Steadlith 1.0.1

Steadlith 1.0.1 fixes state-safety, concurrency, recovery, and planning problems in the local SQLite workflow. The v1 chunk identity schema and runtime dependencies are unchanged.

• Reject overlapping state paths and prevent forced cache exports from replacing configuration, databases, sidecars, manifests, or migration records.
• Roll back interrupted cache/index transactions, keep related reads on one SQLite generation, and prevent delayed operations from publishing an older manifest mirror.
• Reject malformed cache/configuration data with actionable errors and calculate embedding estimates correctly when active vectors can be reused after cache pruning.
• Validate migration target size and the exact requested configuration changes before publishing; accept quoted and padded TOML names.
• Pin OpenAI requests to the official endpoint instead of inheriting an unintended endpoint override.
• Add clean installed-wheel workflow checks to pull-request and release CI, correct contributor setup instructions, and update website dependencies with the reviewed Next.js security fixes.

Install or upgrade:

python -m pip install --upgrade steadlith==1.0.1

This release was checked by the configured Linux/Windows/macOS CI matrix, package installation workflows, documentation builds, and website checks. Live learned-provider execution is not part of these checks; the default hash provider remains an offline lexical retriever. Complex inline/dotted/multiline TOML layouts must be normalized before migration as documented.

See the changelog and documentation for details.

Steadlith 1.0.0

Added

• Added versioned Sphinx documentation for installation, configuration, operations, migrations, JSON output, and the public Python API.
• Added a production website smoke check for key routes, metadata, headers, and assets.
• Added regression coverage for CLI contracts, provider responses and retries, cache pruning, compaction boundaries, migration recovery, and installed entry points.

Changed

• Breaking: status now reads committed index state without scanning source files. The JSON fields index_drift, source_drift, embedding_drift, pending_operations, and pending_embeddings were removed; use plan --json for source and configuration comparison.
• Chunk now rejects non-string content and non-mapping metadata. CDCParams now rejects invalid numeric and boolean fields, unknown keys, and non-CDC strategies.
• Source path metadata now uses project-relative document IDs instead of absolute paths. The first 1.0 plan over a 0.3 index can report metadata-only move operations; applying them does not re-embed unchanged chunks.
• OpenAI requests use Steadlith's retry loop without additional SDK retries.
• The README and website now provide a complete offline quick start and supported deployment and provider scope. The README also provides benchmark commands and support instructions.
• Removed the unused embed_with_cache(..., on_batch=...) callback, duplicate vector serialization code, duplicate website artwork, and an unnecessary development dependency.

Fixed

• State-path collisions now fail with typed errors that name the blocking path and the configuration key to change.
• Configuration and benchmark choice errors now list the accepted values.
• Manifest mirror verification and repair now detect mismatches and preserve a recoverable state after interrupted writes.
• compact --dry-run no longer creates a missing index database.
• Network-consent errors now report the active project configuration instead of assuming the default file.
• The release wheel smoke test now runs in an isolated virtual environment instead of altering the documentation environment.

Steadlith 0.3.0

Full Changelog: v0.2.0...v0.3.0

v0.2.0

Full Changelog: https://github.com/satwiksps/cairn/compare/v0.1.0...v0.2.0

Cairn 0.1.0

Cairn 0.1.0 is the first public release of the incremental indexing toolkit for changing RAG corpora.

Install

python -m pip install cairn-rag

Included

• Rabin content-defined chunking and comparison strategies
• Content-addressed chunk identities, manifests, and Merkle roots
• Cache-aware add, keep, move, and delete planning
• Transactional SQLite indexing with tombstones, verification, and compaction
• Query, migration, rollback, churn, retrieval, and cache workflows
• Offline demo embeddings plus optional OpenAI and sentence-transformers providers

Important limitations

Cairn is alpha software. Public APIs and chunk identities may change before 1.0. The bundled hash embedder is for deterministic testing, not production retrieval. The current stateful TTTD fallback does not guarantee strict fixed-margin locality, and structural snapping remains experimental.

See the README, CLI reference, known limitations, and changelog.