feat: track OpenRouter usage

[wjiayis]

Store usage information in database to track user behavior (if user is not BYOK) as a precursor to billing users, e.g.,

Hourly:

  {
    _id: ObjectId('69cd2e9a926e4b5665909027'),
    user_id: ObjectId('6975da46d6096ac1b07342c2'),
    hour_bucket: ISODate('2026-04-01T14:00:00.000Z'),
    project_id: '69746440592738ba2708cc34',
    success_cost: 0.014026150000000001,
    failed_cost: 0.024026150312000001,
    updated_at: ISODate('2026-04-01T14:41:37.465Z')
  }

Weekly:

  {
    _id: ObjectId('69cd2e9a926e4b5665909028'),
    week_bucket: ISODate('2026-03-30T00:00:00.000Z'),
    project_id: '69746440592738ba2708cc34',
    user_id: ObjectId('6975da46d6096ac1b07342c2'),
    success_cost: 0.014026150000000001,
    failed_cost: 0.024026150312000001,
    updated_at: ISODate('2026-04-01T14:41:37.465Z')
  }

Lifetime:

  {
    _id: ObjectId('69cd2e9a926e4b5665909029'),
    user_id: ObjectId('6975da46d6096ac1b07342c2'),
    project_id: '69746440592738ba2708cc34',
    success_cost: 0.014026150000000001,
    failed_cost: 0.024026150312000001,
    updated_at: ISODate('2026-04-01T14:41:37.465Z')
  }

[wjiayis]

For discussion: do we plan to / how do we track token usage for XtraMCP?

[4ndrelim]

For discussion: do we plan to / how do we track token usage for XtraMCP?

That's a good question. With the current approach of tracking by project ID, it makes it difficult to track tokens across the two backend. There are two approaches:

1. Requests to xtramcp, and the responses, go through PD backend via internal/services/toolkit/toolcall_v2.go. We can track tokens coming in and out.
2. Alternatively, since tokens are tracked on the DB, xtramcp can similarly access it and update tokens tracking.

The latter approach will require passing in project ID to xtramcp and there will be some coupling between xtramcp and DB service for tokens tracking. The former might be easier to manage since all tokens tracking is managed in PD, but we might need additional logic / info passed from xtramcp to get accurate token tracking since not all MCP functions rely on LLM generation (for example, citation verifier or relevant papers search). We may end-up over counting.

Maybe we can leave this for now, and focus on the PD tokens tracking first. There will be some major enhancements (incorporation of skills) for xtramcp. It may even replace some of the existing functionalities. Though i havent gotten around into exploring fully yet. Once this is firmed up, let's revisit.

[4ndrelim]

@wjiayis Thanks for taking this up. Let's test this on staging, and propose a meet with the rest again to decide on a launch timeline. We can propose early on but set a later date, thinking it might be apt to do an upgrade to some of the backend LLM capabilities, and we can launch all these together.

[Junyi-99]

@4ndrelim Hi andre, the staging environment is currently unavailable, i will try to fix it on weekends.

[Junyi-99]

Hi @wjiayis, @4ndrelim , the https://app-dev.paperdebugger.com/ and https://app-stg.paperdebugger.com/ are both online.

[4ndrelim]

Let's find a date this week to run through this on staging.

@kah-seng I believe we havent tested yours on staging as well. Should we also pick a date this week? Not sure if its too much to do both at once or if there is anything to de-conflict. If not, we can split.

[wjiayis]

@4ndrelim All comments have been resolved and this branch is ready for merging into staging. I'll merge it after @kah-seng finish testing his changes on staging, for clarity.