v0.5.0 — Antigravity CLI (agy) compaction continuity

Minor release: a third coding agent joins Claude Code and Codex — Google's Antigravity CLI (agy) — plus the memory-quality work from issues #3 through #9.

Added

• Antigravity CLI (agy) compaction continuity. rag install --agy merges one named hook (SessionStart, PreInvocation, Stop) into ~/.gemini/config/hooks.json with check mode, unique backup, and a target-aware rollback record. SessionStart injects pins, domains, project knowledge, and the checkpoint as an ephemeral message; PreInvocation treats a /compact turn as PreCompact (checkpoint, enrichment, versioned compact prompt), detects an automatic-compaction marker after the fact, and recalls error signatures; Stop stores the /compact summary as the handoff and queues mining. Mining and enrichment understand the Antigravity transcript format. Verified facts about Gemini 3.8 Flash / 3.1 Pro (1,048,576-token window) and Antigravity's hook/compaction behaviour: What’s New in 0.5.0.
• Bounded project context (issue #9), retrieval evidence quality (issue #8), claim evidence and inference status (issue #7), atomic fact validity (issue #6), explicit project scope (issue #5), and the reproducible memory benchmark (issue #3).

Fixed

• Lossless, replay-safe session ingestion (issue #4).

Changed

• Price-aware Codex context policy (350K window, compaction at 250K).

Upgrade

uv sync && uv run rag init-db (migrations 013/014), then uv run rag install --agy --check / uv run rag install --agy for Antigravity. Open a new agy conversation in a trusted workspace and review the agentic-rag hook with /hooks. Automatic compaction has not been observed live yet (BACKLOG 0.4).

v0.4.2 — Enrichment screens values instead of voiding the checkpoint

Patch release for checkpoint enrichment, fixing #2.

Fixed

• Checkpoint enrichment no longer fails on its own subject matter. The enrichment job now screens model output value by value: a transcript-, diff-, or dialogue-shaped value, an evidence item whose claim is not a verbatim part of a digest fragment, or a slug the digest never references is dropped on its own and recorded on the checkpoint as enrichment <field>: N items dropped (<reason>), which the renderer's Warnings: line shows. A credential in the output or a malformed shape still fails the job into last_error, and the store's own gate (validate_enrichment) stays all-or-nothing.
• Word guard exempts paths and identifiers. transcript, diff, and body count only as prose; transcript.py, hooks/transcript, or transcript_delta are facts a checkpoint may state.

625 tests pass. See CHANGELOG.md for the full 0.4.x history.

v0.4.1 — Handoff extraction and tail-preserving truncation

Patch release for the Claude compaction handoff, from a second live /compact smoke on 2026-09-04.

Fixed

• Handoff extraction ignores tags quoted in the prose. The <summary> block of Claude's compact_summary is now bounded by tags on lines of their own (the first <summary> starting a line after the first </analysis> ending one, up to the last </summary> ending a line). In a session that discussed this very mechanism, the previous first-occurrence match stored analysis remainder plus a summary fragment because the prose quoted <summary> and </summary> inline.
• Truncated handoffs keep their tail. Over handoff_max_chars, and over the render budget at SessionStart, the middle of the handoff is cut out instead of its end, so the summary's pending work, current state, and next exact action survive alongside its objective. The …[truncated] marker stands on a line of its own between head and tail.

612 tests pass. See CHANGELOG.md for the full 0.4.x history.

v0.4.0 — Claude compaction continuity

agentic-rag 0.4.0 — Claude compaction continuity

This release brings the session continuity shipped for Codex in 0.3.0 to
Claude Code, bound to Claude's own hook contract, on the same provider-neutral
PostgreSQL + pgvector checkpoint store.

Highlights

• Six Claude lifecycle hooks from the default rag install: PreCompact
  delivers the versioned compact instructions through hook stdout,
  PostCompact stores Claude's compact_summary as a bounded, secret-stripped
  handoff on the checkpoint, SessionStart restores it with an age label, and
  SessionEnd queues the final transcript delta for every Claude reason.
• A managed 1M/500K Claude policy: autoCompactWindow = 500000, a warning
  (never a rewrite) for models without the [1m] suffix, and warnings for
  overriding environment variables.
• SessionStart output capped at Claude's 10,000-character hook limit, with
  the checkpoint shortened into the remaining budget before any section is
  dropped and every cut announced.
• A previewable, recoverable Claude install: rag install --check, unique
  settings.json backups, a mode-0600 rollback record, and a target-aware
  rag install --restore.
• Migration 008 (handoff, handoff_at) and checkpoint handoff: freshness
  in rag status.
• Client detection is payload-driven; Codex behavior is unchanged. The Codex
  sources were checked to confirm PostCompact completes before
  SessionStart(source="compact") runs.

Upgrade

uv sync
uv run rag init-db
uv run rag install --check
uv run rag install

Then, in Claude Code, review the six handlers with /hooks, confirm
/autocompact shows 500000 tokens from settings, and verify with
uv run rag status. Codex users: start a fresh Codex session after
rag install --codex; Codex discovers handlers once per session.

Read What’s New in 0.4.0 for the complete tour.

Verification

• 610 tests passed against PostgreSQL.
• Wheel and sdist passed twine check.
• Migration 008, the installer, and a manual /compact were exercised on a
  real macOS deployment; the two defects that smoke exposed were fixed before
  this release.

v0.3.0 — Codex continuity and native memories

agentic-rag 0.3.0 — Codex continuity

This release adds durable session continuity across Codex context compaction
while keeping PostgreSQL + pgvector as the canonical, audited knowledge store.

Highlights

• Audited continuation checkpoints across PreCompact, PostCompact,
  SessionStart, and SessionEnd.
• A versioned global compact prompt and managed 600K/500K Codex context policy.
• Complementary native Codex memories with Luna extraction and consolidation.
• Provider-neutral Codex/Claude mining, curation, and checkpoint enrichment.
• A lossless, idempotent Codex installer with preview, isolated validation,
  foreign-hook preservation, backups, and conflict-safe rollback.
• Provider outage circuit breaking, bounded retry, and health visibility.
• Secret stripping and evidence validation at provider and persistence
  boundaries.
• A patched cryptography 50.0.1 lock (CVE-2026-69247 is fixed from 50.0.0).

Upgrade

uv sync
uv run rag init-db
uv run rag install --codex --check
uv run rag install --codex

After installation, inspect and trust the six agentic-rag handlers through
/hooks, then verify with uv run rag status.

The 600K/500K long-context policy is configurable and can increase latency and
provider cost for large requests. Measure it against your workload.

Read What’s New in 0.3.0 for the complete tour.

Verification

• 557 tests passed against PostgreSQL.
• Wheel and sdist passed twine check.
• The Wheel was installed in an isolated environment and verified to contain
  the CLI, all seven migrations, and the compact prompt.

SHA-256

• a93a55d10789cb59ebd03033bbced43e149ceee05ebb0995e0b9b768e28bcdda — agentic_rag-0.3.0-py3-none-any.whl
• eccacb578196df9fd2e5d4b6310f4590d8ebff5b78390ccebc8ba80e9ab0e0a9 — agentic_rag-0.3.0.tar.gz

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

v0.2.0 — cross-platform (Windows) portability hardening

Cross-platform (Windows) portability hardening. These changes remove the hard blockers to importing and running the core on non-POSIX platforms and are verified on macOS/Linux. Native Windows is not yet verified end-to-end — background-job scheduling (Task Scheduler) and a real-Windows test pass remain.

Fixed

• Cross-platform worker lock. The single-writer singleton no longer does an unconditional top-level import fcntl, which made agentic_rag.worker unimportable on Windows. The non-blocking exclusive lock is now dispatched by platform — fcntl.flock on POSIX, msvcrt.locking on Windows — behind an identical "return a held handle or None, never crash" contract.
• UTF-8 subprocess decoding. Output of the claude CLI is now decoded as strict UTF-8, so non-ASCII content (e.g. German umlauts) is no longer mangled on platforms whose default text encoding is not UTF-8 (Windows cp1252). A non-UTF-8 byte fails loudly rather than being silently replaced.
• PATH resolution of the claude CLI. The LLM seam and MCP registration resolve the CLI through shutil.which before spawning, so a bare claude also resolves where the executable is a claude.cmd shim (Windows).

See docs/03-quick-start.md → Platform support for how to run on WSL2 or native Windows.

Full Changelog: v0.1.0...v0.2.0

v0.1.0 — initial public release

First public release of agentic-rag: a lean, local-first agentic-RAG long-term memory for Claude Code, backed by PostgreSQL + pgvector.

• Hybrid search (multilingual vector recall via bge-m3 + English/German keyword full-text) over a curated knowledge graph.
• Automatic session mining: a single-writer worker turns your Claude Code transcripts into durable, deduped memories via claude -p.
• Auth-agnostic: runs on your Claude subscription or an API key; embeddings are always local (Ollama). Secret-stripping write gateway; role-based destruction protection; restore-tested backups.
• Generic migrate importer for existing llm-wiki stores; knowledge domains you grow and curate.
• Cross-platform (macOS launchd; Linux cron/systemd), full docs/ handbook.

See CHANGELOG.md for details.