meta-dev

Stand up and drive a campaign runbook — a single orchestration manuscript that sequences multiple related plans by dependency and walks them through the 6-stage waterfall in serial/parallel waves, with a live computed dashboard. Use when coordinating a multi-plan feature arc (not a single plan — that's /meta-dev). Invoked by the /runbook command.

The shared execute→review→fix loop for headless plan execution. A fresh worker per task (DeepSeek default via claude-headless-exec, GLM/Sonnet-200K on flag, Codex review-only) executes and self-runs its task Verify hook; review gates at the PHASE boundary (not per task) via the meta-dev:review-agent Opus subagent, which computes its own git diff and returns a verdict; a fixer worker repairs on non-pass (tier-specific fix ladder). Only N worker lines + one phase verdict return to the conductor — diffs never enter the main thread. Used by /meta-execute --deep|--glm|--sonnet|--codex and /auto-execute. Protocol: references/loop-protocol.md.

Patterns for headless Claude execution via claude -p --output-format json. Tool allowlist tiers, env inheritance, model override for Sonnet routing.

Stand up and maintain a visible stage-level task list for the /meta-dev 6-stage waterfall. Use in autopilot/walk runs so the user can walk away and watch Brainstorm→Design→Plan→Harden→Execute→Review progress live. Stage granularity only — never mirrors /meta-execute's per-task list.

Forward-moving compaction handoff. Use when a long session is getting context-heavy mid-project and you want to compact WITHOUT losing momentum — writes a durable, forward-looking handoff so the post-compaction session resumes the exact next step instead of being left hanging. Also used to proactively flag clean compaction boundaries during work. Triggered on-demand via /meta-compact or proactively when a clean boundary meets a heavy context.

Grug-brain sniff test for plans AND code — detect over-engineering, complexity demons, vague tasks, and bad practices in plans/specs (primary) or code (secondary). Report-only, never edits. Invoked as /sniff-test. By default sniffs the subject under discussion in the current conversation; takes a path, or --diff/--staged/--all. Use before executing a plan, before merging code, or when something "smells off".

Batch engineering changelog. Append entries during work, cut on demand or auto-threshold, draft release post, pair with version bump.

Structured code review across 5 dimensions — correctness, safety, patterns, coverage, scope. Verdict-routed to auto-fix or surface.

Classify tasks as HOTL-safe (auto-execute) or HITL-required (human gate) by blast radius. Use before dispatching plan execution.

Generalized deploy step for the ship pipeline. Ships one target — either push-to-branch (server auto-deploy hook) or rsync+restart over SSH, or publish a built artifact. Config-driven, no project specifics hardcoded.

Design quality rubric scoring — evaluates implementation against design doc on 4 dimensions

Judgment-based plan quality checks — cross-reference integrity, dependency ordering, contract completeness, terminology consistency

Generate definition-of-done contracts from task descriptions. Output structured acceptance criteria, test plan, rollback strategy, and blast radius.

Process inbox items autonomously — fix auto-clearable issues via haiku/sonnet, escalate hard items to opus, surface advisories to user. Goal: drive issue inbox to zero, surface advisories clearly.

Attempt auto-fix with 3-attempt cap, smallest-fix-first strategy, failure dossier to inbox. Use when verify step fails.

Multi-repo, context-aware versioning. Declarative config, semver/calver/custom strategies, follows cascades, atomic version_files updates.