v0.3.0 — the UX release: 14 skills → 8, one front door, ≤5-question setup

The UX release: 14 skills → 8, one front door, ≤5-question setup, plain language on every
user-facing surface. Engine changes are minimal — the CLI checks, scaffolder, deploy generators,
hooks, and templates carry over (one exception: the configure wizard slimmed down); this is a
surface-area consolidation, applying the design system shipped in Ticketwright v2.0.

Changed — the rename map (v0.2 → v0.3)

v0.2	v0.3
start-app + new-app + refine-requirements + add-page + onboard	start-app (the front door — owns spec → scaffold → build → ship; --spec for the requirements phase incl. backfill from existing source, --setup for machine + repo setup, adopt for existing repos)
review-app + apply-review + auto-review-app + sql-review	review-app (--fix applies findings as atomic commits, --auto loops review→fix to convergence, --sql writes the audit companions)
deep-dive-data	audit-lineage
—	feedback-app (new — upstreamed from production use: classify user feedback into BUG / POLISH / UX / NEW-FEATURE / CROSS-CUTTING, apply as atomic per-item commits, follow-up review)
preview-app, validate-app, ship-app, migrate-app	unchanged names, refreshed surfaces

All 8 retired names still work as deprecated alias stubs (commands/); they will be removed in
the next major release.

Added

• Adopt mode (skills/start-app/adopt.md) — /start-app adopt on a repo that already has
  Streamlit apps maps onto the observed layout instead of scaffolding over it, classifies custom
  commands/skills as shadows / extends / unrelated against the plugin's skills, and writes a
  MIGRATION.md checklist. An existing AGENTS.md is never overwritten (renders to
  AGENTS.streamsnow.md for manual merge).
• ≤5-question streamsnow configure — down from ~14 prompts. Asks only runtime, account
  locator, governed database, allowed schemas, and deploy source; everything else (project
  identity derived from the directory name, roles, warehouse, schema names, container objects,
  git-repository deploy fields) is written as an inline-commented default in
  streamsnow.config.yaml — the file is the editing surface, and re-running configure prefills
  from it so hand edits survive. Guarded by a new tests/test_wizard.py contract test.
• Plugin-surface contract test (tests/test_plugin_surface.py) — CI now asserts the 8-skill
  surface, the ≤80-line SKILL.md cap, the 8 alias stubs pointing at their replacements, no retired
  name referenced as live inside skills/, and every relative markdown link resolving.
• Spec backfill (/start-app --spec <slug> on an app with existing code) — reverse-engineers
  REQUIREMENTS.md from st.Page declarations, chart/KPI/filter calls, SQL header blocks, cache
  decorators, and snowflake.yml, marking anything uncertain (inferred) for §10 review.
  Upstreamed from proven production use.
• skills/_shared/runtime-decision.md — the container-vs-warehouse decision in one place,
  neutrally framed (both runtimes are legitimate; the repo default wins absent a concrete reason),
  with the detection rule, trade-off table, and the manifest/connection checklist. Skills now link
  to it instead of re-explaining the choice (previously restated in 5+ skills).
• docs/migrating-a-consumer-repo.md — map a repo's home-grown skills to the plugin surface
  (worked example: a 16-skill production repo), what stays local (branding parity, tracker
  integration, warehouse-specific rules), and the incremental adoption path.

Changed (language & structure)

• SKILL.md ≤80 lines, depth in reference files — every skill's front page is now a scannable
  contract (modes, steps, boundaries), with detail split into per-skill reference files
  (start-app/{spec,scaffold,pages,setup,adopt}.md,
  review-app/{dimensions,fixes,auto-loop,sql-companions}.md, audit-lineage/tracing.md,
  feedback-app/classification.md, validate-app/fixing-checks.md).
• Plain language on user surfaces — skill descriptions lead with the trigger use-case;
  "deterministic PASS/FAIL ship gate" reads "the pass/fail check that must be clean before
  shipping"; report summaries print critical / should-fix / nice-to-have; "legacy" dropped from
  warehouse-runtime framing. Contributor-facing terms (bucket mechanics, check internals) stay in
  reference files.
• Graceful degradation instead of hard failure — missing config: review runs static-only and
  says which findings go unverified; missing Snowflake connection: lineage/companion rows are
  marked unverified with the exact enabler named; missing Playwright MCP: walkthroughs skip
  silently. Skills name the enabler instead of refusing.
• §11 Build Progress simplified — Current phase plus an append-only Sessions log whose last
  line names the next command. The per-page status table is gone (page state is visible in the
  tree and git); existing specs keep working, the table just stops being maintained.
• hooks/session_start.sh discovery line, README, and docs updated to the 8-skill surface.

Carried from the previous Unreleased

• docs/distribution.md — how StreamSnow ships (PyPI CLI + Claude Code plugin)
  and the recorded decision not to add a separate cp -r copyable kit
  (streamsnow init is the config-driven "better cp -r").
• Thickened the plugin skills toward the source's depth — that content now lives in the v0.3
  reference files rather than monolithic SKILL.mds.

v0.2.0

Fixed

• The scaffolded branded_metric now HTML-escapes its label/value/delta before
  rendering with unsafe_allow_html=True, so a database-derived value cannot
  inject markup into the viewer's page (hardening applied to the template and the
  example). Dependency-name matching is PEP 503-normalized, so a manifest that
  spells a package with underscores/dots (snowflake_snowpark_python) is no
  longer reported as missing.
• validate-app now validates the contents of the sibling dependency
  manifest, not just its presence: container apps must declare a
  requires-python that admits the container runtime's Python (PEP 440
  specifier semantics, so >=3.10 is accepted and <3.11 / ==3.10.* are
  correctly rejected) plus streamlit + snowflake-snowpark-python; warehouse
  apps must declare those deps in environment.yml and must not pin python.
• check caching now flags two patterns it previously missed: a public loader
  that hands a named query through a local variable
  (sql = load_sql("x"); conn.query(sql)) and one that delegates a named
  query to a private fetch helper (including transitive helper chains). Only the
  SQL-bearing argument is inspected, so an unrelated string keyword (e.g.
  query_tag="adhoc") no longer trips the generic-executor guard.

Added

• Documentation guides: docs/getting-started.md (try the example with no
  Snowflake, then set up a governed repo), docs/data-discovery.md (find tables
  and wire governed queries), and docs/deploying.md (the end-to-end deploy
  story for both deploy sources). Linked from the README.
• Runnable example app at examples/sample-dashboard/ — a StreamSnow-shaped
  Streamlit dashboard wired to deterministic sample data, so it renders with
  streamlit run and no Snowflake connection. Mirrors the streamsnow init
  structure (st.navigation entrypoint, branding, @st.cache_data loaders).
• packaging runtime dependency (PEP 440 version-specifier parsing in
  validate-app).

StreamSnow v0.1.0

First public release. Build, govern, and ship Streamlit-in-Snowflake apps with Claude Code: a streamsnow CLI (configure/init/new/validate-app/preview/check/deploy) + a Claude Code plugin (14 skills + shared recipes). Container & warehouse runtimes; stage-copy & git-repository deploy sources. Install: uvx streamsnow or pip install streamsnow. See CHANGELOG.md.