v0.6.16

Release v0.6.16

v0.6.15

Fixed

• Knowledge ingest now sends Apollo chunk provenance through context: while retaining metadata compatibility, so source file, heading, section path, chunk index, and token count persist with document chunks.
• Batch embedding now falls back to per-chunk Legion::LLM.embed when embed_batch is unavailable.
• Retired corpus files now emit explicit observation entries tagged retired instead of using an Apollo-unknown document_retired content type.
• Retrieval and synthesis failures are logged through helper-based exception handling, and synthesis returns nil instead of presenting error strings as answers.
• Monitor-only installs now enable the maintenance actor without requiring corpus_path.
• Quality reports count Apollo query access logs with the query action recorded by lex-apollo.

v0.6.14

Changed

• Knowledge defaults are now declared directly in Knowledge.default_settings, and helpers, runners, actors, and JSON sidecar persistence use Legion logging, settings, and JSON helpers end to end.

v0.6.10

Fixed

• Chunker content_hash now uses MD5 + whitespace normalization (matching Legion::Extensions::Apollo::Helpers::Writeback.content_hash) instead of raw SHA-256. This keeps knowledge chunk deduplication aligned with Apollo writeback and avoids insert truncation on deployments whose apollo_entries.content_hash column is still fixed at 32 characters.
• upsert_chunk_with_embedding now requires an explicit {success: true} from handle_ingest before reporting :created/:updated. Failure hashes, missing success keys, and non-Hash returns are reported as :skipped with a warn log instead of false-positive success counts.

v0.6.9

Fixed

• Manifest.scan no longer crashes on Errno::EPERM/EACCES encountered during corpus walk (common on macOS for TCC-protected paths like ~/Library/Accounts). Unreadable subdirs are pruned with a debug log; scan continues. Replaced Find.find with a recursive walker that rescues per-dir; also tolerates Errno::ELOOP and Errno::ENOENT for files that disappear mid-scan.

Version note: 0.6.8 is reserved for the companion fix/content-hash-md5-match-apollo-schema PR (chunker SHA-256 → MD5 hash fix). Both branches target 0.6.7 as their merge base; this PR claims 0.6.9 to avoid intra-batch collision.

v0.6.7

Fixed

• Runners::Query.retrieve_chunks now extracts the entries array from retrieve_relevant's Hash response instead of returning the Hash directly, preventing TypeError: no implicit conversion of Symbol into Integer on knowledge query
• Runners::Maintenance.health now returns { success: false, error: 'corpus_path is required' } when called with path: nil and no settings fallback, instead of raising TypeError: no implicit conversion of nil into String; falls back to Legion::Settings.dig(:knowledge, :corpus_path) when available

v0.6.6

Fixed

• chunk_exists? queries by content_hash column directly instead of LIKE on text[] tags column

v0.6.5

Fixed

• chunk_exists? uses cross-DB Sequel.like instead of PostgreSQL-only pg_array_op

v0.6.4

Changed

• update to rubocop-legion 0.1.7, resolve all offenses

v0.6.3

Fixed

• CorpusWatcher and MaintenanceRunner actors now override time instead of every_interval — the Every actor base class uses time for Concurrent::TimerTask interval, causing both actors to fire every 1 second instead of 300s/21600s