A local-first knowledge skill for Codex CLI and Claude Code. It distills completed coding-agent sessions into durable Obsidian notes without copying raw transcripts, tool output, hidden reasoning or system prompts.
It supports Codex and Claude Code equally; OpenCode ingestion is included as an optional third source.
- incrementally reads local Codex CLI and Claude Code session stores;
- keeps durable user requests and outcome-oriented assistant text;
- redacts common credentials and secret-bearing URLs;
- writes provenance-rich Markdown into an existing Obsidian taxonomy;
- provides evidence-oriented search, coverage status and guarded pruning;
- leaves the original session stores untouched.
git clone https://github.com/aalvsz/agent-knowledge-vault.git
mkdir -p ~/.codex/skills
ln -s "$(pwd)/agent-knowledge-vault/skill/agent-knowledge-vault" \
~/.codex/skills/agent-knowledge-vaultmkdir -p ~/.claude/skills
ln -s "$(pwd)/agent-knowledge-vault/skill/agent-knowledge-vault" \
~/.claude/skills/agent-knowledge-vaultThe skill uses the same SKILL.md workflow in either client. Session ingestion
is controlled explicitly with --provider codex and --provider claude.
python3 skill/agent-knowledge-vault/scripts/knowledge.py doctor
python3 skill/agent-knowledge-vault/scripts/knowledge.py sync \
--provider codex --provider claude
python3 skill/agent-knowledge-vault/scripts/knowledge.py search "deployment decision" --json
python3 skill/agent-knowledge-vault/scripts/knowledge.py status
python3 skill/agent-knowledge-vault/scripts/knowledge.py prune --dry-runDefaults can be overridden with --vault, --state, --codex-root and
--claude-root. Run --help for the complete interface.
Redaction is defense in depth, not a guarantee. Review generated notes before syncing an Obsidian vault to a third-party service. The project intentionally does not include session data, authentication material or machine-local state.
The engine uses the Python standard library. Tests require pytest:
python3 -m pip install pytest
pytest -qLicensed under MIT.