claude-dev-team

Phase-gated bug investigation → root-cause → fix → verify cycle. Enforces root-cause-before-edit, failing-test-first, holistic callsite scan, reopen/redesign gates (SPEC-029), multi-surface done matrix, and self-calibration checklist before any "done" claim. Subcommands: /debug <desc> (full), /debug patch <desc> (fast path), /debug arch <desc> (design-first → /kickoff handoff).

Brutally honest review of staged/modified files — no sugar-coating. Thin wrapper over the council engine with preset `diff-mode`: 5 specialist investigators (logic, security, compliance, quality, simplification) run in parallel, filtered at confidence 80. Blocks commit on critical or compliance findings. Optional path argument saves the review to a file.

Fact-dense rewrite of agent memory prose (tier-0 notes, digests) without losing technical substance. Companion to /memory-distill. Zero external deps.

Adversarial council tribunal engine — reality-checks claims with material evidence. Shared engine for /council and /review-and-commit (diff-mode preset). Blind investigators, evidence-or-silence rule, dual output shapes (verdict[] and finding[]), atomic verdict index at .claude/council/index.json, feedback-memory learning loop. Judge is a dedicated agent with an empty tool allowlist. See specs/core/SPEC-013-adversarial-council-tribunal.md.

Optional host SAST (Semgrep / CodeQL) feed for security review. Fail-open when tools are missing. Agent-internal + review-and-commit / council security flavor. Zero required deps.

Socratic design refinement — structured questioning that forces requirement clarification before any planning or implementation. Use before /kickoff for complex features, or standalone for early-stage ideation. Optional --grill for one-question-at-a-time interviews with recommended answers.

Behavior-preserving polish on recently modified code — clarity only, no logic or API changes. Used by /orchestrate after Tech Lead approve, before QA. Zero external deps (in-plugin; not Anthropic's marketplace plugin).

Full lifecycle orchestrator — fetches issue context, creates worktree, spawns agents end-to-end, enforces tech-lead review loops, and optionally ships a PR. You stay as observer/navigator; agents do all the work. Usage: /orchestrate CDV-1 or /orchestrate

Living project ubiquitous language (CONTEXT.md). Load before naming work; update when terms crystallize during brainstorm/kickoff. Agent-internal — not a user slash command.

Orchestrate the full ticket intake and planning phase — parallel PM+Tech Lead kickoff, spec creation, implementation plan, and TaskCreate task graph. Replaces 7 manual prompts with one command. Usage: /kickoff <TICKET-ID> "<ticket text>" or /kickoff alone to be prompted.

Scaffolds a new project with TDD workflow structure including .claude/plans/, specs/TDD.md, AGENTS.md, and .gitignore. Use when creating new projects, setting up project templates, initializing TDD workflow, or preparing AI agent collaboration structure.

Autonomous CI/test watcher — cron polls a ticket's PR checks (or local tests) every 7 min, spawns a fixer agent on failure (max 3), self-cleans when green. Prefers durable CronCreate; falls back to session-only when the harness denies durable.

Design-first code restructuring that preserves behavior. Enforces design problem written before any edit, characterization tests when coverage is thin, and zero observable behavior change. Subcommands: /refactor <desc> (default), /refactor inline <desc> (approach pre-decided by /debug or /orchestrate).

Manage project backlog items stored in .claude/backlog/ (index at .claude/backlog.md). Supports add, close, list, and init subcommands. Use when adding backlog items, closing/completing backlog items, listing the backlog, or initializing the backlog structure in a project.

Clean up after a ticket ships — verifies all tasks completed, removes the worktree, appends learnings to project memory, marks the plan complete, idempotently re-closes source tracking (backlog/Linear), and prints a Linear close-out checklist. Usage: /wrap-ticket <TICKET-ID>

Phase-2 deep-read protocol for `/retro`. Defines the exact prompt template, input contract, output JSON schema, and validation rules used when `commands/retro.md` spawns a subagent (via the Task tool) to analyze friction anchors flagged by the phase-1 gate. Not user-invoked. Read this file to learn the protocol; the calling command pastes the prompt template into a Task call and validates the returned JSON.

Bootstrap Agent Teams orchestration for any project. Enables bubblewrap sandbox with auto-detected network allowlist, bypassPermissions for zero-prompt agents, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS, and PostToolUse + Stop + TaskCompleted hooks. Creates/updates AGENTS.md with team coordination rules and CLAUDE.md as reference. Run once per project. Safe to re-run — existing files are merged, not overwritten.

Cold-handoff extraction protocol for `/handoff <uuid>`. Defines the orchestration flow, the fan-out invariant, and the five specialized extractor subagent prompt templates (Convergence / Dead-ends / Code-state / Open-threads & conflicts / Basics) with a strict single-object JSON output schema each. Not user-invoked. `commands/handoff.md` reads this file, fills the substitution variables, spawns the five extractors in ONE tool-use block, and hands their JSON to `prepass.sh finalize` for merge. Implements SPEC-018 M4 (sections), M5 (lightweight stated-intent-vs-git flag), and M6 (pointers, not dumps).

Shared read-only parsing seam for Claude Code session transcripts (~/.claude/projects/*.jsonl). Locates a session's canonical file, assembles a deduped chronological timeline, exposes parse primitives, and guards against mid-write files. Consumed by /handoff (SPEC-018) and /retro (SPEC-012).

DevOps-led war-room for production incidents (SPEC-027). Severity triage, parallel RO investigation, append-only timeline, propose-only mitigation, comms drafts, cold postmortem → backlog. User entry: /incident.

Umbrella decomposition and sequenced orchestration (SPEC-025). PM+TL jointly decompose an epic into child tickets + cross-ticket DAG; persist via backlog (Linear optional); walk ready children by handing each to /kickoff or /orchestrate. Composition layer only — never reimplements the ticket lifecycle. Usage: /epic <EPIC-ID> ["text"] | status | complete | block | unblock | --redecompose

Status snapshot for active agent team work — reads TaskList and each agent's context.md, produces a table of who is doing what, what's blocked, and what's ready to claim. Use during Phase 3 implementation to monitor parallel agent progress. Usage: /standup [TICKET-ID]

Write agent memories to the SQLite database (or fall back to .md files). Handles DB detection, SQL-safe INSERT/UPDATE, optional embedding generation (lembed or remote embedding provider), and retry on SQLITE_BUSY. Usage: read this file to learn the protocol, then execute the relevant bash blocks.

Claim extraction, investigation, and cross-agent pair-judge prompt templates for /validate-memory. Defines the LLM-driven per-claim validation pipeline (claim extractor Step 3, investigator Step 4 Tier B) and the --reconcile pair-judge contract (Steps R1–R4). Not user-invoked — consumed by commands/validate-memory.md.

Phase-1 friction gate for /retro. A fast, deterministic, LLM-free heuristic that scans a Claude Code session JSONL and decides whether the session contained enough friction to warrant a deep retrospective. Used by /retro, /kickoff, and /orchestrate to suppress no-op runs on smooth sessions.

Premise→implement→adversarial-refuters pipeline for a known bug ticket. Verifies the premise still holds, implements in a SPEC-016 worktree, spawns N qa refuters in parallel, writes .claude/fix-ticket/ report. Never commits or bumps versions. Usage: /fix-ticket <ticket-id> "<bug>" [--fix "..."]

Internal engine primitive — not user-invoked. Offloads mechanical, machine-verifiable tasks to a local model via OpenCode CLI; driven by /orchestrate and /local-do. Off by default; falls back transparently to Claude when disabled or preflight fails.

Install & config diagnostics for the dev-team plugin (SPEC-022). Read-only by default; --fix applies a narrow allowlist. User-facing entry: /doctor (namespaced dev-team:doctor).

SPEC-026 outcomes ledger writers/readers plus CDV-187 read-only rollup. Helpers: emit-outcome.sh (write), outcome-rates.sh (advisory rates), rollup.sh (display). User-facing entry: /metrics.

Deterministic, LLM-free structural docs-consistency checker (SPEC-010 D1–D8). Checks: cmd-index (README ## Commands ↔ commands/*.md), agent-roster (AGENTS.md + README ↔ agents/*.md), docs-hub (docs/commands links/orphans), manifest-desc (plugin.json description == marketplace plugins[].description). Wired by /release as Step 4.9 after T3. Run manually via: bash skills/docs-drift/check-docs-drift.sh [--root DIR]

Generate unit/integration tests from behavioral specs. Reads MUST/SHOULD requirements from specs/core/ or specs/TDD.md, detects the project's test framework, and writes test files with one test case per requirement. Each test is tagged with its source spec for traceability. Run after /generate-specs or /create-spec to close the spec-to-test gap.

Bump version across all required files (CHANGELOG.md, plugin.json, marketplace.json), commit, tag, and push. Use when releasing any version of this plugin. Ensures all three version files stay in sync — never skips any of them.

Loop-prompt architect protocol — designs reviewed, file-persisted loop programs for the built-in /loop and /goal commands. Library at .claude/loops/ with a journal convention and decision-card escalation. Consumed by /craft-loop (craft, refine, list modes). Ships no runtime. Supports hold/dogfood (no-write) and declared side artifacts under .claude/loops/.

Multi-branch release queue sequencer (SPEC-023). Register ready feature branches, freeze deterministic landing order + slot versions, land each branch as uncommitted tree on master/main, mechanically pre-resolve known conflict classes (TDD index, Version-History, CHANGELOG, version JSON), then invoke /release with the explicit assigned version. Abort-safe, dry-run, never reimplements /release internals.

Multi-team blind peer review with automatic quorum analysis. Spawns N unconstrained reviewer agents + M lens-differentiated reviewer agents in parallel, then clusters their independent findings by semantic similarity to produce a confidence-tiered findings report. Optionally forwards Tier 1 consensus findings to /council for reverse validation. Use when you want adversarial, multi-perspective coverage of an entire codebase, a specific directory, or a set of files. Usage: /blind-review or /blind-review --teams 3 --lenses security,contributor,spec --target path/

Interactive demo of the dev-team workflow. Scaffolds a tiny project in a temp worktree, injects a ticket, and runs the full agent pipeline so you can see real agents working. Cleans up after. Usage: /demo [orchestrate|kickoff|specs]

Reverse-engineer behavioral specs from existing source code. Reads the codebase, groups the public surface by domain, and writes MUST/SHOULD/MUST NOT specs from what the code actually does. All output is marked INFERRED and requires human review. Designed for legacy projects with no existing specs. Run once to establish a baseline, then use /reflect-specs to keep them current.

Search and retrieve agent memories from SQLite DB with semantic or keyword search

Deterministic, LLM-free linter for fenced bash blocks in plugin .md files (SPEC-021). Checks: C1 cross-block variable scope, C2 zsh history-expansion hazard, C3 zsh-fatal unguarded glob, C4 captured inline-PRAGMA sqlite poison. Run by /release as a pre-commit gate (Step 4.8). Not user-invoked directly; run manually via: bash skills/skill-lint/check-skill-bash.sh [FILE...]

Research new/updated Claude Code plugins, MCP servers, and skills released in the last week (or custom time window). Evaluates each against the current dev-team plugin setup, identifies gaps, and proposes enhancements with priority and effort. Usage: /scout-plugins or /scout-plugins 2w or /scout-plugins 30d