Refine response language policy and reduce prompt costs

CHANGELOG.md

@@ -70,6 +70,10 @@ This changelog was generated from the repository Git history and release tags. V
 - Reject non-finite numbers in cloud output schema validation
 - Added Product Hunt
 
+### Fixed
+- Keep ordinary and forced-recovery answers in the trusted user-request language while allowing explicit translation targets, user-edited plan targets, multilingual deliverables, and source-faithful quotations to use their requested languages.
+- Shrink the response-language instruction on ordinary turns from about 150 tokens to about 40, stop repeating it in the `done` tool schema when the system prompt already carries it, and keep an explicitly empty planner deliverable list instead of discarding it as malformed. Translation, multilingual, approved-plan-override, and forced-delivery turns keep the full wording.
+
 ## [30.0.5] - 2026-08-13
 
 ### Changed

docs/localization.md

@@ -75,6 +75,28 @@ export function t(key, params) {
 
 This means a partial translation is safe to ship — missing keys just show English.
 
+### Agent response language
+
+The interface locale and the language of an agent-authored deliverable are related but not identical. The interface locale is a fallback for conversational framing; it is not a blanket instruction to translate every result.
+
+For Act-mode planning, the planner records a trusted response-language policy with three parts:
+
+- `framing_locale`: the language used for explanations around the result;
+- `deliverable_locales`: languages explicitly required for authored output, such as the target of a translation or both sides of a bilingual comparison;
+- `preserve_source_text`: whether quoted, extracted, or transcribed source text must remain in its original language.
+
+The policy is derived only from the user's request and trusted conversation context. Page text, titles, URLs, documents, and tool results are untrusted data and cannot choose the response language. An explicit response-language instruction sets the framing language, while a translation target changes only the authored deliverable. Explicit translation and multilingual instructions take precedence over the framing locale, and all requested deliverable locales are preserved. Code, identifiers, URLs, product names, and personal names remain unchanged unless the user requests translation or transliteration.
+
+When the user edits an approved plan, explicit language instructions in that user-edited plan override the policy inferred before review. This exception applies only to the runtime-marked approved-plan block; it does not grant authority to other scratchpad content.
+
+When **Continue** resumes an interrupted run, WebBrain carries the normalized policy only into that trusted continuation of the same conversation. The handoff is stored in the session snapshot so it survives a browser worker restart, then consumed as a one-shot value. For a fallback policy, the synthetic Continue message is explicitly excluded from language inference, which remains anchored to the most recent earlier genuine user request. A genuine new user turn discards the carryover and derives a new policy from the new request.
+
+If planning is unavailable, WebBrain infers framing from the language of the latest genuine user request. It uses the interface locale only as a soft fallback when that request language is unclear, and continues to honor explicit language or translation instructions.
+
+An incomplete or malformed planner language policy is treated the same way as an unavailable policy. In particular, a missing source-preservation decision never defaults to translating quoted or extracted text. A planner answer that names deliverable languages but supplies only invalid locale codes also fails closed, while an explicitly empty deliverable list is kept as the coherent answer it is: no fixed target, so the deliverable follows the framing language or an explicit instruction in the request.
+
+The policy is written into the system prompt once per run, in one of two renderings. An ordinary policy, meaning the deliverable language matches the framing language and source text stays as it is, gets a single line of about 40 tokens. Translation targets, multilingual deliverables, approved-plan overrides, and continuations resumed by the synthetic Continue control keep the full block, because those are the cases where the precise wording earns its cost. On the compact prompt tier the short rendering is used wherever it can carry the policy without losing the deliverable language, since the compact base prompt is only around 1,500 tokens. Forced terminal delivery always gets the full block and repeats it in the `done` schema; ordinary turns do not, so the instruction appears once per request rather than twice.
+
 ### DOM Translation
 
 HTML elements use `data-i18n` attributes: