refactor(ai-judge): split remaining hotspots and cover CRAP branches

- stage-ify judgeAuthorize into auditEnrollment, runPreflightGates, prepareModelCall, and enforceAndEmit - extract tallyJoinedRows and tallyAttributable from computeMetrics, removing three dead locals - extract validateVerdictResponse from requestStructuredVerdict - extract collectUserTexts and charBudgetStart from buildConversationEvidence - table-drive corpus-replay parseArgs and split main into resolveReplayModel, selectCorpusCases, and replayCorpus - split analyzer cli main into loadReviewEvents, withinWindow, loadAuditEnrolled, and printReport with a run-as-script guard - add 81 tests covering parseEntry, extractBashCommandEvidence, validateVerdictResponse, forcedToolChoice, classifyGit dry-run paths, CLI arg/window/report rendering, and the infra-failure result path - add @vitest/coverage-istanbul for exact per-function CRAP scoring via fallow health --coverage
2026-10-05 11:52:55 +08:00 · 2026-08-22 01:24:50 +08:00
parent 0903119a46
commit c479f51469
18 changed files with 2756 additions and 455 deletions
@@ -0,0 +1,292 @@
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { spawnSync } from "node:child_process";
+import { afterEach, describe, expect, it } from "vitest";
+
+/**
+ * CLI-level tests for arg parsing (`parseArgs`), the `--after`/`--before`
+ * time window (`withinWindow`), and report rendering (`printReport`).
+ * Invalid-arg cases assert the exit contract (stderr + exit 1) that the
+ * analyze-shadow CLI documents.
+ */
+
+const dirs: string[] = [];
+
+function tmp(): string {
+    const dir = mkdtempSync(join(tmpdir(), "ai-judge-cli-args-"));
+    dirs.push(dir);
+    return dir;
+}
+
+afterEach(() => {
+    for (const dir of dirs.splice(0)) {
+        rmSync(dir, { recursive: true, force: true });
+    }
+});
+
+interface RunResult {
+    readonly stdout: string;
+    readonly stderr: string;
+    readonly status: number;
+}
+
+function runExpect(args: readonly string[]): RunResult {
+    const cli = join(import.meta.dirname, "..", "..", "src", "analyzer", "cli.ts");
+    const r = spawnSync("npx", ["tsx", cli, ...args], {
+        encoding: "utf-8",
+        stdio: ["ignore", "pipe", "pipe"],
+    });
+    return {
+        stdout: r.stdout ?? "",
+        stderr: r.stderr ?? "",
+        status: r.status ?? -1,
+    };
+}
+
+function reviewLog(dir: string, lines: readonly object[]): string {
+    const path = join(dir, "review.jsonl");
+    writeFileSync(
+        path,
+        lines.map((l) => JSON.stringify(l)).join("\n") + "\n",
+    );
+    return path;
+}
+
+describe("analyze-shadow CLI — argument parsing (parseArgs)", () => {
+    it("prints usage and exits 0 for --help and -h", () => {
+        for (const flag of ["--help", "-h"]) {
+            const r = runExpect([flag]);
+            expect(r.status).toBe(0);
+            expect(r.stdout.length + r.stderr.length).toBeGreaterThan(0);
+        }
+    });
+
+    it("rejects missing input path with exit 1", () => {
+        const r = runExpect([]);
+        expect(r.status).toBe(1);
+        expect(r.stderr).toContain("missing input path");
+    });
+
+    it("rejects multiple input paths with exit 1", () => {
+        const dir = tmp();
+        const a = reviewLog(dir, []);
+        const b = reviewLog(dir, []).replace("review", "review2");
+        const r = runExpect([a, b]);
+        expect(r.status).toBe(1);
+        expect(r.stderr).toContain("multiple input paths");
+    });
+
+    it("rejects unknown options with exit 1", () => {
+        const dir = tmp();
+        const path = reviewLog(dir, []);
+        const r = runExpect([path, "--bogus"]);
+        expect(r.status).toBe(1);
+        expect(r.stderr).toContain("unknown option: --bogus");
+    });
+
+    it("rejects --after/--before/--audit without a value", () => {
+        const dir = tmp();
+        const path = reviewLog(dir, []);
+        for (const flag of ["--after", "--before", "--audit"]) {
+            const r = runExpect([path, flag]);
+            expect(r.status).toBe(1);
+            expect(r.stderr).toContain(`${flag} requires a value`);
+        }
+    });
+
+    it("rejects invalid --after/--before timestamps", () => {
+        const dir = tmp();
+        const path = reviewLog(dir, []);
+        for (const flag of ["--after", "--before"]) {
+            const r = runExpect([path, flag, "not-a-timestamp"]);
+            expect(r.status).toBe(1);
+            expect(r.stderr).toContain(`invalid ${flag} timestamp`);
+        }
+    });
+});
+
+describe("analyze-shadow CLI — time window (withinWindow)", () => {
+    const events = (ts: string) => [
+        {
+            timestamp: ts,
+            event: "authorizer_chain_resolved",
+            requestId: "req-1",
+            links: ["ai-bash-judge"],
+        },
+        {
+            timestamp: ts,
+            event: "ai_bash_judge.result",
+            requestId: "req-1",
+            resultKind: "judgment",
+            verdict: "allow",
+        },
+        {
+            timestamp: ts,
+            event: "permission_request.approved",
+            requestId: "req-1",
+            resolution: "approved",
+        },
+    ];
+
+    it("keeps events inside --after..--before and drops events outside", () => {
+        const dir = tmp();
+        const path = reviewLog(dir, [
+            ...events("2026-08-18T00:00:05Z"), // inside
+            ...events("2026-08-18T00:00:01Z").map((e, i) => ({ ...e, requestId: `req-old-${i}` })), // before
+            ...events("2026-08-18T23:00:00Z").map((e, i) => ({ ...e, requestId: `req-new-${i}` })), // after
+        ]);
+        const r = runExpect([
+            path,
+            "--after", "2026-08-18T00:00:02Z",
+            "--before", "2026-08-18T12:00:00Z",
+        ]);
+        expect(r.status).toBe(0);
+        expect(r.stdout).toContain("enrollments (N): 1");
+        expect(r.stdout).toContain("joined rows: 1");
+    });
+
+    it("keeps events with a missing timestamp regardless of the window", () => {
+        const dir = tmp();
+        // First event has no timestamp: survives any window.
+        const path = reviewLog(dir, [
+            {
+                event: "authorizer_chain_resolved",
+                requestId: "req-1",
+                links: ["ai-bash-judge"],
+            },
+            ...events("2026-08-18T00:00:05Z").slice(1),
+        ]);
+        const r = runExpect([
+            path,
+            "--after", "2026-08-18T00:00:02Z",
+        ]);
+        expect(r.status).toBe(0);
+        expect(r.stdout).toContain("enrollments (N): 1");
+    });
+});
+
+describe("analyze-shadow CLI — report rendering (printReport)", () => {
+    function joinedLog(dir: string): string {
+        return reviewLog(dir, [
+            {
+                timestamp: "2026-08-18T00:00:01Z",
+                event: "authorizer_chain_resolved",
+                requestId: "req-1",
+                links: ["ai-bash-judge"],
+            },
+            {
+                timestamp: "2026-08-18T00:00:02Z",
+                event: "ai_bash_judge.result",
+                requestId: "req-1",
+                resultKind: "judgment",
+                verdict: "allow",
+                judgeLatencyMs: 5,
+                modelLatencyMs: 3,
+            },
+            {
+                timestamp: "2026-08-18T00:00:03Z",
+                event: "permission_request.approved",
+                requestId: "req-1",
+                resolution: "approved",
+            },
+            {
+                timestamp: "2026-08-18T00:00:04Z",
+                event: "authorizer_chain_resolved",
+                requestId: "req-2",
+                links: ["ai-bash-judge"],
+            },
+            {
+                timestamp: "2026-08-18T00:00:05Z",
+                event: "ai_bash_judge.result",
+                requestId: "req-2",
+                resultKind: "judgment",
+                verdict: "deny",
+                code: null,
+                judgeLatencyMs: 7,
+                modelLatencyMs: 4,
+            },
+            {
+                timestamp: "2026-08-18T00:00:06Z",
+                event: "permission_request.approved",
+                requestId: "req-2",
+                resolution: "approved",
+            },
+        ]);
+    }
+
+    it("renders the comparison matrix and counters for mixed judgment rows", () => {
+        const dir = tmp();
+        const r = runExpect([joinedLog(dir)]);
+        expect(r.status).toBe(0);
+        // Header + source echo.
+        expect(r.stdout).toContain("AI Bash Judge — Shadow diagnostic report");
+        expect(r.stdout).toContain("grade: DIAGNOSTIC");
+        expect(r.stdout).toContain("source:");
+        // Denominator + join counts.
+        expect(r.stdout).toContain("enrollments (N): 2");
+        expect(r.stdout).toContain("joined rows: 2");
+        expect(r.stdout).toContain("joined judgments: 2");
+        // Matrix rows for both verdicts (allow|allow, deny|allow).
+        expect(r.stdout).toContain("allow|allow: 1");
+        expect(r.stdout).toContain("deny|allow: 1");
+        // Conservative deny counter incremented by deny|allow.
+        expect(r.stdout).toContain("conservative: deny 1, defer 0");
+        // Latency lines rendered with p50/p95/max/missing.
+        expect(r.stdout).toMatch(/judge latency: p50=\d+ms p95=\d+ms/);
+        expect(r.stdout).toMatch(/model latency: p50=\d+ms p95=\d+ms/);
+        // No --audit: the audit source line must be absent.
+        expect(r.stdout).not.toContain("(enrollment source, ADR 0006)");
+    });
+
+    it("renders an empty matrix placeholder and no quarantine section when clean", () => {
+        const dir = tmp();
+        const r = runExpect([joinedLog(dir)]);
+        expect(r.status).toBe(0);
+        expect(r.stdout).not.toContain("quarantined rows:");
+        // Build an empty log: no events at all — matrix stays empty.
+        const empty = reviewLog(dir, []);
+        const r2 = runExpect([empty]);
+        expect(r2.status).toBe(0);
+        expect(r2.stdout).toContain("comparison matrix [verdict|human]:");
+        expect(r2.stdout).toContain("(empty)");
+        expect(r2.stdout).toContain("N/A");
+    });
+
+    it("renders the quarantine section with category counts", () => {
+        const dir = tmp();
+        // Two results for one request → duplicate_result quarantine.
+        const path = reviewLog(dir, [
+            {
+                timestamp: "2026-08-18T00:00:01Z",
+                event: "authorizer_chain_resolved",
+                requestId: "req-1",
+                links: ["ai-bash-judge"],
+            },
+            {
+                timestamp: "2026-08-18T00:00:02Z",
+                event: "ai_bash_judge.result",
+                requestId: "req-1",
+                resultKind: "judgment",
+                verdict: "allow",
+            },
+            {
+                timestamp: "2026-08-18T00:00:03Z",
+                event: "ai_bash_judge.result",
+                requestId: "req-1",
+                resultKind: "judgment",
+                verdict: "deny",
+            },
+            {
+                timestamp: "2026-08-18T00:00:04Z",
+                event: "permission_request.approved",
+                requestId: "req-1",
+                resolution: "approved",
+            },
+        ]);
+        const r = runExpect([path]);
+        expect(r.status).toBe(0);
+        expect(r.stdout).toContain("quarantined rows:");
+        expect(r.stdout).toContain("duplicate_result: 1");
+    });
+});
@@ -0,0 +1,214 @@
+import { describe, expect, it, vi } from "vitest";
+import type { ReviewEvent } from "../../src/analyzer/analyze";
+import { analyzeShadowReviewLog } from "../../src/analyzer/analyze";
+import {
+    parseArgs,
+    parseTimestampOption,
+    printReport,
+    withinWindow,
+} from "../../src/analyzer/cli";
+
+/**
+ * In-process unit tests for the analyze-shadow CLI's pure helpers:
+ * argument parsing, the time-window filter, and report rendering. The
+ * end-to-end contract (file IO, dual-log mode, exit codes) lives in
+ * cli-args.test.ts / cli-audit.test.ts; these tests cover branch-level
+ * behavior without spawning tsx subprocesses.
+ */
+
+describe("parseArgs", () => {
+    it("parses the positional path with no options", () => {
+        expect(parseArgs(["node", "cli.ts", "review.jsonl"])).toEqual({
+            path: "review.jsonl",
+            after: null,
+            before: null,
+            audit: null,
+        });
+    });
+
+    it("parses --after, --before, and --audit together", () => {
+        const r = parseArgs([
+            "node",
+            "cli.ts",
+            "review.jsonl",
+            "--after", "2026-08-18T00:00:00Z",
+            "--before", "2026-08-19T00:00:00Z",
+            "--audit", "audit.jsonl",
+        ]);
+        expect(r).toEqual({
+            path: "review.jsonl",
+            after: new Date("2026-08-18T00:00:00Z"),
+            before: new Date("2026-08-19T00:00:00Z"),
+            audit: "audit.jsonl",
+        });
+    });
+
+    it("returns the usage error for --help and -h", () => {
+        expect(parseArgs(["node", "cli.ts", "--help"])).toEqual({
+            error: expect.stringContaining("usage: analyze-shadow") as unknown,
+        });
+        expect(parseArgs(["node", "cli.ts", "-h"])).toHaveProperty("error");
+    });
+
+    it.each([
+        ["missing input path", [] as const],
+        ["multiple input paths", ["a.jsonl", "b.jsonl"] as const],
+        ["unknown option", ["a.jsonl", "--bogus"] as const],
+        ["--audit without value", ["a.jsonl", "--audit"] as const],
+        ["--after without value", ["a.jsonl", "--after"] as const],
+        ["invalid --after timestamp", ["a.jsonl", "--after", "nope"] as const],
+        ["invalid --before timestamp", ["a.jsonl", "--before", "nope"] as const],
+    ])("errors on %s", (_name, argv) => {
+        const r = parseArgs(["node", "cli.ts", ...argv]);
+        expect("error" in r && typeof r.error === "string").toBe(true);
+    });
+});
+
+describe("parseTimestampOption", () => {
+    it("requires a value", () => {
+        expect(parseTimestampOption("--after", undefined)).toEqual({
+            error: "--after requires a value",
+        });
+    });
+
+    it("rejects unparseable timestamps", () => {
+        expect(parseTimestampOption("--before", "yesterday")).toEqual({
+            error: "invalid --before timestamp: yesterday",
+        });
+    });
+
+    it("accepts an ISO instant", () => {
+        expect(parseTimestampOption("--after", "2026-08-18T00:00:00Z")).toEqual({
+            date: new Date("2026-08-18T00:00:00Z"),
+        });
+    });
+});
+
+describe("withinWindow", () => {
+    const evt = (timestamp?: string): ReviewEvent =>
+        ({ event: "e", ...(timestamp === undefined ? {} : { timestamp }) }) as ReviewEvent;
+
+    it("keeps events without a timestamp in any window", () => {
+        const window = {
+            after: new Date("2026-08-18T12:00:00Z"),
+            before: new Date("2026-08-18T13:00:00Z"),
+        };
+        expect(withinWindow(evt(), window)).toBe(true);
+    });
+
+    it("keeps events inside the window and drops those outside", () => {
+        const window = {
+            after: new Date("2026-08-18T12:00:00Z"),
+            before: new Date("2026-08-18T13:00:00Z"),
+        };
+        expect(withinWindow(evt("2026-08-18T12:30:00Z"), window)).toBe(true);
+        expect(withinWindow(evt("2026-08-18T11:59:00Z"), window)).toBe(false);
+        expect(withinWindow(evt("2026-08-18T13:01:00Z"), window)).toBe(false);
+    });
+
+    it("applies each bound independently", () => {
+        expect(
+            withinWindow(evt("2026-08-18T10:00:00Z"), { after: new Date("2026-08-18T09:00:00Z"), before: null }),
+        ).toBe(true);
+        expect(
+            withinWindow(evt("2026-08-18T10:00:00Z"), { after: null, before: new Date("2026-08-18T09:00:00Z") }),
+        ).toBe(false);
+    });
+
+    it("keeps events with an unparseable timestamp regardless of bounds", () => {
+        const window = {
+            after: new Date("2026-08-18T12:00:00Z"),
+            before: null,
+        };
+        // Unparseable -> NaN time -> neither bound applies.
+        expect(withinWindow(evt("not-a-date"), window)).toBe(true);
+    });
+});
+
+describe("printReport", () => {
+    function capture(fn: () => void): string {
+        const chunks: string[] = [];
+        const spy = vi
+            .spyOn(process.stdout, "write")
+            .mockImplementation(((s: unknown) => {
+                chunks.push(String(s));
+                return true;
+            }) as never);
+        try {
+            fn();
+        } finally {
+            spy.mockRestore();
+        }
+        return chunks.join("");
+    }
+
+    const parsed = {
+        path: "review.jsonl",
+        after: null,
+        before: null,
+        audit: null,
+    };
+
+    function analyze(rows: readonly ReviewEvent[]): ReturnType<
+        typeof import("../../src/analyzer/analyze").analyzeShadowReviewLog
+    > {
+        return analyzeShadowReviewLog(rows);
+    }
+
+    const joinedEvents: ReviewEvent[] = [
+        { event: "authorizer_chain_resolved", requestId: "r1", links: ["ai-bash-judge"] },
+        {
+            event: "ai_bash_judge.result",
+            requestId: "r1",
+            resultKind: "judgment",
+            verdict: "allow",
+            judgeLatencyMs: 5,
+            modelLatencyMs: 3,
+        },
+        { event: "permission_request.approved", requestId: "r1", resolution: "approved" },
+    ];
+
+    it("renders header, counts, matrix, and latency lines", () => {
+        const { enrollments, metrics } = analyze(joinedEvents);
+        const out = capture(() => printReport(parsed, enrollments, metrics));
+        expect(out).toContain("AI Bash Judge — Shadow diagnostic report");
+        expect(out).toContain("grade: DIAGNOSTIC");
+        expect(out).toContain("enrollments (N): 1");
+        expect(out).toContain("allow|allow: 1");
+        expect(out).toMatch(/judge latency: p50=5ms/);
+        expect(out).not.toContain("audit:");
+        expect(out).not.toContain("quarantined rows:");
+    });
+
+    it("renders the audit source line in dual-log mode", () => {
+        const { enrollments, metrics } = analyze([]);
+        const out = capture(() =>
+            printReport(
+                { ...parsed, audit: "audit.jsonl" },
+                enrollments,
+                metrics,
+            ),
+        );
+        expect(out).toContain("audit: audit.jsonl (enrollment source, ADR 0006)");
+    });
+
+    it("renders the empty-matrix placeholder for an empty log", () => {
+        const { enrollments, metrics } = analyze([]);
+        const out = capture(() => printReport(parsed, enrollments, metrics));
+        expect(out).toContain("(empty)");
+        expect(out).toContain("N/A");
+    });
+
+    it("renders quarantine categories when integrity faults exist", () => {
+        const duplicate: ReviewEvent[] = [
+            { event: "authorizer_chain_resolved", requestId: "r1", links: ["ai-bash-judge"] },
+            { event: "ai_bash_judge.result", requestId: "r1", resultKind: "judgment", verdict: "allow" },
+            { event: "ai_bash_judge.result", requestId: "r1", resultKind: "judgment", verdict: "deny" },
+            { event: "permission_request.approved", requestId: "r1", resolution: "approved" },
+        ];
+        const { enrollments, metrics } = analyze(duplicate);
+        const out = capture(() => printReport(parsed, enrollments, metrics));
+        expect(out).toContain("quarantined rows:");
+        expect(out).toContain("duplicate_result: 1");
+    });
+});
@@ -83,6 +83,61 @@ describe("loadModelCatalog", () => {
        expect(diagnostics[0]?.key).toBe("file");
        expect(String(diagnostics[0]?.problem)).toMatch(/whole catalog degraded/i);
    });
+
+    it("rejects an entry per invalid field (every parseEntry branch)", () => {
+        const cases: readonly [name: string, entry: unknown][] = [
+            ["non-object entry", "not-an-object"],
+            ["null entry", null],
+            ["array entry", [VALID_ENTRY]],
+            ["empty provider", { ...VALID_ENTRY, provider: "  " }],
+            ["empty model", { ...VALID_ENTRY, model: "" }],
+            ["empty api", { ...VALID_ENTRY, api: "" }],
+            ["unknown status", { ...VALID_ENTRY, status: "experimental" }],
+            ["empty promptVersion", { ...VALID_ENTRY, promptVersion: "" }],
+            ["empty corpusVersion", { ...VALID_ENTRY, corpusVersion: "" }],
+            ["empty testedAt", { ...VALID_ENTRY, testedAt: "" }],
+            ["non-integer corpusCases", { ...VALID_ENTRY, corpusCases: 2.5 }],
+            ["zero corpusCases", { ...VALID_ENTRY, corpusCases: 0 }],
+            ["non-integer matched", { ...VALID_ENTRY, matched: 1.5 }],
+            ["non-integer infrastructureFailures", { ...VALID_ENTRY, infrastructureFailures: 0.5 }],
+            ["latencyMs not an object", { ...VALID_ENTRY, latencyMs: 3000 }],
+            ["latencyMs wrong type", { ...VALID_ENTRY, latencyMs: { p50: "3000", p95: null, max: null } }],
+            ["latencyMs missing key", { ...VALID_ENTRY, latencyMs: { p50: 1, p95: null } }],
+            ["empty reportPath", { ...VALID_ENTRY, reportPath: "" }],
+            ["blank notes", { ...VALID_ENTRY, notes: "   " }],
+        ];
+        for (const [name, entry] of cases) {
+            const { catalog, diagnostics } = run(depsWith(JSON.stringify({
+                version: 1,
+                entries: [entry],
+            })));
+            expect(catalog?.entries, name).toEqual([]);
+            expect(diagnostics.map((d) => d.key), name).toContain("file");
+        }
+    });
+
+    it("accepts latency nulls and omits notes only when absent", () => {
+        const withNulls = {
+            ...VALID_ENTRY,
+            latencyMs: { p50: null, p95: null, max: null },
+        };
+        const { catalog } = run(depsWith(JSON.stringify({
+            version: 1,
+            entries: [withNulls],
+        })));
+        expect(catalog?.entries).toHaveLength(1);
+        expect(catalog?.entries[0]?.latencyMs).toEqual({ p50: null, p95: null, max: null });
+        expect("notes" in (catalog?.entries[0] ?? {})).toBe(false);
+
+        const withNotes = { ...VALID_ENTRY, notes: "qualifying corpus replay report" };
+        const r2 = run(depsWith(JSON.stringify({
+            version: 1,
+            entries: [withNotes],
+        })));
+        expect(r2.catalog?.entries[0]).toMatchObject({
+            notes: "qualifying corpus replay report",
+        });
+    });
 });

 describe("classifyModel", () => {
@@ -0,0 +1,205 @@
+import { describe, expect, it, vi } from "vitest";
+import type { AssistantMessage, Context, Model } from "@earendil-works/pi-ai";
+import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
+import {
+    DEFAULT_TIMEOUT_MS,
+    MIN_TIMEOUT_MS,
+    MAX_TIMEOUT_MS,
+} from "../src/model";
+import { createModelAvailability } from "../src/model";
+import {
+    applyTimeoutOption,
+    replayCorpus,
+    selectCorpusCases,
+    validateCliState,
+    type CliOptions,
+} from "../tools/corpus-replay";
+
+/**
+ * In-process tests for the corpus-replay harness helpers: option
+ * application/validation, case selection, and the replay loop driven by a
+ * stub model availability (no registry, no network). The CLI exit contract
+ * lives in corpus-replay-cli.test.ts.
+ */
+
+function state(overrides: Partial<CliOptions> = {}): { -readonly [K in keyof CliOptions]: CliOptions[K] } {
+    return {
+        provider: "p",
+        model: "m",
+        timeoutMs: DEFAULT_TIMEOUT_MS,
+        cases: null,
+        out: null,
+        strict: false,
+        ...overrides,
+    };
+}
+
+describe("applyTimeoutOption", () => {
+    it.each([
+        [MIN_TIMEOUT_MS, null],
+        [MAX_TIMEOUT_MS, null],
+        [DEFAULT_TIMEOUT_MS, null],
+    ])("accepts the boundary value %i", (value, error) => {
+        const s = state();
+        expect(applyTimeoutOption(s, String(value))).toBe(error);
+        expect(s.timeoutMs).toBe(value);
+    });
+
+    it.each([
+        [String(MIN_TIMEOUT_MS - 1)],
+        [String(MAX_TIMEOUT_MS + 1)],
+        ["12.5"],
+        ["abc"],
+    ])("rejects %s with the documented range error", (value) => {
+        const s = state();
+        const r = applyTimeoutOption(s, value);
+        expect(r).toMatch(/--timeout-ms must be an integer in \[\d+, \d+\]/);
+        expect(s.timeoutMs).toBe(DEFAULT_TIMEOUT_MS);
+    });
+
+    it("rejects a missing value with the range error (the usage check lives in parseArgs)", () => {
+        expect(applyTimeoutOption(state(), undefined)).toMatch(/--timeout-ms must be an integer/);
+    });
+});
+
+describe("validateCliState", () => {
+    it("requires both provider and model", () => {
+        expect(validateCliState(state({ provider: "" }))).toMatch(/^usage:/);
+        expect(validateCliState(state({ model: "" }))).toMatch(/^usage:/);
+    });
+
+    it("rejects --strict combined with a --case subset", () => {
+        expect(
+            validateCliState(state({ strict: true, cases: new Set(["a"]) })),
+        ).toContain("--strict requires the full corpus");
+    });
+
+    it("accepts a complete valid state", () => {
+        expect(validateCliState(state())).toBeNull();
+    });
+});
+
+describe("selectCorpusCases", () => {
+    it("selects the full corpus when no case filter is given", () => {
+        const r = selectCorpusCases(null);
+        expect("selected" in r && r.selected.length).toBeGreaterThan(10);
+    });
+
+    it("selects exactly the requested known ids", () => {
+        const r = selectCorpusCases(new Set(["requested-clean", "extra-push"]));
+        expect("selected" in r ? r.selected.map((c) => c.id) : []).toEqual([
+            "requested-clean",
+            "extra-push",
+        ]);
+    });
+
+    it("errors on unknown ids without selecting anything", () => {
+        const r = selectCorpusCases(new Set(["requested-clean", "no-such-case"]));
+        expect("error" in r && r.error).toContain("unknown case ids: no-such-case");
+    });
+});
+
+describe("replayCorpus", () => {
+    const stderrSpy = () => vi.spyOn(process.stderr, "write").mockReturnValue(true);
+
+    function availabilityWith(
+        respond: () => AssistantMessage,
+    ): ReturnType<typeof createModelAvailability> {
+        const model = {
+            id: "m",
+            provider: "p",
+            api: "openai-codex-responses",
+        } as Model<any>;
+        const registry = {
+            complete: async (
+                _m: Model<any>,
+                _c: Context,
+                _o?: Record<string, unknown>,
+            ) => respond(),
+        } as unknown as ModelRegistry;
+        return createModelAvailability(model, registry);
+    }
+
+    function judgment(verdict: "allow" | "deny" | "defer"): AssistantMessage {
+        return {
+            role: "assistant",
+            content: [
+                {
+                    type: "toolCall",
+                    id: "call-1",
+                    name: "report_verdict",
+                    arguments: { verdict, reason: "stub reason" },
+                },
+            ],
+            api: "openai-codex-responses",
+            provider: "p",
+            model: "m",
+            stopReason: "toolUse",
+            usage: {
+                input: 10,
+                output: 8,
+                cacheRead: 0,
+                cacheWrite: 0,
+                totalTokens: 18,
+                cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+            },
+            timestamp: Date.now(),
+        } as AssistantMessage;
+    }
+
+    it("counts matches over the selected cases", async () => {
+        const spy = stderrSpy();
+        try {
+            // requested-clean expects allow; extra-push expects defer — one
+            // canned judgment cannot satisfy both, so assert row semantics
+            // rather than a perfect score.
+            const availability = availabilityWith(() => judgment("allow"));
+            const selected = selectCorpusCases(
+                new Set(["requested-clean", "extra-push"]),
+            );
+            if (!("selected" in selected)) throw new Error("unreachable");
+            const { rows, matched } = await replayCorpus(
+                selected.selected,
+                availability,
+                DEFAULT_TIMEOUT_MS,
+                new AbortController().signal,
+            );
+            expect(rows).toHaveLength(2);
+            expect(matched).toBe(1);
+            const clean = rows.find((r) => r.case === "requested-clean");
+            expect(clean).toMatchObject({ verdict: "allow", match: true });
+            const push = rows.find((r) => r.case === "extra-push");
+            expect(push).toMatchObject({ verdict: "allow", match: false });
+        } finally {
+            spy.mockRestore();
+        }
+    });
+
+    it("records infrastructure failures as unmatched rows", async () => {
+        const spy = stderrSpy();
+        try {
+            const availability = availabilityWith(() => {
+                const m = judgment("allow");
+                m.stopReason = "error";
+                return m;
+            });
+            const selected = selectCorpusCases(new Set(["requested-clean"]));
+            if (!("selected" in selected)) throw new Error("unreachable");
+            const { rows, matched } = await replayCorpus(
+                selected.selected,
+                availability,
+                DEFAULT_TIMEOUT_MS,
+                new AbortController().signal,
+            );
+            expect(rows).toHaveLength(1);
+            expect(rows[0]).toMatchObject({
+                verdict: null,
+                resultKind: "infrastructure_failure",
+                match: false,
+            });
+            expect(matched).toBe(0);
+        } finally {
+            spy.mockRestore();
+        }
+    });
+});
@@ -28,9 +28,14 @@ describe("classifyHighRisk — data_loss", () => {
    it.each([
        "git clean -nxd",
        "git clean -nd",
+        "git clean -nxfd",
+        "git clean -f -x -n",
+        "git clean -fd",
+        "git clean -xn",
        "git reset --soft HEAD~1",
        "git checkout main",
        "git checkout -- file.txt",
+        "git checkout . file.txt",
        "git restore file.txt",
        "rm -rf build/",
        "rm -rf ./dist",
@@ -542,6 +542,82 @@ describe("AI judge lifecycle", () => {
        harness.shutdown();
    });

+    it("records a provider failure as an infrastructure_failure row and defers", async () => {
+        let authorize: Authorizer["authorize"] | undefined;
+        const service = {
+            registerAuthorizer: vi.fn((_name, callback) => {
+                authorize = callback;
+                return vi.fn();
+            }),
+            checkPermission: vi.fn(),
+            getToolPermission: vi.fn(),
+        } as unknown as PermissionsService;
+        publishPermissionsService(service);
+        publishedService = service;
+
+        const complete = vi.fn(
+            async () => {
+                throw new Error("provider 500");
+            },
+        );
+        const ctx = {
+            hasUI: true,
+            sessionManager: fakeSessionManager(),
+            model: {
+                id: "test-model",
+                provider: "test-provider",
+                api: "openai-codex-responses",
+            } as Model<any>,
+            modelRegistry: { complete },
+            ui: { notify: vi.fn() },
+        } as unknown as ExtensionContext;
+
+        const harness = createFakePi();
+        extension(harness.pi);
+        harness.start(ctx);
+        harness.ready();
+        expect(authorize).toBeDefined();
+
+        const reviews: Array<{
+            event: string;
+            details?: Record<string, unknown>;
+        }> = [];
+        const verdict = await authorize!(
+            ask(),
+            {
+                checkPermission: vi.fn(),
+                getToolPermission: vi.fn(),
+            },
+            {
+                review: (event, details) => reviews.push({ event, details }),
+                debug: vi.fn(),
+            },
+        );
+
+        expect(verdict).toEqual({ kind: "defer" });
+        expect(complete).toHaveBeenCalledTimes(1);
+        expect(reviews).toMatchObject([
+            {
+                event: "ai_bash_judge.result",
+                details: expect.objectContaining({
+                    resultKind: "infrastructure_failure",
+                    verdict: null,
+                    effectiveVerdict: "defer",
+                    modelCalled: true,
+                    code: "model_error",
+                    modelSource: "session",
+                    riskOverride: null,
+                    inputUsage: null,
+                    outputUsage: null,
+                }),
+            },
+        ]);
+        // The failure row must not leak the provider error text.
+        expect(JSON.stringify(reviews)).not.toContain("provider 500");
+
+        harness.shutdown();
+    });
+
    it("does not register from a headless child", () => {
        const service = {
            registerAuthorizer: vi.fn(),
@@ -1,6 +1,8 @@
 import { describe, expect, it, vi } from "vitest";
-import type { AssistantMessage, Context } from "@earendil-works/pi-ai";
+import type { AssistantMessage, Context, Model } from "@earendil-works/pi-ai";
+import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import {
+    createModelAvailability,
    requestStructuredVerdict,
    type ModelAvailability,
 } from "../src/model";
@@ -50,6 +52,18 @@ function ready(
    return { kind: "ready", metadata, complete };
 }

+/** A minimal valid report_verdict tool-call content block. */
+function verdictContent(): AssistantMessage["content"] {
+    return [
+        {
+            type: "toolCall",
+            id: "call-1",
+            name: "report_verdict",
+            arguments: { verdict: "allow", reason: "bounded" },
+        },
+    ];
+}
+
 const evidence = {
    fullCommand: "pnpm test && git push",
    triggeringUnit: "git push",
@@ -231,6 +245,137 @@ describe("requestStructuredVerdict", () => {
        });
    });

+    it("rejects zero and non-finite output usage", async () => {
+        const zeroUsage = await requestStructuredVerdict(
+            ready(async () =>
+                response(verdictContent(), 0),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(zeroUsage).toMatchObject({ code: "model_error" });
+
+        const nonFiniteUsage = await requestStructuredVerdict(
+            ready(async () =>
+                response(verdictContent(), Number.NaN),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(nonFiniteUsage).toMatchObject({ code: "model_error" });
+    });
+
+    it("rejects responses without exactly one report_verdict tool call", async () => {
+        const noCall = await requestStructuredVerdict(
+            ready(async () =>
+                response([{ type: "text", text: "I would allow it." } as never]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(noCall).toMatchObject({ code: "missing_tool_call" });
+
+        const wrongName = await requestStructuredVerdict(
+            ready(async () =>
+                response([
+                    {
+                        type: "toolCall",
+                        id: "call-1",
+                        name: "other_tool",
+                        arguments: { verdict: "allow", reason: "x" },
+                    },
+                ]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(wrongName).toMatchObject({ code: "missing_tool_call" });
+    });
+
+    it("rejects null and array tool-call arguments", async () => {
+        const nullArgs = await requestStructuredVerdict(
+            ready(async () =>
+                response([
+                    {
+                        type: "toolCall",
+                        id: "call-1",
+                        name: "report_verdict",
+                        arguments: null as unknown as Record<string, unknown>,
+                    },
+                ]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(nullArgs).toMatchObject({ code: "invalid_arguments" });
+
+        const arrayArgs = await requestStructuredVerdict(
+            ready(async () =>
+                response([
+                    {
+                        type: "toolCall",
+                        id: "call-1",
+                        name: "report_verdict",
+                        arguments: ["allow", "x"] as unknown as Record<string, unknown>,
+                    },
+                ]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(arrayArgs).toMatchObject({ code: "invalid_arguments" });
+
+        const missingKey = await requestStructuredVerdict(
+            ready(async () =>
+                response([
+                    {
+                        type: "toolCall",
+                        id: "call-1",
+                        name: "report_verdict",
+                        arguments: { verdict: "allow" },
+                    },
+                ]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(missingKey).toMatchObject({ code: "invalid_arguments" });
+    });
+
+    it("rejects non-string and blank-trimmed reasons", async () => {
+        const nonStringReason = await requestStructuredVerdict(
+            ready(async () =>
+                response([
+                    {
+                        type: "toolCall",
+                        id: "call-1",
+                        name: "report_verdict",
+                        arguments: { verdict: "defer", reason: 42 as unknown as string },
+                    },
+                ]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(nonStringReason).toMatchObject({ code: "invalid_reason" });
+
+        const blankReason = await requestStructuredVerdict(
+            ready(async () =>
+                response([
+                    {
+                        type: "toolCall",
+                        id: "call-1",
+                        name: "report_verdict",
+                        arguments: { verdict: "defer", reason: "   " },
+                    },
+                ]),
+            ),
+            evidence,
+            new AbortController().signal,
+        );
+        expect(blankReason).toMatchObject({ code: "invalid_reason" });
+    });
+
    it("maps the bounded deadline to timeout", async () => {
        const waiting = ready(
            async (_context, signal) =>
@@ -319,3 +464,92 @@ describe("requestStructuredVerdict", () => {
        expect(timedOut).toMatchObject({ code: "timeout" });
    });
 });
+
+describe("createModelAvailability — per-API forced tool choice and output cap", () => {
+    it("returns no_model for an undefined model", () => {
+        expect(createModelAvailability(undefined, registryStub())).toEqual({
+            kind: "no_model",
+        });
+    });
+
+    it("returns unsupported_api for an unknown api", () => {
+        const availability = createModelAvailability(
+            modelWithApi("mystery-api"),
+            registryStub(),
+        );
+        expect(availability).toMatchObject({
+            kind: "unsupported_api",
+            metadata: { api: "mystery-api" },
+        });
+    });
+
+    it.each([
+        // [api, expected toolChoice, enforces output cap]
+        ["anthropic-messages", { type: "tool", name: "report_verdict" }, true],
+        ["bedrock-converse-stream", { type: "tool", name: "report_verdict" }, true],
+        ["google-generative-ai", "any", true],
+        ["google-vertex", "any", true],
+        ["openai-completions", { type: "function", function: { name: "report_verdict" } }, true],
+        ["mistral-conversations", { type: "function", function: { name: "report_verdict" } }, true],
+        ["pi-messages", { type: "function", function: { name: "report_verdict" } }, true],
+        ["openai-responses", { type: "function", name: "report_verdict" }, true],
+        ["azure-openai-responses", { type: "function", name: "report_verdict" }, true],
+        // Codex supports required but not named choice, and no output cap.
+        ["openai-codex-responses", "required", false],
+    ])("maps %s to its forced tool choice and cap policy", async (api, toolChoice, cap) => {
+        const calls: Array<Record<string, unknown>> = [];
+        const registry = registryStub({
+            complete: (_model, _context, options) => {
+                calls.push(options as Record<string, unknown>);
+                return Promise.resolve(response(verdictContent()));
+            },
+        });
+        const availability = createModelAvailability(modelWithApi(api), registry);
+        expect(availability).toMatchObject({ kind: "ready", metadata: { api } });
+
+        if (availability.kind === "ready") {
+            await availability.complete({} as never, new AbortController().signal);
+        }
+        expect(calls[0]?.toolChoice).toEqual(toolChoice);
+        if (cap) {
+            expect(calls[0]?.maxTokens).toBeTypeOf("number");
+        } else {
+            expect(calls[0]?.maxTokens).toBeUndefined();
+        }
+    });
+
+    it("forwards retry and cache policy to the registry", async () => {
+        let seen: Record<string, unknown> | undefined;
+        const registry = registryStub({
+            complete: (_m, _c, options) => {
+                seen = options as Record<string, unknown>;
+                return Promise.resolve(response(verdictContent()));
+            },
+        });
+        const availability = createModelAvailability(
+            modelWithApi("anthropic-messages"),
+            registry,
+        );
+        if (availability.kind === "ready") {
+            await availability.complete({} as never, new AbortController().signal);
+        }
+        expect(seen).toMatchObject({ maxRetries: 0, cacheRetention: "none" });
+    });
+});
+
+function modelWithApi(api: string): Model<any> {
+    return {
+        provider: "test-provider",
+        id: "test-model",
+        api,
+    } as unknown as Model<any>;
+}
+
+function registryStub(overrides: Partial<ModelRegistry> = {}): ModelRegistry {
+    return {
+        complete: () => {
+            throw new Error("not called");
+        },
+        ...overrides,
+    } as unknown as ModelRegistry;
+}