81 lines
5.3 KiB
Markdown
81 lines
5.3 KiB
Markdown
# Chat Turn Sequence
|
|
|
|
This diagram traces one conversational chat turn end to end: the frontend POSTs to the real endpoint `POST /api/chat`, which is an SSE stream (the server validates, then calls `reply.hijack()` and writes a raw `text/event-stream`). The route handler (`packages/server/src/local/routes/chat.ts`) assembles the layered system prompt via the per-session `Orchestrator` (`buildSystemPrompt` + `recallMemory`, where recall runs `HybridSearch` over the workspace mind), filters tools by persona/context, then calls `runAgentLoop`, which POSTs to LiteLLM at `${litellmUrl}/chat/completions`. Tool calls pass through the 11-step `executeToolCall` middleware chain (governance, hooks, LoopGuard, injection scan); a `TraceRecorder` wires additive trace callbacks. Tokens and tool events stream back to the UI as named SSE events, and after the loop `autoSaveFromExchange` runs the `CognifyPipeline` to write new memory frames. All claims are grounded in `sections/03a-api-chat-agents.md`, `sections/05a-subsystem-agent-runtime.md`, and `sections/05b-subsystem-memory.md`.
|
|
|
|
```mermaid
|
|
sequenceDiagram
|
|
autonumber
|
|
participant UI as "Frontend (chat UI)"
|
|
participant Route as "POST /api/chat\nchat.ts (SSE)"
|
|
participant Orch as "Orchestrator\nbuildSystemPrompt + recallMemory"
|
|
participant Search as "HybridSearch\n(workspace mind)"
|
|
participant Loop as "runAgentLoop\nagent-loop.ts"
|
|
participant LLM as "LiteLLM\n/chat/completions"
|
|
participant Tools as "executeToolCall\n11-step chain"
|
|
participant Trace as "TraceRecorder"
|
|
participant Cog as "CognifyPipeline\nautoSaveFromExchange"
|
|
participant Disk as "session .jsonl"
|
|
|
|
UI->>Route: "POST { message, workspace, session, model?, persona?, autonomy? }"
|
|
Note over Route: "validate · injection scan score < 0.7 · RBAC · path guard (all pre-hijack)"
|
|
alt rejected
|
|
Route-->>UI: "400 / 403 JSON error (MESSAGE_TOO_LONG, INJECTION_DETECTED, ...)"
|
|
else accepted
|
|
Note over Route: "reply.hijack() · write SSE headers · wire AbortController to client close"
|
|
Route->>Disk: "persistMessage user turn (append .jsonl)"
|
|
Route->>Route: "generateTurnId UUID v4 · resolve model fallback chain"
|
|
|
|
Route->>Orch: "recallMemory(query, turnId)"
|
|
Note over Orch: "catch-up vs semantic · drop temporary/deprecated · injection scan"
|
|
Orch->>Search: "search(query) keyword + vector"
|
|
Search-->>Orch: "SearchResult[] ranked by finalScore = rrfScore * relevanceScore"
|
|
Orch-->>Route: "recalledContext text"
|
|
Route-->>UI: "event: step Recalling relevant memories..."
|
|
Route-->>UI: "event: tool / tool_result (auto_recall)"
|
|
|
|
Route->>Orch: "buildSystemPrompt()"
|
|
Orch-->>Route: "identity + self-awareness + preloaded context"
|
|
Note over Route: "layer profile + runtime facts + activeSpec.rules + skills + Workspace Now + corrections, then composePersonaPrompt"
|
|
Note over Route: "filterToolsForContext + filterAvailableTools + persona allow/deny"
|
|
|
|
Route->>Loop: "runAgentLoop { systemPrompt, tools, messages, stream:true, maxTurns:200, turnId, traceRecording }"
|
|
|
|
loop "each turn up to maxTurns"
|
|
Loop->>LLM: "POST /chat/completions { model, messages, tools, stream }"
|
|
Note over Loop,LLM: "signal = client-abort + 300s timeout · 429/5xx -> backoff, turn--, retry"
|
|
LLM-->>Loop: "stream chunks: content + tool_calls + usage"
|
|
Loop-->>UI: "event: token xN assistant text"
|
|
|
|
alt "tool_calls present"
|
|
Loop-->>UI: "event: tool { name, input }"
|
|
Loop->>Tools: "executeToolCall(name, args)"
|
|
Note over Tools: "parse args -> onToolUse -> governance blockedTools -> pre:tool -> pre:memory-write -> LoopGuard.check -> execute -> scanForInjection -> onToolResult -> post hooks"
|
|
opt "gated tool"
|
|
Loop-->>UI: "event: approval_required { requestId, toolName, input }"
|
|
UI->>Route: "POST /api/approval/:requestId { approved, always? }"
|
|
Route-->>Loop: "resolve(approved) (auto-deny after 5 min)"
|
|
end
|
|
Tools->>Trace: "record tool call + sanitized result"
|
|
Tools-->>Loop: "role:tool result message (sanitized)"
|
|
Loop-->>UI: "event: tool_result { name, result, isError }"
|
|
else "no tool_calls (final answer)"
|
|
Note over Loop: "maybeFireCompletionGate: D3 verification -> D4 phantom-write -> D1 skill-distillation"
|
|
alt "a gate fired"
|
|
Note over Loop: "push corrective directive, continue one more turn"
|
|
else "none fired"
|
|
Loop-->>Trace: "finalize trace"
|
|
Loop-->>Route: "AgentResponse { content, toolsUsed, usage }"
|
|
end
|
|
end
|
|
end
|
|
|
|
Note over Route: "cost tracking · KG entity extraction · disclaimers"
|
|
Route->>Cog: "autoSaveFromExchange(message, result.content)"
|
|
Cog->>Disk: "cognify writes new memory frame (I/P) + FTS + vec index + entities"
|
|
Route->>Orch: "commitSurfacedSignals()"
|
|
Route->>Disk: "persistMessage assistant turn"
|
|
Route-->>UI: "event: done { content, usage, toolsUsed, model, cost? }"
|
|
end
|
|
Note over Route: "on failure -> event: error { message } (raw turn still persisted) · finally: raw.end()"
|
|
```
|