Store

Skills are stateless by accident. The ones that persist anything invent a location (/tmp/session-$$, a bespoke dot-directory, sibling workspace dirs, a markdown file with checkboxes) and a hand-typed JSON schema that the next script silently depends on.

steer store gives every skill a real database with two scopes:

When to use it

• Anything worth keeping between runs: run history, caches, an analyzed profile, a binding the user already confirmed once.
• Real skills pay for the missing primitive by redoing work. Vercel's vercel-cli-with-tokens rediscovers the project and team binding every session (workspace state), and humanizer re-analyzes the user's writing sample on every run (a user-scoped profile). The repo-health example keeps each run's totals at workspace scope, which is what powers its trend arrows.
• Skip it when runs share nothing. State you never read back is debt, not memory.

CLI

# key/value
steer store put last_run '"2026-06-11"'
steer store get last_run
steer store del last_run
steer store keys

# JSON documents
steer store insert runs '{"pr": 142, "ok": true}'
steer store find runs --where ok=true --limit 10
steer store tables

# raw SQL escape hatch
steer store query "SELECT COUNT(*) AS n FROM runs"

All commands take --skill (default: inferred from the nearest SKILL.md or STEER_SKILL), --scope user|workspace, and --workspace <dir>. Values are parsed as JSON when possible, kept as strings otherwise. Reads print JSON, which scripts and agents parse directly.

Python

from steer import Store

with Store("pr-review") as store:                      # user scope
    store.put("last_run", {"at": "2026-06-11", "ok": True})
    store.get("last_run")
    store.insert("runs", {"pr": 142, "findings": 3})
    store.find("runs", {"pr": 142})
    store.query("SELECT json_extract(data,'$.findings') AS n FROM runs")

ws = Store("pr-review", scope="workspace", workspace=".")

Inside a skill, sibling scripts import the same API from the bundled runtime next to them (from steer import Store); the repo-health example's report script does exactly that to record run totals.

In a SKILL.md

steer new --with store generates:

## Memory

This skill persists data between runs with the bundled `store` command
(per-skill SQLite). Examples:

    python3 scripts/steer.py store put last_run '"2026-06-11"'
    python3 scripts/steer.py store get last_run
    python3 scripts/steer.py store insert runs '{"file": "report.pdf", "ok": true}'

Use `--scope workspace` for state that belongs to this project rather
than the user.

Documents land in auto-created tables (id, data JSON, created_at); find filters with json_extract equality and returns dicts with _id / _created_at injected. Table names are validated; everything else is parameterized.

Notes

• SQLite in WAL mode through the stdlib sqlite3 module, so there is nothing to install.
• The raw query API is the escape hatch for migrations, joins, and indexes; steer doesn't hide SQL from you.