Add a labs skill for multi-agent integrated feature review #1

New issue

Open

opened 2026-07-24 06:24:49 -05:00 by deadstyle · 0 comments

deadstyle commented

2026-07-24 06:24:49 -05:00

Owner

Problem

Turn the "full integrated review" pattern into a reusable labs skill (plus supporting
review agents), so a completed feature can be reviewed by an army of parallel agents
without hand-crafting the dispatch each time.

The pattern, as run by hand against the Kirria Slack integration:

Scope the feature. Identify the set of files that make up the completed work
(already-merged is fine — the review target is the current on-disk state, not a diff).
Write a single shared context brief: file list, working directory, architecture
pointers, and a strict findings-output format.
Fan out one agent per review lens, each cold-started against the shared brief,
read-only, told to load the relevant project skills first. The lenses used:
- Code quality (opus) — idiomatic standards, DRY / reuse, unneeded complexity,
  deep modules.
- Test quality & coverage (opus) — coverage of the paths that matter, positive
  and negative cases, tests that don't actually verify behavior.
- Consistency (opus) — layer placement, naming/convention adherence, and
  excessive comments.
- Bugs / logic gaps (opus) — correctness, each finding with a concrete failure
  scenario.
- Observability (sonnet) — can on-call tell it's healthy and diagnose it: metrics,
  logging, failure-mode coverage.
- Security — authn/authz, tenant isolation, injection, secret handling, signature/
  replay verification, input validation at trust boundaries, dependency risk.
- Tech debt — accumulated shortcuts, TODO/ponytail: markers, stale comments,
  deprecated patterns, deferred work, and drift between code and its docs/wiki.
Consolidate. The dispatcher collects each agent's findings, de-dupes overlap
between lenses, and produces one prioritized report.

Design notes worth capturing for whoever builds it:

Lenses should be a configurable set, not hard-coded five — the observability lens was
added mid-run on request.
Each lens is a distinct agent so it stays in its lane; overlap is expected and
de-duped at consolidation, not prevented up front.
Model per lens is tunable (correctness-heavy lenses got opus; observability got sonnet).
The shared context brief is the leverage point — one file every agent reads, so the
briefs stay lean and consistent.
Read-only by contract: agents report findings with file:line + severity, they don't
edit.

This lived as an ad-hoc parallel dispatch; make it a skill so it's repeatable.

## Problem Turn the "full integrated review" pattern into a reusable labs skill (plus supporting review agents), so a completed feature can be reviewed by an army of parallel agents without hand-crafting the dispatch each time. The pattern, as run by hand against the Kirria Slack integration: 1. **Scope the feature.** Identify the set of files that make up the completed work (already-merged is fine — the review target is the current on-disk state, not a diff). Write a single shared context brief: file list, working directory, architecture pointers, and a strict findings-output format. 2. **Fan out one agent per review lens**, each cold-started against the shared brief, read-only, told to load the relevant project skills first. The lenses used: - **Code quality** (opus) — idiomatic standards, DRY / reuse, unneeded complexity, deep modules. - **Test quality & coverage** (opus) — coverage of the paths that matter, positive *and* negative cases, tests that don't actually verify behavior. - **Consistency** (opus) — layer placement, naming/convention adherence, and excessive comments. - **Bugs / logic gaps** (opus) — correctness, each finding with a concrete failure scenario. - **Observability** (sonnet) — can on-call tell it's healthy and diagnose it: metrics, logging, failure-mode coverage. - **Security** — authn/authz, tenant isolation, injection, secret handling, signature/ replay verification, input validation at trust boundaries, dependency risk. - **Tech debt** — accumulated shortcuts, TODO/`ponytail:` markers, stale comments, deprecated patterns, deferred work, and drift between code and its docs/wiki. 3. **Consolidate.** The dispatcher collects each agent's findings, de-dupes overlap between lenses, and produces one prioritized report. Design notes worth capturing for whoever builds it: - Lenses should be a configurable set, not hard-coded five — the observability lens was added mid-run on request. - Each lens is a distinct agent so it stays in its lane; overlap is expected and de-duped at consolidation, not prevented up front. - Model per lens is tunable (correctness-heavy lenses got opus; observability got sonnet). - The shared context brief is the leverage point — one file every agent reads, so the briefs stay lean and consistent. - Read-only by contract: agents report findings with file:line + severity, they don't edit. This lived as an ad-hoc parallel dispatch; make it a skill so it's repeatable.

No labels

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

deadstyle/claude-marketplace#1

No description provided.

Rows
Columns