Skip to content
Trove

Trove / Project

Changelog

Trove releases and notable changes: the v0.1 Claude Code plugin, the v0.2.0 Rust engine, and the unreleased v0.3 work merged on main, with dates.

On this page
  1. First principles
  2. Unreleased (on main)
  3. Behavior changes to know about
  4. 2026-10-23
  5. 2026-10-22
  6. 2026-10-16
  7. Documentation and naming
  8. v0.2.0, 2026-07-21
  9. v0.1, 2026-06-13 (untagged)

Trove has one tagged release, v0.2.0. The repository has no CHANGELOG file; this page is built from the GitHub release notes, the merged pull requests (PRs) and the commit history. Dates are merge dates in Coordinated Universal Time (UTC).

First principles

Each version answered one part of the problem in How it works. v0.1 gave Claude Code a history: decisions and conventions as files loaded every session. v0.2.0 added structure: an index of the code retrieved on a token budget. The work on main makes both usable by any agent, not only Claude Code: memory rendered as skills and pointed to from AGENTS.md, a command line write path, cited answers across every source, and redaction at every point where text leaves a trove.

Unreleased (on main)

The README calls this work v0.3, tracked in the epic #19. It is merged on main but not tagged, and the crates and plugin manifest still carry version 0.2.0. Building from a clone gives you all of it.

Behavior changes to know about

  • trove.sh init with no flag now creates a project trove inside a Git repository. It used to create a user trove.
  • --json and --budget are global flags, and --repo can now come before or after the subcommand.
  • The SessionStart hook can now write: it re-renders project trove skills whose entries changed.
  • The trove about line is now “Memory, semantic and context layer for coding agents”.

2026-10-23

  • trove ship subcommand over the shared ship library, with the full trove-ship flag surface. The trove-ship binary stays as an alias with identical behavior. (#38)
  • trove ask redacts the prompt before any provider call. Secrets are always replaced; other private categories are replaced unless --no-redact-nonsecret is given; a secret that survives means nothing is sent. The footer and --json report what was removed. (#37)
  • Wider secret detection: quoted multi-word passphrases and long or spaced key names (signing, encryption, master, session and similar) now count as secret assignments, in the Rust detectors and in trove.sh guard. (#36)
  • The secret detectors moved into a shared trove-redact crate used by ship, remember and ask, with no rule changes. (#35)
  • A retrieval evaluation harness in evals/ with 16 golden queries, a committed fixture and a measured baseline for ask and query (cargo test -p trove-retrieve --test eval -- --nocapture). (#34)

2026-10-22

  • Continuous integration (CI) now checks formatting and lints across the whole workspace and runs plugins/trove/tests/render.test.sh. Added CONTRIBUTING.md and a command line interface (CLI) reference in the README, and marked PRD.md, ENGINEERING.md and VISION.md as historical. (#33)
  • trove remember: a deterministic write path for any agent or script, with type routing, --section, --update, --tags, a body from --body or standard input, and secret refusal. (#32)
  • trove ask, trove import-historical and trove ship now read the sectioned entries of a project trove. Before this, entries under conventions/, memory/ and context/ were invisible to all three. (#31)
  • README refresh, an animated icon and redrawn preview images. (#29, #30)

2026-10-16

  • Project troves loaded as conventions, memory and context skills, with an AGENTS.md block for other agents. trove.sh gained render, context, sync, guard [--redact], route, agents and project, and grep and list cover the sections. The repository root is the Git top level, so the script works from subdirectories. A secret guard runs on every write and render, the SessionStart hook keeps skills in sync, and a new /trove:render skill was added. (#26)
  • Dense CLI: trove ask across every trove, repository docs, the code index and installed tools, with file:line citations and an optional written answer from Anthropic, Ollama or any command; the interactive prompt when trove runs with no arguments; and trove tools for installed command line tools, in a new trove-tools crate. (#27)
  • trove-ship: handoff docs from a project and its trove with private information removed, in a new trove-ship crate and binary, plus the /trove:ship skill. (#28)
  • Website redesign with link previews and a sitemap. (#25)

Documentation and naming

  • Documentation is published at https://velofy.co/trove/. The repository still publishes its own landing page from index.html; the pull request that would have replaced it with a redirect (#16) was closed without merging.
  • 2026-09-12: the marketplace was renamed from anishfyi-trove to velofy-trove, so the install command is now /plugin install trove@velofy-trove. The same day, repository, site and plugin manifest links were repointed to github.com/anishfyi/trove after the repository moved back from the velofy organization; github.com/velofy/trove redirects to it.

v0.2.0, 2026-07-21

Release notes. Trove became a two-layer system: a Rust repository memory engine plus the existing Claude Code plugin.

Engine (new):

  • Repository indexing into .trove/: symbols, module summaries, subsystem clusters and an architecture map.
  • Progressive retrieval, architecture to subsystem to module to symbol, with context budgets.
  • CLI commands: index, status, query, import-historical, record, patch.
  • Languages: Rust, Python, JavaScript/TypeScript, Go, Bash.
  • The L0 to L6 memory hierarchy documented in VISION.md.

Plugin (v0.1 to v0.2):

  • New /trove:index to index the current repository.
  • New /trove:query for progressive retrieval before editing.
  • New SessionEnd hook with a reflection and re-index reminder.
  • /trove:init, /trove:remember and /trove:recall unchanged.

Breaking changes: none for existing plugin users. Engine features require Rust.

Known limitations listed in the release:

  • Module summaries are heuristic, not generated by a large language model (LLM).
  • The semantic layer uses keyword overlap; embeddings are not implemented.
  • The dependency graph covers imports, not full call graphs.
  • No prebuilt binaries; build from source.

Upgrade:

git pull && cargo build -p trove-cli && cargo run -p trove-cli -- index

For the plugin, run /reload-plugins.

v0.1, 2026-06-13 (untagged)

The first version, a Claude Code plugin only. Its manifest carried version 0.1.0; there is no Git tag for it.

  • Skills init, remember and recall, available as /trove:init, /trove:remember and /trove:recall.
  • SessionStart hook that loads INDEX.md from the project and user troves.
  • trove.sh helper with path, init, list and grep.
  • Marketplace anishfyi-trove, installed with /plugin install trove@anishfyi-trove. The v0.2.0 release used the same name.
  • 2026-06-13: the project was first named Codex, then renamed to Trove the same day to avoid a collision with the OpenAI Codex CLI.
  • 2026-06-14: JavaScript Object Notation (JSON) entries supported alongside Markdown; the remember skill picks the format by the shape of the content.