Cursor rule

1---
2description: "Model compatibility guide: prompt hierarchy, M3-first model selection, tool discipline, read-before-edit safety, and context control across models."
3alwaysApply: false
4---
5 
6# Transferable Agent Behavior (M3-first, model-resilient)
7 
8This repo is tuned for MiniMax M3 first. M3 is a generational shift: 1M-token MSA context, native multimodal input (text, image, video), and higher agentic and coding benchmarks.
9 
10The purpose of this file is to define the transferable execution behavior the agent should copy across models: tool-first work, read-before-edit discipline, short prompts, and context-aware problem solving that also maps well to strong GPT/Codex/Composer-style coding agents.
11 
12## Prompt Hierarchy
13 
14Treat the current environment as the source of truth:
15- system instructions and exposed tools win
16- always-on rules should stay short, durable, and canonical
17- requestable rules should carry runtime-specific guidance
18- user instructions still define the task
19 
20Do not assume an older rule or prompt shape is still valid just because it appears in this repo.
21Prefer a single always-on execution spine over duplicated policy spread across multiple files.
22 
23## Model Selection
24 
25| Need | Default | Why |
26|------|---------|-----|
27| Long context (multi-file refactor, transcript analysis, large retrieval pack, full-repo synthesis) | `MiniMax-M3` | 1M-token MSA context |
28| Visual-fidelity work, design parity, error UI triage, screenshot-driven dev | `MiniMax-M3` | Native multimodal input (text, image, video) |
29| Deep agentic / coding (SWE-Bench-class tasks, long-horizon planning) | `MiniMax-M3` | Higher agentic and coding benchmarks |
30| Anything where the user asks for "frontier" by name | `MiniMax-M3` | Positioned as the new frontier default |
31 
32When the active model is **not** M3 (for example `composer-2.5`, GPT, or Claude), the always-on core continues to apply — tool discipline, read-before-edit, scope control, solver loop, status taxonomy. The M3-specific sections (long-context discipline, multimodal input discipline) become inert, and the agent must not promise multimodal or 1M-context behavior. Confirm the model's actual capabilities from the current runtime before relying on them.
33 
34## Main Failure Modes
35 
36The most common failures are:
37- writing plausible prose instead of using tools
38- guessing file contents instead of reading them
39- teaching stale tool names as if they were universal
40- overloading the prompt with duplicate process text
41- on M3, over-loading context with raw search/fetch output instead of compressing (see `minimax-m3-long-context`)
42- on M3, making visual claims without re-reading the actual attached image/frame (see `minimax-m3-multimodal-input`)
43 
44## Tool Discipline
45 
46- If the prompt exposes a tool for the action, prefer that tool over shell.
47- Follow the exact tool schema shown by the environment.
48- Batch independent reads and searches when helpful.
49- Keep commentary short; spend prompt budget on execution, not narration.
50 
51## Read-Before-Edit
52 
53Canonical workflow lives in the always-on core **Code Discipline** section. In short:
54 
55```text
561. Read the target file in the current session
572. Base the change on exact contents
583. Use the current edit primitive
594. Verify with follow-up reads, build, tests, browser checks, shell output, or multimodal-grounded re-reads
60```
61 
62If the environment offers a patch or search-replace edit tool, prefer it for focused edits (Composer-style agents often expose `StrReplace`; other surfaces may use `ApplyPatch` or similar).
63 
64## Version Handling
65 
66- Never hardcode fast-moving versions in rules.
67- Use web search with the actual current month and year when versions matter.
68- Do not leave placeholders such as `[current year]` in the query.
69- Before adding a new package, framework, or toolchain, verify the latest stable version, compatibility, and official setup path against current authoritative sources.
70- Do not describe guidance as current, official, or best practice unless it is backed by those sources.
71 
72## Context Management
73 
74- Read only what is needed for the task.
75- Prefer targeted searches over broad scans.
76- Keep durable lessons, not raw output, in active context.
77- For large work, push detailed procedures into requestable rules or skills instead of the always-on core.
78- On M3, with 1M tokens available, still compress after each iteration — see `minimax-m3-long-context`.
79