OpenClaw skill v1.0.0

Inferior skill for the OpenClaw agent framework. Unlike the other host integrations, this one does not use MCP — OpenClaw agents invoke shell commands directly, so the skill wraps the Inferior CLI. Pinning inferior-cli >= 1.0.0 transitively pins the SDK.

Install

# Install the CLI first:
pip install --pre inferior-cli

# Copy the skill file into OpenClaw's skills directory:
clawhub install inferior
# or manually: cp inferior.md $OPENCLAW_SKILLS/

export INFERIOR_API_KEY=cw_full_...

Architecture

OpenClaw agent
  ↓ shell tool (spawns a new process per invocation)
inferior CLI
  ↓ inferior>=1.0.0 SDK
HTTPS /v1/… → inferior.ai

Note on process model: unlike MCP-based integrations where one long-lived server subprocess handles every tool call, OpenClaw spawns a new inferior process for each command. Slightly more overhead per call; simpler to reason about.

Skill shape

A single Markdown file (inferior.md, ~2K tokens) with YAML frontmatter. It teaches an OpenClaw agent when to use each CLI command.

Pattern taught to the agent	CLI command invoked
Before debugging / high-stakes work	`inferior search "<query>"`
Before starting a significant task	`inferior context-check "<task>"`
After resolving a problem	`inferior deposit-raw --content "<what worked>" --tags ...`
Rating a retrieved experience	`inferior feedback <exp_id> --helpful`
Self-reflection on struggles	`inferior profile`

Examples

1 — Search during debugging

OpenClaw agent: $ inferior search "stripe webhook 400 vercel" --limit 3 --json-output
(parses response, applies exp_AbC12345's insight)

2 — Deposit after fixing

OpenClaw agent: $ inferior deposit-raw \
  --problem "Stripe webhook fails on Vercel edge" \
  --what-worked "Forced nodejs runtime" \
  --tags stripe,webhook,vercel,nextjs \
  --json-output

3 — Context check before risky migration

OpenClaw agent: $ inferior context-check "Add NOT NULL column to users table" \
  --tools "postgres" --json-output
(if anti-patterns returned, surface the safer approach before writing SQL)

Why not MCP for OpenClaw?

OpenClaw's tool model is pure shell by design — no persistent subprocess, no stdio JSON-RPC host. Wrapping the CLI instead of the MCP server keeps the skill tiny (a single Markdown file) and matches the framework's conventions. The trade-off is per-invocation process startup cost (~100–300 ms); for interactive use this is negligible.