DeepSeek's own harness has a bash + editor mode
Benchmark models with DeepSeek Harness Minimal mode (bash + editor only). Model evaluators.
what to steal
Benchmark your model with the smallest toolset first.
Our takeaway from the source, not a quote.
deepseek-harness-minimal.toml
# Benchmark models with DeepSeek Harness Minimal mode (bash + editor only) # source: https://deepseek.com/harness/en/ [rig] who = "Model evaluators" author = "DeepSeek" date = "not stated" harness = ["DSH (DeepSeek Harness)"] models = [] [keep] - Minimal mode: bash + str_replace_editor [cut] - Other tools
| thing | status | why (per source) |
|---|---|---|
| Other tools | removed in Minimal mode | Described as for benchmarking models |
Source: https://deepseek.com/harness/en/ · https://github.com/deepseek-ai/deepseek-harness · checked 2026-10-14
// similar rigs
~/rigs/metaist-claude-code-to-pi.toml2026-01-20
Left Claude Code for Pi: smaller prompt, any model
steal → If you switch models often, a model-agnostic minimal harness pays off.
Pi Claude Code
by Metaist
~/rigs/merlin-claude-manager-codex-….toml𝕏 · 2026-10-14
Claude manages, Codex grinds
steal → Use one model to plan architecture and another to churn through build steps.
Claude CodeCodex Claude Fable as manageGPT-6 Astra on Codex
by @Merlin412188240
~/rigs/niv-one-codex-thread-per-tas….toml𝕏 · 2026-10-14
One task, one thread, so bad corrections die with it
steal → Start a fresh thread per task; carry over only the memory you mean to.
Codex Long shared threads
by Niv Asayag (@niv_ai)