fix: AI Conversations input/output correlation (0.2.11 + 0.2.12)

[Joshkop]

Summary

Two related fixes to make Sentry AI Conversations show the right input/output pair per turn.

• 0.2.11 (already on branch): Stop hook now closes the current turn so the transcript JSONL flush race doesn't drop gen_ai.output.messages; late-flush retry now triggers when recordOutputs is on and response is null; synthesized sessions refuse ordinal matches.
• 0.2.12 (this commit): align transcript-reader's real-turn index with what UserPromptSubmit actually counts.

0.2.12 root cause

Claude Code's transcript records client-only slash commands (/model, /clear) as a synthetic user-line triple (<local-command-caveat> + <command-name> + <local-command-stdout>), but never fires UserPromptSubmit for them. The transcript reader was treating each of those lines as a real turn, so the collector's turnIndex (incremented per UserPromptSubmit) drifted off-by-N from transcript ordinals. Real UserPromptSubmit hook events don't carry prompt_id, so selectTurn fell back to ordinal — and the ordinal pointed to the wrong transcript turn.

Visible symptoms (conversation 9763be13-…, span 9d1cb6a39ecde5ba):

• Early turns: gen_ai.input.messages set, no gen_ai.output.messages, claude_code.usage_extraction.status = turn_had_no_usage
• Later turn: gen_ai.output.messages present but it was an earlier turn's response

Fix

• src/transcript-reader.ts: pre-pass marks every prompt_id that has any <local-command-*> content as client-only; main loop skips isMeta:true lines, client-only-pid lines, and bare <local-command-*> lines.
• src/server.ts + src/types.ts: accept camelCase promptId on UserPromptSubmit events as defense-in-depth alongside prompt_id.

Test plan

• New failing test in tests/transcript-reader.test.ts reproduces the off-by-N with a /model synthetic triple followed by two real prompts; passes after the fix.
• Full suite: 232 tests passing, tsc --noEmit clean.
• Live verification: run a session that uses /model mid-conversation and confirm in Sentry AI Conversations that each turn's input matches its output.

🤖 Generated with Claude Code