fix: negotiate unsupported temperature capability on current orchestration line

[seonghobae]

Summary

• retry the same chat or raw Responses endpoint once when HTTP 400/422 explicitly reports that temperature is unsupported
• retain the negotiated omission across later transient retries instead of restoring the rejected field
• preserve the provider 4xx body for downstream diagnostics and keep invalid-value/unrelated 4xx responses fail-closed
• add ADR 0012, real local-HTTP transport coverage, and the exact Azure default-only regression

Evidence

• exact base: contextual-orchestrator#765 current head d3f9a9b96523ed572b908c8abba1afa527eb49dc
• exact head: cf4a4501fa5057f89b21cad5033c5925755cd150
• focused transport and sampling tests: 15 passed in 4.59s
• full suite on the refactor: publish gateway-only paper-grounded reasoning contract #765 integration line: 1472 passed in 546.95s
• python -m compileall -q contextual_orchestrator
• git diff --check

This is a successor to #774. It intentionally does not retarget or force-push #774, whose original base is a closed integration branch and therefore has no hosted required Checks.

---

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a2db003c-e864-48b1-b1d6-9aee6eeb7427

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

Current-head review at beb15a7d77d3d47b8de711928faf76ab5d4f7ef7:

• The added regressions preserve unrelated 4xx bodies, reject invalid caller temperatures as non-negotiation, and keep temperature omitted across subsequent transient retries after a provider capability rejection.
• No additional source defect found in the current exact diff. The PR remains stacked on refactor: publish gateway-only paper-grounded reasoning contract #765 with no independent required Checks until its base is integrated.

[seonghobae]

@OpenCode review exact HEAD cf4a450. Verify the Azure default-only temperature diagnostic, 4xx body preservation, and that transient retries retain the negotiated omission. Full local suite: 1472 passed.

[seonghobae]

Live exact-head runtime evidence for cf4a450: the authenticated provider-neutral gateway returned HTTP 200 with one choice, and a direct validated transport call pinned to gpt-5.6-sol with temperature 0.2 completed successfully. Synthetic prompt only; response content was not persisted.

[seonghobae]

[seonghobae]

@opencode-agent Please publish a formal review bound to exact head cf4a4501fa5057f89b21cad5033c5925755cd150. Verify the Azure default-only temperature diagnostic, preservation of unrelated 4xx bodies, and persistence of negotiated temperature omission through later transient retries. Exact local evidence remains 1,472 passed plus an authenticated provider-neutral gpt-5.6-sol synthetic transport success with caller temperature 0.2. This PR is stacked on #765; do not approve stale heads or bypass its base gate.

[seonghobae]

@devin review exact current HEAD 5a6f563. The valid ADR collision finding is fixed: optional sampling capability negotiation is now ADR 0016, while the accepted gateway-only provider contract remains ADR 0012; references were re-scanned and protocol regression tests pass (21 passed). Revalidate this exact HEAD and report only current findings.

[seonghobae]

@devin review exact current HEAD dbf46e5. Fixed the current finding: Responses tool-loop passthrough now runs only after non-empty input and stream validation, so stream=true returns invalid_stream instead of a silent non-streamed 200; missing input returns invalid_input. Added HTTP regressions; focused suite 19 passed with Ruff/diff clean. Revalidate this exact head.

[seonghobae]

@devin review exact current HEAD ef7b1fb. Current remote integration includes the parent stack, #787, the Responses tool-loop stream/input fix, and total inbound-body deadline enforcement. Exact-head focused integration tests pass (38 passed) with changed-file Ruff/diff clean. Revalidate this full SHA only and report any current finding.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🟡 Fully received requests can be rejected as timed out

A complete request body is rejected as timed out (raise RequestError(408,...) at contextual_orchestrator/server.py:5999-6001) after the final chunk that completes the body is read, so a valid, fully-received request is thrown away as an error.
Impact: On a slow connection whose last byte arrives right around the deadline, a request whose entire body was already received is returned a 408 error instead of being processed.

Deadline re-check runs even once the body is complete

The loop condition is while len(chunks) < body_size (contextual_orchestrator/server.py:5990). The deadline check at the top of the loop (5991-5993) correctly only fires when more data is still needed. However, the second check added after chunks.extend(chunk) (5999-6001) runs unconditionally after each read — including the read that completes the body. If time.monotonic() >= read_deadline at that moment, a 408 is raised even though len(chunks) == body_size and the full body is available in chunks. The top-of-loop check already covers the incomplete case on the next iteration, so the post-extend check only adds the spurious rejection of an already-complete body. Guarding it with len(chunks) < body_size avoids discarding a fully-read request.

Suggested change

	if time.monotonic() >= read_deadline:
	self.close_connection = True
	raise RequestError(408, "request_read_timeout", "request body read timed out")
	if len(chunks) < body_size and time.monotonic() >= read_deadline:
	self.close_connection = True
	raise RequestError(408, "request_read_timeout", "request body read timed out")

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[seonghobae]

Current exact HEAD: 5774e1d2e878bf111de3229c3468d2071908ec81.

• Stacked base was merged into the PR head with a normal merge commit; no force-push was used.
• Root cause fixed: a slow byte trickle could extend inbound request reads beyond the configured total deadline; the shared read path now checks the monotonic deadline before and after each chunk.
• Exact-head local proof: 1501 passed in 587.79s; changed test/security regression proof 76 passed in 29.63s; Ruff, compileall, and diff checks pass.
• Current hosted checks: CodeRabbit pass; Devin pending.

@devin-ai-integration please revalidate only this exact HEAD. Normal auto-squash merge is armed; no approval or bypass is being used.