cross-pollinate

name	cross-pollinate
description	Detect a host project's locally-authored coding-agent definitions (skills, subagents, rules, commands, hooks, MCP) and make each available in the formats of the OTHER agents the project supports, as declared in .lisa.config.json. Any-to-any, provenance-tracked, idempotent, never clobbers hand-edited output.
allowed-tools	["Bash","Read","Write","Edit","Glob","Grep"]
argument-hint	[path] [--dry-run] [--write]

Cross-Pollinate: $ARGUMENTS

A host project that installs Lisa may hand-author a definition for one coding agent — a .claude/skills/foo, a .cursor/rules/bar.mdc, an MCP server only Claude sees. /lisa:cross-pollinate detects those locally-authored definitions and regenerates each in the formats of the other agents the project supports, so the whole fleet stays in parity with what a developer wrote for any single one.

The model

any agent's format  ->  Claude-format IR  ->  every OTHER configured agent

Claude format is the canonical intermediate representation (IR) because every Lisa generator already sources from it. So translation is always: normalize the detected definition up to Claude-format IR, then fan that IR out to each target agent using the documented mappings below. Never translate format A → B directly — that multiplies error and loses fidelity on every hop.

The agents a project supports come from .lisa.config.json harness (claude | codex | cursor | agy | copilot | opencode | both | fleet; all is an alias for fleet). Only emit to agents that harness includes.

Provenance is authoritative — the lockfile, not file paths

.lisa/cross-pollination.lock.json (committed) is the source of truth for what this skill generated. Location heuristics cannot work: a generated Cursor rule sits in the same .cursor/rules/ directory as a hand-authored one. The lockfile is the only reliable way to enforce the four invariants:

1. No loops — a path recorded as a target is NEVER treated as a source on the next run. Without this, today's output becomes tomorrow's input and translations ping-pong forever.
2. Garbage collection — when a source is deleted/renamed, its now-orphaned targets are removed.
3. Never clobber edits (Chesterton's fence) — if a target's on-disk hash differs from the recorded generatedHash, a human edited it. Stop and report; do not overwrite. They may have intentionally diverged, or want to adopt the edit as a new source.
4. Idempotent — unchanged source + intact targets => no-op. (This is why it is safe to also run during lisa apply.)

Lockfile shape:

{
  "version": 1,
  "entries": {
    "skill:security-review": {
      "source":  { "agent": "claude", "path": ".claude/skills/security-review", "hash": "…" },
      "targets": [ { "agent": "codex", "path": ".claude/skills/security-review/agents/openai.yaml", "generatedHash": "…" } ]
    }
  }
}

The logicalId is <kind>:<name> — stable across formats. It links a source to its targets and detects collisions: the same logicalId authored independently in two agents (neither generated). On a collision, NEVER auto-translate over either side — report it and let the human pick the source of truth.

How to run

The deterministic core (scan, provenance, loop/GC/drift/conflict detection, and the skill + MCP emitters) is a bundled engine. Always run it first — it does the safe, mechanical work and tells you exactly what is left for you to translate by hand.

# dry-run (default): report only, writes nothing
node "${CLAUDE_PLUGIN_ROOT}/scripts/cross-pollinate.mjs" "$PROJECT_ROOT" --json
# apply the deterministic emits + update the lockfile
node "${CLAUDE_PLUGIN_ROOT}/scripts/cross-pollinate.mjs" "$PROJECT_ROOT" --write

In the Lisa source repo itself the path is node plugins/lisa/scripts/cross-pollinate.mjs (built) or node plugins/src/base/scripts/cross-pollinate.mjs (source).

The JSON report has four lists you act on:

• emits — already handled by --write (skills, MCP). Verify, don't redo.
• conflicts — STOP. Report each to the user; do not translate.
• orphans — GC'd by --write. Confirm the removals look right.
• pending — your job. Definitions whose target format the engine does not emit deterministically. Translate each per the mappings below, then record provenance (see "Recording provenance").

Per-agent format mappings (for pending translation)

Normalize the source to Claude-format IR first, then emit per target:

Rules

• Claude ↔ Cursor: handled deterministically by the engine. Claude reads .claude/rules/<name>.md; Cursor reads .cursor/rules/<name>.mdc with YAML frontmatter (description from the first H1, alwaysApply: true) and intra-rule link extensions rewritten. You do not translate these — verify the engine's emits.
• Codex / OpenCode / agy (pending): project rules are delivered via AGENTS.md, not a per-rule file. MERGE the rule content into AGENTS.md under a stable marked section rather than writing a separate file (so re-runs replace, not append). agy never gets a duplicate rule file (rules-once).
• Copilot (pending): .github/copilot-instructions.md (inline). Merge into a marked section, don't scatter.

Subagents

• Claude / agy / Cursor / OpenCode: agents/<name>.md (or .claude/agents/).
• Copilot: agents/<name>.agent.md (note the .agent.md suffix).
• Codex: does not consume agent files. Report as dropped — unsupported, do not invent a target.

Commands

• Claude / Cursor / Copilot: markdown command files (.claude/commands/…).
• Codex / agy: auto-derive commands from skills; no separate command file. Report as dropped — derived-from-skill.

Hooks (lossy — drop-and-report, never fake support)

• Event-name casing differs per agent (PreToolUse→preToolUse/beforeSubmitPrompt, UserPromptSubmit→userPromptSubmitted, Stop→agentStop/stop, …).
• agy lacks SessionStart/SubagentStart — only PreToolUse (e.g. block-no-verify) is portable. Everything else: dropped — unsupported.
• Codex plugin hooks do not fire in codex exec — install into ~/.codex/hooks.json, not a bundled file.
• Copilot rejects the entire hooks config if a SubagentStart/empty-matcher entry is present — drop that event wholesale.
• For every hook you cannot faithfully translate to a target, REPORT it with the reason. Do not silently omit (no silent caps).

MCP (handled by the engine for Claude↔Cursor; report others)

• Claude: .mcp.json. Cursor: .cursor/mcp.json (same JSON shape).
• Copilot: inline mcpServers object in the plugin manifest (bundled .mcp.json is ignored). Codex: TOML / .codex-plugin pointer. agy/OpenCode: user-global installer / .opencode. The engine reports these as pending; translate per the target's documented shape or report dropped.

Recording provenance for hand-translated pending items

After you write a target by hand, it MUST be recorded in .lisa/cross-pollination.lock.json or the next run will treat it as a source (loop) or fail to GC it. Re-run the engine after staging your translated files — on its next pass it adopts and hashes any target already referenced — OR add the entry yourself following the lockfile shape above (compute generatedHash as the sha256 of the written file/dir). Prefer the engine; only hand-edit the lock if the engine cannot infer the mapping.

Output

Report, concisely:

• harness + resolved target agents
• counts: detected sources, written, skipped-drift, GC'd
• every conflict (with the agents involved)
• every pending item you translated, and every hook/agent/command you dropped with its reason
• confirm .lisa/cross-pollination.lock.json was updated and the working tree otherwise contains only intended changes

Rules

• Treat the lockfile as authoritative; never infer "mine vs theirs" from paths.
• Never overwrite a drifted (hand-edited) target — report it.
• Never auto-resolve a collision — report it.
• Only emit to agents the project's harness includes.
• Drop-and-report any translation a target genuinely cannot support; never fake a location you are unsure of.
• This skill is additive — it does not delete a human's source definitions, only its own orphaned generated targets.