Problem
The Data & Privacy → Strict mode currently claims:
Warn before reading sensitive files (.env, keys, certs)
The page also shows a Sensitive File Guard section with patterns such as .env, *_key*, *.pem, *.p12, *.pfx, id_rsa, and credentials, and says the agent asks for confirmation before reading matches.
This behavior is not implemented. The patterns and warning are display-only: no privacy-level hook intercepts file reads or opens a confirmation flow. The existing sensitive-shield.ts feature is unrelated; it applies sandbox ACL protection to credential directories and does not implement per-file read warnings.
Why this needs product definition
Before implementation, clarify:
- Which access paths should trigger a warning: shell/tool reads, attachment reads, app-owned reads, or all of them?
- Should matching use filename patterns, full paths, known sensitive directories, content detection, or a combination?
- What decisions should the dialog support: allow once, allow for session, always allow, and deny?
- Should Strict block until explicit approval, or only notify/audit?
- How should non-interactive/background and remote-channel requests behave?
- Is the feature controlled only by Strict mode, or should Sensitive File Guard have its own toggle/configuration?
- Should File Access Audit record approvals, denials, and reads, and where should that audit be visible?
- What are the expected failure-safe semantics when the UI is unavailable?
Current impact
Strict mode promises protection that users do not receive, creating a privacy/security expectation mismatch.
Interim/product options
- Define and implement the confirmation behavior end to end; or
- Remove/reword the Strict-mode and Sensitive File Guard claims until the feature exists.
Acceptance criteria
- PM-approved behavior and scope are documented.
- UI copy accurately reflects implemented behavior.
- If implemented, sensitive-file reads are covered across the approved access paths with explicit fail-safe behavior and tests.
Problem
The Data & Privacy → Strict mode currently claims:
The page also shows a Sensitive File Guard section with patterns such as
.env,*_key*,*.pem,*.p12,*.pfx,id_rsa, andcredentials, and says the agent asks for confirmation before reading matches.This behavior is not implemented. The patterns and warning are display-only: no privacy-level hook intercepts file reads or opens a confirmation flow. The existing
sensitive-shield.tsfeature is unrelated; it applies sandbox ACL protection to credential directories and does not implement per-file read warnings.Why this needs product definition
Before implementation, clarify:
Current impact
Strict mode promises protection that users do not receive, creating a privacy/security expectation mismatch.
Interim/product options
Acceptance criteria