vault-tasks

Markdown-file task manager for solo devs building with Claude Code.

Tasks are plain markdown files with YAML frontmatter. They live in your repo (or Obsidian vault), are version-controlled with git, and are readable by humans and LLMs alike. Zero runtime dependencies.

Install

npm install -g vault-tasks

Or use without installing:

npx vault-tasks <command>

Requires Node.js >= 20.

Python projects (pip / uv)

If your project's lockfile is Python-based and you'd rather pin vault-tasks alongside your other dev dependencies:

pip install vault-tasks
# or
uv add --dev vault-tasks

The pip wheel is small (~100 KB) and bundles the same compiled JS that ships on npm. It still requires node (>= 20) on PATH — the wheel does not vendor a Node.js runtime.

Quick start

# Initialize in your repo/vault root
vt init

# Create a task
vt new "Fix login redirect bug" --priority high --tags auth,bug

# See what's open
vt list

# Start working on it
vt start 1

# Mark it done (auto-archives by default)
vt done 1

vt init creates a .vault-tasks.toml config and a backlog/ directory. Tasks are markdown files with ULID prefixes (e.g., 01HYX3KQPD7NG8RRGSSFQ9XNHY-fix-login-redirect-bug.md). Sequential numeric IDs (0001-...) are also supported via config.

Commands

| Command | Description | |---|---| | vt new <title> | Create a task. Options: --priority, --tags, --source, --commit | | vt list | List open tasks. Options: --status, --priority, --tag, --all | | vt search <keyword> | Search titles and body text. --all includes archived. --mode keyword\|bm25\|semantic\|hybrid; --like <id> finds similar tasks; --limit N caps results. See Semantic search | | vt show <id> | Print full task file | | vt start <id> | Set status to in-progress | | vt done <id> | Set status to done (auto-archives) | | vt edit <id> | Update fields: --status, --priority, --tags | | vt stale | List open tasks older than 14 days. --days to customize | | vt archive | Move all completed tasks to the archive directory | | vt tags | List all tags and their counts | | vt lint | Audit the vault: broken wikilinks, orphan evergreens, stale refs, drift. --only, --scope, --json, --quiet, --no-suggestions | | vt init | Initialize config and backlog directory | | vt install-skills | Install Claude Code skills and rules. --install, --list, --update |

Task lookup (<id>) accepts a ULID prefix (e.g., vt done 01HYX), a numeric ID for sequential vaults (e.g., vt done 1), or a substring match against the filename (e.g., vt done login).

Task format

Each task is a markdown file with YAML frontmatter:

---
title: "Fix login redirect bug"
status: open
priority: high
tags:
  - auth
  - bug
created: 2026-04-02
source: "[[2026-04-02 Session Log]]"
---

# Fix login redirect bug

After OAuth callback, users are redirected to `/` instead of the page they came from.

• Status: open, in-progress, done, wont-do
• Priority: high, medium, low
• Tags: freeform, filterable via vt list --tag
• Source: where the task was noticed (supports [[wikilinks]])

Extra frontmatter fields (e.g. due, assignee) are preserved through all operations.

Configuration

vt init creates .vault-tasks.toml at your vault root:

[paths]
backlog_dir = "backlog"           # where task files live
archive_dir = "archive"           # relative to backlog_dir
# journal_dir = "journal"         # build logs and session notes
# projects_dir = "projects"       # project folders with CONTEXT.md
# evergreen_dir = "evergreen"     # evergreen/zettelkasten notes

[task]
# statuses = ["open", "in-progress", "done", "wont-do"]
# priorities = ["high", "medium", "low"]
# default_priority = "medium"
# default_status = "open"
# archive_statuses = ["done", "wont-do"]
# auto_archive = true

[id]
# strategy = "ulid"               # "ulid" | "sequential" | "timestamp"
# pad_width = 4                   # zero-pad width (only used with sequential)

[slugify]
# max_length = 60

The config file is discovered by walking up from the current directory, so it works from any subdirectory.

Claude Code skills

vault-tasks ships with skill templates that teach Claude Code how to work with your task backlog:

vt install-skills --install

This installs into .claude/skills/ and .claude/rules/:

| Skill | What it does | |---|---| | /brief | Pre-session briefing: open tasks, last session context, stale threads | | /build-log | End-of-session log: what was built, learned, decided. Extracts tasks | | /weekly-review | Consolidates journal entries, creates evergreen notes, triages backlog | | /task | Quick task creation/management from within a session | | /lint | Vault health check: broken wikilinks, orphans, stale refs, convention drift |

Skills reference configurable vault paths (journal_dir, projects_dir, evergreen_dir) which are substituted from your .vault-tasks.toml at install time. Customize any skill by creating a SKILL.local.md next to the installed SKILL.md -- local files are never overwritten.

An Obsidian Bases dashboard (backlog.base) is also installed for visual task management.

Library API

vault-tasks also exports a programmatic API:

import { loadConfig, TaskStore } from "vault-tasks";

const config = loadConfig();
const store = new TaskStore(config);

// Create
const task = store.create({ title: "My task", priority: "high", tags: ["api"] });

// Query
const all = store.loadAll();
const results = store.search("login");
const stale = store.stale(14);
const tags = store.allTags();

// Update
store.update(task, { status: "done", priority: "low", tags: ["api", "shipped"] });
store.archiveCompleted();

Exports

• TaskStore -- all CRUD operations
• loadConfig / findConfigFile -- config discovery and parsing
• parseFrontmatter / writeFrontmatter -- YAML frontmatter utilities
• slugify -- title to kebab-case filename
• lintVault -- run all lint checks; returns a LintReport
• Lower-level lint primitives: buildIndex, resolveTarget, collectWikilinks, findBrokenLinks, findOrphanEvergreens, findStaleReferences, findEvergreenDrift, attachSuggestions
• Types: Task, CreateTaskOpts, Config, LintReport, LintOptions, WikiLink, VaultFile, BrokenEntry, Suggestion, SearchHit, SearchMode, SearchOptions

Ranked search (vault-tasks/search)

A separate, zero-dependency subpath export provides BM25-ranked search and task-to-task similarity. Imported only when needed -- the core vault-tasks entry point is unaffected.

import { TaskStore, loadConfig } from "vault-tasks";
import { searchTasks, similarTasks, BM25Index, VectorIndex, tokenize } from "vault-tasks/search";

const store = new TaskStore(loadConfig());

// Free-text query, ranked by BM25
const hits = await searchTasks(store, "auth callback", { mode: "bm25", limit: 10 });

// Vector similarity (uses the configured embedding engine)
const semantic = await searchTasks(store, "login flow", { mode: "semantic" });

// Tasks similar to a given task
const target = store.findIncludingArchive("0042");
const related = await similarTasks(store, target, { mode: "hybrid" });

Available modes:

| Mode | Ranking | Needs an engine? | |---|---|---| | keyword (default) | substring match across title, body, AND tags; priority-sorted | no | | bm25 | BM25 relevance, title-weighted | no | | semantic | cosine similarity over embeddings | yes | | hybrid | bm25 + semantic fused with Reciprocal Rank Fusion | yes |

Semantic search

semantic and hybrid modes rank tasks by meaning rather than shared words. They need an embedding engine to turn text into vectors; keyword and bm25 never do and stay fully offline.

Local-first by default — no data leaves your machine, no API key. vault-tasks talks to a local Ollama server:

ollama serve                      # start the local server
ollama pull nomic-embed-text      # pull the default model (768 dims)
vt search "auth redirect" --mode semantic
vt search --like 0042 --mode hybrid

Vectors are cached at <vault>/.vault-tasks/embeddings.json, so only new or changed tasks are re-embedded on later runs. (Add that path to .gitignore if you don't want to commit it.)

Configure the engine under [search] in .vault-tasks.toml:

[search]
embedding_provider = "ollama"          # ollama | lmstudio | llamacpp | openai-compatible
                                       # | transformers | openai | voyage | gemini
embedding_model = "nomic-embed-text"   # also: mxbai-embed-large (1024d), all-minilm (384d)
# embedding_endpoint = ""              # override server URL; empty = provider default

• Other local servers: set embedding_provider to lmstudio or llamacpp (OpenAI-compatible /v1/embeddings); the default endpoints are http://localhost:1234 and http://localhost:8080.
• In-process, no server: install the optional package and switch provider — npm i @huggingface/transformers, then embedding_provider = "transformers" with embedding_model set to a Hugging Face repo id. It's an optionalDependency, so the core install stays dependency-free.
• Cloud (opt-in): embedding_provider = "openai" (or voyage/gemini). The API key is read from an environment variable you name via embedding_api_key_env — it is never stored in config or on disk.

License

MIT