Auto Eng Review

auto-eng-review

Optional engineering-safety review. Validates that a plan is safe to execute before implementation begins.

First action: run node .agent/.automaton/scripts/get-context.mjs from the project root.

Preamble

Execution safety review. Architecture, data flow, edge cases, test strategy, not product vision. It does not change the plan's content or reopen product scope; its only write is appending its own ## Review: Engineering section. Identifies risks that could cause failure, stalling, or rework.

A good review names the riskiest slice, the most likely failure mode, and whether the test strategy catches it. A bad review lists generic concerns.

Loading discipline: one PLAN.md read, optional DESIGN.md when canonical_design exists, one risk matrix, one verdict. Read source files when assessing technical risk: slice boundaries, dependency assumptions, and blast radius claims are only verifiable against the actual code.

Quality Gate

Before appending the engineering review:

• Ground concerns in slices, file areas, commands, or missing artifacts.
• Separate blockers from follow-up cleanup.
• Avoid reopening product scope unless the plan is unbuildable.
• Read references/quality.md when findings are generic or unactionable.

Do

Do NOT proceed unless:

• canonical_plan is set and PLAN.md is readable.

If the plan is missing or unreadable, set verdict to needs_correction and stop.

Load State

Read the canonical PLAN.md. Read DESIGN.md only when canonical_design is set and resolves to a file. An unset pointer means the plan intentionally has no design artifact; continue without it. A set pointer with a missing file is stale: report it and continue (DESIGN.md is optional here).

Restate the Plan

In engineering terms: what is being built, what systems does it touch, and what is the critical path?

Evaluate Risks

Use this matrix as an internal checklist. Apply standards from references/prime-directives.md while evaluating.

Risk Matrix

Architecture fit, data flow clarity, edge case coverage, test strategy, rollback safety, dependency risk.

A dimension where you can already name the concrete failure mode is a blocking concern, and naming it is the review's job. Surface it explicitly. Blocking concerns return needs_correction. Dimensions that clear the bar but carry named, slice-scoped risks return approved_with_risks. Read references/engineering-sections.md only when the plan carries non-trivial engineering risk.

Render Verdict

Read references/implementation-alternatives.md only when PLAN.md lacks an approach rationale, the user asks for alternatives, or the verdict depends on comparing safer execution paths.

Use strict vocabulary: exactly one of the three approved values, no synonyms.

Outside Voice

Optional. When a second model is reachable from this session and the plan carries non-trivial risk, read references/outside-voice.md after rendering the verdict: a round-capped cross-model loop that arbiters findings with logged reasons, surfaces tension to the user, and never auto-applies anything.

Append Review

Add a ## Review: Engineering section to PLAN.md using the exact template in references/review-template.md.

Update State

Run node .agent/.automaton/scripts/sync-status.mjs --engineering-review "<verdict>" from the project root.

Hand Off

The review always stops. The edge's why: .agent/.automaton/references/ARTIFACT-LIFECYCLE.md (Handoff Contract).

Report the verdict, the riskiest slice, and the one concern execution should watch for. Then end the turn with **Next:** auto-execute, <reason> for approved or approved_with_risks, or **Next:** auto-plan, <reason> for needs_correction.

Output

• PLAN.md with appended ## Review: Engineering section
• .agent/.automaton/state/current.json updated through sync-status.mjs with engineering_review; stage is unchanged by this skill

Rules

• Do not emit the full risk matrix when all dimensions are acceptable. Keep the durable review to the review-template fields.
• Missing DESIGN.md is not a blocker when canonical_design is null, absent, or intentionally skipped by the plan.