muonroi-cli 1.4.0 → 1.5.0

package/dist/src/pil/__tests__/layer16-clarity.test.js

@@ -1,5 +1,9 @@
 import { describe, expect, it } from "vitest";
-import { buildInterviewQuestion, detectClarityGaps, resolveGapsNonInteractive } from "../layer16-clarity.js";
+import { buildInterviewQuestion, resolveGapsNonInteractive } from "../layer16-clarity.js";
+// Phase 2 (2026-06-16): detectClarityGaps + its keyword option-builders were
+// removed (the model now generates every clarification). The surviving helpers
+// — buildInterviewQuestion (render) and resolveGapsNonInteractive (headless
+// default-answer resolution) — are exercised here with model-shaped gaps.
 const EMPTY_PROJECT = {
     language: "typescript",
     framework: null,
@@ -10,122 +14,10 @@ const EMPTY_PROJECT = {
         { path: "src/billing/", name: "billing", entryFiles: [], exportedSymbols: [] },
     ],
     eePatterns: [],
-    relevantModules: [],
+    relevantModules: [{ path: "src/auth/", relevance: "named in prompt", exists: true }],
     scannedAt: Date.now(),
     cwd: "/proj",
 };
-describe("detectClarityGaps()", () => {
-    it("detects outcome gap for vague non-debug prompt", () => {
-        // PIL-L6 fix — debug now joins the autofill set, so vague debug prompts
-        // ("fix auth") no longer trigger an outcome question. Use a generate
-        // prompt instead to still cover the gap-detection path.
-        const gaps = detectClarityGaps("build something", "generate", 0.7, EMPTY_PROJECT);
-        const outcomeGap = gaps.find((g) => g.dimension === "outcome");
-        expect(outcomeGap).toBeDefined();
-    });
-    it("does NOT detect outcome gap for vague debug prompt (autofilled)", () => {
-        const gaps = detectClarityGaps("fix auth", "debug", 0.7, EMPTY_PROJECT);
-        const outcomeGap = gaps.find((g) => g.dimension === "outcome");
-        expect(outcomeGap).toBeUndefined();
-    });
-    it("does NOT detect an outcome gap for a vague general prompt (B2 intent-swallow guard)", () => {
-        // B2 — a `general` prompt's only outcome options are tautological
-        // ("Task completed" / "Issue resolved"). Asking them lets the default
-        // answer overwrite the user's real request, so the intent collapses to
-        // "general: Task completed" and the original prompt is lost. Skip the
-        // askcard so the outcome falls back to the raw request downstream.
-        const gaps = detectClarityGaps("the project feels messy", "general", 0.7, EMPTY_PROJECT);
-        const outcomeGap = gaps.find((g) => g.dimension === "outcome");
-        expect(outcomeGap).toBeUndefined();
-    });
-    it("detects scope gap when no file reference", () => {
-        const gaps = detectClarityGaps("fix auth", "debug", 0.7, EMPTY_PROJECT);
-        const scopeGap = gaps.find((g) => g.dimension === "scope");
-        expect(scopeGap).toBeDefined();
-    });
-    it("returns no gaps for specific prompt", () => {
-        const gaps = detectClarityGaps("fix TypeError in src/auth/login.ts:42", "debug", 0.9, EMPTY_PROJECT);
-        expect(gaps).toHaveLength(0);
-    });
-    it("scope options include matching bounded contexts", () => {
-        const gaps = detectClarityGaps("fix auth", "debug", 0.7, EMPTY_PROJECT);
-        const scopeGap = gaps.find((g) => g.dimension === "scope");
-        expect(scopeGap?.options.some((o) => o.includes("auth"))).toBe(true);
-    });
-    it("does NOT detect a scope gap for a general prompt with no codebase signal (B2-symmetric scope guard)", () => {
-        // Live drive (session 8a87aa060c6a): the pure non-codebase prompt "Reply
-        // with exactly one word: PONG" fired the scope askcard "Which part of the
-        // codebase should this target?" because countFileReferences /
-        // hasExplicitScope / hasOperationalScope were all empty — the detector
-        // assumes every prompt is a codebase task. A general/unclassified prompt
-        // has no codebase dimension to scope, so the question is nonsensical (and
-        // its acceptance card is downstream noise). Skip it, symmetric to the B2
-        // outcome guard; scope falls back to project-root downstream.
-        const gaps = detectClarityGaps("Reply with exactly one word: PONG", "general", 0.6, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeUndefined();
-        // The only candidate gap was scope → general prompt now yields zero gaps,
-        // so discovery never marks interviewed=true and shows no acceptance card.
-        expect(gaps).toHaveLength(0);
-    });
-    it("STILL detects a scope gap for a classified (non-general) task with no file reference", () => {
-        // Guard must stay narrow: a real code task that simply omitted a path still
-        // benefits from the scope-narrowing askcard. Only general/null is skipped.
-        const gaps = detectClarityGaps("implement the search feature", "generate", 0.7, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeDefined();
-    });
-    it("does NOT detect a scope gap for an image-analysis prompt (image is the scope)", () => {
-        // Live drive (PR#34 probe): "Take a screenshot of the homepage and analyze
-        // the diagram.png image to describe its layout" fired the codebase-scope
-        // askcard "Which part of the codebase should this target?" — nonsensical for
-        // an image-analysis task. The image (screenshot / diagram.png) IS the scope,
-        // symmetric to how operational (CI/build) prompts are scoped to the pipeline.
-        const gaps = detectClarityGaps("Take a screenshot of the homepage and analyze the diagram.png image to describe its layout", "analyze", 0.7, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeUndefined();
-        // analyze autofills outcome, so with scope suppressed there are zero gaps →
-        // no interview, no acceptance card.
-        expect(gaps).toHaveLength(0);
-    });
-    it("STILL detects a scope gap for a code task that mentions an ambiguous non-image word", () => {
-        // Narrowness guard: image-scope suppression must not swallow real codebase
-        // tasks. "add a logo to the header" carries no concrete image signal (no
-        // file extension / screenshot / photo), so the scope askcard stays.
-        const gaps = detectClarityGaps("add a logo to the header", "generate", 0.7, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeDefined();
-    });
-    it("does NOT detect a scope gap for a web-search / external-info prompt", () => {
-        // Live drive (tavily probe, session d7a45a2dba30): "search the web for the
-        // latest vitest release notes" classified taskType=analyze fired the
-        // codebase-scope askcard and recorded a wrong scope ("src/mcp"). A
-        // web-search task is scoped to the web, not the codebase — symmetric to the
-        // image-scope and operational-scope guards.
-        const gaps = detectClarityGaps("search the web for the latest vitest release notes", "analyze", 0.7, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeUndefined();
-        expect(gaps).toHaveLength(0);
-    });
-    it("does NOT detect a scope gap for a self-contained computation prompt (data is inline)", () => {
-        // Live drive (deepseek-vs-grok A/B, session 17fc23f0): "Compute f([3,1,2])
-        // where f sorts the list ascending then returns the sum of the first two
-        // elements." classified taskType=analyze (regex:read matched the bare word
-        // "list", conf 0.80 → skipped the brain) fired BOTH the pil-interview scope
-        // askcard ("Which part of the codebase should this target?" → auto "Entire
-        // project") AND the pil-acceptance card. The operand [3,1,2] is supplied
-        // inline — the task has no codebase dimension to scope. Symmetric to the
-        // image / web / operational scope guards.
-        const gaps = detectClarityGaps("Compute f([3,1,2]) where f sorts the list ascending then returns the sum of the first two elements.", "analyze", 0.8, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeUndefined();
-        // analyze autofills outcome, so with scope suppressed there are zero gaps →
-        // no interview, no acceptance card.
-        expect(gaps).toHaveLength(0);
-    });
-    it("STILL detects a scope gap for a code task that embeds a literal but no compute framing", () => {
-        // Narrowness guard: the inline-literal suppression must not swallow real
-        // codebase tasks. "set the default retry delays to [100, 200, 400] in the
-        // config" carries a literal but is scoped to the codebase (no compute verb),
-        // so the scope askcard stays.
-        const gaps = detectClarityGaps("set the default retry delays to [100, 200, 400] in the config", "generate", 0.7, EMPTY_PROJECT);
-        expect(gaps.find((g) => g.dimension === "scope")).toBeDefined();
-    });
-});
 describe("buildInterviewQuestion()", () => {
     it("builds a CouncilQuestionData with pil-interview phase", () => {
         const gap = {
@@ -141,11 +33,35 @@ describe("buildInterviewQuestion()", () => {
         expect(q.options).toBeDefined();
         expect(q.options.some((o) => o.kind === "freetext")).toBe(true);
     });
+    it("surfaces the model's reason (gap.description) as the askcard context", () => {
+        const gap = {
+            dimension: "outcome",
+            description: "answering this changes whether we add OAuth or just API keys",
+            suggestedQuestion: "Which auth method?",
+            options: ["OAuth", "API keys"],
+            defaultIndex: 0,
+        };
+        const q = buildInterviewQuestion(gap, "q-2");
+        expect(q.context).toBe("answering this changes whether we add OAuth or just API keys");
+    });
 });
 describe("resolveGapsNonInteractive()", () => {
-    it("fills gaps with best-effort from project context", () => {
-        const gaps = detectClarityGaps("fix auth", "debug", 0.7, EMPTY_PROJECT);
+    it("fills gaps with best-effort defaults from the model options + project context", () => {
+        const gaps = [
+            {
+                dimension: "outcome",
+                description: "Model-generated clarification #1",
+                suggestedQuestion: "What outcome do you expect?",
+                options: ["Error resolved", "Other (type free answer)"],
+                defaultIndex: 0,
+            },
+        ];
         const resolved = resolveGapsNonInteractive(gaps, EMPTY_PROJECT, "fix auth");
+        expect(resolved.outcome).toBe("Error resolved");
+        expect(resolved.scope.length).toBeGreaterThan(0);
+    });
+    it("falls back to the raw-derived outcome when there is no outcome gap", () => {
+        const resolved = resolveGapsNonInteractive([], EMPTY_PROJECT, "fix the login bug");
         expect(resolved.outcome).toBeTruthy();
         expect(resolved.scope.length).toBeGreaterThan(0);
     });

package/dist/src/pil/__tests__/layer4-gsd.test.js

@@ -75,6 +75,43 @@ describe("layer4Gsd (gsd-native)", () => {
         const result = await layer4Gsd(makeCtx({ raw: "review the pull request" }));
         expect(["review", "discuss", "execute"]).toContain(result.gsdPhase);
     });
+    it("routes a question-shaped analyze/debug prompt to the QUESTION directive (no 'state a plan')", async () => {
+        // De-robotizing: a plain question must not get the STANDARD "state a 2-3 line
+        // plan" scaffold even when L1 classifies it analyze/debug (not "general").
+        const q = "why does the build fail intermittently?";
+        const result = await layer4Gsd(makeCtx({ raw: q, enriched: q, taskType: "debug", intentKind: "task" }));
+        expect(result.enriched).toContain("QUESTION / explanatory");
+        expect(result.enriched).not.toContain("State a 2-3 line plan");
+    });
+    it("treats a genuine general question (general + task) as informational", async () => {
+        const q = "what does the enrichment layer do?";
+        const result = await layer4Gsd(makeCtx({ raw: q, enriched: q, taskType: "general", intentKind: "task" }));
+        expect(result.enriched).toContain("QUESTION / explanatory");
+    });
+    it("does NOT treat an implementation request as informational even if phrased as a question", async () => {
+        // isImplementationIntent guards the question clause: "can you refactor … and
+        // wire up …" is a real edit task → STANDARD scaffold, not the QUESTION directive.
+        const q = "can you refactor the dropdown and wire up the keyboard handlers?";
+        const result = await layer4Gsd(makeCtx({ raw: q, enriched: q, taskType: "refactor", intentKind: "task" }));
+        expect(result.enriched).not.toContain("QUESTION / explanatory");
+    });
+    it("Phase 2b: deliverableKind='answer' is informational even for an imperative (no '?') prompt", async () => {
+        // The raw text is a plain imperative — the legacy regex (isQuestionLike /
+        // isMetaAnalysisPrompt) would NOT mark it informational. The model's
+        // deliverableKind='answer' must override that and route to the QUESTION
+        // directive — proving L4 consumes the model signal, not the regex.
+        const raw = "go over the auth module and tell me what it does";
+        const result = await layer4Gsd(makeCtx({ raw, enriched: raw, taskType: "analyze", intentKind: "task", deliverableKind: "answer" }));
+        expect(result.enriched).toContain("QUESTION / explanatory");
+    });
+    it("Phase 2b: deliverableKind='code' is NOT informational even for a question-shaped prompt", async () => {
+        // The raw text reads as a question — the legacy regex would mark it
+        // informational. The model's deliverableKind='code' must override that so
+        // the STANDARD implement scaffold is used (the deliverable is file edits).
+        const raw = "why not just refactor the dropdown and wire the keyboard handlers?";
+        const result = await layer4Gsd(makeCtx({ raw, enriched: raw, taskType: "refactor", intentKind: "task", deliverableKind: "code" }));
+        expect(result.enriched).not.toContain("QUESTION / explanatory");
+    });
     it("uses ctx.gsdPhase from L1 (unified path) without calling routeTask", async () => {
         const { routeTask } = await import("../../ee/bridge.js");
         vi.mocked(routeTask).mockClear();

package/dist/src/pil/__tests__/layer6-output.test.js

@@ -1,5 +1,5 @@
 import { beforeEach, describe, expect, it, vi } from "vitest";
-import { applyPilSuffix, getResponseToolSet, layer6Output } from "../layer6-output.js";
+import { applyPilSuffix, getResponseToolSet, isImplementationIntent, isQuestionLike, layer6Output, } from "../layer6-output.js";
 // Mock bridge for PIL-03 classifyViaBrain tests
 vi.mock("../../ee/bridge.js", () => ({
     classifyViaBrain: vi.fn().mockResolvedValue(null),
@@ -55,6 +55,38 @@ describe("applyPilSuffix — per-task-type suffixes", () => {
             expect(result).toMatch(/Tôi sẽ/); // bilingual
         }
     });
+    it("de-robotized: NO_PREAMBLE bans only openers, not end-of-turn summary or inter-tool narration", () => {
+        // The summary + inter-tool bans were removed because they stripped natural
+        // connective tissue (the "máy móc" feel). Inter-tool spam is still removed
+        // structurally by stripInterToolNarration() in reasoning.ts. This guards
+        // against the bans silently creeping back into the system prompt.
+        const result = applyPilSuffix("S", makeCtx("debug", "concise"));
+        expect(result).toMatch(/FORBIDDEN OPENERS/);
+        expect(result).not.toMatch(/FORBIDDEN END-OF-TURN SUMMARY/);
+        expect(result).not.toMatch(/FORBIDDEN INTER-TOOL NARRATION/);
+    });
+    it("de-robotized: debug suffix is guidance, not a rigid arrow skeleton", () => {
+        // "Format = Hypothesis → Root cause → Fix → Verify" produced stilted,
+        // label-prefixed answers. It must read as guidance now.
+        const result = applyPilSuffix("S", makeCtx("debug", "concise"));
+        expect(result).toContain("OUTPUT RULES (debug)");
+        expect(result).not.toMatch(/Format = Hypothesis/);
+    });
+    it("E: appends the anti-bookkeeping note on the natural path for non-question turns", () => {
+        // The contract's REPORTING rule leaks as a provenance footer ("evidence only
+        // from this turn") on imperative answer turns; the natural path now guards it.
+        const result = applyPilSuffix("S", makeCtx("analyze", "concise"));
+        expect(result).toMatch(/WRITE FOR THE READER/);
+        expect(result).toMatch(/provenance/i);
+    });
+    it("E: skips the anti-bookkeeping note for question turns (L4 QUESTION directive covers them)", () => {
+        const ctx = { ...makeCtx("analyze", "concise"), raw: "why does the enrichment layer fail?" };
+        expect(applyPilSuffix("S", ctx)).not.toMatch(/WRITE FOR THE READER/);
+    });
+    it("E: response-tools path does not add the natural-path bookkeeping note", () => {
+        const result = applyPilSuffix("S", makeCtx("analyze", "balanced"), true);
+        expect(result).not.toMatch(/WRITE FOR THE READER/);
+    });
     it("PIL-04: response-tools path skips budget+preamble (tool already enforces structure)", () => {
         const result = applyPilSuffix("S", makeCtx("analyze", "balanced"), true);
         expect(result).toContain("respond_analyze");
@@ -74,29 +106,33 @@ describe("applyPilSuffix — per-task-type suffixes", () => {
         expect(result).toMatch(/Do NOT append an evidence-provenance footer/);
     });
 });
-describe("getResponseToolSet — PIL-04 Tier 1.1 gating", () => {
-    it("returns response tool for analyze (list-shaped, JSON wins)", () => {
-        const tools = getResponseToolSet(makeCtx("analyze", null));
+describe("getResponseToolSet — narrow gating (de-robotizing)", () => {
+    // Override raw on a typed ctx so the report/question discriminator is exercised.
+    const ctxRaw = (raw, t) => ({ ...makeCtx(t, null), raw });
+    it("returns response tool for analyze on an explicit report/list request", () => {
+        const tools = getResponseToolSet(ctxRaw("audit the orchestrator and list all cost-leak findings", "analyze"));
         expect(Object.keys(tools)).toContain("respond_analyze");
     });
-    it("returns response tool for plan (list-shaped, JSON wins)", () => {
-        const tools = getResponseToolSet(makeCtx("plan", null));
+    it("returns response tool for plan on an explicit plan request", () => {
+        const tools = getResponseToolSet(ctxRaw("plan the migration to the new auth flow step by step", "plan"));
         expect(Object.keys(tools)).toContain("respond_plan");
     });
+    it("returns response tool for debug only on an explicit report request", () => {
+        const tools = getResponseToolSet(ctxRaw("audit the failing suite and list each root cause", "debug"));
+        expect(Object.keys(tools)).toContain("respond_debug");
+    });
     it("returns empty toolset for generate (code-heavy, markdown wins)", () => {
         expect(getResponseToolSet(makeCtx("generate", null))).toEqual({});
     });
     it("returns empty toolset for refactor (diff-heavy, markdown wins)", () => {
         expect(getResponseToolSet(makeCtx("refactor", null))).toEqual({});
     });
-    it("returns response tool for debug (bounded schema, structural enforcement wins)", () => {
-        const tools = getResponseToolSet(makeCtx("debug", null));
-        expect(Object.keys(tools)).toContain("respond_debug");
-    });
     it("returns empty toolset for documentation (prose-heavy)", () => {
         expect(getResponseToolSet(makeCtx("documentation", null))).toEqual({});
     });
-    it("returns response tool for general when no providerId is passed (back-compat)", () => {
+    it("returns response tool for general regardless of report signal (renders as plain markdown)", () => {
+        // general is exempt from the report/question gate: GeneralSchema is pure text
+        // and its renderer shows plain markdown, so respond_general is never robotic.
         const tools = getResponseToolSet(makeCtx("general", null));
         expect(Object.keys(tools)).toContain("respond_general");
     });
@@ -113,24 +149,78 @@ describe("getResponseToolSet — PIL-04 Tier 1.1 gating", () => {
     it("returns empty toolset when taskType is null", () => {
         expect(getResponseToolSet(makeCtx(null, null))).toEqual({});
     });
+    it("gates the response tool for chitchat turns", () => {
+        const ctx = { ...makeCtx("general", null), intentKind: "chitchat" };
+        expect(getResponseToolSet(ctx)).toEqual({});
+    });
+    it("DROPS respond_<task> for question-style debug/analyze/plan (natural markdown path)", () => {
+        // The de-robotizing change: a plain QUESTION must not be forced into the
+        // rigid respond_* schema + labeled renderer. It falls through to the softened
+        // markdown OUTPUT RULES so the answer reads as natural prose.
+        expect(getResponseToolSet(ctxRaw("why does the build fail intermittently?", "debug"))).toEqual({});
+        expect(getResponseToolSet(ctxRaw("analyze how the enrichment function works", "analyze"))).toEqual({});
+        expect(getResponseToolSet(ctxRaw("what is the cleanest way to structure this module?", "plan"))).toEqual({});
+    });
+    it("KEEPS respond_<task> for explicit report / list / plan requests (EN + VI)", () => {
+        const keep = (raw, t) => Object.keys(getResponseToolSet(ctxRaw(raw, t)));
+        expect(keep("list all cost leaks in the orchestrator", "analyze")).toContain("respond_analyze");
+        expect(keep("review the module and report each finding by severity", "analyze")).toContain("respond_analyze");
+        expect(keep("lập kế hoạch migration sang auth flow mới", "plan")).toContain("respond_plan");
+    });
+    it("DROPS respond_<task> for a QUESTION that merely mentions plan/list (narrow-gate fix)", () => {
+        // Live bug (grok interview): a question that QUOTED the phrase "state a 2-3
+        // line plan" matched the bare word 'plan' in STRUCTURED_REPORT_RE and forced
+        // respond_plan, cramming an introspective answer into a rigid plan schema. A
+        // question-shaped prompt must stay on the natural markdown path even when it
+        // contains plan/list words.
+        expect(getResponseToolSet(ctxRaw("what rules constrain you, e.g. the 'state a 2-3 line plan' directive?", "plan"))).toEqual({});
+        expect(getResponseToolSet(ctxRaw("can you list the main points?", "analyze"))).toEqual({});
+        expect(getResponseToolSet(ctxRaw("how would you plan the rollout?", "plan"))).toEqual({});
+        // Imperative delivery requests are NOT question-shaped → still structured.
+        expect(Object.keys(getResponseToolSet(ctxRaw("plan the rollout step by step", "plan")))).toContain("respond_plan");
+    });
     it("drops respond_<task> on an IMPLEMENTATION-intent prompt (no premature terminal answer)", () => {
         // Live (grok session 19fa8895c41c): an "Improve … implement these fixes"
         // prompt classified `debug` got respond_debug; the model called it mid-task
         // as a plan and the turn ended before the edits completed. Implementation
         // turns must fall through to markdown OUTPUT RULES, not a terminal tool.
+        // Implementation intent takes precedence over a report signal.
         const impl = (raw, t) => ({ ...makeCtx(t, null), raw });
         expect(getResponseToolSet(impl("Improve the story-list screen. Implement these prioritized fixes: …", "debug"))).toEqual({});
         expect(getResponseToolSet(impl("Edit ONLY these two files and fix the empty span", "debug"))).toEqual({});
         expect(getResponseToolSet(impl("refactor the genre dropdown and wire up keyboard handlers", "analyze"))).toEqual({});
         expect(getResponseToolSet(impl("triển khai các cải tiến đã đề xuất", "plan"))).toEqual({});
     });
-    it("KEEPS respond_<task> for pure analysis/plan prompts (narrowness guard)", () => {
-        // The deliverable here IS a structured report — must not be suppressed.
-        const ana = (raw, t) => ({ ...makeCtx(t, null), raw });
-        expect(Object.keys(getResponseToolSet(ana("analyze the orchestrator for cost leaks", "analyze")))).toContain("respond_analyze");
-        expect(Object.keys(getResponseToolSet(ana("why does the build fail intermittently?", "debug")))).toContain("respond_debug");
-        expect(Object.keys(getResponseToolSet(ana("plan the migration to the new auth flow", "plan")))).toContain("respond_plan");
-        expect(Object.keys(getResponseToolSet(ana("review the auth module and explain the design", "analyze")))).toContain("respond_analyze");
+});
+describe("getResponseToolSet — Phase 2b deliverableKind consume (model overrides regex)", () => {
+    const ctxD = (raw, t, deliverableKind) => ({
+        ...makeCtx(t, null),
+        raw,
+        deliverableKind,
+    });
+    it("deliverableKind='code' DROPS respond_* even when the prompt reads as a report/list", () => {
+        // Legacy regex (prefersStructuredReport) would KEEP the tool on "list all …".
+        // The model said the deliverable is code → drop it (edits, not a report).
+        expect(getResponseToolSet(ctxD("list all cost leaks in the orchestrator", "analyze", "code"))).toEqual({});
+    });
+    it("deliverableKind='report' KEEPS respond_* even when the prompt is question-shaped", () => {
+        // Legacy regex (isQuestionLike) would DROP the tool on "why does …?". The
+        // model said the deliverable is a structured report → keep it.
+        const tools = getResponseToolSet(ctxD("why does the suite fail — break it down by cause", "analyze", "report"));
+        expect(Object.keys(tools)).toContain("respond_analyze");
+    });
+    it("deliverableKind='answer' DROPS respond_* for non-general even on a report-shaped request", () => {
+        expect(getResponseToolSet(ctxD("plan the migration step by step", "plan", "answer"))).toEqual({});
+    });
+    it("deliverableKind='answer' KEEPS respond_general (general is exempt — renders as plain markdown)", () => {
+        const tools = getResponseToolSet(ctxD("what does the enrichment layer do?", "general", "answer"));
+        expect(Object.keys(tools)).toContain("respond_general");
+    });
+    it("falls back to the legacy regex when deliverableKind is absent (null)", () => {
+        // No model signal → legacy path: question-shaped analyze drops the tool.
+        expect(getResponseToolSet({ ...makeCtx("analyze", null), raw: "why does the build fail?" })).toEqual({});
+        // …and an explicit report request keeps it.
+        expect(Object.keys(getResponseToolSet({ ...makeCtx("analyze", null), raw: "list all cost leaks" }))).toContain("respond_analyze");
     });
 });
 describe("applyPilSuffix — outputStyle variants", () => {
@@ -294,4 +384,33 @@ describe("layer6Output", () => {
         expect(result.enriched).toBe(ctx.enriched);
     });
 });
+describe("isQuestionLike — Vietnamese yes/no question frames (regression: session f6f7881a5fae)", () => {
+    it("detects the live miss: 'check ... dùng được mcp ... không nhé'", () => {
+        // The exact prompt that was mis-routed to the implement/verify scaffold.
+        expect(isQuestionLike("bạn check xem dùng được mcp muonroi-docs không nhé")).toBe(true);
+        // It is NOT an implementation intent, so layer4-gsd's informational gate fires.
+        expect(isImplementationIntent("bạn check xem dùng được mcp muonroi-docs không nhé")).toBe(false);
+    });
+    it("detects common VI yes/no tails", () => {
+        expect(isQuestionLike("dùng được không")).toBe(true);
+        expect(isQuestionLike("cái này chạy được không vậy")).toBe(true);
+        expect(isQuestionLike("đúng không")).toBe(true);
+        expect(isQuestionLike("phải không nhỉ")).toBe(true);
+        expect(isQuestionLike("test đã pass chưa")).toBe(true);
+        expect(isQuestionLike("xong chưa ạ")).toBe(true);
+        expect(isQuestionLike("có chạy được không?")).toBe(true);
+    });
+    it("does NOT treat a mid-sentence negation as a question", () => {
+        // "không là hỏng" = "or it breaks" — 'không' is not the clause-final particle.
+        expect(isQuestionLike("đừng commit file .env không là lộ key")).toBe(false);
+        // Plain imperative with a 'nhé' softener (no 'không'/'chưa' tail) stays a task.
+        expect(isQuestionLike("sửa giúp tôi cái này nhé")).toBe(false);
+        expect(isQuestionLike("triển khai tính năng login")).toBe(false);
+    });
+    it("still detects the pre-existing EN/VI question shapes", () => {
+        expect(isQuestionLike("why does the build fail?")).toBe(true);
+        expect(isQuestionLike("tại sao build lỗi")).toBe(true);
+        expect(isQuestionLike("explain the pipeline")).toBe(true);
+    });
+});
 //# sourceMappingURL=layer6-output.test.js.map

package/dist/src/pil/__tests__/llm-classify.test.js

@@ -23,6 +23,32 @@ describe("createLlmClassifier (PIL Layer 1 Pass 4)", () => {
         expect(result?.outputStyle).toBe("concise");
         expect(result?.confidence).toBeGreaterThan(0.5);
     });
+    it("parses the three-word reply and marks chitchat from the intent word", async () => {
+        const handle = installMockModel({ fixture: { stream: textOnlyStream("general,concise,chat") } });
+        cleanup = handle.uninstall;
+        const factory = (() => handle.model);
+        const classify = createLlmClassifier(factory, "deepseek-v4-flash");
+        const result = await classify("cảm ơn bạn nhé");
+        expect(result?.taskType).toBe("general");
+        expect(result?.intentKind).toBe("chitchat");
+    });
+    it("treats a general QUESTION as task, not chitchat (keep-tools)", async () => {
+        const handle = installMockModel({ fixture: { stream: textOnlyStream("general,concise,task") } });
+        cleanup = handle.uninstall;
+        const factory = (() => handle.model);
+        const classify = createLlmClassifier(factory, "deepseek-v4-flash");
+        const result = await classify("bạn thử call tool setup_guide xem được không");
+        expect(result?.intentKind).toBe("task");
+    });
+    it("defaults intentKind to task when the model omits the third word (backward compatible)", async () => {
+        const handle = installMockModel({ fixture: { stream: textOnlyStream("debug,concise") } });
+        cleanup = handle.uninstall;
+        const factory = (() => handle.model);
+        const classify = createLlmClassifier(factory, "deepseek-v4-flash");
+        const result = await classify("fix the failing build");
+        expect(result?.taskType).toBe("debug");
+        expect(result?.intentKind).toBe("task");
+    });
     it("returns null when the reply cannot be parsed", async () => {
         const handle = installMockModel({ fixture: { stream: textOnlyStream("¯\\_(ツ)_/¯") } });
         cleanup = handle.uninstall;
@@ -100,14 +126,35 @@ describe("createLlmClassifier (PIL Layer 1 Pass 4)", () => {
         expect(result?.taskType).toBe("debug");
         expect(result?.outputStyle).toBe("concise");
     });
-    it("keeps the tiny 16-token budget for non-reasoning models", async () => {
+    it("keeps a tiny output budget for non-reasoning models (24 — four comma words)", async () => {
         const handle = installMockModel({ fixture: { stream: textOnlyStream("generate,concise") } });
         cleanup = handle.uninstall;
         const factory = (() => handle.model);
         const classify = createLlmClassifier(factory, "Qwen/Qwen3-8B"); // reasoning:false
         await classify("add a new endpoint");
         const call = handle.calls[0];
-        expect(call.maxOutputTokens).toBe(16);
+        expect(call.maxOutputTokens).toBe(24);
+    });
+    it("parses the fourth word as the output deliverable (Phase 2b)", async () => {
+        const handle = installMockModel({ fixture: { stream: textOnlyStream("debug,concise,task,code") } });
+        cleanup = handle.uninstall;
+        const factory = (() => handle.model);
+        const classify = createLlmClassifier(factory, "deepseek-v4-flash");
+        const result = await classify("fix the crash in src/auth/login.ts");
+        expect(result?.taskType).toBe("debug");
+        expect(result?.deliverableKind).toBe("code");
+    });
+    it("recovers the deliverable position-independently and defaults to null when absent", async () => {
+        const reportHandle = installMockModel({ fixture: { stream: textOnlyStream("analyze,concise,task,report") } });
+        cleanup = reportHandle.uninstall;
+        const reportClassify = createLlmClassifier((() => reportHandle.model), "deepseek-v4-flash");
+        expect((await reportClassify("list every env var the CLI reads"))?.deliverableKind).toBe("report");
+        reportHandle.uninstall();
+        // Model omits the 4th word → deliverableKind null (consumers fall back to regex).
+        const bareHandle = installMockModel({ fixture: { stream: textOnlyStream("debug,concise") } });
+        cleanup = bareHandle.uninstall;
+        const bareClassify = createLlmClassifier((() => bareHandle.model), "deepseek-v4-flash");
+        expect((await bareClassify("fix it"))?.deliverableKind).toBeNull();
     });
 });
 //# sourceMappingURL=llm-classify.test.js.map

package/dist/src/pil/agent-operating-contract.d.ts

@@ -37,7 +37,7 @@
  * one imperative line targeting that phase's most damaging failure mode. Kept
  * tight (primacy matters more than detail; tokens are the cost).
  */
-export declare const AGENT_OPERATING_CONTRACT = "[AGENT OPERATING CONTRACT \u2014 read first; applies to every step]\n\n1. BEFORE ACTING: do only what was asked. Never assume scope or facts \u2014 if ambiguous, ask or use defaults; never invent requirements.\n2. READING: base statements on what you read/ran THIS turn. Do not infer contents of files you did not open.\n3. EXECUTING: smallest correct change; never widen scope or mask failures (no `|| true`, skipped tests, or swallowed catch).\n4. WHEN UNSURE: verify and cross-check BEFORE concluding. Bugs need a reproduction; reading code is not proof.\n5. REPORTING: answer ONLY what was asked. Every fact or file:line MUST come from this turn; else label \"unverified\"; do not guess. Synthesize evidence gracefully \u2014 do NOT dump massive verbatim tool outputs into the final answer. Cite concise file:line references. Never claim a build/test ran, or describe edits, you did not actually do this turn; if a check can't run, fix it or say so \u2014 don't imply success.\n\n6. LANGUAGE: Reply in user's detected language for final output. Internal reasoning, tools, and code remain in English.\n\n7. ANTI-M\u00D9 / COMPACTION: After seeing \"[pre-compaction warning at step...\" or \"[context compacted at step...\", decide if you need full prior tool results. Emit PRESERVE_FULL_CONTEXT for full veto this turn, or the lighter KEEP_TOOL_IDS: id1,id2 (ids from prior stub \"(id=...)\") to protect only high-value results (read_file/grep on src/PLAN/error etc are auto-protected). Use the ee_query tool with \"tool-artifact id=XXX\" for on-demand full re-hydrate of elided ones. Self-check \"task finished?\" / \"compacted yet?\". Use EE checkpoints.\n\n[END CONTRACT \u2014 instructions follow]";
+export declare const AGENT_OPERATING_CONTRACT = "[AGENT OPERATING CONTRACT \u2014 read first; applies to every step]\n\n1. BEFORE ACTING: do only what was asked. Never assume scope or facts \u2014 if ambiguous, ask or use defaults; never invent requirements.\n2. READING: base statements on what you read/ran THIS turn. Do not infer contents of files you did not open.\n3. EXECUTING: smallest correct change; never widen scope or mask failures (no `|| true`, skipped tests, or swallowed catch).\n4. WHEN UNSURE: verify and cross-check BEFORE concluding. Bugs need a reproduction; reading code is not proof.\n5. REPORTING: answer ONLY what was asked. Every fact or file:line MUST come from this turn; else label \"unverified\"; do not guess. Synthesize evidence gracefully \u2014 do NOT dump massive verbatim tool outputs into the final answer. Cite concise file:line references. Never claim a build/test ran, or describe edits, you did not actually do this turn; if a check can't run, fix it or say so \u2014 don't imply success.\n\n6. LANGUAGE: Reply in user's detected language for final output. Internal reasoning, tools, and code remain in English.\n\n7. ANTI-M\u00D9 / COMPACTION: After seeing \"[pre-compaction warning at step...\" or \"[context compacted at step...\", decide if you need full prior tool results. Emit PRESERVE_FULL_CONTEXT for full veto this turn, or the lighter KEEP_TOOL_IDS: id1,id2 (ids from prior stub \"(id=...)\") to protect only high-value results (read_file/grep on src/PLAN/error etc are auto-protected). Use the ee_query tool with \"tool-artifact id=XXX\" for on-demand full re-hydrate of elided ones. Self-check \"task finished?\" / \"compacted yet?\". Use EE checkpoints.\n\n8. GIT SAFETY: never push on red \u2014 run the check, await its result in a SEPARATE step, confirm 0 failures, then push. Never `git add -A`/`commit -a`; stage explicitly so secrets (.env, .muonroi-cli/, keys) aren't committed. Never `--no-verify`.\n\n[END CONTRACT \u2014 instructions follow]";
 export interface ContractSectionOptions {
     /** Chitchat turns carry no tools and make no factual claims — skip the contract. */
     chitchat?: boolean;

package/dist/src/pil/agent-operating-contract.js

@@ -49,6 +49,8 @@ export const AGENT_OPERATING_CONTRACT = `[AGENT OPERATING CONTRACT — read firs
 
 7. ANTI-MÙ / COMPACTION: After seeing "[pre-compaction warning at step..." or "[context compacted at step...", decide if you need full prior tool results. Emit PRESERVE_FULL_CONTEXT for full veto this turn, or the lighter KEEP_TOOL_IDS: id1,id2 (ids from prior stub "(id=...)") to protect only high-value results (read_file/grep on src/PLAN/error etc are auto-protected). Use the ee_query tool with "tool-artifact id=XXX" for on-demand full re-hydrate of elided ones. Self-check "task finished?" / "compacted yet?". Use EE checkpoints.
 
+8. GIT SAFETY: never push on red — run the check, await its result in a SEPARATE step, confirm 0 failures, then push. Never \`git add -A\`/\`commit -a\`; stage explicitly so secrets (.env, .muonroi-cli/, keys) aren't committed. Never \`--no-verify\`.
+
 [END CONTRACT — instructions follow]`;
 /**
  * Build the contract block for insertion at the front of the system prompt.

package/dist/src/pil/agent-operating-contract.test.js

@@ -42,8 +42,13 @@ describe("AGENT_OPERATING_CONTRACT", () => {
         expect(AGENT_OPERATING_CONTRACT).toMatch(/AGENT OPERATING CONTRACT/i);
         expect(AGENT_OPERATING_CONTRACT).toMatch(/END CONTRACT/i);
     });
-    it("stays compact (under 1800 chars) to preserve attention budget on every turn (anti-mù section added)", () => {
-        expect(AGENT_OPERATING_CONTRACT.length).toBeLessThan(1800);
+    it("carries the git-safety rule (never push on red; no broad git add of secrets)", () => {
+        expect(AGENT_OPERATING_CONTRACT).toMatch(/GIT SAFETY/i);
+        expect(AGENT_OPERATING_CONTRACT).toMatch(/push on red|never push/i);
+        expect(AGENT_OPERATING_CONTRACT).toMatch(/git add -A|stage explicitly/i);
+    });
+    it("stays compact (under 1900 chars) to preserve attention budget on every turn (git-safety rule added)", () => {
+        expect(AGENT_OPERATING_CONTRACT.length).toBeLessThan(1900);
     });
 });
 describe("buildContractSection", () => {

package/dist/src/pil/cheap-model-playbook.js

@@ -25,41 +25,41 @@
  * Wrapped with the `[CRITICAL TOOL-USE RULES ...]` marker so the model knows
  * to treat these as overrides to anything that follows.
  */
-export const CHEAP_MODEL_PLAYBOOK = `[CRITICAL TOOL-USE RULES — read before invoking any tool; these override defaults that follow]
-
-1. Bash output is AUTOMATICALLY cached. Every \`bash\` call returns a \`run_id\`
-   (e.g. \`bash-1\`) you can re-query via \`bash_output_get(run_id, mode=tail|head|grep|lines)\`.
-   - When you want only the last N lines: do NOT pipe \`| tail -N\`. Run the
-     bare command, then call \`bash_output_get(run_id, mode=tail, lines=N)\`.
-   - Same for \`| head\`, \`| grep PATTERN\`, \`> file\`. Pipes/redirects HIDE
-     the full output from the cache; \`bash_output_get\` reads from the cache
-     without re-running.
-   - This applies to EVERY bash call, not just retries.
-   - To VIEW a file use \`read_file\` (start_line/end_line) — never sed/cat a
-     file. \`bash_output_get\` is for COMMAND output, not files.
-
-2. Before reading more than 3 files to understand a topic, delegate to
-   \`task(agent="explore")\`. The sub-agent returns a compressed summary;
-   you save reading tokens.
-
-3. Use the \`grep\` tool (ripgrep) for content search — NOT \`bash\` with
-   \`grep\` / \`find\` piped.
-
-4. When a tool returns \`ERROR: ...\`, do NOT retry the identical call.
-   Pick a different tool, change inputs meaningfully, or stop and report.
-
-5. Fix the ROOT CAUSE, never mask a failure to make it "pass"
-   (\`continue-on-error\`, swallowed try/catch, skipped/deleted test, \`|| true\`).
-   If a step fails from a missing secret/config, make it CONDITIONAL (skip when
-   absent) so it still runs when present — do NOT blanket-ignore it.
-
-6. For a build / CI / test failure, read the ACTUAL failure log or stack trace
-   BEFORE hypothesizing — fix the real error, not a guess from source alone.
-
-7. ANTI-MÙ / COMPACTION (for long sessions): On pre-warn or "[context compacted at step...", emit PRESERVE_FULL_CONTEXT (full veto) or lighter KEEP_TOOL_IDS: id1,id2 (from stub id=) to protect specific high-value results. read_file/grep/lsp/bash on src/PLAN/error are auto-kept (idea 1). Use ee.query tool with "tool-artifact id=XXX" for on-demand full. Self-check "task finished?" / "compacted yet?". Use EE checkpoints.
-
-[END CRITICAL TOOL-USE RULES — your regular instructions begin below]
-
+export const CHEAP_MODEL_PLAYBOOK = `[CRITICAL TOOL-USE RULES — read before invoking any tool; these override defaults that follow]
+
+1. Bash output is AUTOMATICALLY cached. Every \`bash\` call returns a \`run_id\`
+   (e.g. \`bash-1\`) you can re-query via \`bash_output_get(run_id, mode=tail|head|grep|lines)\`.
+   - When you want only the last N lines: do NOT pipe \`| tail -N\`. Run the
+     bare command, then call \`bash_output_get(run_id, mode=tail, lines=N)\`.
+   - Same for \`| head\`, \`| grep PATTERN\`, \`> file\`. Pipes/redirects HIDE
+     the full output from the cache; \`bash_output_get\` reads from the cache
+     without re-running.
+   - This applies to EVERY bash call, not just retries.
+   - To VIEW a file use \`read_file\` (start_line/end_line) — never sed/cat a
+     file. \`bash_output_get\` is for COMMAND output, not files.
+
+2. Before reading more than 3 files to understand a topic, delegate to
+   \`task(agent="explore")\`. The sub-agent returns a compressed summary;
+   you save reading tokens.
+
+3. Use the \`grep\` tool (ripgrep) for content search — NOT \`bash\` with
+   \`grep\` / \`find\` piped.
+
+4. When a tool returns \`ERROR: ...\`, do NOT retry the identical call.
+   Pick a different tool, change inputs meaningfully, or stop and report.
+
+5. Fix the ROOT CAUSE, never mask a failure to make it "pass"
+   (\`continue-on-error\`, swallowed try/catch, skipped/deleted test, \`|| true\`).
+   If a step fails from a missing secret/config, make it CONDITIONAL (skip when
+   absent) so it still runs when present — do NOT blanket-ignore it.
+
+6. For a build / CI / test failure, read the ACTUAL failure log or stack trace
+   BEFORE hypothesizing — fix the real error, not a guess from source alone.
+
+7. ANTI-MÙ / COMPACTION (for long sessions): On pre-warn or "[context compacted at step...", emit PRESERVE_FULL_CONTEXT (full veto) or lighter KEEP_TOOL_IDS: id1,id2 (from stub id=) to protect specific high-value results. read_file/grep/lsp/bash on src/PLAN/error are auto-kept (idea 1). Use ee.query tool with "tool-artifact id=XXX" for on-demand full. Self-check "task finished?" / "compacted yet?". Use EE checkpoints.
+
+[END CRITICAL TOOL-USE RULES — your regular instructions begin below]
+
 `;
 /**
  * Predicate gating playbook injection.