aonios-agent

name

aonios-agent

description

Route heavy coding tasks to premium multi-model backends (Codex GPT-5.5/5.4 high, GLM-5.2 max, Qwen 3.7 max, Kimi K2.7 code-max, MiniMax M3 high-thinking). Mirrors vertex-coder's structure but shells out to local CLI binaries (codex, opencode) instead of Vertex AI. Two modes: direct (single-file) and interactive (multi-file autonomous with tool use). Use when delegating heavy code generation, debugging, or refactoring tasks to the strongest available model. Triggers on: "delegate to aonios agent", "use the strongest model", "premium model coding", "opencode glm", "codex gpt-5 high", "minimax m3 thinking", "aonios agent direct", "aonios agent interactive". Do NOT use for trivial edits (do them locally) or when the user explicitly forbids external delegation.

Aonios Agent

Premium multi-model coding dispatcher. Mirrors vertex-coder's structure (direct mode + interactive mode) but routes to whichever premium model is best for the job via local CLI binaries. No new Python deps — uses codex exec and opencode run directly.

What this skill IS

1. A model router — picks the right CLI (codex vs opencode) and the right model flag for the requested premium model.
2. A direct coder — aonios_agent_direct.py reads a file, sends it to a premium model, extracts the markdown code block, writes it back.
3. An interactive agent — aonios_agent_interactive.py hands off a multi-step task to a premium model with full tool access (file edit, grep, bash, dependency install).
4. A configuration override surface — ~/.config/aonios-agent/models.json lets users customize model flags without editing code.

What this skill is NOT

• Not a replacement for vertex-coder (different model family).
• Not a general-purpose LLM gateway (use dt serve for that).
• Not a code reviewer (use the review-changes skill).
• Not a memory tool (lives outside the Apeiron memory layer).

---

Architecture

User task
   ↓
Aonios Agent (this skill)
   ↓
model_router.resolve_model(key)
   ↓
   ├─ codex CLI  →  codex exec --model "gpt 5.5 high"      (single-turn)
   └─ opencode CLI → opencode run -m opencode-go/minimax-m3 (multi-turn + tools)
   ↓
Markdown code block (direct mode) OR streamed tool-use (interactive mode)
   ↓
File write OR diff + verification

---

Available models

Key	CLI	Model Flag	Family	Tier	Direct	Interactive
codex-gpt-5.5-high	codex	gpt 5.5 high	OpenAI	high	✓	—
codex-gpt-5.4-high	codex	gpt 5.4 high	OpenAI	high	✓	—
opencode-glm-5.2-max	opencode	opencode-go/glm-5.2	Zhipu	max	✓	✓
opencode-qwen-max	opencode	opencode-go/qwen3.7-max	Alibaba	max	✓	✓
opencode-kimi-k2.7-code-max	opencode	opencode-go/kimi-k2.7-code	Moonshot	code-max	✓	✓
minimax-m3-high-thinking	opencode	opencode-go/minimax-m3	MiniMax	high-thinking	✓	✓

Why codex is direct-only: codex exec runs a single non-interactive turn. Multi-turn tool use via codex requires the experimental codex exec-server which is not yet stable. Use opencode-routed models for interactive work.

---

Quick start

# Health check
python3 inspect_settings.py

# List models
python3 list_models.py

# Direct mode (single file)
python3 aonios_agent_direct.py src/api/auth.ts "Add JWT validation middleware" minimax-m3-high-thinking

# Interactive mode (multi-file)
python3 aonios_agent_interactive.py "Implement OAuth2 PKCE flow in src/auth/" opencode-glm-5.2-max \
  --skills test-driven-development security-and-hardening

# Read prompt from stdin
echo "Refactor for clarity" | python3 aonios_agent_direct.py src/utils.ts - codex-gpt-5.5-high

---

Invocation patterns

Pattern A: Surgical single-file fix

When: One file, clear instruction, no multi-file dependencies.

python3 aonios_agent_direct.py <file> "<instruction>" <model>

Recommended model: minimax-m3-high-thinking (default) — strongest reasoning for code.

Pattern B: Multi-file refactor / feature

When: Touches multiple files, needs exploration, may run tests.

python3 aonios_agent_interactive.py "<task>" minimax-m3-high-thinking \
  --skills test-driven-development

Recommended model: minimax-m3-high-thinking for complex reasoning, opencode-kimi-k2.7-code-max for code-specialized work.

Pattern C: Fast iteration on small tasks

When: Quick fix, low stakes, want speed.

python3 aonios_agent_direct.py <file> "<fix>" opencode-glm-5.2-max

Recommended model: opencode-glm-5.2-max (fast tier).

Pattern D: Adversarial review

When: Need a second opinion on tricky code.

python3 aonios_agent_direct.py <file> "Review this code for race conditions and suggest fixes" codex-gpt-5.5-high

Recommended model: codex-gpt-5.5-high (different family = different blind spots).

---

Configuration override

Default model flags can be overridden via ~/.config/aonios-agent/models.json:

{
  "codex-gpt-5.5-high": {
    "model_flag": "gpt-5-codex-high"
  },
  "minimax-m3-high-thinking": {
    "fallback": "opencode-glm-5.2-max"
  }
}

---

Integration with Apeiron / expert-engineer

When the expert-engineer skill is active, Aonios Agent is invoked when:

1. The task is heavy (multi-file, complex reasoning, broad blast radius)
2. The task benefits from premium-model reasoning (architecture, hard bugs)
3. Local iteration would be expensive (large refactor, many test cycles)

The expert-engineer SKILL.md contains a "Delegate heavy coding → Aonios Agent" routing section that documents when to invoke this skill vs handle locally.

---

Files in this directory

File	Purpose
model_router.py	Maps model keys → CLI invocations. Single source of truth.
aonios_agent_direct.py	Single-file coder (analog of vertex_direct_coder.py).
aonios_agent_interactive.py	Multi-file autonomous agent (analog of vertex_interactive_agent.py).
tools_registry.py	Local tools inventory + persistent memory + skill loader.
list_models.py	Pretty-print model registry with backend status.
inspect_settings.py	Health check (CLIs installed, logged in, memory writable).
requirements.txt	Empty — no new Python deps (uses existing codex + opencode CLIs).
README.md	Same content as this SKILL.md (for dt integration).
SKILL.md	This file.

---

Cross-references

• Sibling backend: ${DELEGATE_TEAM_ROOT}/vertex-coder/ (Gemini-based)
• Orchestration gateway: dt CLI at /opt/homebrew/bin/dt
• Apeiron expert-engineer: ~/.apeiron/agents/apeiron/skills/expert-engineer/SKILL.md
• Global skills: ~/.agents/skills/ (1808 skills available for --skills preload)

---

Last updated: 2026-06-30 — first version, mirrors vertex-coder structure. Maintained by: Mamdouh + Apeiron (collaboratively).