OpenHarness
Recommended
A terminal agent that plans, edits, runs commands and iterates, asking permission before it touches anything. It is the tool we recommend for Devotel AI: the provider model is explicit, switching models is one command, and it has a dry-run mode that costs nothing.
Requirements: Python 3.10+ and Node 18+.
Install
curl -fsSL https://raw.githubusercontent.com/HKUDS/OpenHarness/main/scripts/install.sh | bashmacOS notes, from real installs.
If the installer stops with “Python 3.10+ not found”:
brew install [email protected]
export PATH=/opt/homebrew/opt/[email protected]/libexec/bin:/opt/homebrew/bin:$PATHThen rerun the installer.
The installer links oh into ~/.local/bin, and it may only update your fish config. On zsh
or bash, add that directory yourself:
export PATH=$HOME/.local/bin:$PATHPersist it in ~/.zshrc (or ~/.bashrc) so new shells can find oh.
Add Devotel AI as a provider
oh provider add devotel \
--label "Devotel AI" \
--provider openai \
--api-format openai \
--auth-source openai_api_key \
--model kimi-k3 \
--base-url https://api.devotel.com/v1Give it your key and select the provider
export OPENAI_API_KEY=dvt-YOUR_KEY
oh provider use devotel--auth-source openai_api_key is why the key goes in OPENAI_API_KEY. Put the export in your
shell profile if you do not want to repeat it, or use a per-project .envrc.
Confirm the active provider
oh provider listCheck the * before you start a task. The line marked * is the active provider, and it must
be devotel (kimi-k3 via api.devotel.com). If a different profile is active, OpenHarness
quietly sends your work to that provider instead — no error, no warning, and nothing appears on
your Devotel usage dashboard. Switch back with:
oh provider use devotelUsing it
cd ~/code/my-project
ohDescribe the task in plain language. The agent proposes a plan, then asks before each write and each
command; you approve, edit or reject. Start in a repository with a clean git status — reviewing an
agent’s work is much easier as a diff.
Useful flags:
# Preview the plan without calling the model at all — free, and a good way to check your prompt.
oh --dry-run -p "add retry-with-backoff to the S3 upload helper"
# Non-interactive: run to completion and print the result.
oh -p "fix the failing test in tests/parser_test.py"
# Machine-readable output, for scripts and CI.
oh -p "summarise the TODOs in src/" --output-format jsonWhat a session looks like
Abridged transcript of a real task — write a disk-space monitor, then make it work:
$ oh
› write a script that reports disk usage per mount and warns above 80%
plan
1. create scripts/disk_monitor.py
2. run it against this machine
3. add tests for the threshold logic
create scripts/disk_monitor.py? [y/N] y
✓ wrote scripts/disk_monitor.py (48 lines)
run `python3 scripts/disk_monitor.py`? [y/N] y
✗ exited 1
Traceback (most recent call last):
File "scripts/disk_monitor.py", line 22, in parse_line
used = int(parts[2])
ValueError: invalid literal for int() with base 10: '12Gi'
the df output on this machine is human-readable (`12Gi`), not blocks — I will parse
the suffixes instead of assuming integers.
edit scripts/disk_monitor.py? [y/N] y
✓ added parse_size() handling K/M/G/T suffixes
run `python3 scripts/disk_monitor.py`? [y/N] y
✓ / 41.2% ok
/boot 78.9% ok
/data 91.4% WARNING
add tests/test_disk_monitor.py? [y/N] y
✓ 6 cases: suffix parsing, exactly-80%, above, below, zero-size, missing mount
run `pytest tests/test_disk_monitor.py`? [y/N] y
✓ 6 passed
done — 2 files created, 1 editedThe pattern is the useful part: it wrote code, ran it, read the traceback, understood that df
on this machine emits human-readable sizes, fixed the parser, re-ran, then covered the boundary at
exactly 80%. That loop is what the token estimates on the overview describe — roughly ten
requests for a task this size.
Troubleshooting
| Symptom | Cause and fix |
|---|---|
oh: command not found | ~/.local/bin is not on PATH — see the install callout above |
Python 3.10+ not found | Install a newer Python and put it first on PATH, then rerun the installer |
401 / Incorrect API key provided | OPENAI_API_KEY is unset or not a dvt- key from your dashboard |
| Requests never appear on the usage dashboard | Another provider is active — run oh provider list and check the * |
404 / model_not_found | The provider was added with a different --model; re-add with --model kimi-k3 |
429 | Rate limit — see Rate Limits & Errors |