Skip to content

v1.7.0

Choose a tag to compare

@kaizen-bot kaizen-bot released this 15 Oct 20:18
· 43 commits to main since this release

v1.7.0 (2026-10-15)

Bug Fixes

  • api: Store the repaired guideline category and match separators case-insensitively (#346, 6c04a17)

  • claude: Drop the ignored Write(.evolve/**) allow-rule and strip it on upgrade (#379, 797e6ba)

  • claude: Working doctor import check, drop ignored Write allow-rule (#379, 797e6ba)

  • codex: Probe skills codex actually ships in install.sh status (#337, 5581ee6)

  • doctor: Detect the EVOLVE.md import from the session's own context (#379, 797e6ba)

  • embeddings: Correct type annotations and type-check optional extras in CI (#375, 34cd03e)

  • experiments: Create export files without touching the process umask and skip blank source instructions (#395, 0fa7af1)

  • experiments: Decide a CUGA run by a task with an intent, not by marker files, and reject malformed task manifests (#359, b224bb4)

  • experiments: Let a CUGA run's own files mark it as a run, read 0/1 verdicts and require a string intent (#359, b224bb4)

  • experiments: Parse CUGA string verdicts instead of coercing them with bool() (#359, b224bb4)

  • experiments: Read CUGA child runs past a stray JSON file and skip tasks without an intent (#359, b224bb4)

  • experiments: Reject a CUGA task_ids option that lists no IDs (#359, b224bb4)

  • experiments: Report unhashable evidence cleanly and keep plain-write file modes (#364, 37255e6)

  • experiments: Skip malformed CUGA evaluation checks (#359, b224bb4)

  • experiments: Warn about CUGA runs left unread and tolerate a parent's bad metadata.json (#359, b224bb4)

  • guidelines: Bound the length of each failed check in the outcome prompt (#351, d2140e8)

  • guidelines: Keep segmentation in step with the elision marker (#350, 7fc8cb2)

  • guidelines: Let a segment of a failed run that did not cause the failure return no guidelines (#351, d2140e8)

  • guidelines: Repair multi-valued categories instead of dropping the response (#346, 6c04a17)

  • guidelines: State the full run length in the elision marker (#350, 7fc8cb2)

  • hermes: Keep the empty-id fallback out of the digest (#338, 7bbd835)

  • hermes: Make trajectory filenames injective so provenance cannot misattribute sessions (#338, 7bbd835)

  • hermes: Make trajectory filenames injective so provenance cannot misattribute sessions (#331) (#338, 7bbd835)

  • hermes: Write only the digest-suffixed trajectory name (#338, 7bbd835)

  • hooks: Keep source association identifiers out of PII redaction (#358, 9ece967)

  • install: Register a stable marketplace copy for curl | bash Claude installs (#390, 8b23c88)

  • memory: Carry association times forward only for consolidation (#358, 9ece967)

  • memory: Ignore an untrustworthy stored association time when combining records (#358, 9ece967)

  • memory: Leave guidelines with a non-scalar scope out of consolidation (#358, 9ece967)

  • memory: Preserve scope metadata through consolidation (#358, 9ece967)

  • memory: Reject untrustworthy association times and age legacy members from created_at (#358, 9ece967)

  • milvus: Ensure the index on every first read, bound the Load, and probe load state once per namespace (#344, 982e4fc)

  • milvus: Load the collection before every read (#344, 982e4fc)

  • milvus: Skip the Load request when the collection is already loaded (#344, 982e4fc)

  • packaging: Validate paired release requirements and missing dates (#408, d629e4f)

  • processing: Replay profiles stored before the trajectory window and reject a tail that consumes it (#350, 7fc8cb2)

  • processing: Resolve trajectory window settings from the environment in GuidelineConfig (#350, 7fc8cb2)

Documentation

  • claude: Use the positional form for the sandbox-build target (#337, 5581ee6)

  • config: Say the accurate consistency method keeps its own trajectory window (#350, 7fc8cb2)

  • experiments: Point the CUGA README at the outcome-grounded prompt (#359, b224bb4)

  • experiments: Say a consolidated guideline is recorded only when a member had sources (#358, 9ece967)

  • guidelines: Flag the segment clause in failure prompts as not yet evaluated (#351, d2140e8)

  • hermes: Scope the filename injectivity claim to str ids and pin the on-disk name (#338, 7bbd835)

  • plugins: Address #337 review — qualify publish gap, harden marketplace build, widen excluded-skill guard (#337, 5581ee6)

  • plugins: Describe what the claude, codex, and bob builds actually ship (#337, 5581ee6)

  • plugins: Refresh the claude, codex, and bob plugin docs to match what each build ships (#337, 5581ee6)

Features

  • build: Generate .claude-plugin/marketplace.json from plugin.toml (#337, 5581ee6)

  • experiments: Add a CUGA partitions option (#359, b224bb4)

  • experiments: Add CUGA adapter (#359, b224bb4)

  • experiments: Export playbooks and retrieval indexes (#364, 37255e6)

  • experiments: Scaffold guideline pipeline (#354, ceeee6d)

  • guidelines: Make the trajectory window configurable and keep the end of long runs (#350, 7fc8cb2)

  • guidelines: Use an outcome-grounded extraction prompt when a trajectory has an outcome (#351, d2140e8)

  • memory: Keep source provenance on consolidated guidelines (#358, 9ece967)

  • packaging: Split core and convenience distributions (#408, d629e4f)

  • processing: Accept a task outcome on trajectories and set evidence from it (#354, ceeee6d)

  • processing: Accept a task outcome on trajectories and set evidence from it (#347, 22f99fe)


Detailed Changes: v1.6.1...v1.7.0