README file from
GithubSkill Insight
A desktop-only Obsidian plugin that turns your AI agent sessions into an insight dashboard: which skills you use, how often, in which sessions, and how your skills evolve over time.
What it does
-
Auto-initialization — on startup the plugin scans your configured skill directories (
.agents/skills,.claude/skills,.codex/skillsin your vault) and creates a dashboard card for every installed skill, with no confirmation dialogs. -
Usage mining from session transcripts — on first open it back-fills usage records from the last 7 days of local agent session transcripts (Codex CLI, Claude Code); new skills get their history re-scanned, so usage that happened before the card existed is counted retroactively. After that, it ingests new transcript lines incrementally each time the dashboard opens.
-
Signal detection — counts explicit
Skilltool calls, plus commands that directly readSKILL.mdfiles (each distinct skill counted once per session). Patch-write operations and scripted bulk scans (6+ skill paths in one command) are ignored; directory listings, command outputs and chat prose never participate in matching. -
Insight dashboard with five tabs:
- Overview — usage ranking, an 8-week activity heatmap (click a cell to drill into that day), session volume, and revision activity.
- Skill detail — per-skill timeline of usage and revisions.
- Sessions — usage grouped by agent session, shown as searchable cards with clickable skill chips.
- Revision timeline — how your skills changed over time (r0, r1, r2 …).
- Sources & ingest — read-only discovery report and ingest controls.
-
Settings tab — a settings entry in Obsidian's community-plugins list, with a dashboard shortcut and transcript ingest configuration (global fallback toggles and roots).
Privacy & data safety
- No vault writes. All plugin data lives inside the plugin's own folder (
data.jsonin.obsidian/plugins/skill-insight/). Your notes are never touched. - No network. Everything is parsed locally; the plugin never calls out to the internet.
- Scope by default. Usage mining reads only sessions registered by the Claudian plugin for the current vault, plus sessions whose own metadata (
cwd/ project directory) proves they were started in this vault. A global transcript scan is available but off by default.
How it finds your sessions
| Agent | Source |
|---|---|
| Codex CLI | ~/.codex/sessions/**/rollout-*.jsonl (.zst-compressed archives are skipped) |
| Claude Code | ~/.claude/projects/<encoded-vault-path>/*.jsonl |
| Claudian registry | .claudian/sessions/conv-*.meta.json in the vault, used to map sessions to this vault and to provider backends |
Installation
From the community plugin store (after approval)
- Open Settings → Community plugins.
- Turn off restricted mode, click Browse, search for "Skill Insight" and install.
Manual / BRAT (before approval)
- Install BRAT.
- Run the command BRAT: Add a beta plugin and enter
minybear/skill-insight.
Or copy main.js, styles.css, manifest.json into <vault>/.obsidian/plugins/skill-insight/ and reload.
Limitations
- Desktop only (requires Node.js filesystem access).
- Only Codex CLI and Claude Code transcripts are mined today; other Claudian backends are on the roadmap.
- Transcripts that the agent CLI has already compressed to
.zstarchives cannot be parsed and are skipped.
License
MIT © Awei (minybear)