textmachine/backend/internal/config/echo_mine_test.go

package config

import (
	"fmt"
	"os"
	"path/filepath"
	"strings"
	"testing"
	"time"

	"gopkg.in/yaml.v3"

	"textmachine/backend/internal/llm"
)

// echo_mine_test.go is the §2 regression gate for the DeepSeek echo mine: on a
// provider empirically shown to ECHO the untranslated CJK source when thinking is
// OFF (traced 2026-07-04), no model may be configured to disable/suppress thinking
// at reasoning=off — that re-arms a silent refusal (HTTP 200, no translation) in
// prod. LoadModels must fail-fast on both injection surfaces, while leaving the
// SAFE GLM editor (thinking:disabled over a Russian draft) untouched.
//
// Mutation check (how this guards against a future edit re-arming the mine):
// delete the echoMineViolation call in LoadModels (or the field) and the four
// "armed" cases below stop failing — the test goes red exactly when the guard is
// removed.

// writeAndLoadModels materializes a models.yaml body and runs it through the real
// LoadModels validation (mirrors the config_test.go temp-file pattern).
func writeAndLoadModels(t *testing.T, body string) (*Models, error) {
	t.Helper()
	dir := t.TempDir()
	path := filepath.Join(dir, "models.yaml")
	if err := os.WriteFile(path, []byte(body), 0o644); err != nil {
		t.Fatal(err)
	}
	return LoadModels(path)
}

func TestDeepSeekEchoMineRejected(t *testing.T) {
	today := time.Now().UTC().Format("2006-01-02")
	// provCaps is injected under provider p; modelCaps under model m.
	tmpl := func(echoProne bool, modelExtra string) string {
		flag := ""
		if echoProne {
			flag = "\n    echoes_when_thinking_off: true"
		}
		return fmt.Sprintf(`
prices_checked: %q
default_model: m
providers:
  p:
    kind: openai
    base_url: http://x%s
models:
  m:
    provider: p
    price: { input_per_m: 1, output_per_m: 2 }%s
`, today, flag, modelExtra)
	}

	cases := []struct {
		name       string
		echoProne  bool
		modelExtra string
		wantErr    bool
	}{
		{
			// The exact traced mine: the thinking-disable switch via capabilities.
			name:       "armed via capabilities.reasoning.extra_body_disable",
			echoProne:  true,
			modelExtra: "\n    capabilities: { reasoning: { control: extra_body_disable, off_extra_body: { thinking: { type: disabled } } } }",
			wantErr:    true,
		},
		{
			// The second surface: a thinking-disable smuggled through extra_body,
			// which openAIRequest merges into the wire body — the control-only guard
			// would miss this, so it must be caught too.
			name:       "armed via extra_body thinking key",
			echoProne:  true,
			modelExtra: "\n    extra_body: { thinking: { type: disabled } }",
			wantErr:    true,
		},
		{
			// The NESTED form: chat_template_kwargs.thinking:false is DeepSeek's
			// documented V3.1+ disable — the most likely real arming form. A flat
			// top-level scan misses it; the recursive scan must catch it (ext-review).
			name:       "armed via nested chat_template_kwargs.thinking (DeepSeek V3.1+)",
			echoProne:  true,
			modelExtra: "\n    extra_body: { chat_template_kwargs: { thinking: false } }",
			wantErr:    true,
		},
		{
			// reasoning_effort at off also pushes thinking off the default.
			name:       "armed via capabilities.reasoning.effort",
			echoProne:  true,
			modelExtra: "\n    capabilities: { reasoning: { control: effort, off_effort: minimal } }",
			wantErr:    true,
		},
		{
			// The SAFE GLM case: an IDENTICAL thinking-disable is legal on a provider
			// that is NOT echo-prone (its editor input is a Russian draft).
			name:       "safe: same thinking-disable on a non-echo-prone provider",
			echoProne:  false,
			modelExtra: "\n    capabilities: { reasoning: { control: extra_body_disable, off_extra_body: { thinking: { type: disabled } } } }",
			wantErr:    false,
		},
		{
			// The current benign DeepSeek state: no reasoning capability → off is a
			// no-op → thinking stays at the provider default (ON).
			name:       "safe: echo-prone provider with the default (no) reasoning control",
			echoProne:  true,
			modelExtra: "",
			wantErr:    false,
		},
	}
	for _, c := range cases {
		t.Run(c.name, func(t *testing.T) {
			_, err := writeAndLoadModels(t, tmpl(c.echoProne, c.modelExtra))
			if c.wantErr {
				if err == nil || !strings.Contains(err.Error(), "echo mine") {
					t.Fatalf("expected fail-fast mentioning the echo mine, got %v", err)
				}
				return
			}
			if err != nil {
				t.Fatalf("expected the config to load, got %v", err)
			}
		})
	}
}

// TestBoevoyConfigDeepSeekThinkingStaysOn guards the REAL configs/models.yaml
// against a future edit arming the mine. It asserts the three load-bearing facts
// directly (unmarshalling raw to dodge the prices_checked staleness date-bomb that
// LoadModels enforces): (1) the deepseek provider is still marked echo-prone — the
// guard is inert without the flag; (2) deepseek-v4-flash resolves to ReasoningNone
// so reasoning="off" stays an off-by-omission (thinking ON by default); (3)
// echoMineViolation is empty. Arm the real config and (2)/(3) go red here, while
// tmctl itself fails loudly via LoadModels.
func TestBoevoyConfigDeepSeekThinkingStaysOn(t *testing.T) {
	raw, err := os.ReadFile(filepath.Join("..", "..", "configs", "models.yaml"))
	if err != nil {
		t.Fatalf("read boevoy models.yaml: %v", err)
	}
	var m Models
	if err := yaml.Unmarshal(raw, &m); err != nil {
		t.Fatalf("parse boevoy models.yaml: %v", err)
	}

	const model = "deepseek-v4-flash"
	mod, ok := m.Models[model]
	if !ok {
		t.Fatalf("boevoy config no longer defines %s — update this guard deliberately", model)
	}
	prov := m.Providers[mod.Provider]
	if !prov.EchoesWhenThinkingOff {
		t.Fatalf("provider %s must stay marked echoes_when_thinking_off — the echo-mine guard is inert without it", mod.Provider)
	}
	if got := m.ResolveCapability(model).Reasoning.Control; got != llm.ReasoningNone {
		t.Fatalf("%s must resolve to ReasoningNone so reasoning=off stays a no-op (thinking ON), got %q — echo mine armed", model, got)
	}
	if why := m.echoMineViolation(model); why != "" {
		t.Fatalf("%s re-arms the echo mine: %s", model, why)
	}
}

// TestBoevoyConfigEditorDsproAndGrokReasoning pins two invariants. (1) The D39.22 interim editor:
// the default editor of both boevoy pipelines is deepseek-v4-pro BILINGUAL (supersedes the D30.1
// glm-5 default; the pere-run did not crown glm-vs-dspro, dspro = judge #1 / cheaper / repairable
// defects → owner interim, glm-5 kept as pipeline-arm-glm.yaml reserve) and its reasoning is "off"
// — the LOAD-BEARING echo-mine semantic: on deepseek-v4-pro "off" resolves to ReasoningNone (a
// NO-OP), so thinking stays ON and the echo-mine is NOT armed. (2) grok-4.3 (still the channel-B
// tier / 18+ judge in models.yaml) resolves to an EXPLICIT reasoning off-switch (ReasoningEffortField
// + OffEffort "none"), NOT off-by-omission — grok's OMITTED default is "low" (it thinks, billed
// additively per xAI), and it must not trip the echo-mine guard. c2's select (judge) stage stays glm-5.
func TestBoevoyConfigEditorDsproAndGrokReasoning(t *testing.T) {
	m, err := LoadModels(filepath.Join("..", "..", "configs", "models.yaml"))
	if err != nil {
		t.Fatalf("load boevoy models.yaml: %v", err)
	}
	// (2) grok-4.3 reasoning invariant (channel B / 18+ judge; unchanged by the editor flip).
	const grok = "grok-4.3"
	cap := m.ResolveCapability(grok)
	if cap.Reasoning.Control != llm.ReasoningEffortField || cap.Reasoning.OffEffort != "none" {
		t.Errorf("%s must resolve to explicit reasoning_effort:\"none\" (control=effort, off_effort=none), got control=%q off_effort=%q — control:none leaves grok's \"low\" default (thinks, additive billing)",
			grok, cap.Reasoning.Control, cap.Reasoning.OffEffort)
	}
	if why := m.echoMineViolation(grok); why != "" {
		// The guard must stay silent for grok — NOT because "the editor only sees a Russian draft" (that
		// rationale died with D30.1: the editor is bilingual), but because xai carries no echo-prone flag.
		// Pack-17 deliberately did NOT add one: it would outlaw grok's off-switch, whose money consequence
		// (grok then thinks by default, billed additively) is the owner's call. The exposure is surfaced
		// loudly instead (pipeline.Runner.sourceEchoExposure) and caught by the echo gate (D19.2).
		t.Errorf("%s must not trip the echo-mine guard (xai declares no echoes_when_thinking_off): %s", grok, why)
	}
	// (1) D39.22 interim editor flip: deepseek-v4-pro bilingual editor, reasoning "off" (a ReasoningNone
	// no-op — thinking stays ON), in both boevoy pipelines. The glm-5 default is superseded (kept as a
	// reserve arm); the c2 select (judge) glm-5 slot is a DIFFERENT role and is unaffected.
	for _, pf := range []string{"pipeline-c1.yaml", "pipeline-c2.yaml"} {
		p, err := LoadPipeline(filepath.Join("..", "..", "configs", pf), m, "zh-ru", nil)
		if err != nil {
			t.Fatalf("load %s: %v", pf, err)
		}
		var editModel, editReasoning, selectModel string
		var editFewShot *bool
		for _, st := range p.Stages {
			switch st.Role {
			case "editor":
				editModel, editReasoning, editFewShot = st.Model, st.Reasoning, st.FewShot
			case "judge":
				selectModel = st.Model
			}
		}
		if editModel != "deepseek-v4-pro" {
			t.Errorf("%s edit stage model = %q, want deepseek-v4-pro (D39.22 interim editor; glm-5 is the reserve arm)", pf, editModel)
		}
		if editReasoning != "off" {
			t.Errorf("%s edit stage reasoning = %q, want \"off\" — the echo-mine semantic: on deepseek-v4-pro \"off\"=ReasoningNone (no-op) so thinking stays ON, never armed", pf, editReasoning)
		}
		// The reasoning editor drops the hand-written few-shot block (D38.5 / exp14 §2а: its own CoT is
		// disturbed by the examples). This assertion used to live on pipeline-arm-deepseek-pro.yaml — the
		// arm that swapped the editor TO deepseek-v4-pro. D39.22 made that editor the production one, so
		// the arm became a copy of c1 and was retired; the guarantee moves here, onto the configs that
		// actually ship it.
		if editFewShot == nil || *editFewShot {
			t.Errorf("%s edit stage must set few_shot:false — the reasoning editor drops the hand-written examples (D38.5), got %v", pf, editFewShot)
		}
		if pf == "pipeline-c2.yaml" && selectModel != "glm-5" {
			t.Errorf("%s select (judge) stage must stay glm-5 (Gemini Phase-2 slot), got %q", pf, selectModel)
		}
	}
}

// TestSwapArmConfigs guards the WS6 editor swap-arms (D39.9 — editor = the weak link). The arms are
// CONFIG, not code: each pipeline-arm-*.yaml swaps ONLY the editor model, so a different editor-model
// resolves a different stageSnap.Model → a distinct snapshot_W2 (a separate run). This pins: (1) each arm
// loads and its editor is the arm model; (2) neither arm re-arms the echo mine; (3) neither keeps the
// few-shot block off — both surviving arms run NON-reasoning editors, for which the examples are the
// point; (4) the mistral arm's PRE-CONDITION — mistral-large-2512 carries a rate_limit so the WS1
// per-model rate-guard actually throttles it (§6(б) F4).
//
// ⚠ That every OTHER key matches production is not pinned here and cannot be, because a list of keys can
// only check what somebody wrote into it: it is TestAnArmIsTheProductionConfigWithADifferentEditor, which
// compares the loaded configs whole.
//
// ⚠ pipeline-arm-deepseek-pro.yaml is RETIRED (07.09) rather than absent by oversight: D39.22 made
// deepseek-v4-pro the production editor, so that arm isolated no variable — it was a second copy of c1
// that had to be kept in step forever, and its own pin read as "the arm is wrong" whenever c1 moved. The
// few-shot guarantee it carried moved onto c1/c2 above; the equality gate refuses to let an arm silently
// become a copy again.
// editorKeepsFewShot records, per EDITOR MODEL, whether that editor keeps the hand-written ---FEWSHOT---
// block. It is a decision, not a property a model exposes: a reasoning editor's own chain of thought is
// disturbed by worked examples (D38.5 / exp14 §2а) and drops them; the others keep them, because for a
// non-reasoning editor the examples are the point. Keyed on the model so that a new ARM inherits the
// decision and a new EDITOR has to state one — an unrecorded editor is refused, never defaulted.
var editorKeepsFewShot = map[string]bool{
	"deepseek-v4-pro":    false,
	"glm-5":              true,
	"mistral-large-2512": true,
}

func TestSwapArmConfigs(t *testing.T) {
	m, err := LoadModels(filepath.Join("..", "..", "configs", "models.yaml"))
	if err != nil {
		t.Fatalf("load boevoy models.yaml: %v", err)
	}
	arms, err := filepath.Glob(filepath.Join("..", "..", "configs", "pipeline-arm-*.yaml"))
	if err != nil {
		t.Fatal(err)
	}
	if len(arms) == 0 {
		t.Fatal("no configs/pipeline-arm-*.yaml — this test would be enforcing nothing")
	}
	for _, file := range arms {
		t.Run(filepath.Base(file), func(t *testing.T) {
			p, lerr := LoadPipeline(file, m, "zh-ru", nil)
			if lerr != nil {
				t.Fatalf("load %s: %v", file, lerr)
			}
			var edit *Stage
			for i := range p.Stages {
				if p.Stages[i].Role == "editor" {
					edit = &p.Stages[i]
				}
			}
			if edit == nil {
				t.Fatalf("%s has no editor stage", file)
			}
			// (1) the arm file NAMES its editor, and the name is the only handle a deploy flip has: the
			// slug in pipeline-arm-<slug>.yaml must be part of the editor model it resolves. Derived from
			// the filename rather than listed, so a new arm is covered by being named; without it the
			// glob-driven checks below would pass an arm whose editor was swapped to another arm's model.
			slug := strings.TrimSuffix(strings.TrimPrefix(filepath.Base(file), "pipeline-arm-"), ".yaml")
			if !strings.Contains(edit.Model, slug) {
				t.Errorf("%s resolves editor %q, which does not carry the file's own slug %q — an arm's name is what a deploy flip selects, so it has to say which editor it is", filepath.Base(file), edit.Model, slug)
			}
			// (2) echo-mine safety: the arm editor's model must not re-arm the mine.
			if why := m.echoMineViolation(edit.Model); why != "" {
				t.Errorf("%s editor %q re-arms the echo mine: %s", file, edit.Model, why)
			}
			// (3) few-shot policy, looked up BY EDITOR MODEL rather than by file: the decision belongs to
			// the model, so a new arm on a known editor is covered by existing it, and a new EDITOR is
			// refused until somebody records what it does with the examples.
			keeps, known := editorKeepsFewShot[edit.Model]
			if !known {
				t.Fatalf("%s runs editor %q, whose few-shot decision is not recorded in editorKeepsFewShot — a reasoning editor drops the hand examples and a non-reasoning one keeps them, and defaulting either way would be silent (D38.5)", file, edit.Model)
			}
			switch {
			case keeps && edit.FewShot != nil:
				t.Errorf("%s: editor %q keeps the few-shot block, so few_shot must be left at its default (nil = ON), got %v", file, edit.Model, *edit.FewShot)
			case !keeps && (edit.FewShot == nil || *edit.FewShot):
				t.Errorf("%s: editor %q is a reasoning editor and must set few_shot:false, got %v", file, edit.Model, edit.FewShot)
			}
		})
	}
	// (4) mistral rate-guard pre-condition: mistral-large-2512 must carry a rate_limit so the WS1 wave
	// per-model semaphore actually throttles it (mistral ~48% retry-fails under N-∥ without it).
	if rl := m.Models["mistral-large-2512"].RateLimit; rl.MaxConcurrency <= 0 {
		t.Errorf("mistral-large-2512 must configure rate_limit.max_concurrency>0 (WS6 pre-condition — the swap-arm needs the WS1 rate-guard), got %d", rl.MaxConcurrency)
	}
}