ADR-075 — Nexie v3: the roadmap-study cycle as a skill-level behavior

Status: accepted (2026-07-16). Ships with the whats-next 0.3.0 upgrade (skills rewrite + prompt/effort/budget deltas; no graph change). Extends ADR-060 (conversational single-agent bots) without modifying it.

Context

The 2026-07-15 roadmap session demonstrated the target shape for Nexie (bots/whats-next, the conversational co-CTO): the operator manually led parallel audits (docs/ADR follow-ons, code gaps, operational state) → memory cross-check → verify-before-claim → synthesis into named chantiers with now/next/later tiering, quick-wins and blind spots → grouped arbitrage questions → an epic + framed tickets with a limited ready lot → factory drain observation → an evidence-based bilan. Board ticket native:bbcd96aa captures the referential: Nexie should lead this cycle end-to-end instead of only recommending.

Nexie v2 (ADR-060) is ONE claude_code agent in a 5-node chat loop (seed → nexie → gate → chat ⇄), with all seven board.* capabilities, a schema-validated turn envelope {reply, close, quick_replies, dispatched_ids}, and skills mirrored into .claude/skills/. Three of its skills were still v1-shaped (written for the deleted 15-node form pipeline), and nothing encoded the study cycle, the framed-ticket discipline, or the live-factory operating rules.

Options

A — skills + prompt, no graph change (chosen). The cycle becomes a named on-demand behavior carried by rewritten/new skills (repo-survey fan-out protocol, roadmap-synthesis chantiers + framed tickets, operator-arbitrage decision blocks, factory-ops live-factory rules) plus small prompt deltas. The audit fan-out rides the claude_code backend's native Task tool — reasoning_effort: ultracode adds the standing orchestration consent (model default is already opus-4-8, so diagnostic C089 stays silent). Dispatch stays board-promotion (set_bot + transition_issue → ready); monitoring stays the watched-cards channel (dispatched_ids → server-side auto-subscribe → state-change messages at tool boundaries).

B — subbot orchestration (rejected). Modeling the audits as subbot nodes breaks the conversational contract: subbots are synchronous graph nodes, so the chat parks while they run; parallel fan-out would force isolated: true + lifting worktree: none and max_parallel_branches: 1; parent budgets do not bound children. Every one of those contradicts an ADR-060 invariant pinned by e2e/whats_next_loop_test.go.

C — engine work first (deferred). Give in-run agents a run-monitoring tool (child run status/logs), a first-class "launch bot now" tool, or dispatcher-control MCP surface. Real gaps, but none is required for the cycle: board state transitions are sufficient telemetry for the bilan, and board-promotion is the sanctioned launch lever. Deferred until a dogfooded need proves them.

Decision

Option A, with four operator-arbitrated choices:

1. Full cycle in-run — audits, synthesis, arbitrage, board execution, factory observation and bilan are all Nexie's job; the engine gaps of option C stay out of v3.0.
2. Nexie stays code-read-only — quick-wins are framed tickets routed to worker bots (or listed for the operator), never coded in session; worktree: none and the read-only-outside-the-board guardrail hold.
3. The cycle is doctrine, not graph — it fires on roadmap-scale asks; a small question keeps its small answer. Skills guide, they don't script (the over-framing anti-pattern from ADR-055/058 applies to prompts too).
4. Factory limits are operator-owned — Nexie promotes ≤~5 cards per turn, never raises a cost cap, never spawns a second server/studio on an active store, and recommends (with the exact command) instead of driving the dispatcher.

Consequences

• The turn envelope is unchanged, so the golden replay (nexie_turn_basic, plus the new hand-authored nexie_study_synthesis), the graph-contract e2e, and the studio watched-issues stamp all survive as-is.
• The synthesis lives in reply markdown and on the board (framed tickets: ## Context / ## Done criteria / ## Verify) — no new schema fields, no studio work.
• Per-burst budget grows to 45m / $15 / tool_max_steps: 80 to fit a study turn; ordinary turns use a fraction and the chat pause still costs nothing.
• The horizon: label namespace is re-canonised on the study tiers (now|next|later); legacy spellings remain readable equivalents and are never mass-relabeled silently.
• Known accepted limits: Nexie cannot read child-run internals (bilan evidence = board states + fresh git log/greps), and cannot verify a dispatcher is running other than by observing that promoted cards do not move — both are stated honestly in factory-ops and are the first candidates for option C follow-ups if dogfooding demands them.