Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix(proxy/bedrock): record the real upstream status on InvokeModel outcomes (#4025)
Hill Patel159daa25a088
diff --git a/headroom/proxy/handlers/bedrock.py b/headroom/proxy/handlers/bedrock.pyindex 0a40886c4..24c5d510a 100644--- a/headroom/proxy/handlers/bedrock.py+++ b/headroom/proxy/handlers/bedrock.py@@ -243,6 +243,14 @@ class BedrockHandlerMixin:# contract treats 0 as "not measured") — Bedrock responses are forwarded# byte-faithfully and never parsed. The valuable figure, request-side# compression, is recorded in full.+ #+ # status_code=response.status_code: unlike streaming.py's SDK-based+ # _stream_response_bedrock (which never holds an HTTP response object+ # and so keeps the dataclass default of 200 — see PR #2682), this+ # handler forwards via raw httpx and always has the real upstream+ # status on the returned Response/StreamingResponse. Leaving it out+ # here made every upstream 4xx/5xx (throttling, auth failure, model+ # not found) record as a 200 success in the savings/cost funnel.try:from headroom.proxy.outcome import RequestOutcome@@ -261,6 +269,7 @@ class BedrockHandlerMixin:pipeline_timing=pipeline_timing,tags=tags,client=client,+ status_code=response.status_code,))except Exception:diff --git a/tests/test_proxy/test_bedrock_passthrough.py b/tests/test_proxy/test_bedrock_passthrough.pyindex 45d4ed029..10f23205c 100644--- a/tests/test_proxy/test_bedrock_passthrough.py+++ b/tests/test_proxy/test_bedrock_passthrough.py@@ -450,6 +450,34 @@ def test_outcome_recorded_with_bedrock_provider():assert outcome.tokens_saved == 750[email protected]("upstream_status", [400, 403, 429, 500, 503])+def test_outcome_carries_real_upstream_status_on_error(upstream_status):+ """Regression: handle_bedrock_invoke built RequestOutcome without a+ status_code kwarg, so an upstream throttle/auth/model error recorded as+ the dataclass default of 200 — the failed turn then fed the success+ funnel (savings/cost/request-log) in emit_request_outcome instead of+ being short-circuited by its `status_code >= 400` guard."""+ app = create_app(_make_config())+ with TestClient(app) as client:+ proxy = client.app.state.proxy+ _install_fake_client(+ proxy, _FakeUpstream(status_code=upstream_status, chunks=(b'{"error":true}',))+ )+ proxy._record_request_outcome = AsyncMock()+ proxy.anthropic_pipeline.apply = MagicMock(+ return_value=_FakeResult([{"role": "user", "content": "c"}], 1000, 250)+ )+ resp = client.post(+ INVOKE,+ json={"messages": [{"role": "user", "content": "q" * 3000}], "max_tokens": 8},+ )++ assert resp.status_code == upstream_status+ assert proxy._record_request_outcome.await_count == 1+ outcome = proxy._record_request_outcome.await_args.args[0]+ assert outcome.status_code == upstream_status++# ── env config path ───────────────────────────────────────────────────