minimax-coder

name

minimax-coder

description

Local MiniMax-powered coding agent. Uses the official `mmx` CLI (https://github.com/MiniMax-AI/cli) to call MiniMax models (M3 flagship, M2.7, M2.7-highspeed) for direct single-file edits and multi-file autonomous work with full tool use. Sibling backend to vertex-coder inside delegate-team; mirrors its structure (direct + interactive + tools registry) but routes through mmx instead of an SDK. The Python scripts in this directory are local-agent helpers — Apeiron (M3) can also invoke `mmx text chat` directly via bash for fast ad-hoc calls. Three MiniMax models supported. Use when delegating coding work to MiniMax models. Triggers on: "use minimax coder", "MiniMax M3", "MiniMax M2.7", "minimax direct", "minimax interactive", "mmx text chat". Do NOT use for trivial edits (do them locally).

MiniMax Coder

Local MiniMax-powered coding agent. Sibling to vertex-coder inside delegate-team/. Mirrors vertex-coder's structure (direct mode + interactive mode + tools registry) but uses the official mmx CLI for all API transport. This means:

• No Python SDK dependencies — just mmx on PATH (one npm install away)
• Auth handled by mmx — no key management in our Python code
• You (Apeiron, M3) are the orchestrator — this skill provides the scripts, but you can also call mmx text chat directly via bash for fast ad-hoc sub-agent work

The "boss and director" pattern: Apeiron on M3 calls MiniMax models as sub-agents — M2.7-highspeed for fast isolated work, M3 itself for hard problems with fresh context.

What this skill IS

1. A direct coder — minimax_direct_coder.py reads a file, sends it to a MiniMax model via mmx text chat, extracts the markdown code block, writes it back.
2. An interactive agent — minimax_interactive_agent.py runs a multi-turn tool-use loop with mmx, where the model can read files, grep, run commands, install deps, persist memory.
3. A local-agent surface — the mmx CLI itself is available to Apeiron via bash for ad-hoc calls. The Python scripts here wrap common patterns but aren't required.
4. A multi-model dispatcher — three models supported: M3 (flagship), M2.7 (previous gen), M2.7-highspeed (speed-optimized). All routed through the same mmx text chat call.

What this skill is NOT

• Not a replacement for vertex-coder (different model family — Gemini vs MiniMax).
• Not a memory tool (lives outside the Apeiron memory layer).
• Not a router (use the parent delegate-team dt CLI for cross-backend dispatch).
• Not a replacement for mmx text chat directly — when you need a one-off query, just run mmx text chat --model MiniMax-M3 --message "..." without involving this skill's scripts.

---

Architecture

User task
   ↓
Apeiron (MiniMax-M3) — orchestrator + boss
   ↓
   ├─ Direct call:    mmx text chat --model M3 --message "..."   (fast, ad-hoc)
   ├─ Direct coder:   minimax_direct_coder.py <file> <prompt>    (file → code → file)
   └─ Interactive:    minimax_interactive_agent.py <prompt>       (multi-turn + tools)
                       ↓
                    mmx CLI (handles auth, region, API)
                       ↓
                    https://api.minimax.io (Token Plan API)
                       ↓
                    MiniMax-M3 / MiniMax-M2.7 / MiniMax-M2.7-highspeed

---

Available models

Model	Description	Best for
MiniMax-M3	Flagship (Apeiron's own runtime), 450K context, multimodal, switchable thinking	Default heavy reasoning, complex tasks
MiniMax-M2.7	Previous generation, 200K context, forced thinking	Mature reasoning, good quality/cost balance
MiniMax-M2.7-highspeed	M2.7 optimized for speed, 200K context	Fast isolated sub-tasks, batch processing

All three are whitelisted in ~/.minimax/config.yaml (confirmed by inspect_settings.py).

---

Quick start

Setup (one-time)

# Install mmx CLI globally
npm install -g mmx-cli

# Authenticate with your MiniMax Token Plan API key
mmx auth login --api-key sk-xxxxx

# Verify
mmx quota

Usage

# Direct mode (single file)
python3 minimax_direct_coder.py src/api/auth.ts "Add JWT validation middleware" MiniMax-M3

# Interactive mode (multi-file autonomous with tool use)
python3 minimax_interactive_agent.py "Implement OAuth2 PKCE flow" MiniMax-M3 --skills test-driven-development

# Direct mmx call (fast, ad-hoc — Apeiron can use this without the Python wrapper)
mmx text chat --model MiniMax-M2.7-highspeed --message "Explain async/await in 3 sentences"
mmx text chat --model MiniMax-M3 --system "You are a code reviewer" --message "Review this diff: ..."

# Read prompt from stdin
echo "Refactor for clarity" | python3 minimax_direct_coder.py src/utils.ts - MiniMax-M2.7-highspeed

---

Invocation patterns

Pattern A: Boss-and-director (M3 → M2.7)

Apeiron on M3 keeps the high-level conversation. For a fast isolated subtask, spawn M2.7-highspeed.

# M3 delegates a quick refactor to M2.7-highspeed while M3 keeps thinking
python3 minimax_direct_coder.py src/refactor-target.ts "Apply this exact change..." MiniMax-M2.7-highspeed

Pattern B: Self-orchestration (M3 → M3, fresh context)

Use when you want a hard problem tackled without cluttering the main Apeiron session.

python3 minimax_interactive_agent.py "Solve this hard concurrency bug" MiniMax-M3 --max-turns 20

Pattern C: Adversarial review (MiniMax vs Gemini)

Pair minimax-coder with vertex-coder for a second opinion from a different model family.

# MiniMax opinion
python3 minimax_direct_coder.py src/critical.ts "Review for race conditions" MiniMax-M3 > review-minimax.txt
# Gemini opinion
python3 ${DELEGATE_TEAM_ROOT}/vertex-coder/vertex_direct_coder.py src/critical.ts "Review for race conditions" gemini-3.1-pro > review-gemini.txt
# Diff
diff review-minimax.txt review-gemini.txt

Pattern D: Direct mmx calls (no Python wrapper)

When you just need a quick one-off — skip the Python layer entirely.

# Simple Q&A
mmx text chat --model MiniMax-M2.7-highspeed --message "What's the capital of France?"

# With system prompt and JSON output
mmx text chat --model MiniMax-M3 --system "You are a SQL expert. Reply with just SQL." --message "Write a query to find top 10 customers by revenue" --output json

# Multi-message conversation
mmx text chat --message "What is 2+2?" --message "assistant:4" --message "And 3+3?"

---

Local tools (interactive mode)

The interactive agent exposes 10 local tools via Anthropic tool-use API (mmx native):

Tool	Purpose
list_dir	List directory contents
read_file	Read file contents
write_file	Create/overwrite file (sparingly)
line_replace	PREFERRED — surgical line-range replacement with ... keep markers
grep_search	Regex search across workspace
run_command	Bash execution (60s default, up to 300s)
save_memory	Persist key/value to .agent_memory.json
read_memories	Load all persisted memories
list_global_skills	List installed skills (5 search paths)
load_skill	Load a skill's full instructions into context

Memory file: .agent_memory.json (same name as vertex-coder for parity).

---

Files in this directory

File	Purpose
minimax_direct_coder.py	Single-file coder (analog of vertex_direct_coder.py)
minimax_interactive_agent.py	Multi-file autonomous agent (analog of vertex_interactive_agent.py)
tools_registry.py	Anthropic-format tool schemas + implementations
list_models.py	Model listing
inspect_settings.py	Health check (mmx CLI + auth + quota + memory)
requirements.txt	No runtime deps (Python stdlib + mmx CLI only)
SKILL.md	This file
README.md	Same content as SKILL.md (for dt integration)

---

Integration with Apeiron / expert-engineer

When the expert-engineer skill is active, MiniMax Coder is invoked when:

1. M3 wants to delegate to a MiniMax sub-agent (boss-and-director pattern)
2. The task is MiniMax-specific (testing MiniMax APIs, integrating MiniMax tooling)
3. The user prefers MiniMax model family for any reason (consistency, cost, latency)
4. Adversarial review needs a different family (pair with vertex-coder)
5. Fast isolated sub-task — M2.7-highspeed while M3 keeps thinking

The expert-engineer SKILL.md has a "Delegate heavy coding → /delegate-team" routing section that documents when to invoke minimax-coder vs aonios-agent vs vertex-coder.

Apeiron can also use mmx text chat directly — no need to invoke the Python scripts for one-off queries.

---

Cross-references

• Sibling backends:
  • ${DELEGATE_TEAM_ROOT}/vertex-coder/ — Gemini-based (same structure)
  • ${DELEGATE_TEAM_ROOT}/aonios-agent/ — Multi-CLI (codex + opencode)
• Orchestration gateway: dt CLI at /opt/homebrew/bin/dt
• Apeiron expert-engineer: ~/.apeiron/agents/apeiron/skills/expert-engineer/SKILL.md
• Config source of truth: ~/.minimax/config.yaml
• Auth: mmx auth login --api-key sk-xxxxx (no env var or .env needed)
• Official CLI: https://github.com/MiniMax-AI/cli
• Docs: https://platform.minimax.io/docs/token-plan/minimax-cli

---

Last updated: 2026-06-30 — switched to mmx CLI transport (official, proven). Removed anthropic SDK dependency. Maintained by: Mamdouh + Apeiron (collaboratively).