second-opinion

Use when adding a feature end-to-end — branch, implement, self-review, open a PR, resolve review, and merge.

Use when an agent hits a GUI/desktop wall — a task needs clicking, a desktop app, or a browser flow with no CLI/API.

Use when creating or modifying any scheduled job, reminder, or watchdog.

Use when a change touches a schema, collection, API response, or config that a UI renders — before finishing.

Use when branching, committing, opening or merging pull requests, or replying to code review — the repo-hygiene rules to follow.

Use for work that must run unattended past one turn — "keep going until finished", timed/overnight runs, "money making mode", monitoring a child process, or…

Review a branch/diff on two axes — Standards (repo conventions, code smells, correctness) and Spec (does it do what the issue/PRD asked?) — via parallel…

Use before claiming any task is done, fixed, working, deployed, or collected — including claims about work from an earlier session.

Use whenever a task's success is a message reaching someone — send, post, notify, remind, reply in another channel, or a scheduled message.

Use during any long session before context runs out — long tool loops, large outputs, multi-stage plans, compaction warnings, or resuming from a…

Choose cost-efficient OpenAI models and reasoning effort for subagents, workers, reviewers, coding CLIs, scheduled jobs, and agent configuration.

Diagnosis loop for hard bugs and performance regressions.