yunchan/d3ro-voice
1
0
Fork 0
Code Issues Pull requests Projects Releases 3 Packages Wiki Activity Actions

fix(llm): stop leaking instruction prompts instead of processed text

Browse source 
Running a custom instruction (translate, summarise, rewrite, explain code,
free prompt) inserted the instruction's own wording instead of the result.
Two faults stacked:

The instruction was passed as the text to process, leaving the system-prompt
argument empty. `BASE_SYSTEM_PROMPTS` has no `custom` key, so resolution fell
back to `refine` without saying so, and the model dutifully polished the
instruction it had been handed. The transcript never reached it.

And only `{{text}}` was substituted, which none of the five built-in
instructions use — they carry `{{targetLanguage}}`, `{{userPrompt}}`, or no
placeholder at all. The substitution was a no-op from the day it was written:
the presets landed ten hours before the code that expected them.

- Instruction prompts now go to the system-prompt argument and the transcript
  to the text argument. Instructions that spell out `{{text}}` keep their old
  meaning, so hand-written ones still work.
- `renderInstructionPrompt` resolves `{{text}}`, `{{userPrompt}}` and
  `{{targetLanguage}}` in one place, and warns by name when a placeholder is
  left standing rather than letting it reach the model.
- `resolveSystemPrompt` no longer drops silently to `refine` for `custom`.
- Voice shortcuts no longer die at the `defaultLLMAction === 'none'` gate; an
  explicitly named instruction outranks the default. Without one, `none` still
  passes the transcript through untouched.
- `translate` receives its target language instead of relying on a default two
  call frames away. It is still always English — `AppConfig` has no key for it,
  and neither `language` (UI locale) nor `sttLanguage` (source language) can
  stand in. Choosing a target language needs a setting and is not in this fix.
- Chains ran instructions with placeholders intact; they share the same
  resolution now.
- The command screen's pipeline bench called `llm.generate`, which preload does
  not expose, so every run threw and the catch showed the input back as if it
  had succeeded. It uses `llm.process` now, over the same path production
  takes, and a failure reads as a failure.

Present since the feature shipped: the custom-instruction path has never
worked. Plain actions (refine, summarise, grammar, expand) were unaffected and
are now covered by tests so they stay that way.
This commit is contained in:
Yun Chan 2026-09-21 14:29:43 +09:00
parent 30d51c952f
commit 99f06c253c
14 changed files with 992 additions and 44 deletions
Download patch file Download diff file Expand all files Collapse all files

12
docs/map/10-feature-catalog.md
View file

@ -39,11 +39,11 @@ Status quick-reference: `[x]` done+verified · `[~]` partial/unverified · `[ ]`
|---|---|---|---|---|---|---|
| AI-01 | Local LLM (Ollama) | [x] | [-] | [ ] | [-] | Desktop bundled Ollama |
| AI-02 | Cloud LLM (Claude/OpenAI) | [x] | [x] | [x] | [x] | Desktop `PremiumLLMService`; web/mobile via `llm-proxy`; .NET `LlmProxyService` |
| AI-03 | Auto Polish (cleanup/filler removal) | [x] | [~] | [~] | [x] | Desktop built-in; web/mobile via commands |
| AI-04 | Translate / summarize / rephrase | [x] | [x] | [x] | [x] | Built-in instructions |
| AI-05 | Custom instructions (user commands) | [x] | [x] | [x] | [x] | Desktop `CommandsPage` (Red Team RT-03 verified); web `commands`; mobile `CommandsScreen` |
| AI-06 | Voice keyword commands | [x] | [-] | [ ] | [-] | Desktop `VoiceCommandService` + command popup |
| AI-07 | LLM Chains (multi-step pipelines) | [x] | [ ] | [ ] | [-] | Desktop `ChainService` |
| AI-03 | Auto Polish (cleanup/filler removal) | [x] | [~] | [~] | [x] | Desktop built-in; web/mobile via commands. Desktop Auto Polish is the plain `refine` action (`llm-prompts.ts:14`), not a custom instruction, so it was **not** affected by the 2026-09-21 instruction-prompt fix (AI-05); regression cases now pin `refine`/`summarize`/`grammar`/`expand` (`VoiceModeService.test.ts:435`, `llm-prompts.test.ts:153`) |
| AI-04 | Translate / summarize / rephrase | [~] | [x] | [x] | [x] | Built-in instructions. **Desktop has two paths and only one of them worked.** The plain-action path (Settings → `defaultLLMAction`, `SettingsModal.tsx:653`) reads `BASE_SYSTEM_PROMPTS` directly and was always correct. The built-in *instruction* presets (`CustomInstructionService.ts:26/35/44/53/62`) ran through the custom-instruction path and inserted the instruction's own wording instead of the result — see AI-05. Fixed in `9c2b4d4` (2026-09-21), **not verified in a running app** (`11` GAP-LLM-02). Translate still always targets English: `AppConfig` has no target-language key and neither `language` (UI locale) nor `sttLanguage` (source language) can stand in (`llm-prompts.ts:37-52`, `11` GAP-LLM-01) |
| AI-05 | Custom instructions (user commands) | [~] | [x] | [x] | [x] | Web `commands`; mobile `CommandsScreen` — both go through Edge Functions and are unaffected. **Desktop: the custom-instruction path never worked in any shipped release.** The instruction was passed in the `text` argument of `processText(text, action, targetLanguage, customPrompt)` with the system-prompt argument left empty; `BASE_SYSTEM_PROMPTS` has no `custom` key, so resolution fell back to `refine` silently and the model polished the instruction it was handed — the transcript never reached it. Separately, only `{{text}}` was substituted and **none of the five built-ins use it** (`{{targetLanguage}}`, `{{userPrompt}}`, or no placeholder), so the substitution was a no-op from the day it was written. Introduced `fea923d` (2026-04-05); present `v0.1.0-alpha`..`v1.4.0`; **not a regression**. Three entry points were affected: commands-page activation (`CommandsPage.tsx:102-117`), command-popup selection (`bootstrap.ts:377-381`), voice keyword match (`VoiceModeService.ts:681-686`). Fixed in `9c2b4d4` (2026-09-21) — `llm-prompts.ts` is now the SSOT for placeholder substitution and argument placement (`renderInstructionPrompt:78`, `buildInstructionInvocation:101`, `resolveSystemPrompt:118`). The earlier "Red Team RT-03 verified" claim did not catch this and its cited evidence file (`red_team_log.md`) is not in the repo. `[~]` because unit tests pass (`llm-prompts.test.ts` 21, `llm-handlers.test.ts` 7, `VoiceModeService.test.ts` 22, `ChainService.test.ts` 5) but there is **no running-app verification** and the related `tests/red/{instruction,chain,voice,config}.usecase.test.ts` could not execute (`11` GAP-INFRA-06). See `11` GAP-LLM-01/02 |
| AI-06 | Voice keyword commands | [~] | [-] | [ ] | [-] | Desktop `VoiceCommandService` + command popup. Keyword matching itself works (`VoiceModeService.ts:681-686`), but execution went through the broken custom-instruction path (AI-05), and a **second, separate defect** made the shortcut skip LLM processing entirely whenever `defaultLLMAction === 'none'` — which is exactly the value the commands UI and the command popup write when no command is active (`CommandsPage.tsx:116`, `bootstrap.ts:385`), so an explicitly named instruction was nullified by the default setting. Both fixed in `9c2b4d4` (`VoiceModeService.ts:779`/`:785`); **not verified in a running app** (`11` GAP-LLM-02) |
| AI-07 | LLM Chains (multi-step pipelines) | [~] | [ ] | [ ] | [-] | Desktop `ChainService`. The argument placement was already correct here, but chain steps **never substituted placeholders**, so `{{targetLanguage}}` / `{{userPrompt}}` reached the model verbatim as the system prompt. Now shares the same resolution function as the other two paths (`ChainService.ts:196`). Fixed in `9c2b4d4` (2026-09-21); 5 unit tests GREEN, but `tests/red/chain.usecase.test.ts` could not execute (`11` GAP-INFRA-06) and there is **no running-app verification** (`11` GAP-LLM-02) |
| AI-08 | Screen/context capture for prompts | [x] | [-] | [ ] | [-] | Desktop `ScreenContextService` |
| AI-09 | Streaming responses | [x] | [x] | [x] | [x] | SSE/NDJSON streaming |
| AI-10 | Dictation templates (voice form fill) | [x] | [ ] | [x] | [~] | Desktop `DictationTemplateService`; mobile `TemplatesScreen` |
@ -204,7 +204,7 @@ Status quick-reference: `[x]` done+verified · `[~]` partial/unverified · `[ ]`
| Surface | `[x]` | `[~]` | `[ ]` | Notable strength | Notable weakness |
|---|---|---|---|---|---|
| Desktop | ~40 | 3 | ~8 | Local AI pipeline, meetings, RAG, conversation, key bindings | Ads stubs, no team admin, no email account |
| Desktop | ~40 | 3 | ~8 | Local AI pipeline, meetings, RAG, conversation, key bindings | Ads stubs, no team admin, no email account; custom-instruction/chain LLM path (AI-04..07) fixed 2026-09-21 but unverified in a running app |
| Web | ~22 | 6 | ~14 | Server-shared data UX, billing, meetings, teams | No local AI, limited knowledge upload/search |
| Mobile | ~40 | 12 | ~18 | Cloud + native recording, portability, admin, IAP/ads | External store/console gates, a11y, deep E2E pending |
| Backend | ~45 | 6 | ~4 | RLS, Edge functions, billing, fail-closed AI | Payple webhook signature, some external provider keys |