Orchestrator engine:
- engine.rs: unified Mistral Conversations API tool loop that emits
OrchestratorEvent instead of calling Matrix/gRPC directly
- tool_dispatch.rs: ToolSide routing (client vs server tools)
- Memory loading stubbed (migrates in Phase 4)
Server-side tokenizer:
- tokenizer.rs: HuggingFace tokenizers-rs with Mistral's BPE tokenizer
- count_tokens() for accurate usage metrics
- Loads from local tokenizer.json or falls back to bundled vocab
- Config: mistral.tokenizer_path (optional)
No behavior change — engine is wired but not yet called from
sync.rs or session.rs (Phase 2 continuation).
- gRPC dev_mode config: disables JWT auth, uses fixed dev identity
- Agent prefix (agents.agent_prefix): dev agents use "dev-sol-orchestrator"
to avoid colliding with production on shared Mistral accounts
- Coding sessions use instructions (system prompt + coding addendum)
with mistral-medium-latest for personality adherence
- Conversations API: don't send both model + agent_id (422 fix)
- GrpcState carries system_prompt + orchestrator_agent_id
- Session.end() keeps session active for reuse (not "ended")
- User messages posted as m.notice, assistant as m.text (role detection)
- History loaded from Matrix room on session resume
- Docker Compose local dev stack: OpenSearch 3 + Tuwunel + SearXNG
- Dev config: localhost URLs, dev_mode, opensearch-init.sh for ML setup
new search_web tool calls SearXNG (cluster-internal, free, no tracking)
instead of Mistral's built-in web_search ($0.03/query + rate limits).
returns structured results from DuckDuckGo, Wikipedia, StackOverflow,
GitHub, arXiv, and Brave. no API keys, no cost, no rate limits.
removed Mistral AgentTool::web_search() from orchestrator — replaced
by the custom tool which goes through Sol's normal tool dispatch.
new engagement types: Respond (inline), ThreadReply (threaded),
React, Ignore. LLM returns response_type to decide HOW to engage.
silence mechanic: "shut up"/"be quiet" sets a 30min per-room timer.
only direct @mention breaks through.
structural suppression (A+B):
- reply to non-Sol human → capped at React
- 3+ human messages since Sol → forced passive mode
threads have a lower relevance threshold (70% of spontaneous).
time context injected into evaluator prompt.
the evaluator now receives sol's entire system prompt as a system
message, giving ministral-3b deep context on sol's personality when
scoring relevance. evaluation context window bumped from 25 to 200
messages, room/dm context windows unified at 200.
pre-computed timestamp variables ({ts_yesterday}, {ts_1h_ago},
{ts_last_week}) added to personality template for accurate time
references without LLM math.
Mistral Agents + Conversations API integration:
- Orchestrator agent created on startup with Sol's personality + tools
- ConversationRegistry routes messages through persistent conversations
- Per-room conversation state (room_id → conversation_id + token counts)
- Function call handling within conversation responses
- Configurable via [agents] section in sol.toml (use_conversations_api flag)
Multimodal support:
- m.image detection and Matrix media download (mxc:// → base64 data URI)
- ContentPart-based messages sent to Mistral vision models
- Archive stores media_urls for image messages
System prompt rewrite:
- 687 → 150 lines — dense, few-shot examples, hard rules
- {room_context_rules} placeholder for group vs DM behavior
- Sender prefixing (<@user:server>) for multi-user turns in group rooms
SQLite persistence (/data/sol.db):
- Conversation mappings and agent IDs survive reboots
- WAL mode for concurrent reads
- Falls back to in-memory on failure (sneezes into all rooms to signal)
- PVC already mounted at /data alongside Matrix SDK state store
New modules:
- src/persistence.rs — SQLite state store
- src/conversations.rs — ConversationRegistry + message merging
- src/agents/{mod,definitions,registry}.rs — agent lifecycle
- src/agent_ux.rs — reaction + thread progress UX
- src/tools/bridge.rs — tool dispatch for domain agents
102 tests passing.
Three memory channels: hidden tool (sol.memory.set/get in scripts),
pre-response injection (relevant memories loaded into system prompt),
and post-response extraction (ministral-3b extracts facts after each
response). User isolation enforced at Rust level — user_id derived
from Matrix sender, never from script arguments.
New modules: context (ResponseContext), memory (schema, store, extractor).
ResponseContext threaded through responder → tools → script runtime.
OpenSearch index sol_user_memory created on startup alongside archive.
Matrix bot with E2EE (matrix-sdk 0.9) that passively archives all
messages to OpenSearch and responds to queries via Mistral AI with
function calling tools.
Core systems:
- Archive: bulk OpenSearch indexer with batch/flush, edit/redaction
handling, embedding pipeline passthrough
- Brain: rule-based engagement evaluator (mentions, DMs, name
invocations), LLM-powered spontaneous engagement, per-room
conversation context windows, response delay simulation
- Tools: search_archive, get_room_context, list_rooms, get_room_members
registered as Mistral function calling tools with iterative tool loop
- Personality: templated system prompt with Sol's librarian persona
47 unit tests covering config, evaluator, conversation windowing,
personality templates, schema serialization, and search query building.
