Aigarth

AI Power Grid's Discord participant, running on the Grid's qwen3-27b.

Human message -> current conversation -> one Qwen agent -> silence, lookup or reply.

There is no separate engagement judge, regex audience classifier or reply editor. Every eligible human event reaches the agent through a serialized channel queue. It sees who said what, reply targets, recent messages (including its own), a rolling summary and a short operating brief. A mention is context, not an obligation to talk.

The agent calls finish_turn to choose silence or an exact public reply. Ordinary assistant text remains private. It can look something up and still choose silence. Before finishing, changed channel context is returned to that same agent so it can reconsider an answer that someone else has already given.

Tools

• Public Grid network/model status and enabled validator capabilities.
• Public validator health by ID, plus official releases including previews.
• Curated docs read/search and additional history in the current Discord channel.
• Image generation/remixing, crypto data, web search and optional vision.
• User-controlled safe memory and reminders.
• Human-approved moderation proposals, never autonomous AI bans or slashing.

Operational claims need a current source. Release availability does not prove a Core feature is enabled. Failed lookups mean unknown, not zero or user error.

Run

Requires Node 22.19+ and a Grid service key with appropriate bounded spending.

cp .env.template .env
npm ci
npm run build
npm start

Set DISCORD_TOKEN, GRID_API_KEY, and channel IDs in .env. GRID_CHAT_MODEL=qwen3-27b; GRID_SUMMARY_MODEL defaults to the chat model and is used for background context maintenance, not participation decisions. GRID_VISION_MODEL optionally enables image understanding.

No separate retrieval service is needed for curated markdown docs. Update docs/operating-brief.md with durable operational changes, never live counts or secret configuration. It is loaded each turn and treated as historical orientation. It carries a review date and is marked stale after 48 hours; live lookups remain necessary even when the brief is fresh.

Privacy controls: !memory, !memory on|off, !forget <phrase|all>. Permissions, human voting, deduplication and output rate limits remain mechanical.

Moderation

Qwen proposes a case; moderators use Ban, Delete message, or Dismiss. Buttons verify current Discord permissions, source card and role hierarchy on the server. Ban/dismiss-ban needs Ban Members; delete needs Manage Messages. Reaction quorum is available to human users with the server's existing Members role. Default: four distinct approvals act; three dismissals close. Each voter has one position; role membership and target protections are rechecked before enforcement.

Set the existing Members role ID in COMMUNITY_VOTER_ROLE_IDS. No new role, voter enrollment or native Discord moderation permissions are needed. Set additional staff/protected role IDs in PROTECTED_MODERATION_ROLE_IDS. Owners, bots, configured admins and members with moderation/administration permissions are automatically protected from community votes. Role IDs are used, not names, account age or keyword rules. Empty voter-role configuration keeps community voting disabled while moderator actions continue working. This deliberately trusts Members-role holders: multiple accounts with that role can collude. Distinct Discord user IDs are not proof of distinct people. The bot itself still needs Ban Members and a role above the target. Startup fetches fresh permission/role state and logs moderation readiness; per-target checks remain mandatory. A vote passing is not proof a ban succeeded.

Manual bans close matching cases automatically. Deleting a source message preserves a ban proposal and its redacted evidence. Repeated reports attach to one active case per target. Cases expire after BAN_VOTE_TTL_MS; failed API actions stay open. Redacted originals and latest observed edits/deletions are kept locally for 30 days, not fed into conversational memory or automatic training. Messages never observed by the bot cannot be reconstructed after deletion.

Verify

npm run typecheck
npm test
npm run build
npm audit
npm run eval
STATE_DB_PATH=:memory: npm run eval:moderation
npm run eval:discord-context

The participant evaluation uses real Grid inference with production prompts and tool schemas. Discord actions and generation are stubbed; it never posts to a channel. Its conversation fixtures include human-directed instructions, someone else's thanks, direct questions and validator errors requiring live evidence. It also covers a price question without a mention, an answer arriving mid-turn, scam reporters, and benign deletions. Requests are paced in this evaluator to avoid exhausting the shared service limit; live participation is not throttled by that test pacing.

Diagnose A Turn

Read both stdout and stderr. JSON events share turn_id, message_id, and channel_id: turn queued, turn started, context/model/tool events, then one turn finished. Outcomes are replied, silent, acted, failed, or superseded. Silent means an explicit model decision, not a swallowed error.

queue_ms measures waiting; end_to_end_ms includes ingestion and queue time. model_ms includes model request/stream latency, not just GPU execution. tool_ms includes finish-time context reads, so it overlaps context_ms; do not sum all timing fields. Context source, messages after focus and refresh counts help explain stale work. Partial effects are separately recorded even on failure. Raw messages, tool arguments and private reasoning are not included in turn traces. The process supervisor must retain and rotate both output streams.