Flow Making

Flow-Making

Related (2026-07-22): FlowCreator is now ALSO a plain wrapped chat-agent, chat_agent_flowcreator — a Multi-Turn tool (no ACPX needed) that takes prompt= + flow_filename= and writes the .flw itself. This skill is the heavier, ACPX/invoke_skill-surface path that shells out to scripts/make_flow.py; the wrapped tool is the lighter one-call path. Both convert flow_result.json → .flw with the SAME result_to_flw.py logic — this skill's scripts/result_to_flw.py and the copy vendored at agent/agents/flowcreator/result_to_flw.py (so the pool subprocess can carry it) must be kept in sync. For a simple "make me a .flw from this sentence" ask, prefer chat_agent_flowcreator; use this skill when you need the scripted pipeline.

Produce a canvas-loadable .flw from ${input.objective} by wrapping the FlowCreator engine — which already encodes the full 89-agent catalog, connection contracts, and design rules in agentic_skill.md. Do NOT hand-author the .flw JSON yourself: you do not carry the agent catalog/config-key contracts in context, so a hand-written flow hallucinates agent types and will not load.

Primary path — one deterministic call

Run the shipped driver with execute_command. It copies the FlowCreator template to an isolated runtime dir, writes its config.yaml, runs FlowCreator (which queries the configured Ollama model), and converts the result to a .flw:

python agent/skills_pkg/flow_making/scripts/make_flow.py \
  --objective "${input.objective}" \
  --out "${input.out_path}" \
  --flow-name "${input.flow_name}" \
  --model "${input.llm_model}" \
  --host "${input.llm_host}"

• The path is relative to the chat process's working directory (the Tlamatini app root, where manage.py runs — same cwd execute_command uses for python manage.py ...). If a relative call ever reports "can't open file", retry with the repo-root prefix Tlamatini/agent/skills_pkg/....
• Omit --flow-name / --model / --host when the corresponding input is empty (the driver has sensible defaults).
• Use forward slashes in --out even on Windows (e.g. C:/Users/you/Desktop/Flows/alarm.flw); the driver creates the folder if it is missing.

Reading the result

On success the last stdout line is machine-readable:

agent_count=<N> connection_count=<M> flw_path=<absolute path>

Parse it and return { flw_path, agent_count, connection_count }. Tell the user to open it on the ACP designer via Open ▸ select the .flw (it auto-deploys the agents and draws the connections).

Failure handling

If the driver exits non-zero, its last stdout line begins with ERROR . Surface that message verbatim and act on the common causes — do NOT silently fabricate a .flw:

• ERROR FlowCreator timed out / Cannot reach Ollama — Ollama is not running or the model is not pulled. Ask the user to start Ollama / pull the model, or pass a different --model.
• ERROR FlowCreator: ... — the model returned an unparseable flow; retry once, then report.
• ERROR could not locate the FlowCreator template dir — pass `--template

Fallback — manual two-step (only if the driver is unavailable)

1. Use chat_agent_file_creator to write a config.yaml (JSON is valid YAML) into a copy of agent/agents/flowcreator/ with { "prompt": "${input.objective}", "flow_filename": "...", "llm": {"host": "...", "model": "..."} }, then execute_command: python flowcreator.py in that copy.
2. Convert its flow_result.json to the final .flw: python agent/skills_pkg/flow_making/scripts/result_to_flw.py <runtime>/flow_result.json "${input.out_path}" and read the same agent_count=… connection_count=… flw_path=… summary line.

Notes

• The .flw schema this produces is the schemaVersion-2 nodes/connections shape the canvas loader consumes; see references/flw_schema.md.
• Never mutate the FlowCreator template in place — always run from an isolated copy (the driver does this for you).

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.