fix(token-spy): rebuild headers for decoded responses

[tang-vu]

Why this matters

Token Spy returns resp.content, which HTTPX automatically decompresses, but copied the upstream headers unchanged. A gzip response therefore reached clients as decoded bytes still labeled Content-Encoding: gzip and with the compressed Content-Length; real HTTPX clients fail with DecodingError, and other clients can truncate or double-decode the payload.

Root cause: representation and hop-by-hop headers were reused after the response representation had changed. HTTPX documents that Response.content automatically decodes gzip/deflate while raw streaming is required to preserve wire encoding: https://www.python-httpx.org/quickstart/#binary-response-content

Behavioral invariant: buffered proxy responses describe the body actually sent downstream, while safe end-to-end metadata remains intact.

Overlap check

Searched open and closed upstream PRs for Token Spy content-encoding/content-length/decoded-response/header passthrough terms and inspected current PRs touching ods/extensions/services/token-spy/main.py. Existing changes cover streaming status/error logging, SSE cursors, pricing, persistence, and request auth; none rebuild response headers after HTTPX decoding. This scope is independent.

What changed

• Added one response-header filter for decoded, buffered proxy bodies.
• Removed stale Content-Encoding/Content-Length and standard hop-by-hop fields.
• Applied the contract to Anthropic non-streaming, OpenAI-compatible non-streaming, and catch-all passthrough responses.
• Preserved end-to-end headers such as content type and upstream request IDs; Starlette now emits the correct downstream content length.

Validation

• Red before fix: all three public proxy paths failed at the client boundary; HTTPX attempted a second gzip decode and raised DecodingError.
• pytest tests/test_decoded_response_headers.py -q — 3 passed.
• pytest tests -q — 23 passed, 1 skipped.
• python -m py_compile main.py tests/test_decoded_response_headers.py — passed.
• git diff --check — passed.

Tradeoffs and rollback

This keeps buffered proxy behavior; it does not convert large non-streaming responses to raw streaming. Headers tied to a specific hop or encoded wire body are intentionally not forwarded. Rollback is one commit with no persisted-state impact.

Batch compatibility

Validated as an independent ten-PR batch from upstream main 6ff9b4fc5190099705043acaab7e9b6ad9c8b8f1. The final PR heads merged without conflicts in this order: #2964 -> #2965 -> #2967 -> #2969 -> #2970 -> #2971 -> #2972 -> #2973 -> #2974 -> #2975. The resulting local synthetic merge head is 4ae60eadad9696a9accb735aad71af87d2be802d.

Combined validation on that exact tree:

• Dashboard API boundary suites: 323 passed, 4 skipped.
• Token Spy suite: 36 passed, 1 skipped.
• Privacy Shield suite: 55 passed.
• APE suite: 47 passed.
• Python compile checks and git diff --check: passed.

The scopes are behaviorally independent. The stated order is the tested rollback/merge sequence for shared-file changes; each PR remains individually useful and revertible.