Update to Beeblio

SKILL.md text changes

− removed · + added

--- before
+++ after
@@ -1,27 +1,27 @@
 ---
 name: beeblio-data-analysis
-description: Analyze or clean data from a connected Beeblio project using ChatGPT's own code execution sandbox. Use for tabular cleaning, statistics, reproducible calculations, and research findings based on Beeblio project files.
+description: Analyze or clean data in a connected Beeblio project with its MCP analysis jobs. Use for tabular cleaning, statistics, reproducible calculations, and research findings based on project files.
 ---
 
 # Analyze Beeblio project data
 
-Use Beeblio MCP to identify the project and input files, then use an available ChatGPT code execution sandbox for computation. `run_analysis` is part of Beeblio's in-app agent and is not an MCP tool. Do not claim that code ran in Beeblio or that sandbox files automatically appear in the Beeblio project.
+Use Beeblio MCP to identify the project and input files, then run project analysis with `submit_analysis_job`. The job runs Python or shell commands in a disposable Beeblio compute workspace. It does not access unsaved editor content. Analysis uses account credits.
 
-## Move data into the sandbox
+## Prepare project inputs
 
 1. Resolve the project with `list_projects`; inspect candidate inputs with `list_files` and `stat_file`. Paths sent to MCP tools are project-relative.
-2. For text inputs, call `read_text_file` in successive bounded slices until `nextOffset` is absent. Preserve the bytes and structure of CSV, TSV, JSON, or other source text when placing it in a sandbox file; do not manually retype rows into code. Record the source path and file metadata.
-3. For binary or large inputs, `get_download_url` can provide a short-lived direct URL. Download it into the sandbox only if that sandbox can access the URL. The URL expires after 15 minutes. If transfer is unavailable, explain the limit and use a user-provided file or another available route. Do not claim to have analyzed an inaccessible file.
+2. Inspect the relevant source text with `read_text_file` or use a bounded format reader such as `read_office` or `read_matrix`. Record source paths and metadata. Avoid copying large datasets into the conversation.
+3. Pass required project files in `inputPaths` as absolute `/workspace/` paths, for example `/workspace/2-Data/source.csv`. The compute job mounts those saved inputs. If a source is unavailable, explain the limit rather than claiming to have analyzed it.
 
 ## Compute and verify
 
-- Keep raw input unchanged. Save substantial code as a script in the sandbox, run it, and keep the script with outputs for reproducibility. Use only libraries actually available in that sandbox; Beeblio's batch-image packages and `beeblio_research` are not guaranteed there.
+- Keep raw input unchanged. Save substantial code under `3-Analysis/`, then run it with `submit_analysis_job` using a unique `idempotencyKey`, a suitable `timeoutSeconds` (at most 300), and explicit `inputPaths`. Poll `get_analysis_job` until it succeeds or fails; call `cancel_analysis_job` if the user asks to stop. Check the returned exit code, logs, and output paths. Verify libraries in the compute image before relying on them.
 - For cleaning, inspect schema, units, missing-value markers, duplicates, ranges, and outliers before transforming. Choose and explain missing-data and outlier treatments. Write cleaned data to a new file, never over the raw source. Report input and output row counts and the material changes.
 - For statistics, state the research question, unit of analysis, sample size, missing-data treatment, model or test, assumptions checked, effect size, and uncertainty. Validate key counts and at least one important result independently. Treat observational associations as associations unless the study design supports causal language. A nonsignificant result is not proof of no relationship.
 - Do not paste a full source dataset into the conversation or generated code. Present a compact result table or narrative with the actual computed values and limitations.
 
-## Return outputs to Beeblio
+## Save and report outputs
 
-- Save a requested script under `3-Analysis/` and a Markdown report under `4-Reports/` with `write_text_file` when each is under 2 MiB. Create with `ifNotExists`; read an existing file and use its ETag before replacing it. The current MCP rejects `write_text_file` under `2-Data/`.
-- Pass each derived dataset, chart, or other generated artifact as an actual ChatGPT file to `upload_artifact`. Put cleaned datasets under a new path in `2-Data/derived/`, scripts and figures under `3-Analysis/`, and final reports under `4-Reports/`. The tool accepts files up to 25 MB and never overwrites. Confirm each returned path and size with `stat_file` before saying it is in Beeblio.
-- Report which outputs are in Beeblio and which remain in the ChatGPT sandbox, along with input paths, code path, and any transfer limits.
+- Write job outputs to new paths under `/workspace/2-Data/derived/`, `/workspace/3-Analysis/`, or `/workspace/4-Reports/`. Confirm uploaded outputs with `stat_file` or `read_text_file`. Use `write_text_file` with `ifNotExists` for small scripts and reports; it does not write raw `2-Data/` files.
+- Use `upload_artifact` when an external sandbox produced a file or when a user supplies one. Its 25 MB limit and create-only behavior still apply.
+- Report actual input paths, code path, output paths, status, methods, and limitations. A queued job or saved script alone is not a completed analysis.