| name | virtual-scott-poll |
| description | Scan the sandbox WordPress for comments that virtual-scott should respond to and report a short list. Does NOT post replies — that's virtual-scott-reply's job. Filters out comments by virtual-scott itself, by the blog owner scott, by anonymous authors, comments that already have a virtual-scott reply, and obvious filler. |
| user-invocable | true |
Find comments worth replying to
Produce a short prioritised list of WordPress comments that Virtual
Scott should respond to. This skill does not post anything — it
only discovers and filters. After running, the human (or a scheduled
runner) decides whether to invoke virtual-scott-reply for each id
in the list.
Why split off the reply step: actual reply generation is expensive
(cold OVMS path ≥ 30 s per reply) and reply correctness is graded
post-hoc. Doing discovery + reply in a single agent run risks looping,
high iGPU usage, and shipping many marginal replies on the same
invocation. Discover-then-confirm-then-reply is safer.
Step 1: GET all approved comments
Make this tool call:
- Tool:
web_fetch
- Method: GET
- URL:
http://__WP_HOST__:30080/wp-json/wp/v2/comments?status=approve&per_page=100&order=desc&orderby=date
From the response capture the full array. For each comment, extract:
id, post, parent, author (numeric user_id), author_name,
content.rendered (call this BODY), date.
Step 2: Filter the candidate set
Drop any comment matching ANY of these rules. The remaining set
becomes the candidates.
-
Self (authenticated). author == 2 (virtual-scott user_id).
Don't reply to yourself.
-
Self (anonymous artifact). author_name == "Virtual Scott"
regardless of author. Catches manually-created anon comments
left over from debugging (author=0 + name=Virtual Scott).
-
Blog owner. author == 1 (scott admin user_id). Don't reply
to the blog owner's own comments.
-
Anonymous default. author == 0 AND author_name == "A WordPress Commenter". WP's default seed comment; skip.
-
Already replied. Some other comment in the full array has
parent == comment.id AND its author == 2. (i.e., virtual-scott
has already posted a reply in this comment's thread.)
-
Pure filler. BODY (stripped of HTML tags) matches any of:
- Under 25 characters total
- Matches one of these patterns case-insensitively (regex-style):
^thanks?\!?\.?$, ^nice (post|article|read),
^(great|good|cool|awesome|amazing) (post|article|stuff|read),
^first\!?$, only emoji/punctuation.
Filler comments may be technically polite but the persona rule is
"half a useful answer beats a confident wrong one" — there's
nothing meaningful to answer, so don't.
Step 3: Report the candidates
For each surviving comment, print one row:
{id} post:{post_id} by:{author_name} "{body, truncated to 80 chars}"
Sort by id ascending (oldest first — fairest to commenters who've
been waiting longest).
If the list is empty, print exactly: No comments need a reply.
Cap the report at 10 candidates. If the unfiltered candidates
exceed 10, mention (N additional candidates not listed).
Step 4: Suggest next-step invocations
For each candidate in the report, print a copy-pasteable line for
the human:
To reply: openclaw agent --agent virtual-scott --message 'Read /home/smbaker/.openclaw/workspace-virtual-scott/skills/virtual-scott-reply/SKILL.md and execute its 4 steps for COMMENT_ID={id}.'
The cron-fired virtual-scott-tick skill automates discover-and-
reply for the autonomous loop. This poll skill is the read-only
manual-inspection version: a human runs it to see what would be
replied to before letting the autonomous loop process it.
What this skill deliberately does NOT do
- Does not reply. Discovery only.
- Does not retry on errors. If the GET in step 1 returns
non-2xx, report the status and stop.
- Does not write to WP. Read-only against the REST API. No auth
header needed for step 1 (public read).
- Does not invoke other skills. Agent loops nesting skills are
unreliable on Qwen3-Coder; we keep skills independent.
