Files
personal-wiki/AGENTS.md
Daniel 6bab904d13 feat: v1.4.0 audit fixes, multi-agent compatibility, Bases dashboard
External audit flagged 8 actionable items against current best practices for
Agent Skills, Claude Code hooks, Obsidian v1.9-v1.12, and JSON Canvas 1.0.
This release addresses all of them and adds multi-agent compatibility files
at zero cost to existing users.

Tier 1, critical fixes:

* wiki/meta/dashboard.base: NEW Obsidian Bases dashboard (native, no plugin
  needed). Replaces Dataview as the primary dashboard. Six views: Recent
  Activity, Seed Pages, Entities Missing Sources, Open Questions,
  Comparisons, Sources.
* wiki/meta/dashboard.md: now embeds dashboard.base. Legacy Dataview queries
  retained as optional fallback for users on Obsidian < 1.9.10.
* README.md and skills/wiki/references/plugins.md: Plugins section
  reorganized to recommend Bases (core, no install) primary, Dataview
  optional/legacy.
* skills/canvas/references/canvas-spec.md: added missing JSON Canvas 1.0
  fields. Group nodes now document background and backgroundStyle (cover,
  ratio, repeat). Edges document fromEnd (default 'none') and toEnd
  (default 'arrow') asymmetric defaults. Hex ID convention noted alongside
  descriptive ID alternative.
* .gitignore: track wiki/meta/dashboard.base explicitly.

Tier 2, important improvements:

* hooks/hooks.json: SessionStart now uses both command type
  ([ -f wiki/hot.md ] && cat ...) and prompt type. Command type is the
  canonical safety check that works in non-vault sessions without erroring.
  Matcher: startup|resume.
* hooks/hooks.json: NEW PostCompact hook re-injects hot cache after context
  compaction (hook-injected context does not survive compaction; CLAUDE.md
  does).
* hooks/hooks.json: PostToolUse auto-commit now guarded by [ -d .git ].
* hooks/README.md: NEW documentation including known plugin-hooks STDOUT bug
  (anthropics/claude-code#10875) and workarounds.
* skills/wiki/references/mcp-setup.md: added Option D (Obsidian CLI) for
  v1.12+. Added warning callout above NODE_TLS_REJECT_UNAUTHORIZED line
  explaining process-wide TLS bypass and recommending Option D as the
  secure alternative.
* skills/wiki-ingest/SKILL.md: documented [!contradiction] custom callout
  CSS dependency on vault-colors.css snippet.
* skills/wiki/references/css-snippets.md: full documentation of all four
  custom callouts (contradiction, gap, key-insight, stale) with built-in
  fallback equivalents.

Tier 3, multi-agent compatibility (low complexity, high reach):

* AGENTS.md: Codex CLI / OpenCode bootstrap.
* GEMINI.md: Gemini CLI / Antigravity bootstrap.
* .cursor/rules/claude-obsidian.mdc: Cursor always-on rules.
* .windsurf/rules/claude-obsidian.md: Windsurf Cascade rules.
* .github/copilot-instructions.md: GitHub Copilot conventions.
* bin/setup-multi-agent.sh: idempotent symlink installer for Codex,
  OpenCode, Gemini, Cursor, Windsurf. Wires up the skills/ directory in
  each agent's expected location.

Style cleanup: scrubbed all em dashes from every skill, hook, doc, and
bootstrap file (249 total replacements across 26 files). Skills now use
periods, commas, and colons throughout for cleaner natural prose.

Version: 1.3.0 to 1.4.0 (aligns plugin.json with GitHub release tag format).

Already resolved in v1.1 (no action needed):
* defuddle, obsidian-bases, obsidian-markdown skills shipped
* URL ingestion, vision ingestion, delta tracking docs
* Multi-depth wiki-query (Quick / Standard / Deep)
* PostToolUse auto-commit hook
* allowed-tools field removed from all SKILL.md files
* All templates already use plural tag/alias forms
* Custom callouts CSS already in vault-colors.css

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-08 17:52:03 +03:00

2.5 KiB

claude-obsidian: Agent Instructions

This repo is a Claude Code plugin and an Obsidian vault that builds persistent, compounding knowledge bases using Andrej Karpathy's LLM Wiki pattern. It works with any AI coding agent that supports the Agent Skills standard, including Codex CLI, OpenCode, and similar.

Originally built for Claude Code, the skills now follow the cross-platform Agent Skills spec. Frontmatter uses only name and description (no Claude-specific extensions).

Skills Discovery

All skills live in skills/<name>/SKILL.md. Codex / OpenCode / other Agent Skills compatible agents will auto-discover them when you symlink the directory:

# Codex CLI
ln -s "$(pwd)/skills" ~/.codex/skills/claude-obsidian

# OpenCode
ln -s "$(pwd)/skills" ~/.opencode/skills/claude-obsidian

Or run the bundled installer:

bash bin/setup-multi-agent.sh

Available Skills

Skill Trigger phrases
wiki /wiki, set up wiki, scaffold vault
wiki-ingest ingest, ingest this url, ingest this image, batch ingest
wiki-query query, what do you know about, query quick:, query deep:
wiki-lint lint the wiki, health check, find orphans
save /save, file this conversation
autoresearch autoresearch, autonomous research loop
canvas /canvas, add to canvas, create canvas
defuddle clean this url, defuddle
obsidian-markdown obsidian syntax, wikilink, callout
obsidian-bases obsidian bases, .base file, dynamic table

Key Conventions

  • Vault root: the directory containing wiki/ and .raw/
  • Hot cache: wiki/hot.md (read at session start, updated at session end)
  • Source documents: .raw/ (immutable: agents never modify these)
  • Generated knowledge: wiki/ (agent-owned, links to sources via wikilinks)
  • Manifest: .raw/.manifest.json tracks ingested sources (delta tracking)

Bootstrap

When the user opens this project for the first time:

  1. Read this file (AGENTS.md) and the project CLAUDE.md for full context
  2. Read skills/wiki/SKILL.md for the orchestration pattern
  3. If wiki/hot.md exists, read it silently to restore recent context
  4. If the user types /wiki (or "set up wiki"), follow the wiki skill's scaffold workflow

Reference