X auto-reply

X auto-reply

Kazma can draft a reply when someone mentions the connected account on X. It is off by default. Credentials, ToU caps, and the audit log are the same as X publisher.

A reply is always one subject + one side. Emoji is tone when a Settings card matches, and side when none does.

How a summon is decided

Walk this list. The first hit wins.

1. Keyword on a Settings card — whole word for ASCII, containment for other scripts. Cards are tried top to bottom. Put specific topics above broad ones.
2. Opinion ask — شرايك / what do you think / a ? / ؟ / how-what-why / both 👍 and 👎 (or ❤️ and 😂) in the same mention. Reacts to this post in voice. Does not pick a country card.
3. Summon side — one polarity in the mention:
   • 😂 🤣 💀 🔥 🤬 😡 🙄 👎 or against / roast / ضد / هاجم → criticise this post
   • ❤️ 👏 💯 🙏 👍 or support / defend / دافع / معاه → defend this post
4. Stay silent (default once any card exists). A * keyword still answers everything. “Reply anyway” is voice with no side.

A Settings card only applies when one of its keywords is in the post. A classifier cannot pin an unrelated tweet on a card you wrote for something else.

When a card matches, its side (against / support) is locked. A roast emoji cannot flip a support card. Angry on a support card is anger at its critics, never at the subject.

What counts as a summon

Mention	What Kazma reads
Standalone @handle 😂	The mention itself
Reply under someone else’s post	That parent post
Quote of someone else (even under Kazma’s own tweet)	The quoted post
هذا + x.com/…/status/… or t.co/… in a reply to Kazma	The linked post
Bare reply under Kazma with no quote and no link	You (trusted): walks up to the original post and drafts. Strangers: ignored unless #Open

X only lets this account post a reply on a tweet that mentions it or that it wrote. The public reply therefore sits under your mention, not under the original. The text it drafts against is still the original when a quote or link is present.

Paid X plan required for the mentions poller (GET /2/users/:id/mentions). Free tier 403s; the poller logs that and backs off.

Two ways to fire it

Paste (every X plan). From chat:

/x roast https://x.com/someone/status/1234567890 | Their post text here

Always draft mode. Approve with /x approve <id>. Gateway login is the authorization — you do not need to be in summoners.

Mentions poller (paid plan). Kazma polls its mentions. Save on Settings starts or stops the poller (no restart). X Studio → Conversations → Refresh is the same poll, and it ignores a stuck cursor. /x poll too.

Subjects

Settings → Integrations → X → Auto-reply.

Each card:

Field	Role
Id	Short name (topic)
Side	against (always criticise) or support (always defend)
Keywords	When this card fires. ASCII = whole word. * = catch-all (tried last)
Fallback mood	Used if the mention has no emoji
Optional extra	Colour (who / why) — not a second side
Hard lines	Extra “never say”. Universal lines always apply
Examples	Replies you actually wrote — strongest voice lever

Empty allowlist + allowlist policy = nobody can summon. anyone still has caps, follower floor, screen, and stance check.

Open-thread marker (optional, e.g. #Open): put it in your mention. Without it, only trusted handles get a reply. With it, anyone who mentions Kazma on that post can get a reply. Hashtags match as a whole tag (#Open does not match #OpenAI).

Close-thread marker (optional, e.g. #Close): mention @KazmaAI #Close in that conversation to stop strangers. You can still talk. The close is stored on the X conversation id so later polls stay closed.

Trusted handles (you) can keep talking without extra emoji — thread cooldown and per-target caps do not apply to you, so a general question can be a back-and-forth.

List order is priority: first card whose keyword appears wins.

Modes

Mode	Behaviour
draft	Compose, hold, notify. You post with Approve (Conversations or /x approve)
auto	Posts after rails + screen + stance check. Text you have not read, under your name
off	Master off

/x roast is always draft.

Stance check

The length/threats screen does not know your side. After drafting, a short closed-set call asks argues / contradicts / fence (any language). Sympathy for the other side, “don’t attack them”, or a summary with no side is contradicts or fence and is blocked.

Catch-all * and unmatched voice skip this check (there is no claim to drift from). Content screen and hard lines still run.

auto + a check that cannot run → block. draft + unusable check → allow (you read it).

Knowledge Base (optional)

Off by default. When on, each draft searches the Knowledge Library (hybrid lexical + semantic) and injects up to three fenced snippets as facts. If they conflict with the card’s side, the side wins. Empty or failed lookup does not skip the summon — it drafts without notes.

Opinion-ask / voice-only summons search the post text, not a synthetic id (post / voice). A named Settings card (Kuwait, …) may prefix the query so a short tweet still hits.

1. Ingest a library on Knowledge (/knowledge) until it has chunks.
2. Settings → X → Auto-reply → Ground drafts in the Knowledge Base.
3. Pick one library, or leave All libraries with chunks.
4. Try it with a real post. The preview says how many snippets landed (KB: 2 snippets from … or KB on, no snippets matched).
5. Approve from Conversations only if the draft used those facts the way you want.

Conversations

X Studio → Conversations is the log of summons: parent text, mention, draft, skip reason.

Button	Does
Refresh	Polls X (not a page reload)
Approve	Posts the stored draft, as a reply to the mention
Deny	Parks a held draft
Retry	Re-runs skipped / failed / stuck drafting against current config (always holds for approval)
Delete	Posted: delete on X and drop the row. Else: drop the row

Sort is newest on X first (tweet id), not last touch. A Retry of an old card does not jump it to the top.

Guardrails

On top of ToU policy (length, mentions, duplicates, daily/monthly caps):

• Universal hard lines (no slurs, no threats, criticise institutions not peoples)
• Post-generation screen on what the model wrote
• Daily / per-target / per-thread caps
• Follower floor (default 500) — 0 disables it knowingly
• Reply is posted under the mention (X rule)

Turning it on

1. Settings → X: four OAuth 1.0a values, Test, posting-ready.
2. Auto-reply: trusted handles (or anyone), mode draft, Save (poller starts).
3. Add cards only for topics you want a locked side on. One-offs use the mention emoji/word.
4. Try it (dry run) until the voice is right.
5. Real mention from another account, then Refresh or /x poll. Approve from Conversations.

Commands

/x roast <url> | <text>    draft a reply
/x approve <summon_id>     publish a held draft
/x deny <summon_id>        discard a held draft
/x retry <summon_id>       re-run a skipped/failed summon
/x delete <summon_id>      delete the posted reply on X (or drop the log row)
/x list                    recent summons
/x subjects                declared cards
/x poll                    one mentions poll (paid plan; ignores stuck cursor)

Kill switches

KAZMA_X_REPLY=0 stops auto-reply; scheduled posts still work.
KAZMA_X_POST=0 stops every X write.

Config keys

connectors.x.reply.* in ConfigStore (Settings panel writes these).

key	default	meaning
enabled	false	master switch
mode	off	off | draft | auto
summoner_policy	allowlist	allowlist | anyone
summoners	empty	trusted handles; empty allowlist = nobody
allow_emoji_mood	true	mention emoji may set tone (and unmatched side)
stance_check	true	draft must argue the side
unmatched	skip	skip | voice when nothing matches and the mention sets no side
classify_llm	false	guess a card when keywords miss (can misfire; leave off)
use_knowledge	false	pull fenced KB snippets
knowledge_library	empty	library id, or all
trigger	empty	optional phrase in the mention
max_replies_per_day	5	across targets
max_replies_per_target_per_day	1	per account
cooldown_per_thread_s	3600	one reply per thread per hour
min_target_followers	500	floor; 0 off
poll_interval_s	600	60s floor
subjects	[]	cards (id, side, match, optional extra)

Limitations

• Mentions poller needs a paid X plan. Reads are metered per month.
• Replies fire only while Kazma is running. A mention during downtime is picked up on the next poll if it is still in the mentions window.
• Drafting is outside the supervisor graph: no tools, no chat memory, unless you turn on KB grounding.
• The public reply cannot be attached to a tweet that does not mention this account (X API).
• drafting with no buttons meant a stuck model call — Retry is shown on that status; /x retry <id> works even before a reload.

Testing it safely

1. Try it — nothing leaves the machine.
2. /x roast — holds for approve.
3. Real mention from another account, then Refresh. Score Conversations, not the model’s table.