Repository navigation
fix: backport 5 user-facing fixes to 0.12.1 (batch late-3) - #33725
Conversation
…33645) The FAISS rebuild and the promotion pass both read every stored episodic vector. The rebuild decoded a blob before checking its width, so a blob that is not a whole number of float32s raised and stopped the whole rebuild, leaving the rows after it out of vector search. Promotion clustering raised the same way on such a blob, and on two vectors of different widths, so one odd row ended every promotion pass. Check the rebuild's width on the bytes, as the episodic scoring set already does, and have clustering leave a vector it cannot compare out of every cluster. Every input that worked before gives the same result. (cherry picked from commit 4a91886)
…er (#33641) After a restart the conversation registry is rebuilt from each kept run's state.json, reading its age from `float(updated_at)`. A string or an int past float range raised, so the run was skipped and its conversation was never re-registered for the TTL sweep. `-inf` was read as a time and aged the conversation past its TTL, so the first sweep released it. The stamp is now read through `_finite_time`, the reader the tombstone code already uses, with `started` as the fallback; a stamp that is not a finite number reads as age 0. `_finite_time` itself returns 0.0 for an int past float range instead of raising. (cherry picked from commit 4a6b1e6)
(cherry picked from commit 0207374)
…33649) The Telegram renderer holds code spans and fences behind NUL<index>NUL placeholders while it converts Markdown and strips horizontal rules. Reply text that already carried that shape, such as tool output with raw NUL bytes the agent echoed, was read back as a placeholder. An index past the stash raised IndexError, which aborted the streamed chunk or the final seal, and an index inside it copied a code span into the wrong place. Drop NUL from the text before stashing. Telegram shows no NUL, so the reply looks the same. md_to_telegram_html_safe drops it before the credential scan too, so a key a NUL cuts in two is rejoined before the scan reads it, not after. (cherry picked from commit 813da15)
UX Review (Opus 5.5) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing to check. Each of the five cherry-picks fixes a defect that a test added in this PR shows failing on base, and the Goal covers all five. What this change shipsInventory (7 items) — 7 justifiedIntent: give 0.12 users five bug fixes that already merged to main, through a 0.12.1 patch. FIX.
[FIRST-PRINCIPLES-REVIEWED] 50c0fb3 Description read[DESCRIPTION-READ] 8c54694fe56ed89906b381528c8187dfa6a21b2aea558b266b47730194875481 That sha256 names the PR description this verdict read. Recompute it with |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: five small cherry-picks of fixes already merged to main, each with tests, and the one stricter check (the region validator) runs only on add or edit, not when stored records load. Notes
[DESIGN-REVIEWED] 50c0fb3 Description read[DESCRIPTION-READ] 749c89f7046a0be6bc9417561168db656468692171fe646bf8312853c527ad15 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
Problem / Motivation
Goal: Ship 5 user-facing fixes that merged to main after the 0.12 cut to 0.12.1 users.
Each fix is a bug that 0.12 users can hit today. Each one cherry-picks onto
release/0.12.0with no conflict.Why it matters
0.12 users keep hitting these bugs until a patch release carries them. A single bad stored vector breaks episodic memory. A kept subagent run with a bad timestamp is lost. A failed WhatsApp image upload gives no reason. A region name with non-ASCII digits is accepted, and a NUL byte breaks a Telegram reply.
Not a goal
git cherry-pick -xof the main merge commit, unchanged.What changed (motivation → approach → change)
The 5 fixes are applied in main merge order, one
cherry-pick -xof each main merge commit, onrelease/0.12.0at8e30b9d6c0, which is the head after #33574 (s3b1) merged.4a918862de4a6b1e628160d72b9aa20207374048813da15dd4Dropped: none. All 5 picked cleanly in order, and none failed its tests.
Backwards compatibility
Compatible: each commit is the main commit unchanged, already reviewed and merged there. Each main PR states its own compatibility. The only stricter check is #33642's region validator, and every ASCII region name that passed before still passes.
Removes nothing: no field, API, key or config is removed. #33642 refuses a region with a non-ASCII digit only when a record is written or edited, and such a region could never reach AWS. Stored records still load.
Tests
Run on this branch at
50c0fb3edb, in a clean Python 3.12 venv (uv venv -p 3.12,uv pip install -e . -r pyproject.toml --group dev), withenv -u KIROCREW_NATIVE_SKILL_PROJECTION PYTHONNOUSERSITE=1and one pytest run per file:test/test_continuable_followups.py: 24 passedtest/test_instances_ssm.py: 133 passedtest/test_telegram_renderer_nul.py: 6 passedtest/test_vector_memory_embedding_lifecycle.py: 46 passed, 8 skippedtest/test_whatsapp_turn_renderer.py: 40 passedNo file under
website/changed, so vitest and tsc were not run.Proof: not provable -- other: a cherry-pick batch of already-merged main fixes, each proven in its own main PR
Manual verification
N/A: each fix was verified in its own main PR, and this batch carries them unchanged.
Related Issues
0.12.1 backport batch late-3: #33645, #33641, #33646, #33642, #33649.
Checklist
cherry-pick -xof a main commit with a Conventional Commits titleContribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.