fix(parser): fall back to message.id when requestId is absent

[estelledc]

Summary

Fixes ~2.6× token over-counting on Claude Code installs that point at a proxied API endpoint (corporate Bedrock gateways, third-party LLM relays). Direct-API users are unaffected.

Root cause

dedup_by_request_id groups assistant chunks by requestId and deltaizes cumulative snapshots. When requestId is absent the event falls through to the without_req path, which deduplicates by line_uuid. But Claude Code assigns a unique uuid per JSONL line — every content-block line of one streaming response gets a different uuid — so this fallback never collapses anything. Modern Claude Code also writes the same final usage snapshot on every content-block line, so an N-block response gets summed N times.

The proxy case is the trigger: corporate gateways often strip the requestId request header but preserve message.id (it lives in the response body). On a real user's archive (511 sessions, 28,581 assistant-with-usage lines, 12,619 unique message.ids) this produced inflation factor 2.265× (token-weighted: 2.635×, since longer responses have more chunks). The 2.0.0 changelog claim "totals now match what Anthropic actually billed" silently regresses to 2-3× over-count for these users.

Fix

request_id = raw.request_id.or_else(|| msg.id.clone())

Both id-spaces (req_… and msg_…) are globally unique per Anthropic API call, so substituting message.id when requestId is missing reuses the existing canonicalization machinery (deltaize, ghost-chunk extraction, canonical-stream picker, mirror dedup) without touching any of it. Direct-API users keep request_id semantics exactly as before — the fallback only kicks in when the field is None.

Verification

	tokens (sum of all events)
Pre-patch CCMeter	4,437,331,044
Post-patch CCMeter (simulated)	1,690,796,103
CCCostMonitor reference (dedup by message.id in its py implementation)	1,683,903,949
Delta	0.4% (residual: null-msg.id line handling differences, not algorithmic)

Tests

• All 30 existing parser tests pass.
• Adds 3 regression tests:
  • dedups_assistant_chunks_by_message_id_when_request_id_absent — same-file streaming, identical cumulative snapshots
  • mirror_dedup_uses_message_id_when_request_id_absent — parent + sub-agent files mirroring three cumulative chunks
  • prefers_request_id_over_message_id_when_both_present — sanity that requestId still wins when present

Test plan

• cargo test parser:: → 33 pass (30 existing + 3 new)
• Real-data sanity: post-patch totals on a 511-session proxy archive match CCCostMonitor's reference within 0.4%
• Reviewer to confirm direct-API behavior unchanged (covered by existing captures_request_id, deltaized_tokens_sum_to_final_snapshot_across_streams, etc.)

🤖 Generated with Claude Code