fix(token_count): add OpenAI Responses API usage extraction

[yossiovadia]

Summary

• Fix token_count filter to extract usage from OpenAI Responses API streaming events
• Without this fix, all Responses API clients (Codex CLI, OpenAI SDK with /v1/responses) report 0 tokens

Fixes #713

Root cause

parse_openai only handled Chat Completions format (usage.prompt_tokens). The Responses API puts usage in response.completed events with a different shape:

{"type": "response.completed", "response": {"usage": {"input_tokens": N, "output_tokens": M}}}

Two differences vs Chat Completions:

1. Usage nested under response.usage, not top-level usage
2. Field names input_tokens/output_tokens, not prompt_tokens/completion_tokens

Fix

Extend parse_openai to try the Responses API format as a fallback. 3 new structs, one fallback branch. The existing SSE streaming path (try_complete_usage) already calls parse_openai on each SSE data payload — no changes needed in the streaming handler.

What's included

• filters/src/token_usage/providers.rs — 3 new Responses API structs + fallback in parse_openai
• 3 new tests: response.completed event, missing total, null response

Test plan

• 3 new tests for Responses API format
• All existing provider parser tests pass unchanged (39 total)
• cargo clippy -p praxis-ai-filters -- -D warnings — zero warnings
• Validated end-to-end: Codex CLI → Praxis → metering dashboard shows real token counts

[praxis-bot]

Review Summary

Clean, well-scoped fix for streaming Responses API token extraction. The fallback approach in parse_openai is correct for SSE response.completed events and the test coverage for that path is solid.

One gap: non-streaming Responses API responses are silently missed (see inline comment). 1 medium finding.

---

Automated review by praxis-bot

[praxis-bot]

[Medium] Non-streaming Responses API responses silently return 0 tokens. When stream: false, the /v1/responses endpoint returns usage at the top level with Responses API field names:

{"id": "resp_123", "object": "response", "usage": {"input_tokens": 150, "output_tokens": 42, "total_tokens": 192}}

Neither parse path matches this:

• OpenAiResponse fails because OpenAiUsage requires prompt_tokens/completion_tokens, which are absent.
• ResponsesApiEvent succeeds but response is None (no response key at the top level).

Add a third fallback using a struct with top-level usage that expects input_tokens/output_tokens:

#[derive(Deserialize)]
struct ResponsesApiDirectResponse {
    usage: Option<ResponsesApiUsage>,
}

Then after the ResponsesApiEvent branch:

if let Ok(resp) = serde_json::from_slice::<ResponsesApiDirectResponse>(body)
    && let Some(usage) = resp.usage
{
    return Some(TokenUsage::new(
        usage.input_tokens,
        usage.output_tokens,
        usage.total_tokens,
    ));
}

This reuses ResponsesApiUsage and covers both streaming and non-streaming Responses API paths.