self-evolving-opencode

CLI pipeline that reads OpenCode's SQLite DB, extracts user goals via LLM, clusters them, and automatically synthesizes + evolves Claude skills from conversation patterns.

Setup

uv sync

Requires GCP auth for Vertex AI:

gcloud auth application-default login

Redis (for LLM call caching — degrades gracefully if down):

podman compose up -d    # starts seo-redis on localhost:6380

Initialize the skills database (once):

uv run python scripts/init_skills_db.py

Env vars (defaults work if your zshrc sets them):

• ANTHROPIC_VERTEX_PROJECT_ID (default: itpc-gcp-ai-eng-claude)
• GOOGLE_VERTEX_LOCATION (default: global)

Skill Evolution

The main feature. Mines OpenCode conversations to create and improve Claude skills automatically.

# 1. Dry run — inspect what it would do (no writes to disk or DB)
DRY_RUN=1 uv run python play.py --evolve 5

# 2. Real run — writes to ~/.claude/skills/ and ./skills.db
uv run python play.py --evolve 5

# 3. Check what it wrote
ls ~/.claude/skills/
cat ~/.claude/skills/*/SKILL.md

--evolve [N] runs two sequential queues (default N=50 sessions per queue):

1. Synthesizer (oldest-first): extract goals → cluster → semantic search existing skills → LLM decide new/update → synthesize SKILL.md
2. Evolve (newest-first): detect skill invocations → reflect per thread (tag rules + extract insights) → curate per skill (ADD new rules)

Options:

• --concurrency M — max concurrent LLM calls (default: 5)
• DRY_RUN=1 — runs all LLM calls, prints SKILL.md content to stdout, zero disk/DB writes

Skills are written to ~/.claude/skills/<name>/SKILL.md. Rules are tracked in ./skills.db.

Usage

List sessions

uv run python play.py                              # most recent 30
uv run python play.py -n 10                         # limit
uv run python play.py --agent auto-accept           # filter by agent
uv run python play.py --dir my-project              # filter by directory
uv run python play.py --agent auto-accept --dir my-project -n 5

Extract goals

uv run python play.py --goals ses_abc123...         # by session ID
uv run python play.py --agent auto-accept --goals 6 # by row index
uv run python play.py --goals ses_abc123... --check  # check if achieved
uv run python play.py --goals ses_abc123... --summarize  # summarize thread

Batch + clustering

uv run python play.py --goals-file sessions.txt              # batch extract
uv run python play.py --goals-file sessions.txt --cluster     # cluster goals
uv run python play.py --goals-file sessions.txt --cluster --summarize  # + summaries

Testing

uv run pytest                    # all except @live (default)
uv run pytest -m live --override-ini="addopts="   # hits real LLM API
uv run pytest -m redis           # only Redis-dependent tests
uv run pytest -k test_opencode   # single test file

How it works

1. List mode — queries session table in ~/.local/share/opencode/opencode.db
2. Goals mode (--goals) — sends conversation transcript to Claude Opus via Vertex AI, extracts goals via forced tool use
3. Check mode (--check) — evaluates whether each goal was achieved
4. Evolve mode (--evolve) — full pipeline: extract goals → cluster → synthesize skills → reflect on threads → curate rules