[FEATURE] Client-side token estimate fallback when a provider omits usage stats (Ollama Cloud kimi-k3:cloud shows 0 tokens)

Summary

When a provider's response omits the usage block entirely, OpenCode records 0 tokens / 0% / $0.00 for the session and shows it in the TUI — with no client-side fallback estimate. New / minimal providers that don't yet emit usage stats are rendered invisible in the token counter and context-% meter.

OpenCode version

1.18.7 (TUI, Linux)

Reproduction

Provider: Ollama Cloud (ollama provider), model kimi-k3:cloud (released ~1 hour before this report).

1. Start a session on ollama/kimi-k3:cloud and run a few turns (e.g. a /start bootstrap that triggers several model generations).
2. The TUI Context panel shows 0 tokens, 0% used, $0.00 spent regardless of how much was generated.
3. Inspect the session DB:

sqlite3 -readonly ~/.local/share/opencode/opencode.db \
  "SELECT substr(id,1,16), model, tokens_input, tokens_output, cost
   FROM session ORDER BY time_created DESC LIMIT 6"

Actual (one kimi-k3 session that completed 19 real assistant turns):

ses_05bc182d5ffe | {"id":"kimi-k3:cloud","providerID":"ollama"} | 0  | 0  | 0.0
ses_05bbb9584ffe | {"id":"glm-5.2:cloud","providerID":"ollama"} | 3271199 | 21691 | 0.0

At the message level the per-turn tokens object is {"total":0,"input":0,"output":0,"reasoning":0,"cache":{"write":0,"read":0}} — i.e. no usage was populated, not even a partial one. The model responded normally (19 assistant messages with content were persisted); only the usage telemetry is missing.

Contrast (same provider, same OpenCode build, different model)

ollama/glm-5.2:cloud over the same ollama provider records full usage correctly (millions of tokens across many sessions). So this is not a provider-adapter parse regression in general — it's specific to a model whose response omits the usage block. kimi-k3:cloud is brand-new; Ollama Cloud appears to ship new models before wiring usage reporting on the server side.

The gap

OpenCode records exactly what the provider returns. When the usage block is absent, every field defaults to 0 and the TUI displays a misleading 0% / 0 tokens, with no indication that the provider simply didn't report usage. (Cost being $0.00 is expected for subscription providers like Ollama Cloud; the token/percent zeros are the problem.)

Proposal

When a provider returns no usage stats for a turn, fall back to a client-side estimate so the context-% meter and token counter keep working:

• Estimate input tokens from the prompt via a tokenizer (or a cheap heuristic — chars/4). Output tokens can be counted from the streamed completion text directly (this is exact, no tokenizer needed).
• Mark the estimate distinctly in the UI (e.g. an ~ prefix or a tooltip "estimated — provider did not report usage") so it's not confused with provider-reported numbers.
• Alternatively (lower-effort): when usage is absent, show "usage not reported by provider" instead of 0, so users aren't misled into thinking nothing happened.

Related

• #39092 — same symptom (tokens frozen at 0, $0, 0%) but on Desktop with the Synthetic provider; the reporter's own hypothesis is "the provider may not return token usage metadata." Same root gap, different provider — confirms this is a general cross-provider issue, not kimi-k3-specific.
• #34022 — OpenAI-compatible providers report zero tokens_cache_read despite upstream returning cached_tokens (a parsing gap on the openai-compatible path). Adjacent: both are about OpenCode's usage-tracking fidelity for openai-compatible/Ollama-style providers.