Debugging Mac AI Agent Failures: A Layered Checklist for Codex, Claude Code, OpenClaw, and Computer Use

Most "the model refuses to act" failures on a Mac are not model-quality failures.\n\nThey are usually setup failures across one of five layers:\n\n1. Provider/model routing: the selected model is unavailable, blocked by provider policy, or not the runtime that owns the tool you are trying to use.\n2. Auth state: OpenRouter, Claude Code, Codex, Gemini, or local gateway credentials are stale or attached to the wrong profile.\n3. Session isolation: one TUI/session is reused across unrelated project directories, so the agent answers from the wrong state.\n4. macOS permissions: Screen Recording, Accessibility, and per-app approval can disagree with each other.\n5. Runtime visibility: Computer Use or browser-use can exist on disk while the active agent runtime cannot actually call the MCP/tool.\n\n## A practical triage order\n\nDo these in order. Do not start by changing prompts.\n\n### 1. Verify the active model route\n\nRead the model shown in the active session and run a small provider smoke test. If the error says no allowed providers are available, the prompt is irrelevant. Fix the route first.\n\n### 2. Prove session isolation\n\nStart sessions from two different project directories and confirm they get separate session IDs and memory/search scopes. If both directories attach to the same old conversation, fix that before debugging behavior.\n\n### 3. Separate permission state from tool state\n\nFor Computer Use, "Accessibility is enabled" is not enough. Verify the actual runtime can list apps, inspect the active UI, or request access for a target app.\n\n### 4. Check native-module/runtime path errors\n\nFor Claude Code Computer Use crashes such as checkAccessibility or registerEscape being undefined, inspect whether the native Computer Use module exists and whether the CLI knows where to load it from.\n\n### 5. Reduce verbosity before increasing intelligence\n\nIf the agent is verbose, freezing, and burning context, lower default thinking on routine tasks, isolate specialist sessions, and reserve expensive/deep models for code review, complex refactors, and business planning.\n\n## I made a low-friction paid read\n\nIf you have one failing OpenClaw, Codex, Claude Code, Computer Use, OpenRouter, Gemini, or macOS-permission symptom and want a quick outside read, I opened a $19 async quick read.\n\nQuick read page: https://igorganapolsky.github.io/openclaw-mac-ai-workstation-setup/quick-read.html?utm_source=devto&utm_medium=article&utm_campaign=quick_read_20260505\n\nCheckout: https://buy.stripe.com/aFaeVd3Ug3n05pLfSH3sI0u\n\nNo secret pasting, no macOS security bypasses, no revenue guarantees. Redact tokens and customer data before sending logs.