Skip to content

CHANGELOG.md

Latest commit

 

History

History
90 lines (83 loc) · 5.94 KB

CHANGELOG.md

File metadata and controls

90 lines (83 loc) · 5.94 KB
Raw

Changelog

All notable changes to dig are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Added

  • Canary release channel — every push to main publishes bleeding-edge builds of all three artifacts: a rolling canary GitHub prerelease of the cross-compiled CLI (canary.yml; dig --version reports X.Y.Z-canary.<sha>), @vllnt/dig@canary on npm (npm.yml), and dig-client .devN on PyPI (pypi.yml). npm and PyPI keep canary + release in one file each (OIDC trusted publishing, no tokens), gated on CANARY_ENABLED. Provenance/attestations are on now that the repo is public. Stable releases are tag-driven and gated behind maintainer approval — canary mode is the default (see .claude/rules/release.md). Runbook: docs/RELEASING.md.
  • dig retain [file] — the agent-memory capture primitive: writes content (a file argument, stdin, or a rendered session via --transcript) into the KB at a dated, content-addressed memory/ path (--as to override, --date for reproducible captures), then scans + indexes it as a reversible changeset, so dig find/dig recall surface it. Path-escape guarded.
  • Session retention — dig retain --transcript <session.jsonl> renders a Claude Code transcript to readable markdown (user + assistant turns, tool calls summarized; thinking, tool output, system reminders, and injected skill bodies dropped). A SessionEnd plugin hook (hooks/retain-session.sh) auto-captures finished sessions into memory/sessions/ — double opt-in (DIG_RETAIN_SESSIONS=1 and a .dig KB at the session's directory) and fail-open, so it can never block or break a session.
  • dig recall <query> — the agent-memory recall primitive: a token-budgeted (--budget), provenance-tagged context pack ranked from the KB (text or --json), so an agent loads relevant memory without overflowing its context. Snippets land on the query-relevant window of each matched document (not its head), so recalling a long captured session returns the matching exchange.
  • Semantic + hybrid retrieval — opt-in vector index behind a [retrieval] policy, embeddings via any OpenAI-compatible endpoint, stored in .dig/vectors.db as a derived view with a blob-keyed cache. dig find --mode fts|vector|hybrid; hybrid fuses FTS and vector rankings with Reciprocal Rank Fusion. FTS stays the deterministic default.
  • Background semantic indexing — scans queue unseen blobs instantly; dig embed drains the backlog with per-file commits (interruptible, resumable) and dig watch drains it per tick. An unreachable endpoint degrades gracefully and never blocks the deterministic spine.
  • Multilingual / cross-lingual recall — validated via config only (model = "bge-m3"): a query in one language retrieves documents written in another.
  • Benchmark eval harness (tools/eval) — LongMemEval, LoCoMo, and BEAM adapters scoring retrieval through the real pipeline (recall@k, hit@k, NDCG@10, MRR). Full LongMemEval-S: hybrid hit@5 98.0% vs the published 96.6% bar. Scoreboard in docs/evals.md.
  • dig mcp — run dig as a Model Context Protocol server over stdio, exposing the CLI surface (find, recall, drift, log, export read-only; retain to capture into memory; org/reconcile preview-by-default with an apply flag; undo) as tools any MCP client can drive. dig_retain + dig_recall make dig an agent's memory layer over MCP — capture a session, load a budgeted pack back, all reversible.
  • dig serve — localhost HTTP+JSON daemon over the CLI contract (GET /find /recall /drift /log /export, POST /retain /org /reconcile /undo, apply-gated), so apps and SDKs embed dig without shelling out — including dig as a memory layer (/retain captures, /recall loads a budgeted pack). Binds loopback only — never public.
  • @vllnt/dig TypeScript SDK (clients/typescript) — dependency-free typed client over the daemon, incl. recall() / retain() memory methods (typed RecallPack); CI builds + tests it against a real dig serve. A single npm.yml workflow publishes both the canary (push) and the release (tag) via OIDC trusted publishing — no token.
  • dig-client Python SDK (clients/python) — stdlib-only client over the daemon, same surface incl. recall() / retain(); CI-tested against a real dig serve; a PyPI-publish workflow ships it on release (gated on PYPI_TOKEN).
  • Claude Code plugin (.claude-plugin/) — /plugin marketplace add vllnt/dig then /plugin install dig@dig bundles the dig skill + the dig mcp server.
  • AI SDK tools (@vllnt/dig/ai) — digTools(client) returns Vercel AI SDK tool() definitions for the dig surface, so an agent can search/organize a KB and use it as memory via dig_recall + dig_retain (write a decision, recall a budgeted pack later — mutations apply-gated, reversible). ai + zod are optional peer deps.
  • Configurable retrieval primitives — [retrieval] policy gains rrf_k, candidate_factor, chunk_size, chunk_overlap tuning knobs (0 = default, reproducing shipped behavior); changing chunk size/overlap re-embeds the KB.
  • Event sinks — [[event_sink]] policy entries fire on every committed changeset: webhook POSTs the event JSON; exec runs a command (off unless DIG_ALLOW_EXEC_SINKS=1). Sinks observe — a sink failure warns, never rolls back the commit.
  • dig --version — build metadata (version, commit, date).
  • Release tooling — GoReleaser cross-compiles checksummed binaries for linux/darwin/windows × amd64/arm64; a vX.Y.Z tag publishes a GitHub release.

Changed

  • main branch protection hardened (enforce_admins, required CI checks) ahead of going public.