[senamakel]

Summary

This is the last cluster of vendor-extraction wave 4. It moves web-chat error text matchers, emoji extraction, voice MIME mapping, reply-speech parsing and the external speech-to-text (STT) and text-to-speech (TTS) HTTP clients into tinyinference.

crates/openhuman-core/src: 594,536 → 593,405 lines (−1,131), measured on top of #6838.

Stacked on #6838, which is itself stacked on #6829. Merge those first.

Depends on:

• feat: move string-level failure predicates, emoji extraction and voice clients from OpenHuman tinyinference#46 (c2386549), which stacks on chore: remov unwanted/mobile references and update skills submodule #45.
• chore(deps): bump tinyinference for Wave 4 inference extraction tinyagents#253 (1804d6f4), which stacks on #251 and bumps the tinyinference gitlink.

What moved (tinyinference)

• tinyinference-llm failure::text: the pure predicates from web_chat/web_errors/{response_predicates,retry,provider_detail}.rs.
  • This covers extract_provider_name, extract_provider_error_detail, with_provider_detail and is_fallback_chain_exhausted.
  • parse_retry_after_secs is added upstream as its own function, next to parse_retry_after_ms. It keeps the host's semantics: camelCase keys, quote stripping and rounding up.
  • User-facing copy stays in the host: malformed_history_user_message, retry_after_hint and the error classification table.
• tinyinference-llm classification::emoji: extract_first_emoji, is_emoji_start and is_regional_indicator, from inference reactions.
• tinyinference-voice mime: mime_for_extension and extension_for_mime.
• tinyinference-voice reply:
  • AlignmentFrame, ReplySpeech and the viseme and alignment parsers, with a literal-JSON test of the shape.
  • VisemeFrame now uses the vendor type.
  • synthesize_reply stays in the host.
• tinyinference-voice: the ExternalSttClient and ExternalTtsClient classes (OpenAI-compatible, Deepgram and ElevenLabs).
  • The caller passes in the reqwest::Client, so proxy and TLS stay under host control.
  • SttApiStyle and TtsApiStyle move upstream behind an optional schemars feature, which the host enables. The published config schema and the TOML shape are unchanged.
• Cron: is_local_provider_no_model_loaded_message now calls body_indicates_no_model_loaded instead of carrying its own copy.
• Deleted: TranscriptionResult, which had no users.

API note: openhuman::config::schema::{SttApiStyle, TtsApiStyle} no longer exists. The runtime-boundary check forbids re-exporting tinyinference types, so callers import them from tinyinference_voice directly. openhuman-cli gains a dev-dependency for its config e2e test.

Error strings are byte-identical.

Verification

• Vendor:
  • tinyinference-llm: 448 tests pass.
  • tinyinference-voice: 41 tests pass, with and without schemars.
  • The workspace passes fmt and clippy -D warnings.
• OpenHuman:
  • Every test target builds with product features, as do the app manifest and embed/tinyhumans.
  • Full lib suite: 9,570 passed, 1 failed. The failure, credential_scrub::browser_confirmation_token_survives_without_exempting_page_credentials, is known on base.
• CI scripts:
  • Feature forwarding, the gated-test allowlist and submodule monotonic pass.
  • rust:layout and the runtime-boundary check fail only on the entries that already fail on base.
  • No module pins moved.

Summary by CodeRabbit

• Refactor
  • Agent memory guidance, run controls, progress tracking, voice processing, network access, and usage reporting now use shared implementations. Existing workflows are routed through updated integrations.
  • Network tools retain host-level security checks, including privacy and approval policies.
• Documentation
  • Updated technical guides to reflect current tool responsibilities and integrations.

The merge-base changed after approval.