@ilya-lesikov/pi-pi 0.6.0 → 0.7.0

package/extensions/orchestrator/phases/review.test.ts

@@ -0,0 +1,54 @@
+import { describe, it, expect } from "vitest";
+import { reviewSystemPrompt } from "./review.js";
+
+describe("reviewSystemPrompt apply_feedback wording", () => {
+  it("autonomous plan/implement mandates re-calling pp_phase_complete and does not tell the agent to wait for the user", () => {
+    const prompt = reviewSystemPrompt("/tmp/task", 1, "plan", "autonomous");
+    expect(prompt).toContain("pp_phase_complete");
+    expect(prompt).not.toContain("Present the synthesis to the user");
+    expect(prompt).not.toContain("A new review pass will begin");
+  });
+
+  it("guided plan/implement keeps the user-facing synthesis behavior", () => {
+    const prompt = reviewSystemPrompt("/tmp/task", 1, "plan", "guided");
+    expect(prompt).toContain("Present the synthesis to the user");
+    expect(prompt).not.toContain("Call pp_phase_complete again to finalize");
+  });
+
+  it("brainstorm prompt is unaffected by mode (never instructs pp_phase_complete)", () => {
+    const auto = reviewSystemPrompt("/tmp/task", 1, "brainstorm", "autonomous");
+    const guided = reviewSystemPrompt("/tmp/task", 1, "brainstorm", "guided");
+    expect(auto).toBe(guided);
+    expect(auto).toContain("BRAINSTORM REVIEW CYCLE");
+    expect(auto).not.toContain("pp_phase_complete");
+  });
+
+  it("autonomous plan feedback folds into a new synthesized plan, not a separate fix-plan file", () => {
+    // Re-review and the transition read only the latest `*synthesized*` plan, so
+    // plan-phase feedback must land there.
+    const plan = reviewSystemPrompt("/tmp/task", 1, "plan", "autonomous");
+    expect(plan).toContain("synthesized.md");
+    expect(plan).not.toContain("do NOT modify the original synthesized plan");
+
+    // The implement phase keeps the fix-plan/implement pattern (synthesized plan
+    // is code guidance there, not the reviewed artifact).
+    const impl = reviewSystemPrompt("/tmp/task", 1, "implement", "autonomous");
+    expect(impl).toContain("Create a fix plan");
+    expect(impl).toContain("Implement the fixes");
+  });
+
+  it("points each phase at the directory its reviewers actually write to", () => {
+    // plan reviewers write to plan-reviews (planning.ts) and outputs load from
+    // plan-reviews (context.ts); the prompt must match, not code-reviews.
+    const plan = reviewSystemPrompt("/tmp/task", 1, "plan", "autonomous");
+    expect(plan).toContain("plan-reviews/");
+    expect(plan).not.toContain("code-reviews/");
+
+    const impl = reviewSystemPrompt("/tmp/task", 1, "implement", "autonomous");
+    expect(impl).toContain("code-reviews/");
+    expect(impl).not.toContain("plan-reviews/");
+
+    const brainstorm = reviewSystemPrompt("/tmp/task", 1, "brainstorm", "guided");
+    expect(brainstorm).toContain("brainstorm-reviews/");
+  });
+});

package/extensions/orchestrator/phases/review.ts

@@ -12,8 +12,14 @@ function isEnabled(value: { enabled?: boolean } | undefined): boolean {
   return value?.enabled !== false;
 }
 
-export function reviewSystemPrompt(taskDir: string, pass: number, phase?: string): string {
-  const reviewsDir = phase === "brainstorm" ? join(taskDir, "brainstorm-reviews") : join(taskDir, "code-reviews");
+export function reviewSystemPrompt(taskDir: string, pass: number, phase?: string, mode?: "guided" | "autonomous"): string {
+  // Each phase writes/loads its review outputs in a distinct directory:
+  // brainstorm -> brainstorm-reviews, plan -> plan-reviews, everything else
+  // (implement/review) -> code-reviews. The apply_feedback prompt must point the
+  // agent at the SAME directory the reviewers wrote to (see planning.ts /
+  // context.ts), otherwise it synthesizes against the wrong (empty) directory.
+  const reviewsDirName = phase === "brainstorm" ? "brainstorm-reviews" : phase === "plan" ? "plan-reviews" : "code-reviews";
+  const reviewsDir = join(taskDir, reviewsDirName);
   const plansDir = join(taskDir, "plans");
 
   if (phase === "brainstorm") {
@@ -42,26 +48,62 @@ export function reviewSystemPrompt(taskDir: string, pass: number, phase?: string
     ].join("\n");
   }
 
+  // In autonomous plan/implement the phase does NOT complete until the agent
+  // re-calls pp_phase_complete: that is what finalizes the pass and advances the
+  // phase. Telling it to "present to the user" and wait causes the plan phase to
+  // stall after applying feedback. Guided mode keeps the user-facing behavior.
+  const finalStep =
+    mode === "autonomous"
+      ? "3. Call pp_phase_complete again to finalize this review pass. The phase is NOT complete until you do — do NOT stop or wait for the user"
+      : "3. Present the synthesis to the user";
+
+  // In the plan phase the synthesized plan IS the reviewed artifact: re-review
+  // and the phase transition both read only the LATEST `*synthesized*` file (see
+  // getLatestSynthesizedPlan). So autonomous plan feedback must be folded into a
+  // new synthesized plan, NOT a separate fix-plan file (which those readers
+  // ignore). In the implement phase the synthesized plan is code guidance, so
+  // the fix-plan/implement/afterImplement pattern is correct there.
+  const tail =
+    mode === "autonomous"
+      ? phase === "plan"
+        ? [
+            "",
+            "If the reviewers require changes:",
+            `1. Fold the required changes into a NEW synthesized plan at ${plansDir}/<timestamp>_synthesized.md (re-review and the phase transition read only the latest \`*synthesized*\` plan, so the fixes MUST land there — do not write them to a separate fix-plan file)`,
+            "2. Then call pp_phase_complete again — the extension will start a new review pass or advance the phase as appropriate. Do NOT wait for the user.",
+          ]
+        : [
+            "",
+            "If changes are needed:",
+            `1. Create a fix plan at ${plansDir}/<timestamp>_<description>.md (do NOT modify the original synthesized plan)`,
+            "2. Implement the fixes",
+            "3. Run afterImplement commands",
+            "4. Then call pp_phase_complete again — the extension will start a new review pass or advance the phase as appropriate. Do NOT wait for the user.",
+          ]
+      : [
+          "",
+          "If changes are needed:",
+          `1. Create a fix plan at ${plansDir}/<timestamp>_<description>.md (do NOT modify the original synthesized plan)`,
+          "2. Implement the fixes",
+          "3. Run afterImplement commands",
+          "4. A new review pass will begin",
+        ];
+
   return [
     `[PI-PI — REVIEW CYCLE (pass ${pass})]`,
     "",
-    "Code reviewer outputs are ready.",
+    "Reviewer outputs are ready.",
     `Read them from ${reviewsDir}/, synthesize feedback, and implement fixes if needed.`,
     "",
-    "You are a SYNTHESIZER: merge the reviewer outputs. Do NOT write your own code review from scratch.",
-    "- Do NOT create the code-reviews/ directory yourself — the extension manages it.",
+    "You are a SYNTHESIZER: merge the reviewer outputs. Do NOT write your own review from scratch.",
+    `- Do NOT create the ${reviewsDirName}/ directory yourself — the extension manages it.`,
     "- Do NOT call plannotator_submit_plan.",
     "",
     "# Your job (in this order):",
     `1. Read ALL reviewer outputs from ${reviewsDir}/`,
     `2. Synthesize into ${reviewsDir}/<timestamp>_final_pass-${pass}.md`,
-    "3. Present the synthesis to the user",
-    "",
-    "If changes are needed:",
-    `1. Create a fix plan at ${plansDir}/<timestamp>_<description>.md (do NOT modify the original synthesized plan)`,
-    "2. Implement the fixes",
-    "3. Run afterImplement commands",
-    "4. A new review pass will begin",
+    finalStep,
+    ...tail,
   ].join("\n");
 }
 

package/extensions/orchestrator/pp-menu.test.ts

@@ -3,10 +3,83 @@ import { join } from "path";
 import { getDefaultConfig, GLOBAL_CONFIG_PATH, parseDuration } from "./config.js";
 import * as configModule from "./config.js";
 import * as flantInfra from "./flant-infra.js";
-import { formatDuration, formatSourceTags, getConfigSourceInfo, pickMaxReviewPasses } from "./pp-menu.js";
+import { formatDuration, formatSourceTags, getConfigSourceInfo, pickMaxReviewPasses, showUsage } from "./pp-menu.js";
+import { createUsageTracker } from "./usage-tracker.js";
+
+const USAGE_TRACKER_SYMBOL = Symbol.for("pi-pi:usage-tracker");
 
 afterEach(() => {
   vi.restoreAllMocks();
+  delete (globalThis as any)[USAGE_TRACKER_SYMBOL];
+});
+
+function renderUsage(tracker: ReturnType<typeof createUsageTracker>): string {
+  (globalThis as any)[USAGE_TRACKER_SYMBOL] = tracker;
+  let captured = "";
+  showUsage({ ui: { notify: (text: string) => { captured = text; } } });
+  return captured;
+}
+
+describe("showUsage subscription rendering", () => {
+  it("labels subscription model and agent rows and excludes their dollars", () => {
+    const tracker = createUsageTracker();
+    tracker.recordTurn("openai/gpt-5", "openai", 100, 50, 0, 0, 0.4, false);
+    tracker.recordTurn("sub/claude-opus-4-6", "pp-flant-anthropic-sub", 200, 100, 0, 0, 9.9, false);
+    tracker.recordSubagentCompletion({ input: 30, output: 10 } as any, 5.0, {
+      description: "Explore", agentType: "explore", modelId: "sub/claude-haiku-4-5",
+    });
+
+    const out = renderUsage(tracker);
+
+    expect(out).toContain("sub/claude-opus-4-6:");
+    expect(out).toContain("subscription");
+    expect(out).toContain("explore");
+    expect(out).toContain("$0.40");
+    expect(out).not.toContain("$9.90");
+    expect(out).not.toContain("$5.00");
+    expect(out).toContain("Cost: $0.40");
+  });
+
+  it("shows the four-bucket input breakdown and processed-input totals", () => {
+    const tracker = createUsageTracker();
+    // uncached 84, output 27k, cacheRead 1000, cacheWrite 200 on a sub model.
+    tracker.recordTurn("sub/claude-opus-4-8", "pp-flant-anthropic-sub", 84, 27000, 1000, 200, 0, true);
+    // A cache-using subagent, to exercise the inline By-agent breakdown.
+    tracker.recordSubagentCompletion({ input: 10, output: 500, cacheRead: 300, cacheWrite: 90 } as any, 0, {
+      description: "Explore", agentType: "explore", modelId: "sub/claude-haiku-4-5",
+    });
+
+    const out = renderUsage(tracker);
+
+    // Total "Input" is the processed input across main + subagent:
+    // main 84+1000+200=1284, subagent 10+300+90=400 → 1684 → "1.7k".
+    expect(out).toContain("Input: 1.7k tokens");
+    expect(out).toContain("uncached:    94");    // 84 + 10
+    expect(out).toContain("cache read:  1.3k");  // 1000 + 300
+    expect(out).toContain("cache write: 290");   // 200 + 90
+    expect(out).toContain("Output: 28k tokens"); // 27000 + 500
+    // Hit rate = 1300 / 1684 = 77% (rounded).
+    expect(out).toContain("⚡77% hit rate");
+    // Cost is always shown, even at $0.00 for subscription sessions.
+    expect(out).toContain("Cost: $0.00");
+    // Per-model row is a one-liner: processed input (↑1.3k) with an inline
+    // uncached / cache read / cache write breakdown.
+    expect(out).toContain("sub/claude-opus-4-8: ↑1.3k (u84 r1.0k w200)");
+    // By-agent row carries the same inline breakdown (400 processed = 10+300+90).
+    expect(out).toContain("explore: ↑400 (u10 r300 w90)");
+  });
+
+  it("does not inflate paid model share when a subscription model is present", () => {
+    const tracker = createUsageTracker();
+    tracker.recordTurn("openai/gpt-5", "openai", 100, 0, 0, 0, 0.5, false);
+    tracker.recordTurn("sub/claude-opus-4-6", "pp-flant-anthropic-sub", 100, 0, 0, 0, 2.0, false);
+
+    const out = renderUsage(tracker);
+
+    expect(out).toContain("openai/gpt-5:");
+    expect(out).toContain("$0.50");
+    expect(out).not.toContain("$0.25");
+  });
 });
 
 describe("settings helpers", () => {