Records Claude Code sessions as agent trajectories in cxdb
Records every Claude Code session as structured turns in cxdb — a turn-DAG context store for AI agent trajectories. Browse full session histories, tool calls, token usage, and reasoning in the cxdb UI.
In the cxdb repo:
make dev-docker-up
Starts cxdb in Docker. Frontend at http://localhost:9080 · API at http://localhost:9010
git clone https://github.com/prime-radiant-inc/claude-plugin-cxdb-integration
cd claude-plugin-cxdb-integration
make install
Sessions started after installation are recorded automatically. Open http://localhost:9080 to watch them appear in real time.
Seven Claude Code hooks record the full session lifecycle:
| Hook | cxdb item_type | Notes |
|---|---|---|
SessionStart | system (info) | Creates context, publishes type bundle, writes ContextMetadata + Provenance |
UserPromptSubmit | user_input | Captures user message text (async) |
PostToolUse | (buffered) | Buffers tool call + result to session state — no cxdb write yet |
Stop | assistant_turn | Flushes buffered tool calls; parses transcript for text + token metrics |
SubagentStop | handoff | Records agent-to-agent handoffs for Task tool calls |
PreCompact | system (warning) | Notes when context window compaction occurs (async) |
Notification | system (info) | Records Claude Code notifications (async) |
Tool call results are captured per tool type — each tool's response has a different structure:
| Tool | Content captured |
|---|---|
Bash | stdout (+ stderr if non-empty) |
Read | File contents (capped at 4000 chars) |
Glob | Matched file paths |
Grep | Matched file paths or content lines |
Edit | "Edited <filepath>" |
Task | Subagent's final text output (capped at 4000 chars) |
WebFetch / WebSearch | Generic fallback |
All cxdb writes are fire-and-forget — if cxdb is unreachable, your Claude Code session continues normally.
Create ~/.claude/cxdb-observability.local.md with YAML frontmatter:
url: http://cxdb.your-tailnet.ts.net:9010
Or set the CXDB_URL environment variable before starting Claude Code.
URL resolution order:
$CXDB_URL environment variableurl: field in ~/.claude/cxdb-observability.local.mdhttp://localhost:9010 (default)Each session produces a chain of cxdb.ConversationItem turns:
system (info) ← Session Started, with ContextMetadata + Provenance
user_input ← "Write a function that..."
assistant_turn ← Text + tool calls (Bash, Read, Edit, ...) + token metrics
user_input ← "looks good, now add tests"
assistant_turn
...
Provenance fields captured per session:
make install # registers plugin in ~/.claude/plugins/installed_plugins.json
make uninstall # removes it
Apache 2.0 — see LICENSE.
Matches all tools
Hooks run on every tool call, not just specific ones
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
npx claudepluginhub prime-radiant-inc/claude-plugin-cxdb-integrationReverse engineer clean behavioral specs from any codebase
An iterative implementation methodology that pairs with superpowers. Extracts requirements with proof obligations from large spec collateral, defines a walking skeleton that closes a real journey, then loops through audited sprints building a behavior evidence corpus until an auditor confirms the product matches the spec. Designed for comprehensive or ambiguous specs where the upfront writing-plans flow loses the plot.
Streamlined Linear integration - one tool, seven actions, zero bloat
Two skills for adversarial code review. parallel-adversarial-review dispatches two same-model reviewer subagents in parallel under a competitive scoring frame, aggregating findings and taking the worst severity on disagreement. multi-model-adversarial-review extends this to N installed coding-agent CLIs (claude, codex, gemini, opencode, amp, aider) with a cross-critique stage that catches hallucinations and severity inflation, then synthesizes a deduplicated report. Includes a fixture-based eval suite that scores precision and recall against planted defects.
A documentation department in a box: studies a project and its readers, decides the doc portfolio with you, writes contributor/user/adopter docs with verified claims, and keeps them true as the code moves
Observability platform for Claude Code and Agent SDK — monitor, debug, and control AI coding agents
Real-time observability dashboard for Claude Code agents
Automatic tracing of Claude Code sessions to Arize AX or Phoenix with OpenInference spans. Supports 9 hooks for comprehensive observability.
Session analytics hook for Claude Code — 15 configurable sections (models, cost, cache, tools, files, git diff, RTK savings...)
Tracks Claude Code session usage (tokens, costs, skills, MCP, subagents) and sends data to a self-hosted dashboard.
MCP server that saves 98% of your context window with session continuity. Sandboxed code execution in 11 languages, FTS5 knowledge base with BM25 ranking, and automatic state restore across compactions.