execute-review

Execute Review

Orchestrate a structured codebase review. You drive a two-phase pipeline -- scout then review -- and present the output to the user. Context discipline is critical throughout: every tool call and subagent dispatch should serve the review, nothing else.

REQUIRED: You should have been provided with a scope (directory, file, glob, or project root). If you have not, ask and immediately exit.

The Pipeline

1. Scout the Codebase

Dispatch a Haiku-powered subagent to produce a structural inventory of the codebase within the given scope. The scout maps structure without reading file contents in detail.

<invoke name="Agent">
<parameter name="subagent_type">ed3d-basic-agents:haiku-general-purpose</parameter>
<parameter name="description">Scout codebase structure</parameter>
<parameter name="prompt">
You are a codebase scout. Your job is to produce a compact structural inventory of a codebase without reading full file contents. Be fast and precise.

**Scope:** [INSERT SCOPE]

Use Glob and Bash (for line counts, directory listing) to map:

1. **Languages**: which programming languages are present and their relative proportion (file count per extension)
2. **Frameworks/Libraries**: identify from config files (package.json, Cargo.toml, go.mod, requirements.txt, etc.)
3. **Directory layout**: top-level structure and what each directory appears to contain (src, lib, tests, config, scripts, docs, etc.)
4. **Entry points**: main files, index files, CLI entry points, server startup files
5. **Test locations**: where tests live (test/, __tests__/, *_test.go, *_spec.ts, etc.) and rough count
6. **Configuration**: build configs, CI/CD files, environment configs
7. **File counts**: total files by type within scope

Output a compact markdown inventory. Do NOT read file contents -- only names, paths, and metadata. Keep it under 300 words.

Format:

Structural Inventory

Languages: Go (45 files), TypeScript (23 files), ... Frameworks: Echo, React, ... Layout:

• src/ -- main application code
• pkg/ -- shared packages
• ...

Entry Points: cmd/server/main.go, src/index.ts, ... Test Locations: *_test.go alongside source (38 test files), src/tests/ (12 files) Config: Makefile, docker-compose.yml, .github/workflows/ci.yml Total Files: ~150 within scope

</parameter>
</invoke>

When the scout returns, print the structural inventory to the user so they can see what was found.

2. Select Files to Review

Using the structural inventory, select a representative sample of files to review. The goal is to cover the most important code without exceeding what a single reviewer agent can handle effectively.

Selection priorities (in order):

1. Entry points -- main files, server startup, CLI entry points. These reveal the application's shape.
2. Core business logic -- the modules that do the actual work, not glue or boilerplate.
3. Shared utilities/packages -- code that many other files depend on. Problems here cascade.
4. Configuration and infrastructure -- build configs, CI, Docker. These shape reliability.
5. Tests -- a sample of test files, especially for core modules. Test quality reveals code quality.
6. API boundaries -- route definitions, handler files, API schemas. Trust boundary code.

Selection constraints:

• Aim for 15-30 files for a single-pass review. Fewer for small codebases, more only if fan-out is used.
• Prefer depth over breadth -- it's better to thoroughly review 15 files than to skim 40.
• If the scope is a single file or small directory, review everything in it.
• Skip generated files, vendored dependencies, lock files, and binary assets.

Use Glob and Read (first few lines only, to assess relevance) to make your selections. List the selected files.

3. Decide: Single Pass or Fan-out

Estimate whether the selected files fit in a single reviewer's context:

• Single pass (default): 30 or fewer files, codebase is cohesive (one primary language/framework). Dispatch one reviewer.
• Fan-out: More than 30 files after sampling, OR the codebase has clearly separate modules (e.g., frontend + backend, multiple microservices). Partition files by module/directory and dispatch multiple reviewers in parallel.

Fan-out heuristic: Be conservative. If in doubt, fan out. It's better to dispatch two reviewers with clean context than one reviewer with truncated context.

If fanning out, partition files into groups where each group:

• Contains files from the same module/directory
• Has its own entry points and tests where possible
• Gets a copy of the structural inventory for orientation

4. Dispatch Reviewer(s)

Single Pass

Dispatch the blank-slate-review:codebase-reviewer agent with the structural inventory and file list.

<invoke name="Agent">
<parameter name="subagent_type">blank-slate-review:codebase-reviewer</parameter>
<parameter name="description">Review codebase: [scope summary]</parameter>
<parameter name="prompt">
## Structural Inventory

[PASTE SCOUT INVENTORY HERE]

## Files to Review

[LIST ALL SELECTED FILE PATHS, ONE PER LINE]

## Scope

[SCOPE DESCRIPTION]

Read every listed file and produce your review following the output contract in your instructions.
</parameter>
</invoke>

Fan-out

Dispatch multiple blank-slate-review:codebase-reviewer agents in parallel, one per module partition. Each gets the full structural inventory (for orientation) but only its partition's files.

<invoke name="Agent">
<parameter name="subagent_type">blank-slate-review:codebase-reviewer</parameter>
<parameter name="description">Review module: [module name]</parameter>
<parameter name="prompt">
## Structural Inventory

[PASTE SCOUT INVENTORY HERE]

## Files to Review

[LIST THIS PARTITION'S FILE PATHS ONLY]

## Scope

This is a partitioned review. You are reviewing the [module name] module as part of a larger codebase review. Focus on your assigned files.

Read every listed file and produce your review following the output contract in your instructions.
</parameter>
</invoke>

When all reviewers return, merge the results:

1. Concatenate findings from all reviewers
2. Re-number finding IDs sequentially (C1, C2, ..., H1, H2, ..., etc.)
3. Deduplicate: if two reviewers flagged the same file:line with the same issue, keep one
4. Merge remediation themes: combine overlapping themes from different reviewers, update finding ID references
5. Recalculate metrics

5. Present Output

Print the complete review to stdout. Use the exact output contract format from the reviewer agent. If results were merged from fan-out, the merged output should be seamless -- the user should not need to know how many reviewers were dispatched.

After printing the review, offer to write the findings to a file:

<invoke name="ToolSearch">
<parameter name="query">select:AskUserQuestion</parameter>
<parameter name="max_results">1</parameter>
</invoke>

Then ask:

Would you like me to write these findings to a file?
- Yes, write to [suggest a path like docs/review-YYYY-MM-DD.md or review.md]
- No, stdout is fine

If the user says yes, write the review output to the specified file. Do not commit it -- let the user decide.

Context Discipline

This skill is designed around context cleanliness. Follow these rules:

• No extraneous tool calls. Every Glob, Read, Grep, or Bash call should directly serve the scout or sampling phases. Do not explore out of curiosity.
• Be smart about tool use. Use tools intelligently. Use tools like ast-grep (sg) to surgically and thoroughly work your way through the codebase, use treesitter or LSPs if appropriate. For heavyweight tools spin out a subagent to prevent from polluting main context.
• No commentary between phases. Move from scout to sampling to dispatch without narrating your thought process at length. Brief status updates only.
• No redundant reads. If the scout already told you what languages are present, do not re-scan for languages.
• Subagent prompts are self-contained. The reviewer agent should have everything it needs in its prompt -- structural inventory and file list. Do not assume it can see your conversation history.
• Merged output is clean. If fan-out was used, the final output should read as a single coherent review, not a concatenation of separate reports.

Common Rationalizations -- STOP

Excuse	Reality
"Let me read a few more files to be thorough"	No. The sampling phase selected files for a reason. Trust the sample.
"I should review this file too, it looks interesting"	No. Stick to the sample. Adding files mid-review wastes context.
"I'll skip the scout and just start reading code"	No. The scout inventory drives intelligent sampling. Never skip it.
"Fan-out is overkill for this"	If you're unsure, fan out. Truncated context is worse than an extra subagent.
"I'll add some suggestions for improvements"	No. This is observation only. Findings and remediation, not implementation.