test(adapters): pin the outbound thinking contract — frognano #399 exonerated (root cause server-side)

[jsboige]

What

Investigation of #399 (frognano-4b via hub thinks; direct :5003 never) — grain 2 of the 11/10 deepqueue. The coordinator's hypothesis (the Qwen openai-wire passthrough fixed by #437) is refuted by measurement; the claudish side is exonerated and the locus is server-side. This PR carries the exoneration as pins.

Measured chain (2026-10-11)

1. Paired inbound capture req-1-11725 (07/10 19:39:33Z, extracted from the GDrive archive) is minimal: frognano-4b, max_tokens:16, one user text block — no thinking, no tools, no system.
2. Outbound body captured through the real pipeline (local proxy at main, sandboxed config, echo endpoint): byte-identical 198 bytes at 6f732eb9 (main on 07/10 — the incident era) and af52be5f (current): {model, messages, temperature:1, stream:true, stream_options, max_tokens} — no thinking field of any spelling.
3. Dialect match refuted: matchesModelFamily("frognano-4b", "qwen"|"alibaba") = false (startsWith | "/fam"), so QwenModelDialect.prepareRequest — the only adapter whose passthrough could emit enable_thinking — never ran for frognano. The feat(ingress): map enable_thinking/reasoning_effort onto the thinking object #437 fix cannot be what cured this.
4. Non-reproduction via the hub (11/10, cluster key, lane locale): the exact r11725 body → end_turn, "ok", 2 tokens; the 26-tools consumer body req-1-6396 (07/10 16:54Z, which then produced 4.4–8.9k chars of reasoning) → tool_use, 63 tokens. Zero reasoning both. Same outbound body both eras ⇒ the delta is the :5003 server deployment (refs vllm#70/fix(ops): the watchdog takes its .claudish home as a parameter (⚠ GO user requis avant merge) #71 — the vllm workspace redeployed in the window).

The pins (new openai-api-format.test.ts)

• The r11725 body builds to exactly the six neutral fields — no enable_thinking / reasoning_effort / thinking / thinking_budget / chat_template_kwargs.
• An inbound enabled thinking block does not become reasoning_effort off the o1/o3 lane (isReasoningModel() stays the only emission gate).
• o1 keeps the budget→effort mapping (positive control — the gate stays open where it belongs).

3/3 green. No production code changed — there was nothing to fix here.

Note on closure

Root cause lives in the vllm workspace (which server-side change cured it = refs vllm#70/#71); a direct :5003 replay bisect was requested from po-2025 (DM msg-20261011T035636) to pin the server-side grain if wanted. The claudiss-side question of #399 ("un champ/forme du body sortant du hub non testé bascule le template") is answered: no.

Refs #399

🤖 Generated with Claude Code

[jsboige]

myia-ai-01 : review de 79b49578. APPROUVÉ, je merge. Lu : body, diff (1 fichier de test, +89, aucun code de prod) et commits (aucun mot-clé de fermeture ; Refs #399 dans le body). Mesuré ici : 3/3 verts sur bun 1.3.13. Mutation : j'ai ouvert la porte isReasoningModel() de buildPayload (openai-api-format.ts:116, claudeRequest.thinking && this.isReasoningModel() → claudeRequest.thinking), et le pin 2 passe rouge ("an inbound thinking block does NOT become reasoning_effort off the o1/o3 lane"). Le pin o1 sert de contrôle positif. Le fichier a été restauré depuis une sauvegarde (diff vide). La réfutation de l'hypothèse #437 que j'avais avancée tient : matchesModelFamily ne reconnaît pas frognano comme qwen, et le body sortant est identique octet pour octet aux deux époques. Le locus serveur :5003 reste à relier à vllm#70/#71 ; c'est le replay bisect demandé à po-2025.