| name | opencode-ollama-agent-guardrails |
| description | Enforce builder/reviewer discipline for OpenCode and Ollama workflows, including no commits, no raw prompt publishing, relative paths only, validation honesty, and human approval after signoff. |
OpenCode Ollama Agent Guardrails
Use this skill when configuring, running, reviewing, or publishing an OpenCode + Ollama multi-agent coding workflow.
Required Agent Split
Builder agent:
- plans the current task;
- implements only approved scope;
- runs validation;
- writes implementation notes and test results;
- fixes blocking issues inside scope;
- stops for human review and commit approval.
Reviewer agent:
- reviews builder plans, diffs, tests, generated output, screenshots, and security posture;
- blocks scope creep and weak tests;
- writes only review and signoff files under the current session folder;
- never edits product code.
Hard Rules
- Do not commit, push, rebase, force-push, or start another prompt.
- Do not publish raw prompts, raw OpenCode logs, session exports, or private screenshots.
- Do not use local absolute paths in public-bound files.
- Do not claim success unless the required validation commands passed.
- Do not weaken tests to pass.
- Do not let reviewers edit production code.
- Do not let agents edit resolver-owned prompt model locks ending in
_MODELS.json.
- Stop after final signoff so a human can inspect and decide whether to commit.
Session Rules
Use:
agent-sessions/<project-slug>/<prompt-id>/
Product edits belong only under the configured product repo, usually:
product-repo/
Raw logs may stay local in the session folder, but they must remain ignored.
Decision Vocabulary
Reviewer decisions must be exactly one of:
APPROVEDCHANGES_REQUESTEDSIGNED_OFFNOT_SIGNED_OFF
