Tlamatini Flow From Objective

Flow from objective

Produce a canvas-loadable .flw for the user's stated objective.

Superseded by the flow-making skill. Prefer flow-making: it drives the FlowCreator engine (full 89-agent catalog + connection contracts) and emits a validated, schemaVersion-2 .flw. This skill is kept as an alias/entry point — do NOT hand-author the .flw JSON, because you do not carry the agent catalog in context and a hand-written flow hallucinates agent types and will not load.

Procedure (delegate)

1. Invoke the flow-making skill with the same inputs: invoke_skill('flow-making', { "objective": "${input.objective}", "out_path": "${input.out_path}" }).
2. Return its result verbatim: { flw_path, agent_count, connection_count }.

If you must run it directly

Use the shipped driver — it copies the FlowCreator template to an isolated dir, runs it, and writes the .flw:

python Tlamatini/agent/skills_pkg/flow_making/scripts/make_flow.py \
  --objective "${input.objective}" --out "${input.out_path}"

The last stdout line is agent_count=<N> connection_count=<M> flw_path=<path>.

Correct .flw shape (schemaVersion 2)

If you ever emit .flw JSON by hand, it MUST match the loader contract (acp-file-io.js::loadDiagram / flow_spec.py) — NOT a {version, agents, connections:[{from,to,kind}]} shape (that is obsolete and will not load):

{
  "schemaVersion": 2,
  "nodes": [
    {"id": "starter-1", "text": "Starter", "left": "50px", "top": "50px",
     "agentPurpose": "", "configData": {"target_agents": ["monitor_log_1"]}}
  ],
  "connections": [
    {"sourceIndex": 0, "targetIndex": 1, "inputSlot": 0, "outputSlot": 0}
  ],
  "artifacts": {}
}

See agent/skills_pkg/flow_making/references/flw_schema.md for the full contract.

Current installed-agent contract — 2026-09-15

Use agent/agents/flowcreator/flow_catalog.json for canonical names, current config schemas, output/input slots, lifecycle flags and structured fields for all 89 installed types. GUI-Manager is design only. After changing a template/contract/reference, run python scripts/update_flow_catalog.py and its --check mode in the repository. Deployment refreshes runtime snapshots.

FlowCreator selects capabilities before detailed design, validates the generated plan, and uses bounded repair. Declare Ender input connections explicitly; Ender target_agents is a kill list. Counter uses L/G slots; source dependencies do not choose a conditional output branch. Generated Parametrizers require valid _parametrizer_mappings, one source and one target. Do not maintain a separate hardcoded Parametrizer producer list.

For desktop flows, use explicit physical/screenshot geometry and verified target windows. input_sent is input delivery only; errors may be partial and must not be blindly replayed. Read docs/desktop-input-and-flow-contracts.md and docs/agent-coverage.md for the full contract and verification scope.

Model choices in generated flows

Preserve quoted "@config" and missing registered model fields so generated agents follow Config → Models. Keep a literal model/engine/voice only when an explicit override is intended; never fill inheritance with a guessed tag. The registry in agent/agents/model_settings.py maps all 21 model-backed agents to 38 global settings. Wrapped-chat globals are seeded before explicit tool arguments, so do not manufacture model arguments when translating a request into a flow. Optional empty Whisperer cloud model and LaTeXer repair model values have distinct meanings. Video analysis_type remains a per-agent task choice; its local audio model is separate from Whisperer's engine. See docs/model_configuration.md and the current generated flow catalog for exact field paths and defaults.