--- name: prepare-dragon-run description: Prepare a Helio Dragon Assess or Enhance plan without starting a thermal simulation or optimization. --- Use this skill when the user wants to assess or enhance valid G-code, or asks what Dragon needs before a run. Call `prepare_dragon_run`; preparation may register and validate G-code, but it must never create a simulation or optimization. 1. Inspect the tool schema and preserve the user's original G-code, exact printer/material IDs, operation, and explicit options. Use a fresh valid slice, not an STL or 3MF. If a Helio result replaced a working file and an `original_` exists, use that original. Never overwrite or modify it. Use `dragon_catalog` (`resource: printers` or `materials`) and returned IDs when discovery is needed; follow `resource.pageInfo.hasNextPage` with `page`/`page_size` (default 1/100, maximum 100), rather than inventing a fuzzy matcher. `SLICE`, `PARSED_GCODE`, and `USER` describe caller input only; they are not proof of a value or identity. 2. For a new local source, stream the complete original file through SHA-256\. Call the core `get_presigned_url` tool with the literal filename `test.gcode`. Stream the file bytes with HTTP PUT to the returned upload URL using the exact returned `mimeType` as the `Content-Type` header, and require a successful HTTP response before registering. Do not hardcode a MIME type: the signed upload may require the returned value. If `mimeType` is missing, report the incomplete upload response rather than guessing. Do not add authorization headers to the signed upload URL, print or persist the URL, or put file contents in chat. An upload is not registration or run creation. 3. Set `preparation.context.source` using the tool's returned `key` as `immutable_key`, the actual original filename as `original_name`, the calculated `sha256`, and an empty `id` for a new registration. For an already registered original source, verify and use its actual ID/key instead of uploading again. Let the MCP verify exact catalog IDs and registered mappings; a filename, source label, or embedded profile label is not proof. Obtain other facts from existing slicer/reader output or explicit user input; never invent them, build a parser, or claim a backend change is required merely because a fact is missing. 4. Choose one stable preparation `identity`, retain the exact source context, and call `prepare_dragon_run`. Preparation may register and poll G-code but never starts Assess/Enhance. Reuse that identity and unchanged preparation on retries or in-progress registration; do not upload/register a replacement after an ambiguous response. The service may promote the upload key internally; never guess a registered resource to recover it. 5. Ask only for fields the planner reports as missing. Never manufacture an ID, temperature, layer range, speed, flow bound, or timing statistic. For Assess, omit unavailable slice statistics and environment fields so backend defaults apply; the printer may provide only its exact `printer_id`. An explicit environment override needs only the original slot-0 first-layer bed temperature for the actual bed and the known original chamber mode; do not infer either. 6. For Enhance, obtain `object_layer_count` from existing slicer/reader output or explicit user input and default layers to `2..object_layer_count` when none are requested. Explicit advanced mode requires all four speed/flow bounds but not nozzle/layer dimensions. With `HELIO_DEFAULT`, a known 0.2 mm nozzle, layer height below 0.2 mm, or reference printer/material mapping already forces advanced mode; obtain the bounds without asking for unrelated missing dimensions. Otherwise obtain original nozzle diameter and process layer height to decide whether defaults are permitted. Timing statistics are optional and used only for an estimate. 7. For `NEEDS_CONFIRMATION`, explain each substitution or warning and obtain its required acknowledgment. Do not describe a reference printer/material as the user's actual selection. 8. For `BLOCKED` or `NEEDS_INPUT`, report the reason and smallest user-correctable next step. Do not attempt a degraded standard run. A prepare-only request stops with the plan; when the user already requested execution and the plan is ready, continue through the start skill without requesting redundant permission. For Enhance, do not silently change planner-selected layer bounds or limit modes. For Assess, do not expose privileged result capabilities. Return the plan identity and resolved summary so a later explicit start can use exactly that plan.