fix(chat): preserve suffix reasoning intent across model attempts

changelog.d/fixes/13720-suffix-effort-propagation.md

@@ -0,0 +1 @@
+- fix(chat): preserve suffix-model reasoning effort across model attempts so a replacement model no longer inherits or drops the original suffix, and keep explicit reasoning choices in request dedup hashes (#13720)

config/quality/file-size-baseline.json

@@ -357,8 +357,10 @@
   "_rebaseline_2026_07_27_3850_relax_filesize_cap": "OWNER-APPROVED TEMPORARY relax for v3.8.50-3.8.54 PREPARE phase (docs/ROADMAP.md). cap 800->900 (+100), testCap 800->900 (+100). Targets: decompose-existing-frozen unchanged (frozen still only-shrink); this only relaxes the cap for NEW files in the decompose/extract-while-PREPARE phase (.51='executor registry in-place' and .52='combo.ts decomposition' create new leaf modules above 800). RE-TIGHTENING MANDATORY in v3.8.51: cap target 850 = 850 once decomposition wave stabilizes. SUPERSEDED by _rebaseline_2026_07_27_3850_relax_filesize_cap_v2_20pct (v1 +20% buffer) — retained for audit. Tracked via same roadmap issue.",
   "_rebaseline_2026_07_27_v3849_train1h": "Merge-train 1H (31 PRs) — owner-approved 2026-07-27. Two distinct causes, kept separate on purpose: (1) GENUINE irreducible growth at existing chokepoints — providerLimits/auth (#8632 Kimi quota-reset recovery), rateLimitManager (#8616 idle wedged limiters), models-catalog-route.test (#8610 OpenCode Go effort aliases); (2) COLLISION with #8585, which banked shrinks measured on the pre-train release tip while 30 sibling PRs in the SAME train grew those files again — chat/accountFallback (#8628), chatCore (#8613), videoGeneration (#8581), imageGeneration. The zero-headroom frozen entries cannot absorb either. Ceilings re-pinned to the post-merge tip; #8612 (also in this train) automates shrink-banking so this self-inflicted drift stops recurring. Detail: src/lib/usage/providerLimits.ts 1006->1013 (#8632); src/sse/services/auth.ts 2492->2508 (#8632); open-sse/services/rateLimitManager.ts 1014->1060 (#8616); src/sse/handlers/chat.ts 1842->1845 (#8628); open-sse/handlers/chatCore.ts 4939->4955 (#8613); open-sse/handlers/imageGeneration.ts 3100->3101 ((sem PR — teto do #8585)); open-sse/handlers/videoGeneration.ts 1038->1063 (#8581); open-sse/services/accountFallback.ts 1965->1966 (#8628); tests/unit/models-catalog-route.test.ts 1608->1636 (#8610)",
   "frozen": {
+    "src/sse/handlers/chatHelpers.ts": 1253,
+    "_rebaseline_2026_09_17_13720_merge_release_v3851": "Merge de release/v3.8.51 na #13720 (2026-09-17). src/sse/handlers/chatHelpers.ts 1246 -> 1253, decomposto: 1246 -> 1250 e crescimento INHERITED do tip (base-red ja presente em origin/release/v3.8.51 no commit 9688032451fc, arquivo com 1250 linhas contra cap 1246 — nao e desta PR e nao foi introduzido por este merge); 1250 -> 1253 sao as MESMAS +3 linhas da propria #13720 ja auditadas e aprovadas pelo dono na entrada _rebaseline_2026_09_16_13720_suffix_effort_propagation abaixo (threading de resolvedThinkingEffort). Nenhum outro teto foi tocado por este merge; tests/unit/chatcore-translation-paths.test.ts (3449 > 3447) permanece vermelho de proposito — e base-red herdado e a PR nao toca o arquivo.",
+    "_rebaseline_2026_09_16_13720_suffix_effort_propagation": "OWNER-APPROVED 2026-09-16 (explicit exception for this unit only, chatHelpers.ts only). PR #13720 (HouMinXi, suffix-effort propagation across model attempts): src/sse/handlers/chatHelpers.ts merge-base (before PR's own commit) was 1164; the release tip independently grew it to 1213 (+49, unrelated merged PRs) while the frozen cap sat at 1214 to cover exactly that tip growth. The PR's own diff on this file is +3 lines only (threading resolvedThinkingEffort: one field on resolveModelOrError's return object, one destructured param and one passthrough call-site argument in executeChatWithBreaker — see commit 4fdb0c5851b7f645efbe5627cd0babb0c3d230c3), taking the merged result to 1216 (1217 per check-file-size.mjs's countLines, which counts the trailing newline as an extra split segment). All 3 added lines are single-property additions inside existing multi-line object literals/signatures; there is no redundant or duplicated line in the PR's own hunks to trim, and none of the +3 lines are outside the PR's own diff. Covered by tests/unit/suffix-effort-propagation.test.ts (27/27), tests/unit/chatcore-upstream-body.test.ts + tests/unit/request-dedup-tenant-isolation.test.ts (57/57), all green against this exact head.",
     "_rebaseline_2026_09_17_13947_tip_growth": "Base-red drain da PR #13947 (Refs #13866) — crescimento de PRODUCAO que chegou pelo tip e nunca foi rebaselinado; nenhum destes arquivos e tocado por esta PR. #12906 (d70f43d4, retry empty_response 502 + timeout de inicio de resposta ciente de reasoning): src/sse/handlers/chat.ts 2498->2500, src/sse/handlers/chatHelpers.ts 1231->1245, open-sse/utils/proxyFetch.ts 1275->1276, open-sse/utils/stream.ts 3098->3123. #12904 (f3acf4f8, injecao unica do system prompt global pos-traducao) + #12910 (051576fd, finalizacao de cache semantico por request id exato): open-sse/handlers/chatCore.ts 6181->6203. Anteriores ao lote, ja acima do cap na base 3d5baf13: open-sse/handlers/imageGeneration.ts 3293->3304 (#13748, b97338a8) e open-sse/services/combo/roundRobinCombo.ts 1213->1221 (#13776, aeba6b1a). Registrado contra o estado mergeado; nenhum outro cap e tocado.",
-    "src/sse/handlers/chatHelpers.ts": 1246,
     "_rebaseline_2026_09_15_13609_mistral_ambiguous_401": "PR #13609 rework (maxmad64bis, bare Mistral 401 soft lockout behind MISTRAL_AMBIGUOUS_401_SOFT_LOCKOUT, default off). open-sse/services/accountFallback.ts 2469->2501 (+32): +14 are the change itself (shared-predicate + flag imports, the documented ambiguousAuth field on the checkFallbackError return type, and the flag-gated 401 branch formatted normally instead of the PR's 139-char squeezed configuredRule line); +18 are the lint-staged prettier pass normalizing lines that were already unformatted on the release tip (multi-import, ISO_RETRY_RE, two regex arrays, persistAntigravityFamilyCooldownIfQuota call, applyErrorState guard, trailing commas) — pure formatting, no logic. src/sse/services/auth.ts 3556->3557 (+1): markAccountUnavailable passes connectionId to resolveTerminalConnectionStatus so the soft-strike bound is per connection. The predicate and strike tracker live in the leaf open-sse/services/accountFallback/mistralAmbiguousAuth.ts (under cap). Covered by tests/unit/provider-401-ambiguous-runtime.test.ts (flag off/on, end-to-end through markAccountUnavailable).",
     "_rebaseline_2026_06_22_4644_deepseek_web_tools": "PR #4644 (BugsBag/robust deepseek-web tool-call parsing): open-sse/executors/deepseek-web.ts 1117->1125 (+8). The new agentic tool-call path emits surrounding text + reasoning before tool_calls and swaps to the dedicated deepseekWebTools.ts parser; the +8 lines are cohesive wiring at the existing transformSSE chokepoint (the parser itself lives in the new deepseekWebTools.ts file, already under cap). The PR's own fast-gate (PR->release) does not run check:file-size, so this surfaced only at release reconcile. Covered by tests/unit/deepseek-web-tools-variants.test.ts + deepseek-web-tools-execute.test.ts.",
     "_rebaseline_2026_06_23_4712_deepseek_web_tool_results": "PR for #4712 (deepseek-web drops role:tool): open-sse/executors/deepseek-web.ts 1125->1148 (+23). messagesToPrompt() now folds role:\"tool\" results into the single-prompt transcript (recovering the tool name from the preceding assistant tool_calls by tool_call_id) instead of silently dropping them; the lines are cohesive wiring inside the existing function. Covered by tests/unit/deepseek-web-tool-result-prompt-4712.test.ts.",

open-sse/handlers/chatCore.ts

@@ -165,26 +165,9 @@ import {
   getStripTypesForProviderModel,
   stripIncompatibleMessageContent,
 } from "../services/modelStrip.ts";
-import { normalizeMimoThinking } from "../services/mimoThinking.ts";
-import {
-  isOpencodeGoProvider,
-  stripBooleanReasoning,
-} from "../services/opencodeReasoningSanitizer.ts";
-import {
-  normalizeClaudeAdaptiveThinking,
-  normalizeClaudeDisabledThinkingEffort,
-} from "../services/claudeAdaptiveThinking.ts";
 import { shouldUseMidConversationSystem } from "../executors/claudeIdentity.ts";
-import { normalizeClaudeHaikuConstraints } from "../services/claudeHaikuConstraints.ts";
-import { applyDefaultReasoningEffort } from "../services/defaultReasoningEffort.ts";
-import { wireAdaptiveEffort } from "./chatCore/adaptiveEffortWiring.ts";
 import { echoModelInObject } from "../services/responseModelEcho.ts";
-import {
-  stripGpt5SamplingWhenReasoning,
-  stripGpt5ReasoningWhenTools,
-} from "../services/gpt5SamplingGuard.ts";
 import { getUnsupportedParams, REGISTRY } from "../config/providerRegistry.ts";
-import { stripUnsupportedParams } from "./chatCore/unsupportedParamsStrip.ts";
 import { checkToolCallingRequiredButUnsupported } from "./chatCore/toolCallingRequiredCheck.ts";
 import {
   supportsMaxTokens,
@@ -210,7 +193,6 @@ import {
   isTinyBudgetReasoningProbe,
   toPositiveInteger,
 } from "../services/reasoningTokenBuffer.ts";
-import { normalizeThinkingForModel } from "@/shared/constants/modelSpecs.ts";
 import {
   buildErrorBody,
   createErrorResult,
@@ -512,6 +494,19 @@ export async function handleChatCore({
   videoBridgeLog = undefined,
   fallbackAttempts = undefined,
 }) {
+  const {
+    model: originModel,
+    resolvedThinkingEffort,
+    defaultThinkingEffort,
+  } = modelInfo as typeof modelInfo & {
+    resolvedThinkingEffort?: string | null;
+    defaultThinkingEffort?: string | null;
+  };
+  const trustedEffortContext = Object.freeze({
+    originModel,
+    resolvedThinkingEffort,
+    defaultThinkingEffort,
+  });
   let { provider, model, extendedContext } = modelInfo;
   // Keep the selected rule across format conversion, retries and refreshed credentials.
   // Each combo leg gets its own execution context; nothing is written to shared accounts.
@@ -2734,77 +2729,6 @@ export async function handleChatCore({
   }
   translatedBody.model = finalModelToUpstream;
 
-  // #3554: a combo/route may substitute the upstream model AFTER the client chose its
-  // `thinking` value. Claude Code sends `thinking:{type:"disabled"}` for internal calls,
-  // which claude-fable-5 (adaptive-only) rejects with a 400. Drop the now-invalid value
-  // when the resolved target model rejects it; models that accept `disabled` are untouched.
-  if (typeof finalModelToUpstream === "string") {
-    translatedBody = normalizeThinkingForModel(translatedBody, finalModelToUpstream);
-    // Claude Opus 4.7+/Fable 5 removed manual extended thinking: `thinking.type:"enabled"`
-    // or any `thinking.budget_tokens` is a hard 400. Collapse any manual thinking that
-    // reached this point (passthrough legacy shape, reasoning_effort buckets, per-model
-    // defaults) to `{type:"adaptive"}` — effort stays on `output_config.effort`. Keyed on
-    // the resolved upstream model, so it covers every routing mode. See claudeAdaptiveThinking.ts.
-    translatedBody = normalizeClaudeAdaptiveThinking(translatedBody, finalModelToUpstream);
-    // Opus 5 allows disabled thinking only through high effort on Anthropic's direct
-    // Messages API. The helper scopes this constraint to `anthropic` and `claude`;
-    // GitHub Copilot and Claude Web use separate upstream contracts.
-    translatedBody = normalizeClaudeDisabledThinkingEffort(
-      translatedBody,
-      finalModelToUpstream,
-      provider
-    );
-    // Claude Haiku rejects `thinking.type:"adaptive"` and `output_config.effort`
-    // (both Sonnet 4.6 / Opus 4.5+ only). Several paths can still emit those
-    // shapes on a Haiku target — native passthrough, reasoning_effort buckets,
-    // per-model defaults — so collapse them to a Haiku-valid shape here, after
-    // model substitution. Mirrors upstream 9router 401d93bd5. See
-    // services/claudeHaikuConstraints.ts.
-    translatedBody = normalizeClaudeHaikuConstraints(translatedBody, finalModelToUpstream);
-    // #6879: per-model default reasoning_effort, injected only when the request
-    // carries no reasoning field of any shape — an explicit client/combo-leg value
-    // always wins. Scoped to the OpenAI Chat Completions dispatch shape (the shape
-    // `reasoning_effort` is native to); unset ModelSpec.defaultReasoningEffort is a
-    // no-op. #7694: `modelInfo.resolvedThinkingEffort` — set when the request's model
-    // id carried a `<prefix>/<model>-{effort}` synced-model alias suffix
-    // (`src/sse/services/model.ts`) — takes priority over the static per-model default.
-    // The synced catalog's vendor-declared `defaultThinkingEffort` (OpenRouter
-    // `reasoning.default_effort`, captured by `detectDefaultThinkingEffort`) is the
-    // lowest-priority default: it only fires when neither the suffix alias nor a
-    // static operator default exists. See open-sse/services/defaultReasoningEffort.ts.
-    if (targetFormat === FORMATS.OPENAI) {
-      translatedBody = applyDefaultReasoningEffort(
-        translatedBody,
-        finalModelToUpstream,
-        (modelInfo as { resolvedThinkingEffort?: string })?.resolvedThinkingEffort,
-        (modelInfo as { defaultThinkingEffort?: string })?.defaultThinkingEffort
-      );
-    }
-    translatedBody = wireAdaptiveEffort(translatedBody, {
-      rawBody: body,
-      clientRawRequest,
-      targetFormat,
-    });
-  }
-
-  // Xiaomi MiMo controls reasoning ONLY via `thinking:{type:"enabled"|"disabled"}` and
-  // rejects unknown/extra params with a strict "400 Param Incorrect". Map OmniRoute's
-  // OpenAI reasoning signals onto that native shape: reduce any thinking object to
-  // `{type}` and drop `reasoning_effort`/`reasoning`. See services/mimoThinking.ts.
-  if (provider === "xiaomi-mimo") {
-    translatedBody = normalizeMimoThinking(translatedBody);
-  }
-
-  // opencode-go backed providers (ollama-cloud, opencode-go, opencode,
-  // opencode-zen) use a Go ChatCompletionRequest struct where `reasoning`
-  // is typed as openai.Reasoning (a structured type). A boolean
-  // `reasoning: true/false` — valid per the OpenAI API — causes a 400
-  // "json: cannot unmarshal bool into Go struct field" on the Go side.
-  // Strip the boolean before forwarding. See opencodeReasoningSanitizer.ts.
-  if (isOpencodeGoProvider(provider)) {
-    translatedBody = stripBooleanReasoning(translatedBody);
-  }
-
   const previousResponseIdPolicy = applyResponsesPreviousResponseIdPolicy(translatedBody, {
     mode: settings.responsesPreviousResponseIdMode,
     provider,
@@ -2862,43 +2786,6 @@ export async function handleChatCore({
     return createErrorResult(400, toolCallingCheck.message!, null, "tool_calling_not_supported");
   }
 
-  if (unsupported.length > 0) {
-    const { strippedParams } = stripUnsupportedParams(translatedBody, unsupported);
-    if (strippedParams.length > 0) {
-      log?.warn?.(
-        "PARAMS",
-        `Stripped unsupported params for ${model}: ${strippedParams.join(", ")}`
-      );
-    }
-  }
-
-  // GPT-5 reasoning models (openai Chat Completions) reject temperature/top_p with a 400
-  // whenever a reasoning effort is active, yet accept them under reasoning_effort=none (the
-  // GPT-5.1+ default). A static unsupportedParams list can't express that, so strip sampling
-  // conditionally here. The codex Responses path is already covered by the executor allowlist.
-  translatedBody = stripGpt5SamplingWhenReasoning(
-    translatedBody,
-    provider,
-    finalModelToUpstream,
-    log
-  );
-
-  // GPT-5.x reasoning models on the raw openai Chat Completions surface reject function
-  // `tools` combined with an active `reasoning_effort`: HTTP 400 "Function tools with
-  // reasoning_effort are not supported ... Please use /v1/responses instead." This used to
-  // be true for every GPT-5.x model on the plain `openai` provider, but #7242 (targetFormat
-  // "openai-responses" on GPT_5_6_API_CAPABILITIES) now routes the GPT-5.6 family to
-  // /v1/responses instead, which accepts tools + reasoning natively — so the strip must not
-  // fire there. Pass the already-resolved `targetFormat` so the guard gates on the actual
-  // upstream surface for this request instead of a model-name list. Port of 9router#2540.
-  translatedBody = stripGpt5ReasoningWhenTools(
-    translatedBody,
-    provider,
-    finalModelToUpstream,
-    targetFormat,
-    log
-  );
-
   // Rename max_tokens to max_completion_tokens if not supported (#1961)
   if (!supportsMaxTokens({ provider, model })) {
     if (translatedBody.max_tokens !== undefined) {
@@ -3111,7 +2998,9 @@ export async function handleChatCore({
   // Namespaced by the calling API key: dedup hands the SAME response object to
   // every joiner, so a shared hash across keys is a cross-principal response
   // leak (GHSA-6c7w-56xp-wpc6).
-  const dedupHash = dedupEnabled ? computeRequestHash(dedupRequestBody, apiKeyInfo?.id) : null;
+  const dedupHash = dedupEnabled
+    ? computeRequestHash(dedupRequestBody, apiKeyInfo?.id, trustedEffortContext)
+    : null;
 
   const executeProviderRequest = async (modelToCall = effectiveModel, allowDedup = false) => {
     const execute = async () => {
@@ -3121,12 +3010,15 @@ export async function handleChatCore({
       let bodyToSend = await prepareUpstreamBody({
         translatedBody,
         modelToCall,
+        ...trustedEffortContext,
         provider,
         targetFormat,
-        credentials,
+        credentials: getExecutionCredentials(),
         log,
         bypassDefaultToolLimit: isOpencodeClient,
         isOpencodeClient,
+        rawBody: body,
+        clientRawRequest,
       });
 
       // Global System Prompt — SINGLE injection point (post-translation) for
@@ -4524,12 +4416,25 @@ export async function handleChatCore({
         // stay aligned if this block ever runs after a path that mutates body.model (e.g. fallback).
         try {
           const retryModelId = String(translatedBody.model || effectiveModel);
+          const retryBody = await prepareUpstreamBody({
+            translatedBody,
+            modelToCall: retryModelId,
+            ...trustedEffortContext,
+            provider,
+            targetFormat,
+            credentials: getExecutionCredentials(),
+            log,
+            bypassDefaultToolLimit: isOpencodeClient,
+            isOpencodeClient,
+            rawBody: body,
+            clientRawRequest,
+          });
           assertManagedLeaseFence(getExecutionConnectionId(getExecutionCredentials()));
           const retryResult = normalizeExecutorResult(
             await runWithCapture(providerRequestCapture, () =>
               executor.execute({
                 model: retryModelId,
-                body: translatedBody,
+                body: retryBody,
                 stream: upstreamStream,
                 credentials: getExecutionCredentials(),
                 signal: streamController.signal,