models.infer without tools emits no assistant_reply; harness hosts cannot obtain the run's final text

[gregjkal]

Summary

A Lua models.infer(prose) call in a section that advertises no tools returns the model's text to Lua, but the engine emits only model_turn_completed for that turn. No assistant_reply event is emitted. A host that follows session events through harness-api therefore never sees the reply text, and since harness-api exposes no run outcome value either (RunOutcome::Completed { final_text } lives only in the private run log), there is no way for a host to obtain the run's final text.

Observed at master b64c1c9d.

Where

• crates/promptforge-api-runtime/src/execute/tools.rs, accept_infer_completion: reports MODEL_TURN_COMPLETED, optionally request/response debug events, then returns Ok(text) for CompletionResult::Text. Nothing emits assistant_reply.
• Compare crates/promptforge-api-runtime/src/execute/scheduler/chat.rs, Round::text_reply and accept_chat: the tool-advertising chat path emits assistant_reply (and assistant_tool_calls) for the same kind of turn.
• crates/harness/sessions/src/session.rs forwards every engine event, so the gap is in the engine, not the harness.

Reproduction

Prompt:

---
name: echo
description: Reply with the first line of the paper
promptforge: 0
models:
  writer:
    min_context: 32768
    description: any chat model
---

# Echo

```lua
models.default("writer")
```

## Only

Reply with exactly the first line of this text and nothing else:

{{ args }}

```lua
return models.infer(prose)
```

Launch it as an agent session through harness-api against any chat model and follow the session events. Observed sequence for the Only section:

section_started, lua_shared_load_started, lua_shared_load_succeeded,
lua_chunk_started, model_turn_completed, lua_chunk_succeeded,
lua_teardown_started, lua_teardown_succeeded, section_finished

then run_succeeded. No assistant_reply, no thinking.

The text does arrive in Lua. Replacing the last chunk with

local r = models.infer(prose)
error("PROBE: infer returned " .. #r .. " bytes: " .. r:sub(1, 60))

fails the run with PROBE: infer returned 66 bytes: Everyone needs this. Field experience at three companies sho, so the completion was non-empty and the model turn completed normally.

For contrast, the same run's tool-advertising section emits model_turn_completed followed by assistant_tool_calls, as expected.

Expected

One of:

1. accept_infer_completion emits assistant_reply for a text result, the same way Round::text_reply does, so hosts see every model reply regardless of whether tools were advertised.
2. harness-api exposes the run outcome (final_text) to the host after run_succeeded.

Option 1 also keeps the event stream consistent: today a host cannot tell from events whether a model_turn_completed produced text, an empty reply, or a tool batch unless tools were advertised.

Context

Found while porting papergate (wg21-paperflow) to harness-api. papergate reads the run's last assistant_reply event as its report because no outcome value is exposed; with this gap, a prompt whose final step is a tool-less models.infer can never report. The debug response event does carry the raw completion body, which is the only current workaround.