Build 26341799

First, something Dollar got wrong: deleting a folder inside your user folder counted as a catastrophe. In 2.0.1, recursively deleting anything under C:\\Users\\<you> — a project's node_modules, say — was classed with wiping the whole user folder, so it stopped and waited for you in every tier (the same for paths written with ~ or an environment variable). Now only the whole user folder is in that class. A folder inside it gets a red notice like any other risky delete, and your agent is not held.

When Dollar's own service was failing, ordinary commands were blocked. If the local service returned an error or had not finished starting, the hook blocked far more than the irreversible class — a plain git push, a del, an rmdir — and said they matched the irreversible floor. They did not. It now blocks only what is truly irreversible and lets the rest through with a local record. The hook is refreshed when Dollar starts, so this applies once you have launched 2.1.

Some costs were computed at the wrong rate. A model the built-in price catalogue does not know is costed at a generic default rate and marked unconfirmed. Several newer models were landing there, so their dollar figures were off — high for some, low for others. Added: Claude Opus 5.5 and Sonnet 5.5; GPT-6 (astra, sol, luna) and GPT-6.1 sol; the o-series (o1, o1-pro, o3, o3-pro, o3-mini, o4-mini). Codex's built-in auto-review is now treated as part of the Codex plan: tokens counted, dollar amount zero. This applies to usage recorded after you update — figures already recorded are not recalculated — and totals remain local estimates, not your provider's bill.

Other things on screen that were not true.

• Spend page: "all-time" tokens left out cache tokens while today / week / month included them, so "this month" could exceed "all-time". The per-task average divided by successful tasks only; it now uses all tasks and says so.
• Full screen: the sidebar's agent count read zero healthy on every page, whatever the real state, and the Escape Pod badge said "LOCAL · REDACTED" before anything had been done.
• Local records: a command containing "Authorization: Bearer …" had only the word "Bearer" masked, so a token without a well-known key prefix stayed readable. Fixed; records written before the update are not rewritten.

The usage report could not be generated at all. Full screen → Toolkit → Usage report → "Generate report" failed every time in 2.0.1. That was our bug, not your machine; it works now. It was also written to fetch its collector (ccusage 20.0.19) from npm with npx when you clicked — and, because of that bug, never got that far. 2.1 ships the same pinned version inside the app and runs it directly; nothing is downloaded to run it. To compute costs, the collector downloads a public model price table each time you generate a report — from GitHub, and from models.dev if a model is missing there. It only downloads; none of your usage is sent. If the table cannot be fetched, some newer models' costs come out as $0 (token counts are unaffected) — we are fixing that. Dollar has no upload path: the report is a file saved where you choose.

The floor holds in more places. A full review closed gaps in what always stops and waits: delete-style syncs aimed at a root, home or drive; remote deletions through hosting and package CLIs; roundabout spellings of root and home; more force-push shapes; destructive SQL passed through MCP tool arguments. Edits to Dollar's own guard files are recognised when Codex writes them as a patch, and are blocked even while Dollar's service is down. In the other direction, a URL that merely contains "token" or "secret" no longer blocks. Fixed in both the online classifier and the offline hook.

New.

• Task receipt and black box — the Recovery Center (Settings → Toolkit → Open Recovery Center) opens each task with three lines: what got done, what did not run or finish, whether it can still be recovered — each marked observed or inferred. Below is a step-by-step record in which "allowed" and "actually finished" are separate entries. Tasks from before the update show only what the old record kept, and say so. Rolling back is unchanged: preview, then confirm.
• Appearance page — Settings → Appearance. Save Dollar's current look as a pack (one local file), edit the images, import it back, apply it without restarting, export it to share. Import reports frames that are out of alignment. The library starts empty; nothing here is uploaded.
• Guided tour — 14 short stops (12 in Local toolkit mode), starting with what onboarding never said: right-click the cat. Offered once after first-run setup. Already set up? Settings → Replay the guided tour. It steps aside for approvals and notices.
• "Needs you" — a third tab in the Missed notices window for what is waiting on you now; its count joins the unread line in the right-click menu. It never pops up on the cat.
• Advice, for Codex too — a word when a Codex session's context window is filling or its quota window is nearly used. Advice now shows what was seen, your latest restore point and a next step, with buttons for things Dollar can do itself. At most two an hour; dismiss one kind three times running and it mutes itself for seven days.

Also — links and "show in folder" no longer open hidden beneath Dollar's always-on-top window. Local toolkit mode was a one-way door — Settings said to replay the guide to switch back, and the guide had no way back; it does now. The token-pace window has a gear to add CPU, memory, disk and spend. AI Highway lanes draw one car per recent record (up to 12), moving only while the reading is live. With the cat hidden, the status strip has a 🐾 button to call it back. Spend sits next to Settings; the one-click checkup now lives on Agent Health only.

More fixes.

• The card shown after an update now appears just above the cat and can be clicked. Before, it sat at the very top of the pet window, far from the cat, and clicks on it went through to the desktop.
• That card no longer lists older features as new in this version.
• Six more key formats are now masked before Dollar writes its local records: AWS access keys, Google API keys, bare JWTs, GitHub ghu_/ghr_ tokens, private key blocks, and the user:password part of database connection strings. The approval card masks them too. Lines written by earlier versions are not rewritten.
• The spend page now says the right thing about models with no confirmed price: they are estimated at the default rate and included in the totals. It used to say they were left out, which was wrong.
• Added Claude Haiku 5.5 to the built-in price list, which used to estimate it at the default rate. Claude Sonnet 5.5 cache reads now use Anthropic's lower price from October 7.
• The sidebar's last line is no longer cut off at 1280×680, which is a 1080p screen at 150% scaling.
• Security: Dollar's hook now only talks to a server on your own machine. Before, a setting inside a project could redirect it to another address.
• Security: removing a skill can no longer delete the whole skills folder.
• Security: restoring an old sprite version now rejects invalid version names.
• Added GPT-6.1 Terra and GPT-6 Rosalind to the built-in price list.

Platform note: this update is Windows only. macOS stays on 2.0.1 for now; 2.1 for macOS will follow as its own update and will be announced here.