How Claude Code and aider react

Question: "do skills and hooks always work?" Answer: yes, they are client-side. What can degrade is the model behind, not the client.

Roles today: Claude Code through the proxy (px-claude) is the daily agent; aider is the cheap editor for targeted changes; OpenCode is benchmarked as a third tool (Phase 1).

Claude Code with ANTHROPIC_BASE_URL → llm_brain

brain setup claude-code --proxy prints the px-claude shell function (--docker for the containerised variant): ANTHROPIC_BASE_URL = the proxy, ANTHROPIC_AUTH_TOKEN = the dev client key, main model brain/auto (Auto routing), background ("haiku") model brain/fast, experimental betas and adaptive thinking disabled. Plain claude keeps the Anthropic subscription. Pin a tier with ANTHROPIC_MODEL=brain/agent px-claude or /model brain/agent.

Always works (runs locally in the CLI)

Feature	Why
Hooks (PreToolUse, PostToolUse, Stop, UserPromptSubmit…)	Shell scripts executed by the CLI
Skills	Markdown injected into the context: reaches the model as text
CLAUDE.md, slash commands, permissions	Client-side
MCP servers	Tools run locally; only the schema reaches the model
Subagents	Just more HTTP requests
claude-kit (submodule .claude-kit/)	Portable hooks/skills/agents: no impact, reused as-is

What the proxy has to get right

SSE passed through unbuffered, count_tokens forwarded, claude-* ids mapped to a tier, anthropic-beta / anthropic-version passed on, and the sanitizer: drop context_management, output_config and adaptive thinking, turn Claude Code's mid-conversation system turns into user turns (without it non-Claude models answer with an empty reply). Details and reasons: Client compatibility.

Degrades with the model behind (doesn't break, gets worse)

• Tool-call reliability: weak models produce malformed tool calls → retries → more tokens → more cost. The agent tier (deepseek-v4-pro) exists for this: it is the tier with a measured reliable Claude Code loop, and brain/auto's fallback.
• Skill adherence: long prompts; a weak model partly ignores them.
• Extended thinking: only if the model behind has reasoning.
• Context: the limit is the model's behind (1M on the current tiers).
• Prompt caching: Claude Code leans on it heavily; a turn on a new backend re-reads the whole history (Cache layers).

aider with OPENAI_API_BASE → llm_brain

• brain setup aider --proxy prints px-aider: architect brain/reasoning, editor brain/fast, auto-commits, the repo's CLAUDE.md read as conventions. It is for targeted changes, not the agent loop.
• brain setup aider also writes .aider.model.metadata.json (context and prices of the aliases, otherwise aider misestimates) and the model settings: diff / editor-diff edit formats (aider defaults unknown models to whole, expensive and fragile) and the tier's OpenRouter fallback chain.
• No hooks or skills: the equivalents are --read, --lint-cmd, --test-cmd, --auto-test. --auto-test is the natural signal for escalation, which is not built yet (API layer).