Agents catalog

8 agents. Each is a markdown file under agents/ with frontmatter declaring its tools and model.

Agent	Model	Tools	When it fires	What it produces
engineer	sonnet	Read, Grep, Glob, Bash, Write, Edit	Default for every /frappe-stack:build. Turns PM intent into a Stack Blueprint payload.	A live blueprint on staging, status=Applied.
reviewer	sonnet	Read, Grep, Glob, Bash	After every build, before every promote.	A CRITICAL/HIGH/MEDIUM/LOW report. Blocks promote on any CRITICAL or HIGH.
tester	sonnet	Read, Grep, Glob, Bash, Write, Edit	After every build that adds or modifies code.	FrappeTestCase tests + Playwright E2E + a coverage report. Gate ≥ 80% on changed lines.
deployer	sonnet	Read, Grep, Glob, Bash	Owns /frappe-stack:promote end-to-end.	Pre-promote checklist run, branch + commit + PR + merge watch + smoke-test.
analyst	haiku	Read, Grep, Glob, Bash	Read-only reports — counts, A/B results, audit-log queries.	A table + 1-line interpretation + the exact filter used.
migrator	sonnet	Read, Grep, Glob, Bash	Schema changes — rename, fieldtype change, naming-rule change.	A migration plan with patches.txt entry, rollback DDL, test plan.
documenter	haiku	Read, Grep, Glob, Write, Edit	After every successful build, on /pull, before every promote.	Auto-generated docs/doctypes/<name>.md, docs/workflows/<name>.md, CHANGELOG entry.
orchestrator	sonnet	Read, Grep, Glob	Multi-step requests spanning multiple agents.	An ordered plan (sequential vs parallel steps); dispatches the chain.

How they pair up

Trigger	Sequence (→ = sequential, ‖ = parallel)
/frappe-stack:build doctype X	engineer → (reviewer ‖ tester) → documenter
/frappe-stack:build chain (DocType + Workflow)	engineer (DocType) → engineer (Workflow) → (reviewer ‖ tester) → documenter
/frappe-stack:review	reviewer ‖ tester (combined report)
/frappe-stack:promote	deployer (which calls reviewer + tester internally)
Schema change	migrator → engineer → tester → reviewer → documenter
"Show me the experiment results"	analyst

Model choice rationale (per ~/.claude/rules/common/performance.md)

• Haiku (analyst, documenter): high-frequency, read-only or templated. 90% capability at 3× cost savings.
• Sonnet (everything else): code generation + reasoning, where wrongness is expensive.
• Opus (none): reserved for architectural decisions and deep analysis. None of these agents need it.

Pairing with global agents

The plugin's agents wrap and extend global ones from ~/.claude/agents/:

Plugin agent	Calls global agent
reviewer	frappe-reviewer (hard rule layer) + security-reviewer (when novel attack surface)
migrator	database-reviewer
tester	tdd-guide

What they refuse

Each agent's .md file lists its refusals. Quick highlights:

• engineer: refuse blueprints failing local guardrails (reserved name, fieldtype whitelist, workflow shape). Refuse to run on is_production=1 site. Refuse to skip user confirmation step.
• reviewer: refuse to approve any build with even one CRITICAL or HIGH finding. Refuse to run without semgrep available.
• tester: refuse to mock the database. Refuse to skip negative-path tests. Refuse to delete hard-to-test code to game coverage.
• deployer: refuse Friday-after-14:00 without --emergency. Refuse bundling >5 unrelated changes. Refuse force-merge bypassing branch protection.
• migrator: refuse one-shot renames. Refuse migrations without a backup verified in last 24h.
• analyst: refuse to mutate. Period.
• orchestrator: refuse to run agents in parallel when they share state. Refuse to skip reviewer "because the build was straightforward."

Adding an agent

1. Create agents/<name>.md with frontmatter:

   ---
   name: <name>
   description: <when to invoke — proactively / when triggered / never>
   tools: <minimum needed; smaller is safer>
   model: haiku | sonnet | opus
   ---
   

2. Body: workflow steps, refuses, pairs-with.
3. Add a row to the table above.
4. If it wraps a global agent, document the wrap in the row.