ask
Ask the expensive Side-Eye judge about THIS session. --all = full sighted review, --turn N = recent exchanges, free text = scoped opinion, --validate = free setup check. User-invoked; spends real money except --validate and --help.
معلومات المصدر
- المستودع
- redhat-et/sideeye
- آخر نشاط في المصدر
- ١٥ سبتمبر ٢٠٢٦ في ٠٢:٣٢
- لغة SKILL.md المكتشفة
- الإنجليزية
- النجوم
- ٠
- التفرعات
- ٠
خيارات التثبيت
يُحدَّد Prompt الذي يراجع المصدر أولًا بشكل افتراضي. يمكنك التبديل إلى أمر مباشر أو تنزيل نسخة محلية.
مراجعة ملفات المصدر
اقرأ SKILL.md وأي ملفات مرافقة يعرضها SkillsMP قبل أن تقرر التثبيت.
عرض SKILL.md
SKILL.md
تعليمات المصدر · معاينة للقراءة فقط- name
- Ask
- description
- Ask the expensive Side-Eye judge about THIS session. --all = full sighted review, --turn N = recent exchanges, free text = scoped opinion, --validate = free setup check. User-invoked; spends real money except --validate and --help.
- disable-model-invocation
- true
- allowed-tools
- Bash(sideeye review *), Bash(sideeye advise *), Bash(sideeye route), Bash("$HOME/.local/bin/sideeye" review *), Bash("$HOME/.local/bin/sideeye" advise *), Bash("$HOME/.local/bin/sideeye" route), Bash(command -v sideeye*), Bash(uv tool install *), Bash(pipx install *), Bash(python3 -m pip install *)
# /sideeye:ask — ask the expensive model
This session runs on a cheap model. This command sends this session (or part of
it) to a real Claude judge for an independent second opinion. The user types what
they want checked; the judge's answer comes back verbatim.
## Decide the mode (first match wins, top to bottom)
| User typed | Mode | Command to run |
|---|---|---|
| nothing, or `--help` / `help` | usage | print the Usage section below and stop (no spend) |
| `--validate` (with or without other text) | setup check | `sideeye route` (free) |
| `--all` | full review | `sideeye review --current --yes` |
| anything else (free text, `--turn N`) | scoped opinion | `sideeye advise --current --yes --question "<their text>"` |
Routing rules:
- **`--validate` wins over everything** — it is the free diagnostic; run
`sideeye route` alone, print its stdout verbatim, stop.
- **`--all` wins over free text** — if both appear, run the review and mention
in one line that the text was ignored.
- `--turn N` or `--turns N` → add `--turns N` to the advise command (default 1).
- Free text becomes `--question "<text verbatim>"`. It may be empty
(e.g. `/sideeye:ask --turn 2` with no text is a valid scoped opinion).
- Pass through these flags **unchanged** in whichever mode applies:
`--model <fable|opus|sonnet|haiku|full-id>`, `--no-code`, `--max-cost N`.
Do not invent flags the user didn't type.
## Engine (one-time, first use only)
The judge runs on the `sideeye` engine — usually already on PATH. Check:
```
command -v sideeye
```
If it prints a path, skip to "Run it". If it prints nothing, install it once
(~30s; the first available tool wins):
```
uv tool install git+https://github.com/yossiovadia/sideeye || \
pipx install git+https://github.com/yossiovadia/sideeye || \
python3 -m pip install --user git+https://github.com/yossiovadia/sideeye
```
Two things can go wrong:
- Install succeeded but `command -v sideeye` is still empty (a fresh
`~/.local/bin` is not on this shell's PATH) → run the command as
`"$HOME/.local/bin/sideeye" <same subcommand and flags>` instead.
- The install line failed (none of uv / pipx / pip is available) → tell the user
to install one (`brew install uv`) and stop.
## Run it
Run the command from the mode table with a **generous Bash timeout
(≥ 300000 ms)** — a full-session judge call can take a few minutes.
Then:
1. **Present the command's stdout VERBATIM**, inside a fenced code block. Do not
summarize, soften, reorder, re-score, or "clean up" any of it. That output is
the judge's verdict — the user must see it exactly as written. After the
verbatim block, you may add one short line offering to address the issues it
lists.
2. **On a nonzero exit code:** show the command's error output verbatim and
**STOP**. Do not retry with different flags or arguments. Exception: in
`--validate` mode the diagnostic report is on **stdout** (exit 1 just means
"not yet ready") — show stdout there, not stderr.
3. **NEVER write the verdict yourself.** This is the whole point of the tool: a
verdict invented here would be produced by the *cheap* model this session
runs on — the exact plausible-but-fake output Side-Eye exists to catch. Only
the real stdout of `sideeye review` / `sideeye advise` / `sideeye route` is a
verdict. If the command fails, returns nothing, or is blocked, say so plainly
and stop — do not fabricate scores, issues, or a summary under any
circumstances.
## Usage (print when the user typed --help or nothing)
```
/sideeye:ask — ask the expensive model (the Side-Eye judge) about this session
/sideeye:ask --validate free setup check: is the judge route
configured and unblocked? (free)
/sideeye:ask --help this text (no spend)
/sideeye:ask <question> the judge's opinion on your question,
scoped to recent work (~2¢–$1)
/sideeye:ask --turn 3 <question> same, over the last 3 exchanges
/sideeye:ask --all full sighted review of the whole session
+ code diff (median ~75¢, up to ~$3.50)
extra flags: --model fable|opus|sonnet|haiku · --no-code (skip the diff)
· --max-cost N (abort above this ceiling)
```
## Notes
- The judge must hit a real Claude route — never this session's own (cheap)
model. A "no judge route" or BLOCKED error means the route isn't configured:
surface the message verbatim, don't work around it. `--validate` (free) shows
exactly what resolves and why.
- If the estimate exceeds the cost ceiling, the command aborts with a message
telling the user how to raise `--max-cost`. Show it; don't bypass it.
- The authoritative flag lists are always `sideeye review --help` and
`sideeye advise --help`.
عرض على GitHub