| name | anything-to-spatial-app |
| description | Use when creating or materially updating a PICO Spatial Android/Kotlin app from Figma, screenshot/mockup, PRD, intent, hybrid sources, or a bounded panel patch, and the task requires choosing or preserving container, window model, panel hierarchy, and layout regions. NOT for SDK upgrades, empty-dir bootstrap, legacy Android porting, pure code review/refactor, old-baseline D2C A/B evaluation, or performance diagnosis. |
| license | Apache-2.0 |
Anything → PICO Spatial App
🔴 Contract touchpoints (progressive loading; read at the phase that needs them):
references/workflow-contract.md — Phase 1 / 4 / 5 / 7 artifact schemas, reflection fields, Backtrack table.references/architecture-conventions.md — Phase 6 before Kotlin/Compose code: layered packages + ViewModel/UseCase/Repository + unit-test floor.../spatial-ui-design-style/SKILL.md — Phase 6 before Compose UI, especially screenshot/visual codegen: PicoTheme / Material / hover / click+haptics / R1–R8 rules. This is a generation-time contract, not a final lint suggestion.
Boundary
| Situation | Use instead |
|---|
| Upgrade SDK / migrate deprecated APIs | spatial-sdk-update |
| First-time scaffold from empty dir | spatial-app-onboarding |
| Migrate 2D Android app to spatial | porting-android-app |
| Performance diagnosis | spatial-app-perf-diagnose |
| 3D bbox / placement planning | spatial-sdk-scene-builder |
Reference index (read on demand, never preemptively)
| Read when … | File |
|---|
| Phase 1 / 3 — input mode + evidence | references/input-normalization.md, references/evidence-extraction.md |
| Phase 3 / 5 — visual layout inference, region decomposition, repeated/state mapping | references/layout-inference.md |
| Phase 1.5 — adapter dispatch | adapters/_registry.json + the selected adapters/*-adapter.md |
| Phase 1.5 — validating/editing adapter contract or registry | ADAPTER_ROADMAP.md |
| Phase 4 — container + window model | references/container-decision.md, references/window-model-decision.md |
Phase 4 — spatial_features includes anchor / env_mesh | references/spatial-anchor.md (BLOCK if anchor in WindowContainer) |
| Phase 5 — layout schema | references/layout-schema.md |
| Phase 6 — entry chain + manifest | references/manifest-and-entry.md + (window-container.md OR stage.md) |
| Phase 6 — gradle setup errors | references/gradle-setup.md |
| Phase 6 — SpatialUI component whitelist | references/spatial-ui-components.md |
| Phase 6 / figma-adapter — Figma tokens + visual feature mapping | references/figma-mapping.md |
| Phase 6 / figma-adapter — SpatialUI import lookup | references/spatial-api-imports.md |
| Phase 4 / 6 — window-level components, Subwindow, floating layers | references/spatial-windows-guide.md |
| Phase 6 — architecture + tests | references/architecture-conventions.md |
7-phase flow
| # | Name | Artifact | Gate |
|---|
| 1 | Frame | Input Envelope | inputs explicit |
| 1.5 | Adapter Selection | selected adapter + hook plan; execution may wait for Phase 2 workspace facts | one adapter max per current input_mode |
| 2 | Read | inspection notes | (none) |
| 3 | Spec | Evidence Packet + Normalized Spec + Assumption Ledger | normalization complete |
| 4 | Decide | Container Decision + Window Model Decision | legality + singularity |
| 5 | Plan | Spatial Layout Contract (or Patch Contract) | contract complete |
| 6 | Build | code edits / scaffold | (verified in 7) |
| 7 | Verify | 11-step gate / adapter hooks (see Phase 7) | machine-driven + Gradle sync + runtime launch + agent-owned MCP hooks |
incremental_patch mode skips Phase 4 (inherit from existing module) and Phase 5 emits a Patch Contract.
Operating protocol
-
Sequential & gated. No skipping artifacts; no proceeding past a failed gate. On failure: fix the artifact, apply a conservative default and log it in Assumption Ledger, or ask the user only if the unresolved issue materially changes the architecture.
-
Output language. Match the user's natural language for prose; artifact JSON keys, command names, and Step Output labels stay in canonical English.
-
Persistence. Artifacts live under <target>/.scratch/ with canonical filenames (input_envelope.json, evidence_packet.json, normalized_spatial_spec.json, assumption_ledger.json, spatial_layout_contract.json or patch_contract.json). assumption_ledger.json is always present; use [] only when no assumptions exist.
-
Resume. If <target>/.scratch/ already contains valid artifacts when the skill starts, do NOT re-emit them — re-run the scripts (cheap) and resume at the first missing/failing artifact. Delete .scratch/ only on explicit "redo from scratch".
-
Step Output (Phases 1, 4, 5, 7 only).
Step Output
- Artifact: <name>
- Summary: <1-3 lines>
- Key fields: <bullets>
- Reflection: <citation per workflow-contract.md>
- Gate result: PASS | BLOCKED
- Next action: proceed | revise | conservative default | ask user
Phase 1 — Frame
Apply input-mode routing in order, first match wins:
Step A — incremental_patch?
IF user names a specific file/panel/module
AND scope ≤ 1–2 regions or components
AND root container + window model do NOT need to change
THEN input_mode = incremental_patch. STOP.
Step B — otherwise classify by strongest source:
Figma URL → visual_design
screenshot / mockup image → visual_reference
PRD / long-form spec → product_doc
one-line ask only → intent_only
more than one of the above → hybrid
| User says | Routing |
|---|
"add a search field to MainPanel.kt in myapp" | incremental_patch |
"use this Figma to redesign myapp's home page" | visual_design |
"use this Figma to add a close button to DetailPanel" | incremental_patch |
Decide generation_mode (existing_module / new_project) and emit Input Envelope per workflow-contract.md §1. Gate: input_mode, generation_mode, target / output, input_sources[] all explicit.
Phase 1.5 — Adapter Selection
Read adapters/_registry.json, select exactly one active adapter for the current input_mode, then read only that adapter. Zero / multiple active matches = BLOCKED. This phase selects the adapter and installs any hooks.verify / hooks.cleanup plan for Phase 7; it does not force evidence extraction before workspace facts exist. If the selected adapter requires target platform, root container, or existing window model, execute that adapter after Phase 2 inspection and feed its output into Phase 3. Adapters are limited to existing evidence_packet.json / normalized_spatial_spec.json / assumption_ledger.json schema fields; their seven-field contract lives in ADAPTER_ROADMAP.md.
Adapter failure_mode may explicitly reroute to another input_mode (for example Figma → screenshot fallback). In that case, revise input_envelope.json, return to Phase 1.5, and select exactly one adapter for the new mode. Do not chain a second adapter under the old input_mode, and do not silently fall back.
Phase 2 — Read (workspace inspection, only in existing_module / incremental_patch)
Inspect: module build.gradle.kts, AndroidManifest.xml, Main.kt, platform/SpatialApplication.kt, platform/LaunchActivity.kt, existing res/. Prefer editing existing files; reuse package, manifest wiring, resources. Preserve current root container unless user explicitly asks otherwise or the requested feature is impossible. No JSON artifact — carry notes into Phase 3.
Phase 3 — Spec
Emit, in order:
- Evidence Packet —
facts / unknowns / conflicts / confidence. Facts only.
- Normalized Spatial Spec —
request_context / product_intent / spatial_intent / window_intent / layout_intent / ambiguities / evidence_trace. On disagreement, pick one and explain the tie-break in evidence_trace (no parallel truths).
- Assumption Ledger — every architecture-impacting default with
assumption / impact / confidence.
Visual input guardrails (HARD)
For screenshots/mockups, keep the main skill path concise and classify evidence by responsibility before planning. Detailed schemas and examples live in references/workflow-contract.md §5.
For visual_reference work, read references/layout-inference.md before finalizing layout_intent, content_layout_metrics, or visual_content_contract; use it to decompose regions, catch repeated/stateful structures, and sanity-check overlay relationships.
- Split visual evidence into app-owned window, window ornaments, page content, temporary floating layer, and spatial environment context.
- Do not treat passthrough / skybox / floor / scenery / system safety lines as app content unless facts prove the app owns them.
- Edge-pinned long-lived rails / tabs / toolbars are
window_chrome_ornaments[] (TabBar / Toolbar / Subwindow), not page children. Not multi_window does not mean page content.
visual_reference must carry reference_frame, content_layout_metrics, and visual_content_contract; derive window size from app_owned_bbox_px, not the full screenshot.- Bind measured sizes and visual semantics to Phase-6 constants/components (
SideNavigation, SearchField, fixed grids, asset-backed image cards) instead of magic dp or placeholder UI.
Gate: enough evidence to choose generation_mode, propose one container_candidate, compare ≥1 window_model_candidate. No hidden assumptions.
Phase 4 — Decide (skip in incremental_patch)
Read references/container-decision.md and references/window-model-decision.md. If spatial_features includes anchor / env_mesh, also read references/spatial-anchor.md and resolve legality now.
4a. Container Decision
Required: container / container_reason / container_evidence[] / rejected_near.{alternative, rejection_reason} / rejected_far.{alternative, rejection_reason}.
4b. Window Model Decision
Choose one of: single_panel, single_panel_with_popup, sidebar_content, master_detail, window_plus_subwindow, multi_window. Same field requirements as 4a. Apply the Subwindow-vs-multi_window escalation rule from window-model-decision.md.
Reflection (HARD): rejected_near neighbouring, rejected_far distant; both rejection_reason MUST cite a concrete Evidence Packet.facts.<key>, a row of the legality table, or an escalation rule number. "not needed" / "not applicable" / "no evidence" = BLOCK. Apply legality inline — do NOT defer to Phase 7.
Phase 5 — Plan
The contract IS the layout tree (no separate "internal" step). Required fields per workflow-contract.md §5:
If the input is screenshot/mockup-driven, consult references/layout-inference.md to sanity-check (do not re-derive) regions[], repeated_structures[], states[], and window_chrome_ornaments[]; layout_intent stays frozen from Phase 3.
container / container_reason / window_model / window_reasonreference_frame (visual_reference only) — screenshot px, app-owned bbox, target window dp, scale policycontent_layout_metrics (visual_reference only) — panel padding, measured region rects, repeated item sizes/gaps. This is mandatory when generating screenshot-based page content.visual_content_contract (visual_reference only) — sidebar surface/search/chip semantics, tab visible count/style, card content/overlay/asset policy.window_chrome_ornaments[] (when present) — id, type (TabBar / Toolbar / Subwindow), placement, role; ornaments are siblings of the main page, not main-page children.windows[] — id, role, anchor, default_visibility, childrenregions[] — hierarchy / states / alignment / sizerepeated_structures[], states[]evidence_trace[] — ≥1 entry per primary window, each fact_ref citing a concrete Evidence Packet.facts.<key> or a Phase 4 decision field. "because the design says so" = BLOCK.
Persist to <target>/.scratch/spatial_layout_contract.json. For incremental_patch, emit Patch Contract instead.
Phase 6 — Build
Module mode rules
-
Existing module: keep namespace/package, manifest wiring, entry chain. Allowed: new Compose files, new drawables/strings, new state holders. NOT allowed without explicit escalation: switching root container, changing pico.spatial.windowcontainer.* meta, introducing Stage-only APIs (anchor / ECS / env_mesh) inside a WindowContainer.
-
New project: scaffold the project with pico-cli project create (it owns the bundled template, entry chain, and manifest), then refine the Stage variant if needed.
pico-cli project create \
--dir <output-dir> --name <name> --package <com.example.foo> \
--template <planar|volumetric|stage>
python3 -m scripts.inject_container \
--output <output-dir> \
--container <STAGE_MIXED|STAGE_PROGRESSIVE|STAGE_FULL>
Map the Phase-4 container to a CLI template: ON_PLAIN → planar, IN_VOLUME → volumetric, STAGE_MIXED|STAGE_PROGRESSIVE|STAGE_FULL → stage. pico-cli project create produces a complete, runnable base project — package layout, Main.kt entry chain, and a fully-populated AndroidManifest.xml with the container meta-data already in place. Do NOT re-scaffold, rewrite mainApp, or re-insert manifest meta; build your UI on top of the generated entry point. The only post-CLI step is for Stage: all three STAGE_* variants share --template stage, so run inject_container --container STAGE_* to set the pico.spatial.stage.style / immersion values (it edits the meta already present, and is a no-op for the WindowContainer cases).
Entry chain: Application.onCreate { launch(::mainApp) } → mainApp(scope: SpatialAppScope) → DefaultWindowContainer {} or DefaultStage {} → SpatialLaunchActivity.
Android Studio sync is mandatory for new modules. After creating or including
a new module (settings.gradle.kts changed), trigger Android Studio
Sync Project with Gradle Files before claiming the app can be run from the
IDE. If no IDE-sync API is available to the agent, run the Gradle project
discovery proxy in Phase 7 and explicitly tell the user that Android Studio sync
is still required before the first IDE run/configuration selection.
UI rules (layered + minimal)
- Generate UI from the Phase-5 contract only: root container → window ornaments → windows → regions → reusable components.
- Read
../spatial-ui-design-style/SKILL.md before Compose UI; generated code must pass design-style admission without skip/degraded mode.
- Prefer SpatialUI built-ins and documented imports; never invent SDK names.
PicoTheme {} wraps windowed UI; windowConstraints(...) is resize bounds, not first-open size.
- Keep business UI 2D unless Phase 5 justifies 3D / Stage behavior; Stage-only APIs stay out of WindowContainer flows.
- Implement
window_chrome_ornaments[] with window-level fittings, and implement reference_frame / content_layout_metrics / visual_content_contract through named constants, state, and components.
- Custom interactive components must follow spatial-ui-design-style indication + haptics rules, or use a built-in component that provides them.
Architecture rules (HARD)
Read references/architecture-conventions.md before code. The checker enforces layered packages, thin Main.kt, MVI-lite state, repository boundaries, mandatory ViewModel tests, and UseCase + UseCase tests when the screen has non-trivial business rules, filtering, sorting, selection, or data transformation.
Phase 7 — Verify (machine-driven)
Run bash scripts/validate_workflow_and_build.sh <target>. The canonical 11-step order, skip semantics, JSON outputs, and Figma hook ordering are in references/workflow-contract.md; smoke-build diagnosis is in references/gradle-setup.md. spatial-ui-design-style admission is non-optional for generated Compose UI: missing verifier, missing source root, verifier failure, or --skip-design-style must fail the run.
Do NOT paraphrase JSON results — read passed / summary.errors / failures_or_explicit_none literally. verification_summary.json.clean: false means a degraded run; it exits non-zero unless --allow-degraded was explicit. Disclose every warnings[] / skips[] entry and never imply skipped gates passed.
Backtrack: after 2 consecutive failures at the same check, return to the originating phase per the Backtrack table in references/workflow-contract.md. Edit the offending artifact first; do not silently rewrite code that contradicts an unchanged contract.
Structural review (LLM-owned, semantic): repeated structures preserved as templates; selected/disabled/highlighted states represented in state holders; no UI added beyond input; in existing_module mode, reuse module resources before adding new ones.
Exit checklist
Run is complete only when ALL hold:
validate_workflow_and_build.sh <target> exits 0 and <target>/.scratch/verification_summary.json.clean == true. Only Gradle sync / runtime launch may be environment-degraded, and degraded runs are not complete unless explicitly accepted in handoff.- For new modules, Android Studio Sync Project with Gradle Files has been triggered, or the final handoff explicitly states that the user must trigger it before first IDE run because no IDE sync API was available.
legality_check_result.json, implementation_scan_result.json, gradle_sync_result.json, architecture_check_result.json, unit_tests_result.json all → "passed": true.design_style_result.json.passed == true and design-style verifier → 0 errors; no --skip-design-style / degraded bypass.- If the selected adapter declares hooks,
adapter_hooks_result.json.passed == true and verify / cleanup hooks ran in the registry order.
- None of the hard-fail conditions in
workflow-contract.md triggered.
Final handoff:
- Container: <chosen + why>
- Window model: <chosen + why>
- Mode: <existing module update | new scaffold | incremental_patch>
- Path: <module path | output path>
- Android Studio sync: <done | user must run Sync Project with Gradle Files>
- Assumptions: <explicit list or 'none'>
- Remaining inferred/mock parts: <list>
- Workflow artifacts: Input Envelope / Evidence Packet / Normalized Spec / Assumption Ledger / Spatial Layout Contract (or Patch Contract)
Pitfalls (do not)
- skip workspace inspection when the user names a module (default to
existing_module)
- restate Phase 3 in Phase 5 (the contract IS the layout tree)
- defer container × feature legality to Phase 7 (decide inline in Phase 4)
- confuse overlay with window — popup menus stay in one panel
- escalate to
multi_window without independent launcher / lifecycle / placement memory evidence
- invent SDK names; only the whitelist
- over-spatialize a 2D settings UI (most belong in
ON_PLAIN)
- implement screenshot spatial background, passthrough, floor/trees/skybox, or system safety lines as app content
- fill Step Output with mechanical PASS — Reflection must cite a fact-key or legality-table row
- run the full 7-phase flow on a small patch; use
incremental_patch mode
- emit code before reading
architecture-conventions.md and spatial-ui-design-style/SKILL.md
Honesty: A 2D reference under-specifies a spatial app. State explicitly when passthrough / skybox / depth / hover / haptics / gestures were inferred; flag that anchors and Full Space behaviors require a device.
Clarification: ask the user only when one of these is truly unresolved — no visual reference is available, target module / output cannot be inferred, multiple window interpretations are equally plausible and materially change the app structure, or package / namespace conflict cannot be resolved safely. Otherwise proceed with the safest default and state the assumption.
