codecartographer-pi 0.23.0 → 0.24.1

package/dist/core/broadside-verify.js

@@ -0,0 +1,433 @@
+// Broad-Side verification pass (#143): the hybrid the roadmap kept coming
+// back to. The batch sweep is cheap and reads files without being able to
+// look anything up, and its measured weakness is precision, not coverage —
+// on this repository the top twelve findings by severity were two real
+// defects and ten claims that a look at the guard, the caller, or the
+// tsconfig would have dismissed. So after collect, one sync-priced call per
+// finding, with three read-only tools confined to the repository, reads the
+// cited code and says whether the failure is reachable.
+//
+// Measured on 2026-09-13 (the #143 comparison run): twelve findings, the
+// rubric below, `google/gemini-3.7-flash` at low reasoning effort —
+// twelve-for-twelve agreement with a reviewer's ground truth, precision of
+// the confirmed set from 17% to 100%, both true findings kept, $0.13 in all
+// (about a cent a finding). The rubric mattered more than the model: a first
+// draft without the `not-a-defect` verdict "confirmed" two type casts that
+// every caller satisfies, because they were literally true of the code.
+//
+// Read-only on purpose. A tool-using pass that could mutate would need the
+// headless-agent retry rule (retry only before the first tool call); one
+// that only reads keeps the batch property that re-running is always safe.
+import { readdir, readFile, stat, writeFile } from "node:fs/promises";
+import { isAbsolute, join, relative, resolve } from "node:path";
+import { BROADSIDE_DIR, BROADSIDE_LENS_IDS, broadsideDirFor, defaultReasoningFor, isSlurpable, listRepoFiles, loadBroadsideState, loadSavedLensResults, parseLensJson, persistBroadsideRunMerging, } from "./broadside.js";
+export const BROADSIDE_CHAT_URL = "https://openrouter.ai/api/v1/chat/completions";
+/** How many findings `verify` reads by default, most severe first. */
+export const BROADSIDE_VERIFY_DEFAULT_TOP = 10;
+/** Tool calls one finding may spend before it must answer. */
+export const BROADSIDE_VERIFY_MAX_TOOL_CALLS = 8;
+/** The lenses whose findings carry a file:line and a claim to check. */
+export const BROADSIDE_VERIFIABLE_LENSES = ["defect", "security"];
+const READ_FILE_MAX_LINES = 200;
+const GREP_MAX_MATCHES = 40;
+const GREP_MAX_FILE_BYTES = 1_000_000;
+const TOOL_OUTPUT_MAX_CHARS = 12_000;
+const SEVERITY_RANK = { critical: 0, high: 1, medium: 2, low: 3 };
+/** The sync-priced id behind a `:batch` model id (`vendor/name:batch` → `vendor/name`). */
+export function syncModelFor(batchModel) {
+    return batchModel.replace(/:batch$/, "");
+}
+/** The findings a run's saved lens results carry, most severe first. */
+export function rankVerifiableFindings(stored) {
+    const out = [];
+    for (const result of stored) {
+        if (!BROADSIDE_VERIFIABLE_LENSES.includes(result.lensId) || result.truncated)
+            continue;
+        const parsed = parseLensJson(result.content);
+        for (const finding of parsed?.findings ?? []) {
+            const location = typeof finding.location === "string" ? finding.location : "";
+            const title = typeof finding.title === "string" ? finding.title : "";
+            if (!location || !title)
+                continue;
+            out.push({
+                lensId: result.lensId,
+                customId: result.customId,
+                severity: typeof finding.severity === "string" ? finding.severity.toLowerCase() : "unknown",
+                title,
+                location,
+                description: typeof finding.description === "string" ? finding.description : "",
+                pattern: typeof finding.pattern === "string" ? finding.pattern : typeof finding.category === "string" ? finding.category : "",
+            });
+        }
+    }
+    // Stable: severity, then lens order, then the order the lens listed them.
+    return out
+        .map((finding, order) => ({ finding, order }))
+        .sort((a, b) => (SEVERITY_RANK[a.finding.severity] ?? 9) - (SEVERITY_RANK[b.finding.severity] ?? 9)
+        || BROADSIDE_LENS_IDS.indexOf(a.finding.lensId) - BROADSIDE_LENS_IDS.indexOf(b.finding.lensId)
+        || a.order - b.order)
+        .map(({ finding }) => finding);
+}
+/**
+ * Three read-only tools over the repository's own file list — the same
+ * listing the lenses scan (tracked and untracked, ignore rules applied) minus
+ * everything {@link isSlurpable} keeps out of a lens: credential stores, build
+ * output, binaries. A path outside the repository, or one the listing does
+ * not contain, is an error the model sees, not a read.
+ */
+export async function createRepoReader(cwd) {
+    const { files } = await listRepoFiles(cwd);
+    const readable = new Set(files.filter(isSlurpable));
+    const confine = (path) => {
+        const abs = resolve(cwd, path);
+        const rel = relative(cwd, abs).split("\\").join("/");
+        if (!rel || rel.startsWith("..") || isAbsolute(rel))
+            throw new Error(`path is outside the repository: ${path}`);
+        return rel;
+    };
+    return {
+        async readFile(path, startLine, endLine) {
+            const rel = confine(path);
+            if (!readable.has(rel))
+                return `error: ${rel} is not a readable source file of this repository`;
+            const lines = (await readFile(join(cwd, rel), "utf8")).split("\n");
+            const start = Math.max(1, Math.floor(Number(startLine ?? 1)) || 1);
+            const requestedEnd = Math.floor(Number(endLine ?? start + READ_FILE_MAX_LINES - 1)) || start;
+            const end = Math.min(lines.length, requestedEnd, start + READ_FILE_MAX_LINES - 1);
+            if (start > lines.length)
+                return `error: ${rel} has ${lines.length} lines`;
+            return lines.slice(start - 1, end).map((line, i) => `${start + i}: ${line}`).join("\n");
+        },
+        async grep(pattern, pathPrefix) {
+            let regex;
+            try {
+                regex = new RegExp(pattern);
+            }
+            catch (error) {
+                return `error: invalid pattern (${error instanceof Error ? error.message : String(error)})`;
+            }
+            const prefix = pathPrefix ? confine(pathPrefix) : "";
+            const matches = [];
+            for (const rel of readable) {
+                if (prefix && rel !== prefix && !rel.startsWith(`${prefix}/`) && !rel.startsWith(prefix))
+                    continue;
+                let text;
+                try {
+                    if ((await stat(join(cwd, rel))).size > GREP_MAX_FILE_BYTES)
+                        continue;
+                    text = await readFile(join(cwd, rel), "utf8");
+                }
+                catch {
+                    continue;
+                }
+                const lines = text.split("\n");
+                for (let i = 0; i < lines.length && matches.length < GREP_MAX_MATCHES; i++) {
+                    if (regex.test(lines[i]))
+                        matches.push(`${rel}:${i + 1}: ${lines[i]}`);
+                }
+                if (matches.length >= GREP_MAX_MATCHES)
+                    break;
+            }
+            return matches.length > 0 ? matches.join("\n") : "(no matches)";
+        },
+        async listDir(path) {
+            const rel = path === "." || path === "" ? "" : confine(path);
+            try {
+                const entries = await readdir(join(cwd, rel), { withFileTypes: true });
+                return entries
+                    .filter((entry) => {
+                    const child = rel ? `${rel}/${entry.name}` : entry.name;
+                    return entry.isDirectory() ? [...readable].some((f) => f.startsWith(`${child}/`)) : readable.has(child);
+                })
+                    .map((entry) => (entry.isDirectory() ? `${entry.name}/` : entry.name))
+                    .sort()
+                    .join("\n") || "(empty)";
+            }
+            catch (error) {
+                return `error: ${error instanceof Error ? error.message : String(error)}`;
+            }
+        },
+    };
+}
+const TOOLS = [
+    { type: "function", function: { name: "read_file", description: "Read a range of lines (1-based, inclusive) from a source file of the repository; at most 200 lines per call.", parameters: { type: "object", properties: { path: { type: "string" }, start_line: { type: "integer" }, end_line: { type: "integer" } }, required: ["path"] } } },
+    { type: "function", function: { name: "grep", description: "Search the repository's source files for a regular expression; returns at most 40 matching lines as path:line: text.", parameters: { type: "object", properties: { pattern: { type: "string" }, path_prefix: { type: "string", description: "optional directory or file to search under" } }, required: ["pattern"] } } },
+    { type: "function", function: { name: "list_dir", description: "List a directory of the repository.", parameters: { type: "object", properties: { path: { type: "string" } }, required: ["path"] } } },
+];
+const VERDICT_SCHEMA = {
+    name: "broadside_verification",
+    strict: true,
+    schema: {
+        type: "object",
+        properties: {
+            verdict: { type: "string", enum: ["confirmed", "not-a-defect", "discarded", "unclear"] },
+            confidence: { type: "string", enum: ["high", "medium", "low"] },
+            evidence: {
+                type: "array",
+                items: { type: "object", properties: { file: { type: "string" }, lines: { type: "string" }, note: { type: "string" } }, required: ["file", "lines", "note"], additionalProperties: false },
+            },
+            reasoning: { type: "string" },
+        },
+        required: ["verdict", "confidence", "evidence", "reasoning"],
+        additionalProperties: false,
+    },
+};
+/**
+ * The rubric. The order is deliberate — it is the order a reviewer settles a
+ * claim in — and `not-a-defect` is the verdict that separates "the code does
+ * what the claim says" from "and that is a bug": without it, a cast every
+ * caller satisfies gets confirmed because it is literally there.
+ */
+export const BROADSIDE_VERIFY_SYSTEM_PROMPT = "You verify a scouting finding produced by a one-shot batch scan against the real source code. " +
+    "The scan saw files without being able to look anything up; you can. Use the tools to read the cited location " +
+    "and whatever else the claim depends on (callers, the definition of a helper, error handling around it). " +
+    "Then decide, in this order: " +
+    "'discarded' — the claim is wrong about the code (the guard exists, the value cannot be what the claim assumes, " +
+    "the cited line does something else, the condition it fears is ruled out by the project's config or runtime); " +
+    "'not-a-defect' — the claim is literally true of the code but no caller, input, or state can reach the failure it " +
+    "describes: a TypeScript cast every caller satisfies, a hypothetical about an environment the project does not " +
+    "target, a style or type-hygiene observation; " +
+    "'confirmed' — the failure is reachable: name the concrete input, call site, or sequence that triggers it, and what " +
+    "then goes wrong; " +
+    "'unclear' — settling it needs runtime behaviour or specification knowledge the code does not contain. " +
+    "Be strict: a real but different problem than the one claimed is 'discarded' with the difference noted, and " +
+    "'confirmed' without a trigger you found in the code is not allowed. " +
+    "Cite line ranges you actually read. When you are done, reply with only the JSON verdict object.";
+function findingPrompt(index, finding) {
+    return (`Finding ${index} (${finding.lensId} lens, severity ${finding.severity}):\n` +
+        `Title: ${finding.title}\nLocation: ${finding.location}\nPattern/category: ${finding.pattern || "-"}\n` +
+        `Description: ${finding.description}\n\n` +
+        "Verify it. Reply with a JSON object {verdict, confidence, evidence:[{file,lines,note}], reasoning}.");
+}
+async function chat(fetcher, apiKey, model, body) {
+    const resp = await fetcher(BROADSIDE_CHAT_URL, {
+        method: "POST",
+        headers: { Authorization: `Bearer ${apiKey}`, "Content-Type": "application/json" },
+        body: JSON.stringify({ model, reasoning: defaultReasoningFor(), usage: { include: true }, ...body }),
+        signal: AbortSignal.timeout(120_000),
+    });
+    const data = (await resp.json());
+    if (!resp.ok || data.error) {
+        const detail = data.error?.message ?? JSON.stringify(data.error ?? data).slice(0, 300);
+        throw new Error(`OpenRouter chat: HTTP ${resp.status}: ${detail}`);
+    }
+    return data;
+}
+function parseVerdict(text) {
+    const trimmed = text.trim();
+    const fenced = /```(?:json)?\s*([\s\S]*?)```/.exec(trimmed);
+    try {
+        const parsed = JSON.parse(fenced ? fenced[1] : trimmed);
+        return typeof parsed?.verdict === "string" ? parsed : null;
+    }
+    catch {
+        return null;
+    }
+}
+/** Verify one finding: up to the tool budget, then a verdict. */
+export async function verifyFinding(finding, index, reader, apiKey, model, fetcher) {
+    const messages = [
+        { role: "system", content: BROADSIDE_VERIFY_SYSTEM_PROMPT },
+        { role: "user", content: findingPrompt(index, finding) },
+    ];
+    let toolCalls = 0;
+    let cost = 0;
+    const usageCost = (data) => {
+        const usage = data.usage;
+        return typeof usage?.cost === "number" ? usage.cost : 0;
+    };
+    try {
+        // The +2 leaves room for the final schema-forced call after the budget.
+        for (let step = 0; step < BROADSIDE_VERIFY_MAX_TOOL_CALLS + 2; step++) {
+            const budgetSpent = toolCalls >= BROADSIDE_VERIFY_MAX_TOOL_CALLS;
+            if (budgetSpent)
+                messages.push({ role: "user", content: "Tool budget spent. Emit the verdict JSON object now from what you have read." });
+            const data = await chat(fetcher, apiKey, model, budgetSpent
+                ? { messages, response_format: { type: "json_schema", json_schema: VERDICT_SCHEMA }, max_tokens: 2000 }
+                : { messages, tools: TOOLS, tool_choice: "auto", max_tokens: 4000 });
+            cost += usageCost(data);
+            const choice = data.choices?.[0];
+            const message = choice?.message ?? { role: "assistant", content: "" };
+            messages.push(message);
+            const calls = message.tool_calls;
+            if (calls && calls.length > 0 && !budgetSpent) {
+                for (const call of calls) {
+                    toolCalls += 1;
+                    let args = {};
+                    try {
+                        args = JSON.parse(call.function.arguments || "{}");
+                    }
+                    catch {
+                        // Malformed arguments: the model sees the error and can retry.
+                    }
+                    let output;
+                    try {
+                        output = call.function.name === "read_file"
+                            ? await reader.readFile(String(args.path ?? ""), args.start_line, args.end_line)
+                            : call.function.name === "grep"
+                                ? await reader.grep(String(args.pattern ?? ""), typeof args.path_prefix === "string" ? args.path_prefix : undefined)
+                                : call.function.name === "list_dir"
+                                    ? await reader.listDir(String(args.path ?? "."))
+                                    : `error: unknown tool ${call.function.name}`;
+                    }
+                    catch (error) {
+                        output = `error: ${error instanceof Error ? error.message : String(error)}`;
+                    }
+                    messages.push({ role: "tool", tool_call_id: call.id, content: output.slice(0, TOOL_OUTPUT_MAX_CHARS) });
+                }
+                continue;
+            }
+            const verdict = parseVerdict(typeof message.content === "string" ? message.content : "");
+            if (verdict)
+                return finish(verdict, toolCalls, cost);
+            if (budgetSpent)
+                break;
+            // Prose instead of JSON: one schema-forced call, no tools.
+            const final = await chat(fetcher, apiKey, model, {
+                messages: [...messages, { role: "user", content: "Emit the verdict JSON object now." }],
+                response_format: { type: "json_schema", json_schema: VERDICT_SCHEMA },
+                max_tokens: 2000,
+            });
+            cost += usageCost(final);
+            const forced = parseVerdict(String((final.choices?.[0]?.message?.content) ?? ""));
+            if (forced)
+                return finish(forced, toolCalls, cost);
+            break;
+        }
+        return { verdict: "error", confidence: "low", evidence: [], reasoning: "no verdict within the tool budget", toolCalls, cost };
+    }
+    catch (error) {
+        return { verdict: "error", confidence: "low", evidence: [], reasoning: error instanceof Error ? error.message : String(error), toolCalls, cost };
+    }
+}
+function finish(verdict, toolCalls, cost) {
+    const known = ["confirmed", "not-a-defect", "discarded", "unclear"];
+    const value = String(verdict.verdict);
+    const evidence = Array.isArray(verdict.evidence)
+        ? verdict.evidence.map((e) => ({ file: String(e.file ?? ""), lines: String(e.lines ?? ""), note: String(e.note ?? "") }))
+        : [];
+    return {
+        verdict: (known.includes(value) ? value : "unclear"),
+        confidence: typeof verdict.confidence === "string" ? verdict.confidence : "low",
+        evidence,
+        reasoning: typeof verdict.reasoning === "string" ? verdict.reasoning : "",
+        toolCalls,
+        cost,
+    };
+}
+/**
+ * Verify the top findings of a collected run against the repository.
+ *
+ * `maxCost` is a running cap, not a pre-flight estimate: a sync call's cost
+ * is only known when it returns, so the pass stops *before* starting the next
+ * finding once the cap is reached and reports `partial`. On the default
+ * model a finding costs about a cent.
+ */
+export async function runBroadsideVerify(cwd, apiKey, opts = {}) {
+    const broadsideDir = broadsideDirFor(cwd);
+    const state = await loadBroadsideState(broadsideDir);
+    const run = opts.runId ? state.runs.find((candidate) => candidate.id === opts.runId) : state.runs[state.runs.length - 1];
+    if (!run) {
+        throw new Error(opts.runId ? `No Broad-Side run with id ${opts.runId}.` : "No Broad-Side run recorded. Submit and collect one first.");
+    }
+    const runDir = join(broadsideDir, run.outputDir);
+    const stored = await loadSavedLensResults(runDir, run.lenses);
+    const ranked = rankVerifiableFindings(stored);
+    if (ranked.length === 0) {
+        throw new Error(`Run ${run.id} has no verifiable findings on disk: the defect and security lenses either did not run, are not collected yet, or found nothing. Collect the run first.`);
+    }
+    const top = Math.max(1, Math.floor(opts.top ?? BROADSIDE_VERIFY_DEFAULT_TOP));
+    // The run's model is a `:batch` variant; the sync endpoint wants the base id.
+    const model = opts.model ?? syncModelFor(run.model);
+    const maxCost = opts.maxCost ?? 0;
+    const fetcher = opts.fetcher ?? fetch;
+    const reader = await createRepoReader(cwd);
+    const selected = ranked.slice(0, top);
+    const findings = [];
+    let totalCost = 0;
+    let stoppedByCost = false;
+    for (const [i, candidate] of selected.entries()) {
+        if (opts.signal?.aborted)
+            break;
+        if (maxCost > 0 && totalCost >= maxCost) {
+            stoppedByCost = true;
+            break;
+        }
+        const outcome = await verifyFinding(candidate, i + 1, reader, apiKey, model, fetcher);
+        const finding = {
+            index: i + 1,
+            lensId: candidate.lensId,
+            customId: candidate.customId,
+            severity: candidate.severity,
+            title: candidate.title,
+            location: candidate.location,
+            ...outcome,
+        };
+        findings.push(finding);
+        totalCost += outcome.cost;
+        opts.onProgress?.(finding);
+    }
+    const status = findings.length === selected.length ? "completed" : "partial";
+    const entry = {
+        status,
+        model,
+        top,
+        verified: findings.length,
+        confirmed: findings.filter((f) => f.verdict === "confirmed").length,
+        cost: totalCost,
+        at: new Date().toISOString(),
+    };
+    const result = {
+        runId: run.id,
+        outputDir: join(".codecarto", BROADSIDE_DIR, run.id),
+        model,
+        status,
+        candidates: ranked.length,
+        findings,
+        totalCost,
+        ...(stoppedByCost && { stoppedByCost: true }),
+    };
+    await writeFile(join(runDir, "verified.json"), `${JSON.stringify({ ...entry, run_id: run.id, candidates: ranked.length, findings }, null, "\t")}\n`, "utf8");
+    await writeFile(join(runDir, "verified.md"), renderVerifiedMarkdown(result), "utf8");
+    run.verify = entry;
+    await persistBroadsideRunMerging(broadsideDir, run);
+    return result;
+}
+const VERDICT_MARK = { confirmed: "✓", "not-a-defect": "–", discarded: "✗", unclear: "?", error: "!" };
+export function renderVerifiedMarkdown(result) {
+    const lines = [
+        `# Verified findings — run ${result.runId}`,
+        "",
+        `${result.findings.length} of ${result.candidates} verifiable finding(s) read against the source on \`${result.model}\` (most severe first), $${result.totalCost.toFixed(4)}.` +
+            (result.stoppedByCost ? " Stopped by the cost cap before the rest." : ""),
+        "",
+        "A **confirmed** finding names the input, call site, or sequence that reaches the failure. **not-a-defect** means the claim is",
+        "literally true of the code but nothing can reach the failure it describes; **discarded** means the claim is wrong about the",
+        "code; **unclear** needs runtime or specification knowledge. Every verdict is still a model's reading — a confirmed finding",
+        "is a lead worth a human's next look, not a validated claim.",
+        "",
+    ];
+    for (const f of result.findings) {
+        lines.push(`## ${VERDICT_MARK[f.verdict] ?? "?"} ${f.index}. [${f.severity}] ${f.title}`, "", `- **verdict**: ${f.verdict} (${f.confidence})`, `- **location**: ${f.location}`, `- **lens**: ${f.lensId} (${f.customId})`);
+        if (f.evidence.length > 0)
+            lines.push(`- **evidence**: ${f.evidence.map((e) => `${e.file}:${e.lines} — ${e.note}`).join("; ")}`);
+        lines.push(`- **reasoning**: ${f.reasoning.replace(/\s+/g, " ").trim()}`, `- **cost**: $${f.cost.toFixed(4)} (${f.toolCalls} tool call(s))`, "");
+    }
+    return lines.join("\n");
+}
+export function verifyResultText(result) {
+    const counts = { confirmed: 0, "not-a-defect": 0, discarded: 0, unclear: 0, error: 0 };
+    for (const f of result.findings)
+        counts[f.verdict] = (counts[f.verdict] ?? 0) + 1;
+    const lines = [
+        `Broad-Side verify — run ${result.runId}: ${result.status}`,
+        `  ${result.findings.length} of ${result.candidates} verifiable finding(s) read on ${result.model} | cost: $${result.totalCost.toFixed(4)}` +
+            (result.stoppedByCost ? " (stopped by the cost cap)" : ""),
+        `  confirmed ${counts.confirmed} · not-a-defect ${counts["not-a-defect"]} · discarded ${counts.discarded} · unclear ${counts.unclear}` + (counts.error ? ` · error ${counts.error}` : ""),
+    ];
+    for (const f of result.findings) {
+        lines.push(`  ${VERDICT_MARK[f.verdict] ?? "?"} [${f.severity}] ${f.title} @ ${f.location} — ${f.verdict}`);
+    }
+    lines.push(`Details in ${result.outputDir}/verified.md. A confirmed finding is a lead for a human's next look, not a validated claim.`);
+    return lines.join("\n");
+}

package/dist/core/broadside.d.ts

@@ -244,6 +244,17 @@ export type BroadsideTriageEntry = {
     cost?: number;
     error?: string;
 };
+/** Recorded on the run once a verification pass has run (#143); see core/broadside-verify.ts. */
+export type BroadsideVerifyEntry = {
+    /** `completed`: every selected finding got a verdict; `partial`: the cost cap or an abort stopped it early. */
+    status: "completed" | "partial";
+    model: string;
+    top: number;
+    verified: number;
+    confirmed: number;
+    cost: number;
+    at: string;
+};
 /** The truncation retry pass of one run: one batch per model (#206). */
 export type BroadsideRetryEntry = {
     status: "submitted" | "completed" | "failed";
@@ -275,6 +286,8 @@ export type BroadsideRun = {
      * run cannot both submit it (#322). Absent until a collect claims it.
      */
     retry?: BroadsideRetryEntry;
+    /** The verification pass over the top findings, when one has run (#143). */
+    verify?: BroadsideVerifyEntry;
     totalCost?: number;
     pricing?: ModelPricing;
     maxCost?: number;
@@ -505,12 +518,16 @@ type LensDefinition = {
     skipTestFiles?: boolean;
     globsFor: (info: RepoInfo) => string[];
     /**
-     * Where to look when `globsFor` matches nothing (#319). The security and
-     * api lenses target server/, auth, and middleware paths because that is
-     * where the trust boundary usually lives; a service whose server is
-     * `src/server.js` matched none of them and got no security review at all.
-     * The fallback is the language's whole source set — priced as such, and
-     * said so in the estimate, the run record, and the prompt.
+     * Where to look when `globsFor` matches no source file (#319). The
+     * security and api lenses target server/, auth, and middleware paths
+     * because that is where the trust boundary usually lives; a service whose
+     * server is `src/server.js` matched none of them and got no security
+     * review at all. A match that is only documents is the same starvation:
+     * `SECURITY.md` satisfied the security lens on CodeCartographer itself,
+     * which then reviewed a policy and reported zero findings. The fallback
+     * is the language's whole source set, added to whatever did match —
+     * priced as such, and said so in the estimate, the run record, and the
+     * prompt.
      */
     fallbackGlobsFor?: (info: RepoInfo) => string[];
     systemPrompt: (info: RepoInfo) => string;
@@ -520,18 +537,36 @@ export declare function getLens(lensId: BroadsideLensId): LensDefinition;
 export declare function listLenses(): LensDefinition[];
 /** The languages Broad-Side can scan; anything else is refused at submit. */
 export declare const BROADSIDE_LANGUAGES: readonly ["go", "python", "rust", "typescript", "javascript"];
+/**
+ * The files a run scans, and where they came from. Contents are always read
+ * from the working tree, so the list is the working tree's too: tracked files
+ * plus untracked ones git does not ignore, minus files deleted on disk. The
+ * list used to come from `git ls-tree HEAD`, so a run mixed the committed
+ * file list with uncommitted contents and never saw an untracked file (#248).
+ * A target that is not a git repository gets a bounded walk.
+ */
+export declare function listRepoFiles(targetDir: string): Promise<{
+    files: string[];
+    snapshot: RepoSnapshotSource;
+}>;
 export declare function collectRepoInfo(targetDir: string, opts?: {
     redact?: boolean;
 }): Promise<RepoInfo>;
+export declare function isSlurpable(relPath: string): boolean;
 type CollectedFile = {
     relPath: string;
     moduleName: string;
 };
 /**
- * The files a lens will read: its targeted globs, or — when those match
- * nothing and the lens declares a fallback — the fallback globs, with a
- * sentence saying so (#319). The sentence travels to the estimate, the
- * batch entry, and the prompt, so a fallback scan is never a silent one.
+ * The files a lens will read: its targeted globs, or — when those match no
+ * source file and the lens declares a fallback — the fallback globs on top
+ * of whatever did match, with a sentence saying so (#319). The sentence
+ * travels to the estimate, the batch entry, and the prompt, so a fallback
+ * scan is never a silent one.
+ *
+ * "No source file" rather than "no file": a policy document or a config
+ * file under a targeted path satisfies the globs and leaves the lens with
+ * nothing to review, and the coverage note it writes back is the only sign.
  */
 export declare function selectLensFiles(allFiles: string[], lens: LensDefinition, info: RepoInfo): {
     files: CollectedFile[];

package/dist/core/broadside.js

@@ -749,7 +749,19 @@ const LENSES = {
                 "Return a JSON object following the defect_scan_report schema. " +
                 "Cite file:line for every finding. List which patterns you checked. " +
                 "If the code looks clean for a pattern, say so rather than staying silent. " +
-                "Prefer precision over volume — 3 solid findings beat 15 vague ones.");
+                "Prefer precision over volume — 3 solid findings beat 15 vague ones.\n\n" +
+                // The verification pass (#143) confirmed 2 of the 12 top findings a
+                // scan produced with the paragraph above alone; the other ten were
+                // casts and assertions every caller satisfied, guards that lived one
+                // call away, or environments the project does not target. The rubric
+                // the verifier applies is asked of the scan itself, up front.
+                "A finding is a reachable failure: name in the description the concrete input, call site, or sequence " +
+                "that reaches it and what then goes wrong. A cast, assertion, `any`, or non-null `!` that every caller " +
+                "you can see satisfies, a hypothetical about a runtime or environment the project does not target, or a " +
+                "style or type-hygiene observation is not a defect — leave it out, or if it is worth a note, report it " +
+                "at severity low under the pattern name `type-hygiene` so it ranks apart from reachable failures. " +
+                "When the guard you looked for may live in another module, say which check you could not find " +
+                "rather than asserting it is absent; severity high or medium is for failures you traced to a trigger.");
         },
         userPrompt: (info, source, moduleName) => `Scan this ${info.language} module for mechanical defects.\n\n` +
             `Module: ${moduleName}\n\n` +
@@ -920,7 +932,7 @@ const SOURCE_SPECS = {
  * file list with uncommitted contents and never saw an untracked file (#248).
  * A target that is not a git repository gets a bounded walk.
  */
-async function listRepoFiles(targetDir) {
+export async function listRepoFiles(targetDir) {
     try {
         const listed = await execFileAsync("git", ["-C", targetDir, "ls-files", "-z", "--cached", "--others", "--exclude-standard"], { maxBuffer: 64 * 1024 * 1024, timeout: GIT_TIMEOUT_MS });
         const deleted = await execFileAsync("git", ["-C", targetDir, "ls-files", "-z", "--deleted"], {
@@ -1200,7 +1212,7 @@ function matchesAnyGlob(path, globs) {
     }
     return false;
 }
-function isSlurpable(relPath) {
+export function isSlurpable(relPath) {
     // A credential store is never a lens input, whatever its globs say (#252).
     if (isSecretFile(relPath))
         return false;
@@ -1254,25 +1266,51 @@ function collectFilesMatching(allFiles, lens, globs) {
     }
     return out;
 }
+/** Code in any language Broad-Side scans as, whatever this repo's is. */
+const SOURCE_EXTENSIONS = new Set(Object.values(SOURCE_SPECS).flatMap((spec) => spec.exts));
+function isSourceFile(relPath) {
+    const dot = relPath.lastIndexOf(".");
+    return dot > relPath.lastIndexOf("/") && SOURCE_EXTENSIONS.has(relPath.slice(dot).toLowerCase());
+}
+/** `a, b, c and 4 more` — a matched-file list short enough for a status line. */
+function listSome(paths, max = 3) {
+    if (paths.length <= max)
+        return paths.join(", ");
+    return `${paths.slice(0, max).join(", ")} and ${paths.length - max} more`;
+}
 /**
- * The files a lens will read: its targeted globs, or — when those match
- * nothing and the lens declares a fallback — the fallback globs, with a
- * sentence saying so (#319). The sentence travels to the estimate, the
- * batch entry, and the prompt, so a fallback scan is never a silent one.
+ * The files a lens will read: its targeted globs, or — when those match no
+ * source file and the lens declares a fallback — the fallback globs on top
+ * of whatever did match, with a sentence saying so (#319). The sentence
+ * travels to the estimate, the batch entry, and the prompt, so a fallback
+ * scan is never a silent one.
+ *
+ * "No source file" rather than "no file": a policy document or a config
+ * file under a targeted path satisfies the globs and leaves the lens with
+ * nothing to review, and the coverage note it writes back is the only sign.
  */
 export function selectLensFiles(allFiles, lens, info) {
     const globs = lens.globsFor(info).filter(Boolean);
     const targeted = collectFilesMatching(allFiles, lens, globs);
-    if (targeted.length > 0 || globs.length === 0 || !lens.fallbackGlobsFor)
+    if (globs.length === 0 || !lens.fallbackGlobsFor)
+        return { files: targeted };
+    if (targeted.some((f) => isSourceFile(f.relPath)))
         return { files: targeted };
     const fallbackGlobs = lens.fallbackGlobsFor(info).filter(Boolean);
-    const files = collectFilesMatching(allFiles, lens, fallbackGlobs);
-    if (files.length === 0)
-        return { files };
+    const matched = new Set(targeted.map((f) => f.relPath));
+    const sources = collectFilesMatching(allFiles, lens, fallbackGlobs).filter((f) => !matched.has(f.relPath));
+    if (sources.length === 0)
+        return { files: targeted };
+    const excluded = lens.skipTestFiles ? "test files excluded" : "";
+    const scanned = `scanned all ${info.language} sources (${fallbackGlobs.join(", ")})`;
     return {
-        files,
-        fallback: `no files matched ${globs.join(", ")}${lens.skipTestFiles ? " (test files excluded)" : ""}; ` +
-            `scanned all ${info.language} sources (${fallbackGlobs.join(", ")}) instead`,
+        // What did match rides first: the policy the model is about to check
+        // the code against, ahead of the code.
+        files: [...targeted, ...sources],
+        fallback: targeted.length === 0
+            ? `no files matched ${globs.join(", ")}${excluded ? ` (${excluded})` : ""}; ${scanned} instead`
+            : `no source files matched ${globs.join(", ")} (only ${listSome(targeted.map((f) => f.relPath))}` +
+                `${excluded ? `; ${excluded}` : ""}); ${scanned} as well`,
     };
 }
 function collectLensFiles(allFiles, lens, info) {
@@ -1396,8 +1434,8 @@ export function buildBatchRequest(lens, info, slice, index, sliceCount, model =
                     // so the model judges the trust boundary wherever it appears
                     // and does not report the missing server/ as a finding (#319).
                     (slice.fallback
-                        ? `NOTE: this repository has no files under the paths this lens usually reads (${slice.fallback}). ` +
-                            "What follows is every source file it has; locate the trust boundary and the request-handling code wherever they live.\n\n"
+                        ? `NOTE: this repository has no source files under the paths this lens usually reads (${slice.fallback}). ` +
+                            "What follows is every source file it has, after anything those paths did match; locate the trust boundary and the request-handling code wherever they live.\n\n"
                         : "") + lens.userPrompt(info, slice.content, slice.moduleName),
                 },
             ],
@@ -1624,6 +1662,10 @@ export async function persistBroadsideRunMerging(broadsideDir, run) {
                 run.triage = onDisk.triage;
             if (retryEntryRank(onDisk.retry) > retryEntryRank(run.retry))
                 run.retry = onDisk.retry;
+            // A verification pass another process recorded is never dropped by
+            // a collect that never knew about it; a newer pass replaces an older.
+            if (onDisk.verify && (!run.verify || onDisk.verify.at > run.verify.at))
+                run.verify = onDisk.verify;
             for (const [lensId, theirs] of Object.entries(onDisk.batches)) {
                 if (theirs && batchEntryRank(theirs) > batchEntryRank(run.batches[lensId]))
                     run.batches[lensId] = theirs;
@@ -3545,6 +3587,9 @@ export function statusText(state) {
         }
         lines.push(`  synthesis: ${run.synthesis.status}`);
         lines.push(`  triage: ${run.triage?.status ?? "pending"}`);
+        if (run.verify) {
+            lines.push(`  verify: ${run.verify.status} — ${run.verify.confirmed} confirmed of ${run.verify.verified} read on ${run.verify.model}, $${run.verify.cost.toFixed(4)}`);
+        }
         if (run.totalCost !== undefined)
             lines.push(`  total cost: $${run.totalCost.toFixed(6)}`);
     }