Adaptive Review

Adaptive code review that routes to haiku/sonnet/opus based on diff complexity signals. Use instead of requesting-code-review for cost-efficient reviews.

Adaptive Code Review

Review code changes with model depth proportional to change complexity. No wasted opus tokens on trivial diffs.

Step 1: Collect Signals

Run these commands to gather diff signals:

# Get diff stats (against HEAD~1 or origin/main, whichever makes sense)
BASE=$(git merge-base HEAD origin/main 2>/dev/null || echo "HEAD~1")
git diff --stat $BASE..HEAD
git diff --numstat $BASE..HEAD

Extract:

• lines_changed: total added + deleted
• files_changed: number of files
• dirs_changed: number of unique top-level directories touched (cross-module indicator)

Then scan for high-risk patterns — only in code files (exclude .md/.txt/.json/.yaml from grep):

git diff $BASE..HEAD -- '*.ts' '*.js' '*.py' '*.go' '*.rs' '*.java' '*.c' '*.cpp' '*.rb' '*.sh' | grep -ciE '(password|secret|token|auth|session|cookie|sql|inject|exec\(|eval\(|lock|mutex|semaphore|atomic|concurrent|unsafe)'

• risk_hits: count of matches (0 if only docs/config changed)

Step 2: Route

Condition	Depth	Model
lines_changed < 50 AND files_changed <= 1 AND risk_hits == 0	fast	haiku
lines_changed < 200 AND dirs_changed <= 1 AND risk_hits <= 2	medium	sonnet
Everything else (>200 lines OR dirs_changed >= 2 OR risk_hits > 2)	deep	opus

Announce the routing decision:

Review depth: [fast|medium|deep] (N lines, N files, N dirs, N risk hits)

Step 3: Dispatch

Fast (haiku)

Spawn agent with model: "haiku", subagent_type of your code-review agent:

Prompt focus: formatting, naming conventions, obvious bugs, unused imports. Skip architecture analysis. Keep it under 30 seconds.

Medium (sonnet)

Spawn agent with model: "sonnet", subagent_type of your code-review agent:

Standard code review: correctness, error handling, test coverage, code quality.

Deep (opus)

Spawn agent with model: "opus", subagent_type of your code-review agent:

Full review: architecture, security, performance, cross-module impact. If language-specific reviewers exist (python-reviewer, go-reviewer, database-reviewer), spawn them in parallel.

Step 4: Report

Present results with depth label so the user knows what level of review was applied:

## Adaptive Review: [FAST|MEDIUM|DEEP]
Signals: {lines} lines, {files} files, {dirs} dirs, {risk_hits} risk hits

[reviewer output]

If fast review finds anything concerning, suggest upgrading: "Fast review flagged potential issues. Run /adaptive-review --deep for thorough analysis."

Overrides

User can force depth:

• /adaptive-review --fast — force fast regardless of signals
• /adaptive-review --deep — force deep regardless of signals
• /adaptive-review --medium — force medium