expose per-message tokens/cost + a real 'list all base models' surface

[adrianipopescu]

What

Two closely related asks for the assistant subcommand, both blocking downstream tooling on 0.16.0:

1. Re-expose per-message token counts and $ cost in assistant --stream --stream-output json events and in assistant thread get <id> output.
2. Ship a first-class "list all base models available to this account" command that works without pre-existing custom assistants.

Context

I run a small OpenAI-compatible shim in front of kagi assistant --stream --stream-output json so that Bifrost / Open WebUI / anything speaking POST /v1/chat/completions can route through my Kagi Assistant. The shim needs three pieces of data out of the CLI:

• Every call: prompt/completion token counts + upstream $ cost, so the shim can populate the OpenAI usage envelope and export a per-model cost counter for Grafana.
• Startup: the full base-model catalog, so GET /v1/models returns the real slug set (not just ki_quick / ki_research / ki_deep_research).

On 0.8.0 the metadata was in message.metadata_html on the final stream event — parseable, though HTML — so I could reconstruct tokens and cost per call. On 0.16.0 those fields are gone from both --stream-output json and assistant thread get, so downstream callers see usage: {prompt_tokens: 0, completion_tokens: 0, total_tokens: 0} for every reply. I'm currently faking it with ceil(len_utf8 / 4), which is fine for trending but useless for reconciliation.

Repro (tokens)

$ echo hi | kagi assistant --model ki_quick --stream --stream-output json | tail -1 | jq
{
  "meta": {},
  "thread": { ... },
  "message": {
    "id": "...",
    "state": "done",
    "prompt": "hi",
    "reply_html": "...",
    "markdown": "Hi there! ...",
    "documents": [],
    "profile": { "model_name": "ki_quick", "model_version": "ki_quick-2026-05-22" }
  },
  "md_delta": "..."
}

No metadata_html, no usage, no tokens, no cost. assistant thread get <id> is the same story.

Repro (model catalog)

$ kagi assistant models
{"models":[]}

$ kagi assistant custom list
[]

assistant models documents itself as "List Assistant base-model slugs available to custom assistants". On my Ultimate account with no saved custom assistants, it returns an empty list — even though kagi assistant --model claude-4-5-sonnet (etc.) works. There's no other exposed way to enumerate the full model catalog, so shims either hard-code a list that goes stale or scrape the settings HTML.

Auth is fine (kagi auth check passes; assistant thread list returns the account's real threads; prompts stream successfully). This isn't a plan gap, it's a missing surface.

Suggested

Tokens + cost:

• Add usage: {prompt_tokens, completion_tokens, total_tokens, cost_usd} to the final state:"done" stream event and to assistant thread get <id>'s messages. Kagi's own UI shows these numbers per message, so the API is producing them — CLI just needs to surface.
• Even a minimal cost_usd per call would be enough for cost telemetry; token counts are a bonus.

Model catalog:

• kagi assistant models --all (or a rename to kagi assistant base-models) that lists every model slug the account can use with --model, independent of whether custom assistants exist.
• Include labels + optional per-model pricing if the API returns it, so shims don't have to maintain a separate table.

Workaround:

• Shim: hard-coded fallback catalog (30-ish slugs), token estimation from char length, static per-model pricing table, /metrics Prometheus endpoint counting requests + estimated USD per model. Works but drifts as Kagi adds/renames models.

Happy to test a preview build or send a PR if you can point me at the code paths.