admin/openhuman

Fork 0

mirror of https://github.com/tinyhumansai/openhuman.git synced 2026-07-27 21:08:00 +00:00

Files

T

History

3c247a2439 Feat/humanlike replies (#168 )

* fix(chat): prevent stacked socket listeners on reconnect

subscribeChatEvents was declared async despite having no awaits, so the
cleanup function was returned in a microtask after React's synchronous
cleanup had already run. Each socket reconnect added another layer of
listeners that were never removed, causing chat:done to fire N times and
produce duplicate message bubbles.

- Remove async keyword from subscribeChatEvents; return cleanup fn directly
- Update useEffect call site to store cleanup synchronously and return it
  to React (eliminates the mounted/then race entirely)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(local-ai): add multi-turn chat via Ollama /api/chat

Expose openhuman.local_ai_chat RPC method so the UI can run full
conversation-history chat directly through the bundled Ollama model
without touching the cloud inference API.

- ollama_api.rs: add OllamaChatMessage / OllamaChatRequest / OllamaChatResponse
  types for the /api/chat endpoint
- service/public_infer.rs: add LocalAiService::chat_with_history() — sends
  multi-turn message array to Ollama, updates latency/TPS status on response
- ops.rs: add LocalAiChatMessage struct and local_ai_chat async op
- schemas.rs: register local_ai_chat controller (schema, handler, params)
  in all_controller_schemas + all_registered_controllers

Zero cloud tokens are consumed on this path; the call never reaches the
backend socket or the /openai/v1/chat/completions endpoint.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(conversations): local-model chat gate with multi-bubble delivery

When Ollama is ready (isLocalModelActive), handleSendMessage bypasses
the cloud socket entirely and routes through openhumanLocalAiChat.
Response is segmented and delivered as multiple typed bubbles with
natural pauses — human-like reply behaviour at zero cloud token cost.

UI / delivery
- deliverLocalResponse(): segments full reply via segmentMessage(),
  dispatches each bubble with getSegmentDelay() pause between them;
  typing indicator (isDelivering) shows between segments
- Socket-connected guard skipped on local path so offline local use works
- Cloud socket path (chatSend → chat:done) fully unchanged

Frontend RPC
- tauriCommands: openhumanLocalAiChat(messages, maxTokens?) wraps
  openhuman.local_ai_chat via core RPC; LocalAiChatMessage type exported

Tests (402 passing)
- messageSegmentation: 4 new edge-case tests (whitespace, 80-char
  boundary, paragraph split, delay scaling)
- localChatGating (new file): 9 tests — segmentation correctness, delay
  bounds [500, 1400] ms, sender→role mapping for message history build

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(prompts): replace vague emoji guidance with explicit contextual rules

The previous "Minimal — match the user's style" instruction was too
loose, causing the model to stack decorative emojis on every message
(e.g. "Hey! 😄 Just cooking up some AI magic! 🚀🔥✨").

SOUL.md — add Emoji Rules section:
- Hard cap: one emoji maximum per message; none is always acceptable
- Contextual, not decorative: emoji must reinforce the specific content
  (🔥 for exciting news, 🤔 for uncertainty, ✅ for confirmations)
- Never open a sentence with an emoji
- Skip entirely in error/warning/technical/long responses
- Mirror the user's own emoji usage pattern
- Concrete good/bad examples so the model can calibrate

BOOTSTRAP.md — tighten the Communication Preferences entry to reference
the same rules rather than the old vague one-liner.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* chore: scope PR_DESCRIPTION.md to feat/humanlike-replies only

Remove unrelated package manager distribution content (was from a
different branch). Description now covers only the 4 commits on this
branch: socket listener fix, Rust local_ai_chat RPC, frontend local
chat gate + multi-bubble delivery, and emoji prompt rules.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* refactor(local_ai): improve code formatting and readability in chat operations

- Adjusted formatting in local_ai_chat function for better readability by adding line breaks.
- Simplified the mapping of messages to OllamaChatMessage in ops.rs.
- Streamlined the await syntax in schemas.rs for clarity.
- Enhanced formatting in public_infer.rs for consistency in API request construction.

* refactor(conversations): enhance code readability and structure in Conversations component

- Improved formatting and consistency in the Conversations component, including better alignment of dispatch calls and message handling logic.
- Removed redundant imports and streamlined the mapping of stored messages for clarity.
- Adjusted conditional rendering for improved readability in the UI logic.

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>