Context

Part of #415. Both applyMediaUnderstanding() and applyLinkUnderstanding() pipelines must be gutted.

Architectural decision: Middleware routes media and links to CLI runtimes — it does NOT process them itself. Text extraction, image vision, video vision, and URL content fetching are the CLI agent's job, not ours. The multimodal contract (#385) and runtime implementations (#397, #396) ensure the path from channel → CLI exists. STT/TTS is the only exception because it's channel-level infrastructure.

Blast radius

src/media-understanding/ — 49 files, ~5.5k lines

Consumers to decouple/remove:

Config schema to remove: MediaUnderstandingScopeSchema, MediaUnderstandingCapabilitiesSchema, MediaUnderstandingAttachmentsSchema, MediaUnderstandingModelSchema, ToolsMediaUnderstandingSchema in src/config/zod-schema.core.ts. Types in src/config/types.tools.ts.

src/link-understanding/ — 6 files, ~333 lines

Consumers to remove:

Config schema to remove: LinkToolsConfig, LinkModelConfig types, tools.links config field, schema labels/help entries.

Note: link-understanding imports shared utilities from media-understanding (CLI_OUTPUT_MAX_BUFFER, resolveTimeoutMs, resolveMediaUnderstandingScope) — both modules die together, no relocation needed for these.

Work order

1. Relocate STT shared utilities — move what src/stt/ actually needs out of media-understanding/ into src/stt/
2. Switch discord/voice — use src/stt/ directly instead of media-understanding runner
3. Remove both pipeline calls from get-reply.ts (applyMediaUnderstanding, applyLinkUnderstanding)
4. Remove auto-reply consumers — media decision types, status display, media note, link understanding context field
5. Remove config schema — tools.media (image/video/audio vision config), tools.links
6. Delete src/media-understanding/ and src/link-understanding/ entirely
7. Verify — all tests pass, STT still works, voice messages work, no orphan imports

Depends on

• refactor(media): extract STT from media-understanding into src/stt/ #424 — STT extraction (done ✅)

Does NOT depend on

• feat(runtime/gemini): multimodal I/O support #397 / feat(runtime/claude): multimodal I/O support #396 — Gemini/Claude multimodal are independent. The middleware already routes media or notifies users (feat(middleware): communicate multimodal limitations to users when runtime can't handle media #400). Gutting the vision/link pipelines doesn't need runtimes to implement their multimodal support first.

Related

• feat: unified auth + multimodal AgentRuntime — implementation plan #415 — implementation plan (parent)
• research(auth+media): auth profiles and media understanding architecture #376 — original gut proposal
• feat(middleware): add multimodal media attachments to AgentRuntime contract #385 — multimodal contract (done ✅)
• feat(middleware): communicate multimodal limitations to users when runtime can't handle media #400 — unsupported media notification (done ✅)