tracked in todos; a request finished in one turn gets a plain answer, without lines such as Next: none. Thanks to @cynkai. (#9616, #9676)
A provider that sets its own anthropic-beta header keeps the betas a request needs, so Claude 5.5 models through such a route no longer fail their first call. (senpi#2957, senpi#2963)
Deprecating platform packages for a release that never shipped rechecks the registry before failing, so a deprecation that is still propagating no longer fails the workflow. (#9769)
- bb9a255 Merge pull request #9773 from code-yeongyu/docs/5.1.27-notes
- 6469097 docs(changelog): 5.1.27 release notes
- 498c508 Merge pull request #9771 from code-yeongyu/docs/9770-anthropic-prompt-cache
- 42a542f docs(configuration): document the Anthropic 1-hour prompt cache and the cache keep-alive
- 007dbce Merge pull request #9769 from code-yeongyu/ci/deprecate-readback-retry
- fe3db35 Merge pull request #9750 from code-yeongyu/fix/9749-memory-lock-reclaim-and-backoff
- 44b16c2 ci(npm-deprecate-version): recheck the read-back with backoff for about 2 minutes before failing
- 2c52bdf Merge pull request #9676 from cynkai/fix/9616-ledger-only-for-todo-work
- 158b641 Merge remote-tracking branch 'origin/dev' into fix/9749-memory-lock-reclaim-and-backoff
- bb11c19 Merge pull request #9765 from code-yeongyu/release/v5.1.26-source-state
- 4cc4939 Merge remote-tracking branch 'origin/dev' into fix/9749-memory-lock-reclaim-and-backoff
- 69649bc test(memory-core): inject the unreadable-lock failure instead of chmod, so the sweep test holds on Windows and as root (#9749 review)
- 7892ac6 refactor(memory-core): split the lock module by responsibility, sweep each dead-owner lock in isolation, and bound the backoff test by its contract (#9749 review)
- d644aa2 fix(memory-core): read before publishing the recovery lock, back off lock polls with jitter, and reclaim dead-owner locks and candidates (#9749)
- ea8fec9 test(memory-core): lock waits back off, a live recovery holder costs no publishes, and dead-owner locks and candidates are reclaimed (#9749)
- 2c4009e fix(prompts): keep the progress-ledger block out of single-turn answers
Thank you to 1 community contributor:
- @cynkai:
- fix(prompts): keep the progress-ledger block out of single-turn answers
bun add -g omo-ai