Imported from rjmurillo/ai-agents (src/claude/AGENTS.md). Install upstream with npx skills add rjmurillo/ai-agents --skill claude. Copyright stays with the author.

src/claude/

31 hand-maintained Claude Code agent prompts; the claude-agents plugin source in the rjmurillo/ai-agents repository, installed by Claude Code as .claude/agents/<name>.md.

Matters

• 31 agent files (<name>.md), plus claude-instructions.template.md (no frontmatter, not an agent) and security/references/ (docs the security agent points at). None of it is generated: the drift checker's own comment states "Claude agents have unique content and are NOT generated from templates."
• name, description, argument-hint appear in all 31 files. metadata.role in 25 of 31. A tools: frontmatter list in only 2 of 31 (analyst, security). A model: pin in exactly 1 of 31 (code-reviewer: haiku, with a required model-rationale: line, per ADR-080).
• No enforced section structure. Measured across the 31 files: ## Core Identity appears in 17, ## Constraints in 10, ## Memory Protocol in 13, ## Handoff Options in 13, ## Output Format in 7. Grep the specific file before assuming a heading exists.
• Changing shared agent behavior means editing this file AND, in the rjmurillo/ai-agents repository, templates/agents/<name>.shared.md, .claude/agents/<name>.md, and .github/agents/<name>.agent.md in the same change, then regenerating. The automated co-change check only confirms the diff touched the required siblings; nothing compares their text.
• All 31 files are byte-identical to their .claude/agents/ installed copies today (verified by diff). That is upheld by convention plus the co-change check, not a content gate, so re-diff rather than assume it still holds.

Entry points

• <name>.md: edit directly for Claude-only behavior (MCP tool ids, Serena calls, Task syntax).
• Task(subagent_type="<name>", prompt="..."): how another Claude Code agent invokes one of these.
• In the rjmurillo/ai-agents repository: uv run python build/generate_agents.py, run after any shared-behavior edit to refresh the generated Copilot CLI and VS Code mirrors.

Where to look

Skip

• Nothing under this tree is generated, so there is no output copy to avoid editing here; the trap runs the other way (see Dangerous assumptions).
• .claude/agents/<name>.md, .github/agents/<name>.agent.md (in the rjmurillo/ai-agents repository): sibling installed copies, edited alongside this tree, never instead of it.

Constraints

• Cross-harness behavior: read agent-harness-reference first; hook, event, or generated-Copilot changes run through ai-agents-portability-campaign.
• An agent file here MUST NOT reference .agents/, build/, or scripts/ paths that will not exist for a downstream installer of this plugin.
• A model: field MUST NOT be added without an ADR-080 KEEP_PIN manifest entry, or the haiku cost exception plus a model-rationale: line.
• Changing shared behavior MUST also touch templates/agents/<name>.shared.md in the repository; the repository's parity check fails a solo template edit but does not fail a solo edit to this file alone, so the discipline is on the author, not the gate.
• A script path built from ${CLAUDE_PLUGIN_ROOT:-.claude} resolves correctly only two ways: unset, falling back to .claude/ (in-repo dev use), or set to a plugin root that itself ships a skills/ directory. src/claude/ ships none, so this pattern only works when the project-toolkit plugin (.claude/) is also installed alongside claude-agents.
• GitHub operations should route through a skills/github/scripts/... script rather than raw gh, and memory search through a skills/memory/scripts/search_memory.py script, subject to the availability constraint above.

Dangerous assumptions

• Assuming a passing drift check means an agent here agrees with its src/vs-code-agents/ counterpart is wrong: the check is a word-set similarity floor over an allowlist of section names, not equality, and several comparisons score a hardcoded 100.0 when the agent has none of the allowlisted sections at all.
• Assuming the shared-template "required sections" list (templates/README.md: Core Identity, Core Mission, Key Responsibilities, Constraints, Memory Protocol, Handoff Options; Activation Profile is common but not on that list) holds for files in this tree is wrong: none of those headings appears in more than 17 of the 31 files, and nothing enforces the list here.
• Assuming ${CLAUDE_PLUGIN_ROOT:-.claude}/skills/... always resolves for a claude-agents-only install is wrong: this plugin ships no skills/ directory of its own (see Constraints).
• Assuming a solo edit here is safe because "the agent still works" ignores that no automated check requires the matching template edit in the same change; only the co-change diff check runs, and it does not require this direction.

Dependencies

• Feeds .claude/agents/<name>.md (hand-copy, currently identical content) and is read, by filename only, by the repository's drift checker for the src/claude vs src/vs-code-agents comparison.
• The repository's parity check (co-change gate, blocks a solo template edit) and drift checker (similarity floor, weekly cron) are the only two automated checks over this tree; neither proves content agreement.
• Consumed by the claude-agents marketplace plugin entry, whose source is exactly this directory; nothing above it (docs, governance, build tooling) ships to an installer.

Architecture

• This tree is a hand-maintained fork of the shared template body, not a generated copy: the shared source lives in a separate file per agent in the repository, and the two are kept in step by author discipline plus the co-change check, never by a tool that compares their content.
• Every one of the 31 agent files has a templates/agents/<name>.shared.md counterpart; claude-instructions.template.md has none, and the drift checker's skip list (_NON_AGENT_FILENAMES) covers only AGENTS and CLAUDE, so it is reported as NO COUNTERPART.

Commands

# All of the following run from the rjmurillo/ai-agents repository root, not from an installed plugin.
uv run python build/generate_agents.py                                              # refresh Copilot CLI + VS Code mirrors after a shared edit
uv run python build/scripts/validate_install_parity.py --files src/claude/<name>.md  # check the co-change requirement
uv run python build/scripts/detect_agent_drift.py --changed src/claude/<name>.md     # scope the similarity check to one agent