@mjasnikovs/pi-task 0.38.3 → 0.38.5

package/dist/task/deep-render-check.d.ts

@@ -1,3 +1,4 @@
+import WebSocket from 'ws';
 export type DeepRenderOutcome = {
     outcome: 'pass';
     detail: string;
@@ -132,6 +133,31 @@ export declare function deriveLegacyFacts(log: SessionRequest[]): Pick<DeepSessi
 export declare function judgeDeepSession(f: DeepSessionFacts): DeepRenderOutcome;
 /** Whole-session wall-clock cap, including browser launch (I4). */
 export declare const DEEP_RENDER_TIMEOUT_MS = 45000;
+/** Minimal DevTools-protocol client: request/response ids over one socket, plus
+ *  event fan-out. Everything the driver needs and nothing more.
+ *
+ *  @internal Exported for its own tests: the id/response matching and the
+ *  send-failure path are the parts of the driver that need no browser. */
+export declare class Cdp {
+    private readonly ws;
+    private nextId;
+    private readonly pending;
+    private readonly handlers;
+    constructor(ws: WebSocket);
+    on(method: string, cb: (params: Record<string, unknown>) => void): void;
+    send(method: string, params?: Record<string, unknown>, sessionId?: string): Promise<Record<string, unknown>>;
+}
+/** Page-side fill: native value setters + input/change events, so a controlled
+ *  React/Vue/Svelte input actually updates its state (a bare `el.value = …` does
+ *  not, and the form then submits empty).
+ *
+ *  @internal Exported so the credential ESCAPING is testable: a password with a
+ *  quote or a backslash in it must not break out of the injected expression. */
+export declare function fillExpr(identifier: string, password: string): string;
+/** Wait until `ms` of silence pass with no new request, or `cap` elapses.
+ *  @internal Exported for its own tests — the quiet/cap arithmetic decides
+ *  whether a slow page is judged half-loaded. */
+export declare function settle(lastActivity: () => number, cap: number, quiet?: number): Promise<void>;
 /**
  * Drive one authenticated session against `url` and judge it. `cwd` is the project
  * whose dotenv declares the account. Every failure mode of the DRIVER itself
@@ -146,4 +172,8 @@ export declare function runDeepRenderCheck(url: string, cwd: string, opts?: {
     /** Recorder hook: receives the facts the verdict was made on. Used by the
      *  corpus builder; the gate itself never passes it. */
     onFacts?: (f: DeepSessionFacts) => void;
+    /** @internal Settle quiet window, for tests only. The three settle windows
+     *  cost QUIET_MS each against a fake browser that answers instantly, which
+     *  is the whole runtime of a driver test. The gate never passes it. */
+    quietMs?: number;
 }): Promise<DeepRenderOutcome>;

package/dist/task/deep-render-check.js

@@ -348,8 +348,11 @@ const POST_SUBMIT_CAP_MS = 12_000;
  *  small so the three settle windows together stay inside DEEP_RENDER_TIMEOUT_MS. */
 const RE_NAV_CAP_MS = 6_000;
 /** Minimal DevTools-protocol client: request/response ids over one socket, plus
- *  event fan-out. Everything the driver needs and nothing more. */
-class Cdp {
+ *  event fan-out. Everything the driver needs and nothing more.
+ *
+ *  @internal Exported for its own tests: the id/response matching and the
+ *  send-failure path are the parts of the driver that need no browser. */
+export class Cdp {
     ws;
     nextId = 1;
     pending = new Map();
@@ -418,8 +421,11 @@ const INSPECT_EXPR = `(() => {
 })()`;
 /** Page-side fill: native value setters + input/change events, so a controlled
  *  React/Vue/Svelte input actually updates its state (a bare `el.value = …` does
- *  not, and the form then submits empty). */
-function fillExpr(identifier, password) {
+ *  not, and the form then submits empty).
+ *
+ *  @internal Exported so the credential ESCAPING is testable: a password with a
+ *  quote or a backslash in it must not break out of the injected expression. */
+export function fillExpr(identifier, password) {
     return `(() => {
     const setValue = (el, v) => {
         const proto = el instanceof HTMLTextAreaElement ? HTMLTextAreaElement.prototype : HTMLInputElement.prototype
@@ -457,8 +463,10 @@ const SUBMIT_EXPR = `(() => {
     if (submit) { submit.click(); return {ok: true, how: 'click'} }
     return {ok: false, reason: 'no submit control'}
 })()`;
-/** Wait until `ms` of silence pass with no new request, or `cap` elapses. */
-async function settle(lastActivity, cap, quiet = QUIET_MS) {
+/** Wait until `ms` of silence pass with no new request, or `cap` elapses.
+ *  @internal Exported for its own tests — the quiet/cap arithmetic decides
+ *  whether a slow page is judged half-loaded. */
+export async function settle(lastActivity, cap, quiet = QUIET_MS) {
     const deadline = Date.now() + cap;
     for (;;) {
         const idleFor = Date.now() - lastActivity();
@@ -514,7 +522,7 @@ export async function runDeepRenderCheck(url, cwd, opts = {}) {
             child = c;
         }, s => {
             socket = s;
-        }, opts.onFacts), budget);
+        }, opts.onFacts, opts.quietMs), budget);
     }
     catch (e) {
         const why = e instanceof Error ? e.message : String(e);
@@ -540,7 +548,7 @@ function withTimeout(p, ms) {
         });
     });
 }
-async function drive(url, bin, userDataDir, credentials, holdChild, holdSocket, onFacts) {
+async function drive(url, bin, userDataDir, credentials, holdChild, holdSocket, onFacts, quietMs) {
     const origin = new URL(url).origin;
     /** Every verdict goes through here, so a recorder sees the same facts the judge
      *  does — the corpus is what the gate itself read, not a reconstruction. */
@@ -626,7 +634,7 @@ async function drive(url, bin, userDataDir, credentials, holdChild, holdSocket,
         return r.result?.value;
     };
     await cdp.send('Page.navigate', { url }, sessionId);
-    await settle(() => lastActivity, SETTLE_CAP_MS);
+    await settle(() => lastActivity, SETTLE_CAP_MS, quietMs);
     const before = await evaluate(INSPECT_EXPR);
     if (!before)
         throw new Error('the page could not be inspected');
@@ -698,7 +706,7 @@ async function drive(url, bin, userDataDir, credentials, holdChild, holdSocket,
     if (!submitted?.ok)
         return judge(unsubmitted({ submitted: false }));
     lastActivity = Date.now();
-    await settle(() => lastActivity, POST_SUBMIT_CAP_MS);
+    await settle(() => lastActivity, POST_SUBMIT_CAP_MS, quietMs);
     // The sign-in request: the first same-origin non-GET issued by the submit. Its
     // own 2xx is the precondition for judging anything, and it is excluded from the
     // data evidence (STEP 0: the broken build satisfies "≥1 same-origin 2xx" with
@@ -731,7 +739,7 @@ async function drive(url, bin, userDataDir, credentials, holdChild, holdSocket,
     if (authAccepted && leftAuthWall) {
         await cdp.send('Page.navigate', { url }, sessionId);
         lastActivity = Date.now();
-        await settle(() => lastActivity, RE_NAV_CAP_MS);
+        await settle(() => lastActivity, RE_NAV_CAP_MS, quietMs);
     }
     return judge(facts(sessionLog(authId, authAt), {
         submitted: true,

package/dist/task/gate-deps.d.ts

@@ -38,6 +38,19 @@ export declare function collectChangedFiles(cwd: string, signal?: AbortSignal):
  * never a blocker. The `.pi-tasks/` bookkeeping is excluded from every git command.
  */
 export declare function collectAddedLines(cwd: string, signal?: AbortSignal): Promise<AddedLine[]>;
+/**
+ * Deterministic sandbox-path-leak pass (see foreign-path.ts, mx5 run 13 PROMPT 4
+ * item 1): find absolute paths the task committed that resolve nowhere on this
+ * machine while the real file sits in the repo, REPAIR the ones whose relative
+ * form provably resolves, and return verify findings for whatever is left.
+ *
+ * The repair runs here, before the verify child, for the same reason lint-fix
+ * does: the defect is mechanical and the correct target is already known, so
+ * spending an AUTOFIX round (or a human) on a path substitution is waste. What it
+ * cannot repair still reaches the child under rule 4e. Failures degrade to no
+ * findings — a sharpener, never a blocker.
+ */
+export declare function collectForeignPathFindings(cwd: string, signal?: AbortSignal, logDebug?: (m: string) => void): Promise<string[]>;
 /**
  * Deterministic neutered-check-script pass (see script-escape.ts, mx5 run 13 PROMPT
  * 4 item 4): check-class scripts that cannot report failure, in a manifest THIS

package/dist/task/gate-deps.js

@@ -61,6 +61,18 @@ export function truncateToolResult(text, limit = TOOL_RESULT_LOG_LIMIT) {
 }
 /** Keep the gate machinery's own artifacts out of every git pathspec below. */
 const EXCLUDE_TASKS_DIR = ':(exclude).pi-tasks';
+/**
+ * Pin the diff header prefixes on any command whose output we PARSE for paths.
+ *
+ * `parseAddedLines` reads the file out of the `+++ b/…` header, but the prefix is
+ * user-configurable: `diff.mnemonicPrefix=true` emits `i/` and `w/` instead of
+ * `a/` and `b/`, and `diff.noprefix=true` emits none. Both are common developer
+ * settings, and under either one every added line was attributed to a path like
+ * `w/src/app.ts`, which exists nowhere — so the probe-gaming and sandbox-path
+ * probes silently read a diff of files that are not in the repo. Passing the
+ * prefixes explicitly makes the output independent of the host's git config.
+ */
+const DIFF_PREFIX_ARGS = ['--src-prefix=a/', '--dst-prefix=b/'];
 const splitLines = (s) => s
     .split('\n')
     .map(l => l.trim())
@@ -110,7 +122,7 @@ export async function collectChangedFiles(cwd, signal) {
  * never a blocker. The `.pi-tasks/` bookkeeping is excluded from every git command.
  */
 export async function collectAddedLines(cwd, signal) {
-    const tracked = await git(cwd, ['diff', 'HEAD', '--', '.', EXCLUDE_TASKS_DIR], signal);
+    const tracked = await git(cwd, ['diff', ...DIFF_PREFIX_ARGS, 'HEAD', '--', '.', EXCLUDE_TASKS_DIR], signal);
     const lines = tracked.exitCode === 0 ? parseAddedLines(tracked.stdout) : [];
     const untracked = await git(cwd, ['ls-files', '--others', '--exclude-standard', '--', '.', EXCLUDE_TASKS_DIR], signal);
     for (const name of splitLines(untracked.exitCode === 0 ? untracked.stdout : '')) {
@@ -124,7 +136,7 @@ export async function collectAddedLines(cwd, signal) {
         }
     }
     if (lines.length === 0) {
-        const last = await git(cwd, ['diff', 'HEAD~1..HEAD', '--', '.', EXCLUDE_TASKS_DIR], signal);
+        const last = await git(cwd, ['diff', ...DIFF_PREFIX_ARGS, 'HEAD~1..HEAD', '--', '.', EXCLUDE_TASKS_DIR], signal);
         return last.exitCode === 0 ? parseAddedLines(last.stdout) : [];
     }
     return lines;
@@ -141,7 +153,7 @@ export async function collectAddedLines(cwd, signal) {
  * cannot repair still reaches the child under rule 4e. Failures degrade to no
  * findings — a sharpener, never a blocker.
  */
-async function collectForeignPathFindings(cwd, signal, logDebug) {
+export async function collectForeignPathFindings(cwd, signal, logDebug) {
     const lines = await collectAddedLines(cwd, signal);
     if (lines.length === 0)
         return [];

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@mjasnikovs/pi-task",
-    "version": "0.38.3",
+    "version": "0.38.5",
     "description": "Deterministic task planning and spec-orchestration for local models — crash-safe /task pipelines with verify/enforce gates, a real-time remote web view, and web/docs/fetch/worker subagent tools.",
     "type": "module",
     "main": "./dist/index.js",