מה השתנה

CHANGELOG-v2.8.md

The download on this site is v2.8.3. A monthly checkup is coming in a later version.

v2.8.3 (2026-10-25)

• No questions in the chat. Every question is a waiting item on the ticket page, including the questions at the start and after closing. The chat only says that the page was updated.
• Plain words before codes. You always read a plain description first. A code such as R1, A1 or X1 appears only as a small tag after it, on the page, in the chat and in the summary.
• The closing code matches what happened. The ticket records plain facts (was there a real fault, was a source removed, did a change fix it), and a fixed decision list turns them into the closing code.
• Unique ticket numbers. A new helper reserves the next free ticket number the moment a ticket opens, so two sessions never share a number. Continuing an open ticket happens only when you ask, and keeps its original opened time.
• "In short" at the top of the summary. The summary starts with 2-3 plain lines: the problem, what was found, what was done and what is left.
• Declined follow-ups are kept. A follow-up you decline or put off appears on the closed page under "to follow up".
• One step before every page update. A new helper stamps every time, updates the ticket file and runs the page check. The page is published only after the check passes.
• Required offers are recorded. The blanket-permission offer, the restore-point outcome and the scan check are recorded before any change; otherwise the page check blocks the page.
• A Defender scan is offered only when Defender is the active antivirus. Otherwise the skill names the active product from Windows Security Center.
• Closed timer. A closed ticket shows its closing time and total time, never a pause.
• The skill logs its own failures. Failures, time-outs and stalls of the skill's own commands go to its error log.
• Working folders. Any folder outside AppData can hold the working page.
• Stray backups. The helpers warn about leftover backup files in the tickets folder.
• Script-block logs. A fixed read-only helper reads them by time and by plain path parts and prints only short excerpts. The skill no longer writes its own query for them.
• The version is shown in the ticket page header.

v2.8.2 (2026-10-24)

• No guard script and no hooks. The skill's deny and ask categories now live in references/settings-permissions.json, as Claude Code permission rules. It is data, not a script, and has no allow rules, so it never approves anything by itself.
• One optional manual step. A skill cannot install permission rules itself. To turn them on, merge the permissions block into <your IT folder>\.claude\settings.json (recommended) or %USERPROFILE%\.claude\settings.json. Back up the file first and check the result with /permissions. The skill offers this once and never edits a settings file itself.
• Without the merge, the skill's own rules and Claude Code's default of asking before every command that changes something still apply.
• Known limit. Rules match the command as written. Running a PowerShell script found on the PC stays a written rule: it is shown first and waits for your yes.

v2.8.1 (2026-10-24)

• Security-safe operation. The skill uses only signed, built-in Windows tools and cmdlets, in plain readable commands and visible windows. It never hides or obscures command text, never weakens antivirus, firewall or SmartScreen protection or adds scan exclusions, never reads the sign-in process memory or exports credential stores, never clears event logs or deletes shadow copies or restore points, and never changes accounts or boot configuration.
• Any legitimate action that could still raise a security alert (for example a Defender scan that quarantines a file, or a bulk cleanup) gets its own approval with a one-line warning and is never covered by blanket permission. If a security product blocks a step, the skill explains it and lets you decide.
• The skill's own helper scripts no longer start hidden windows, and admin steps run in one visible window.

v2.8 (2026-10-24)

New: - Blanket permission, opt-in. You can let the skill fix what it finds in the current ticket without asking before each action. Nothing is selected by default, and silence never counts as consent. - Scope: the current ticket only, and only reversible low- or medium-risk changes. It expires when the ticket closes, when the session ends, or after 60 minutes of work. - Revoke: a banner on the page has a revoke button, or say so in the chat. - Still applies: every action is still shown on the page and logged. Backups and exports come first, and a restore point is offered before the first change. - Never covered (each always needs its own approval): deleting or overwriting your data, disk, partition or boot changes, drivers, BIOS or firmware, uninstalling software, security settings, anything that needs a UAC prompt, deleting registry keys under HKLM, running scripts found on the PC, and anything whose undo is unknown. - A timer and progress bar for each task. After 5 minutes the row turns red and says the task is taking longer than expected but is still running. The progress bar moves only on real counts. - Security offers. When a check finds a significant security issue, the skill says so and offers to handle it, only if you agree. - Automatic ticket save. When the ticket closes, the page and the ticket file are saved to Documents\IT-Agent-Quick\Reports without asking. An existing file is backed up first. - Restore point before the first change, with a clear result: created, skipped by Windows' one-per-24-hours limit, System Protection off (the skill recommends turning it on as its own approved action, and you decide), or UAC declined with a retry option. - Data-usage wording corrected. The skill no longer says that nothing leaves the PC. Its files stay on the PC; the conversation, including command output, is sent to Anthropic under its data-usage terms. - Offered scans and repairs, each with its own approval: a Microsoft Defender quick or full scan (offered at most once every 5 days per PC), and the system-file repair path (DISM /RestoreHealth, then sfc /scannow), both as timed long operations.

Fixes: - Windows Update checks run in their own process with a hard time limit, so they can no longer hang the session. - The disk-usage scan has a time budget per section and lists folders it could not finish. - App updates (winget) run one package at a time, each with its own limit; before and after versions are read so the result is reliable, and stalled downloads are stopped. - Before a long operation, the skill shows a realistic upper bound for its time. - A registry key is created only when it is missing. - Service restarts use time limits and report a pending stop instead of waiting forever. - The local ticket page no longer uses a meta refresh; it reloads only while a task runs, never while you type, and keeps your scroll position. - With the Claude desktop app, the local page and related files live under Documents\IT-Agent-Quick, never in AppData. - Shortcuts with Hebrew names are read correctly. - The Wi-Fi signal is read in any Windows display language, and the network name is never shown. - Answers on the page are stored in the ticket file; an answered item stays locked after a reload, so an approval can't be clicked twice. - Invisible and bidirectional control characters are removed from anything sent to a command. - The page check runs on every publish; a check that can't run counts as a failure. - A declined or unanswered UAC prompt leaves a "try again" item instead of a silent skip.

Known limits

• Tested on Windows 11 with Windows PowerShell 5.1; not tested on Windows 10.
• Whether the ticket page displays in Claude's built-in panel depends on that panel's sign-in.