architect-review

name	architect-review
description	Run iterative architect review passes on a design spec, tracking convergence. Use when a spec is written and needs hardening before implementation — each pass finds fewer issues until the design stabilizes.
allowed-tools	Agent, Read, Edit, Bash, Glob, Grep

Architect Review

Run multiple architect review passes on a design spec, with each pass reading the spec as modified by the previous. Track changes in a changelog file to measure convergence.

When to use

• After writing or updating a design spec, before moving to implementation planning
• When you want independent review of architectural decisions
• When the user says "review the plan", "harden the spec", "architect review"

Do NOT use when:

• The spec is still a rough draft the user is actively editing -- wait until they say it's ready for review
• The document is implementation code, not a design spec -- use code review instead
• The spec is under 1 page -- a single careful read is enough, no iterative passes needed

How it works

Each pass is a background Opus agent that reads the current spec, makes concrete edits, and reports what it changed. Passes run sequentially — each reads the output of the previous. A changelog file tracks every change across passes — see step 1 for where it lives and how it's named.

The pattern converges: pass 1 finds many issues, pass 2 finds fewer, pass 3 finds edge cases, pass 4 typically finds nothing. Stop when a pass makes 0-2 changes.

Process

1. Pick the changelog location and create it

Two locations, picked by one question: is the architect review being run on a single existing plan/spec file in a repo?

• Yes → beside the plan, as a plain rolling changelog. Path: <spec-dir>/<spec-name>-changelog.md. No datestamp in the filename, no architect-review in the filename — it's the changelog for this plan. If the user re-runs an architect review later, append a new dated section to the same file (don't create a second file). The changelog travels with the plan in git, gets committed alongside spec edits, and forms the canonical record of how the spec evolved before any PR exists.

• No → ~/tmp/architect-review/<slug>-<YYYY-MM-DD-HHMM>.md. This covers multi-file reviews, scratch input, external repos you don't want to pollute, and anywhere there isn't one obvious "the spec file" to sit next to. The datestamp lives in the filename here because there's no rolling file to append to — multiple reviews need to coexist as separate files. Run mkdir -p ~/tmp/architect-review first.

When in doubt, ask the user: "I'm putting the architect review changelog at <path> — sound right?" A wrong guess that drops a file into the reviewed repo is more annoying than a one-line check.

Check whether the file already exists before seeding it. This is the common case for the beside-spec mode on a re-run — the prior review's changelog is still there.

• File does not exist → seed it with this template:

  # <Spec Title> — Changelog
  
  Spec: <absolute path to spec>
  
  ## Architect Review — <YYYY-MM-DD HH:MM>
  
  ### Convergence Tracking
  
  | Pass | Changes |
  | ---- | ------- |
  

• File already exists → do NOT overwrite. Append a new ## Architect Review — <YYYY-MM-DD HH:MM> section (with its own ### Convergence Tracking table) at the end of the file. Prior sections stay untouched. This is how multiple reviews accumulate on the same plan.

2. Run passes as background agents

Launch each pass using the Agent tool with model: "opus" and run_in_background: true. Wait for each pass to complete before launching the next. Each agent gets this prompt structure:

You are a senior software architect doing pass <N> on a design spec.
Previous passes made [X, Y, Z] changes respectively (see the changelog).

Read the changelog at <changelog-path> first to understand what previous passes changed.
Then read the spec at <spec-path> carefully.

Context: <brief description of the system>

Review for:
- [pass 1] Architectural soundness, interface design, missing concerns, unnecessary complexity, consistency
- [pass 2] Anything pass 1 missed or got wrong, contradictions from pass 1 edits, over-engineering
- [pass 3+] Subtle implementation bugs, remaining ambiguity, over-engineering from previous reviewers

Make specific, concrete edits using the Edit tool. Each edit is a "change" — cosmetic rewording does NOT count. If the spec is solid, make zero changes — don't change for the sake of changing.

Report:
1. Numbered list of changes made and why
2. Things reviewed but left unchanged
3. Assessment: is this ready for implementation?

3. After each pass completes

1. Read the agent's final report and extract the numbered list of changes.
2. Print the delta summary to the main session before doing anything else. Numbered list of the concrete changes the pass made, each with a one-line rationale. The user wants to see what moved between passes — not just "pass N done." This is mandatory: if you skip the summary, the user can't tell whether the pass was useful without manually diffing the spec.
3. Append the same numbered changes + rationale to the changelog file (the durable record).
4. Update the convergence table row: | Pass N | <count> |.
5. If changes > 2: launch the next pass (background).
6. If changes <= 2: stop — the spec has converged.

4. Report convergence

Tell the user the final state and always point at the persisted changelog (whichever location you picked in step 1) so they can revisit pass-by-pass diffs later:

Architect review complete — 4 passes, converged.
Pass 1: 21 changes | Pass 2: 13 | Pass 3: 5 | Pass 4: 1
Full changelog: <absolute path to the changelog file>

If the changelog landed beside the spec (the in-repo iteration case), also remind the user to git add it — it's part of the iteration record, not transient noise.

5. (Optional) Post the convergence summary to the PR

If the spec lives in a git repo on a branch with an open pull request, post the convergence summary as a PR comment so reviewers see the architect's findings without digging into the orchestrating session. This is best-effort — every failure mode here is a silent skip, never an error.

Detect:

gh pr view --json url,number 2>/dev/null

If that returns nothing — no PR for the current branch, gh not installed, gh not authenticated, not in a git repo — skip this step entirely. The architect review is still successful; the PR comment is a bonus, not a requirement.

If a PR is found, build the comment body (a temp file is easiest):

## Architect Review — <YYYY-MM-DD HH:MM>

**Result:** converged in 4 passes (or: stopped at 5 passes, still open: <list>)

| Pass | Changes |
| ---- | ------- |
| 1    | 21      |
| 2    | 13      |
| 3    | 5       |
| 4    | 1       |

**Final assessment:** <ready for implementation / needs more work because …>

Full changelog: `<absolute path>`

Then post:

gh pr comment <number> --body-file <body-path>

Always include the run timestamp in the comment header. Don't try to update an existing comment in place — surfacing the full review history (one comment per re-run) is more valuable than keeping the comment count low. The timestamp makes it obvious which review is which.

Requires the gh CLI installed and authenticated. If either is missing, this whole step is a silent no-op — do not nag the user, do not fail the run.

Key rules

• Always run agents in background — don't block the conversation
• Sequential, not parallel — each pass must read the previous pass's edits
• Opus model for all passes — architecture review needs the strongest reasoning
• Changelog is mandatory — without it, you can't measure convergence
• Beside the plan as <spec>-changelog.md when reviewing one in-repo plan file; ~/tmp/architect-review/<slug>-<datestamp>.md otherwise — see step 1
• Multiple reviews on the same beside-the-spec changelog append as new dated sections — don't make a second file
• Check before seeding — if the beside-spec changelog already exists, append a new dated section; never overwrite
• PR comment is best-effort — every failure mode in step 5 is a silent skip; the architect review never fails because gh isn't installed
• Surface deltas after every pass — the main session must print the numbered change list, not just announce that the pass finished
• Don't over-run — 4 passes is typical. Stop at 5 max even if not fully converged. Report what's still open.
• Read the spec's repo CLAUDE.md first — understand project conventions before reviewing
• A "change" is a substantive architectural edit — fixing a typo or rewording a sentence for style does not count toward the convergence threshold

Anti-patterns

• Running passes in parallel (each would review stale state)
• Making cosmetic/wordsmith changes (wastes a pass)
• Continuing past convergence (diminishing returns)
• Not creating the changelog (can't track progress)