Changelog - v1.35.6

🐛 Bug Fixes

A second launch can no longer touch the installation state of the running app. Opening
TensorPM again while it was already running — by double-click, by a link, or through the
installer — let the second process run the version check against the same data directory, in rare
cases entering the uninstall path. Only the instance that actually starts does this now.
The project agent no longer receives its instructions twice. Relevance checks, wizard feedback
and document analysis prepended their instructions to the request while the app also sent its
general ones, which cost tokens and blurred the task. Each request now carries exactly one set of
instructions.
Claude- and Codex-subscription accounts get the right instructions again. On these two
connection types the task-specific instructions for relevance checks, wizard feedback and document
analysis were not applied.
The daily automatic analysis no longer starts twice. Right after startup, while sign-in and
entitlements were still settling, a second run could begin in parallel with the first.
Updates download once instead of twice. Every update first attempted a partial download that
could not succeed with our packages, logged an error, and then downloaded the full package anyway.

Changing many work packages at once is now a single step. Shifting all due dates by a week
previously produced one request per item — eighteen for a mid-sized project, each one repeated in
every following round of the conversation. The agent now sends one request covering all items,
including items that need different values. This is noticeably faster and consumes fewer AI
credits.
A single bad value no longer discards a whole batch. Valid changes are applied and the
problematic ones are reported back, instead of throwing everything away and retrying.
Wizard feedback prompts are shorter. The field under review was sent twice — once as the value
to review and once as project context. On long descriptions this alone cost about 2,000 tokens per
review.