toolroll 0.9.82 → 0.9.84

package/CHANGELOG.md

@@ -1,5 +1,46 @@
 # Changelog
 
+## 0.9.84 — 2026-11-01
+
+- **An honest Spend page, in Settings.** See API dollars, plan usage and
+  unknown cost separately — never added together, and unknown never shown
+  as $0 — by project, model, role and person, with the most expensive runs
+  and a CSV export. Settings → Spend holds your monthly limits.
+- **One Models editor (Settings → Models).** "Sized by job" is the default:
+  Haiku 5.5 for small tasks, Sonnet 5.5 for routine ones, Opus 5.5 for large
+  or risky ones, and Opus 5.5 plans and reviews. "Opus for everything" is one
+  click away. Every change shows its exact before → after first, needs your
+  password, and is logged; approved work keeps its models. "Switch to it"
+  appears when a newer model ships. A fresh install now starts on Sized by
+  job instead of needing a terminal command.
+- **A safer first-account page.** The setup code page now locks after five
+  wrong tries (with how to get a new code), and only accepts submissions from
+  the page itself.
+- **No more dead ends on first run.** "Connect a builder" opens the steps to
+  do it, the setup checklist knows when a builder is connected, model names
+  read plainly, and `toolroll up` prints the exact command to read your
+  sign-in details.
+- **Settings, rebuilt on one frame.** A clearer layout and one consistent
+  way to confirm sensitive changes: each one shows exactly what it asks for,
+  can't be sent twice, and keeps what you typed if it's refused.
+- **The console answers right away.** Projects are checked in the
+  background; a project only becomes visible once its folder is confirmed.
+- **CI** jobs now have time limits.
+
+## 0.9.83 — 2026-10-31
+
+- **A folder that doesn't answer can't freeze the console.** A disconnected
+  drive, an iCloud-synced Documents folder or a pending macOS privacy prompt
+  used to stall the whole console at startup. Folder reads now run in a
+  separate, time-limited process: the console keeps answering, and the one
+  page that needed the folder says it didn't answer.
+- **Projects come back by themselves.** A saved project whose folder was
+  away when Toolroll started connects as soon as the folder returns, with no
+  restart. It reconnects only to the same folder it was saved as, and a
+  project you remove stays removed.
+- **Safer project paths.** A project's access is decided from its real,
+  resolved location, never from how its path is spelled.
+
 ## 0.9.82 — 2026-10-31
 
 - **A failing nightly holds your releases.** Turn it on with

package/README.md

@@ -286,9 +286,10 @@ the git tests on 2.41 and on 2.47.
 `npm install -g toolroll` gives you the bare `toolroll` command for later
 (`standing-orders`, the older name, still works).
 
-`up` prints your login once (and saves it beside the database as
-`up-login.txt`), opens the app in your browser, and connects this machine as
-the builder. Add an existing folder or a GitHub repository from **Projects**;
+`up` saves your password beside the database as `up-login.txt` (every
+start prints where, and the `cat` command that shows it; never the password
+itself), opens the app in your browser, and connects this machine as the
+builder. Add an existing folder or a GitHub repository from **Projects**;
 the builder and unified chat pick it up while the app keeps running. The
 projects folder and every added repository are remembered. Later,
 `toolroll up` can be run from any directory and reconnects all of them.

package/dist/assignment-ui.js

@@ -1,6 +1,7 @@
 import { assignmentPresentationOf, shortenedMaterialReason } from './assignment-presentation.js';
 import { plainReasonWords } from './proof.js';
 import { chatResultHref } from './chat-controls.js';
+import { startBuilderHref } from './needs-you.js';
 import { statusDetailsHtml, statusWhyHtml } from './task-status.js';
 import { html } from './html.js';
 /** The existing verified receipt reader may discover damage after a verdict
@@ -39,13 +40,16 @@ export function assignmentActionHref(assignment) {
     }
     if (runId !== null && (action.code === 'inspect-run' || action.code === 'reconcile-run'))
         return `/r/${runId}`;
+    // The task page only repeats that no builder is connected: lead to the command that connects one.
+    if (action.code === 'start-worker')
+        return startBuilderHref(taskId);
     const anchor = action.code === 'approve-scope' ? '#approve'
         : action.code === 'inspect-stop' || action.code === 'resume-run' ? '#task-control'
             // Retry (and Build again) is the task's status card itself, at the top: no anchor to its options.
             : action.code === 'unhold' ? '#task-actions'
                 : action.code === 'write-scope' || action.code === 'select-agent' ? '#scope'
                     : action.code === 'inspect-hold' ? '#holds'
-                        : action.code === 'start-worker' || action.code === 'repair-dependency' ? '#run-status' : '';
+                        : action.code === 'repair-dependency' ? '#run-status' : '';
     return taskHref(assignment, taskId) + anchor;
 }
 /** Every version of the assignment, oldest first, as the card lists them. */