Randomize model-to-instance assignment

[ScuttleBot]

Shuffles the model list before distributing across instances.

Why:
Previously, models were assigned in order (alphabetically if from YAML). This meant anthropic/* models always landed on early instances, z-ai/* on later ones. If there are any systematic differences between instances (boot timing, network, etc.), this could introduce bias.

Changes:

• distribute_models() now shuffles by default before round-robin assignment
• Added --no-shuffle flag to restore deterministic behavior if needed

Example:

# Default: randomized
uv run orchestrate_vultr.py --count 5

# Deterministic (alphabetical order preserved)
uv run orchestrate_vultr.py --count 5 --no-shuffle

[kilo-code-bot]

SUGGESTION: Docstring examples show deterministic round-robin output, but with the new default shuffle=True, the actual distribution will be randomized. Consider noting that these examples assume shuffle=False, e.g.:

Examples (with shuffle=False, showing round-robin logic):

This avoids confusing readers who might expect the documented output from the default behavior.

[kilo-code-bot]

Code Review Summary

Status: 1 Issue Found | Recommendation: Merge with minor improvement

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	1

Issue Details (click to expand)

SUGGESTION

File	Line	Issue
orchestrate_vultr.py	241	Docstring examples show deterministic output but default behavior is now shuffled

Other Observations (not in diff)

• Reproducibility: random.shuffle() uses the global RNG without an explicit seed. For benchmark reproducibility, consider adding an optional --seed argument that calls random.seed(value) before shuffling. This would allow users to reproduce exact model distributions across runs while still defaulting to random behavior. Not a bug, but a useful enhancement for a benchmarking tool.

Files Reviewed (1 file)

• orchestrate_vultr.py - 1 suggestion

---

Overall: Clean, well-structured change. The implementation correctly copies the input list before shuffling (avoiding mutation of the caller's list), the --no-shuffle flag provides an opt-out, and the shuffle keyword-only parameter keeps the API clean. Good use of list(models) to avoid side effects.

Fix these issues in Kilo Cloud