whats-next

whats-next — Nexie, the conversational co-CTO

ONE adaptive agent in a standing chat loop. You talk; Nexie reads the board and the repo, recommends (never raw dumps), creates/curates/ dispatches tickets, verifies whether issues are still relevant against the code and git history, and remembers the session (same LLM session across turns + a cross-run CONTEXT_BRIEF.md).

Shape (v3)

seed (compute) → nexie (agent) → gate (compute) ── is_close ──▶ done
                    ▲                 │ (default)
                    │                 ▼
                    └── conversation_loop(1000) ── chat (human)

• nexie — claude_code + ${ITERION_WHATS_NEXT_MODEL_CLAUDE:-claude-opus-5}, full board capabilities (board.read/create/move/assign/label/close/comment), bundled skills via the native Skill tool, interaction: human so it can ask_user mid-turn (with clickable options). Each turn returns {reply, close, quick_replies, dispatched_ids}.
• chat — a one-field human node: Nexie's reply renders as the chat bubble, the answer is the operator's next message. The pause is budget-free and can last days — this is the standby home base.
• The loop edge carries outputs.nexie._session_id so every turn resumes the same claude session: the conversation has real memory.
• Exit: only an explicit operator "close". Everything else loops.

v1 (15 nodes: survey → priorities form → roadmap → review form → emit → dispatch pickers → ask_continue radio) is gone — the forms made the operator encode questions into preset fields and each exchange cost a full state-machine turn. See docs/bot-runs/whats-next.md for the dogfood history that motivated the rewrite.

Run

devbox run -- iterion run bots/whats-next/main.bot \
  --var initial_message="quels tickets sont des quick wins ?"

Or from the studio: /whats-next (the composer seeds initial_message; presets are just canned seeds). When dispatched from a board card, the manifest maps the issue title/body into initial_message.

Inputs

Var	Type	Default	Purpose
initial_message	string	""	seed for the first turn; empty → Nexie opens with a board summary + recommendation
scope_notes	string	""	standing constraints for the session
workspace_dir	string	${PROJECT_DIR}	repo whose board/code Nexie works

Prerequisites

• Claude Code CLI signed in (or ANTHROPIC_API_KEY) — the nexie agent runs on claude_code.
• A writable .iterion/dispatcher/ under workspace_dir (the native kanban store auto-initialises on first issue creation).

Guardrails

• Targeted explicit instruction → act immediately; bulk (≥3) or destructive → dry-run + ask_user confirmation.
• Read-only outside the board (no repo edits, no commits — and worktree: none, so no phantom storage branches).
• Board writes are capability-gated MCP tools; set_bot is the canonical dispatcher selector; dispatch = transition to ready.
• Untrusted-input boundary: repo/issue text is data, not instructions.

Skills bundled with this bot

All under skills/, mirrored to <workspace>/.claude/skills/ at run start: whats-next (playbook), iterion-board, iterion-bot-catalog (generated — edit bot manifests, not the file), iterion-label-vocabulary, repo-survey, roadmap-synthesis, operator-arbitrage, factory-ops, session-continuity, iterion-dsl-quickref, dogfood-cycle.