# Bambu cloud: what we have actually proved
An agent can observe a printer without running a new service. On **2026-11-07**, a normal account login, existing Python libraries and the installed Bambu media runtime produced fresh P1S/X1E telemetry and a real P1S camera frame. X1E video remains unresolved. This is a connection experiment, not a shipped printer-control command.
Production work belongs inside the existing [manufacturing hierarchy](https://github.com/Valpatel/cad-agent/blob/dev/software/cadagent/manufacturing/README.md#where-machine-automation-belongs): separate artifact preparation, device adapters, execution and evidence. Bambu is one integration; its cloud protocol must not define the workflow for every machine.
```mermaid
flowchart TD
A["account login"] --> O["owned printer"]
O --> S["fresh status"]
O --> G["camera grant"]
G --> R["printer ready"] --> F["decoded frame"]
classDef done fill:#eef1e4,stroke:#173d3c,color:#173d3c
classDef check fill:#fff1df,stroke:#bd3a12,color:#173d3c
class A,O,S done
class G,R,F check
```
| Evidence | P1S | X1E |
|---|---|---|
| Normal cloud login and owned-device listing | Verified | Verified |
| Fresh status and matched firmware reply | Verified | Verified |
| Actual decoded cloud camera frame | 1280 × 720 JPEG | Not obtained |
| Motion, heating, print submission | Not tested | Not tested |
The X1E returned a camera grant, then rejected preparation of that **same grant** with `84033543`. Its camera works in the official app, according to the owner. A grant is therefore insufficient evidence of a working stream. The error's precise meaning in this camera context is not established by the official error tables.
The intended application is a manufacturing experiment loop: generate technique-test specimens, stamp their part
numbers, pack plates, verify the printer and a clear bed, submit the job, and connect physical feedback to each
specimen's exact design and settings. An occupied or uncertain bed must stop submission. The observations above
establish some prerequisites; they do not establish that the complete loop is implemented.
Each stamped short ID should resolve to a full UUID and a version-controlled experiment record: immutable source
repository/commit, model parameters, geometry identity and manufacturing settings. Check short IDs for collisions.
Record print attempts and human function/durability feedback against the exact specimen. Source identity remains
fixed as later observations guide the next design; the source commit precedes the generated record, avoiding a
self-referential commit hash.
Repeatable investigation method
1. Use normal sign-in and verification. Keep credentials in an owner-only local session outside the repository. Verify the owned-device list before selecting a device.
2. Connect to the correct regional cloud MQTT broker with verified TLS. Subscribe only to the selected owned device's report topic. A subscription acknowledgement or empty heartbeat does not prove current printer state.
3. After acknowledgement, request one standard `pushing.pushall` and one `info.get_version`. Require a fresh, nonretained state and a version response matching the request's sequence ID. Merge sparse deltas; retain model-specific temperature decoding.
4. Select the camera protocol from actual firmware and the official model compatibility data. Observe status while requesting a normal cloud camera grant. Correlate preparation replies by the grant identifier; timing alone is insufficient.
5. Start a bounded reader only with readiness evidence. Decode the actual image and inspect it. Preserve unsuccessful stages separately. A URL, byte buffer or successful mock is not a frame.
Keep raw device reports, grants and camera images private. Record dated, sanitized outcomes and exact source identities publicly. These steps describe the reviewed experiment; there is no production CLI that automatically performs them yet.
Existing code and agent skills considered
| Source | Useful part and observed limit |
|---|---|
| [Bambu Studio](https://github.com/bambulab/BambuStudio/tree/355d71eee5a9ebc794693fe57e63674806019770) | Authoritative request shapes, model compatibility and media ABI. Installed media runtime decoded the P1S frame. Its networking SDK did not establish our fresh local session; that is not a cloud-login failure. |
| [Bambuddy](https://github.com/maziggy/bambuddy/blob/75c9c2db740b74d17803c82119803f6ae7088783/backend/app/services/bambu_cloud.py) | Normal account/verification flow informed the successful small probe. No server installation was needed. |
| [voska's Bambu skill](https://github.com/voska/bambu/blob/72fde5d67e2ceba27d1ef03c9031800f2dca2722/skills/bambu/SKILL.md) | Existing Go CLI and LAN workflow; inspected version does not supply this cloud-camera path. |
| [Printing Press Bambu skill](https://github.com/mvanhorn/printing-press-library/blob/a672844a2e7dca35b6e0a7828e00d61ecb32302d/library/devices/bambu/SKILL.md) | Read-only local telemetry patterns; inspected version has no cloud camera/control. |
| [Text-to-CAD Bambu skill](https://github.com/earthtojake/text-to-cad/blob/4099cac3ddc56a5e696138c53fd7bee685e8ad9d/skills/bambu-labs/SKILL.md) | Official-client model handoff, with the person choosing the printer and starting the job. |
None of these skills was installed. Their source inspection does not qualify an untested printer or justify transferring their dependencies, licenses or transport assumptions into cad-agent.