task-eligibility

name	task-eligibility
description	Task Eligibility Skill

Task Eligibility Skill

Tags

eligibility, task-class, routing, local, gemini, codex, claude, thermal, batch, background, MLFQ

When to Use

Deciding which agent to route a task to; understanding why a task was refused by local model; thermal-critical conditions; choosing between delegate-to-local modes; understanding batch vs background task class restrictions.

---

1. Agent Capability Matrix

Capability	Claude	Gemini (auto_edit)	Gemini (yolo)	Codex	Local (direct)	Local (agent)
File read	✓	✓	✓	✓	✗	✓
File edit	✓	✓	✓	✓ (patch)	✗	✓
Shell exec	✓	✗	✓	✓	✗	limited
Web fetch	✓	✓	✓	✗	✗	✗
RAG query	✓	✗	✗	✗	✗	✓ (MCP)
Live service query	✓	✗	✓	✓	✗	✓
Multi-turn reasoning	✓	✓	✓	✓	✓	limited
Git operations	✓	✗	✓	✓	✗	✗

---

2. Local Model Task Eligibility

The local model (Qwen3-35B) is bounded by hardware and MLFQ scheduling.

Mode Selection

Mode	Best for	Constraints
--mode direct	Reasoning, analysis, debate, structured JSON output	No tool access; cannot call live services; max ~512 tokens
--mode agent	Tasks needing RAG/memory/coordinator queries	Tool calls via MCP; slow at 1 tok/s; 180 token output ceiling (coordinator)
--mode hybrid	RAG-grounded lookups ONLY	Coordinator intent classifier may defer "planning" tasks to Claude (empty response)
--mode auto	Default; heuristic selects direct/agent/hybrid	May mis-classify edge cases

Task Class Filter (MLFQ)

Task Class	Eligibility	Notes
background	✓ always	Lowest priority, runs when CPU idle
batch	✗ under thermal=critical	CPU temp ≥85°C blocks batch; use background instead
interactive	✓ L0 priority	Set agent_type=human in coordinator delegate payload
coding	✓ implementer only	Detected by _CODE_TASK_RE in agent_executor.py

Thermal gate: when thermal_tier == critical (≥85°C), batch tasks are refused by MLFQ. Use background task class or defer to remote agent.

---

3. Gemini Eligibility Rules

Gemini has two modes with different eligibility:

auto_edit mode (default for most delegation)

• File read/edit/search: ✓
• Shell commands: ✗ (run_shell_command does not exist — wastes a turn)
• Live service queries: ✗
• Git operations: ✗
• Web fetch: ✓ (web_fetch tool available)

Eligible task types in auto_edit:

• Code editing within declared slice
• File content verification (acceptance criteria check via grep_search)
• Research and synthesis (reading files, searching patterns)
• Review (using file-verifiable criteria only)
• Writing new files

Not eligible in auto_edit:

• Running tests (aq-qa, pytest)
• Curling live endpoints
• Anything requiring shell execution

yolo mode (explicit --mode yolo)

• Full shell access: ✓
• All git operations: ✓
• Live service queries: ✓
• Same eligibility as Codex

---

4. Codex Eligibility Rules

• Shell access: ✓ (but requires < /dev/null stdin)
• Large prompts: must use --prompt-file /tmp/file.txt (not inline)
• Edit format: uses apply_patch (unified diff) — request patches, not replace calls
• Git operations: ✓
• RAG queries: ✗ (no MCP access)

Eligible for: large implementation slices, complex multi-file refactors, integration work, final acceptance when orchestrator needs a clean diff.

---

5. Task → Agent Routing Recommendations

Task type	Preferred agent	Fallback
Quick analysis, reasoning	Local (direct)	Claude
RAG lookup, memory recall	Local (agent)	Claude
Code edit (bounded slice)	Gemini (auto_edit)	Codex
Code edit (needs shell validation)	Gemini (yolo) or Codex	Claude
Large refactor (multi-file)	Codex	Claude
Review (file-verifiable)	Gemini (auto_edit)	Claude
Review (needs runtime check)	Claude	Gemini (yolo)
NixOS module changes	Claude (needs rebuild context)	Codex
Emergency fix (runtime needed)	Claude	Gemini (yolo)

---

6. Refusing a Task (Implementer Self-Check)

If you receive a task that exceeds your eligibility:

ELIGIBILITY_REFUSED: <task type> exceeds this agent's capability class
Agent: gemini/auto_edit
Reason: task requires shell execution (aq-qa validation) — not available in auto_edit mode
Recommendation: re-delegate to Gemini yolo mode or Claude

Do NOT attempt a task you can't complete and then fail silently — state the limitation early so the orchestrator can re-route efficiently.

---

7. Lane Selection (Cost-Aware Routing)

From CLAUDE.md Rule 4: "Prefer local inference for bounded tasks; remote only when task value justifies cost."

Local model:  0 API cost, 1-2 tok/s, 180 token output ceiling, thermal-sensitive
Gemini:       API cost, fast, no shell in auto_edit
Codex:        API cost, fast, needs /dev/null stdin
Claude:       API cost, highest quality, full tool access

Decision threshold:

• Task output < 180 tokens AND no shell needed → local (direct or agent)
• Task needs file edits + is bounded → Gemini (auto_edit)
• Task needs shell validation → Gemini (yolo) or Codex
• Task needs codebase understanding + high quality → Claude