Agent Lint

Relies on BDK foundation (STARTUP_INSTRUCTIONS.md). Assumes environment discovery has already run.

Agent Lint

Verify BDK agent file meets portability and quality conventions.

Determine target

• $ARGUMENTS: one or more space-separated paths.
• No arguments: ask user which file, or infer from context.
• Agents live at agents/<name>.md — flat file, no subdirectory.

Per path: read file, run all checks. Print separate report block per file.

Run checks

Report each check as PASS, FAIL, or WARN with one-line reason.

1. Frontmatter present

• File must start with ---
• Must contain name: and description: fields

2. Model field — REQUIRED for BDK agents

FAIL if missing or empty. Valid: haiku, sonnet, opus, inherit, or full model ID (e.g. claude-sonnet-4-6, claude-haiku-4-5).

• inherit = uses parent model (valid but WARN — pin for predictable cost/behavior)
• Omitting model: = defaults to inherit — FAIL for BDK agents
• WARN if not standard alias or known model ID

3. No absolute paths

FAIL if file contains:

• /Users/, /home/, /root/, /opt/, /var/, C:\
• Any hardcoded filesystem path

4. No project-specific file references

WARN if file refs files only in one project:

• Patterns: src/foo.py, app/models/user.rb, internal/auth/handler.go
• Exception: generic illustrative examples (e.g. <your-file>, path/to/file)

5. No project-specific instructions

WARN if file contains phrases tied to specific project:

• "in this project", "this repo", "our codebase", specific org/team names
• Hardcoded branch names, database names, service names

6. No language-specific commands

FAIL if file hardcodes language/framework tooling:

• Test runners: pytest, go test, npm test, yarn test, cargo test, rspec, jest, mocha
• Build tools: mvn, gradle, cargo build, make (as build command)
• Linters: ruff, eslint, golangci-lint, rubocop, flake8

Exception: agents explicitly scoped to one language (name contains language, or description states it).

7. Skill cross-references use full namespace

Refs to BDK skills must use /bdk: prefix.

• FAIL: /commit, /debug
• PASS: /bdk:commit, /bdk:debug

8. name format valid

Must match ^[a-z0-9-]{1,64}$. FAIL on uppercase, underscores, spaces, or >64 chars.

9. description length

WARN if description: >250 chars.

10. Invalid frontmatter fields

FAIL if unsupported frontmatter fields used.

Valid agent fields: name, description, model, tools, disallowedTools, permissionMode, maxTurns, skills, mcpServers, hooks, memory, background, isolation, color, initialPrompt, effort

Skill-only fields invalid in agents (FAIL): argument-hint, disable-model-invocation, user-invocable, allowed-tools, context, agent, paths, shell

10a. Plugin-unsafe fields

If agent in plugin's agents/ directory (not .claude/agents/ or ~/.claude/agents/):

• WARN if frontmatter uses hooks, mcpServers, or permissionMode — silently ignored in plugin agents. Tell user to copy to .claude/agents/ if needed.

11. tools field format

If tools: present: must be string or YAML list of known tool names.

• Valid: Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch, Agent, mcp__*, Agent(name) syntax
• WARN on non-standard tool names
• PASS if absent (agent inherits default toolset)
• NOTE: disallowedTools can be used alongside or instead of tools

12. $ARGUMENT typo

FAIL if body uses $ARGUMENT (no S) — correct is $ARGUMENTS or $ARGUMENTS[N].

13. Agent file placement

Must be flat .md directly under agents/ — not in subdirectory.

• FAIL: agents/foo/bar.md
• WARN if filename doesn't match name: field

14. External agent dependencies

If agent invokes other agents by name:

1. Check agents/<name>.md exists in BDK repo.
2. Not found → WARN: unknown agent dependency, not shipped with BDK.

15. Safety rules declared

Read-only agents (analysis, review, exploration) should declare safety constraints.

• WARN if no file-modification tools in tools: AND body lacks "MUST NOT modify" or "read-only".
• PASS if tools: has only read-only tools (Read, Grep, Glob, WebFetch, WebSearch; Bash excluded).

16. Terminal output pattern

Agents with terminal output should define start/complete banners.

• WARN if no output format defined — hard to track in multi-agent workflows.
• PASS if body has output format block or states "no output required".

Output format

One report block per file.

AGENT LINT: agents/test-runner.md
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PASS  Frontmatter present (name, description)
PASS  model: haiku defined
PASS  No absolute paths
PASS  No project-specific file references
PASS  No project-specific instructions
WARN  Language-specific command: pytest (line 31) — use "run the project's test suite" or scope agent to specific language
PASS  Skill references use /bdk: namespace
PASS  name format valid
PASS  description length OK
PASS  No invalid frontmatter fields
PASS  tools: field format valid
PASS  No $ARGUMENT typos
PASS  Agent file placement correct
PASS  No unknown external agent dependencies
WARN  No safety rules declared — add "MUST NOT modify files" if read-only intent
PASS  Terminal output pattern defined
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
0 failures, 2 warnings

All files pass → end with: All checks passed.

No fixes beyond check results unless user asks.

Checklist before responding

Run ALL 16 checks (not just seemingly relevant ones):

• 1. Frontmatter present
• 2. Model field — required (FAIL)
• 3. No absolute paths
• 4. No project-specific file references
• 5. No project-specific instructions
• 6. No language-specific commands
• 7. Skill cross-references use full namespace
• 8. name format valid
• 9. description length
• 10. Invalid frontmatter fields
• 10a. Plugin-unsafe fields (hooks/mcpServers/permissionMode in plugin agents)
• 11. tools field format
• 12. $ARGUMENT typo
• 13. Agent file placement
• 14. External agent dependencies
• 15. Safety rules declared
• 16. Terminal output pattern

Don't output checklist. Use internally only.