CHANGELOG.md

@@ -2,6 +2,16 @@
 
 All notable changes to **History** are listed here, newest first.
 
+## v0.31.0 — 2026-10-15 10:22
+- **#52: dynamic scenario generation — the "Create scenario" path.** An additive second way to start a game alongside the 5 static presets: the user picks one of **3 light era templates** (`ww2` 1939, `present` 2026, `near-future` 2035) and writes a free-text **description** (plus an editable start year), and the LLM **generates** a draft world + an advisor **checkpoint list** (step 1). The player reviews the checkpoints (approve/remove) and picks a nation to lead, and a second LLM call **finalizes** the world — constrained to the *approved* checkpoints — into a saved, playable `GameState` (step 2): a normal game at `#/game/<id>` that plays with the existing turn engine, map, and panels untouched. The 5 static presets and the turn stub (`server/stub.ts`) are 100% unchanged; a new deterministic **scenario stub** (a distinct transport discriminating the two calls by marker) makes the whole path reproducible under `LLM_STUB=1`.
+  - **Era templates (`presets/templates/`).** Three small JSON seeds + an `index.json` manifest — each `{ id, name, year, eraDescription, suggestedPowers: [{ name, region (ISO_A3), note }], tensions: [] }` (e.g. near-future suggests a North American Union/USA, a Chinese Federation/CHN, a European Confederation/FRA, a Russian Union/RUS, with tensions like "autonomous weapons and the new arms race"). `server/templates.ts` is a tiny store mirroring `createPresetStore` (manifest + per-file load; a missing/broken file degrades to absent).
+  - **Shared contract (`shared/scenario.ts`, dependency-free).** `EraTemplate` / `TemplateMeta` / `SuggestedPower`; `WorldDraft = { nations: WorldNation[]; events? }` where `WorldNation.territory` holds **ISO_A3 country codes** (not region keys) and `relations` are numeric sentiments (−100..100); `Checkpoint = string`; the four message shapes (`ScenarioDraftRequest/Response`, `ScenarioCommitRequest/Response`). Pure transforms: `validateWorldDraft(raw, { validCountryCodes })` — defensive parse/clamp/normalize that never throws (name-required, id kept-or-slug fallback, duplicate ids suffixed, territory upper-cased/deduped and filtered to known ISO_A3 codes when a valid set is given, relations re-keyed to assigned ids via id or case-insensitive name match with self/unknown refs dropped, optional backstory events kept only when they name a known nation; returns `null` when the result is unusable — no nations, or no valid territory to render); `relationFromScore` (≥60 ally · 0..59 neutral · −40..−1 rival · −80..−41 hostile · ≤−81 war); `worldDraftToGameState(draft, expand, playerNationId, opts)` — expands ISO_A3 territory to region keys via the injected `expand` (`expandCountriesToKeys` in production), seeds `regionStats` (so each owned region starts with its owner's stats and the derived aggregate equals them on turn 1), a clean turn-1 state (`version:1`, `turn:1`, empty `unowned`/`markers`/`pendingNationOps`/`pending`), `presetId: 'dynamic'`, and a guard that re-points a dangling `playerNationId` to the first nation so the game is always playable.
+  - **Scenario LLM steps (`server/scenario.ts`).** Two plain `chat()` calls (no tools, no separate advisor endpoint — the advisor is embodied in step 1's checkpoint list): `generateDraft({ template, description, year })` → `{ world, checkpoints }` (the prompt carries the era description, suggested powers, tensions, the user's description, the year, and the valid ISO_A3 reference), and `finalizeWorld({ …, approvedCheckpoints, draft, playerNationId })` → final `GameState` (re-asked to produce the final world constrained to the approved checkpoints, based on the step-1 draft the client re-sends so the chosen nation survives; falls back to the draft itself when the model returns nothing usable). Same defensive parsing as the engine (strip fences → first balanced object → `JSON.parse` → `validateWorldDraft`), one nudge-retry when the answer is unusable; a transport failure (model down) throws → HTTP 502, a bad *answer* degrades to the fallback and never crashes.
+  - **Scenario stub (`server/stubScenario.ts`).** Deterministic transport for the two scenario calls only (discriminated by `SCENARIO_GENERATE_MARKER` / `SCENARIO_REGENERATE_MARKER` in the system message, mirroring how the turn stub keys off `ADVISOR_MARKER`): a fixed 3-nation world (Alba/FRA 70·60·65·55, Boreas/RUS 55·70·50·60, Gamma/ESP 30·45·70·40) + one backstory event ("The new order takes shape") + a fixed 3-checkpoint list; the finalize step refines the world deterministically from the approved checkpoints (each naming a nation by its 4-letter name prefix raises that nation's stability +5, clamped) — so "finalize respects the approved checkpoints" is observable end-to-end. The turn stub stays 100% unchanged.
+  - **Routes (`server/index.ts`).** `GET /api/scenarios/templates` (the 3 manifest entries), `POST /api/scenarios/draft` (`{ templateId, description, year? }` → `{ world, checkpoints }`), `POST /api/scenarios/commit` (`{ templateId, description, year?, draft?, approvedCheckpoints, playerNationId }` → `{ id }` — the finalized world is saved to SQLite with `startDate = <year>-01-01` and title = the template name). A separate `scenarioLlm` client (the scenario stub under `LLM_STUB=1`, the same live `qwen3.8-27b` client otherwise) — the turn `llm` is untouched. `POST /api/games/:id/jump` synthesizes a minimal `Preset` (`dynamicPresetFromGame`) for a saved dynamic game (`presetId: 'dynamic'` has no static file on disk — `runTurn` only consumes title/startDate/worldBackground/rules) so a created scenario plays with the normal turn engine; a genuinely-missing static preset still 404s.
+  - **Client.** New `ScenarioForm` (3 template cards, a pre-filled editable year, a description textarea, Generate → `/draft`; a 502 toasts and stays on the form) and `CheckpointReview` (the draft world as nation cards with stats, the checkpoint checklist with per-item approve/remove, a pick-a-nation-to-lead step, Finalize → `/commit` → navigates to `#/game/<id>`; a 502 toasts and stays on review). `PresetPicker` gains a "Create scenario" card, and `App.tsx` wires the `home → form → review → navigate` flow (a distinct `scenarioBusy` flag that never gates "Jump forward" or the turn flow).
+  - **Tests.** Unit: the template store (manifest + all 3 files' shape); `validateWorldDraft` / `relationFromScore` / `worldDraftToGameState` (valid → clean, stat/score clamping, bad country codes dropped, garbage → null without throwing, id dedup, territory expansion, `regionStats` seeding, player fallback, turn-1 invariants); `generateDraft` / `finalizeWorld` against the scenario stub (deterministic output, finalize respecting the approved checkpoints, a saved playable game). E2E: a new `e2e/scenario.spec.ts` (2 tests — the templates manifest over the API, and the full deterministic flow under `LLM_STUB=1`: near-future (2035) + description → draft → remove one checkpoint → lead Boreas → finalize (checkpoints shift stats) → a playable game at the `#/game/<id>` deep link with turn 1, correct date, dossier stats, map legend, and the backstory event; plus an empty jump proving the dynamic game runs the normal turn loop). The existing `e2e/sim.spec.ts` is unchanged. Totals: **307 unit tests (14 files)**, **16 e2e tests (2 scenario + 14 sim)**; typecheck, lint, and build green.
+
 ## v0.30.0 — 2026-10-15 08:22
 - **#50: make the four 0–100 stats consequential — a hard military conquest gate, soft stability takeovers, and economy growth/decay.** The four stats now actually drive outcomes instead of being decorative. **Military is a hard gate:** a contested invasion (a region's sole owner changing hands) is repelled whenever the attacker's military is **more than `CONQUEST_MILITARY_MARGIN = 15`** below the defender's — equal or within 15 still wins. The engine enforces this in `applyTurnResult` and emits a new `repelled` event when it fires; **re-claims of unowned (no man's land) regions are not gated** (there is no defender). **Stability is a soft lever:** the system prompt documents that a nation whose stability is below 30 *and* that hosts a `gathering` marker can suffer a **non-military takeover** (coup / revolution / faction seizure) — narrated by the LLM with a political/social success event, a gathering marker, and a downward stability/economy shock. **Economy is a soft lever:** a strong (rising) economy trends the other stats up (growth) and a collapsing economy drags them down (decay), reflected via `stateChanges` / `regionChanges`. A **frontier military-balance hint** and a **takeover-risks** hint are emitted in the user payload (each only when present) so the model — and the player, via the same data — sees the levers before acting. The stub stays 100% deterministic and untouched; e2e is unaffected (the UK/wwii-1939 e2e has an empty frontier, so the gate is vacuous there).
   - **Protocol (`shared/protocol.ts`).** New `CONQUEST_MILITARY_MARGIN = 15` and a pure `canConquer(attackerMil, defenderMil, margin = CONQUEST_MILITARY_MARGIN)` — `!(attackerMil < defenderMil - margin)` (proceeds unless the attacker is strictly more than the margin below the defender).

package.json

@@ -1,6 +1,6 @@
 {
   "name": "history",
-  "version": "0.30.0",
+  "version": "0.31.0",
   "private": true,
   "type": "module",
   "description": "AI-driven alternate-history simulator — the LLM is the world engine (Pax Historia style).",