fix: attribute error/aux rows to the requested model, not 'unknown'

[acastellana]

From the 2026-07-15 dashboard review: the "By model family" panel showed 857 requests / 20% err in a single opaque unknown bucket. Cause: failed and non-chat requests carry no served model_family (it comes from the upstream x_router, absent on error), so they all collapsed to unknown with no idea which model they targeted. requested_model was captured (auth_proxy.py:3337) but never used as a fallback.

Fix

Fall back to the requested model when the served family is absent, in all three aggregation paths — kept in lockstep (the SQL-vs-Python parity test enforces it):

• host_store.usage_aggregate family_k — persistent all-time view
• auth_proxy._stats live family buckets — recent feed
• auth_proxy._aggregate_usage_rows — historical replay fold

Served family stays the primary key for 2xx rows; the stored model_family column is untouched (still null on error) — only the display bucketing changes. policy_ir/flow_ir routed requests now show under their route label instead of unknown too.

Tests

New test_error_rows_attribute_to_requested_model_not_unknown pins it; full suite 501 passed, 2 skipped.

Deploy

Ships with the pending router deploy (main 6f80677, devops).

Summary by CodeRabbit

• Bug Fixes
  • Improved dashboard usage reporting for error and auxiliary events without model-family data.
  • These events are now attributed to the requested model or route instead of being grouped under “unknown.”
  • Runtime and historical usage views now provide more consistent model-family totals.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 91dbb8ba-4488-4c2a-9e9b-c6cbadde9044

📥 Commits

Reviewing files that changed from the base of the PR and between 6f80677 and f0d1935.

📒 Files selected for processing (3)

• auth_proxy.py
• host_store.py
• tests/test_metering.py

---

📝 Walkthrough

Walkthrough

Usage aggregation now attributes error and auxiliary rows without model_family to their requested model or route across runtime counters, history-shaped aggregation, and SQL results. A regression test verifies the requested-model bucket and excludes an unknown bucket.

Changes

Usage attribution

Layer / File(s)	Summary
Runtime family bucketing auth_proxy.py	_record_request derives a fallback family key and applies it to global, caller-scoped, and key-scoped model-family counters.
Persistent aggregation and regression coverage auth_proxy.py, host_store.py, tests/test_metering.py	History-shaped and SQL aggregation fall back to requested_model or route; the regression test verifies error-row attribution to gpt-5.5 without an unknown bucket.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: jmlago, muncleuscles

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: attributing error/aux rows to the requested model instead of "unknown".
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/error-model-attribution

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f0d1935977

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Use the fallback key when filtering model buckets

When a window has error/aux rows with model_family NULL and requested_model='gpt-5.5', this new family_k puts gpt-5.5 into filter_options.models; selecting that model sends model=gpt-5.5 back to /dashboard/api/stats, but host_store._usage_where(..., model_family=model) still filters on raw model_family = %s. Those same fallback-attributed rows are then excluded from totals/recent, so drilling into the new bucket shows too few requests or a lower error rate. Please make the model filter use the same fallback expression, or avoid exposing fallback-only keys as filter options.

Useful? React with 👍 / 👎.