/x-engage

Thin wrapper over the local x-engage checkout.

Set X_ENGAGE_DIR in your environment (or this skill's caller config) to the absolute path of your local clone. Defaults to ~/Work/x-engage if unset.

Architecture (May 2026 rebuild)

Drafts are produced by corpus retrieval + minimal positive-spec prompt + post-filter lint, not by stacking rules in the prompt. Key components:

• voice-profile.personal.md (~80 lines) — six positive specs + confirmed your voice patterns
• voice-corpus.md — tagged voice-matched reply examples, top-3 retrieved by source-post keyword
• voice-receipts.md — static verifiable user facts (your GitHub stats, your company numbers), top-2 retrieved by keyword
• scripts/lib/safety.py — deterministic post-filter lint (em-dash ban, aphorism patterns, AI-stock phrases, tbh/kinda cadence gate)
• scripts/lib/voice.py — drafter orchestrator + Claude CLI invocation
• SKILL_DISCIPLINE.md — anti-bloat guardrails (read this before adding rules)

Run /x-engage verify after any change to catch bloat or staleness early.

Disambiguation (read before routing)

If the user says vaguely "start the daemon" / "start the bot" / "run x", default to autopilot start since it auto-launches the scan-bg pool feeder too — one command covers the full autonomous mode.

Only route to run-bg (advanced) if the user explicitly says "scan only" / "pool only" / "no publishing" / "don't auto-reply". run-bg does NOT publish — it just feeds the pool for later manual fetch.

autopilot start is a one-way door for the day's account safety. If unsure whether the user wants autonomous publishing vs. manual review, ASK.

Examples

Map user phrasings to subcommands:

• "/x-engage" or "draft some x replies" or "scan x" → fetch
• "draft 30 replies" or "fetch 25" → fetch 30 / fetch 25 (count arg)
• "show me the x queue" or "review x drafts" → review
• "approve 1, 3, 5" or "approve all" → approve 1,3,5 / approve all
• "redraft 2 shorter, drop the question" → redraft 2: shorter, drop the question
• "kill #4" or "drop draft 4" → kill 4
• "save #5 as a good one" or "mark 5 good" → good 5
• "ship it" or "publish approved" → publish
• "x status" or "how many replies left today" or "how is autopilot doing" or "is the daemon running" → status (unified — covers queue, scan-bg, autopilot)
• "verify x-engage" or "is the skill healthy" or "check for bloat" → verify
• "start autopilot" or "run x autonomously" or "autopilot start" → autopilot start (auto-starts scan-bg)
• "stop autopilot" or "kill autopilot" → autopilot stop
• "start autopilot target=30 until=17:00" → autopilot start target=30 until=17:00
• "start autopilot don't sleep" / "start autopilot stay awake" → autopilot start --keep-awake
• "show me today's replies" / "list published" / "what did autopilot send" → autopilot list
• "show me the last 50 replies" → autopilot list 50
• (advanced) "start scan-bg only" / "pool only no publish" → run-bg
• (advanced) "stop scan-bg" → stop-bg

Subcommands

Parse the first word of the user's input as the subcommand, then run the matching command via Bash. All commands start with cd "$X_ENGAGE_DIR" && python3 -m scripts.x_engage — shown abbreviated as … below.

Stream output to the user. Use long timeout (600000ms) for fetch and publish.

Review-phase rendering

When review returns rows, render them in this exact format so the user can act fast:

#<id>  @<author> (<follower_count> followers) · <age_min>min ago · score <score>
  Source: "<first 140 chars of source post>"
  Draft:  "<draft text>"

After listing, prompt:

Reply with: approve <ids|all>, redraft <id>: <feedback>, kill <id>, good <id>, or publish

After completion

• fetch: pulled / drafted / skipped counts.
• review: show drafts as above; do not summarize.
• approve / redraft / kill / good: confirm action in 1 line ("Approved #1, #3", "Redrafting #2…", "Killed #4", "Saved #5 as vibe reference")
• publish: published / failed / deferred counts. Surface any safety signals.
• status: phase, today's published count, paused state, queue counts.
• verify: relay the report verbatim. If warnings, suggest concrete next action.

When user flags a draft as "not my voice"

The protocol (full details in SKILL_DISCIPLINE.md):

1. Identify the failure — which specific phrase / shape / register collision is wrong?
2. Add to BANNED_ANYWHERE (or BANNED_OPENERS if position-specific) in scripts/lib/safety.py. Comment with date + draft ID.
3. Test the lint — verify the failed draft rejects + a clean voice-matched variant still passes.
4. Kill the draft in the live queue.
5. Run /x-engage verify to confirm bloat is in check.

Do NOT:

• Add a "don't do this" instruction to the prompt or voice-profile.personal.md (Pink Elephant — re-anchors the model on the forbidden pattern).
• Add a new template, register quota, or shape steer.
• Lower the post-filter threshold to "give the model more room."

When user flags a draft as "yes, this is me"

1. Run /x-engage good <id> — pins into good-drafts.md.
2. If the draft demonstrates a NEW pattern (new register, new POV, new receipt shape), also append manually to voice-corpus.md so future retrieval picks it up.

References

The references that the drafter actually reads at runtime:

• voice-corpus.md — tagged voice-matched replies. Top-3 retrieved per draft.
• voice-receipts.md — verifiable static facts. Top-2 retrieved per draft.
• voice-profile.personal.md — six positive specs + anti-pattern reminders.
• references/guardrails.md — human-readable rate-limit doc (operational caps, cooldown reasoning). The executable rules are in scripts/lib/safety.py + state.py.

Anti-bloat: references/_archive/ holds legacy files from the pre-rebuild architecture (T1-T7 templates etc.). Not loaded. Don't restore.

Autopilot mode

/x-engage autopilot start boots an autonomous loop that drafts AND publishes replies without per-draft approval. Designed for fire-and-forget engagement velocity.

How it works:

• launchd plist com.x-engage.autopilot fires autopilot-tick every 60s
• Each tick: check halt conditions → pull 1 fresh candidate from pool → draft → lint+score (same bar as manual: ≥0.45) → insert as approved → publish via Playwright
• Tick is idempotent — safe to crash and resume
• Requires scan-bg daemon running in parallel to keep the pool fresh. Autopilot warns if it's not. Start it with /x-engage run-bg.
• Does NOT auto-publish pre-existing pending drafts. The pending queue from manual mode is untouched — autopilot only ships drafts it created in the current tick. To clear a stale manual queue, use /x-engage review + kill or approve + publish as normal.

Stop conditions (any triggers self-unload):

1. Daily target hit (default 50, panic ceiling 50)
2. stop_at time reached (default 18:00 local, uses tz from settings.yml)
3. ~/.x-engage/PAUSED flag exists (any safety signal writes this)
4. X_ENGAGE_HALT=1 env var

What it does NOT bypass:

• 90s min gap between publishes
• Per-handle 24h cooldown
• 4 replies / 30d lifetime cap per author
• Voice-match threshold (0.45)
• Safety lint
• CAPTCHA / ACCOUNT_PAUSED / RESTRICTION → writes PAUSED, auto-unloads daemon, exit 2

Args:

• target=N — override daily target (1 ≤ N ≤ 50). Panic ceiling refuses higher.
• until=HH:MM — override stop time (local TZ from settings.yml).

Manual /x-engage publish is unaffected — it still requires require_explicit_approval=true.

Resilience contract (what survives what)

How to verify it's actually alive (not just "installed"): run /x-engage status — the heartbeat: Xs ago [ALIVE] line proves the tick fired recently. Anything over 120s = STALE = something's wrong; over 600s = DEAD = check tail logs/autopilot.err.

Critical: safety + auth signals

If output contains ACCOUNT_PAUSED, RESTRICTION, CAPTCHA, LOGIN_REQUIRED, or exit code 2 (from publish):

• DO NOT retry. Halt.
• Tell the user: "X account safety check failed. Open the screenshot at ~/Downloads/x-incident-*.png. After confirming the account is healthy, delete ~/.x-engage/PAUSED to resume."
• Do not run any further X-touching commands until the user explicitly resumes.

If output contains COOKIES_EXPIRED (from fetch, exit code 2):

• DO NOT retry. Halt.
• Tell the user: "Your X session cookies expired. To fix:
  1. Open x.com in Chrome, log out, log back in.
  2. DevTools (Cmd+Opt+I) → Application → Cookies → https://x.com
  3. Copy auth_token and ct0 values into .env (replace old AUTH_TOKEN= and CT0= lines).
  4. Delete ~/.x-engage/PAUSED to resume. Then run /x-engage setup to verify, and re-run fetch."
• Do not run any further X-touching commands until the user has updated cookies and removed the PAUSED file.

Unknown args

If arg is not in the table above, print:

Usage: /x-engage [fetch|review|approve|redraft|kill|good|publish|status|setup|verify|autopilot]
  fetch [N]      — fetch candidates + draft + queue (default 15)
  review         — show pending drafts in chat
  approve <ids>  — mark drafts approved (e.g. "approve 1, 3" or "approve all")
  redraft <id>   — re-draft one with feedback (e.g. "redraft 2: shorter")
  kill <id>      — reject a draft
  good <id>      — save a draft as a vibe reference for future drafting
  publish        — ship approved drafts via Playwright
  status         — unified snapshot (queue + scan-bg + autopilot + flags)
  setup          — first-time setup
  verify         — skill health check (line counts, bloat, staleness)
  autopilot start [target=N] [until=HH:MM]
                 — autonomous publish daemon (bypasses approval, auto-starts scan-bg)
  autopilot stop — stop autonomous daemon

Advanced (rarely needed — autopilot manages scan-bg automatically):
  run-bg         — install scan-bg pool feeder only (no publishing)
  stop-bg        — unload scan-bg daemon