fix: prevent raw JSON dump in respond-to-comment PR replies

[warp-dev-github-integration]

Problem

When the Oz agent completes work via tool calls (e.g., editing files) without producing a final text message (type='agent'), the "Reply to Comment" step in the respond-to-comment workflow falls back to posting the entire raw NDJSON stream as the PR comment body. This results in comments like:

@user: {"type":"system","event_type":"run_started","run_id":"...","run_url":"..."}
{"type":"agent_reasoning","text":"..."}
{"type":"tool_call","tool":"edit_files",...}
...

See: https://github.com/warpdotdev/warp-server/pull/10070#discussion_r3029369447

Root Cause

The JSONL parsing logic in the "Reply to Comment" step only looks for events with type === 'agent' (the agent's direct text output). When no such event exists — which happens when the agent completes its work purely through tool calls and reasoning — lastText remains empty and the code falls back to raw:

const content = (lastText || raw || '').trim();

Fix

Updated the JSONL parsing in examples/respond-to-comment.yml (and regenerated .github/workflows/respond-to-comment.yml) to:

1. Extract agent_reasoning text as fallback — when there's no type: 'agent' event, the last agent_reasoning text is used instead
2. Extract run URL — system events with run_url are captured and appended as a "View full agent run" link
3. Never fall back to raw NDJSON — the isJsonl flag tracks whether the output was JSONL format; if so, only structured content is used
4. Still handle text format — non-JSONL output (e.g. output_format: text) is used as-is

Fixes REMOTE-1224

Conversation: https://staging.warp.dev/conversation/bf17e8da-31fa-4af4-b487-5d11a0b8df35
Run: https://oz.staging.warp.dev/runs/019d4f93-ca88-7dbd-9bd1-3d6b87ff75ad

This PR was generated with Oz.