Repository navigation
docs: say where an agent can do harm, and add a Claude Code skill - #5
Conversation
|
The text is good. The This depends on #1 through #4, so I'll rebase it once those land. If #4 takes longer, the firmware section splits out so the rest isn't held up. One change I'm making: I'm leaving "you are expected to act, not to ask permission for every step" out of the MCP instructions. Everything else in the project leans the other way, and the two places where an agent acts on the account or the firmware already say "only when the user asks". The skill file can keep its version of the sentence, since someone installs that deliberately. |
Depends on `pr/status-timing`, `pr/rtsp-camera`, `pr/connect-firmware` and `pr/manual-login`. They are merged below this commit. The skill describes their commands, and this commit edits the camera and firmware tool descriptions. When those four have landed, review only this commit. The MCP instructions and tool descriptions said what each tool does. They did not say where a tool can cause harm. Now they say it where it matters: - `plate_clear` and `set_ready` are the claim of the agent, and nothing checks them. Look at a fresh `get_camera_snapshot` or ask the user. Never set the flag only to pass a refusal. - `run_gcode` is not a console. It overwrites `/usb/prusactl-macro.gcode` and starts it as a real print job. It does not check the content. - `api_request` is a raw escape hatch. It has no safety checks. - Deletes of printer files, Connect files and queue entries are permanent. - If a print looks wrong, pause first. Stop is final. Use it for a clear failure or when the user says so. `skills/prusactl/SKILL.md` is a Claude Code skill. It covers the status, start, monitoring, login and firmware workflows, with the same hazards. The README points to the skill and lists the limit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Petr Štetiar <ynezz@true.cz>
The agent guidance added "You are expected to act, not to ask permission for every step" to the instructions every MCP client reads. Everything else in prusactl leans the other way: the plate question, the camera look before a start, and "only when the user asks" for sign-in and firmware. The sentence stays in the skill file, which someone installs on purpose; the server's own instructions keep to the cautions. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
52cdd06 to
a2961ab
Compare
|
Merged in v0.3.0. As said, I dropped "you are expected to act, not to ask permission for every step" from the server instructions and left it in the skill, which people install on purpose. The rest went in as you wrote it. Thanks for all five. |
2 participants
hi, my clankers found this nice project during research, I liked it so decided to make you aware about the changes I've made for my use cases, so everyone can profit from that, but feel free to close it or implement it differently, happy to help with testing, cheers, Petr
Depends on #1, #2, #3 and #4. Until they are merged, this PR also shows their commits. Please review only the last commit.
The MCP instructions and tool descriptions said what each tool does. They did not say where a tool can cause harm. Now they say it where it matters:
plate_clearandset_readyare the claim of the agent, and nothing checks them. The agent must look at a fresh snapshot or ask the user. It must never set them only to pass a refusal.run_gcodeoverwrites/usb/prusactl-macro.gcodeand starts it as a real print job.api_requesthas no safety checks.skills/prusactl/SKILL.mdis a Claude Code skill. It covers these workflows, with the same hazards:login --manualfor agentsTesting
go test ./.... The changes are descriptions and docs only.