| name | dte-deep-reviewer |
| description | Deep multi-lens review of a pull request, a branch diff, or a whole codebase, producing a written findings doc. Runs compound-engineering's cross-model code review, the intent-engineering lenses, majestic's Rails code-review orchestrator, and cubic review/scan, then adversarially cross-checks where they disagree. Use when the user says "review this PR / branch / diff", "deep review", "check this codebase", or hands over PR numbers/branch names. Loops over a list of PRs or branches. Not for architecture-level structure (use dte-arc-review) or for writing/fixing code (use dte-arc-work). |
dte-deep-reviewer — PR / branch / codebase review → findings doc
Read ../../references/conventions.md first. Output is a findings doc; it does not change code.
Input
- A target: a PR (number/URL), a branch (diff vs base), or a codebase/path.
- Or a list of PRs/branches → review each (parallel sub-agents;
--serial to slow it down).
- Read the project
CLAUDE.md/AGENTS.md first.
Procedure (per target)
- Resolve & contextualize. Get the diff (
gh pr diff, git diff <base>...<branch>, or the path).
Use Augment codebase-retrieval to pull the surrounding code the diff touches (callers, callbacks,
tests) so review isn't diff-blind.
- Run the review lenses in parallel (one sub-agent each; collect structured findings):
/ce-code-review — correctness, security, reliability, performance, testing personas + the
cross-model second opinion (the 3.15.0 win). Pass the plan/base when known.
/ie-review — predictability, convention, simplicity, experience, architecture lenses.
- majestic
review:rails-code-review — Rails-specific orchestrated review.
- cubic
/run-review + /scan (if available) — team-learned patterns + security scan.
- Front-end lens (conditional). If the diff touches FE files (
.erb/.haml/.slim, ViewComponent,
JS/TS, Stimulus, CSS/Tailwind — see conventions "Front-end lens"), add ie-experience-reviewer +
frontend-design/ui.sh (interaction states, a11y, keyboard/focus, look-and-feel). No FE files → say so, skip.
- Adversarial cross-check. Compare the lenses: where two or more agree → confirmed; where they
disagree on severity or existence → disputed (surface both, don't merge or average).
- Synthesize. Dedupe, severity-rank (🔴/🟠/🟡), each finding with
file:line + why + concrete fix.
Apply owned-convention (interactors-not-services). Heuristic findings → "suspected"; only
single-line-decidable ones get a hard claim.
- Write
docs/reviews/deep-review-<target>-<date>.md in the conventions shape, with the
Confirmed vs disputed section and tools used / unavailable.
For a list
One sub-agent per PR/branch → per-target docs + a roll-up docs/reviews/deep-review-rollup-<date>.md
ranking targets by risk and listing cross-PR themes.
Done when
Each target has a findings doc: one-line verdict, confirmed findings with evidence + fixes, the disputed
section, and (for PRs) a ready-to-paste summary comment. Hand fixes to dte-arc-plan / dte-arc-work.