refactor(backends): extract shared LLM call logging helper

[aaight]

Summary

• Extracts the duplicated storeLlmCall wrapper pattern from all three native-tool engines (Claude Code, Codex, OpenCode) into a shared logLlmCall() helper at src/backends/shared/llmCallLogger.ts
• The helper guards on runId (no-op when absent), fires storeLlmCall asynchronously, and catches/logs any storage errors using a per-engine label
• Each engine now uses the shared helper instead of maintaining its own inline wrapper

Changes

• src/backends/shared/llmCallLogger.ts (new): logLlmCall() helper accepting runId, callNumber, model, optional token counts, optional costUsd, optional response, and an engineLabel for error messages
• src/backends/claude-code/index.ts: replaced inline logLlmCall() function with shared helper call; renamed local function to logClaudeCodeLlmCall to avoid naming collision
• src/backends/codex/index.ts: replaced storeLlmCall call inside trackUsage() with shared helper; removed unused logger import
• src/backends/opencode/index.ts: replaced async storeUsage() function with synchronous call to shared helper; cleaned up now-redundant .catch() at call site
• tests/unit/backends/shared-llmCallLogger.test.ts (new): 9 unit tests covering guard-on-runId, fire-and-forget behaviour, and error-catch logging with engine label

Test plan

• npm run typecheck passes (zero errors)
• npm run lint passes (zero errors)
• New unit tests pass (shared-llmCallLogger.test.ts — 9 tests)
• Full backend unit test suite passes (554 tests across 27 files)

Trello card: https://trello.com/c/69b572d0893fa82bfa5c6dd8

🤖 Generated with Claude Code

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

[nhopeatall]

LGTM - the shared logLlmCall() extraction preserves the existing storage contract and engine behavior while removing duplicated fire-and-forget error handling. I verified the new helper test, the affected backend unit suites, and npm run typecheck locally with no regressions.

🕵️ codex · gpt-5.4 · run details