Code Sessions

A VS Code sidebar that gives you a central command center for coding-agent CLI activity: every session from Claude Code (~/.claude/projects/) and Grok Build (~/.grok/sessions/), grouped by source and by day. Tokens, cost, subagents, projects-touched, duration, topic classification, plus the file changes your work produced (knowledge base + project repos) — all in one place. Click a Claude session to resume it inside the official Claude Code extension panel. Click 🔍 View conversation for a per-turn timeline.

Upgrade note. The extension id is now code-sessions (previously claude-sessions, then coder-sessions). On first activation it imports your existing session index + topic classifications from the prior global-storage dir (so you won't trigger a full reclassification), and a one-time shim migrates any settings you customized under the old claudeSessions.* / coderSessions.* keys into the current codeSessions.* namespace — your data and config carry over automatically.

[Activity bar icon] Code Sessions
├─ Sessions
│   ▼ Claude Code — 1,373 sessions
│       ▼ Today — 4 sessions · $812 · 1.2M tok · 🪄7
│           ▼ Research cookie management in Claude search tools  💬1561 · $709 · ⏱26h · 12m ago
│               💬 1561 msgs · $709.20 · 398M tok · 🪄7 · ⏱26h · 12m ago
│               📁 docs, unpolarize, ai/otelo
│               🔍 View conversation       ← opens timeline webview
│               ▶ Resume in Claude         ← opens in Claude Code extension panel
│               📜 Open raw JSONL
│       ▶ Yesterday — 9 sessions · $187 · 14M tok
│       ▶ Last 7 days — 47 sessions · $1,420 · 250M tok
│       ▶ Older — 1,315 sessions
│   ▼ Grok Build — 1,424 sessions
│       ▼ Today — 18 sessions
│           ▶ Coder rebrand + Grok support · model=grok-build · 💬41 · 🛠32
│       ▶ Yesterday — 22 sessions
│       ▶ Older — 1,384 sessions
├─ KB changes
│   ▼ Today — 14 files · 3 commits
│       knowledge/tech/security/fetcher-exposure-unpolarize.md       [M]
│       ...
│   ▶ Yesterday — 6 files · 2 commits
├─ Projects
│   ▼ Today — 26 files · 8 commits
│       ▼ unpolarize (12)
│           backend-py/app/telemetry/header_capture.py   [M]
│           ...
│       ▼ ai/otelo (10)
│           ...
│   ▶ Yesterday — 4 files · 1 commit

What it shows

Sessions pane

A tree of every Claude Code session (read from ~/.claude/projects/*/), grouped by day:

• Today / Yesterday / Last 7 days / Older buckets, each labelled with totals (session count, total cost, total tokens, total subagents).
• Each session shows on a single line:
  • Title (auto-generated; falls back to the first user prompt)
  • 💬 messages · $cost · ⏱ session-span · time-ago (12m / 5h / 3d / 2w / 2mo)
• Sessions modified in the last 2 minutes have a pulse icon and auto-expand.
• Expand a session to see:
  • Wider metric row (with token count + cache breakdown)
  • 📁 Projects touched (derived from Edit/Write file paths in the session, not just the working directory)
  • 🔍 View conversation — opens the per-turn timeline webview
  • ▶ Resume in Claude — uses claude-vscode.primaryEditor.open(sessionId) to open the session in the official Claude Code extension panel; falls back to a terminal claude --resume <uuid> if the official extension isn't installed
  • 📜 Open raw JSONL — opens the underlying session file in a regular editor

KB changes pane

A list of file changes in your knowledge base (default ~/docs), grouped by day:

• Committed changes from git log --name-status --since=<lookback>.
• Uncommitted working-tree changes ((uncommitted) subject).
• Status icons: A added, M modified, D deleted, R renamed.
• Click → open the file. Right-click → "Show diff against HEAD~1" (uses VS Code's built-in vscode.diff).

Projects pane

The same git log view, but two-level: day → project → files.

• Watches an explicit list (codeProjectsActivity.repoPaths) by default: ~/projects/unpolarize, ~/projects/ai/otelo.
• Auto-discovers additional repos: walks ~/projects/<depth-2> and includes any git repo with commits in the lookback window (so openclaw, inference-service, etc. show up automatically as you touch them).
• Same open/diff actions as KB.

Insights dashboard (webview)

Click the 📊 graph icon in the Sessions title bar to open a per-account dashboard:

• KPI row: total cost, total tokens, total messages, total subagents, median user thinking time (gap between Claude finishing and your next reply), burst rate (% of replies in <5s — flow-state indicator) — across the lookback window.
• Daily cost bar chart.
• Daily tokens by type stacked bars (input / output / cache read / cache write).
• "When you Claude" heatmap — 7 days × 24 hours, cell intensity = session count. Reveals your real work rhythm.
• Cost distribution histogram — how many sessions in each $-bucket.
• Top projects by cost horizontal bar chart (uses projects_touched, not just session cwd).
• Tool usage horizontal bar chart — top 12 tools by call count, computed by deep-parsing the most-recent N sessions (default 20). Bash usually wins.
• Top 10 expensive sessions clickable table.

All charts are inline SVG. No external libraries, no scripts, no fetches. Colors come from VS Code theme variables so dark and light themes both render correctly.

Conversation viewer (webview)

Click 🔍 View conversation on any session → opens an editor tab with a full timeline:

• Header summary: turn count, tool count, subagent count, session span (wall-clock first → last), time in tools, time outside tools, first/last activity timestamps.
• One card per turn:
  • #N · <wall-clock> · duration Xs · 3 tools (Bash×2 · Edit×1)
  • USER prompt
  • ASSISTANT response (with its own duration)
  • TOOLS section (collapsed by default — click to expand the calls list)
• One <details> per tool call (collapsed by default):
  • Tool icon (🔧 standard / 🪄 subagent)
  • Name + duration
  • Started / ended timestamps
  • Input as pretty-printed JSON
  • Output (truncated at 4 KB)
• Subagent calls (Agent tool) get a distinct purple-ish background and show subagent_type + description inline.
• Errors get a red border + error pill.

The webview uses VS Code theme variables so dark and light themes both render correctly. enableScripts=false (static HTML — no XSS surface).

How it works & what it stores

Code Sessions is a read-only dashboard over your local coding-agent CLI history. It reads three external sources it does not own and indexes them into a local cache — nothing leaves your machine.

What it reads (sources of truth, never modified):

• ~/.claude/projects/*/*.jsonl — Claude Code session transcripts
• ~/.grok/sessions/<cwd>/<uuid>/ — Grok Build sessions (chat_history.jsonl, summary.json, signals.json)
• Git history of your KB and ~/projects/* repos (via git log / git status, held in memory — not cached)

What it stores (a derived, disposable index): A single SQLite database (better-sqlite3, WAL mode) at:

<globalStorage>/sessions-cache.db
# macOS: ~/Library/Application Support/Code/User/globalStorage/zhirafovod.code-sessions/sessions-cache.db

It holds tables for sessions, turns, topic classifications, embeddings + 2D layout coordinates, starred sessions, and a migration ledger. The schema is versioned via PRAGMA user_version. This DB is purely a cache — delete the zhirafovod.code-sessions global-storage folder and the extension rebuilds it from disk on next launch (topics and embeddings are recomputed on demand). An incremental indexer only re-parses transcripts whose (mtime, size) changed, triggered by a 1.5 s-throttled file watcher.

Local-only compute:

• Embeddings for the agent graph use a local Ollama daemon (127.0.0.1:11434, nomic-embed-text), falling back to a built-in hashed bag-of-words vector when Ollama isn't running.
• Topic classification uses local Ollama by default, or optionally claude -p (your Claude subscription). When spawning claude, the extension passes a curated env (PATH/HOME/USER only) and deliberately withholds ANTHROPIC_API_KEY so it never switches to metered API billing.

There are no analytics, no telemetry, and no cloud calls from the extension itself. Webviews run under a strict Content-Security-Policy (static views execute no scripts; interactive views use nonce-gated scripts and whitelist no remote origins).

To inspect or reset the cache, and for the full table-by-table schema, see docs/DATA-STORES.md. Deeper internals live in docs/architecture.md and docs/embedding-setup.md.

Install

The fastest path uses the bundled scripts:

git clone git@github.com:unpolarize/code-sessions-vscode.git
cd code-sessions-vscode
npm install
./scripts/build-install.sh      # compile, package, install the .vsix
./scripts/ollama-setup.sh       # install Ollama if missing, start service, pull models

Reload VS Code → click the new Code Sessions icon in the Activity Bar.

The scripts are thin wrappers over standard commands; if you'd rather run them by hand:

npm install
npm run compile
npx @vscode/vsce package         # → code-sessions-X.Y.Z.vsix
code --install-extension code-sessions-*.vsix --force

For build / contribute / release flow see CONTRIBUTING.md.

Ollama dependency

The extension uses a local Ollama daemon for two things:

• Topic classification of conversation turns (the background classifier + the Classify all topics button on the agent graph). Default model: llama3.2:3b (~2 GB). Optional — if Ollama isn't running, the daemon idles and classification just doesn't happen automatically.
• Embeddings for the agent-graph 2D + 3D layouts. Default model: nomic-embed-text (~270 MB). Required if you want a real embedding-based scatter; the extension falls back to a hashed bag-of-words layout otherwise (much less informative).

Both run locally — no API tokens, no network calls beyond 127.0.0.1:11434.

One-time setup

# 1. Install Ollama
brew install ollama                                   # macOS
# curl -fsSL https://ollama.com/install.sh | sh       # Linux

# 2. Start the daemon (foreground)
ollama serve                                          # or: brew services start ollama

# 3. Pull the models the extension uses
ollama pull llama3.2:3b
ollama pull nomic-embed-text

# 4. Sanity check
curl -s http://127.0.0.1:11434/api/tags | jq '.models[].name'

Or just run ./scripts/ollama-setup.sh — it does all four steps and is safe to re-run (it skips what's already done).

Choosing different models

The defaults are tuned for small machines. If you have headroom, edit settings:

After changing the embedding model, run Claude: Drop cached embeddings and re-embed from the palette so the agent graph rebuilds from scratch.

Turning auto-classification off

If you don't want the daemon running:

• codeSessions.classify.autoBackground → false — disables the background daemon.
• codeSessions.classify.autoOnOpen → false — disables the on-open auto-classify when you view a conversation.

Manual classification still works via the Classify all topics button on the agent graph or Analyze conversation topics in the command palette.

Settings

Related

• sessions shell alias — the same data, instantly, from any terminal. bash ~/.claude/skills/sessions/session-center.sh recent 30 (or sessions recent 30).
• /sessions Claude Code skill — same script, invoked through the Claude Code slash-command system. Slower because the LLM is involved; use the shell alias or this extension when you want a "fully programmatic" view.
• Design + investigation notes (in the Sergey docs workspace): ~/docs/knowledge/tech/agents/claude-code-command-center.md.

Why this exists

The official Claude Code VS Code extension (anthropic.claude-code) declares a claude-sessions-sidebar view container, but it is gated behind context flag claude-vscode.sessionsListEnabled — not user-settable and not shipped yet. This extension fills the gap, adds analytics the official one will not (token / cost / subagents / projects-touched / per-turn trace), and integrates with the official one (resume opens the conversation in the Claude panel via claude-vscode.primaryEditor.open).

License

MIT — see LICENSE.