PLATFORM_SPEC.md — Xiaomi MiMo-Code¶

New to gald3r on this platform? Start at Use gald3r with Xiaomi MiMo-Code for a short overview before the full breakdown below.

Terminal-native AI coding agent with cross-session memory. Fork of OpenCode. Launched June 2026.

MiMoCode is a terminal-native AI coding assistant built by Xiaomi as a fork of OpenCode. It inherits all core OpenCode capabilities (TUI, LSP, MCP, multiple providers, plugins) and adds persistent cross-session memory, intelligent context management, subagent orchestration, goal-driven autonomous loops, Compose mode for specs-driven development, and self-improvement via Dream/Distill.

Why this matters for gald3r early-mover advantage: MiMo-Code launched June 10, 2026, gained 7.6k GitHub stars and 610 forks within days, supports AGENTS.md + CLAUDE.md natively, has an MCP layer, and ships with a Compose mode very similar to gald3r's g-go pipeline. Early gald3r support = high-value SEO/GEO traffic and developer adoption signal.

1. Folder Hierarchy¶

gald3r writes to the repo-root .mimocode/ folder. Verified layout from source/docs:

<project-root>/
├── AGENTS.md                          ← primary instruction file (read natively)
├── CLAUDE.md                          ← Claude Code compat file (also read)
├── MEMORY.md                          ← persistent project knowledge (auto-maintained)
├── checkpoint.md                      ← session state snapshot (auto)
├── notes.md                           ← scratch notes (auto)
└── .mimocode/
    ├── mimocode.json                  ← primary config (project-level)
    ├── agents/
    │   └── *.md                       ← custom agent definitions (YAML frontmatter + prompt)
    ├── prompts/
    │   └── *.txt                      ← prompt files referenced from mimocode.json
    └── tasks/
        └── <id>/progress.md           ← per-task progress logs (auto)

~/.config/mimocode/
├── mimocode.json                      ← global config
└── agents/                            ← global custom agent definitions

2. CRASH Primitive Support¶

Primitive	Support	Notes
Hooks	⚠️ Partial	Inherits from OpenCode. Session hooks exist; lifecycle events need verification
Rules	✅ Native	`AGENTS.md` at project root read natively. Also reads `CLAUDE.md`
Skills	⚠️ Via Compose	Compose mode workflows = Skills equiv. `/distill` auto-generates reusable skills
Commands	✅ Native	`/goal`, `/dream`, `/distill`, `/voice`, custom slash commands via Compose
MCP	✅ Native	`mcp` section in `mimocode.json`; inherits full OpenCode MCP layer
Agents	✅ Native	`.mimocode/agents/*.md` (project) or `~/.config/mimocode/agents/` (global)

3. Instruction Files (Rules)¶

MiMo-Code natively reads both AGENTS.md and CLAUDE.md at the project root:

# AGENTS.md — gald3r rules injected here
<!-- Injected by gald3r install -->
<!-- Platform: mimo -->

## Project Context
...all gald3r AGENTS.md content...

Additionally, MEMORY.md serves as a persistent knowledge file that the agent maintains automatically across sessions — this is MiMo-Code's equivalent of the gald3r vault.

Load order (verified from docs + source): 1. ~/.config/mimocode/mimocode.json global config 2. .mimocode/mimocode.json project config (overrides global) 3. AGENTS.md → CLAUDE.md at project root 4. MEMORY.md injected into context window automatically

4. Skills (Compose Mode + Distill)¶

MiMo-Code's Compose mode provides structured specs-driven development with built-in sub-workflows: planning, execution, code review, TDD, debugging, verification, and merging.

The /distill command auto-generates reusable skills from repeated workflows:

/distill
→ Scans recent session traces
→ Packages repeated workflows into .mimocode/agents/{name}.md skill files
→ Immediately available as sub-agents

gald3r SKILL.md compatibility: Place SKILL.md files in .mimocode/agents/ with the MiMo-Code frontmatter format. gald3r skills can be adapted with a thin wrapper:

---
description: "Gald3r skill: g-skl-tasks"
mode: subagent
model: mimo/mimo-v2.5-pro
tools:
  write: true
  edit: true
  bash: false
---
<!-- SKILL.md content here -->

5. Commands (Slash Commands)¶

Built-in commands:

Command	Description
`/goal <condition>`	Set a stopping condition; judge model validates completion
`/dream`	Extract recent knowledge into `MEMORY.md`, remove stale entries
`/distill`	Auto-package repeated workflows into reusable skills/agents
`/voice`	Enable voice input (MiMo logged-in users)

Custom commands via Compose mode config in mimocode.json:

{
  "compose": {
    "workflows": {
      "review": {
        "description": "Full code review workflow",
        "steps": ["plan", "review", "verify"]
      }
    }
  }
}

6. Hooks¶

MiMo-Code inherits OpenCode's hook architecture. OpenCode hooks fire on session events. Pending verification: confirm .mimocode/hooks.json or equivalent hook registration.

Likely hook events (from OpenCode heritage): - session_start — inject context from MEMORY.md + AGENTS.md - before_tool_call — gate tool permissions - session_end — trigger checkpoint save

gald3r hook files: Target .mimocode/hooks.json once doc crawl confirms the schema.

7. MCP Configuration¶

MCP is configured in mimocode.json under the mcp key (OpenCode-compatible syntax):

{
  "$schema": "https://mimo.xiaomi.com//config.json",
  "mcp": {
    "gald3r": {
      "type": "remote",
      "url": "https://your-gald3r-instance/mcp",
      "enabled": true
    },
    "filesystem": {
      "type": "local",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
    }
  }
}

gald3r MCP wire-up: Add the gald3r MCP endpoint to .mimocode/mimocode.json during gald3r install --platform mimo. This gives MiMo-Code agents access to all 42 gald3r tools.

8. Agents¶

Custom agents are defined as markdown files with YAML frontmatter:

.mimocode/agents/gald3r-reviewer.md

---
description: "gald3r code review agent"
mode: subagent
model: mimo/mimo-v2.5-pro
temperature: 0.1
tools:
  write: false
  edit: false
  bash: false
---

You are a gald3r-aware code reviewer. Apply the g-go-review criteria:
security, quality, acceptance criteria, and hallucination guard.
Follow .gald3r/CONSTRAINTS.md rules. Report findings as BUG[BUG-NNN] entries.

Built-in primary agents (press Tab to switch):

Agent	Mode	Description
`build`	primary	Full tool permissions for development
`plan`	primary	Read-only analysis and solution design
`compose`	primary	Specs-driven development orchestration

9. Memory System (Unique to MiMo)¶

MiMo-Code's persistent memory is a key differentiator from other OpenCode forks:

Memory File	Purpose	Auto-managed?
`MEMORY.md`	Persistent project knowledge, rules, architecture decisions	via `/dream`
`checkpoint.md`	Session state snapshot (task progress, recent context)	Auto on context limit
`notes.md`	Temporary scratch area for agents	Auto
`tasks/<id>/progress.md`	Per-task progress logs	Auto

gald3r + MiMo memory synergy: MEMORY.md is MiMo's equivalent of gald3r's vault notes. Consider wiring the gald3r @g-memory-capture workflow to write summaries to both .gald3r/vault/ AND MEMORY.md when running in a MiMo-Code project.

10. gald3r Installation¶

Install Path¶

<project-root>/
├── AGENTS.md                ← gald3r injects rules here (natively read)
├── CLAUDE.md                ← gald3r injects Claude Code compat layer
└── .mimocode/
    ├── mimocode.json        ← gald3r adds MCP entry here
    └── agents/
        └── g-agnt-*.md      ← gald3r agents adapted to MiMo format

Install Command (when implemented)¶

gald3r install --platform mimo

This should: 1. Write AGENTS.md (instruction file) at project root 2. Add gald3r MCP endpoint to .mimocode/mimocode.json 3. Place adapted gald3r agents in .mimocode/agents/ 4. Write MEMORY.md with initial project memory seed 5. Confirm readback: mimo read AGENTS.md should load rules

Verification Checklist¶

[x] AGENTS.md read natively at session start ✅ (confirmed from docs)
[x] CLAUDE.md read natively ✅ (visible in MiMo-Code repo root)
[x] MCP servers configurable in mimocode.json ✅ (confirmed from docs)
[x] Custom agents via .mimocode/agents/*.md ✅ (confirmed from docs)
[ ] Hooks schema confirmed (verify .mimocode/hooks.json or equivalent)
[ ] Skills/Compose auto-discovery from .mimocode/agents/ confirmed
[ ] Full gald3r install test completed

11. Relationship to OpenCode¶

MiMo-Code is a fork of OpenCode and is architecturally compatible. gald3r's existing g-skl-platform-opencode knowledge is the starting point. Key additions over OpenCode:

Feature	OpenCode	MiMo-Code
Cross-session memory	❌	✅ (`MEMORY.md`, `checkpoint.md`)
Subagent orchestration	Basic	✅ (full lifecycle, parallel, cancellable)
Dream/Distill	❌	✅ (self-improvement loop)
Goal / stop condition	❌	✅ (`/goal` with judge model)
Voice input	❌	✅ (TenVAD + MiMo ASR)
Max Mode	❌	✅ (parallel best-of-N with judge)
Default model	Provider agnostic	`mimo/mimo-v2.5-pro`

gald3r parity: OpenCode parity target applies. MiMo-Code gains the OpenCode full parity tier PLUS the memory/agent additions above. Priority: ship MiMo support before competitors.

12. Official Sources¶

Docs: https://mimo.xiaomi.com/mimocode
Agents: https://mimo.xiaomi.com/mimocode/agents
GitHub: https://github.com/XiaomiMiMo/MiMo-Code
MiMo model: https://mimo.mi.com/
News: https://www.neura.market/news/xiaomi-mimo-v2-5-pro-open-model-compiler-coding-benchmarks

Verified 2026-06-13 against GitHub README + mimo.xiaomi.com/mimocode docs.

Reference page — canonical source: g-skl-platform-mimo/PLATFORM_SPEC.md in the engine repo, generated by scripts/generate_platform_support_docs.py. Pages marked curated are hand-assessed and not yet live-verified.