translate-incoming

Wrapper for translate_incoming on the local mcp-translation server. Takes a verbatim incoming message and returns a structured decode useful for autistic / audhd readers who want the implicit corporate-speak made explicit.

Authoritative schema: packages/mcp-translation/schemas/translate_incoming.schema.json. Design rationale: docs/decisions/0005-translation-tool-design.md — ambiguity spans MUST anchor verbatim to the input.

When to use

• The user pastes a Slack / email / Linear / GitHub message and says

"what does this actually mean" or "decode this".

• The browser extension's right-click action.
• Before drafting a reply, when the user is unsure what is actually being

asked.

What it does

Calls mcp__neurodock-translation__translate_incoming. The server itself performs no LLM calls — it orchestrates a prompt template against the caller's MCP-client LLM (local Ollama or cloud, per the user's profile) and returns the typed response.

The response contains:

• explicit_ask — the surface-level request, paraphrased plainly. Null

if the message is a status update with no ask.

• likely_subtext[] — up to 5 ranked hypotheses about implicit meaning,

each with a 0..1 confidence score.

• ambiguity — a report of spans whose meaning is unclear, with

{start_char, end_char, reason, note}. The quoted span MUST equal `text[start_char:end_char]` exactly. This is the verbatim-anchor invariant from ADR 0005.

• recommended_next_action — a single coarse action + a short

justification.

• model_provenance — which model ran the call (local vs cloud, model id).

How to invoke

Input shape:

{
  "text": "Hey — can we revisit the rollout timeline? I'm not sure everyone is aligned.",
  "channel": "slack",
  "thread_context": [
    "We agreed last sprint to ship the rollout by end of May.",
    "Two engineers flagged risk on the migration script."
  ]
}

Field rules:

• text (required): 1–8000 chars. Stored only in-memory; never logged.
• channel (optional): one of the v0.1.0 enum

(slack, email, teams, linear, github, generic). Influences register expectations. Use generic if unsure.

• thread_context (optional): up to 20 prior messages, oldest first.

Each entry is a verbatim quote, 1–2000 chars.

• target_language (optional): BCP-47 tag (e.g. en-IE, de, ja).

Falls back to English prompts if no language pack is registered.

The verbatim-anchor invariant

Every ambiguity span the caller receives satisfies:

input.text[span.start_char : span.end_char]  ==  (the substring shown in note context)

The server's VERBATIM_ANCHOR_FAILED error guarantees this — the LLM cannot fabricate a span. If the response is invalid, the server rejects and the caller retries; it does not "fix up" mismatches.

When rendering ambiguity to a user, always quote the verbatim substring, not a paraphrase. The whole point of the schema's anchor is that the user can trust the quote.

Errors

• TEXT_REQUIRED — empty text.
• TEXT_TOO_LONG — over 8000 chars. Split or excerpt.
• CHANNEL_UNKNOWN — channel not in the v0.1.0 enum. Use generic.
• LANGUAGE_PACK_MISSING — target_language has no pack. Retry without

it to fall back to English.

• MODEL_UNAVAILABLE — caller's MCP client could not reach its LLM.

Server does not retry; caller decides to surface or queue.

• VERBATIM_ANCHOR_FAILED — LLM fabricated a span. Retry the call.

Limitations

• The translation quality depends on the caller's configured model.

Local 7B Ollama is more likely to hallucinate than a frontier cloud model, and the verbatim-anchor enforcement will reject those responses. Acceptable failure mode — see ADR 0005 trade-offs section.

• The server does not store transcripts. If the user wants to come back

to a decode later, the caller must persist it.

• Eval coverage lives at

packages/evals/corpora/translation/incoming/v0.1.0/general.jsonl — prompt changes ship through the eval pipeline.

Voice

When surfacing subtext to the user, lead with the top-confidence hypothesis and label it as a hypothesis ("most likely they mean X"), not as a fact. Quote ambiguous spans verbatim. Do not perform the recommended next action without confirmation — the user reads the brief; the user decides.