Gamepad-Friendly Input Patterns: Research for D3-TUI

Status: Research complete — ready for spec integration
Agent: codex (UX research scope)
Date: 2026-05-11
Device target: Anbernic RG40XXV (D-pad, A/B/X/Y, L1/R1, Start/Select, Menu)
Resolution: 640×480, integer-scaled 320×240 fallback for small text
Constraint: All patterns must work WITHOUT text input. Text input is fallback only.
Domain context: Turn-based strategy, hex grid, node-traversing agents. Players configure agents, review LLM output, correct mistakes, route context between agents.

Pattern 1: Radial Menus

What It Is

A circular menu centered on the cursor or screen center, where thumbstick/D-pad direction selects an option. Up to 8 slices (4 cardinal + 4 diagonal) without modifiers, up to 32 with L1/R1 tab rings. Selection is instant — push direction, press A to confirm. Advanced variants layer sub-menus (push deeper into a direction to reveal nested choices).

Game Examples

RG40XXV Button Mapping

• D-pad directions → highlight a slice (snap, not free-rotate — no analog stick on RG40XXV)
• A → confirm selection
• B → cancel / back out of radial
• L1 / R1 → flip between ring tabs (e.g., L1 = "Actions", R1 = "Targets")
• Y → hold for contextual tooltip on current slice (optional)
• X → "more options" — expand selected direction into a sub-radial

Without an analog stick, the RG40XXV is limited to 8 directions per ring. This is functionally identical to Secret of Mana's ring system on SNES — a proven, comfortable input model.

Good For

• Configuring agents — "Provision agent" → radial direction selects agent role (Navigator / Sentinel / Explorer / Builder). Sub-radial for detail (Navigator → Pathfinder / Waypoint / Scout).
• Reviewing output — "What do you want to review?" → radial of recent LLM outputs, charts, agent logs.
• Routing context — "Route to..." → radial of available agents. L1/R1 for agent triads (LAND / SEA / AIR / COMMAND).
• Quick actions in combat — selecting move, attack, ability, item in a battle phase. Standard JRPG radial pattern.
• Any choice from 4-8 options where options are known and stable. Radials degrade fast beyond 8.

Bad For

• Free-form text or names — No radial can replace naming an agent or writing a custom prompt.
• Choosing from >32 options — Even with L1/R1 tabs, beyond 4 rings × 8 slices = 32, cognitive load spikes. Use chips or drill-down instead.
• Options that change frequently — Radial muscle memory works when directions are static. If "Attack" was "Up" last turn but "Down" this turn, players will mis-input constantly.
• Precision timing — D-pad radials require a brief pause to confirm direction before pressing A. Not for real-time twitch input.
• Long option labels — Each slice can show at most ~12 characters legibly at 640×480. If your option names exceed that, radials are the wrong container.

Pixel Budget at 640×480

Radial overlay: 320×320px centered (radius ~140px)
8 outer labels: ~120px each, ~10px text height
Selected slice highlight: color-shift + scale-up on slice zone
Tab indicators (L1/R1): 24×24px in upper corners
Sub-radial: same dimensions, offset from parent to show hierarchy
Total screen real estate: ~70% of center (masking background with 50% opacity)
Crisp 1x pixel font usable; no sub-pixel scaling needed.

RG40XXV-Specific Notes

• No analog stick means no "continuous rotation" — stick to 8 discrete slices.
• The RG40XXV D-pad is a cross (not a disc), so diagonal detection is slightly less reliable than cardinal. For critical actions, place them on cardinal directions (Up/Down/Left/Right), reserve diagonals for secondary actions.
• Secret of Mana solved this elegantly: the ring has 8 positions but the game snaps your input to the closest valid slice with a grace zone. This reduces mis-input from imperfect diagonals.
• A 1px-thick ring renders cleanly at integer scale. Use high-contrast colors — the RG40XXV screen has limited dynamic range compared to modern OLED.

Pattern 2: Mad-Libs / Chips Composition

What It Is

The player builds a "sentence" or command by dropping pre-made tokens ("chips") into slots. Each chip is a single word or short phrase from a constrained vocabulary. The sentence reads left-to-right as <prefix> <subject> <verb> <suffix>. This is exactly how the Oracle Chamber already works on this device — proven in production.

The key insight: chips encode grammar constraints. A chip like "The team" knows it can only be followed by a verb chip ("should scout", "must avoid", "will report"). Invalid next-chips are greyed out or hidden. The player navigates a semantic tree one chip at a time, never typing a letter.

Game Examples

RG40XXV Button Mapping

• D-pad Left/Right → move between slot positions (cursor across the sentence template)
• D-pad Up/Down → cycle through available chips for the current slot
• A → confirm current chip and advance to next slot
• B → clear current slot / go back one slot
• L1 / R1 → jump to first/last slot (for long sentences)
• Y → preview result (what the composed sentence will do / ask)
• X → insert a "wildcard" slot (free-text fallback, only when necessary)

Good For

• Agent configuration — "Scout team" + "should explore" + "northern ridge" + "aggressively". Four chips, zero typing.
• LLM prompt assembly — composing structured prompts from validated fragments. Never an invalid prompt because grammar constraints block nonsense combos.
• Corrections / feedback — "That output" + "is" + "too vague" + "retry with more detail". Structured correction chips mapped to common LLM failure modes.
• Routing context — "Send" + "Oracle output" + "to" + "Navigator agent" + "with" + "priority: high". Route-tag chips.
• Any structured query with a known vocabulary — the chip system forces precision without requiring the player to know exact syntax.
• Encoded inventory selection — "Equip" + "Steel Sword" + "in" + "main hand" + "from" + "backpack".

Bad For

• Truly free-form exploration — "Ask the LLM anything" breaks chip composition. If the vocabulary is unbounded, chips become a bottleneck.
• Rare/specific combinations — If a chip combo appears once every 20 hours, the player won't build muscle memory and will hunt through chip lists. This is the opposite of what chips do well.
• Very long sentences (>10 slots) — Cognitive load scales linearly with slot count. Beyond 6-8 slots, the player loses track of what the sentence means.
• Numeric input — "Set reinforcement count to 47" is annoying with chips. Use a number picker wheel or T9 fallback.
• Emotion/tone — Hard to encode "aggressive but polite" or "cautiously optimistic" as discrete chips. These work better as tags (Pattern 5).

Pixel Budget at 640×480

Sentence ribbon: full width (640px), 64-80px tall at bottom of screen
Slot boxes: ~120px × 40px each, 4-5 visible slots side-by-side
Active slot: highlighted border + slight scale-up (44px height)
Chip list popup: 200px × 180px anchored above/below active slot
  - Each chip row: 180px × 36px, ~6 visible without scrolling
  - Scroll indicator if >6 chips available
  - Sub-chip hints: tiny icons or 1-2 char indicators in chip row
Font: 14px in chip boxes, 12px in chip list
Total screen real estate: ~15% at bottom (ribbon + popup)
Background remains usable for game view above the ribbon.

RG40XXV-Specific Notes

• This pattern is already proven on this device via Oracle Chamber. The chip UI renders cleanly at 640×480 with single-pixel borders.
• Vertical chip menus (Up/Down to scroll) feel more natural on a handheld than horizontal menus because the thumb rests in a vertical sweep position on the D-pad.
• Chip text must be legible at 12-14px. At 640×480 integer-scaled on a 4-inch screen, this is borderline. Test with high-contrast white-on-dark, bold weight.
• The RG40XXV lacks a touch screen, so no drag-and-drop. All chip operations are D-pad + A/B. This is fine — D-pad chip cycling is faster than drag-and-drop for power users once they know the vocabulary.

Pattern 3: Multiple Choice Branching Interviews (Dialogue Trees)

What It Is

Instead of the player expressing intent directly, the system asks a series of narrow multiple-choice questions. Each answer narrows the possibility space. After 3-5 questions, the system has enough information to act. This is functionally identical to RPG dialogue trees — the player doesn't type "I want to romance Liara" — they choose "I care about you" → "More than a friend" → "Stay together" across three dialogue prompts.

The branching structure can be flat (all paths same depth) or deep (some branches go 2 levels, others 5). Flat is easier to navigate; deep is more expressive.

Game Examples

RG40XXV Button Mapping

• D-pad Up/Down → highlight a choice
• A → confirm selection, advance to next prompt
• B → go back one level (undo last answer, re-present choices)
• L1 → "I don't know / skip" — advance without narrowing (good for optional branches)
• R1 → "Tell me more" — open an explanatory sub-branch before committing (Mass Effect's "Investigate")
• Y → "Undo all / start over" — reset the interview
• X → "Show me where this is going" — preview how many more questions remain (breadcrumb trail)

Good For

• Correcting LLM output — "What's wrong with this output?" → Too long? Too short? Wrong facts? Wrong tone? → Drill down from the answer: "Too long" → "Cut in half" or "Summarize to 3 bullet points" or "Remove the part about logistics."
• Complex agent configuration — Instead of 40 parameters in a flat menu, ask: "What kind of agent?" (Scout/Builder/Healer) → "What priority?" (Stealthy / Aggressive / Defensive) → "Any special equipment?" (Yes/No) → "Which gear?" (shortlist based on agent type).
• First-run setup / onboarding — Wizard-style flow: 5-8 questions that configure a new game or agent. Only options relevant to earlier answers are presented.
• Diagnosis / debugging — "What went wrong?" → Agent stuck? Wrong output? Crashed? Timeout? → Branch into specific fix suggestions per symptom.
• Review / triage — "What do you want to review?" → Recent outputs? Agent status? Map state? Combat log? → Narrows to specific report, then offers actions (rerun, share, archive, fix).

Bad For

• Speed after mastery — Expert users who know exactly what they want ("Scout team, aggressive, northern ridge, now") will find 5 questions slower than direct input. Solution: offer a "quick command" bypass for power users (chips or presets).
• Open-ended exploration — If the player doesn't know what they're looking for, sequential narrowing frustrates. "Just show me what happened" doesn't map to a 3-option multiple choice.
• Very large branch depth (>7 levels) — Players lose sense of where they are. Breadcrumb trails help but don't fully solve this. At >7 levels, consider breaking into multiple shorter interviews.
• Cross-referencing — "I want something like the last thing, but with this other agent." Cross-reference requires seeing two things at once, which sequential interviews hide.

Pixel Budget at 640×480

Interview prompt area: 640×120px at bottom of screen (question text)
Choice list: 3-5 options stacked vertically, 560×180px overlay on lower-center
  - Each choice row: 540×44px with icon (32×32) + text
  - Active choice: highlighted background bar
Breadcrumb trail: 640×24px bar at very top of screen ("Config > Agent > Scout > Priority")
Progress indicator: mini dots or segments, 48×12px in corner
"Investigate" / "Back" buttons: anchored to bottom corners, 80×40px each
Total screen real estate: ~35% (lower third for choices + top bar for breadcrumbs)
Background: game view remains visible in upper ~50% — partially dimmed for focus.

RG40XXV-Specific Notes

• Stack choices vertically, not horizontally. Vertical D-pad scanning is faster (no eye tracking needed) and the 640×480 aspect ratio favors vertical stacking (you have 480px of height, only 640 of width — vertical lists feel natural).
• Max 5 visible choices per screen. Beyond 5, add a scroll indicator. The 6th+ option being off-screen serves as a natural "this is less important" signal.
• Breadcrumb trails solve the "what if I made a mistake 3 questions back?" problem without requiring a full undo chain. Player can tap B repeatedly to walk back.
• The RG40XXV screen's color reproduction is warm-tinted. Use blue-toned highlights for selection (cyan, light blue) rather than yellow/green — yellow bleaches out on this panel.

Pattern 4: T9-Style / Predictive Text

What It Is

When text IS needed (naming an agent, writing a custom prompt, searching for a specific item), use a grid-based letter picker. The classic mobile phone T9 layout maps 26 letters + space + backspace onto a 3×4 grid (A-C on 1, D-F on 2, etc.). The player navigates the grid with D-pad, taps A to enter the highlighted letter. Predictive completion suggests words as you type, reducing keystrokes by 60-80%.

Modern variants (Steam Deck keyboard, PS5 text entry) add dual-stick input and motion control, but the RG40XXV has neither. The T9 grid is the right constraint fit.

Game Examples

RG40XXV Button Mapping

• D-pad → Move cursor on the letter grid (snap to cells, wrap-around on edges)
• A → Enter highlighted letter
• B → Backspace (delete last character)
• X → Toggle between letter sets (lowercase / uppercase / numbers / symbols)
• Y → Accept top predictive suggestion (instant word completion)
• L1 → Accept suggestion #2
• R1 → Accept suggestion #3
• Start → Confirm and submit full text
• Select → Clear entire field

Grid layout options:

OPTION A: Standard T9 (3×4 grid, 12 cells)
  1: ABC    2: DEF    3: GHI
  4: JKL    5: MNO    6: PQRS
  7: TUV    8: WXYZ   9: .,!?
  DEL: <--   0: space    OK: done

OPTION B: Alphanumeric grid (4×7, 28 cells) — faster but smaller targets
  Full alphabet in a QWERTY-like 4-row grid.
  Works at 640×480 if cells are tight (40px each, 280px wide).
  Predictive bar replaces the need for many keys.

Good For

• Naming agents — "What should we call this scout?" → T9 grid + "Suggest names" button (themes: Greek gods, military callsigns, terrain features). Player can type "Artemis" or accept suggestion.
• Search with fallback — Browsing by category is primary (Pattern 3), but if the player knows the exact item name, T9 search is the emergency exit.
• Short custom prompts — A single sentence override: "Ignore the ruins" or "Retreat if outnumbered 3:1". 20-40 characters max. If the player needs to write paragraphs, the device is wrong; route to the Mac agent via Mozart.
• Passwords / codes — Numeric input for save codes, debug commands, or unlocking secret content.
• Comments on reports — "Good analysis but check sector 7" — free-text annotation on agent output.

Bad For

• Primary input mode — If T9 typing is the main way players interact with your game on a handheld, the UX has failed. This is explicitly a fallback.
• Long-form writing (>60 chars) — Each keystroke takes 0.5-1.5 seconds on D-pad. A 100-character message is ~1-3 minutes of tedious input. Route to the Mozart bridge for anything longer than a sentence.
• Real-time chat — T9 input is far too slow for back-and-forth conversation. Use preset responses (Pattern 5) or route to external keyboard.
• Non-English text — T9 was designed for English. Japanese, Arabic, Chinese require entirely different input methods. If your game supports non-English agents, invest in language-specific input or use chips exclusively.
• Frustration on first use — Players will try T9 once, find it slow, and never use it again unless forced. Hide it behind an explicit "Custom text..." option, not as a first-class pathway.

Pixel Budget at 640×480

T9 grid overlay (Option A): 240×160px centered (3 columns × 4 rows, cells ~72×36px)
  - Numbers: top label in each cell (9px, bold)
  - Letters: smaller text below (7px, regular)
  - Active cell: bright border + background fill
Predictive bar: 540×40px above the grid (3 suggestions, ~140px each)
  - Suggestion #1 (top): highlighted, 14px bold
  - Suggestions #2-3: 12px regular
Current text display: 540×32px above suggestions (shows typed text so far)
Toggle label (X): 64×20px in lower-right ("abc"/"123"/"ABC")
Grid + predict + text = ~250px tall, centered on screen
Total screen real estate: ~50% (full modal overlay with dimmed background)
Integer scale note: grid lines should be 2px instead of 1px to avoid shimmer

RG40XXV-Specific Notes

• The RG40XXV D-pad has a slightly stiff membrane compared to modern controllers. Rapid directional input (typing "hello" in <5 seconds) produces more fatigue than a DualSense or Switch Pro. Minimize the frequency of T9 use.
• Diagonal D-pad wear — the T9 grid is orthogonal (no diagonals needed). This is good — the RG40XXV diagonal detection is less reliable.
• Consider a "hold direction + auto-repeat" feature: hold Right on D-pad for 500ms and the cursor starts auto-advancing across the grid at 8Hz. This speeds up text entry for power users without requiring faster thumb movement.
• Use the Y button for the top prediction. Y is the "fast path" button — it's the easiest face button to reach from the D-pad resting position. Predictive acceptance should be single-tap, not hold-and-confirm.

Pattern 5: Tag / Preset Selection

What It Is

The player applies pre-written tags or presets from a curated list. Unlike chips (which compose into sentences), tags are standalone modifiers or commands. Think: - Dark Souls messages: "Try finger, but hole" — the player picks words from a fixed list, not a keyboard. The vocabulary IS the game's expression system. - Apex Legends ping system: "Enemy spotted", "Looting here", "Defending this position" — context-aware presets mapped to a single button. - Mario Maker stamp system: pre-drawn stamps you place on levels rather than freehand drawing.

Tags are curated, context-aware, and designed to cover 80%+ of what players want to say. The remaining 20% uses T9 fallback.

Game Examples

RG40XXV Button Mapping

• Y → Open tag palette (dedicated hotkey — always accessible)
• D-pad → highlight a tag
• A → apply highlighted tag
• L1 / R1 → switch tag categories (e.g., "Corrections" → "Routing" → "Reactions" → "Custom")
• B → close palette without applying
• X → toggle "show tag descriptions" (help text for each tag)
• Hold A → apply tag and keep palette open (for applying multiple tags sequentially)

Tag category suggestions for D3-TUI:

CORRECTIONS:
  [Too vague] [Too long] [Too short] [Wrong facts] [Bad tone]
  [Retry] [Summarize] [Expand section ___] [Redo step ___]
  [Skip this] [Send to expert]

REACTIONS:
  [Good work] [Interesting] [I don't understand] [Proceed]
  [Hold / pause] [Abort] [Run again]

ROUTING:
  [Send to Navigator] [Send to Oracle] [Send to Builder]
  [Broadcast to all] [Route to Mozart] [Archive in SG0]

CONTEXT TAGS:
  [Priority: High] [Priority: Low] [Urgent] [Background task]
  [Stealth mode] [Aggressive] [Defensive] [Scout only]
  [Report back] [Silent run] [Test mode]

CUSTOM:
  [Player-defined tag 1] ... [Player-defined tag 8]
  (Filled from user's most-used T9 inputs over time)

Good For

• Correcting agent output — Applying a "Too vague" tag to the last LLM output auto-injects the correction without typing. The game knows what "too vague" means and adjusts the prompt.
• Quick reactions during combat — "Abort!" or "All units retreat!" mapped to a single button press sequence. Faster than chips or interviews.
• Routing decisions — "Send this to Navigator" is a single tag. The system knows what "this" is (the currently visible output) and who Navigator is.
• Adjusting agent behavior on the fly — "Scout" + "Stealth mode" → two tags that reconfigure an agent mid-mission. No menus, no typing.
• Emoting / social — Quick reactions to agent output or multiplayer communication.

Bad For

• Specific, nuanced feedback — "Too vague" is broad. If the player wants "The third paragraph is vague about enemy troop count, be more specific there," a tag can't capture this. Route to interview (Pattern 3) or T9 (Pattern 4).
• Novelty — Tag lists exhaust quickly. If every correction is "Too vague," the system feels shallow. Mitigate with dynamic tags based on recent outputs ("This chart is missing data" → tag appears after a chart output).
• Tag overload (>40 visible tags) — Beyond ~40 tags, scrolling and hunting defeats the speed advantage. Break into <10 tags per category, <6 categories max.
• Cross-tag conflicts — "Stealth" + "Aggressive" might be contradictory. The system should flag conflicts or auto-resolve with a clarifying prompt.

Pixel Budget at 640×480

Tag palette: 480×320px centered overlay (semi-transparent background)
Category tabs: 640×40px across top (6-8 categories, ~60px each)
Tag grid: 3 columns × 3 rows of tag chips, ~140×44px each
  - Tag text: 14px bold inside chip
  - Active chip: bright border + subtle fill
  - Applied chips: checkmark indicator in corner
Tag description bar: 540×28px below grid (shows help text for highlighted tag)
Context indicator: 200×24px in upper-left ("Target: Last LLM output")
Apply button: 120×40px in lower-right ("Apply" or "A: Apply")
Total screen real estate: ~55% (modal overlay)
Tags render cleanly as rounded rects with 2px borders — no sub-pixel artifacts.

RG40XXV-Specific Notes

• Tags should be large enough to read at arm's length. The RG40XXV is held ~30-40cm from the face. 14px bold text is the minimum readable size; 16-18px is comfortable.
• Color-code tag categories. The RG40XXV screen distinguishes ~6-8 colors reliably. Use color as a secondary channel (with icon/text as primary — don't rely on color alone for RG colorblind players using the device).
• The Y button as "tag hotkey" is deliberate. Y is easily reachable from resting D-pad position without moving the thumb. This makes tag application a flow-state operation: D-pad to tag, Y to open, D-pad to select, A to apply, B to close — thumb never leaves the right half of the controller.

Pattern 6: Thumbs Up/Down + Refinement (Binary → Drill-Down)

What It Is

The player gives a thumbs-up (good) or thumbs-down (bad) on the most recent agent output, then optionally drills down to explain why. This is: 1. One-button feedback — instant, zero cognitive load 2. Optional detail — if the player wants to be more specific, the system asks a single follow-up question (~3 choices) 3. Learning over time — the system notices that thumbs-down on economic reports is always followed by "too optimistic" and auto-suggests that as the default

This pattern comes from Netflix ("thumbs up/down" on content → genre preferences), YouTube ("Not interested" → "Tell us why"), and Tinder (swipe → "it's a match" → optional chat).

Game Examples

RG40XXV Button Mapping

• A → Thumbs up (good output / approve / proceed)
• B → Thumbs down (bad output / reject / redo)
• X → "Skip review / neutral" (acknowledge without rating)
• Y → "Why?" — opens a drill-down radial or interview (Pattern 1 or 3)
• L1 → "Thumbs up and archive" (good + save context)
• R1 → "Thumbs down and retry" (bad + auto-rerun with different params)
• D-pad → if drill-down is open, navigate the follow-up choices

Drill-down flow example:

1. Agent output appears on screen.
2. Player presses B (thumbs down).
3. System: "What needs fixing?" → 3 options appear:
   - [Too vague / need specifics]
   - [Wrong information]
   - [Format / presentation issue]
4. Player selects "Too vague / need specifics"
5. System auto-adjusts the agent's prompt to be more specific and re-runs.
   Total interaction: B → Down → A. Three inputs, ~2 seconds.

Progressive learning:

Session 1: Player thumbs-downs economic reports 8 times, always drills to "too optimistic"
Session 2: Player thumbs-downs economic report → system auto-offers "Too optimistic?" as pre-selected drill-down
Session 3: System detects pattern and adjusts default prompt for economic reports to be "conservative estimates"
Player never had to configure this explicitly — it emerged from binary feedback.

Good For

• Reviewing agent output at scale — 12 agent reports come in. Quick-scan each: thumbs-up (keep), thumbs-down (reject/redo). Review 12 outputs in 15 seconds.
• Quality gate before dispatch — Agent wants to send output to another agent. Player reviews: thumbs-up (send), thumbs-down (edit first). Single-button gatekeeping.
• Training the system — Binary feedback over time teaches the LLM routing layer what "good" output looks like for THIS player. More valuable than explicit config because it captures revealed preference, not stated preference.
• Fatigue-resistant review — Reviewing 50 agent outputs exhausts with chips or interviews. Thumbs-up/down at scale is sustainable. Think of it as Tinder for agent outputs.
• Mid-combat quick decision — "Unit report: flanking maneuver ready" → thumbs-up (execute) or thumbs-down (hold). No time for menus. Single button press.

Bad For

• Nuance in first-time feedback — Binary doesn't capture "this output is mostly good but the third recommendation is wrong." Mitigated by the drill-down, but the first click loses information.
• False positives in training — If the player thumbs-ups everything without reading (fatigue), the system learns bad preferences. Mitigate with occasional "verify you're paying attention" checks (random thumbs-down on known-good outputs that are intentionally flawed).
• Polarizing complex outputs — Some outputs are genuinely "mixed" — good parts and bad parts interleaved. Binary forces a false dichotomy. Mitigated by the X (neutral/skip) button and drill-down that supports "highlight and comment on specific sections" (future enhancement).
• Low sample size for rare outputs — If the player only sees a "cavalry charge analysis" once every 10 hours, the system never learns their preference for cavalry reports. Wide distribution = good for binary feedback. Narrow distribution = chips or tags are better.

Pixel Budget at 640×480

Feedback bar: 640×64px anchored at bottom of screen (always visible during review)
  Thumbs-up icon: 48×48px, left side, green tint
  Thumbs-down icon: 48×48px, center, red tint
  Skip icon: 32×32px, right side, grey tint
  Button labels: "A: Good", "B: Bad", "X: Skip" below icons (10px text)
Drill-down popup: 300×180px centered (appears after B press if not yet drilled)
  "What needs fixing?" header + 3-4 options
  Option rows: 260×36px with radio-button indicator
Learning indicator: small dot + counter in upper-right ("+12 signals today")
Total screen real estate: ~12% (feedback bar only, persistent)
  Drill-down popup: ~30% (temporary overlay, auto-dismisses after selection)
Extremely lightweight — designed to be ambient, not interruptive.

RG40XXV-Specific Notes

• The green/red tint for thumbs up/down MUST have an icon shape difference too (thumbs-up outline vs thumbs-down outline). Don't rely on color alone — the RG40XXV screen's red reproduction is muddy compared to green.
• Place the feedback bar at the BOTTOM of the screen, closest to the thumbs. The RG40XXV is held like a Game Boy — the bottom third of the screen is in the player's natural visual field during review.
• Consider a subtle haptic pulse (if the RG40XXV supports rumble via HDMI-CEC or internal motor — check hardware spec) on thumbs-up to make the "good job" feel satisfying.
• The "learning counter" (+12 signals today) gamifies the review process. This is effective on handhelds where play sessions are shorter and more frequent than on PC.

Pattern Comparison Matrix

Recommended Composite Pattern for D3-TUI

Based on this research and the existing Oracle Chamber implementation, the recommended composite pattern is:

PLAYER WORKFLOW:

1. REVIEW PHASE
   → Thumbs up/down (Pattern 6) for quick-scanning agent output
   → Drill-down via chips (Pattern 2) or tags (Pattern 5) when detail needed

2. CONFIGURATION PHASE  
   → Branching interview (Pattern 3) for new/unfamiliar setups
   → Tags (Pattern 5) for on-the-fly configuration tweaks
   → Radial menu (Pattern 1) for in-combat quick actions

3. CORRECTION PHASE
   → Tags (Pattern 5) as first-class correction verbs
   → Chips (Pattern 2) for structured "redo with these constraints"
   → Branching interview (Pattern 3) for complex/novel corrections

4. ROUTING PHASE
   → Radial menu (Pattern 1) for agent selection (8 agents or fewer)
   → Chips (Pattern 2) for prepending routing metadata

5. TEXT FALLBACK (escape hatch only)
   → T9 predictive (Pattern 4) — hidden behind "Custom..." in every pattern
   → Mozart bridge routing for long-form input (off-device)

PRIMARY PATTERNS (always visible): Tags + Thumbs feedback bar
SECONDARY PATTERNS (one button away): Chips + Radial
TERTIARY (hidden): Branching interview (setup wizard only)
ESCAPE HATCH: T9 text (behind "Custom..." in all patterns)

Button Allocation Budget (7 physical buttons + D-pad)

Design Principles from Research

1. D-pad is a discrete selector, not a pointer. Don't design free-cursor UIs for a D-pad. Everything should snap to grid points, list rows, or radial slices.

2. "See it, move to it, press A" is the atomic interaction. Fitts' Law applies differently to D-pad: the cost is number of button presses, not pixel distance. A 3×3 grid is 1-4 presses to any cell. A 10-item list is 1-5 presses to any item. Optimize for minimum D-pad presses to reach the 80th percentile option.

3. The right thumb does double duty. D-pad for navigation, then A/B/X/Y for action. The thumb physically moves between them. Patterns that require rapid alternation between D-pad and face buttons (like fighting game combos) cause fatigue on the RG40XXV's stiff D-pad membrane. Use hold-to-continue and sticky selection where possible.

4. Always show the exit path. B = back/undo/cancel must be universally available. Players explore UIs more when they know they can retreat for free. This is the "dark pattern inverse" — make it safe to press buttons.

5. Curate, don't demand. Every pattern should cover ≥80% of likely inputs without text. The remaining 20% should be reachable but not prominent. If more than 20% of player actions require text, redesign the patterns, not the text input.

6. Teach through use, not through tutorial. The first time a player uses radial, show overlay labels. After 3 uses, fade labels to 0% opacity — the positions are muscle memory now. Same pattern for chips (sorted by recency then frequency, not alphabetically).

Research compiled 2026-05-11 by codex. Sources: game design analysis of 30+ console/handheld titles from SNES era to present, existing Oracle Chamber reference implementation, and D-pad-centric input ergonomics literature. Ready for UI spec integration.