read_file

Read a file from the MemDrive memory library. TOTAL RESPONSE BUDGET — THREE LIMITS ON THREE DIFFERENT LAYERS. (1) HOST: the MCP host (Claude Desktop, the only host verified so far) rejects a whole tool result over about 1 MB ("Tool result is too large. Maximum size is 1MB"). That limit lives in the host, outside this server, and is NOT a model or context-window limit; other hosts (ChatGPT, Grok) are untested, so this server assumes the same ~1 MB ceiling for all of them. (2) SERVER: text bytes plus the base64 bytes of every image block in one response must together fit inside a single total budget — default 950,000 bytes, overridable with MEMDRIVE_RESPONSE_BUDGET_BYTES (clamped to 100,000–50,000,000) — a safety margin under the host limit. Text comes first and is guaranteed; images share only the budget the text leaves. An image that does not fit is downscaled in memory to fit it, or skipped with a note — a read never fails as a whole because of an image. The file on disk is never modified. Previews are built at full resolution first, trying higher JPEG quality before shrinking, so a preview uses the budget for sharpness instead of stopping far below it. At the default budget the effective per-image cap is about 712,000 binary bytes (after base64 inflation), so the separate 3.5 MB per-image preview ceiling only binds when the budget is raised for a host with a larger limit. HISTORY: before v0.1.1 this server had NO total budget, so a bundled read whose base64 exceeded the host limit was rejected wholesale by the host — that is what happened to the first read of an 813.5 KB image (~1.11 MB as base64). (3) MODEL: once the model receives an image it scales it by pixels into tokens under its own limits — a separate layer from the byte limits above, not something this server controls. TWO-STEP READING — USE THIS WHENEVER A DOCUMENT REFERENCES SEVERAL IMAGES, OR ONE LARGE IMAGE: Step 1, call read_file with include_images=false to read the text and get the list of referenced image paths. Step 2, call read_file once per referenced image, passing that image path itself as path, so the image alone gets the whole response budget and keeps the best quality the budget allows. Reading an image file directly is also limited by the same total budget: a single image is returned untouched when it fits, and downscaled in memory only when the image itself exceeds the budget. Text files: returned as text (truncated at the smaller of the 1 MB text cap and the total budget, with a note when truncated). Markdown files: with include_images=true (the default), the text is returned TOGETHER WITH every locally-referenced image that fits the remaining budget in the same response — each image becomes an image content block, so the model sees words and pictures together. Remote (http) and data: images are skipped with a note. Images over the 3.5 MB preview cap are downscaled in memory only; the file on disk is never modified. Image files: returned as a single image content block (downscaled in memory only when over the preview cap or the total budget). If an image cannot be fitted within the budget at all, the result is text explaining that, not an error — the original is untouched. SVG files are returned as text (they are XML), never as image blocks. Paths are relative to the library root (MEMDRIVE_ROOT). Absolute paths are accepted only when they stay inside the root.