---
description: "MiniMax M3 self-evolution harness: iterative refinement loops, compress-before-iterate, autonomous debugging, and recursive improvement patterns."
alwaysApply: false
---

# MiniMax M3 Self-Evolution Harness

M3 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.

## Core Loop

```
Iterate until evidence shows the problem is solved or the approach is exhausted:
  1. Run the smallest diagnostic check
  2. Read the failure output directly
  3. Compress prior raw evidence that you no longer need (see below)
  4. Make ONE targeted fix based on evidence
  5. Re-run the exact check that failed
  6. If fixed, verify the broader surface
  7. If not fixed, form a new hypothesis from the NEW evidence only
```

Do NOT repeat the same fix twice on the same hypothesis. Do NOT assume the cause without reading the evidence.

## Compress Before Next Iteration (M3)

On M3 the failure mode shifts from "ran out of room" to "ran too many parallel hypotheses without compressing." Before each new iteration:

- Replace raw search/fetch output from previous iterations with a 2–4 line summary.
- Drop evidence that is no longer relevant to the current hypothesis.
- Keep one canonical "current best hypothesis" line at the top of your scratchpad.
- For very large work, use the `minimax-m3-long-context` skill to plan the loader.

## When to Use This Harness

| Task | Harness Approach |
|------|------------------|
| Debug runtime error | Read error → one fix → re-run |
| Optimize performance | Profile → smallest change → measure |
| Fix failing test | Read test → read code → one fix → run test |
| Investigate unexpected behavior | Log/check → hypothesis → targeted probe |
| Iterative code improvement | Current state → one improvement → verify |
| Triage a visual bug from a screenshot | Read the screenshot, form hypothesis, one fix, re-read the post-change frame (`multimodal-grounded`) |

## Iteration Limits

- **Diagnostic phase**: 3 iterations to identify root cause
- **Fix phase**: 2 iterations per hypothesis before switching strategy
- **Overall**: If 5 total iterations pass without progress, summarize evidence and ask
- **Context check**: if you have run 3+ raw search/fetch operations without compressing, compress before continuing

## Evidence Rules

Always prefer direct evidence over inference:
- Runtime errors: Read the actual error message
- Test failures: Read the test output
- Performance: Read the actual measurement
- Behavior: Read the actual output/log
- Visual claims: Read the actual attached image/frame, not a memory or guessed description

Never say "likely caused by" without reading the evidence first.

## Self-Correction Pattern

When a fix doesn't work:

```
Previous hypothesis: [what I thought was wrong]
Evidence against it: [what actually happened]
New hypothesis: [what the evidence suggests instead]
Next action: [specific check to confirm new hypothesis]
```

## Anti-Patterns

- Do not run the same diagnostic command twice expecting different results
- Do not make multiple changes between diagnostic cycles
- Do not assume a fix works without re-running the exact check
- Do not expand scope mid-iteration (finish the loop first)
- Do not accumulate raw search output across iterations without compressing
- Do not make a visual-fidelity claim without re-reading the post-change frame