Architecture review

Find choices that could make a technical proposal wrong, unsafe, or impossible to prove. Review the proposed behavior and tradeoffs, not the document's size or format.

Use one fresh subagent that did not write the proposal. Give the reviewer the complete context. Tell the reviewer to work directly without delegating. If you are that reviewer, review directly. Stay read-only.

If fresh subagents are unavailable, stop and report that independent review is blocked. Continue only if the user explicitly accepts a documented self-review.

Process

1. Read the goal, proposal, repository instructions, relevant current code,

tests, schemas, configuration, and linked material. Treat claims about the current system as unverified until code, tests, schemas, configuration, infrastructure, or relevant runtime evidence supports them.

1. State the problem, affected user, intended outcome, success measure, scope,

constraints, and main tradeoff. Report any that the proposal leaves unclear.

1. Trace one real case from input to observable outcome. Include ownership,

validation, state changes, side effects, response timing, failure, retry, cleanup, and what the user sees where they matter.

1. Challenge the chosen design with the review focus below. Look for a simpler

choice that reaches the same outcome and proof with less state, coupling, duplication, or operational work.

1. Surface material open questions. Recommend an answer when evidence supports

one. Do not invent questions that cannot change the design.

1. Return findings, open questions, a short assessment, and one verdict.

Do not rewrite the proposal, plan the work, review implementation code, or implement changes.

Review focus

• Check fit with the current system, ownership, boundaries, interfaces, data,

compatibility, migration, rollout, and rollback.

• Trace partial failure, retry, cancellation, concurrency, startup, shutdown,

and recovery where they affect the proposal.

• Check claimed scale, limited resources, latency, throughput, storage, cost,

dependency failure, and operator recovery only where they can change the choice.

• Check identity, authorization, untrusted input, credentials, destructive

authority, and sensitive data handling.

• Require observable acceptance criteria and proof for important rules and

failure paths. Do not let implementation invent user-visible behavior, interfaces, data rules, security policy, or failure behavior.

Material questions

Report an open question only when two capable implementations could answer it differently in a way that affects users, data, interfaces, security, scale, performance, operations, cost, compatibility, or proof.

• Blocking: implementation should not start without the answer.
• Important: the proposal should record the answer, but the reviewer can

recommend a safe default from available evidence.

Omit questions that are stylistic, safely local to implementation, outside the stated scope, or speculative beyond the scale the proposal claims to support.

Findings

Report only flaws that can change the design or its safety:

• Blocker: the choice is unsafe, contradicts the goal or current system, or

cannot recover from an important failure.

• Important: the proposal permits materially different implementations or

leaves a meaningful risk in behavior, scale, performance, security, operations, compatibility, or proof.

For each finding or open question include its location, concrete failure or ambiguity, impact, evidence, and the smallest correction or recommended answer. Report unclear wording when it prevents a new teammate from explaining, evaluating, or implementing the design. Omit other writing preferences.

Use plain words. State the exact condition, failure, and effect. Do not hide an unknown behind vague language such as may have issues or could be risky.

Verdict

• Approve: no blocker or important findings or open questions remain.
• Request changes: the proposal has a fixable blocker or important finding or

open question.

• Blocked: the review lacks required context, repository evidence, specialist

coverage, or an independent reviewer.

State what remains unverified. Do not approve because the document is detailed or because every template section exists.