fix: echo empty reasoning_content on tool-bearing DeepSeek requests

[jkyberneees]

Problem

DeepSeek thinking mode returns HTTP 400 on tool loops:

llm: deepseek: HTTP 400 (invalid_request_error):
The `reasoning_content` in the thinking mode must be passed back to the API.

It appears at an arbitrary iteration and then recurs on every later request, so one bad turn kills the rest of the session.

Root cause

DeepSeek requires reasoning_content to be replayed on every assistant turn of a request that carries tools — including turns where the provider itself returned no reasoning, which it does on some turns. The assistant message was tagged json:"reasoning_content,omitempty", so such a turn serialized with no key at all.

Enforcement is shape-dependent: a request whose history continues a tool loop (ends with a tool result) is rejected when the key is missing, while a terminal request is accepted. That is why the failure surfaces at an arbitrary iteration rather than a fixed one, and why it then persists.

Fix

• oaMessage.ReasoningContent → *string, so "key absent" and "key present and empty" are both expressible on the wire.
• New Quirks.EchoReasoningWithTools, enabled for deepseek — an explicit provider flag, never URL sniffing.
• When a request carries tools, the key is echoed on every assistant turn, empty value included.

Blast radius

• Applies to every assistant turn of a chat-completions request that carries tools, including turns that made no tool call themselves. It follows the tools, not the thinking preset.
• Without tools, the key is sent only when there is reasoning to replay.
• Providers without the quirk serialize byte-identically to before. zai carries no flag: no z.ai documentation confirms the same requirement.
• A custom provider id inherits no quirks, and WithQuirks replaces the struct rather than merging into it — both documented.

Verification

• Live probe (tag-gated, never in CI): with the same history and request, the pre-fix shape is rejected where this change is accepted. A terminal request accepts both shapes, which is why the echo is keyed on tool-carrying requests rather than applied unconditionally.
• Unit tests: empty echo with tools; key absent without tools; key absent without the quirk; the echo follows the tools rather than the thinking preset; and the built-in registry entry actually puts the echo on the outbound body.
• Dropping the flag from the registry fails the suite — a regression that previously passed unnoticed.
• gofmt, go vet (with and without the e2e tag), golangci-lint, go test -race -count=1, and CI are all green.

Notes

• README.md and AGENTS.md are updated in the same commits, along with the openai.go and responses.go doc comments.
• A consumer pinning a released version of this module receives the fix once a new version is tagged and the dependency is bumped.

[jkyberneees]

Live verdict — the empty echo is the correct layer (controlled experiment)

The tag-gated probe settles the one open premise. Four live calls, thinking enabled, tools advertised, identical history — only the shape and the echo vary:

shape	EchoReasoningWithTools	result
continuation (history ends with a tool result)	false — the pre-fix wire shape	REJECTED — HTTP 400 (invalid_request_error): The 'reasoning_content' in the thinking mode must be passed back to the API.
continuation	true — this PR	ACCEPTED (finish="stop")
terminal (history ends with a user message)	false	ACCEPTED
terminal	true	ACCEPTED

VERDICT [continuation]: present-and-empty is accepted where omission is rejected —
                       the shipped echo is the correct layer. Ship it.
VERDICT [terminal]:     DeepSeek accepts both a present-empty and an omitted key here.

What this proves

1. The reported 400 reproduces exactly — byte-identical error text, on the pre-fix shape, with the continuation history. This is the production failure, reduced to four messages.
2. The fix resolves it. The only difference between the rejected and accepted continuation requests is the echoed key. One variable, opposite outcomes.
3. Enforcement is shape-dependent, which explains the original symptom. The terminal request accepts a missing key; the continuation does not. So a session dies as soon as a turn's reasoning is elided and the next request continues the tool loop — which is why the failure lands at an arbitrary iteration rather than at a fixed one.
4. The earlier "harmless but not load-bearing" reading was an instrument error, not a result. That run ended its history with a user message (terminal), so it never exercised the enforcing path. Recorded here because a false negative that was believed for a while is part of the evidence trail.

This also closes the round-1 hypothesis that omission on older turns must have been tolerated (inferred from the session reaching iteration ~16). It was not tolerated; that session simply had not yet hit a continuation whose history carried an elided turn — and once it did, the 400 recurred on every later request, exactly as reported.

Still true, unchanged

• The echo applies to every assistant turn of a chat-completions request carrying tools. The probe does not contradict that, and no reason has appeared to gate it on model or thinking preset.
• Providers without the quirk are untouched; zai carries no flag (no documentation confirms the requirement).
• Reach: odek's go.work means local development already builds this branch, but CI and released odek still resolve the pinned v0.3.2 — a tag + bump is required before this ships.