claude-smart 0.1.0

Self-improving Claude Code plugin — learns from corrections via reflexio

claude-smart

Self-improving Claude Code plugin — learns from your corrections, not just remembers them.

Quick Start • How It Works • Installation • Slash Commands • Configuration • Troubleshooting • License

claude-smart turns your Claude Code corrections into durable rules that shape future sessions. Instead of replaying past observations as context, it distils them into a project playbook and per-session preferences — so Claude stops repeating the same mistakes and adapts to how your codebase actually wants to be written.

---

Why Learning, Not Memory

Plain memory solutions preserve what happened — they re-inject transcripts, summaries, or observations from prior sessions. That works for continuity, but it has two real limits:

• It grows with every session. More memory means more tokens, more context dilution, and diminishing returns as Claude has to re-read a history it can't fully use.
• It doesn't change behavior. If you corrected Claude yesterday about a library choice, a test framework, a deployment region — a memory system remembers the correction happened, but Claude may still make the same default choice today because nothing updated its decision-making.

claude-smart takes a different approach: extract, don't accumulate. Each session's corrections and successful patterns are distilled by an LLM into two small, structured artifacts:

• User profile — short, session-scoped preferences Claude should respect right now.
• Project playbook — cross-session behavioral rules keyed to your project, with an explicit trigger (when the rule applies) and rationale (why). Rules are deduplicated, updated, and archived as they evolve.

The result is a compact, always-up-to-date set of instructions Claude reads at the start of every session — measured in dozens of tokens rather than thousands, and actually capable of changing behavior.

---

Quick Start

# 1. Clone with the reflexio backend submodule
git clone --recurse-submodules https://github.com/ReflexioAI/claude-smart.git
cd claude-smart

# 2. Install dependencies (creates a uv-managed venv, pulls reflexio as a path dep)
uv sync

# 3. Turn on the local providers inside reflexio — no API key required at all
mkdir -p ~/.reflexio
grep -q '^CLAUDE_SMART_USE_LOCAL_CLI=' ~/.reflexio/.env 2>/dev/null \
  || echo 'CLAUDE_SMART_USE_LOCAL_CLI=1' >> ~/.reflexio/.env
grep -q '^CLAUDE_SMART_USE_LOCAL_EMBEDDING=' ~/.reflexio/.env 2>/dev/null \
  || echo 'CLAUDE_SMART_USE_LOCAL_EMBEDDING=1' >> ~/.reflexio/.env

# 4. Start the local reflexio backend (storage + search + extraction orchestrator)
uv run reflexio services start --only backend --no-reload

# 5. Install the plugin into Claude Code (project-level)
mkdir -p .claude && cat > .claude/settings.local.json <<'JSON'
{
  "extraKnownMarketplaces": {
    "claude-smart-local": {
      "source": { "source": "directory", "path": "." }
    }
  },
  "enabledPlugins": { "claude-smart@claude-smart-local": true }
}
JSON

Restart Claude Code in this workspace. The first time you correct Claude on something project-specific ("no, don't use pytest-asyncio — use anyio with trio"), a playbook rule will be extracted. Every subsequent session in the project starts with that rule injected — automatically, without you asking.

---

Key Features

• 🧠 Learn, don't just remember — Corrections become structured, deduplicated rules, not transcript replays.
• 🎯 Two-tier scope — Per-session profiles for the current conversation; cross-session playbooks for the whole project.
• 🔌 Fully local — no external API keys needed — Generation runs through your local claude CLI; semantic search runs on an in-process ONNX embedder (all-MiniLM-L6-v2). The whole stack works offline.
• 🔎 Hybrid search — Playbooks and profiles are indexed with vector + BM25 search for fast, robust retrieval.
• 📥 Automatic hook ingestion — SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd all wired up; you don't run anything manually.
• 🏷️ Correction-aware — Corrective phrasings ("no, don't", "actually", "stop", "wrong") are detected and weighted during extraction.
• 🧪 Offline resilience — If the reflexio backend is down, hooks buffer to disk; the next successful publish drains them.
• 🧰 Three slash commands — /show, /learn, /tag for on-demand control.

---

How It Works

Core components:

1. 5 lifecycle hooks (plugin/hooks/hooks.json)
   • SessionStart — fetches the project playbook from reflexio and injects it as additionalContext.
   • UserPromptSubmit — buffers each user turn, heuristically flags corrections.
   • PostToolUse — records tool invocations for later extraction.
   • Stop — finalizes the assistant turn from the transcript, publishes to reflexio.
   • SessionEnd — flushes the remaining buffer with force_extraction=True.
2. Local state buffer — JSONL per session at ~/.claude-smart/sessions/{session_id}.jsonl. Offline-safe.
3. Reflexio backend (submodule at reflexio/) — SQLite storage, hybrid search, profile/playbook extraction, dedup, status lifecycle (CURRENT → ARCHIVED). Runs on localhost:8081.
4. Claude Code LLM provider — a LiteLLM custom provider registered inside reflexio. Every generation call (extraction, update, dedup, evaluation) subprocesses claude -p --output-format json, so no OpenAI/Anthropic key is needed for the learning loop.
5. Three slash commands — /show, /learn, /tag.

Data flow:

Claude Code session
  ├─ UserPromptSubmit ─┐
  ├─ PostToolUse  ─────┤  → JSONL buffer ─→ Stop ─→ reflexio publish_interaction
  └─ Stop         ─────┘                              │
                                                      ▼
                                        ┌─────────────────────────┐
                                        │ reflexio extractors     │
                                        │  (run via claude -p)    │
                                        │  → profiles + playbooks │
                                        └────────────┬────────────┘
                                                     │
                                                     ▼
Next session → SessionStart → search_user_playbooks(agent_version=project_id)
              → additionalContext injected into Claude's system prompt

Mapping to reflexio:

Reflexio field	claude-smart value
user_id	Claude Code session_id — scopes profiles to the current conversation
agent_version	project_id (git-toplevel basename) — stable across sessions, so playbooks accumulate project-wide
session_id	Claude Code session_id — for reflexio's deferred success evaluation

Cross-session playbook retrieval uses search_user_playbooks(agent_version=project_id, user_id=None) — playbooks written from any prior session in this project surface for every future session.

---

Installation

One-command install

If you just want the plugin wired into Claude Code (marketplace added, plugin installed, ~/.reflexio/.env seeded with the local-provider flags), run one of:

# uvx — pulls the Python package straight from git, no clone required
uvx --from git+https://github.com/ReflexioAI/claude-smart claude-smart install

# npx — same thing via the published npm wrapper
npx claude-smart install

Both do the same three things:

1. claude plugin marketplace add ReflexioAI/claude-smart
2. claude plugin install claude-smart@yilu
3. Append CLAUDE_SMART_USE_LOCAL_CLI=1 and CLAUDE_SMART_USE_LOCAL_EMBEDDING=1 to ~/.reflexio/.env (idempotent).

You still need to start the reflexio backend yourself the first time (uv run reflexio services start --only backend --no-reload from a clone of this repo). Everything else — submodule init, uv sync — runs inside Claude Code's Setup hook on first session.

For a manual, step-by-step walkthrough, see below.

Prerequisites

Tool	Purpose
Claude Code	The host CLI — also used as the LLM backend for extraction
uv	Python package manager (Python 3.12+)
git	Needed to clone with submodules and to derive the project id

No external API keys needed. Generation runs through your local claude CLI (via a LiteLLM custom provider). Embeddings run through an in-process ONNX model (all-MiniLM-L6-v2, bundled by chromadb). Both are opt-in; when enabled, reflexio refuses to fall back to paid APIs.

Step 1 — Clone the repository (with the reflexio submodule)

git clone --recurse-submodules https://github.com/ReflexioAI/claude-smart.git
cd claude-smart

# If you forgot --recurse-submodules
git submodule update --init --recursive

Step 2 — Install Python dependencies

uv sync

This creates .venv/, pulls reflexio-ai as a path dependency from the reflexio/ submodule, and registers the claude-smart and claude-smart-hook console scripts.

Step 3 — Enable the local providers in reflexio

Two env flags turn on the local generation backend (Claude Code CLI) and the local embedder (in-process ONNX). Both live in ~/.reflexio/.env:

mkdir -p ~/.reflexio
grep -q '^CLAUDE_SMART_USE_LOCAL_CLI=' ~/.reflexio/.env 2>/dev/null \
  || echo 'CLAUDE_SMART_USE_LOCAL_CLI=1' >> ~/.reflexio/.env
grep -q '^CLAUDE_SMART_USE_LOCAL_EMBEDDING=' ~/.reflexio/.env 2>/dev/null \
  || echo 'CLAUDE_SMART_USE_LOCAL_EMBEDDING=1' >> ~/.reflexio/.env

On first use, the embedder downloads the ~80 MB ONNX model once and caches it at ~/.cache/chroma/onnx_models/. Subsequent starts reuse the cache and stay offline.

Step 4 — Start the reflexio backend

Run this from the claude-smart repo root (not the reflexio/ subdir) so that uv run uses the claude-smart venv — which already has chromadb installed for the local embedder and reflexio-ai available as a path dep with the reflexio CLI script registered:

uv run reflexio services start --only backend --no-reload

You should see a log line like:

Registered claude-code LiteLLM provider (cli=/path/to/claude)
Local embedding provider enabled (model=local/minilm-l6-v2)
Auto-detected LLM providers (priority order): ['claude-code', 'local']
Primary provider for generation: claude-code
Embedding provider: local
Application startup complete.

Health check:

curl http://localhost:8081/health
# {"status":"healthy"}

Leave this running in a separate terminal. Stop it later with:

uv run reflexio services stop

Step 5 — Install the plugin into Claude Code

Project-level (recommended while you evaluate):

mkdir -p .claude
cat > .claude/settings.local.json <<JSON
{
  "extraKnownMarketplaces": {
    "claude-smart-local": {
      "source": { "source": "directory", "path": "$PWD" }
    }
  },
  "enabledPlugins": { "claude-smart@claude-smart-local": true }
}
JSON

User-level (all projects):

Put the same JSON into ~/.claude/settings.json, using an absolute path for the marketplace path.

Restart Claude Code. On the next session start you should see reflexio logs show a search_user_playbooks call — that's the SessionStart hook fetching the (currently empty) playbook.

Step 6 — Sanity check

Inside Claude Code:

/show

On a fresh project you'll see _No playbook or profiles yet for project ._ — correct. Have a conversation, include at least one genuine correction ("no, don't use X — use Y"), then:

/learn

That forces immediate extraction. Run /show again after ~20–30 seconds; the extracted rule should appear.

---

Slash Commands

Command	What it does
/show	Print the current project playbook plus the current session's user profiles (same markdown that SessionStart injects). Use it to audit what rules and preferences Claude is being told to follow.
/learn	Force reflexio to run extraction now on the current session's unpublished interactions. Without this, extraction runs at the end of the session or on reflexio's batch interval.
/tag [note]	Tag the most recent turn as a correction, for cases the automatic heuristic missed. The note becomes the correction description the extractor sees.

---

Configuration

Environment variables

Variable	Default	Purpose
CLAUDE_SMART_USE_LOCAL_CLI	0	Set to 1 in ~/.reflexio/.env to route generation through the local claude CLI.
CLAUDE_SMART_USE_LOCAL_EMBEDDING	0	Set to 1 to use the in-process ONNX embedder (requires chromadb).
CLAUDE_SMART_CLI_PATH	shutil.which("claude")	Override the path to the claude binary.
CLAUDE_SMART_CLI_TIMEOUT	120	Per-call subprocess timeout (seconds). Raise for slow prompts.
CLAUDE_SMART_STATE_DIR	~/.claude-smart/sessions/	Where the per-session JSONL buffer lives.
REFLEXIO_URL	http://localhost:8081/	Point the plugin at a non-local reflexio backend.

Where data lives

Path	What
~/.reflexio/data/reflexio.db	Source of truth — profiles, user_playbooks, interactions, FTS5 indexes, and vec0 embedding tables (plus .db-shm / .db-wal WAL sidecars). Inspect with sqlite3.
~/.reflexio/.env	Provider config — CLAUDE_SMART_USE_LOCAL_CLI, CLAUDE_SMART_USE_LOCAL_EMBEDDING, any optional API keys.
~/.claude-smart/sessions/{session_id}.jsonl	Per-session buffer. User turns, assistant turns, tool invocations, {"published_up_to": N} watermarks. Safe to inspect and safe to delete — everything past the latest watermark has already been written to reflexio's DB.
~/.cache/chroma/onnx_models/all-MiniLM-L6-v2/	Cached ONNX weights (~86 MB, downloaded once). Delete to force a re-download.

Scope: profile vs. playbook

• Profile (user_id = session_id) — session-scoped preferences. Does not persist across sessions, but is reinjected if you resume the same session (/resume, /clear, /compact).
• Playbook (agent_version = project_id) — cross-session. Every session in the same project — identified by git-toplevel basename — sees the accumulated playbook.

Embeddings

claude-smart uses an in-process ONNX embedder (Chroma's all-MiniLM-L6-v2, 384-dim, zero-padded to reflexio's 512-dim schema). The model weights are downloaded on first use (~80 MB, cached under ~/.cache/chroma/onnx_models/) — after that, no network calls for embedding. Runtime cost is a few milliseconds per short document on CPU.

If you still want to use a cloud embedding provider (OpenAI, Gemini, etc.), omit CLAUDE_SMART_USE_LOCAL_EMBEDDING and set the corresponding API key in ~/.reflexio/.env — reflexio will fall back to its standard provider-priority chain.

---

How the Claude Code Provider Works

claude-smart ships a small patch to reflexio (reflexio/server/llm/providers/claude_code_provider.py) that registers a LiteLLM CustomLLM named claude-code. Every time reflexio wants to generate, evaluate, or dedup, it ends up in litellm.completion(model="claude-code/default", ...) — which routes to our handler. The handler:

1. Splits LiteLLM's messages into (system_prompt, dialogue).
2. Subprocesses claude -p --output-format json --append-system-prompt "<system>" with the dialogue on stdin.
3. Parses the JSON stdout into a LiteLLM ModelResponse with populated usage tokens.

Registration is opt-in (CLAUDE_SMART_USE_LOCAL_CLI=1) and idempotent, so enabling it does not affect users who still want OpenAI/Anthropic — reflexio's normal provider-priority chain stays intact.

---

Troubleshooting

SessionStart injects nothing after a correction. Extraction is async by default. Run /learn to force it, wait ~20–30s, then run /show — no new session needed. /show shows whether the rule was actually extracted.

Reflexio refuses to boot with "no embedding-capable provider". Check that CLAUDE_SMART_USE_LOCAL_EMBEDDING=1 is in ~/.reflexio/.env and that chromadb is installed in the venv (uv run python -c "import chromadb" should print nothing). If you'd rather use a cloud embedder instead, drop the env flag and set OPENAI_API_KEY or GEMINI_API_KEY in the same file.

claude-smart doesn't see my interactions. Check ~/.claude-smart/sessions/. If your current session's JSONL has no User/Assistant rows, the plugin isn't receiving hook events — verify .claude/settings.local.json has the right path and that enabledPlugins is true.

Hooks appear to time out. Each hook is capped at 15–60s. If you see long pauses, check uv is on PATH (hooks shell out to uv run). Set CLAUDE_SMART_CLI_TIMEOUT=180 to give the LLM provider more headroom.

A different LLM is being used. Reflexio's provider priority is claude-code > local > anthropic > gemini > ... > openai. If you have CLAUDE_SMART_USE_LOCAL_CLI=1 and an Anthropic key set, claude-code still wins for generation; local sits above openai/gemini for embeddings. Check the startup log line Primary provider for generation: <name> and Embedding provider: <name> to confirm.

I want to wipe everything and start over.

rm -rf ~/.claude-smart/sessions/
rm -rf ~/.reflexio/data/           # reflexio SQLite store

---

Dashboard (web UI)

A Next.js management UI lives in dashboard/. Use it to browse local session buffers, inspect extracted user profiles, edit and archive project playbooks, and tweak the claude-smart environment. It connects to the same reflexio backend the plugin uses, so run that first.

# 1. reflexio backend on :8081 (see Step 4 above)
uv run reflexio services start --only backend --no-reload

# 2. install and run the dashboard
cd dashboard
npm install
npm run dev   # http://localhost:3001

The dashboard reads ~/.claude-smart/sessions/*.jsonl directly (server-side) for in-flight session transcripts and proxies everything else through reflexio. All state lives where the CLI already keeps it — the dashboard does not introduce a second source of truth.

---

Development

Run the test suite:

# reflexio patch unit tests
cd reflexio
uv run pytest tests/server/llm/ -q -o 'addopts='

# claude-smart package tests
cd ..
uv run pytest tests/ -q

Exercise a hook handler directly (useful for debugging without a live Claude Code session):

echo '{"session_id":"dev-1","source":"startup","cwd":"'"$PWD"'"}' \
  | uv run python -m claude_smart.hook session-start

---

License

This project is licensed under the Apache License 2.0. The bundled reflexio/ submodule is also Apache 2.0. Claude Code is Anthropic's and not covered by this license.

See the LICENSE file for details.

---

Support

• Issues: open one on GitHub describing the symptom and include the reflexio startup log (stdout of uv run reflexio services start) and the relevant lines of ~/.claude-smart/sessions/{session_id}.jsonl.
• Architecture notes: see the plan file in .claude/plans/ (if present), which walks through each design decision and the rationale.

---

Built on reflexio · Runs on Claude Code · Written in Python 3.12+