Agent Git permission boundary Requirements

Overview

Users can edit files and read Git status in a restricted agent mode, then see the agent fail when it runs git add or git commit. The failure can make the repository or Kandev appear broken, even though the Changes panel can commit the same work through a different permission path.

Requirements

REQ-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001: Agent Git permission boundary

Intent: Users can edit files and read Git status in a restricted agent mode, then see the agent fail when it runs git add or git commit. The failure can make the repository or Kandev appear broken, even though the Changes panel can commit the same work through a different permission path.

Acceptance criteria

AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.1: Public Git documentation explains that Changes-panel Git operations and agent-shell Git commands use different permission paths.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.2: The documentation explains that git status can work while a shell commit fails because the agent cannot write Git metadata such as .git/index.lock.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.3: The documentation directs users to commit from the Changes panel without changing the agent mode.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.4: The documentation explains that a user who needs the agent shell to commit must select an agent mode that allows Git metadata writes.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.5: The documentation states that mode names come from the installed agent and uses Codex mode names only as examples.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.6: GIVEN an agent edits files in a restricted mode, WHEN git status works but git commit fails with an .git/index.lock error, THEN the public Git guide explains the permission boundary and the Changes-panel recovery path.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.7: GIVEN a user opens the Changes panel with agent edits present, WHEN the user chooses Commit Changes, THEN the public Git guide explains that Kandev performs the Git operation outside the agent shell sandbox.
AC-AGENTS-GIT-OPERATIONS-PERMISSION-BOUNDARY-001.8: GIVEN a user wants the agent shell to commit, WHEN the user reads the public Git guide, THEN the guide explains that the user must choose an agent-specific mode that permits Git metadata writes.

Migrated source detail

Why

What

Public Git documentation explains that Changes-panel Git operations and agent-shell Git commands use different permission paths.
The documentation explains that git status can work while a shell commit fails because the agent cannot write Git metadata such as .git/index.lock.
The documentation directs users to commit from the Changes panel without changing the agent mode.
The documentation explains that a user who needs the agent shell to commit must select an agent mode that allows Git metadata writes.
The documentation states that mode names come from the installed agent and uses Codex mode names only as examples.

Scenarios

GIVEN an agent edits files in a restricted mode, WHEN git status works but git commit fails with an .git/index.lock error, THEN the public Git guide explains the permission boundary and the Changes-panel recovery path.
GIVEN a user opens the Changes panel with agent edits present, WHEN the user chooses Commit Changes, THEN the public Git guide explains that Kandev performs the Git operation outside the agent shell sandbox.
GIVEN a user wants the agent shell to commit, WHEN the user reads the public Git guide, THEN the guide explains that the user must choose an agent-specific mode that permits Git metadata writes.

Out of scope

Changing agent sandbox behavior or permission modes.
Changing the Changes-panel Git implementation.
Adding a new Kandev-wide permission-mode name.
Changing Git error handling or automatically changing a session mode.

4.4 KiB Raw Blame History