feat(backends): add beforeExecute/afterExecute lifecycle hooks to AgentEngine interface

[aaight]

See PR description

🕵️ claude-code · claude-sonnet-4-6 · run details

[nhopeatall]

Summary

LGTM — well-designed lifecycle hooks that cleanly separate engine-specific setup/cleanup from the adapter orchestration layer. The interface extension is backward-compatible (optional methods), all engines are updated consistently, and all 571 backend tests pass.

Architecture & Design

The pattern is sound: beforeExecute/afterExecute in the AgentEngine interface give each engine a clean injection point for pre/post work, while the adapter's nested try/finally ensures afterExecute always runs even on failure. The backward compatibility story for direct execute() callers (e.g., tests) is well handled — ClaudeCode removes internal cleanup entirely, Codex uses _adapterLifecycleActive to avoid double-cleanup, and LLMist simply doesn't implement the hooks.

Should Fix

• src/backends/opencode/index.ts:812 (afterExecute) — Unlike ClaudeCode (which fully moved cleanup to afterExecute) and Codex (which uses _adapterLifecycleActive to skip internal cleanup when the adapter lifecycle is active), OpenCode's execute() still calls cleanupContextFiles in its finally block (line 883-885 of the full file) and afterExecute calls it again unconditionally. This results in duplicate cleanup when called through the adapter. While cleanupContextFiles is idempotent and this won't cause failures, it's an inconsistency with the other engines. Consider either (a) removing the if (hasOffloadedContext) block from execute()'s finally, or (b) adding the _adapterLifecycleActive guard pattern from Codex.

🕵️ claude-code · claude-opus-4-6 · run details