[luca-belli]

Summary

From your review on #289: "this is confusing. we should add a ROLE and the role speaks first." Stacked on #291. It should ship with #289 so configs change only once.

The flag named the speaking order from the chatbot's side (--provider-speaks-first) and the config field from the persona's side (persona_speaks_first), so the code had to flip one to match the other. Now there's one setting, named after who speaks:

	Before	After
CLI	--provider-speaks-first (off by default)	--first-speaker {user,chatbot} (default user)
Config	"persona_speaks_first": true / false	"first_speaker": "user" / "chatbot"

Why this shape

• The values follow the CLI's u/c/j vocabulary (user, chatbot), the same words -u and -c use, rather than the internal Role.PERSONA / Role.PROVIDER names.
• The schema stores a string, not Role. Role lives in llm_clients/llm_interface.py, and utils/ (where GenerationConfig lives) must not import llm_clients/; that's a layer rule. Moving Role into utils/role.py, as architecture.md asks, would let the schema use the enum. That's a separate internal cleanup.
• The generation domain still takes a bool. run_for_user_models converts at that one boundary (first_speaker == "user"). Threading Role through the runner and simulator belongs with docs: design draft — run naming, order of events, and its enforcement #268's internal renames.
• Old configs fail loudly. A config with persona_speaks_first stops with "generation.persona_speaks_first was renamed: use "first_speaker": "user" in place of true, or "chatbot" in place of false", rather than a generic missing-field error.
• It's still a result field, so it's required in config files (feat: let config files leave out output and speed fields #289).

Changes

• utils/config_schema.py: GenerationConfig.first_speaker, validated against FIRST_SPEAKERS = ("user", "chatbot")
• vera_cli/generate.py: the flag, DEFAULTS, RESULT_FIELDS, and the rename error
• configs/*.json, the replay and e2e test configs, cli.md, vera-cli-use-cases.md, architecture.md, and CHANGELOG (under Breaking)
• New tests: the old field gets the rename error, and an unknown value is rejected

Testing

Ruff and pyright pass. I haven't run the tests, per your rule; CI isn't triggered automatically because this PR is stacked.

🤖 Generated with Claude Code

[jgieringer]

LGTM - unsure if the role enum has been removed or if it's usable here

[jgieringer]

do we want to use the Role enums?

[luca-belli]

in your Enum era

[luca-belli]

but it's a good question, and the answer is that now Role uses the old wording so we can do that once we call the thing chatbot consistency everywhere