fix(goals): fire goal judge after streamed turns (Ralph loop stuck at turns_used=0)

[Shinigallo]

Problem

When a /goal is set on a platform with streaming enabled (the default on Telegram), the Ralph loop never advances: the goal stays active at turns_used=0, the goal_judge model is never called, and no [Continuing toward your standing goal] continuation is ever enqueued. The agent does the work for the first turn, but the loop silently stalls.

This affects any streamed /goal, not an edge case.

Root cause

The goal-continuation hook in _handle_message reads the turn's final_response from the return value of _handle_message_with_agent:

if _final_text.strip():
    ...
    await self._post_turn_goal_continuation(..., final_response=_final_text)

When the reply is streamed, _handle_message_with_agent returns None (the text was already delivered out-of-band via the stream consumer; the dict it would otherwise return is response, which ends up None on that path). So _final_text is empty, the if _final_text.strip(): gate is false, and _post_turn_goal_continuation — the only place that calls the judge and increments turns_used — is never reached.

Fix

gateway/run.py:

1. Always invoke the goal-continuation hook when a session resolves, instead of gating on non-empty final_response. It is a cheap no-op when no goal is active (_post_turn_goal_continuation returns early on not mgr.is_active()).
2. In _post_turn_goal_continuation, when final_response is empty, recover the turn's last assistant message from the persisted transcript (SessionDB.get_messages) via a new _recover_last_assistant_text() helper before judging. Genuinely empty turns (interrupted / errored) are still skipped so the judge isn't asked to evaluate nothing.

No behavior change for non-streamed turns or for non-goal messages.

Testing

Verified end-to-end over Telegram with a multi-turn /goal:

• turns_used advances (was stuck at 0, now 0 → 4+)
• the goal_judge model is loaded/called on every turn
• [Continuing toward your standing goal] continuations are enqueued between turns
• judge verdicts are correct (continue while incomplete, with sensible reasons)
• transcript recovery returns the exact streamed final response (confirmed against the logged response_len)

Non-goal messages and non-streamed goals are unaffected.

🤖 Generated with Claude Code

[tonydwb]

Fixes a real bug where streamed /goal loops stalled at turns_used=0 (68 additions). Adds transcript recovery to read the actual assistant response when final_response is empty. Well-commented with clear explanation of the failure mode. Clean fix.

Reviewed by Hermes Agent

[Victornovikov]

Confirmed on v0.18.0 over Telegram. Goal state from state_meta:

{"turns_used": 0, "last_verdict": null, "last_turn_at": 0.0}

Judge never fires. Streaming returns None from _handle_message_with_agent, _final_text is empty, the strip() gate blocks the continuation hook entirely. First turn runs, then dead silence.

The manual "send continue" workaround is not a fix — it defeats the entire point of an autonomous loop. /goal is broken on every streaming gateway. Would be good to get this merged.