AI Work Quality v0.5 — enabled organizations get updated session scorecards, including Agent leverage and Security, evidence-backed judgments, and team and engineer trends without minimum-sample gates. Outcomes are shown separately from scores, and coverage shows how much evidence each view uses.
Rulebook editing — change a rule's audience, repositories, paths, exclusions, and trigger through inline controls or a plain-language request. Review the before-and-after changes before saving a proposal or activating an allowed edit; searchable repository lists and path presets make scope selection easier.
Improvements2
Rule relevance — the rule judge now applies its relevance decision by default, withholding low-fit advice and lifting a gate when the current turn is clearly unrelated.
Execution sandbox — updated the spreadsheet-writing library used by code execution, alongside the backend runtime library refresh.
Fixes2
Session scores — the Sessions list refreshes scores as later judge batches arrive and clears stale cached scores. Quality session tables keep titles readable, with outcome, security, and boost details on the session page.
Captured checkpoints — preserve validated rewind checkpoint metadata while redacting the rest of an imported transcript, so capture no longer removes the boundaries needed to compare checkpointed sessions.
Security5
Cookie choices — an English/French banner offers Accept all and Refuse all, with Cookie settings available from the app, login, and developer portal and a public Cookie Policy. Analytics cookies and advertising attribution require current consent; refusal permits only cookieless PostHog counting.
Consent lifetime — choices sync across tabs, are requested again after six months, and keep a fixed thirteen-month tracker lifetime. Withdrawing consent clears tracking storage and advertising attribution; accepting again does not extend existing tracker expiry.
Transcript privacy — redact recognized credentials and personal information from newly captured transcripts, with additional secret-redaction checks when messages are stored. Existing history is not automatically rewritten.
Dependencies — upgraded the libraries behind request validation, request tracing, and domain-name handling to their patched releases, closing published advisories: requests with unexpected fields are rejected again instead of slipping past validation, and crafted trace headers can no longer slow the server down.
Document uploads — replaced the end-of-life PDF library used to read uploaded PDFs with its maintained successor, addressing known text-extraction vulnerabilities in the old dependency.