Route: Pick the Right Workflow Pattern

Based on Anthropic's "Building Effective Agents" research. The right pattern depends on the problem — using the same approach for everything is the #1 efficiency killer.

Analyze the task described in $ARGUMENTS and recommend ONE pattern:

Pattern Decision Tree

1. Simple Fix (Direct Execution)

When: Single file, clear scope, you can describe the diff in one sentence. Examples: Typo fix, rename a variable, add a log line, small bug fix. Action: Skip planning. Just do it. Verify with a quick test or build.

2. Prompt Chaining (Sequential Phases)

When: Multi-file change with known files. Clear what needs to happen. Examples: Add a new API endpoint, refactor a function used in 3 places. Action: Explore → Plan (list exact files + changes) → Implement → Verify.

3. Routing (Explore First)

When: Uncertain scope. You don't know which files are involved yet. Examples: "Fix the auth bug" (which auth? where?), "Improve performance" (of what?). Action: Use a subagent to explore the codebase first. Based on findings, pick pattern 1 or 2.

4. Orchestrator-Workers (Parallel Subagents)

When: Large feature with independent subtasks. Multiple files that don't depend on each other. Examples: Add validation to 10 API endpoints, migrate 5 components to new pattern. Action: Create plan, break into independent tasks, dispatch subagents in parallel.

5. Evaluator-Optimizer (Iterative Refinement)

When: Quality-sensitive work where the first version won't be good enough. Examples: Complex algorithm, security-critical code, public API design. Action: Implement → Self-critique → Iterate → Verify. Use /challenge from self-critic plugin.

Adaptive Retrieval Strategy

After choosing a workflow pattern, recommend a retrieval strategy — what information to gather and how. Different task types benefit from different retrieval approaches (Meta-Harness, arXiv 2603.28052: domain-specific routing outperforms one-size-fits-all retrieval).

Bug Fix Retrieval

• git log --oneline --since="2 weeks" -- <file> — recent changes that may have introduced the bug
• git blame <file> on the broken section — who changed what, when
• Grep for related error messages in traces (~/.claude/traces/*.jsonl)
• Check test files for the affected code — are there tests? Do they cover this case?

New Feature Retrieval

• Grep the codebase for similar existing features — match their patterns
• Read the closest analogous component end-to-end before writing anything
• Check for shared utilities, base classes, or traits you should extend
• Look for migration patterns if the feature needs schema changes

Refactor Retrieval

• Find all callers/dependents: grep -rn "functionName\|ClassName" --include="*.php"
• Check test coverage for affected code — are there tests that will catch regressions?
• git log --oneline --diff-filter=M -20 -- <file> — understand change velocity
• Identify the blast radius: how many files does this touch?

Performance Fix Retrieval

• Check for N+1 queries, missing indexes, or unbounded loops in the affected path
• Look at similar performance fixes in git history: git log --all --oneline --grep="perf\|slow\|optimize"
• Profile data if available (logs, traces, APM)

Documentation/Config Retrieval

• Read existing docs for the affected area — match the style
• Check related config files for patterns to follow
• Verify the change is consistent with README and any API specs

Output Format

RECOMMENDED PATTERN: [Pattern Name]
REASON: [One sentence why this pattern fits]
RETRIEVAL: [Which retrieval strategy above + specific commands to run]
FIRST STEP: [What to do right now]

Key principle from Anthropic: "Start simple. Add complexity only when it demonstrably improves outcomes."