Merge pull request 'fix: disable hidden thinking for bounded local advice' (#176) from codex/wp-200-medium-thinking-control into main

adapters/models/openai_compatible_local.go

@@ -96,11 +96,15 @@ func (a *OpenAICompatibleLocalAdapter) Discover(ctx context.Context) ([]Discover
 }
 
 type compatibleChatRequest struct {
-	Model       string              `json:"model"`
-	Messages    []compatibleMessage `json:"messages"`
-	Temperature int                 `json:"temperature"`
-	Stream      bool                `json:"stream"`
-	MaxTokens   int                 `json:"max_tokens"`
+	Model              string                   `json:"model"`
+	Messages           []compatibleMessage      `json:"messages"`
+	Temperature        int                      `json:"temperature"`
+	Stream             bool                     `json:"stream"`
+	MaxTokens          int                      `json:"max_tokens"`
+	ChatTemplateKwargs compatibleTemplateKwargs `json:"chat_template_kwargs"`
+}
+type compatibleTemplateKwargs struct {
+	EnableThinking bool `json:"enable_thinking"`
 }
 type compatibleMessage struct {
 	Role    string `json:"role"`
@@ -121,7 +125,7 @@ func (a *OpenAICompatibleLocalAdapter) Advise(ctx context.Context, request cogni
 	if err := ValidateAdvisoryRequest(request); err != nil || request.ProviderModelID == "" || len(request.Context) != 0 {
 		return cognitive.AdvisoryResponse{}, ErrInvalidAdvisory
 	}
-	payload := compatibleChatRequest{Model: string(request.ProviderModelID), Messages: []compatibleMessage{{Role: "system", Content: "Provide non-canonical advisory text only. Never claim that an action occurred."}, {Role: "user", Content: request.Message}}, Temperature: 0, Stream: false, MaxTokens: request.MaxOutputBytes / 4}
+	payload := compatibleChatRequest{Model: string(request.ProviderModelID), Messages: []compatibleMessage{{Role: "system", Content: "Provide non-canonical advisory text only. Never claim that an action occurred."}, {Role: "user", Content: request.Message}}, Temperature: 0, Stream: false, MaxTokens: request.MaxOutputBytes / 4, ChatTemplateKwargs: compatibleTemplateKwargs{EnableThinking: false}}
 	var result compatibleChatResponse
 	status, err := a.requestJSON(ctx, http.MethodPost, "/chat/completions", payload, &result)
 	if err != nil {

adapters/models/openai_compatible_local_test.go

@@ -2,15 +2,20 @@ package models
 
 import (
 	"context"
+	"encoding/json"
 	"io"
 	"net/http"
 	"strings"
 	"testing"
 	"time"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
 )
 
 type captureRoundTripper struct {
 	authorization string
+	body          []byte
 }
 
 func TestOpenAICompatibleLocalAdapterUsesConfiguredTimeout(t *testing.T) {
@@ -31,7 +36,14 @@ func TestOpenAICompatibleLocalAdapterRejectsNonPositiveTimeout(t *testing.T) {
 
 func (t *captureRoundTripper) RoundTrip(request *http.Request) (*http.Response, error) {
 	t.authorization = request.Header.Get("Authorization")
-	return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(`{"data":[]}`)), Header: make(http.Header), Request: request}, nil
+	if request.Body != nil {
+		t.body, _ = io.ReadAll(request.Body)
+	}
+	body := `{"data":[]}`
+	if strings.HasSuffix(request.URL.Path, "/chat/completions") {
+		body = `{"choices":[{"finish_reason":"stop","message":{"role":"assistant","content":"ok"}}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`
+	}
+	return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(body)), Header: make(http.Header), Request: request}, nil
 }
 
 func TestOpenAICompatibleLocalAdapterUsesResolvedAPIKey(t *testing.T) {
@@ -57,3 +69,26 @@ func TestOpenAICompatibleLocalAdapterDoesNotSendAuthorizationWithoutKey(t *testi
 		t.Fatalf("unexpected authorization header")
 	}
 }
+
+func TestOpenAICompatibleLocalAdapterDisablesTemplateThinkingForBoundedAdvisory(t *testing.T) {
+	transport := &captureRoundTripper{}
+	adapter, err := NewOpenAICompatibleLocalAdapter("http://gateway/v1", transport)
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = adapter.Advise(context.Background(), cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: "request-1", ActorID: "actor", PrincipalID: "principal", ScopeID: "scope", Message: "help", Locale: "ru-RU", MaxOutputBytes: 256, ProviderBindingID: "binding-1", ProviderBindingRevision: 1, ProviderInstanceID: "provider-1", ModelDescriptorID: "model-1", ProviderModelID: capability.ProviderModelID("medium-model")})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var payload struct {
+		ChatTemplateKwargs struct {
+			EnableThinking bool `json:"enable_thinking"`
+		} `json:"chat_template_kwargs"`
+	}
+	if err := json.Unmarshal(transport.body, &payload); err != nil {
+		t.Fatal(err)
+	}
+	if payload.ChatTemplateKwargs.EnableThinking {
+		t.Fatal("expected bounded advisory to disable template thinking")
+	}
+}

dev/ACTIVE_TASK.json

@@ -1,5 +1,5 @@
 {
-  "task_file": "dev/tasks/DEV-000198-development-gate-live-evidence.json",
-  "task_id": "DEV-000198",
+  "task_file": "dev/tasks/DEV-000199-medium-advisory-thinking-control.json",
+  "task_id": "DEV-000199",
   "task_revision": 1
 }

dev/tasks/DEV-000199-medium-advisory-thinking-control.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000199",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c7170ee751b715b0cf6721d8491530c185870b2a",
+  "risk_tier": "E",
+  "intent": "Disable provider template reasoning for bounded OpenAI-compatible local advisory requests so a loaded MEDIUM model can return bounded advisory content instead of consuming the entire output cap with hidden reasoning.",
+  "requirement_ids": ["WP-200", "model fleet/failover", "bounded advisory output"],
+  "allowed_change_scope": {"modules": ["models adapter"], "paths": ["adapters/models/openai_compatible_local.go", "adapters/models/openai_compatible_local_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000199-medium-advisory-thinking-control.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "contracts/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing OpenAI-compatible local advisory contract"], "may_change": ["provider request serialization adds typed template control"]},
+  "semantic_owner": "cognitive adapter",
+  "canonical_state": [],
+  "state_transitions": ["bounded advisory request asks compatible runtimes to disable hidden template reasoning"],
+  "side_effects_and_privileges": ["bounded local provider inference only; no activation or promotion"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["provider ignoring the optional request field remains an explicit qualification result"],
+  "assumptions": ["llama.cpp-compatible OpenAI endpoints ignore or honor the typed chat_template_kwargs field without breaking the existing contract"],
+  "acceptance_scenarios": ["request contains enable_thinking=false", "adapter still validates bounded completion", "provider failure remains typed and failover-compatible"],
+  "required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl/independent -task dev/tasks/DEV-000199-medium-advisory-thinking-control.json", "git diff --check", "CI", "production medium bounded smoke and public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}