Sync ML Reports

Copy reports with the skore CLI. Switch the default destination only when the user asked to. Do not evaluate, audit, or invent Project.sync Python.

Human-facing prose

Details: setup-workspace references/human_facing_prose.md. Ask where reports live (disk, Hub, MLflow) in those words. Do not name G-SKORE-MODE, skill ids, or the wrapper CLI in the question. skore sync output may appear in the close as the sync table.

Procedure

1. Run python -m skore_skills status. Read policy.skore_mode and skills. Open experiments/ and audit/ for the skore.Project(...) init block: name=, Hub workspace=, MLflow tracking_uri=. Local store is always the workspace reports/ directory (absolute path). Never omit --from-workspace / --to-workspace on a local endpoint.
2. If policy.skore_mode is unset: STOP. First pick is G-SKORE-MODE in evaluate-ml-pipeline. Do not ask local / hub / mlflow here. Load that skill only if status.skills.evaluate-ml-pipeline is true and the user asked to evaluate; else one-line skip. The close is only this stop: the destination is not chosen yet and is picked when a report is stored. Do not list local, Hub, or MLflow, a workspace name, or a tracking URI. Do not use the source-to-destination close below.

3. AskUserQuestion for any answer not already in the request. Ahead of each question, state in 2–4 lines what the answer authorizes — which reports move where, whether skore_mode and the Project init lines get rewritten — and the facts it rests on: the current mode, the discovered report count, the endpoint. A file link is an addition, never the context.

   • Intent. Switch default destination (sync, then persist skore_mode and rewrite every Project init) vs copy only (sync, leave policy and experiment files).
   • Destination — same three options as G-SKORE-MODE: local (disk, no account), hub (https://skore.probabl.ai), mlflow (tracking server). Hub: ask the workspace name; it MUST NOT contain /. MLflow: ask tracking_uri; confirm a bare host:port as http://host:port. Do not default the URI.
   • Project name if name= is missing or disagrees across files.
   • Dry-run first vs transfer now.

   If intent is switch and destination equals policy.skore_mode, stop in one line.

4. Destination extras: load add-python-package only if status.skills.add-python-package is true, for Skore at the destination mode (env add-skore --mode <dest>). If that skill is missing, name Skore for the destination and stop. Do not splice pip install / skore[...] here.
5. If Hub is source or destination: require SKORE_HUB_API_KEY in the environment. Missing → name it and stop. Do not open a browser login. Do not read .skore for the key.

6. Build skore sync. If skore is not on PATH, name skore-cli and stop. Do not call skore.Project.sync in Python.

   skore sync <project> --from=<source_mode> --to=<dest_mode>

   Source mode is policy.skore_mode. Add:

   Endpoint	Flags
   local	--from-workspace or --to-workspace = resolved reports/
   hub	--from-workspace or --to-workspace = Hub workspace name (required)
   mlflow	--tracking-uri=...; never *-workspace

   --to-project only if the destination name differs. --hub-url only when SKORE_HUB_URI (or the user) names a non-default Hub. --both only if the user asked to copy missing reports both ways. Otherwise one-way.

   If the user picked dry-run first, run with --dry-run, show the plan, then ask to transfer. Live run omits --dry-run.

   Usage/auth/backend errors: name stdout/stderr and stop. Do not invent a Python fallback. Empty output No reports to synchronize. is success (nothing to copy; a switch may still continue).

7. Switch intent only, after a successful (or empty) sync: python -m skore_skills policy set skore_mode <dest>. Rewrite every Project init in experiments/ and audit/ to the destination form in evaluate-ml-pipeline/references/g_skore_mode.md (audit must match the paired experiment, byte-for-byte modulo formatting). If dest is local, mkdir reports (exist_ok); no README. If hub or mlflow, do not create reports/. If journal/JOURNAL.md exists, insert a --- under History and one line skore_mode: <old> → <dest> (sync-ml-reports). Missing journal → skip in one line; do not paste a JOURNAL body.

   Copy-only: do not policy set, rewrite init, mkdir, or edit JOURNAL.

Stop conditions

• Do not steal first G-SKORE-MODE when skore_mode is unset.
• Do not silently change a recorded mode; switch requires the switch intent (or an explicit user request to switch).
• Do not git commit.
• Do not evaluate, project.put, or audit.
• Do not pass *-workspace on an MLflow endpoint.
• Do not omit *-workspace on a local or Hub endpoint.

End of turn

User-facing close

This close applies only after a sync. An unset policy.skore_mode uses the step-2 stop instead.

Short story: source → destination, whether policy changed, and the skore sync table or No reports to synchronize. Do not dump experiment files.

Then python -m skore_skills git end-turn --stage evaluate (the persist bucket for this work; this is not a CV run). If JSON action is invoke, load persist-ml-git only if status.skills.persist-ml-git is true and stop; that skill returns to triage. If persist is missing, name the pending staged paths and stop. Otherwise load triage-ml-task only if status.skills.triage-ml-task is true; else stop. No git commit.