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
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.
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.
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.
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.
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.
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).
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.
1---2name: sync-ml-reports3description: >4 Copy skore reports between local, Hub, and MLflow, and optionally switch the5 upload destination. Trigger when the user asks to sync or migrate reports,6 switch skore mode, upload reports to Hub or MLflow, or pull Hub/MLflow7 reports onto disk. First G-SKORE-MODE pick stays evaluate-ml-pipeline.8metadata:9 source: https://skillmd.com/skills/probabl-ai/sync-ml-reports10 modelTier: medium11---1213# Sync ML Reports1415Copy reports with the `skore` CLI. Switch the default destination16only when the user asked to. Do not evaluate, audit, or invent17`Project.sync` Python.1819## Human-facing prose2021Details: `setup-workspace` `references/human_facing_prose.md`.22Ask where reports live (disk, Hub, MLflow) in those words. Do not23name `G-SKORE-MODE`, skill ids, or the wrapper CLI in the24question. `skore sync` output may appear in the close as the25sync table.2627## Procedure28291. Run `python -m skore_skills status`. Read `policy.skore_mode`30 and `skills`. Open `experiments/` and `audit/` for the31 `skore.Project(...)` init block: `name=`, Hub `workspace=`,32 MLflow `tracking_uri=`. Local store is always the workspace33 `reports/` directory (absolute path). Never omit34 `--from-workspace` / `--to-workspace` on a **local** endpoint.35362. If `policy.skore_mode` is unset: STOP. First pick is37 G-SKORE-MODE in `evaluate-ml-pipeline`. Do not ask local / hub38 / mlflow here. Load that skill only if39 `status.skills.evaluate-ml-pipeline` is true and the user40 asked to evaluate; else one-line skip. The close is only this41 stop: the destination is not chosen yet and is picked when a42 report is stored. Do not list local, Hub, or MLflow, a43 workspace name, or a tracking URI. Do not use the44 source-to-destination close below.45463. **AskUserQuestion** for any answer not already in the request.47 Ahead of each question, state in 2–4 lines what the answer48 authorizes — which reports move where, whether `skore_mode` and49 the Project init lines get rewritten — and the facts it rests50 on: the current mode, the discovered report count, the51 endpoint. A file link is an addition, never the context.5253 - **Intent.** Switch default destination (sync, then persist54 `skore_mode` and rewrite every Project init) vs copy only55 (sync, leave policy and experiment files).56 - **Destination** — same three options as G-SKORE-MODE:57 `local` (disk, no account), `hub` (https://skore.probabl.ai),58 `mlflow` (tracking server). Hub: ask the workspace name; it59 MUST NOT contain `/`. MLflow: ask `tracking_uri`; confirm a60 bare `host:port` as `http://host:port`. Do not default the61 URI.62 - **Project name** if `name=` is missing or disagrees across63 files.64 - **Dry-run first** vs transfer now.6566 If intent is **switch** and destination equals67 `policy.skore_mode`, stop in one line.68694. Destination extras: load `add-python-package` only if70 `status.skills.add-python-package` is true, for Skore at the71 **destination** mode (`env add-skore --mode <dest>`). If that72 skill is missing, name Skore for the destination and stop. Do73 not splice `pip install` / `skore[...]` here.74755. If Hub is source or destination: require `SKORE_HUB_API_KEY` in76 the environment. Missing → name it and stop. Do not open a77 browser login. Do not read `.skore` for the key.78796. Build `skore sync`. If `skore` is not on PATH, name80 `skore-cli` and stop. Do not call `skore.Project.sync` in81 Python.8283 ```bash84 skore sync <project> --from=<source_mode> --to=<dest_mode>85 ```8687 Source mode is `policy.skore_mode`. Add:8889 | Endpoint | Flags |90 |---|---|91 | local | `--from-workspace` or `--to-workspace` = resolved `reports/` |92 | hub | `--from-workspace` or `--to-workspace` = Hub workspace name (required) |93 | mlflow | `--tracking-uri=...`; never `*-workspace` |9495 `--to-project` only if the destination name differs.96 `--hub-url` only when `SKORE_HUB_URI` (or the user) names a97 non-default Hub. `--both` only if the user asked to copy98 missing reports both ways. Otherwise one-way.99100 If the user picked dry-run first, run with `--dry-run`, show101 the plan, then ask to transfer. Live run omits `--dry-run`.102103 Usage/auth/backend errors: name stdout/stderr and stop. Do not104 invent a Python fallback. Empty output `No reports to105 synchronize.` is success (nothing to copy; a **switch** may106 still continue).1071087. **Switch intent only**, after a successful (or empty) sync:109 `python -m skore_skills policy set skore_mode <dest>`.110 Rewrite every Project init in `experiments/` and `audit/` to111 the destination form in112 `evaluate-ml-pipeline/references/g_skore_mode.md` (audit must113 match the paired experiment, byte-for-byte modulo formatting).114 If dest is **local**, `mkdir reports` (`exist_ok`); no README.115 If **hub** or **mlflow**, do not create `reports/`.116 If `journal/JOURNAL.md` exists, insert a `---` under History117 and one line `skore_mode: <old> → <dest> (sync-ml-reports)`.118 Missing journal → skip in one line; do not paste a JOURNAL119 body.120121 **Copy-only:** do not `policy set`, rewrite init, mkdir, or122 edit JOURNAL.123124## Stop conditions125126- Do not steal first G-SKORE-MODE when `skore_mode` is unset.127- Do not silently change a recorded mode; switch requires the128 switch intent (or an explicit user request to switch).129- Do not `git commit`.130- Do not evaluate, `project.put`, or audit.131- Do not pass `*-workspace` on an MLflow endpoint.132- Do not omit `*-workspace` on a local or Hub endpoint.133134## End of turn135136### User-facing close137138This close applies only after a sync. An unset `policy.skore_mode`139uses the step-2 stop instead.140141Short story: source → destination, whether policy changed, and142the `skore sync` table or `No reports to synchronize.` Do not143dump experiment files.144145Then `python -m skore_skills git end-turn --stage evaluate`146(the persist bucket for this work; this is not a CV run). If147JSON `action` is `invoke`, load `persist-ml-git` only if148`status.skills.persist-ml-git` is true and stop; that skill149returns to triage. If persist is missing, name the pending150`staged` paths and stop. Otherwise load `triage-ml-task` only if151`status.skills.triage-ml-task` is true; else stop. No `git152commit`.