Second Opinion is a local orchestration helper. It does not proxy prompts through a hosted service and does not store API keys.

• Installs a CLI in ~/.local/bin.
• Writes one managed skill file per selected agent.
• Updates its own CLI file when second-opinion update downloads a newer published version.
• Runs installed agent CLIs in non-interactive mode when asked.
• Stores task metadata and logs under ~/.second-opinion/jobs/ so either local manager can show and steer delegated work.
• Stores a small, public Artificial Analysis benchmark cache under ~/.second-opinion/cache/. Benchmark fetches send no repository content, task text, credentials, or job metadata and require no API key.
• Routes explicit image tasks through an image tool already exposed by a selected coding harness, attaches only the reference paths the user supplied, and verifies the requested workspace artifact before reporting success.
• Opens an optional native Tk task-manager process for background tasks by default. --manager terminal uses a terminal window and --manager none opens nothing.

• It does not modify model settings or pin a default model.
• It does not read or write provider credentials.
• It does not call image APIs directly, install image generators, or request image API keys.
• It does not install MCP servers.
• It does not install hooks.
• It does not change agent permission settings.
• It does not run a server, browser runtime, or always-on daemon. Background jobs and the optional manager are explicit local child processes started only for Second Opinion usage.
• It does not modify, inject code into, or control the Codex, Claude Code, OpenCode, Grok, or Antigravity applications.

Installed skill files include a managed-file marker. If a target file exists without that marker, install skips it unless --force is supplied.

Uninstall removes only marked files:

second-opinion uninstall --all

second-opinion update refreshes only marked, managed skill files by default. Use --all-skills only when you intentionally want every supported Second Opinion skill regenerated.

work mode is the CLI default and may edit files. Use explicit --mode consult when the user wants a read-only review, plan, or risk analysis.

Parallel team runs create multiple native harness processes. Give work-mode workers non-overlapping file or subsystem ownership. Second Opinion adds collision-avoidance instructions, but it cannot make arbitrary simultaneous edits conflict-free. Review strategies default to consult mode, and the parent agent remains responsible for synthesis and verification.

The generated skill tells parent agents not to pass secrets and not to create recursive delegation chains unless explicitly requested.

Steering messages and model changes are written only to the selected task record. A steering turn runs through the task's original native agent harness and its existing local authentication. Changing a model affects a future turn; it does not mutate a provider process already running. Stopping a task terminates its process group but does not revert workspace edits that the task already made.

Image outputs must resolve inside the selected workspace and use a supported raster extension. Existing outputs are preserved unless --force is explicit. Reference files may live elsewhere because Codex and OpenCode support native attachment flags and some other harnesses support additional readable directories; Second Opinion passes only paths named on the command line. A successful image task must create or update the artifact and pass a small constant-memory file-signature check.