commit-capture

name	commit-capture
description	Captures conversation context around git commits to Obsidian vault. Fires automatically via PostToolUse hook.
version	2.2.0

Commit Capture

The value here is not commit metadata — git already has that. The value is the conversation context: what you were investigating, what you tried, what you decided, and why.

Architecture

Detection and metadata extraction are handled by scripts/commit-capture.sh (shell script, zero AI cost). This skill is invoked when the hook outputs a line starting with obsidian-commit-capture: — all metadata is inline, no file read needed.

The actual write goes through the keeper CLI (scripts/keeper append) — the keeper's deterministic write primitive. The CLI creates the file on absence, makes parent dirs, and appends the section; this skill does not Read/Write/ mkdir directly. The CLI is a plain command (zero AI cost, no subagent), so commit-capture stays as cheap as before — the keeper is used, via command. This skill's only job is to synthesize the context bullets (which require the conversation, so they must be built here in the agent turn) and hand the write to the CLI.

Config

Vault path is passed inline by the hook (parsed once from the resolved config — $OBSIDIAN_LOCAL_MD, else ${XDG_CONFIG_HOME:-$HOME/.config}/claude-obsidian/obsidian.local.md, else the plugin dir). Do not Read the config file from this skill — every read costs ~1-2K tokens per commit. If vault_path is empty in the hook output, skip silently (the config is missing or malformed).

Steps

1. Parse inline metadata from the hook output line: obsidian-commit-capture: hash=<h> | msg=<m> | branch=<b> | files=<f> | org_repo=<o> | repo_name=<r> | ticket=<t> | date=<d> | time=<ti> | vault_path=<v> Extract: hash, msg, branch, files, org_repo, repo_name, ticket, date, time, vault_path.

2. Target (relative to vault_path): Projects/Development/<org_repo>/<date>.md.

3. Build the session context — this is the primary output and the only part that needs the conversation (so it must be done here, in the agent turn).

   Review the full conversation since the last commit (or session start if first commit). Capture:

   • Goal — what task or problem was being worked on
   • Investigation — what was explored, what files were read, what was searched for
   • Decisions — choices made, alternatives considered and rejected, tradeoffs
   • Debugging — if applicable: symptoms, hypotheses tested, root cause found
   • Loose ends — anything unresolved, flagged for follow-up, or noted for later

   3-8 bullet points. Dense and specific. Written so you can reconstruct the reasoning months later.

4. Write the body to a temp file — the section content the keeper will append:

   **Branch:** <branch>
   **Message:** <msg>
   **Files:** <files>
   
   ### Context
   
   - <bullet points from step 3>
   
   ---
   

5. Write the new-file header to a second temp file — used by the keeper only if the dated note doesn't exist yet:

   ---
   date: <date>
   repo: <org_repo>
   tags: [<repo_name>, auto-captured]
   source: claude-code
   ---
   
   # <repo_name> — <date>
   

   If ticket is non-empty, add ticket-<ticket> to the tags array.

6. Hand the write to the keeper CLI:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/keeper" append \
     --vault "<vault_path>" \
     --target "Projects/Development/<org_repo>/<date>.md" \
     --section "## <time> — <hash>" \
     --body-file "<body-temp>" \
     --init-file "<header-temp>"
   

   The CLI creates parent dirs, writes the header on first commit of the day, and appends the section. No Read/Write/mkdir, no subagent.

7. Confirm silently — output only: Captured <hash> → <org_repo>/<date>.md

Important

• Do NOT open the note in Obsidian GUI
• Do NOT modify the daily note
• The context section is the whole point — never skip it
• If the hook output line is missing or malformed, skip silently