@koda-sl/baker-bridge 0.164.0-dev.d9feffe34 → 0.168.0-dev.d78500250

package/dist/hono/bridge-job.test.js

@@ -1,5 +1,5 @@
 import { readFileSync } from "node:fs";
-import { mkdir, mkdtemp, rm, utimes, writeFile } from "node:fs/promises";
+import { mkdir, mkdtemp, readFile, rm, utimes, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { afterEach, describe, expect, it, vi } from "vitest";
@@ -9,11 +9,19 @@ vi.mock("./env.ts", () => ({
 // The skills allowlist is read off the workspace, so the project root has to be
 // steerable to a fixture directory. Defaults to the real root for every other test.
 const projectRootMock = vi.hoisted(() => ({ value: process.cwd() }));
+// Where the per-company instructions import is written; kept out of the real ~/.baker.
+const instructionsDirMock = vi.hoisted(() => ({
+    value: `${process.env.TMPDIR ?? "/tmp"}/baker-job-instructions-${process.pid}`,
+}));
 vi.mock("./runtime-files.ts", async (importOriginal) => {
     const actual = await importOriginal();
     return {
         ...actual,
-        bridgeRuntimeFiles: { ...actual.bridgeRuntimeFiles, projectRoot: () => projectRootMock.value },
+        bridgeRuntimeFiles: {
+            ...actual.bridgeRuntimeFiles,
+            projectRoot: () => projectRootMock.value,
+            instructionsDir: () => instructionsDirMock.value,
+        },
     };
 });
 import { AVATAR_CASTING_TOOL_NAME } from "./avatar-casting.js";
@@ -254,6 +262,8 @@ function createExecutor(options) {
         ...(options.catchUpWithMain ? { catchUpWithMain: options.catchUpWithMain } : {}),
         ...(options.keepCliInStep ? { keepCliInStep: options.keepCliInStep } : {}),
         ...(options.holdsNoWork ? { holdsNoWork: options.holdsNoWork } : {}),
+        // Never the network: a turn nobody judges runs at the default effort.
+        effortJudge: options.effortJudge ?? (async () => null),
     });
 }
 function memoryInFlightTurns(initial) {
@@ -1006,6 +1016,39 @@ describe("BridgeJobExecutor", () => {
             ],
         });
     });
+    // Tool results are most of what a turn writes to the prompt cache, and nothing
+    // recorded their size: the turn now reports it next to its model usage.
+    it("reports how much each tool's results put in the context this turn", async () => {
+        const executor = createExecutor({
+            queryResults: [
+                scriptedMessages({
+                    type: "assistant",
+                    parent_tool_use_id: null,
+                    uuid: "assistant_tool",
+                    session_id: "session_tools_size",
+                    message: {
+                        id: "msg_tool",
+                        role: "assistant",
+                        content: [{ type: "tool_use", id: "tool_1", name: "Bash", input: { command: "baker ads meta ads" } }],
+                    },
+                }, {
+                    type: "user",
+                    parent_tool_use_id: null,
+                    uuid: "user_tool_result",
+                    session_id: "session_tools_size",
+                    message: {
+                        role: "user",
+                        content: [{ type: "tool_result", tool_use_id: "tool_1", content: "r".repeat(800) }],
+                    },
+                }, resultMessage(0.1, "session_tools_size")),
+            ],
+        });
+        const outcome = await executor.execute({ threadId: "thread_tools_size", turnId: "turn_tools_size", prompt: "How are the ads?", history: history([]) }, activeControls());
+        expect(outcome).toMatchObject({
+            kind: "completed",
+            toolResults: [{ tool: "baker ads meta", agent: "main", calls: 1, bytes: 800, tokens: 200 }],
+        });
+    });
     // OpenRouter answers under its own slug. A CLI resuming a transcript that names
     // the slug drops every earlier thinking block from its first request, which
     // breaks the prompt cache right after the system prompt (BAKER-331).
@@ -1288,16 +1331,19 @@ describe("BridgeJobExecutor", () => {
             "x-session-id: thread_env",
         ].join("\n"));
     });
-    // OpenRouter has no Bedrock endpoint for a per-message effort, so while the CLI
-    // sends one every Opus 5.5 request skips Bedrock and bills OpenRouter credits.
-    it("keeps the agent's model requests servable by Bedrock", async () => {
+    // With the capability on, the CLI restates the effort in its mid-conversation
+    // system message, so a turn on another effort level keeps the conversation's
+    // prompt cache. Without it, a changed top-level effort re-writes everything
+    // after the first tool exchange. Pinned on rather than left to the CLI's own
+    // default, which a remote feature gate can switch off.
+    it("turns on per-turn effort, so a turn on another effort level keeps the prompt cache", async () => {
         const queryInputs = [];
         const executor = createExecutor({
             queryInputs,
             queryResults: [scriptedMessages(resultMessage(0, "session_bedrock"))],
         });
         await executor.execute({ threadId: "thread_bedrock", turnId: "turn_bedrock", prompt: "Build a landing", history: history([]) }, activeControls());
-        expect(queryInputs[0]?.options.env).toMatchObject({ CLAUDE_CODE_MODEL_CAPABILITIES: "-per_turn_effort" });
+        expect(queryInputs[0]?.options.env).toMatchObject({ CLAUDE_CODE_MODEL_CAPABILITIES: "per_turn_effort" });
     });
     it("silences the CLI task reminders and ANSI colour in the agent subprocess", async () => {
         const queryInputs = [];
@@ -1397,6 +1443,62 @@ describe("BridgeJobExecutor", () => {
         expect(disallowed).not.toContain("Task");
         expect(disallowed).not.toContain("SendMessage");
     });
+    // Developer tools a marketer's chat never calls (0 calls across 358
+    // production chats, Oct 2026). Each definition rides every request a session
+    // makes — DesignSync's alone is ~9k characters — whenever the CLI does not
+    // defer it.
+    it("removes the developer tools a marketer's chat never calls", async () => {
+        const queryInputs = [];
+        const executor = createExecutor({
+            queryInputs,
+            queryResults: [scriptedMessages(resultMessage(0, "session_dev_tools"))],
+        });
+        await executor.execute({ threadId: "thread_dev_tools", turnId: "turn_dev_tools", prompt: "Build a landing", history: history([]) }, activeControls());
+        const disallowed = queryInputs[0]?.options.disallowedTools ?? [];
+        expect(disallowed).toEqual(expect.arrayContaining(["DesignSync", "ReportFindings", "NotebookEdit"]));
+    });
+    // The scaffold's CLAUDE.md is the same text for every company on a scaffold
+    // version, but loaded as a memory file it lands in the first user message with
+    // no cache breakpoint after it, so every chat wrote its ~24k tokens cold. In the
+    // system prompt it ends on a breakpoint and one chat reads another's copy.
+    it("sends the workspace's shared instructions as the system prompt and loads COMPANY.md on its own", async () => {
+        const workspace = await mkdtemp(join(tmpdir(), "baker-job-instructions-"));
+        await writeFile(join(workspace, "CLAUDE.md"), "# Project Instructions\n\n## Client Context\n\nOne client per repo. Everything known about them:\n\n@COMPANY.md\n\n## Rules\n\nStage, don't ask.\n");
+        projectRootMock.value = workspace;
+        const queryInputs = [];
+        const executor = createExecutor({
+            queryInputs,
+            queryResults: [scriptedMessages(resultMessage(0, "session_instructions"))],
+        });
+        await executor.execute({ threadId: "thread_instructions", turnId: "turn_instructions", prompt: "Build a landing", history: history([]) }, activeControls());
+        const options = queryInputs[0]?.options;
+        expect(options.systemPrompt).toEqual({
+            type: "preset",
+            preset: "claude_code",
+            append: "# Project Instructions\n\n## Rules\n\nStage, don't ask.\n",
+            snapshot: true,
+        });
+        expect(options.settings).toMatchObject({ claudeMdExcludes: [join(workspace, "CLAUDE.md")] });
+        expect(options.additionalDirectories).toEqual([instructionsDirMock.value]);
+        expect(options.env?.CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD).toBe("1");
+        expect(await readFile(join(instructionsDirMock.value, "CLAUDE.md"), "utf8")).toBe(`@${join(workspace, "COMPANY.md")}\n`);
+    });
+    it("lets the CLI load a CLAUDE.md it cannot split, exactly as before", async () => {
+        const workspace = await mkdtemp(join(tmpdir(), "baker-job-instructions-old-"));
+        await writeFile(join(workspace, "CLAUDE.md"), "# Project Instructions\n\nNo company import.\n");
+        projectRootMock.value = workspace;
+        const queryInputs = [];
+        const executor = createExecutor({
+            queryInputs,
+            queryResults: [scriptedMessages(resultMessage(0, "session_instructions_old"))],
+        });
+        await executor.execute({ threadId: "thread_instructions_old", turnId: "turn_instructions_old", prompt: "Hi", history: history([]) }, activeControls());
+        const options = queryInputs[0]?.options;
+        expect(options.systemPrompt).toBeUndefined();
+        expect(options.additionalDirectories).toBeUndefined();
+        expect(options.settings ?? {}).not.toHaveProperty("claudeMdExcludes");
+        expect(options.env?.CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD).toBeUndefined();
+    });
     it("allowlists the workspace's own skills so the CLI's built-in developer skills stay out of the session", async () => {
         const workspace = await mkdtemp(join(tmpdir(), "baker-job-skills-"));
         for (const skill of ["landing", "brand-build", "docx"]) {
@@ -1575,11 +1677,10 @@ describe("BridgeJobExecutor", () => {
         await executor.execute({ threadId: "thread_nobeta", turnId: "turn_nobeta", prompt: "Hi", history: history([]) }, activeControls());
         expect(queryInputs[0]?.options).not.toHaveProperty("betas");
     });
-    // Effort governs thinking depth, preamble and tool-call count, and the API
-    // default is `high`. A silent drift back to the default costs money twice:
-    // more tokens per turn, and — because the resolved effort is rendered into
-    // the prompt — a cache-breakpoint invalidation on the turn it changes.
-    it("sends a single constant effort so the prompt cache is never invalidated by it", async () => {
+    // Effort governs thinking depth, preamble and tool-call count. A turn nobody
+    // could judge runs at Opus 5.5's own default, set explicitly so an SDK or
+    // model default change cannot move it.
+    it("runs a turn Jev could not judge at medium effort", async () => {
         const queryInputs = [];
         const executor = createExecutor({
             queryInputs,
@@ -1592,8 +1693,57 @@ describe("BridgeJobExecutor", () => {
         await executor.execute({ threadId: "thread_effort", turnId: "turn_effort_1", prompt: "Hi", history: history([]) }, activeControls());
         await executor.execute({ threadId: "thread_effort", turnId: "turn_effort_2", prompt: "Again", history: history([]) }, activeControls());
         expect(queryInputs[0]?.options.effort).toBe("medium");
-        // Constant across turns — a per-turn value is what would break the cache.
-        expect(queryInputs[1]?.options.effort).toBe(queryInputs[0]?.options.effort);
+        expect(queryInputs[1]?.options.effort).toBe("medium");
+    });
+    it("runs a turn Jev clearly reads as small at low effort, and reports why", async () => {
+        const queryInputs = [];
+        const asked = [];
+        const executor = createExecutor({
+            queryInputs,
+            effortJudge: async (questions) => {
+                asked.push(...questions);
+                return [0.92, 0.05];
+            },
+            queryResults: [scriptedMessages(resultMessage(0, "session_effort_low"))],
+        });
+        const outcome = await executor.execute({
+            threadId: "thread_effort_low",
+            turnId: "turn_effort_low",
+            prompt: "¿Cuál fue el CPL de ayer?",
+            history: history([
+                { role: "user", text: "Revisa la campaña de verano", createdAt: 1 },
+                { role: "assistant", text: "Listo: la campaña de verano va bien.", createdAt: 2 },
+            ]),
+        }, activeControls());
+        expect(queryInputs[0]?.options.effort).toBe("low");
+        expect(asked[0]).toMatchObject({
+            instructions: { message: "¿Cuál fue el CPL de ayer?", previous_reply: "Listo: la campaña de verano va bien." },
+        });
+        expect(outcome).toMatchObject({
+            kind: "completed",
+            effortDecision: { effort: "low", reason: "small", small: 0.92, complex: 0.05 },
+        });
+    });
+    it("keeps a plan-mode turn at medium without asking Jev", async () => {
+        const queryInputs = [];
+        let asked = false;
+        const executor = createExecutor({
+            queryInputs,
+            effortJudge: async () => {
+                asked = true;
+                return [0.99, 0];
+            },
+            queryResults: [scriptedMessages(resultMessage(0, "session_effort_plan"))],
+        });
+        await executor.execute({
+            threadId: "thread_effort_plan",
+            turnId: "turn_effort_plan",
+            prompt: "Plan Q4",
+            planMode: true,
+            history: history([]),
+        }, activeControls());
+        expect(queryInputs[0]?.options.effort).toBe("medium");
+        expect(asked).toBe(false);
     });
     it("pins the auto-compact window to the context window it resolved for the model", async () => {
         const queryInputs = [];
@@ -3880,6 +4030,8 @@ describe("BridgeJobExecutor stall recovery", () => {
             kind: "completed",
             costUsd: 0.05,
             model: "claude-test",
+            // The tool result the stalled query had already received still counts.
+            toolResults: [expect.objectContaining({ calls: 1 })],
             cleanBoundary: expect.objectContaining({ sessionId: "session_a" }),
         });
         expect(queryInputs).toHaveLength(2);
@@ -4618,6 +4770,9 @@ describe("BridgeJobExecutor turn end with background tasks", () => {
     });
 });
 describe("BridgeJobExecutor review mode", () => {
+    afterEach(() => {
+        projectRootMock.value = process.cwd();
+    });
     const review = {
         reviewId: "rev_1",
         apiKey: "bk_review_key",
@@ -4675,6 +4830,21 @@ describe("BridgeJobExecutor review mode", () => {
         await executor.execute(reviewJob(), activeControls());
         expect(queryInputs[0]?.options.settings).toMatchObject({ permissions: { ask: ["Bash"] } });
     });
+    // The brief has the reviewer read the rulebook as the Session had it
+    // (`git show <sha>:CLAUDE.md`); the working tree's copy may already be newer.
+    // Auto-loading it put ~24k tokens of the marketer's instructions in front of
+    // every review request, written cold and re-read on each one.
+    it("does not load the workspace's CLAUDE.md into the reviewer, as a memory file or as the system prompt", async () => {
+        const workspace = await mkdtemp(join(tmpdir(), "baker-review-instructions-"));
+        await writeFile(join(workspace, "CLAUDE.md"), "# Project Instructions\n\n## Client Context\n\nOne client per repo. Everything known about them:\n\n@COMPANY.md\n\n## Rules\n");
+        projectRootMock.value = workspace;
+        const queryInputs = [];
+        const executor = createExecutor({ queryInputs });
+        await executor.execute(reviewJob(), activeControls());
+        expect(queryInputs[0]?.options.settings).toMatchObject({ claudeMdExcludes: ["**/CLAUDE.md"] });
+        expect(queryInputs[0]?.options.systemPrompt).toBeUndefined();
+        expect(queryInputs[0]?.options.additionalDirectories).toBeUndefined();
+    });
     it("removes every editing tool on top of the list an ordinary turn is denied", async () => {
         const queryInputs = [];
         const executor = createExecutor({ queryInputs });
@@ -4738,6 +4908,7 @@ describe("BridgeJobExecutor review mode", () => {
         expect(options.env?.BAKER_DEBUG_LOG).toBeUndefined();
         expect(options.settings ?? {}).not.toHaveProperty("disableAllHooks");
         expect(options.settings ?? {}).not.toHaveProperty("permissions");
+        expect(options.settings ?? {}).not.toHaveProperty("claudeMdExcludes");
     });
     // BAKER-3462: after an English skill body, the next line went out in English to a Spanish chat.
     const spanishTurn = "Prepara cuatro anuncios nuevos para la campaña de verano y súbelos hoy mismo";