Harbor Framework integration

Harbor Framework 0.21.x can invoke Argus itself as an installed agent. One Harbor trial maps to one bounded Argus project:

Harbor task + sandbox
        |
        v
Argus Manager -> Planner -> Engineer <-> Reviewer
        |
        v
Harbor verifier

The adapter does not recreate an Argus-like loop in Harbor. Harbor installs Argus inside the task environment and calls Argus's normal headless runtime with the task instruction in a private UTF-8 file. The complete team then works in the Harbor task workspace until the Planner certifies project_done or Harbor's trial timeout stops the run.

Requirements

• Python 3.12 or newer on the Harbor host
• Harbor Framework >=0.21,<0.22
• Docker or another Harbor environment provider
• an OpenAI model and credentials accepted by Harbor's Codex integration

Install Argus and Harbor in the same host environment:

python3.12 -m venv .venv
.venv/bin/python -m pip install -e '.[harbor]'

Run

export OPENAI_API_KEY=...

.venv/bin/harbor run \
  --dataset terminal-bench@2.0 \
  --agent argus_skill.integrations.harbor:ArgusHarborAgent \
  --model openai/gpt-5.4-mini \
  --ak reasoning_effort=high

Harbor calls ArgusHarborAgent.run(...). The adapter then:

1. installs Codex CLI and Argus in the task environment;
2. configures Harbor's model credentials for Argus's Codex backend;
3. uploads the Harbor instruction as argus-objective.txt, without placing the objective in process arguments;
4. starts argus-skill --daemon-fg --continuous --bounded;
5. treats the one-shot task as stage-closing and requires the native independent Reviewer even for low-risk verticals;
6. waits for the complete Argus runtime to finish;
7. reports Argus token/cost totals and completion metadata to Harbor.

The task environment needs Python 3.11+. When its system Python is older, the installer uses uv to provision an isolated Python 3.12 runtime.

Agent arguments

Argument	Default	Meaning
argus_package	wheel built from current checkout	optional pip requirement installed instead of the local source
codex_version	latest	Codex CLI version installed by Harbor
reasoning_effort	high	effort used by all four Argus roles
timeout	Harbor trial timeout	optional inner Argus timeout in seconds

By default, a source checkout builds its current wheel on the Harbor host, uploads it, and installs that exact wheel in the task environment. This means uncommitted adapter/runtime changes are evaluated too.

When the adapter is loaded from an installed package rather than a source checkout, point argus_package at an immutable wheel or Git revision:

.venv/bin/harbor run \
  --dataset terminal-bench@2.0 \
  --agent argus_skill.integrations.harbor:ArgusHarborAgent \
  --model openai/gpt-5.4-mini \
  --ak 'argus_package=argus-skill @ https://packages.example/argus_skill.whl'

The package reference is shell-quoted before it is passed to pip. Prefer an immutable version, wheel digest, or commit instead of a moving branch.

Trial artifacts

Harbor's agent log directory contains:

• argus-objective.txt — exact task given to Argus;
• argus-runtime.log — headless Argus stdout/stderr;
• argus-state/ — complete persistent Argus project state, including continuous.json, events.jsonl, backlog, checkpoints, and transcripts;
• codex-home/sessions/ — native model sessions created by Argus roles.

After the trial, Harbor's AgentContext.metadata.argus records whether the bounded project completed, the Planner's done reason, state path, model call count, and pricing status. Token and cost fields are populated from Argus's own usage ledger.

Compatibility boundary

The initial direct adapter uses Argus's Codex backend because Harbor already has a maintained Codex installer and model-credential contract. The invoked Argus runtime is otherwise the normal complete runtime: Manager, Planner, Engineer, Reviewer, persistent state, checkpoints, and completion gates are not replaced.

The optional dependency is pinned to Harbor 0.21.x. Test the installed-agent contract before widening that range.