Instructions file CodexOpenCode
AGENTS.md instructions for lemoncrow-lab/lemoncrow: You are operating as lemoncrow:code; apply its full discipline and route all file I/O, search, edits, and shell work through LemonCrow's MCP tools.
Instructions file
Claude Code instructions for lemoncrow-lab/lemoncrow, covering claude.md, environment, common commands, test and lint / format / typecheck.
Agent
Unattended software engineer: run tasks end to end, autonomously — no approval, no questions, ever. Ambiguous → smallest reasonable interpretation, stated as assumption: in the task report.
Agent
Software engineer on a lean toolset (token-heavy tools stripped): run tasks end to end.
Agent
Software engineer: ship the asked-for change end to end — locate, edit, verify, report.
Agent
Implementation specialist: complete an accepted plan or scoped task in one verified pass; sole builder, no design handoff or executor delegation.
Agent
Read-only explorer: locate the code answering the question, cite it by stable anchor, report fast.
Agent
Catch-all agent: work fitting no specialized role — mixed research+implementation, ad hoc investigation, multi-step chores across code and shell. Never assume every task is a code change. Pure code change → code; accepted plan → execute; checked deliverable → solve.
Agent
Planner: inspect only what's needed, produce the smallest viable plan another agent can execute without guessing — smallest trims padding, never steps the spec's properties require.
Agent
External researcher: fetch primary sources, synthesize, cite every claim.
Agent
Adversarial reviewer: find what's wrong; don't validate that work was done. Never edit source files.
Agent
Autonomous solver: own a concrete, verifiable task end to end — no planning handoff.
Agent
Fully autonomous unattended agent.
Agent
Minimal-toolset coding agent where context overhead matters.
Agent
Main coding agent. Edits, refactors, fixes bugs, and ships features with the LemonCrow task loop.
Agent
Always use for focused isolated code executor.
Agent
Always use for fast read-only explore agent for locating code.
Agent
Always use for all other tasks, catch-all agent.
Agent
Always use for for read-only implementation planner.
Agent
Always use for external web researcher.
Agent
Always use for adversarial read-only reviewer.
Agent
Always use for autonomous verified task solver.
Skill Claude CodeCodex
Measures what LemonCrow actually buys you on your repo and prompts — offline from session history (free) or online as a real A/B run against the host you are in, vanilla vs LemonCrow (real API spend). Does not benchmark LemonCrow internals — dev suite commands at the bottom for that.
Skill Claude CodeCodex
Thin pass-through to the real lc CLI (already built, already validated) — lc agent / lc skill / lc set / any other documented subcommand. Run the right CLI command, relay its output. Never invent cost numbers, validation, or confirmation logic here; the CLI already owns all of that. Not the cross-host bulk installer…