App ↔ research bridge

When to read this: You know either the production Chatty app or a research module, and need to see how they connect — memory, context window, agent loop, traces, sub-agents, and where SOTA work lands.

Parent frame: Paper → experiment → product.

Two views of the same system

flowchart TB
  subgraph app [Production Chatty — ships today]
    Loop[Agent loop<br/>llm_service + AgentFactory]
    Mem[Memory + skills<br/>memory_service]
    Ctx[Context window<br/>token_budget]
    Stream[StreamManager + ATIF export]
    Tools[~60 tools + MCP]
    SubAg[Sub-agents<br/>invoke_agent_tool]
    Settings[Settings models<br/>ModelConfig, Execution, Training]
  end

  subgraph research [Research domains — experiment → promote]
    M0[M0 Trace contract]
    M1[M1 ReAct fidelity]
    M2[M2 AFlow topology]
    M3[M3 GEPA prompts]
    M4[M4 ACE playbook]
  end

  Loop --> M1
  Loop --> M0
  Stream --> M0
  Mem --> M4
  Ctx --> M3
  Ctx --> M4
  Settings --> M3
  SubAg --> M2
  Tools --> M1
  Tools --> M2
  M0 --> M2
  M0 --> M3
  M0 --> M4

Read left → right to find which research module touches a production component. Read right → left to find where a promoted mechanism lands in the app.

Master mapping

Full settings field map: settings integration map.

Agent loop & streaming

Production today

The ReAct-shaped loop runs in chatty-core: AgentFactory builds a rig-agent client, llm_service streams multi-turn tool calls, and GPUI’s StreamManager owns lifecycle (cancel, token usage, trace JSON) without duplicating response text.

Research touchpoints

Reserved: Strategy backoff rule (RESERVED.md) — affects loop termination only; loop body is not rewritten.

Memory, skills & playbook (M4)

Production today

Persistent memory uses memvid (memory.mv2). The agent gets remember and search_memory; skills are stored with a [SKILL] title prefix or as filesystem SKILL.md files. There is no automatic recall: the preamble tells the model to call search_memory itself. How that retrieval works, stage by stage, is in M6 Retrieval Part B.

Research touchpoints (ACE)

Promotion: playbook runtime likely a setting (AGE-47 scope: global / per-model / per-conversation).

Context window & token budget

Production today

Before each LLM call, Chatty BPE-estimates preamble + tools + history against ModelConfig.max_context_window. The footer bar shows fill level; actual token counts patch in after the stream completes.

Research touchpoints

Future product link: conversation compaction (the per-call guard in services/context_shaper.rs and the persisted summarize_oldest_half) is the natural place to apply ACE grow-and-refine eviction when playbook + history exceed pressure thresholds — the compaction ships; the playbook hook into it does not exist yet.

Model config & prompts (M1, M3)

Production today

Each model has a persisted ModelConfig: preamble (system prompt), temperature, context limit, per-million token pricing, multimodal flags.

#![allow(unused)]
fn main() {
// crates/chatty-core/src/settings/models/models_store.rs
pub struct ModelConfig {
    pub preamble: String,
    pub temperature: f32,
    pub max_context_window: Option<i32>,
    pub cost_per_million_input_tokens: Option<f64>,
    pub cost_per_million_output_tokens: Option<f64>,
    // ...
}
}

Research touchpoints

Promotion: GEPA output → preamble per model (setting); apply via AGE-45.

Traces, ATIF & training export (M0)

Production today

Conversations export to ATIF v1.6 (exporters/types.rs). StreamManager holds trace_json per stream. TrainingSettingsModel toggles auto-export after each assistant response.

Research touchpoints

Deep dive: M0 Trace module page, chatty-trace README.

Tools, sub-agents & workflow (M1, M2)

Production today

~60 native tools plus MCP integrations. Sub-agent tools compose other agents from the registry. Execution settings control workspace path, approval mode, and sandbox.

Research touchpoints (AFlow)

Monty code mode (sandbox/monty_bridge.rs) could become PythonRepr for AFlow. Today MontySandbox (sandbox/monty.rs) runs stdlib-only Python through a python3 subprocess under time/memory limits, and SandboxManager tries it before Docker; the bridge’s ToolBridge interface exists but the pause/resume loop that would let scripts call Chatty tools is not wired. Source: crates/chatty-core/src/sandbox/.

Optimizer & offline runs (M2, M3)

Not in the request path. Search runs via chatty-optimize (planned CLI: chatty-tui --optimize, AGE-44).

Stage B evals run in harbor-chatty, not inside the app binary.

Quick lookup: research module → app landing

Related

• Component map — crate/module/entity diagrams
• System overview — three-layer mental model
• Module pages — per-paper fidelity and eval protocol
• Experiment protocol — Stage A/B bar and promotion inputs