[edmundmiller]

Summary

Adds a Claude Code Stop hook that runs nextflow lint once, after the agent finishes a turn, on uncommitted .nf and .config files.

This is the small version of the lint hooks from nextflow-io/agent-skills #5 and #10. Those ran on every Write/Edit (PostToolUse), which linted half-finished files and started a JVM per edit. A Stop hook sees the working tree after the turn's edits are done, and it runs one lint for the whole set of changed files.

The hook is registered in hooks/hooks.json (the file Claude Code loads for a plugin) and stays idle until someone opts in:

• one repository: mkdir -p .claude && touch .claude/nextflow-lint-hook
• every repository: NEXTFLOW_LINT_HOOK=1

Until then it exits immediately and does not call Nextflow. It also skips when nextflow is missing, when the tree has no uncommitted .nf or .config files, and when a stop hook has already continued the turn (so a lint failure is reported once).

Plugin version is unchanged. No language-server or .lsp.json changes.

Test plan

• python3 -I tools/check_sources.py (114 files, 0 errors)
• python3 -I tools/test_consolidate_skills.py
• claude plugin validate --strict on the marketplace and plugin manifests (Claude Code 2.1.295)
• claude plugin install reports Hooks (1) Stop plus the existing 28 skills and the seqera MCP server
• Hook script: no opt-in leaves nextflow uncalled; NEXTFLOW_LINT_HOOK=1 and a .claude/nextflow-lint-hook file both enable it; lint failure exits 2 with the report on stderr; stop_hook_active: true skips; missing nextflow and a non-git directory exit 0; a clean tree exits 0; quoted paths with spaces are passed through and deleted files are not
• CI validate workflow