dialogue_turns | SignalWire

Extract the real dialogue from a call log. Drops everything that is machinery rather than speech: entries whose role isn’t in roles, entries carrying tool_calls, and empty content. The system, system-log, tool, and assistant-manual roles are all excluded by default.

drop_echo handles one engine behavior. The chat engine appends its own post-prompt output to call_log as a bare role: assistant entry, indistinguishable by role from real speech. Replayed into another medium, the agent would narrate a summary of itself. It is identifiable only by content, being identical to post_prompt_data.raw, which is what this parameter compares against.

Parameters

call_log

AnyRequired

The log, as delivered in call_log, raw_call_log, or raw_messages. A non-list value yields [].

roles

tuple[str, ...]Defaults to ("user", "assistant")

Roles to keep. Keyword-only. The default is exported as DIALOGUE_ROLES.

drop_echo

str | NoneDefaults to None

Exact content to treat as the summary echo and drop. Keyword-only.

Returns

list[dict[str, str]] — {"role", "content"} pairs in order.

Example

from signalwire.core.post_prompt import dialogue_turns
raw_summary = raw_body.get("post_prompt_data", {}).get("raw")
turns = dialogue_turns(
    raw_body.get("call_log", []),
    drop_echo=raw_summary,
)
for turn in turns:
    print(turn["role"], turn["content"])

Part of the post-prompt normalization module.