Subagents

Specialist subagents

Every specialty in the constellation has a dedicated subagent. Invoke it through the Task tool (subagent_type: <name>) instead of improvising outside your expertise — specialty work done inline is a defect even when it happens to be correct. The subagent's returned report is the evidence for that slice of the card.

The roster

Specialty	subagent_type	Call it for
Data & databases	athena-database-agent	datasets (HF/Kaggle/Braintrust), schema design, ingestion/transformation pipelines, data QA, ML data preparation
HuggingFace & data science	lucius	HF downloads/uploads/edits, datasets/transformers, EDA, statistical analysis, model training/eval, database upkeep
Prompt engineering	prompt-engineer	run-failure diagnosis, GEPA mutations, prompt A/Bs across the llm-entity-extraction + llm-mailroom surfaces
Docs & board	atom	READMEs/wikis/changelogs, doc-drift detection, Kanban board upkeep, inter-agent coordination, repo restructuring
Software & UI	hazel-ui-software-master	UI/UX, fullstack, HTML/JS/CSS, bug fixes, local network testing, app architecture
Systems & repo hygiene	jarvis-systems-maximizer	performance triage, memory/CPU hogs, disk usage, cache cleanup, repo bloat, process prioritization
vLLM serving	vllm-specialist	vllm serve configs, quantization, multi-LoRA, structured outputs, speculative decoding, parallelism, KV-cache/memory tuning, local or Modal-hosted vLLM behind DEFAULT_PROVIDER=vllm
Modal compute	modal-specialist	Modal apps (@app.function/@app.cls/@app.server), Images/Volumes/Secrets/Sandboxes, GPU choice, scale-to-zero, modal serve/deploy, the repo's deploy/modal_vllm.py apps
Codebase exploration	explore	fast read-only searches ("where is X", "how does Y work") before you edit
General multi-step	general	research/execution spanning several specialties with no single owner

Project agent files

Project agents are committed under .opencode/agents/ — version-controlled and reviewed like code:

File	Role
.opencode/agents/vllm-specialist.md	vLLM serving playbook (vLLM stable-grounded) + repo wiring + guardrails
.opencode/agents/modal-specialist.md	Modal SDK-grounded primitives/CLI/GPU discipline + the repo's deploy apps
.opencode/agents/prompt-engineer.md	the GEPA diagnostic evaluator / prompt engineer
.opencode/agents/PROMPT_ENGINEER_GEPA_PROVENANCE.md	provenance documentation for the prompt-engineer agent — not a callable subagent

Restart opencode after adding a project agent so the Task tool picks it up.

vllm-specialist, modal-specialist, and prompt-engineer are mode: all (DMR-020 for the first two) — callable through the Task tool as subagents and selectable directly as primary agents. Verify with opencode agent list.

Rules that make the roster work

• One specialist per concern. Do not ask a subagent to do another specialty's job; chain them (e.g. explore → vllm-specialist → modal-specialist → prompt-engineer).
• Brief the specialist like a card. Pass the card ID, the exact scope, the files/seams involved, and the evidence contract (what must be green, what the report must name). A vague prompt returns a vague report.
• The caller owns the work. Subagent output is a report, not a merged change: the calling agent writes the files, runs the gates, updates governance/TASKS.md, and commits — the specialist is evidence, not an author of record.
• Specialists verify current upstream docs (vLLM stable, Modal SDK, HF Hub APIs, Langfuse) before writing configuration; a report that quotes a version without checking it is incomplete.
• Skills are the second layer. Load the huggingface, langfuse, or customize-opencode skill when a task matches its description, and the package-local .cursor/skills/ (openrouter, ollama, modal, langfuse, apache-phoenix, braintrust, huggingface, langgraph, dojo-scoring, legalbench) for provider/sink depth inside a package.
• Out of scope stays out of scope. The prompt-engineer mutates prompts from evidence; it never changes what "correct" means, the scorer, the ground truth, or the schema — hand those to the owning surface.

Provenance: DMR-008 (2026-09-09) — the roster landed in AGENTS.md with the dedicated vllm-specialist + modal-specialist project agents; issue #12 closed with the card. The prompt-engineer project agent and its GEPA provenance file predate the roster.