Reply Handler

Drafts replies to comments on the user's own (or anyone else's) LinkedIn post — either one comment at a time, or a full-thread sweep. Same lightweight, mostly ephemeral shape as Comment Drafter (/draft-comment): not part of the Drafts/ approval pipeline. Output is copy-paste text in the response; there is no Draft Note, no vault schema, no file write.

Reading third-party post/comment content: see REQUIREMENTS.md §27 — pasted text is the primary, reliable input; a post URL is optional metadata only, and a WebFetch attempt on it is best-effort convenience, never ground truth.

Modes

• Single-reply — the user pastes one comment (optionally plus the original post's text/topic for context) and wants one reply drafted.
• Sweep — the user pastes a full thread export (all top-level comments and their replies, as copied from the post) and wants a reply drafted for every comment/reply worth answering, in one batch. A post URL may come along as best-effort/reference-only metadata per §27 — never relied on to reconstruct thread content that wasn't actually pasted.

If it's ambiguous which mode applies (e.g. a short paste with no visible reply structure), ask, or default to single-reply for one comment / sweep for anything with multiple comments.

Flattening-handling design (sweep mode)

LinkedIn's UI only nests one level deep: a reply to a reply gets flattened into the same top-level-reply list, with LinkedIn auto-inserting an "@Name" prefix on the reply's text to show who it was actually meant for — it is not real nesting, and the export the user pastes will preserve that flattened, one-level shape.

On parse, build a flat list per top-level comment:

{author, text, replied_to}

• replied_to is inferred from an explicit "@Name" at the start of a reply's text (LinkedIn's own flattening marker).
• If no such @mention is present, replied_to defaults to the top-level comment's author (the reply is assumed to address the thread starter, not some other repliers further down).
• Never guess a replied_to target beyond these two rules — if the text is ambiguous even with the @mention convention, say so rather than picking a guess silently.

When drafting, label every reply output as:

Reply to [Author] (responding to [X] in this thread)

and have the drafted reply text itself open by naming or otherwise clearly addressing that specific person's point — never a generic reply that could float under any comment in the thread. If two different people replied to two different targets within the same top-level thread, keep their drafted outputs visually grouped under that thread for one scan-through, but each still carries its own explicit target label so nothing gets misattributed when the user copies them back individually.

Low-value filter (sweep mode only)

Single-reply mode always drafts what's asked — the filter only runs in sweep mode, to avoid wasting a drafted reply on a comment not worth answering. Skip, and log as skipped with a reason, anything that is:

• Pure emoji/reaction-only — "🔥🔥", "👏", no words.
• Generic praise with zero specific reference to post content — "Great post!", "So true", "Well said" and nothing else.
• Off-topic self-promotion/spam — an unrelated link drop, "check out my profile/product".
• A near-duplicate of a comment already answered earlier in the same thread — same point, no new angle.
• A tag/mention with no added content — "@friend check this out" and nothing else.

Keep everything else — including a short-but-substantive disagreement, a real question, a correction, or a personal anecdote tied to the post, even if brief. When in doubt, keep it rather than skip it.

Always report which comments were skipped and why. Never silently drop a comment from the output — the skipped-comment log is a required part of every sweep report.

Process

Single-reply mode

1. Take the pasted comment, plus the original post's text/topic if given, per §27 (pasted text primary; a URL alone is not reliable).
2. Read Content-Learnings/voice-guide.md in full, fresh — never rely on a cached memory of it.
3. Draft one reply: short, specific to what the comment actually said, in the user's own voice, no throat-clearing, no generic filler.
4. Output as copy-ready text.

Sweep mode

1. Ingest the pasted full thread export.
2. Parse into top-level comments plus attributed replies per the flattening design above — build the {author, text, replied_to} list for each top-level thread.
3. Apply the low-value filter to every comment and reply, logging what's skipped and why.
4. Read Content-Learnings/voice-guide.md in full, fresh.
5. Draft a reply for each surviving comment/reply, each carrying its explicit "Reply to [Author] (responding to [X])" addressee label.
6. Present the full batch grouped by original top-level thread, so the user can scan one thread at a time.
7. Report counts: total parsed / filtered out / drafted.

Report back

• The drafted repl(ies) as copy-ready text.
• In sweep mode: each reply labeled with who it's addressed to, grouped by top-level thread; the skipped-comment log with reasons; and the parsed/filtered/drafted counts.

Hard rules

• Never auto-post a reply, and never imply one was posted — there is no comment-posting API anywhere in this repo, same limitation as /draft-comment (REQUIREMENTS.md §27, §33). Output is always copy-paste text for manual posting.
• Never fabricate a comment's content, author, or thread structure beyond what was actually pasted. If the pasted thread is ambiguous or incomplete, say so rather than inventing the missing shape.
• Never invent context about the original post beyond what the user actually gave.
• Every drafted reply follows Content-Learnings/voice-guide.md's tone — conversational, direct, no AI-writing tells.
• Never silently drop a low-value comment from the sweep report — always state that it was skipped and why.
• No vault artifact: no Draft Note, no Drafts/ entry, no approval- pipeline lifecycle, for either mode.