argus-reviewer-e2e 0.3.1 → 0.4.0

package/dist/cli.d.ts

@@ -1,16 +1,18 @@
 #!/usr/bin/env node
 import { Config } from './config.js';
-import { type ExecFn } from './detect.js';
+import { type ExecFn, type ProbeFn } from './detect.js';
 import { BrowserDriver } from './driver/browser.js';
 import { VisionClient } from './engine/loop.js';
 import { type PrMeta } from './evidence/ci.js';
 import { type Evidence } from './evidence/link.js';
+import { type ValidationAudit } from './review/validate.js';
 import { type SecretsScanResult } from './review/secrets.js';
 import { type TriageRecord } from './review/triage.js';
 import { type FindingAdjudicationAudit } from './review/adjudicate.js';
 import { type ProbeRecord } from './probe/queue.js';
 import { type HeadBinding } from './report/manifest.js';
 import { CallCost } from './vision/cost.js';
+import { Message } from './vision/openrouter.js';
 export interface CliDeps {
     cwd?: string;
     env?: NodeJS.ProcessEnv;
@@ -22,6 +24,15 @@ export interface CliDeps {
     launchDriver?: (config: Config) => Promise<BrowserDriver>;
     /** Inject a subprocess runner (tests stub `a0`/`gh` detection + delegation). */
     exec?: ExecFn;
+    /** Inject the host reachability probe (tests stub a0 detection). */
+    probe?: ProbeFn;
+    /**
+     * Whether output goes to a terminal. Defaults to process.stdout.isTTY
+     * when `out` is not injected, and to false when it is.
+     */
+    isTTY?: boolean;
+    /** Terminal width for the summary block (default process.stdout.columns, else 80). */
+    columns?: number;
 }
 export declare function main(argv: string[], deps?: CliDeps): Promise<number>;
 interface PrFile {
@@ -35,7 +46,7 @@ export interface ReviewFinding {
     severity: string;
     category?: string;
     message: string;
-    /** U8 — Jev true-positive probability (absent = unadjudicated). */
+    /** U8: confidence-model true-positive probability (absent = unadjudicated). */
     p?: number;
     /** R1 — committable replacement lines for the commented range (parse-bounded). */
     suggestion?: string;
@@ -51,9 +62,28 @@ export interface ReviewComment {
     start_side?: 'RIGHT';
     side: 'RIGHT';
     body: string;
-    /** R10 — path:line:bodyFirstLine:hash8(suggestion); a corrected suggestion re-posts. */
+    /** KTD4: path:line:severity:normalizedMessage:hash8(suggestion); a corrected suggestion re-posts. */
     dedupKey: string;
 }
+/** Per-finding audit record for a finding the post-parse filters removed. */
+export interface DroppedFinding {
+    file: string;
+    line?: number;
+    severity: string;
+    category?: string;
+    message: string;
+    reason: 'outside-diff' | 'revert-nit';
+}
+export interface ReviewScope {
+    /** Changed files in the PR with a patch. */
+    totalFiles: number;
+    /** Files that reached the review model. */
+    reviewedFiles: number;
+    /** Files kept out by `review.exclude`. */
+    excludedFiles: number;
+    /** Up to 5 excluded paths, for the Diagnostics line. */
+    excludedSample: string[];
+}
 interface CodeReviewReport {
     ok: boolean;
     skipped: boolean;
@@ -66,7 +96,7 @@ interface CodeReviewReport {
     reviewEvent: 'comment' | 'request_changes';
     /** Blocker-severity findings a sandbox probe reproduced. */
     provenBlockers: number;
-    /** Blocker-severity findings at/above the Jev P(true-positive) gate. */
+    /** Blocker-severity findings at/above the confidence-model P(true-positive) gate. */
     highConfidenceBlockers: number;
     /** KTD3 — eligibility-filtered, severity-sorted, sanitized, capped. */
     reviewComments: ReviewComment[];
@@ -80,10 +110,24 @@ interface CodeReviewReport {
     secretsScan?: SecretsScanResult | {
         skipped: string;
     };
-    /** U7 triage record — Jev pre-review signals (annotate/route, never gates). */
+    /** U7 triage record: confidence-model pre-review signals (annotate/route, never gates). */
     triage?: TriageRecord;
     /** U8 adjudication audit — per-finding p + suppressed records. */
     findingAdjudication?: FindingAdjudicationAudit;
+    /** Findings dropped for citing a file/line the diff never shows. */
+    droppedUnanchored?: number;
+    /** nit/q findings dropped for asking to revert text the diff added. */
+    droppedReverted?: number;
+    /** Per-finding audit of dropped findings (capped at 50); counters stay total. */
+    droppedFindings?: DroppedFinding[];
+    /** Synthesis verdict when it diverges from the post-filter derived verdict. */
+    modelVerdict?: 'pass' | 'needs_changes' | 'approve';
+    /** How much of the PR the review covered, and what was left out. */
+    scope?: ReviewScope;
+    /** Findings dropped by deterministic validation, with reasons. */
+    validation?: ValidationAudit;
+    /** Test-file findings capped at nit (bug/risk with no non-test citation). */
+    testFileCapped?: number;
     calls: CallCost[];
     visionCostUsd: number;
     tokens: number;
@@ -91,6 +135,14 @@ interface CodeReviewReport {
     budgetExceeded: boolean;
     /** Identity relationship between the report source and checkout. */
     headBinding?: HeadBinding;
+    /** Workflow-run nonce (GITHUB_RUN_ID) — see runNonceFrom. */
+    runNonce?: string;
+    /**
+     * Base64 HTML-comment payload (`argus-probe-persist`) carrying reproduced
+     * probe source — the sticky poster embeds it verbatim so `@argus persist`
+     * can commit the probes later from a base-only checkout (E1.U3).
+     */
+    persistPayload?: string;
 }
 /**
  * Split `git diff` text into per-file PrFile entries — the local-diff
@@ -113,6 +165,7 @@ export declare function loadFixture(dir: string, exec?: ExecFn): Promise<{
     skipped: string;
 }>;
 export declare function buildPatchChunks(files: PrFile[], contexts?: Record<string, string>): string[];
+export declare function buildCodeReviewMessages(repo: string, pr: string, patchText: string, chunkIndex?: number, totalChunks?: number, profiles?: readonly string[]): Message[];
 export declare function parseCodeReview(content: string): {
     summary: string;
     verdict: 'pass' | 'needs_changes' | 'approve';
@@ -126,7 +179,50 @@ export declare function parseCodeReview(content: string): {
  */
 export declare function carryForwardSuggestions(findings: CodeReviewReport['findings'], originals: CodeReviewReport['findings']): CodeReviewReport['findings'];
 /**
- * R3/KTD2 — Jev P(true-positive) at/above which a blocker-severity finding
+ * New-side (RIGHT) line ranges covered by each file's diff hunks — the
+ * only lines a finding can anchor to (and the only ones it could have
+ * seen).
+ */
+export declare function diffLineRanges(files: readonly {
+    filename: string;
+    patch?: string | undefined;
+}[]): Map<string, [number, number][]>;
+/**
+ * New-side line number -> line text for every line the diff shows
+ * (added and context). Lets post-parse checks compare a finding's claim
+ * against what the cited line actually says.
+ */
+export declare function diffLineTexts(files: readonly {
+    filename: string;
+    patch?: string | undefined;
+}[]): Map<string, Map<number, string>>;
+/**
+ * Drop nit/q findings that ask to remove or revert text the cited diff
+ * line itself contains — i.e. findings that would undo wording the PR
+ * deliberately added ("remove `inconclusive`", "replace 'self-reported'
+ * with 'self-reported'"). Verdict-driving findings (bug/risk, security-
+ * category, configured blocking severities) are never touched: if the
+ * claim is real, severity stays the reviewer's call.
+ */
+export declare function filterRevertNits(findings: readonly ReviewFinding[], textsByFile: Map<string, Map<number, string>>, blockSeverities?: readonly string[]): {
+    kept: ReviewFinding[];
+    dropped: ReviewFinding[];
+};
+/**
+ * Drop findings whose line isn't visible in the file's diff. A finding on
+ * a file the diff doesn't touch, or at a line outside every hunk, is
+ * unverifiable and unpostable — misnumbered and fabricated citations land
+ * here. Line-less (file-level) findings always survive. Verdict-driving
+ * findings (bug/risk, security-category, configured blocking severities)
+ * are never dropped — a misnumbered cite on a real defect must still
+ * gate; the post-time isOnDiff check keeps its comment off the PR.
+ */
+export declare function filterToDiffLines(findings: readonly ReviewFinding[], rangesByFile: Map<string, [number, number][]>, blockSeverities?: readonly string[]): {
+    kept: ReviewFinding[];
+    dropped: ReviewFinding[];
+};
+/**
+ * R3/KTD2: confidence-model P(true-positive) at/above which a blocker-severity finding
  * counts as proven for the REQUEST_CHANGES gate. This is a different axis
  * from `review.findingThreshold` (P(false-positive) for nit/q suppression)
  * — never reuse that knob. 0.7: high-confidence without demanding
@@ -140,7 +236,7 @@ export declare const P_TRUE_POSITIVE_THRESHOLD = 0.7;
  * code-review.json; posters read `reviewEvent`, never recompute.
  * Unadjudicated blockers (no p, not reproduced) never escalate —
  * degrade-open by design. The two counts overlap deliberately: a
- * reproduced AND Jev-confident finding is reported under both.
+ * reproduced AND high-confidence finding is reported under both.
  */
 export declare function computeReviewEvent(findings: ReviewFinding[], blockSeverities: string[], allowRequestChanges: boolean): {
     reviewEvent: 'comment' | 'request_changes';