Configuration - OpenClaude Loop - openclaude-loop

Arguments

Argument	Default	Meaning
`mode`	`full`	`full` includes recon and requirements; `review` starts from an existing plan
`plan` / `PLAN_FILE`	`PLAN.md`	plan path, used for review, build and approval binding
`log` / `LOG_FILE`	`PLAN-REVIEW-LOG.md`	append-only audit trail
`builder`	`claude`	`claude` or `opencode`
`rounds` / `MAX_ROUNDS`	`5`	maximum completed plan-review rounds
`MAX_FIX_ROUNDS`	`2`	bounded build-fix rounds
`MAX_INSPECTION_ROUNDS`	`2`	initial inspection plus one after fixes
`inspect`	`on`	`off` is a logged explicit opt-out only
`research`	proportionate	`none`, `web`, or explicitly opted-in `deep`
`reviewer_model`, `builder_model`, `inspector_model`	OpenCode default	`provider/model`
`reviewer_variant`, `builder_variant`, `inspector_variant`	none	sent as `provider/model#variant`
`PROOF_CMD`	derived from the repo	exact verification command

Example:

/openclaude-loop:openclaude-loop mode=review plan=docs/rfc.md builder=opencode rounds=3 reviewer_model=anthropic/claude-sonnet-5

Models

This matters for reading any verdict honestly.

Claude Code — the orchestration host, your current session.
OpenCode — an external coding-agent CLI. A program, not a model.
The OpenCode model — whatever provider and model you configured for OpenCode.

So the independence of a review depends entirely on how you configured OpenCode. Point it at the same provider as your session and you get a second agent with a fresh context and no authorship bias — genuinely useful, but not a second provider. The plugin calls OpenCode the independent agent, the external reviewer or the second agent, and claims a cross-provider review only when the configured model demonstrably comes from a different provider.

Requested and observed models are recorded separately, and an unresolved default is reported as unresolved. If you name a model OpenCode does not have, the run stops. Nothing is ever silently substituted.

Testing the plugin itself

python scripts/validate.py                 # plugin metadata, skill structure, references
python -m unittest discover -s tests -v    # 73 contract tests
claude plugin validate .                   # Claude Code manifest validation

The suite drives a fake OpenCode CLI and disposable git repositories, so it consumes no provider quota. Fixtures deliberately use repository paths containing spaces and non-ASCII characters.