Capability Diagnostic

When to Use

Run this diagnostic when an agent:

• Fails a task (test failures, wrong output, incomplete work)
• Produces output that doesn't match the expected schema
• Gets blocked by enforcement hooks repeatedly
• Takes significantly longer than expected on a straightforward task

Failure Analysis Decision Tree

Step 1: Did the agent have the right tools?

• Check allowed-tools in agent spec (if any)
• Check if enforcement hooks blocked required tool calls
• Fix: Add missing tools to agent spec, or adjust hook configuration

Step 2: Did the agent have the right context?

• Check which skills were preloaded (agent spec skills field)
• Was the task specification complete and unambiguous?
• Did the agent need information from files it didn't read?
• Fix: Add missing skills to agent spec, improve task prompt

Step 3: Did the agent have sufficient model capability?

• Haiku struggles with: design decisions, cross-module reasoning, ambiguous specs
• Sonnet struggles with: subtle bugs, architecture-sensitive changes, performance optimization
• Opus handles: complex reasoning, cross-cutting concerns, deep analysis
• Fix: Escalate to next model tier (see escalation skill)

Step 4: Was the task specification ambiguous?

• Did the agent ask clarifying questions? (sign of ambiguity)
• Did the agent make assumptions that turned out wrong?
• Could two reasonable agents interpret the spec differently?
• Fix: Rewrite task spec with explicit inputs, outputs, and constraints

Step 5: Was it an infrastructure issue?

• Hook blocking (enforce-orchestrator-delegation, protect-hooks)
• Context window exhaustion (agent ran out of context mid-task)
• Transient errors (classifyHandoffIfNeeded, MCP connection drops, network timeouts)
• Git worktree conflicts (locked files, stale worktrees from crashed agents)
• SQLite lock contention (multiple agents accessing ReviewDb simultaneously)
• MCP server unavailable (Serena LSP not responding, Linear API rate limited)
• Fix: For transient errors, retry with fresh context. For persistent infrastructure, report and fix root cause.

Output Format

After diagnosis, report:

Diagnosis: [tool gap | context gap | model gap | spec gap | infrastructure]
Root cause: [one sentence]
Fix: [specific action to take]

Common Patterns