Cursor rule

1---
2description: "MiniMax M3 self-evolution harness: iterative refinement loops, compress-before-iterate, autonomous debugging, and recursive improvement patterns."
3alwaysApply: false
4---
5 
6# MiniMax M3 Self-Evolution Harness
7 
8M3 has a 1M-token MSA context and high skill adherence, and supports autonomous self-improvement through iterative reinforcement learning loops. Use this rule when the task involves debugging, optimization, or recursive code improvement.
9 
10## Core Loop
11 
12```
13Iterate until evidence shows the problem is solved or the approach is exhausted:
14  1. Run the smallest diagnostic check
15  2. Read the failure output directly
16  3. Compress prior raw evidence that you no longer need (see below)
17  4. Make ONE targeted fix based on evidence
18  5. Re-run the exact check that failed
19  6. If fixed, verify the broader surface
20  7. If not fixed, form a new hypothesis from the NEW evidence only
21```
22 
23Do NOT repeat the same fix twice on the same hypothesis. Do NOT assume the cause without reading the evidence.
24 
25## Compress Before Next Iteration (M3)
26 
27On M3 the failure mode shifts from "ran out of room" to "ran too many parallel hypotheses without compressing." Before each new iteration:
28 
29- Replace raw search/fetch output from previous iterations with a 2–4 line summary.
30- Drop evidence that is no longer relevant to the current hypothesis.
31- Keep one canonical "current best hypothesis" line at the top of your scratchpad.
32- For very large work, use the `minimax-m3-long-context` skill to plan the loader.
33 
34## When to Use This Harness
35 
36| Task | Harness Approach |
37|------|------------------|
38| Debug runtime error | Read error → one fix → re-run |
39| Optimize performance | Profile → smallest change → measure |
40| Fix failing test | Read test → read code → one fix → run test |
41| Investigate unexpected behavior | Log/check → hypothesis → targeted probe |
42| Iterative code improvement | Current state → one improvement → verify |
43| Triage a visual bug from a screenshot | Read the screenshot, form hypothesis, one fix, re-read the post-change frame (`multimodal-grounded`) |
44 
45## Iteration Limits
46 
47- **Diagnostic phase**: 3 iterations to identify root cause
48- **Fix phase**: 2 iterations per hypothesis before switching strategy
49- **Overall**: If 5 total iterations pass without progress, summarize evidence and ask
50- **Context check**: if you have run 3+ raw search/fetch operations without compressing, compress before continuing
51 
52## Evidence Rules
53 
54Always prefer direct evidence over inference:
55- Runtime errors: Read the actual error message
56- Test failures: Read the test output
57- Performance: Read the actual measurement
58- Behavior: Read the actual output/log
59- Visual claims: Read the actual attached image/frame, not a memory or guessed description
60 
61Never say "likely caused by" without reading the evidence first.
62 
63## Self-Correction Pattern
64 
65When a fix doesn't work:
66 
67```
68Previous hypothesis: [what I thought was wrong]
69Evidence against it: [what actually happened]
70New hypothesis: [what the evidence suggests instead]
71Next action: [specific check to confirm new hypothesis]
72```
73 
74## Anti-Patterns
75 
76- Do not run the same diagnostic command twice expecting different results
77- Do not make multiple changes between diagnostic cycles
78- Do not assume a fix works without re-running the exact check
79- Do not expand scope mid-iteration (finish the loop first)
80- Do not accumulate raw search output across iterations without compressing
81- Do not make a visual-fidelity claim without re-reading the post-change frame
82