Add MCP-managed dictation context for Deepgram keyterms #240

Problem

Agent Code wraps speech-to-text input in <stt ...> because dictation can contain transcription mistakes. That warning is useful, but it is reactive: the model sees a possibly-wrong transcript after Deepgram has already guessed.

Deepgram supports vocabulary/context biasing (keyterm for current Nova-3/Flux flows, and keywords with boosts for older model flows). Agent Code should expose a safe feedback loop where agents can inspect and improve the dictation context used for future speech-to-text.

Example failures this should address:

• MCP being heard as unrelated words
• pull request / PR being mangled
• project names, branch names, issue numbers, provider names, and repo-local terms being misheard
• casing/formatting knowledge such as Codex, Claude, Deepgram, OpenAI, GitHub, TypeScript

Proposed shape

Add an optional built-in MCP domain, tentatively dictation_context, that lets an active agent read and update a bounded dictation vocabulary/context store.

The MCP server should expose tools like:

• dictation_context_read
  • returns currently active keyterms/keywords, formatting rules, model/provider mode, and limits
• dictation_context_add_terms
  • adds short technical/domain terms with optional aliases and rationale
• dictation_context_remove_terms
  • removes stale or bad terms
• dictation_context_record_correction
  • records that a misheard phrase should map to a preferred term in future context
• optionally dictation_context_suggest_from_workspace
  • proposes terms from repo names, branches, package names, open issues, and common Agent Code concepts

Deepgram integration

Current code path:

• streaming Deepgram URL is built in packages/agent-voice-dictation/src/speech/deepgram/streaming.ts
• Agent Code starts streaming in src/main/ipc/dictation.ts
• current streaming URL only sets model

Desired behavior:

• For Flux / Nova-3-compatible flows, append repeated keyterm=<term> query params.
• For older Deepgram model flows, use keywords=<term>:<boost> only where that API is appropriate.
• Batch fallback should receive equivalent context, not only the streaming preview.
• Dictation debug logs should record which context version/hash was used, but must not leak API keys or full private audio.

Storage / safety constraints

• Store context under Agent Code state, not inside project files by default.
• Keep a small, bounded active list, for example top 100 terms.
• Terms should be short words or noun phrases, not arbitrary prompt text.
• Normalize duplicates and casing.
• Track provenance: user-added, agent-added, inferred, timestamp, optional rationale.
• Agents should be allowed to improve vocabulary, not inject broad behavioral instructions into dictation.
• Provide a way to inspect and clean up stale terms.

Acceptance criteria

• A built-in MCP domain exists for dictation context and can be enabled for selected agents.
• Agents can read the active dictation vocabulary/context through MCP.
• Agents can add/remove bounded terms through MCP with validation.
• Deepgram streaming and batch transcription receive the active context.
• Debug traces show the context version/hash and term count used for each dictation session.
• Existing dictation works unchanged when no context terms are configured.

Related code

• packages/agent-voice-dictation/src/speech/deepgram/streaming.ts
• packages/agent-voice-dictation/src/speech/deepgram/index.ts
• src/main/ipc/dictation.ts
• src/mcp/runtime/createBuiltInMcpServer.ts
• src/mcp/shared/types.ts