Repository navigation
docs: clarify shell permission limits - #56676
sayedrisat wants to merge 2 commits into
Conversation
|
The following comment was made by an LLM, it may be inaccurate: |
There was a problem hiding this comment.
The new section is accurate and consistent with the rest of the page: same config shape and same 'scanner-produced command string' term as the Actions table. I read it against the base page; it's docs only, so there's nothing to run. The public API is unchanged.
One thing worth fixing: where the section sits (see the inline comment).
| | `opencode_read_mcp_resource` | `<server>:<uri>` | | ||
| | `execute` | `*`; a deny removes Code Mode and the tools reached only through it, while nested tools enforce their rules | | ||
|
|
||
| ### Shell rules do not restrict files a command reads |
There was a problem hiding this comment.
Putting the section here splits the Actions table from its existing follow-up example: "For example, allow one skill and deny all other skills:" (line 144) now sits under the "Shell rules do not restrict files a command reads" heading, so it reads as a shell example. Could you move this section after the skill example (e.g. just before the doom_loop/lsp note), or next to the other shell material further down the page?
Issue for this PR
Closes #56545
Type of change
What does this PR do?
Clarifies that shell permission patterns match the command string, not every file a command reads. The example shows why a filename deny is not a sandbox or a reliable way to protect secrets from recursive shell searches.
How did you verify your code works?
Website typecheck, generated-file check, and build passed. The repository pre-push check passed.
Screenshots / recordings
Not applicable.
Checklist
Added by the opencode bot (AI-generated).
Related: #2242, #55232, #17073