agents opus

Quick reference

Field	Value
Command	/brewcode:agents
Input	free-form prompt; empty defaults to read-only status
Specialist	brewcode:agent-creator
List source	.claude/agents/, ~/.claude/agents/, ${CLAUDE_PLUGIN_ROOT}/agents/ (shipped, read-only)
Model	opus

When to use

• Check the status of all agents across project and global scopes
• Create a new agent from a plain-English description
• Improve an existing agent’s trigger phrases or system prompt
• Review an agent for quality, focus, and trigger coverage
• Sync agent files against real code/CLAUDE.md/rules after a repo, session, or commit changed (memory sync)
• Batch-create or batch-improve a set of agents in one pass

Examples

/brewcode:agents
/brewcode:agents status of my agents
/brewcode:agents create a backend validator agent
/brewcode:agents improve reviewer
/brewcode:agents sync
/brewcode:agents синк по текущей сессии

How it works

Modes

Mode	Trigger	Action
status	default for any “show me” intent — “статус”, “что есть”, “состояние”	Inventory by scope, state, overlaps, health flags
list	explicit only — “list” / “список” / “перечисли”	Globs *.md across .claude/agents/, ~/.claude/agents/, ${CLAUDE_PLUGIN_ROOT}/agents/ (shipped, read-only), plus brewcode/agents/ only inside the plugin workspace (test -d brewcode/.claude-plugin)
create	”создай” / “create” / “new” / “добавь” / “scaffold”	Scaffolds a new agent file via agent-creator; asks scope, model, CLAUDE.md update
improve	”улучши” / “improve” / “refactor” / “fix” / “почини”, or a bare existing <name|path>	Tunes description triggers and system prompt quality
review	”ревью” / “review” / “validate” / “проверь корректность”	Single reviewer — the project’s agent from .claude/agents/, else general-purpose — two-phase: review, then double-check the findings
sync	sync, синк, memory sync, актуализируй, обнови знания, приведи в соответствие с кодом	Memory sync: rewrites agent files against ground truth via the shared reference; non-growth guaranteed
batch	auto-detected fan-out flag	Not a menu item — auto-detected from plural form, “все” / “all”, or multiple names/paths; triggers parallel delegation when creating or improving multiple agents

Reference files

agent-creator and hook-creator used to carry all of this inline; it now lives in this skill’s own references/ directory and each agent reads only the file it needs, on demand. 11 files, each under 200 lines.

File	Used by	Holds
agent-frontmatter-fields.md	agent-creator	AG file format, frontmatter field reference (incl. experimental.cacheTtl)
agent-scope-and-tools.md	agent-creator	Available tools, model precedence, scope rules
agent-context-and-execution.md	agent-creator	Subagent context inheritance, execution modes, resource limits
agent-template.md	agent-creator	Description budget, system-prompt structure, guardrails, validation checklist
agent-known-issues.md	agent-creator	Known bugs, limitations, version history, debugging
hooks-events.md	hook-creator	Session lifecycle order, 33-event table, matcher pattern syntax
hooks-io-contract.md	hook-creator	Stdin fields, message-routing matrix, exit codes, output schemas
hooks-types-config.md	hook-creator	5 hook types and their config fields, precedence
hooks-env.md	hook-creator	Env vars, sensitive-path prompt, project-root resolution
hooks-templates.md	hook-creator	Bash/JS hook skeletons, fail-safe design rules, pre-ship checklist
hooks-changes.md	hook-creator	Changes through 2.1.285 and known contract limits

Delegation

Ordinary subagents receive bounded context through the Agent prompt and cannot use AskUserQuestion; unresolved user decisions return to the main caller. Conversation forks retain parent history and tools. A skill’s context: fork is an ordinary subagent, not a conversation fork. Upstream nested-agent depth support does not override this project’s main-caller ownership of delegation, decisions, and acceptance. A partial maxTurns return is unfinished work, even when files were written; inspect checkpoint evidence before resuming or declaring completion.

Plugin agents ignore frontmatter hooks, mcpServers, permissionMode, and initialPrompt. Place plugin hooks in hooks/hooks.json; use the appropriate local/session surface when those agent fields are required. Local files and CLI JSON also have different supported fields. The hook references cover 33 events and five handler types with event-specific output and type support.

One subagent = ONE bounded unit — here, ONE agent definition: roughly 5 files, roughly 10 steps. “Create the whole roster” is never one task; a batch is split into N spawns fired in ONE message, one agent-creator per agent file.

Every spawn prompt carries six fields. A bare one-line task is never enough:

Field	Content
GOAL	the overall task and why it exists — the point beyond the file edit
ROLE	what this agent owns; what it must NOT touch
SCOPE	exact paths/commands in bounds + explicit out-of-bounds
CONTEXT	what is already done, by whom, what runs in parallel — trimmed to what THIS agent needs
CONSUMER	who or what uses the result next, and the shape it must fit
DONE	acceptance criteria + the exact report shape you want back

For create, that means the specialist is told the scope and model were already decided (do not re-ask), which agents already exist and must not be duplicated, how many siblings run in parallel, and that its description line drops verbatim into the CLAUDE.md agents table row written right after it finishes.

Create / improve parameters

Parameter	Options	Notes
Scope	Project / Global / Plugin	Project → .claude/agents/; Global → ~/.claude/agents/; Plugin → brewcode/agents/, offered only when test -d brewcode/.claude-plugin succeeds (the dev/plugin workspace) — elsewhere the option is dropped, since it would write a stray <cwd>/brewcode/agents/
Model	sonnet / opus / haiku / inherit, or a verified supported model id	Inherit uses the caller’s model; choose from the current supported catalog rather than an assumed alias
CLAUDE.md update	yes / no	Adds or updates the agent row in the agents table; appends a new section if none exists
Improve focus	triggers / system prompt / both / full review	Asked for improve mode only
Description	Canonical budget ≤150 tokens, approximately 600 characters	Drives discovery; ≤100 characters is an optional brevity target, not a second required cap
Specialist	brewcode:agent-creator	Handles file scaffolding and prompt engineering; never writes under ${CLAUDE_PLUGIN_ROOT} — the installed plugin copy is read-only

improve resolves the target across the writable scopes (project / global / plugin workspace) only. A name that only matches under the shipped, read-only ${CLAUDE_PLUGIN_ROOT}/agents/ is reported and the skill stops — it is never copied or edited.

Sync mode

Scope	Meaning
repo (default)	Whole working tree
session	This conversation only — decisions, user corrections, bugs hit
commit	git show/git diff <ref>, default HEAD

Announced as Sync scope: <scope> — <evidence> | targets: <N> agents. Implemented in brewcode/skills/skills/references/mode-sync.md (steps S1-S6) via the constant SYNC_REF = ${CLAUDE_SKILL_DIR}/../skills/references/mode-sync.md — the same file /brewcode:skills sync runs, so the two rosters cannot drift apart.

Non-growth directive: each synced file ends <= its original line count; total delta <= 0. Any positive delta needs per-line justification. Order: DELETE first, then FIX, then ADD last.

Auto-dispatch after create/improve: once a create or improve completes, the same reference runs again scoped to only the file just written — S3 ground truth through S5 verdicts, folded into one row of the final report. No second agent-creator spawn: the coordinator applies every verdict itself with targeted Edit calls, since S4 fan-out is skipped for this scoped pass. A clean file reports one line: sync: no drift.

Verdict	Action
STALE	Fix to minimal true form
DEAD	Delete, plus anything that only existed to explain it
DUPLICATE	Keep the single best location
OBVIOUS	Delete
DRIFT	Compress prose to a table row
MISSING	Add <= 1 line, only if verified against a real source and its absence costs a real failure

Report: per-target table (File | Lines before → after | Fixed | Deleted | Added | Key change), stale facts corrected, added items with source, skipped agents, total delta.