| name | go-review-ultra |
| description | Run every go-review-* skill (style, nit, skeptic, and legibility if applicable) in parallel against a target Go directory and merge the findings. Use when the user wants a full review pass — e.g. "ultra-review @pkg/foo", "run all the go reviews on internal/search", "give me everything you've got on this package". Fans out via the Agent tool, deduplicates findings, and returns one combined report. Does not write code. |
Fan out the three go-review skills in parallel against a target directory, then merge.
Input
The user gives you a target directory (@pkg/foo or a path). If the user has a claim to scrutinize (the go-review-nit phase 1 input), capture it too. If the directory is missing, ask. If the claim is missing, proceed without it and tell go-review-nit to skip phase 1.
How to fan out
Spawn three Agent tool calls in a single message (parallel), using the dedicated agent types:
- go-review-style (
subagent_type: go-review-style): apply Go style and judgment conventions.
- go-review-nit (
subagent_type: go-review-nit): phase 1 (verify claim) + phase 2 (nit exported API). If no claim, tell it to skip phase 1.
- go-review-skeptic (
subagent_type: go-review-skeptic): phase 1 (bug hunt) + phase 2 (deletion sweep). No praise.
Prompt each agent with the target directory and any claim. They already have their instructions built in; just tell them what to review.
go-review-legibility: only run if the user supplied a list of inspected false-positive findings. This one stays as a skill (invoked via the Skill tool) because it needs user-provided input. If skipped, note it in the merged report.
Model selection
The Agent tool accepts an optional model parameter (sonnet / opus / haiku) that overrides the subagent's default per call. Support both a default and a per-invocation override:
- Default: no
model override. Subagents inherit the parent session's model. Keeps cost predictable.
- Per-invocation override. Parse a model hint from the user's prompt — e.g. "ultra-review @pkg/foo with opus", "…on sonnet", "…using haiku for the lighter passes". If a model is named, pass
model: <name> on every fanned-out Agent call. If the user gives a per-lens hint ("opus for skeptic, sonnet for the rest"), honor it.
- Recognized phrasing: literal model name (
opus / sonnet / haiku), optionally with on / with / using / via. Reject ambiguous hints by asking which model the user means.
Merging the findings
Each subagent returns a list. Merge into one report:
- Group by skill. Keep the three sections separate — the lenses are different and collapsing them loses signal.
- Dedupe by
file:line within a section. If style and nit both flag the same line for related reasons, keep both. They're different lenses. Only dedupe exact restatements.
- Per finding, include a "why" line. Don't just say what's wrong. Explain what goes wrong in practice if the pattern isn't fixed. The reader should learn the principle, not just the fix. Example: "Exported error variable instead of a type. Callers will compare with
== instead of errors.Is, which breaks when the error is wrapped."
- Combined severity table at the end, drawn from all three sections. Promote any finding the subagent marked high to the top.
- Top 1-3 to action first across all skills, opinionated. Don't list ties.
If go-review-nit's phase 1 pushed back on the claim, surface that before the merged findings. The user needs to redirect before the rest of the report matters.
Gotchas
- Agents don't inherit conversation context. Pass the target dir, the claim (if any), and any constraints in the spawn prompt.
- 3x the token cost of a single review. The win is parallelism and keeping intermediate searches out of the main context.
What this skill does not do
- Write code or apply fixes. Output is a merged findings report; the user picks what to fix.
- Re-review with a fourth lens. The three sub-skills are the contract.
- Run
go-ultra-optimize — that skill is invoke-explicitly-only and is not part of the standard review pass.
