Settings integration map

When to read this: You need to know which user-facing settings correspond to which research mechanisms — and which are still blocked on product decisions.

Parent: Paper → product pipeline.

Today (shipping)

Settings model	Field(s)	Research link	Promotion today
`ModelConfig`	`preamble`	M1 few-shot block; M3 GEPA target	User-editable default
`ModelConfig`	`temperature`, `cost_per_million_*`	M3 cost accounting	Ships
`ModelConfig`	`supports_images`, `supports_pdf`, …	—	Ships (unrelated)
`TrainingSettingsModel`	`atif_auto_export`	M0 trace export toggle	Setting
`ModuleSettingsModel`	`virtual_agents`	Named worker roles for the small-teams experiment (ADR-0011 C10, `dev/design/fleet.md` §2)	Ships (JSON only, no GPUI page)
`ModuleSettingsModel`	`team.verification`	The command behind the evidence envelope every delegation reply carries (ADR-0011 C12)	Ships (JSON only, no GPUI page)

Location: crates/chatty-core/src/settings/models/.

virtual_agents is a list of { name, model?, disable_tools?, extra_args? } in module_settings.json; empty means the single local-agent. The team a leader delegates to is a roster plus a skill file, e.g. local-coder on qwen3:4b and local-reviewer on gemma4:26b with disable_tools: ["fs-write", "shell", "git"]. team.verification sits beside it and belongs to the whole roster: the runner runs it in each worker’s worktree when the worker’s task ends and puts its exit code into the evidence envelope the leader reads. Schema and the resolution rules: Agents and specs, “Named virtual agents” and “The evidence envelope”.

Planned (product integration project)

Settings model	Field(s)	Research link	Gate issue	Status
`FlowSettingsModel`	`workflow_enabled`, `active_workflow_id`	M2 AFlow IR	AGE-54	Not implemented
Playbook storage	scope TBD	M4 ACE	AGE-47	Blocked — human product decision
Optimizer launcher	CLI vs GPUI vs CI	M2, M3 offline runs	AGE-44	Blocked
Artifact apply policy	auto / review / export	M2 IR, M3 preamble, M4 deltas	AGE-45	Blocked
Dataset paths	bundled vs user-supplied	Stage A loaders	AGE-46	Blocked

Optimizer artifacts → settings (when gates close)

flowchart LR
  GEPA[GEPA run] --> Preamble[ModelConfig.preamble]
  AFlow[AFlow search] --> IR[Saved WorkflowRepr JSON]
  ACE[ACE offline] --> Bullets[Playbook sections]
  IR --> FlowSettings[FlowSettingsModel]
  Bullets --> PlaybookStore[Playbook scope TBD]
  Preamble --> Apply[Apply policy AGE-45]
  FlowSettings --> Apply
  PlaybookStore --> Apply

Ablation flags (research / CLI, not GPUI v1)

In chatty-optimize — for reproduction, not end-user settings until promoted:

Flag area	Purpose	Module
`SelectBestCandidate` vs Pareto	GEPA ablation	M3
Merge on/off	GEPA+Merge	M3
Strategy enum	ReAct / CoT / CoT-SC / hybrids	M1
DGM archive disable	M5 in agenticloop	—

See crates/chatty-optimize/src/ablation.rs.

Persistence paths

Data	Typical path
Models + preambles	JSON via `models_repository`
General settings	JSON via `general_settings_repository`
Conversations	SQLite
Agent memory / skills	memvid via `memory_service`
Playbook (future)	TBD — AGE-47

Experiment protocol — how the setting-vs-default verdict is reached
App ↔ research bridge — component-level mapping
Module pages
Chatty agentic product integration (Linear)

Keyboard shortcuts