Outsider

Outsider — Quick Opinion From Another Agent

Purpose

Get a fast opinion from an agent CLI running outside this session. One reviewer, no synthesis step — a sanity check, not an exhaustive review. Use review:consilium when you want a board.

This skill writes one temp file per run — the question — and nothing else. The responder is launched read-only: it reports, and it does not edit the repo.

The value is structural: the responder shares none of this conversation's context and runs in its own process, usually on a different model family. Which vendor answers is incidental, which is why this skill resolves the agent instead of naming one.

When to Use

• "ask codex", "ask claude", "what would another agent say", "get an outside opinion"
• /outsider, /outsider review, /outsider codex review
• A quick external sanity check on an approach or decision
• An independent review of the current code changes

Requirements

timeout or gtimeout must be on PATH — both ship with GNU coreutils. Stock macOS has neither; brew install coreutils provides gtimeout. Without one the script skips instead of running an agent CLI unbounded, because a hung agent would hang the caller's turn. run-outsider.sh list reports it when it is missing.

Agent Selection

The script picks the agent. Pass --host <the agent you are> on every invocation — claude, codex, opencode, or pi — so it never asks you to review your own work. Add --agent <name> only when the user named one; that overrides everything, host included.

bash <skill-dir>/scripts/run-outsider.sh list --host claude

list prints what is installed and what would be selected, and spends nothing.

Default order is codex claude opencode pi, minus the host. A missing CLI is a skip, not a failure: the script prints why and exits 0. Callers can treat this leg as droppable.

No model or reasoning level is set by default — the chosen agent runs on whatever it is already configured to use. To pin one, or to change the preference order, edit ~/.config/edloidas/outsider/config (OUTSIDER_AGENTS, OUTSIDER_MODEL_<AGENT>, OUTSIDER_EFFORT_<AGENT>, OUTSIDER_ARGS_<AGENT>); the same names work as environment variables and override the file. references/agents.md has the full registry, the config keys with examples, and how to add an agent.

Ask Mode

1. Prepare a focused question with context. Extract only the relevant code or plan excerpt — don't dump the conversation. Keep it under 2000 words; past that, use review mode, which extracts the diff itself.

2. Resolve the temp directory (once per session):

   bash <skill-dir>/scripts/resolve-tmp.sh
   

   Use the output as <TMP>. Pick a run id <ID> once too — a session identifier the host already exposes, or the current YYYYMMDD-HHMMSS — and reuse the same literal string for every file in this run so concurrent runs don't overwrite each other.

3. Write the question with a file-write tool, not a Bash heredoc — heredocs with markdown headers trip some hosts' shell security heuristics. Write to <TMP>/outsider-<ID>-question.md:

   ## Question
   <clear, specific question>
   
   ## Context
   <relevant code, plan excerpt, or description>
   

4. Run it, after printing one line naming the agent that will answer and the wait — Asking codex (ask mode, up to 2 min).

   bash <skill-dir>/scripts/run-outsider.sh ask --host claude <TMP>/outsider-<ID>-question.md
   

Custom preamble

Both modes prepend a prompt file to whatever you pipe them — references/prompt.md for ask, references/review-prompt.md for review. A caller with its own prompt for the responder passes --preamble <file> to swap it:

bash <skill-dir>/scripts/run-outsider.sh ask --host claude \
  --preamble <caller-skill-dir>/references/its-own-prompt.md <TMP>/outsider-<ID>-question.md

The preamble replaces the default entirely, so it has to carry the responder's whole brief — including its output shape. review:consilium uses this for its outside board seat.

Review Mode

Pick the scope first:

Review takes 3–10 minutes. Print one line before it starts, naming the agent, the scope, and the wait — Asking codex to review 6 uncommitted files (up to 9 min). Pass 540 so the script's own timer fires first and can print its timeout message, and set the surrounding command timeout to the highest value the host allows:

bash <skill-dir>/scripts/run-outsider.sh review --host claude --uncommitted 540

Presenting Output

The first line of the output is [outsider] agent: <name>. Name that agent in your report — the response is not interpretable without knowing who gave it.

• Lead with "Codex's take:", "From pi:", and so on, using the agent the script actually ran
• Evaluate the findings against your own context rather than adopting them
• Highlight agreements; flag disagreements and explain which side you land on and why
• The responder only ever saw what you piped it. Discount findings that are really requests for context it could not see
• Empty output or an error line means it had nothing to offer — say so and move on

Codex's take: it agrees the cache key is wrong and would key on tenant + URL,
same as the fix on the branch.

One disagreement: it wants `resolveTenant()` awaited at the call site. It could
not see `src/router.ts`, where the await already happens — discounting that one.

One thing it caught that I had not: the seeded test passes against the buggy
key, so the case as written would not have failed before the fix.

Then stop. Do not run a second agent, and do not re-ask the same question with more context — one outside opinion is the whole deliverable.

Edge Cases

• No agent installed, or only the host is — the script says so and exits 0. Don't retry.
• Timeout — the script prints a timeout message. Note it and proceed without.
• No timeout binary — the script skips and says which tool is missing, per Requirements. Report that and proceed without an outside opinion; do not retry.

Notes for Callers

This skill stays model-invocable on every host. The predecessor claude skill set disable-model-invocation: true only because calling Claude from Claude was recursive; selection now handles that structurally, so there is nothing left to suppress.

Other skills invoke this one as a skill, not by script path — a repo-relative path only resolves inside one checkout.

review:consilium runs its outside board seat through this skill, and review:doubt runs its outside verification seat the same way — both pass their own seat prompt as the preamble. That was the last duplicate runner in the collection; there is now one implementation of "call an agent CLI that is not the host".