Starting kit for interns and students: Reasonix setup, Reasonix vs OpenCode, and DeepSeek Harness (tool-use and reasoning tracking, heartbeats, embedded prompts, clarifying questions).
Find a file
justinaquino fe1ca82532 Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide
Setup and API keys:
- 01: Node >= 22 install (nvm on Linux, LTS/winget on Windows); DeepSeek key
  section split into get-and-fund (402 Insufficient Balance), wizard, env var
  for bash and PowerShell (history-safe input), key test, gitignore .env
- 01: fix troubleshooting curl (no auth header always returned 401); add rows
  for EACCES, npm.ps1, 402, per-window env vars, and the `reasonix setup --help`
  trap (it writes a file named --help)
- 01: PowerShell versions of the first-session commands and Get-FileHash
- 03: global pinned install of dsh, bare-dsh needs --profile, how dsh reads
  DEEPSEEK_API_KEY, Models page / apiKeyEnv, telemetry switch, pnpm prereq
- README: prerequisites and a PowerShell quickstart, including dsh with its key
- .gitignore: .env

Local models (08, rewritten):
- four engines: Ollama (easiest), LM Studio, llama.cpp, Unsloth (now runs and
  trains), with Linux and Windows installs
- tiny text models: Granite 4.0 350M/1B, Granite 4.2 3B, Qwen3.5 0.8B/2B/4B,
  LFM2.5 1.2B, Nemotron 3 Nano 4B, with Ollama tags, HF and GGUF links, sizes
- smallest vision models: SmolVLM 256M, LFM2.5-VL 450M/1.6B, SmolVLM2 500M,
  Qwen3.5 0.8B, Moondream 2, Qwen3-VL 2B, Granite Vision 4.1 4B, Gemma 4 E2B
- wiring into Reasonix, dsh and OpenCode; a three-model lab
- removes stale claims ("Qwen 3.5 does not exist", Unsloth as training-only)

SOURCES: everything checked on 2026-10-16; PowerShell commands marked [R]
(not executed on Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-16 12:50:44 +08:00
examples Intern-Tools: Reasonix setup, Reasonix vs OpenCode, and a DeepSeek Harness guide 2026-10-16 10:38:31 +08:00
.gitignore Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide 2026-10-16 12:50:44 +08:00
01-reasonix-setup.md Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide 2026-10-16 12:50:44 +08:00
02-reasonix-vs-opencode.md Intern-Tools: Reasonix setup, Reasonix vs OpenCode, and a DeepSeek Harness guide 2026-10-16 10:38:31 +08:00
03-deepseek-harness.md Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide 2026-10-16 12:50:44 +08:00
04-heartbeats.md Intern-Tools: Reasonix setup, Reasonix vs OpenCode, and a DeepSeek Harness guide 2026-10-16 10:38:31 +08:00
05-embedded-prompts-and-steering.md Intern-Tools: Reasonix setup, Reasonix vs OpenCode, and a DeepSeek Harness guide 2026-10-16 10:38:31 +08:00
06-clarifying-questions.md Intern-Tools: Reasonix setup, Reasonix vs OpenCode, and a DeepSeek Harness guide 2026-10-16 10:38:31 +08:00
07-labs.md Intern-Tools: Reasonix setup, Reasonix vs OpenCode, and a DeepSeek Harness guide 2026-10-16 10:38:31 +08:00
08-cpu-models-and-engines.md Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide 2026-10-16 12:50:44 +08:00
README.md Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide 2026-10-16 12:50:44 +08:00
SOURCES.md Linux/Windows setup and API keys for Reasonix and dsh; rewrite local-models guide 2026-10-16 12:50:44 +08:00

Intern-Tools

A starting kit for interns and students: how to install and run an AI coding agent, and how modern agent harnesses actually work.

This pack covers three things, in order:

# Topic File
1 Reasonix — install, configure, and the exact commands to run 01-reasonix-setup.md
2 Reasonix vs OpenCode — the real differences, and when to use which 02-reasonix-vs-opencode.md
3 DeepSeek Harness (dsh) & the web harness — tool-use and reasoning tracking, heartbeats, embedded prompts, and clarifying questions 03-deepseek-harness.md

Deeper dives on the parts of the harness that matter most:

Runnable example files live in examples/.


The one-paragraph version

An AI coding agent is three things stacked, and people constantly confuse them:

  1. The model — DeepSeek, or Claude, or GPT. It predicts text. It has no hands and no memory between requests.
  2. The harness — the program around the model (Reasonix, OpenCode, DeepSeek Harness, Claude Code). It gives the model tools (read a file, run a command, edit code), decides what context the model sees on each turn, and drives the loop.
  3. The workflow around both — your instructions, your review, your decision about when to let it run unattended.

Most beginners focus entirely on (1) and are then disappointed. The harness is where the difference between "a chatbot that talks about code" and "an agent that gets work done" actually lives — and it is the part you can configure. That is what this pack teaches.


Prerequisites

Need Why Check
Node.js ≥ 22 both Reasonix and DeepSeek Harness ship through npm node --version — install: 01 §2 (nvm on Linux, LTS installer on Windows)
A DeepSeek API key, with a balance the model behind the agent — get one at https://platform.deepseek.com/api_keys; the account is prepaid, so top up first or every call returns 402 01-reasonix-setup.md §3 (Linux and Windows)
A terminal bash/zsh on Linux/macOS, PowerShell or Git Bash on Windows —
Git (recommended) agents change files; git is your undo button git --version

Budget note: an API key is a paid account. Use it deliberately — start with small tasks, watch the token receipt Reasonix prints after each turn, and read §8 of the setup guide before leaving anything running unattended.


20-minute quickstart

Do this in an empty folder you do not mind the agent editing. Install Node.js ≥ 22 first (01-reasonix-setup.md §2) and have a funded DeepSeek key ready (§3.1).

Linux / macOS (bash/zsh)

# 1. Install the agent (pulls a prebuilt native binary)
npm i -g reasonix
reasonix --version          # expect a version line, e.g. reasonix v1.31.3

# 2. Point it at the model, once
reasonix setup              # interactive wizard; writes your key to ~/.reasonix/.env

# 3. Prove it is wired up before trusting it
reasonix doctor             # health check: Node, API key, MCP wiring

# 4. Do a real, tiny task in a scratch repo
mkdir -p ~/scratch/hello && cd ~/scratch/hello && git init
printf 'def add(a, b):\n    return a + b\n' > calc.py
reasonix -p "add a docstring to add() in calc.py and show me the diff"
git diff                    # <- the actual evidence the work happened

Windows (PowerShell) — if npm is blocked by execution policy, use npm.cmd and reasonix.cmd instead (see 01 §2, Windows notes).

# 1. Install the agent
npm i -g reasonix
reasonix --version

# 2. Point it at the model, once (the wizard stores the key in Reasonix's own .env)
reasonix setup

# 3. Prove it is wired up
reasonix doctor

# 4. A real, tiny task in a scratch repo
New-Item -ItemType Directory -Force "$HOME\scratch\hello" | Out-Null
Set-Location "$HOME\scratch\hello"
git init
Set-Content calc.py "def add(a, b):`n    return a + b"
reasonix -p "add a docstring to add() in calc.py and show me the diff"
git diff

Then move on to the DeepSeek Harness, which is a different (and complementary) tool. It has no setup wizard — it reads DEEPSEEK_API_KEY from the terminal that starts it, so set that first (03-deepseek-harness.md §2):

npm i -g @deepseek-ai/dsh@0.1.1-rc.2
read -rs -p "DeepSeek key: " DEEPSEEK_API_KEY && export DEEPSEEK_API_KEY && echo
dsh web                      # Web UI, opens at http://127.0.0.1:3080
npm i -g @deepseek-ai/dsh@0.1.1-rc.2
$env:DEEPSEEK_API_KEY = Read-Host "DeepSeek key" -MaskInput   # PowerShell 7; or set the User variable (01 §3.3)
dsh web

House rules for this pack

These are the habits that separate an intern who is trusted with an agent from one who is not.

  1. Never report success you did not observe. "The agent said it finished" is not evidence. Run the command, read the file, check git diff. An agent's own summary is a claim; your check is the evidence. An empty or timed-out check is a failure, not a pass.
  2. Pin the version you validated. These tools ship fast. Record reasonix --version and dsh --version in your notes, and re-check this pack when you upgrade — commands move.
  3. Read before you run. Anything you install here is third-party code that will execute on your machine with your permissions and can reach your API key. Prefer the released binary and verify its checksum, or build from source, before pointing an agent at anything sensitive.
  4. Never paste secrets into a prompt or a config file. API keys belong in an environment variable or the tool's own .env file, never in a chat, a screenshot, a git commit, or an example config.
  5. Keep agents out of production and off other people's data until you have been shown how the access controls work. An agent with a shell can do everything you can do.
  6. Say when you guessed. If you proceeded on an assumption instead of asking, write the assumption down. That single habit is the subject of 06-clarifying-questions.md.

What is verified here, and what is not

Honesty about evidence is a technical skill, not a personality trait, so this pack practices it:

  • Every version number, command, filename, and quoted mechanism in this pack was read from the installed software or the upstream repository on the date in SOURCES.md, and the source is named there.
  • Star counts and dates come from the public GitHub API on the same date. They are a snapshot.
  • Where a figure comes from a vendor's own claim, it is labelled as such.
  • Judgements are labelled as judgements. Where this pack says one tool is a better fit, it says why, and it says what the other tool is better at. No tool here is "the best" — a claim like that is only meaningful relative to a job.

Third-party components documented here are MIT-licensed projects (Reasonix, DeepSeek Harness, OpenCode), each with its own repository. They are not part of this pack and are not supported by it. This guide itself is Comfac IT material.