Manage ML Backlog

Replace iterate-as-cadence. Do not own setup, exploratory data analysis, build, smoke, evaluate, or audit methodology.

Human-facing prose

Details: setup-workspace references/human_facing_prose.md. JOURNAL rows, design-note Status / Results, and # comments describe this experiment's outcome — not the skills framework, the CLI, or the command that produced an output. Questions and replies use the same data-science language — not skill ids, G-* names, or the wrapper CLI. <!-- results-embed: … --> is a site marker. Authoring hints stay in this skill. style is ruff only.

The CLI writes JOURNAL with sections in order: Status, Data understanding, Modeling decisions, History, Ideas, Backlog. History, Ideas, and Backlog start as header-only tables. Column contracts stay here (Stem, Intent, Status, Headline result, Report, Design note; Question, Status, Experiment, Source; #, Item, Source). Do not put those contracts back into HTML comments in the file.

Model-entry selection mode

When model-ml-pipeline calls with the backlog array from python -m skore_skills model choices:

1. When the user has not named a B<N>, present exactly those rows in their returned order and AskUserQuestion for one pick. One row, or a message that already names a B<N>, is the pick: do not ask again. Carry each row's Item and Source as the option context, and say in 2–4 lines what the pick authorizes (a design note to approve) and what it does not (no model code yet). A file link is an addition, never the context. Do not rescan into a different menu and do not add an idea.
2. Turn the selected row's Item + Source into a Proposal. Ask only for missing shaping facts; do not invent a Method from a one-line item. Do not ask Yes / No on the proposal.
3. Return that proposal to model-ml-pipeline. Do not ask to confirm it. After the model stage creates and populates the design note, remove only the selected Backlog row and add the planned History row. Preserve every other stable B<N> index.

This mode does not require a report/audit digest and does not run the outcome-recording procedure below. Empty Backlog is a routing error: return to model-ml-pipeline; do not fabricate B1.

Record-outcome mode

This mode is only a dispatch from model-ml-pipeline, evaluate-ml-pipeline, or audit-ml-pipeline at that caller's end of turn. A user saying the run finished, or "record it", is a direct Procedure turn: record the outcome, then End of turn (site build --if-stale when site policy is true, then git end-turn --stage backlog). Do not treat that user request as this mode.

When model-ml-pipeline, evaluate-ml-pipeline, or audit-ml-pipeline calls at end of turn with the normalized G-REPORT-LOCATOR, optional headline, and G-AUDIT-FINDING. This is the only path that records an outcome without a full backlog turn. Audit may have been skipped; the locator remains required and the finding becomes n/a — audit not run. If the caller omitted the locator, run python -m skore_skills loop locator --stem <stem> and paste JSON locator. If it omitted the finding, run python -m skore_skills audit finding --stem <stem> (stop → n/a — audit not run / n/a — audit digest unavailable). Do not rephrase either string.

Run Procedure steps 1-3 and nothing else:

1. Step 1 — python -m skore_skills status; require an approved stem.
2. Step 2 — read journal/JOURNAL.md; scaffold the index if it is missing.
3. Step 3 — update the matching History row and design-note Status block from the digest or user-supplied headline when available. Paste G-REPORT-LOCATOR and G-AUDIT-FINDING verbatim into their separate Status lines. Do not change State or Approved by user on. Also refresh the journal/JOURNAL.md Status rows Last experiment and Last result. Insert or replace ## Results between Status and Notebooks from digest text, not HTML. ### Metrics is a heading only.

Then return to the caller. Do not read journal/ideas/, do not edit the Ideas table, and do not open the idea-triage menu — the caller did not ask what to try next.

Do not dispatch audit-ml-pipeline in this mode; the digest is already in hand and dispatching would bounce back here. Do not run this skill's End of turn either: the caller owns convert / site / git end-turn and the User-facing close. This mode writes History; it does not replace the caller's chat close.

The Procedure guards still bind. Never mark done while smoke is red, and never invent a metric — no digest and no user-supplied value means use n/a, not a guess. Never construct a missing backend URL. Record n/a — backend did not expose a locator in both markdown destinations when the digest has no authoritative locator.

Procedure

1. Run python -m skore_skills status. When recording a done outcome, run python -m skore_skills design consent --stem <stem>. ask / stop → do not mark done. Also require green smoke evidence and a normalized report locator. An audit digest is optional. G-AUDIT-FINDING is required as one of: the value returned by audit, n/a — audit not run, or n/a — audit digest unavailable.
2. Read journal/JOURNAL.md History and Backlog. If the index is missing, run python -m skore_skills scaffold --journal. Do not write or paste the file. The CLI writes sections in order: Status, Data understanding, Modeling decisions, History, Ideas, and Backlog. If that command cannot run this turn, name it and stop. After the file exists, edit the existing History and Backlog tables (columns: Stem, Intent, Status, Headline result, Report, Design note; and #, Item, Source). Ideas columns are Question, Status, Experiment, Source; triage owns that table. A planned History row uses n/a in Report. Stable B<N> indices. Do not renumber on removal. Record-outcome does not edit Ideas.
3. If recording a run: copy the headline metric from the audit digest or the user's value. Do not invent numbers. With no digest and no user headline, skip the headline in one line and leave the History status unchanged. Do not write done with headline n/a. Update the matching History row (planned → done only if smoke passed and a headline result exists). Headline metric remains the source for History and Last result; never substitute G-AUDIT-FINDING for performance. Copy the digest's persisted-report locator into the History Report cell and the design note's Persisted report Status line. Paste that string verbatim; do not paraphrase it as "normalized" or rewrite the Hub URL. If the digest has none, write n/a — backend did not expose a locator in both places; do not derive or guess a URL. Copy G-AUDIT-FINDING verbatim into the design note's Audit findings line. Audit skipped → n/a — audit not run; missing/errored digest → n/a — audit digest unavailable. The Status update copies Headline result, Persisted report, and Audit findings. Do not change State or Approved by user on. Then insert or replace ## Results in the design note, between ## Status and ## Notebooks. Summarize from the audit digest — its cell outputs carry repr(report), ## Checks summary, and ## Metrics summary as text. With no audit this turn, fall back to scratch/results/<stem>/report.txt, which evaluate writes. Write ### Report overview from the report text, then ### Checks when that section exists. ### Metrics is a heading only when ## Metrics summary exists: do not transcribe the metric table or its values. The site embeds scratch/results/<stem>/metrics.html under that heading. History and Last result still get the single headline metric. Evaluation-only (audit skipped): Report overview only — do not invent Checks or Metrics subsections. After Metrics, add one ### subsection per extra Display cell the audit appended, using a human title and a <!-- results-embed: <slug> --> comment with the accessor name as <slug> so site build can inject the viewer. Report overview, Checks, and each extra subsection are 2–4 sentences from that cell's output. Do not copy G-AUDIT-FINDING, do not parse *.html, and do not paste iframes (site build injects those). Do not add <!-- results-embed: audit -->. The audit notebook viewer is audit/<stem>.nb.html, which site build places under ## Notebooks. If no subsection has a source, skip the Results section.
4. Idea triage, separate from record-outcome. Read journal/ideas/*.md and the Ideas table. If ## Ideas is missing, insert that header-only table between History and Backlog. Do not rewrite the other sections. A file's Triage line is open, promoted, discarded, or aside. A missing line is open. An open file whose Question and Source already match one Backlog row is set to promoted without asking, its Ideas row is removed, and no duplicate row is appended. Match both fields so two user ideas do not collapse. An open file with no Ideas row gets one: Question as plain text, Status open, Experiment from the file, Source copied verbatim. Ask the other open files: promote / discard / set aside. Write that value on the Triage line and keep the file. Promote removes that Ideas row and appends a stable B<N> row (Item from Question, Source copied verbatim) and sets promoted. Discard sets discarded on the file and the Ideas row and adds no Backlog row. Set aside sets aside on the file and the Ideas row and adds no Backlog row. A promoted idea is absent from Ideas. promoted, discarded, and aside leave the default queue; a later pass does not ask about them. Do not create a design note here. Do not delete an idea file. Changing Triage does not remove or renumber a Backlog row. An empty folder is a one-line skip: there are no idea files to triage, and it does not fabricate B1. When no file is open and tagged files remain, say in one line how many are promoted, discarded, and set aside, and offer to revisit. Do not retag until the user picks a file. On revisit, the same three choices apply. Promoting then appends B<N> only when that Question and Source are not already a row, and removes the Ideas row. The only other follow-up is offering to shape an idea or search the literature when those skills are installed. Do not load either skill, and do not start a search or a shaping menu, until the user picks one. Missing skill → one-line skip; do not invent that skill's search or shaping steps. After the user picks, that skill writes the idea file and returns here; triage the new file in this same mode. When the user picks an existing B<N> to draft, return that row to model-ml-pipeline, which can create its design-note shell with python -m skore_skills scaffold --journal --stem <NN_short_name>. Do not draft that template in this backlog turn.

Stop conditions

• Do not design or implement the next experiment in this turn.
• Do not dispatch setup or audit by skill id. Returning a selected row or proposal to model-ml-pipeline is required. Do not ask Yes / No on that proposal.
• In model-entry selection mode, do not invent a Backlog row or remove it before the paired design note exists.
• Do not invent metrics.
• Do not derive, shorten, or merge G-AUDIT-FINDING with the headline metric. Copy each into its owned field.
• Do not change design-note State or Approved by user on. Record-outcome copies Headline result, Persisted report, and Audit findings.
• Do not parse scratch/results/<stem>/*.html when writing ## Results. Summarize Report overview and Checks from the digest, or from report.txt on the evaluation-only path. ### Metrics is a heading only.
• Do not paste a JOURNAL.md body or recreate the index from memory.
• Do not mark done while smoke is red.
• Do not delete an idea file. Triage writes its Triage line and moves or updates its Ideas row.
• Design approval is owned by model-ml-pipeline; this skill only returns a proposal or selected Backlog row.

End of turn

Direct turns reach here, including a user request to record a finished run. Dispatched record-outcome mode returns before this section.

If policy.site is true, export-ml-site is installed, run python -m skore_skills site build --if-stale. Do not run notebook convert. Skip in one line otherwise. If site build errors with mkdocs-material is required, load add-python-package for mkdocs-material (agent) and build once more. Do not pixi add / uv add. If that skill is missing, or the retry still fails, name the error in one line. Name a build error; do not fail the backlog turn.

Run python -m skore_skills git end-turn --stage backlog. 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. Do not run git commit in this skill.