Spec — AI multimodal input composer

Voice integration via voice-mode.kmd (#121) e voice/wake-word.kmd. Capability gating via model-selector.kmd (#113). Impl ticket voice-side: services/ai/ai#115 (já aberto).

Princípios

1. One composer, four modes — text, image, file, voice. No produto separated UI per mode.
2. Drag-and-drop + paste + attach — todo o composer aceita drops; paste from clipboard auto-attaches.
3. MIME validation gated by model — accept only what active model can consume.
4. Voice as transition — mic button enters dedicated voice mode (#121), não inline transcription.
5. Storage scoped — uploads em path-prefix per workspace.

R1 — Anatomia

┌─────────────────────────────────────────────────┐
│ [📷 photo.jpg · 240KB · ✗]                      │  ← upload preview chips
│ [📄 spec.pdf · 1.2MB · ✗]                       │
├─────────────────────────────────────────────────┤
│ [🖼] [📎] [text input area...........] [🎙] [➤]│  ← composer row
└─────────────────────────────────────────────────┘

Slots:

Drop zone: ENTIRE composer area accepts drops (visual highlight on dragover).

R2 — Upload preview chips

Per file attached:

┌────────────────────────────────────┐
│ [thumb] filename · MIME · size  ✗ │
└────────────────────────────────────┘

Thumb:

• Image: actual thumbnail (96×96 max).
• PDF: first page render (futuro; v1 = generic icon).
• Audio: waveform mini.
• Other: MIME icon.

Validation states:

R3 — MIME validation per model

Current model capability (from model-selector.kmd #113):

Attach button filters file picker by accepted MIMEs.

Drag-drop: invalid MIME shows red overlay + reject sound.

R4 — Drag-and-drop

• dragover entire composer: visual highlight (border accent + tint).
• drop: validates each file MIME; valid → adds to chips; invalid → rejection toast.
• dragend without drop: removes highlight.

Mobile equivalent: long-press file in OS file manager → "Share to" → Koder app target.

R5 — Paste hook

• Image in clipboard: auto-attach as pasted-{timestamp}.png.
• URL in clipboard: if model supports web-fetch, offer "Attach URL preview" chip (chip = fetched + summarized).
• Text: standard paste in text input (default behavior).

R6 — Voice button → voice mode

Tap mic button:

1. Check voice.enabled toggle from voice/wake-word.kmd R1 — if off, prompt user to enable in settings.
2. Request mic permission (if not granted).
3. Transition to fullscreen voice mode per voice-mode.kmd (#121).
4. Voice mode user closes → back to composer; transcription (if transcript_to_text enabled) populates text input.

NÃO renderiza waveform inline no composer; voice mode é separated UI.

R7 — Storage path-prefix per workspace

Per policies/multi-tenant-by-default.kmd:

• Object store path: koder://workspaces/<workspace_id>/users/<koder_user_id>/uploads/<conversation_id>/<file_uuid>.{ext}
• Cross-tenant access returns 404.
• Retention per identity-data-retention.kmd: orphaned uploads (conversation deleted) cascade after 24h grace.

R8 — Limits

Limits configurable per workspace/product.

R9 — Surface bindings

R10 — Acessibilidade

• Composer: role="form" aria-label="Message composer".
• Buttons: aria-label per button.
• Drop zone: announce drop state via aria-live.
• Chips: <li> semantically; remove button per chip.
• Keyboard: Tab cycle; Cmd/Ctrl+Enter to submit.
• Reduced-motion: no slide animations.

R11 — i18n

R12 — Per-preset variation

Cosmetic. Composer is always functional.

T-suite

• T1 Mount: composer renders all 4 input modes.
• T2 Text submit: type + Enter (or Send) → onSubmit callback with text only.
• T3 Attach image: file picker → image added to chips → submit with attachment.
• T4 Drag-and-drop: simulate dragover + drop file → chip added.
• T5 Paste image: simulate clipboard image paste → chip auto-attached.
• T6 MIME rejection: drop non-supported MIME → red chip + rejection toast.
• T7 Mic transitions: tap mic → voice-mode.kmd UI invoked.
• T8 Limits enforced: try to attach 11th file → blocked; try to attach > 25MB file → blocked.
• T9 Multi-tenant: upload from workspace A → path-prefix correct; workspace B can't access.
• T10 Keyboard: Cmd+Enter submits; arrow keys nav between chips.
• T11 A11y: drop announcement via aria-live; buttons aria-labeled.
• N1 Mic without permission: graceful prompt + degrade (no voice option).

Cross-link

• Companion: voice-mode.kmd (#121), model-selector.kmd (#113 — MIME gating), chat-message-bubble.kmd (submit destination)
• Voice base: voice/wake-word.kmd
• Storage: policies/multi-tenant-by-default.kmd
• Impl voice ticket: services/ai/ai/backlog/pending/115-cli-desktop-voice-input.md
• Refs: shadcn AI Input Group, LangChain agent-chat-ui multimodal