fix(codex): normalize LLM call logging to per-turn boundaries

[aaight]

Summary

• Root cause: Codex logged one storeLlmCall row per usage-bearing event (e.g. response.completed), instead of one row per completed turn — resulting in a single giant aggregate call or duplicate rows per run
• Fix: Introduces a turn-scoped accumulator (CodexTurnAccumulator) that collects text, tool names, and usage across JSONL events within a turn; persistence fires exactly once on turn.completed
• Schema unchanged: No database migrations needed; existing agent_run_llm_calls and runs API/UI contract is preserved

Key changes

src/backends/codex/index.ts

• Added CodexTurnAccumulator type to CodexLineContext to track in-progress turn state
• accumulateTurnUsage(): merges intermediate usage events (e.g. response.completed) into the accumulator without persisting
• persistTurnLlmCall(): called exactly once on turn.completed; writes a compact, turn-scoped payload (text summary + tool names + usage) rather than a raw event dump
• handleStructuralEvent(): turn.started/thread.started reset the accumulator; turn.completed accumulates any final usage then persists
• handleParsedLine(): intermediate usage events now call accumulateTurnUsage() instead of trackUsage() (which is removed)
• Usage ordering: turn.completed aggregate totals supersede intermediate response.completed per-response values

tests/unit/backends/codex.test.ts

• Updated 2 existing tests to emit proper turn.started + turn.completed lifecycle events (reflecting correct behavior)
• Added 4 new tests:
  • Multi-turn stream: verifies 2 turns produce 2 storeLlmCall rows with stable sequential call numbers
  • Duplicate-usage prevention: both response.completed and turn.completed carry usage — exactly 1 row stored
  • Compact payload shape: stored payload contains turn, tools, usage and is < 2 KB
  • No-turn-completed: bare usage events without turn.completed produce no rows

Test plan

• All 63 Codex unit tests pass
• All 9 shared-llmCallLogger tests pass
• TypeScript type check passes (zero errors)
• Biome lint passes (zero errors)
• Pre-existing test failures in unrelated files (postComment, github) confirmed pre-existing on base branch

Card: https://trello.com/c/ueUZYl4d/359-it-looks-like-for-codex-runs-we-store-only-one-gigantic-llm-call-in-our-agent-runs-llm-calls-lets-look-how-it-looks-with-other-e

🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

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

[nhopeatall]

Summary

LGTM — Clean refactor that introduces a turn-scoped accumulator to normalize Codex LLM call logging from one-row-per-usage-event to one-row-per-completed-turn, with solid test coverage for the key scenarios.

The implementation correctly:

• Initializes the accumulator both at context creation (for streams that lack turn.started) and on turn.started/thread.started events
• Uses "last value wins" override semantics in accumulateTurnUsage, so turn.completed aggregate totals correctly supersede intermediate response.completed per-response values
• Persists exactly once per turn boundary and resets the accumulator afterward
• Removes the raw JSONL line from the stored payload in favor of a compact, structured turn summary (text + tools + usage)
• Preserves the existing logLlmCall contract with llmCallLogger.ts — no downstream schema or API changes needed

The 4 new tests cover the critical scenarios well: multi-turn sequencing, duplicate-usage prevention, compact payload shape, and the no-turn-completed edge case. CI passes cleanly.

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