dark-factory

Open a new gnome-terminal window running claude /remote-control

Watch CI checks on a PR until completion or failure. Spawn fix handlers for failures, retry up to maxIterations. Return pass/fail status.

Watch for and resolve review comments on a PR. Handle each unresolved thread, re-run CI after resolving, retry up to maxIterations. Return resolved/failed status.

Debug a non-obvious bug. Creates a worktree, runs debugger-agent, and opens a PR with the fix.

Terminates all other running Claude / dark-factory terminal sessions and spawns one fresh factory terminal, leaving the user with exactly one active terminal.

Orchestrates automated code review by spawning high-level and low-level reviewers in parallel, then running the resolver in a loop until all issues are resolved.

Reviews code against a plan file for structural and architectural conformance. Appends high-level IssueItems to tmp/issues.md.

Reviews code at the function level for bugs, untested paths, inter-agent conflicts, and refactor opportunities. Appends low-level IssueItems to tmp/issues.md.

Reads tmp/issues.md, applies fixes for each unchecked item, checks them off, and returns whether any items remain unresolved.

Orchestrator for the investigation command. Invokes investigation-agent, then loops calling claim-validator-agent until all claims verified. Commits verified doc via SubagentStop hook.

When writing hooks or other structured data into .claude/settings.json, always read the existing file, merge new entries, and deduplicate — never overwrite the whole file.

Use this skill when building a post-manufacture auditor that scans agent execution logs and transcripts for pipeline instruction violations expressed in the agents' own words or tool call patterns.

Use this skill when writing an agent that must retry a blocking external check (CI, polling, review threads) in a loop — structure the loop with a named max-iteration guard, a BREAK-after-first-fix pattern, and re-entrance of upstream checks after downstream fixes.

How to have an agent run a helper Python script, capture its stdout URL or value, and send a PushNotification only when the script succeeds — skipping silently on failure.

When a Claude Code PreToolUse hook intercepts an Agent tool call, the prompt and subagent_type live under .tool_input, not at the top level of the hook stdin JSON.

1 hook across 1 event