AI Diff Reviewer addon

Every Deep Work Plan closes the same way: a mandatory Final Review that reads the plan’s entire accumulated change set before the work can be called done. Its security pass is the last point at which anything gets caught. Without help, the only reader at that point is the same agent that wrote the code.

This addon puts a second reader on that diff. It wires the AI Diff Reviewer — listed on the marketplace as “AI Diff Reviewer”, current release v3.3.0 — into the security pass, where it returns something structured rather than prose: a verdict, a findings table, and a severity on each finding. Since v3, a critical finding means the addon’s verifier confirmed it with a second, code-grounded model call; only verified criticals block completion until fixed or explicitly accepted. The review is a gate, not a comment.

Since standard 2.3.0 that local review is part of the baseline, not an extra. Onboarding installs it; every Final Review runs it. What stays optional is the CI surface — Flow B, where the same review gates pull requests through the GitHub Action.

The boundary that makes this safe to adopt is deliberately narrow. The reviewer is an MIT-licensed, tag-pinned skill executed by the coding agent you already run, so no Deep Work Plan flow depends on a commercial service, a CI provider, or a secret. Flow A (local-only) is what every onboarded repository gets; Flow B is offered explicitly and never installed unrequested. A developer may also decline the reviewer outright — that decline is recorded as a declared exception, and verify reports the repository as non-conformant on that point until it is installed.

When to use it

Two adoption flows

Detection for the local review requires skill + an extension file at one of: .review/extension.md, .github/ai-diff-reviewer/extension.md, or .github/ai-pr-reviewer/extension.md. Skill alone is not enough.

What this addon wires (narrow by design)

The DWP addon does not reinvent the reviewer. It defers install, methodology, the CI wizard, extension authoring, PR drafting, and post-CI review loops to the upstream skill’s six sub-skills (parent default flow, generate-extension, setup, open-pr, apply-review, address-review).

The required local review

create adds the local review step to every Final Review’s security pass and execute runs it. Output appends under ## AI Diff Reviewer local review in the plan-local analysis_results/SECURITY_REVIEW.md (inside the plan’s own folder, never the repository root).

• Missing reviewer — recorded, never skipped silently: a missing skill or extension becomes a local reviewer not installed finding; Final Review runs the local pass when the skill is present and otherwise carries the finding into the completion report — installation belongs to the onboarding consent or an explicit addon invocation, never a surprise bootstrap.
• Soft-fail (invocation only): a review that could start but errors → warn once, record, continue; never fail the task for that miss.
• Gate after a completed pass: verified critical findings block Final Review completion until fixed or explicitly accepted (BC-07). Unverified critical claims arrive as annotated warnings — visible, non-blocking unless strict-unverified-criticals: true restores claim-based gating. A review that hit its turn cap (incomplete) or its wall clock (timeout) is not a clean pass under blocking strictness (BC-04). warning / info are documented but non-blocking.
• Flow A needs no CI secret. An unset provider key must not suppress the local pass.

Flow B CI gate (optional)

Action DailybotHQ/ai-diff-reviewer@v3, typically label-gated (ready), with a stable-named AI review gate job for branch protection and opt-in skip-review-label: skip-ai-review. Since v3 the review budget follows the change’s deterministic risk tier — 8/20/30/40 turns from low to critical under budget-profile: auto (fixed restores the pre-v3 constants during the transition) — and a push that changes no code runs a verifier-only round. Shared prompt.md + extension align methodology and severity; local and CI stay methodology-identical, while CI round 2+ may be shorter under Iteration-Aware Review and the local pass stays full.

Optional review companions

Two developer-invoked sub-skills close the loop after CI posts a review; neither is ever a plan task file (that would break mandatory final-task order).

• apply-review walks findings per-finding (apply / defer / skip) with consent. Read-only by default; never commits or pushes.
• address-review (new in v3.1.1) is the one-invocation loop: find the branch’s open PRs, check the review is fresh for the current head, present the findings with an apply/defer/skip plan, then — on one yes — apply, commit in small Conventional Commits batches, push, and re-arm the reviewer the way the repo triggers it (label-gated → toggle the label off/on; push-triggered → confirm the new run). Unlike apply-review, it commits and pushes; that is the loop’s point. On aggregated ensemble reviews it reads the aggregate document and the ai-pr-reviewer-aggregate marker.
• The machine path is the structured output, not the review body: the review-output/3.0 document (.aiprr/review-output.json, located through the structured-output-path and structured-output-sha256 outputs) carries the run record, findings with evidence and verification, refuted findings, and the gate. A review body that says Recommendation: approve is not evidence the check passed — read the tracking marker’s Highest severity / Strictness gate / Check status block first.

What changed in v3

Three releases landed on 2026-09-24 (v3.0.0, v3.0.1, v3.1.0) and v3.1.1 followed with the address-review sub-skill; v3.2.0 and v3.2.2 (2026-09-26) then made that loop CI-aware and able to arm a PR the reviewer never ran on. None of them changes how this addon wires the reviewer — Flow A, the three detection paths and the never-block ladder are unchanged — but they change what an adopter gets.

Two of these matter more than the rest for the methodology.

The verified-criticals gate hardens the security pass. Before v3, a model could claim a critical freely and the gate followed the claim. Now a critical finding in a Final Review means a second model call confirmed it against the code, and the release campaign measured the difference: 771 paid evaluation runs across ten campaigns at roughly $83 total, with the critical tier reaching 63/63 recall at adjudicated precision 1.0, and incremental rounds cutting input tokens by 62–76 %. These figures are upstream’s published measurements, not Deep Work Plan’s own.

economy is deliberately not cheaper. The upstream benchmark of 2026-09-16 measured grok-4.3 at 0 of 5 known defects — it approves without reviewing — while grok-4.5 matched grok-4.6 at 3 of 5 with no false positives, at the same cost and a quarter of the wall time. Since no cheaper xAI model still reviews, economy resolves to the same model as balanced rather than being a tier that finds nothing; on the grok runner, the risk tier’s turn budget, not the model choice, is what scales cost. The xAI path therefore moves from roughly $0.07 to roughly $0.40–0.75 per review through the CLI; Since the v3.0 prompt a 63-run re-measurement moved grok-4.3 from 0 of 5 to 83 % recall with zero false positives at ~31 % lower cost — but it still fails to write its findings file about three times as often as grok-4.5, so grok-4.5 remains the default and model: grok-4.3 stays an explicit, accepted-trade-off choice. These figures are upstream’s published measurements, not Deep Work Plan’s own.

Behavior

• Flow A is the baseline; Flow B is asked, never guessed. Installing a workflow unrequested is a larger footprint than staying on Flow A.
• Reconcile, don’t clobber. Existing skill, extension, or review workflow are preserved; fill gaps only.
• Auth deferred. Provider secrets for CI are maintainer-configured; this addon never stores credentials.
• Vendor-neutral. No commercial service, CI provider, or secret is ever required; the CI surface is the only piece that touches a provider.

Notes

Local review required since standard 2.3.0; CI surface optional. Upstream skill: DailybotHQ/ai-diff-reviewer. Upstream migration guide: docs/MIGRATION_v3.md. Spec page: Add-ons.