feat/220 token counting integration tests

[mkoushni]

test(examples): add token_counting integration tests and example config (#220)

Summary

Adds an end-to-end integration test suite and example configuration for
the token_count filter (merged in #262), covering all five provider
response formats, both streaming and non-streaming delivery, and
response-body passthrough guarantees, as described in
proposal #220.

Rebased onto current main: uses the actual merged filter names
(token_count, token_usage_headers) and header names
(Praxis-Token-Input/Output/Total) rather than the placeholder names
from the original proposal draft (x_token_headers, X-Token-*).

---

What changed

examples/configs/token-counting.yaml

Rewrote the example pipeline to chain
router → token_usage_headers → token_count → access_log → load_balancer,
matching the filters that actually exist on main. Documents that
response headers are committed before the response body phase, so
token_usage_headers (declared before token_count to run after it
in the reverse response-hook order) has nothing to inject for any of
the currently-supported providers (openai, anthropic, google, bedrock,
azure) — they all extract counts from the response body in
on_response_body, after headers have already been sent. Counts remain
available in filter_metadata for other response-phase filters.

tests/integration/tests/suite/examples/token_counting.rs

12 integration tests covering the provider/mode matrix from proposal
#220 that's actually supported by the merged filter:

Test	Provider	What is verified
openai_non_streaming_extracts_token_counts	OpenAI	200 OK + body passthrough
anthropic_non_streaming_extracts_token_counts	Anthropic	200 OK + body passthrough
google_non_streaming_extracts_token_counts	Google	200 OK + body passthrough
bedrock_converse_non_streaming_extracts_token_counts	Bedrock Converse	200 OK + body passthrough
azure_non_streaming_extracts_token_counts	Azure	200 OK + body passthrough
openai_streaming_extracts_token_counts	OpenAI	200 OK + SSE body passthrough
anthropic_streaming_split_events_extracts_token_counts	Anthropic	200 OK + SSE body passthrough (input/output tokens split across message_start/message_delta)
google_streaming_no_done_sentinel_extracts_token_counts	Google	200 OK + SSE body passthrough (no [DONE] sentinel)
non_json_response_body_passes_through_unchanged	OpenAI	Body byte-for-byte identical; no Praxis-Token-* headers
missing_usage_fields_no_token_headers_injected	OpenAI	Praxis-Token-* headers absent when no usage
openai_streaming_whitespace_and_comments	OpenAI	Noisy SSE (keep-alive comments, empty lines, pretty-printed JSON) passes through unchanged
example_config_token_counting_openai	OpenAI	Example config smoke test — config parses and proxy returns 200

Why Praxis-Token-* headers are never asserted present here:
Response headers are committed before the response body filter phase
begins. token_usage_headers runs its on_response hook in reverse
pipeline order — after token_count.on_response but before
token_count.on_response_body. Every provider covered here extracts
counts from the response body inside on_response_body, at which
point headers have already been sent downstream, so
token_usage_headers never has anything to inject. Header-based
extraction (e.g. a future Bedrock InvokeModel provider using
upstream response headers) is the one case where header injection
would be observable; that provider isn't part of this PR and is
tracked separately.

Wiring

• tests/.../examples/mod.rs — adds mod token_counting;
• examples/README.md — adds token-counting.yaml entry

---

Test plan

• make lint passes (clippy, nightly fmt, separators, filter-docs)
• make test passes (full unit + integration + doctest suite)
• All 12 integration tests pass locally

Related

• Proposal: docs/proposals/00220_token-counting-integration-tests.md

[praxis-bot-app]

PR too large: 768 lines added (limit: 750, excludes Cargo files, tests, docs, examples, and benchmarks). Please split into smaller PRs. Add skip/pr-conventions label to override.

[leseb]

please wait for #244 so we can run these tests in CI, thanks!

[praxis-bot-app]

Missing Signed-off-by: a623efd. All commits require sign-off (via git commit --signoff).

#244 merged

[praxis-bot]

Review: Token Counting Integration Tests (#243)

Adds integration tests and example config for the token counting pipeline. The test matrix is well structured and the SSE/non-streaming coverage across providers is thorough.

Findings

#	Severity	File	Description
1	Critical	mod.rs	Duplicate mod token_counting; — will not compile
2	Large	token_counting.rs	Missing negative test: Bedrock InvokeModel with absent upstream token-count headers
3	Medium	token_counting.rs	Doc comment on test function violates project convention
4	Medium	token_counting.rs	Separator says "Helpers" — convention requires "Test Utilities"

[praxis-bot-app]

Unsigned commits: 6fae4bc. Please sign your commits.