obsidian-second-brain

Maintain an Open Knowledge Format (OKF) v0.2 wiki in an Obsidian vault from Claude Code or Pi, without moving the agent out of its current repository. The package supplies a wiki skill, destination-based file protection, lifecycle maintenance, and local retrieval tools.

Obsidian Git owns commits, pulls, and pushes. Install and enable it for the configured workflow; the agent integration never performs those operations. Markdown, local retrieval, and page maintenance remain usable while Obsidian is closed. Automatic sync requires the running app and plugin.

The vault remains yours. Setup previews proposed changes before confirmation, preserves existing notes and unrelated settings, and refuses conflicts. See NOTICE.md and LICENSE for attribution and license terms.

Requirements

• Claude Code or Pi; Python 3.14 or newer with PyYAML; markdown-it-py for evidence reports, outlines, and section-selected reads; jq for lifecycle commands.
• Node.js 18 or newer for the package CLI launcher. Use a supported Node.js LTS release; Node 18 is end-of-life. Pi has its own Node.js requirement, which can be higher than this package minimum.
• Obsidian with the owner-installed Obsidian Git plugin for version control. Setup validates its installed manifest and enabled-plugin list before applying the configuration profile; it does not download plugin binaries.
• Git for the vault repository and package installation. Repository creation, remote selection, identity, and authentication remain owner-controlled.
• A repository checkout or installed package directory for setup commands. Installing the package does not add a setup CLI to the shell path.

Ensure python3 --version reports Python 3.14 or newer. Install the Python runtime dependencies from the active package checkout (or unpacked package):

python3 -m venv "$HOME/.venvs/obsidian-second-brain"
. "$HOME/.venvs/obsidian-second-brain/bin/activate"
python3 -m pip install -r requirements.txt

markdown-it-py is used by evidence-report, outline, and section-selected read commands. Doctor, search, and explicit-span reads do not require it. If a parser-dependent command cannot import it, the CLI prints an installation instruction instead of silently using incomplete Markdown scanning.

On Linux, install your distribution's python3-venv package if necessary. Keep the environment active when running Python commands.

Install the agent package

For Pi, install the published npm release:

pi install npm:@soulwaxx/obsidian-second-brain

To track unreleased source instead, use pi install git:github.com/soulwaxx/obsidian-second-brain.

For Claude Code:

claude plugin marketplace add soulwaxx/obsidian-second-brain
claude plugin install obsidian-second-brain@obsidian-second-brain --scope user

The GitHub marketplace is a catalog; Claude fetches the plugin from the same versioned npm artifact as Pi. The catalog deliberately does not pin a plugin version. Claude Code and npm must support npm plugin sources. Existing Git-backed installations retain the same plugin identity. Refresh the catalog and update the plugin once to migrate:

claude plugin marketplace update obsidian-second-brain
claude plugin update obsidian-second-brain@obsidian-second-brain --scope user

Restart Claude Code after updating. For unpublished local development, load the checkout with claude --plugin-dir /absolute/path/to/obsidian-second-brain; adding its marketplace still selects the published npm plugin.

Use an existing checkout or create one for the following setup examples:

git clone https://github.com/soulwaxx/obsidian-second-brain.git
cd obsidian-second-brain
PLUGIN_DIR="$(pwd -P)"
VAULT="$HOME/Notes/MyVault" # Replace with the selected absolute vault path.

Scaffold, then configure

First preview the minimal wiki scaffold:

python3 "$PLUGIN_DIR/scripts/bootstrap-vault.py" --vault "$VAULT"

Review the plan, then use its exact hash:

python3 "$PLUGIN_DIR/scripts/bootstrap-vault.py" --vault "$VAULT" --apply --confirm 'PASTE_SCAFFOLD_HASH_HERE'

This creates missing .obsidian/, wiki/quickstart.md, and generated navigation. It preserves existing notes, logs, and settings. Scaffolding can precede plugin installation; it does not yet create or migrate the agent integration config.

Open the directory as an Obsidian vault, deliberately create/use its Git repository, and install/enable Obsidian Git. Review Git setup before the first backup. Then preview the supported configuration profile:

python3 "$PLUGIN_DIR/scripts/bootstrap-vault.py" --vault "$VAULT" --configure

Review the selected configuration/plugin differences and ignore additions, then confirm this separate plan:

python3 "$PLUGIN_DIR/scripts/bootstrap-vault.py" --vault "$VAULT" --configure --apply --confirm 'PASTE_CONFIGURE_HASH_HERE'

The reviewed configuration selects the vault and sets the deprecated agent features.autoCommit field to false. It preserves unrelated settings, existing plugin schedules, and an existing choice to disable pushing. Disabling plugin commit squashing is an explicit proposal, not a hidden reset. Private settings/preimages are bound to the plan hash but not dumped in the preview.

The default config is ~/.config/obsidian-second-brain/properties.json. OBSIDIAN_AGENT_CONFIG selects another file; OBSIDIAN_VAULT_PATH selects a vault for an invocation. Setup never silently retargets a config pointing to a different vault. Re-preview when inspected files change. See the setup guide for migration and prerequisite diagnostics.

Work from any folder

Start the agent in the repository or folder where the main task belongs. Use /obsidian-second-brain:wiki in Claude Code or /skill:wiki in Pi as the general entry point. Focused intents are available as /obsidian-second-brain:wiki-query, /obsidian-second-brain:wiki-save, /obsidian-second-brain:wiki-ingest, /obsidian-second-brain:wiki-research, and /obsidian-second-brain:wiki-health in Claude Code, or /skill:wiki-query, /skill:wiki-save, /skill:wiki-ingest, /skill:wiki-research, and /skill:wiki-health in Pi. Query and health are read-only; save only files an explicitly selected conversation insight; ingest/research do not imply raw source capture or automatic filing. Every intent follows the shared wiki write rules. The skill resolves the configured vault and uses absolute runtime paths when necessary. Ordinary relative coding paths keep their current-directory meaning. Larger coordinated wiki edits may optionally use the reviewed recoverable batch CLI; small edits retain the normal lifecycle. An explicitly selected local UTF-8 text source may also be preserved through a separate reviewed, immutable capture into .raw/agent-captures/; selected sources can optionally use reviewed byte observations and page reconciliation baselines without archiving source content. Neither is mandatory, and capture does not provide web/media extraction or authorize wiki filing. See the focused workflow reference and baseline contract for approval and recovery details.

Startup supplies a small vault locator rather than injecting the entire personal index into unrelated work. The skill searches and reads relevant pages when a task needs them. An approved ingest or research pass may update relevant canonical pages; ordinary questions remain read-only unless filing was requested. Keep one wiki writer at a time.

File-tool guards apply to selected-vault destinations regardless of agent cwd. They protect generated index.md and lifecycle-owned log.md at every wiki depth, curated sources, settings, and unsafe paths. Shell writes are not sandboxed; follow the skill's explicit guard/validation workflow if the file tool lifecycle is unavailable.

The lifecycle validates changed pages and finalizes writing batches with navigation, logging, and one optional BM25 refresh per validated changed batch. A failed refresh leaves a durable retrieval-only retry; content/navigation/log publication and search freshness are reported separately. A permitted writer finalization or explicit BM25 build can settle it without duplicating the log. With features.retrievalRefresh: false, no automatic build occurs. Outside-lifecycle Obsidian/manual/shell edits are diagnosed as stale, not automatically maintained. Failed changed pages are retained for repair; an unrelated historical defect is not a vault-wide veto. Unrelated startup/read-only shutdown does not regenerate the vault. Obsidian Git handles version control independently, including for legacy configs with autoCommit: true.

Query and diagnose

Pi also registers four direct, read-only wiki_search, wiki_outline, wiki_read, and wiki_doctor tools. They invoke the currently installed package's Python retrieval/read operations and do not inject wiki content into prompts; see the Pi tool contract.

When the obsidian-second-brain bin is on PATH, use the package-owned CLI for read-only structured search and doctor commands from any folder:

obsidian-second-brain search "query terms" --json
obsidian-second-brain doctor --json
obsidian-second-brain evidence-report --json
obsidian-second-brain source-check --root-id research-notes --source-root "/absolute/path/to/selected sources" --json

Pi and Claude plugin installation does not necessarily put that npm bin on shell PATH. If command -v obsidian-second-brain fails, do not install a second copy: invoke the active package's Node launcher using its package-relative path. The shared workflow reference shows how to derive an absolute package root from Claude's loaded skill base directory or Pi's loaded skill location; this works from any caller cwd and selects the version already loaded by that client.

Search uses the selected vault and does not rebuild the index; doctor reports configuration, dependencies, vault, cache-exclusion, lifecycle, and retrieval status without repairing them. evidence-report is also read-only; optional source/claim records are not migrated or required, and its structural, freshness, and relationship output does not establish factual truth or source verification. source-check is a separate, explicitly selected, read-only comparison of recorded local sources; it reports review signals, not truth or reconciliation. See the source-check contract. Citation scanning follows markdown-it-py's CommonMark parsing; it does not enable Obsidian-specific Markdown extensions. Use the wiki's generated index hierarchy when search is unavailable or has no useful result. Focused workflow limits and selected-conversation save rules are documented in the shared workflow reference.

For an explicitly requested Obsidian compatibility audit, run the separate read-only scripts/obsidian_compat.py report. It keeps OKF resolution unchanged and labels filesystem results separately from native resolver evidence. Reviewed link migration covers authored Markdown batches, opt-in lifecycle-owned historical destinations, and final graph checks. Neither diagnostics nor normal finalization implicitly migrate old history.

Local retrieval

Retrieval ships with the package; no vault-local helper installation, Ollama, or network service is required. From any folder:

python3 "$PLUGIN_DIR/scripts/bm25-index.py" build --vault "$VAULT"
python3 "$PLUGIN_DIR/scripts/retrieve.py" "query terms" --vault "$VAULT"

The derived cache lives under .vault-meta/retrieval/. Keep it and .vault-meta/lifecycle/ effectively ignored in Git. Lifecycle refresh uses package-owned scripts; a missing/corrupt cache falls back to generated index navigation. Vault-local helper copying is no longer supported; use the active package's scripts instead. Package updates never delete existing vault-local files.

Current retrieval conservatively masks clear single-line Markdown strikethrough outside literal/code/link contexts; authored history and original source offsets remain unchanged. ISO-date queries use the historical lexical view. Unmarked contradictions still require human consolidation. Search/doctor never rebuild old caches: explicitly rebuild BM25 and optional semantic chunks for the new projection/input scheme before assuming current coverage.

Document embedding builds are manual and input-incremental: compatible exact raw-unit inputs reuse verified vectors across edits/moves/duplicates; only missing inputs are sent to the installed model. Chunk pooling/boundaries and RRF remain unchanged. Legacy pooled caches cannot recover individual unit vectors; their first explicit compatible build populates the input cache. Builds report real input/character counts, duration and cache size. Cached semantic searches embed only the query, never missing documents or the legacy page-head shortlist.

The read-only evaluator uses privately reviewed paired EN/IT gold and prepared disposable snapshots. Its live quality results are separate from deterministic CI and do not authorize a model/fusion change.

Embedding operations default to qwen3-embedding:4b; --ollama-model selects another already-installed model. Changing models does not make old-model cache vectors compatible: rebuild semantic chunks explicitly after the selected model is available. Embedding reranking is a separate opt-in. Install/start Ollama and explicitly download the model before querying:

ollama pull qwen3-embedding:4b
python3 "$PLUGIN_DIR/scripts/retrieve.py" "query terms" --vault "$VAULT" --rerank

Normal retrieval makes no Ollama request. Reranking accepts loopback endpoints, ignores proxies, rejects redirects, and falls back to BM25 on failure. Exact topic-title matches receive a ranking boost; ISO-date queries stay lexical and prefer matching dated notes/relevant dated sections. See retrieval details.

Optional specialist

Claude Code supports @agent-obsidian-second-brain:wiki-vault. For Pi, optionally install pi-subagents with pi install npm:pi-subagents, then ask the parent to run obsidian-second-brain.wiki-vault, or use /run obsidian-second-brain.wiki-vault "your task". The specialist resolves the same configured vault; it does not require the parent to start there.

Neither packaged specialist pins a model or effort. Claude Code 2.1.292+ adds /obsidian-second-brain:wiki-agent "your task": select Vault agent effort in the plugin's /config settings to request low for that explicit delegation without changing parent effort. The default is inherit; direct agent mentions bypass the preference. The plugin's option picker requires Claude Code 2.1.271+. Pi uses its existing per-agent thinking override instead of a package setting. See agent configuration for precedence, model restrictions, and unverified live-client checks. Pi without the subagent runner still supports the skill and lifecycle.

Verify the package

Run from the repository checkout:

npm run release:check
npm test
npm run test:release
npm run test:package
npm run test:pi-runtime -- --require-runtime
claude plugin validate .
bash tests/install-smoke.sh
bash -n hooks/obsidian-session.sh tests/install-smoke.sh
shellcheck tests/install-smoke.sh

Installation tests use temporary agent settings and a disposable vault, not the real profile, and do not send a model prompt. CI checks macOS/Linux with Node 26 and Python 3.14/3.15. Use Conventional Commit PR titles: feat: releases a minor version, a breaking change a major version, and other types a patch. Full code checks run before merge; title/body edits run only metadata validation. After a squash merge, publishing verifies that the merged tree passed PR CI, smoke-tests the prepared release, and uses npm OIDC to publish with provenance. There is no repeated main test matrix or manual client-verification gate.

Require CI passed and PR title passed in the main ruleset. Publishing refuses direct pushes and mismatched or expired verification evidence. See CI/CD maintenance for rollout, locked release tools, the temporary critical-blocking audit policy, and recovery of partially published tags.