# Follow the initiating Codex route for auxiliary work

When a Harness Agent uses this bundle's `codex-local` provider, context-summary generation and `web_search` must use the same selected Codex model instead of a deployment-wide DeepSeek fallback. Compaction remains a one-shot Codex generation whose completed text becomes the durable Harness checkpoint; `maxTokens` is advisory because App Server exposes no output-cap equivalent. Web search is a conditional `tools/execute` takeover: when enabled, it runs one isolated App Server process per query with native live search, using either the settings-owned search model override or the initiating model, then lets the existing `web_search` output contract normalize citations. Disabled takeover and non-Codex Agents delegate unchanged to the installed provider chain. A separate process is required because the main Codex thread is waiting for the Harness tool result and cannot accept a concurrent turn.