feat(#57): hybrid war — TurnResult.hybrid ops (spy/sabotage/assassinate/cyber/economic/information/proxy) with exposure backfire

Promote the bare 'war' relation to a named War entity in GameState.wars:
- War: { id, sides[], declaredTurn, declaredDate?, aims?, status, front? }
- TurnResult.wars: { declare: WarDeclareOp[], end: WarEndOp[] }
- syncWarRelations: derive 'war' from active wars, fallback to 'hostile' on end
- Mutual-defense drag-in: belligerent allies added to sides (one level)
- Bare 'war' back-compat: never downgraded, never force-created
- Stub: empty wars (e2e unaffected); ~30 new unit tests

v0.33.0

AGENTS.md

@@ -327,6 +327,57 @@ history/
   `server/protocol.test.ts`, `server/db.test.ts`, `server/regions.test.ts`,
   `scripts/region-geo.test.ts`). **PR 1 of 2 — deliberately NOT deployed** (the user deploys
   when PR 2, the per-region LLM lever, ships).
+- **Wars — `GameState.wars`, a first-class conflict entity (v0.33, #56).**
+  Since v0.33 the `war` relation is no longer a bare boolean flag but a named `War` entity in
+  `GameState.wars` (required, `[]` at creation; a legacy SQLite row without the field loads as
+  `[]` — additive, never wipes). Each `War` has a stable id (`w1`, `w2`, …), a `sides` list
+  (≥2 distinct nation ids — any two distinct members are at war with each other), `declaredTurn`,
+  `declaredDate`, optional `aims`, a `status` (`active` / `armistice` / `capitulation`), and an
+  optional `front` (region keys). `TurnResult.wars = { declare: WarDeclareOp[]; end: WarEndOp[] }`
+  (both required, default `[]`): a **declare** creates a new active `War` (sides resolved via
+  `resolveNationRef` — id first, then case-insensitive unique name; a declare with <2 resolved
+  distinct sides is dropped); an **end** finds a war by id (unknown → no-op, never throw) and sets
+  its `status` to the outcome (`armistice` or `capitulation`). After declare/end,
+  `syncWarRelations(prev, next)` enforces the derived `war` relation: any two nations on opposite
+  sides of ≥1 **active** war get `war` in both directions (overriding lesser statuses — you
+  cannot "make peace" without ending the war); when a war ends, the pairs it caused that are not
+  covered by another active war fall back to `hostile`. A **bare** `war` relation with no matching
+  active war (legacy/back-compat) is never downgraded and never force-created. **Mutual-defense
+  drag-in**: when a war is declared, each belligerent's `ally`-relation partners are added to the
+  war's `sides` (one level, no cascade; deterministic — sides in declared order, allies in sorted
+  id order). The system prompt documents the field + declare/end ops (at most 1–2 declarations
+  per turn, each narrated in `events`); `buildUserPrompt` carries the current **active** wars
+  (as `WarInfo[]`: id + sides + aims) so the model can `end` by id. The **stub never declares or
+  ends** a war (`wars: { declare: [], end: [] }`) — e2e (CI, `LLM_STUB=1`) is unaffected and the
+  path is unit-tested only (`server/engine.test.ts`).
+- **Hybrid war — `TurnResult.hybrid`, a deterministic non-kinetic operations spectrum (v0.34, #57).**
+  `TurnResult.hybrid = { ops: HybridOp[] }` (required; default `{ ops: [] }`) adds a covert /
+  hybrid-warfare spectrum alongside the kinetic `regionChanges` lever: `HybridOp =
+  { type, from, target, detail?, outcome? }` with `type ∈ spy | sabotage | assassinate |
+  cyber | economic | information | proxy` and `outcome ∈ success | exposed` (the LLM
+  declares its *intended* outcome; the engine applies the deterministic effect).
+  `from`/`target` are refs resolved by the shared `resolveNationRef` (id first, then
+  case-insensitive unique name); a `target` that is not a resolvable nation is treated
+  as a region key — it hits that region's SOLE owner (a key with 0 or 2+ owners — or the
+  acting nation itself — is dropped, never thrown). Per type: `sabotage` dips `economy`
+  (−3), `economic` dips `economy` (−5), `cyber` dips `military` (−3), `assassinate` dips
+  `stability` (−5); `information` nudges the from↔target relation one step apart (capped
+  at `hostile`), `proxy` likewise but capped at `rival`; `spy` is informational only.
+  A region-target dip lands on the single `regionStats[regionKey]` entry (hybrid ops
+  never move territory — `regionChanges` is the only lever that does); a nation-target
+  dip lands on every region the nation owns; afterwards the target's derived `stats` is
+  re-derived (0-region / no-entry guards), so the effect persists and the sole-ownership
+  invariant is untouched. **Exposed** (`outcome: "exposed"`) is a backfire: the intended
+  effect is NOT applied; instead the from↔target relation moves one step *toward*
+  `hostile` (capped; `war` is never created or escalated by an exposed op) and a
+  diplomatic/failed event is auto-appended to the turn's `events`.
+  **Additive only — no `GameState.hybrid` field, no `db.ts` change** (the ops are
+  consumed inside `applyTurnResult` after territory + wars and before markers).
+  `isMeaningful` counts hybrid ops; the system prompt documents the field + per-type
+  effects (the "unclear" fallback and the stub both return `hybrid: { ops: [] }`), and the
+  user payload carries the player's `relations` map so the model can reason about the
+  relation caps. The **stub never** proposes a hybrid op, so e2e (CI, `LLM_STUB=1`) is
+  unaffected and the path is unit-tested only (`server/engine.test.ts`).
 - **Map click → dossier target (owner resolved from the RegionMap, by region key).**
   Clicking a sub-region does NOT open a popup — it switches the dossier to that sub-region's
   owning nation (`focusNation(ownerId, {name})` in `App.tsx`, which also records which region was

CHANGELOG.md

@@ -2,6 +2,49 @@
 
 All notable changes to **History** are listed here, newest first.
 
+## v0.34.0 — 2026-10-15 16:35
+- **#57: Espionage & Hybrid War — a non-kinetic operations spectrum with a deterministic per-type effect and an exposure backfire.** Every prior turn lever was kinetic (conquest, state-change, region-change, wars, takeovers) or a player-confirmed proposal (nation ops). This version adds `TurnResult.hybrid = { ops: HybridOp[] }` (required, default `{ ops: [] }`) so the model can run **covert, non-kinetic** operations: spies, sabotage, assassination attempts, cyberattacks, economic pressure, disinformation, and proxy support. Unlike territory these **never move borders** — they dip a target's stats or sour a relation. The engine is **fully deterministic (no hidden randomness)**: the LLM declares an intended `outcome` (`success` by default, or `exposed`); a successful op applies its small per-type effect, an `exposed` op **backfires** (no effect, the actor's relation with the target sours one step toward hostile, and an auto `diplomatic`/`failure` event is emitted).
+  - **Protocol (`shared/protocol.ts`).** New `HybridOpType` (the seven types) + `HYBRID_OP_TYPES` const, and the `HybridOp` interface (`{ type; from; target; detail?; outcome? }`). `from` is the acting nation (id or unique name — resolved at apply time); `target` is **either** a nation (id/name) **or** a region key (hits that region's sole owner). `TurnResult.hybrid = { ops: HybridOp[] }` is required, placed between `wars` and `markers`.
+  - **Engine (`server/engine.ts`).**
+    - `isHybridOpType(t)` — exported guard over the seven types.
+    - `validateHybridOps(raw)` — new exported, defensive validator (never throws): accepts the `{ ops: […] }` object **or** a bare array; each op needs a known `type`, a non-empty `from` (≤200) and non-empty `target` (≤200); `outcome:'exposed'` is kept (success/missing is omitted to keep the shape minimal); `detail` capped at 300.
+    - `HYBRID_STAT_EFFECT` — exported table of the deterministic per-type stat dip: `sabotage`→economy −3, `assassinate`→stability −5, `cyber`→military −3, `economic`→economy −5; `spy`/`information`/`proxy` have no stat effect.
+    - `shiftRelationTowardHostile(state, a, b, cap)` — exported; nudges the bilateral relation **one** rung up the hostile ladder (`ally→neutral→rival→hostile`), symmetrically, capped at `cap`; a missing relation starts at `neutral`; a `war` relation is never touched (a covert op neither fabricates nor downgrades a war). Used for the `information`/`proxy` nudge and the `exposed` backfire.
+    - `resolveHybridTarget(state, target, validKeys?)` — internal; resolves a target to the concrete nation it affects: a nation ref via `resolveNationRef`, else a region key's **sole** owner (0 or 2+ owners → not targetable); `validKeys` (when supplied) gates the region path.
+    - `applyHybridOp(next, op, geoOf?, validKeys?, regionNameOf?, actingTurn?)` — internal; resolves `from`/`target` (unknown/ambiguous actor, unresolvable target, and self-targets are dropped — never throw), then applies the per-type effect: an `exposed` op backfires (relation sours one step, cap hostile, auto `diplomatic`/`failure` event); a successful stat op dips the target's `regionStats` (the single named region for a region target, otherwise **every** region the target owns) and re-derives the target nation's aggregate stats (same 0-region/no-entry guards as the main recompute) so the effect persists; `information`/`proxy` nudge the relation (caps `hostile`/`rival`); `spy` is informational only.
+    - `applyTurnResult` — new **hybrid block** (after the wars block, before markers): applies each validated op against the post-wars state. Sole ownership is preserved (a hybrid op affects stats, never ownership).
+    - `isMeaningful` — a non-empty `hybrid.ops` makes a turn meaningful.
+    - `validateTurnResult` — initializes and validates `hybrid` (default `{ ops: [] }`).
+    - `buildSystemPrompt` — documents the `hybrid` field and a HARD CONSTRAINTS bullet (non-kinetic, never moves borders, the per-type effects, the `exposed` backfire, region-key targets hit the sole owner, at most a few per turn, narrate each in `events`) in both tool and plain modes.
+    - `buildUserPrompt` — new optional `relations` param (the player's `nationId → RelationStatus` map) so the model sees the ladder its ops nudge and the starting point before a backfire; emitted only when provided.
+    - `runTurn` — computes the player's `relations` from state and passes them as the last `buildUserPrompt` arg; the "unclear" fallback carries `hybrid: { ops: [] }`.
+  - **Additive only.** No `GameState` field, **no** DB migration, **no** `db.ts` change — `hybrid` lives entirely in the `TurnResult` and is consumed by `applyTurnResult`.
+  - **Stub (`server/stub.ts`).** `deterministicTurn` returns `hybrid: { ops: [] }` — the stub **never** runs a covert op (mirrors `nationOps: []`, `takeovers: []`, `wars`). `StubPayload` gains an optional `relations` field (ignored). E2E (CI, `LLM_STUB=1`) is fully unaffected; the path is unit-tested only.
+  - **Unit tests (`server/engine.test.ts`).** ~30 new tests: `isHybridOpType`; `validateHybridOps` (both raw shapes, trim/cap, exposed-vs-success, drop rules); `validateTurnResult` — hybrid; `isMeaningful` — hybrid; `shiftRelationTowardHostile` (ladder, caps, war-untouched, missing=neutral); `applyTurnResult` — hybrid (each stat type's dip, nation-vs-region target, the per-type relation nudge & caps, `exposed` backfire + auto event, exposed-on-war, drop rules for unknown actor / self-target / unknown region / double-owned region, the `validKeys` gate, borders-untouched invariant, `HYBRID_STAT_EFFECT` magnitudes); `buildSystemPrompt` — hybrid contract (both modes); `buildUserPrompt` — relations; stub — hybrid.
+  - **Client.** No change — hybrid ops are an engine/LLM concern; their effects surface through the existing stats, relations, and event feed.
+  - **AGENTS.md.** New convention entry documenting the `hybrid` field, the deterministic per-type effects, the `exposed` backfire, the region-vs-nation targeting, and the stub/test coverage.
+
+## v0.33.0 — 2026-10-15 15:45
+- **#56: War as a first-class state — the `war` relation becomes a named `War` entity.** The v0.22+ `war` relation was a bare boolean flag on the `Relation` type with no lifecycle, no identity, and no way to distinguish "at war since 1939" from "at war since 1941" or to narrate the end of a conflict. This version promotes it to a first-class `War` entity in `GameState.wars` with a stable id, sides, declaration date/turn, optional aims, a lifecycle status (`active` / `armistice` / `capitulation`), and an optional front (region keys). The LLM declares and ends wars via `TurnResult.wars = { declare, end }` (both required, default `[]`); the engine creates/updates the entity and **derives** the `war` relation from active wars (any two nations on opposite sides of ≥1 active war get `war` in both directions, overriding lesser statuses — you cannot "make peace" without ending the war). When a war ends, the pairs it caused that are not covered by another active war fall back to `hostile`. A **bare** `war` relation with no matching active war (legacy/back-compat) is never downgraded and never force-created. **Mutual-defense drag-in**: when a war is declared, each belligerent's `ally`-relation partners are added to the war's `sides` (one level, no cascade; deterministic — sides in declared order, allies in sorted id order), so the dragged-in ally becomes `war` with all belligerents including its former ally.
+  - **Types (`shared/types.ts`).** New `War` interface (`id`, `sides: string[]` ≥2 distinct, `declaredTurn: number`, `declaredDate?: string`, `aims?: string[]`, `status: 'active'|'armistice'|'capitulation'`, `front?: string[]`); `GameState.wars: War[]` (required, `[]` at creation; a legacy SQLite row without the field loads as `[]` — additive, never wipes).
+  - **Protocol (`shared/protocol.ts`).** New `WarDeclareOp` (`{ sides: string[]; aims?: string[] }`) and `WarEndOp` (`{ id: string; outcome: 'armistice'|'capitulation'; terms?: string }`); `TurnResult.wars = { declare: WarDeclareOp[]; end: WarEndOp[] }` (both required, default `[]`). New `WarInfo` interface (`{ id; sides; aims? }`) for the user-prompt payload.
+  - **Engine (`server/engine.ts`).**
+    - `validateWarOps(raw)` — new exported, defensive validator (never throws): non-object → empty; `declare` — sides trimmed/capped 200/deduped, ≥2; aims ≤5×200; `end` — id ≤200, outcome exactly `armistice`|`capitulation`, terms ≤400.
+    - `activeWarPairs(wars): Set<string>` — NUL-joined sorted pair keys for active-only wars.
+    - `atWar(wars, a, b)` — convenience check.
+    - `syncWarRelations(prev, next)` — exported; enforces the derived `war` relation (active pairs → `war` both directions; ended pairs not covered by another active war → `hostile`; bare `war` never downgraded, never force-created).
+    - `freshWarId(taken)` — first free of `w1`, `w2`, … (next to `freshNationId`).
+    - `applyTurnResult` — new **wars block** (after takeovers, before markers): declare → create `War` (sides resolved via `resolveNationRef`; <2 distinct → dropped); end → find by id (unknown → no-op); then `syncWarRelations(state, next)`.
+    - `isMeaningful` — a non-empty `declare` or `end` makes a turn meaningful.
+    - `buildSystemPrompt` — documents the `wars` field (declare/end, at most 1–2 per turn, each narrated in `events`; the `war` relation is derived — do not emit it directly for active-war pairs).
+    - `buildUserPrompt` — carries current **active** wars as `WarInfo[]` so the model can `end` by id.
+    - `runTurn` — computes active-only `wars` from `state.wars`, passes as last arg; the "unclear" fallback carries `wars: { declare: [], end: [] }`.
+  - **DB (`server/db.ts`).** `load` defaults `state.wars` to `[]` when absent (after the markers default, before `migrateRegionStats`).
+  - **Stub (`server/stub.ts`).** `deterministicTurn` returns `wars: { declare: [], end: [] }` — the stub never declares or ends a war (mirrors `nationOps: []`, `takeovers: []`). E2E (CI, `LLM_STUB=1`) is unaffected; the path is unit-tested only.
+  - **Unit tests (`server/engine.test.ts`).** ~30 new tests: `freshWarId`, `validateWarOps`, `validateTurnResult` (wars), `activeWarPairs` / `atWar`, `syncWarRelations`, `applyTurnResult` — wars (declare creates war + derives relations; end removes active status + un-derives relations; mutual-defense drag-in; unknown-id end no-op; duplicate declare skip; unresolvable-ref drop; bare-war back-compat; relations-override-forced-back; same-turn declare+end leaves prior relation untouched), `isMeaningful` — wars, `buildSystemPrompt` — wars contract, `buildUserPrompt` — wars, stub — wars.
+  - **Client.** No change — wars are an engine/LLM concern; the UI already renders the `war` relation from the `Relation` type (which the engine now keeps in sync).
+  - **AGENTS.md.** New convention entry documenting the `War` entity, the `syncWarRelations` algorithm, the mutual-defense drag-in, and the stub/test coverage.
+
 ## v0.32.0 — 2026-10-15 10:56
 - **#54: non-military takeovers become a visible regime change — a `Takeover` in the turn renames the nation.** The v0.30.0 stats work made stability a *soft* lever: the system prompt told the model that a low-stability nation (`stability < 30`) hosting a `gathering` marker "can suffer a non-military takeover — a coup, revolution or faction seizure of power", narrated with an event and a stat shock — but the model had no actual *field* to express the regime change itself, so the nation's name never changed and the takeover was invisible in the world. This version closes that gap with a new `TurnResult.takeovers` field: `Takeover = { nationRef, newRegimeName }`. Each entry is an **automatic, in-world** regime change applied directly by `applyTurnResult` (unlike `nationOps`, which are player-confirmed proposals) — the target nation's display `name` becomes `newRegimeName`, while its stable `id`, territory, relations (keyed by id), stats, and `playerNationId` are all preserved, so a player whose nation is overthrown keeps leading it under the new regime.
   - **Protocol (`shared/protocol.ts`).** New `Takeover` interface (`nationRef` + `newRegimeName`, both strings). `TurnResult.takeovers: Takeover[]` is **required** (default `[]`), placed between `nationOps` and `markers`.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "history",
-  "version": "0.32.0",
+  "version": "0.34.0",
   "private": true,
   "type": "module",
   "description": "AI-driven alternate-history simulator — the LLM is the world engine (Pax Historia style).",

server/db.test.ts

@@ -31,6 +31,8 @@ function makeState(id: string): GameState {
     regionStats: {
       GBR: { military: 80, economy: 70, stability: 60, population: 50 },
     },
+    // (v0.33.0, #56) no wars.
+    wars: [],
   };
 }
 

server/db.ts

@@ -49,6 +49,10 @@ export function openDb(filePath: string): Db {
         if (state.pendingNationOps === undefined) state.pendingNationOps = [];
         // (v0.27, #44) Same for map markers: a legacy game has none placed.
         if (state.markers === undefined) state.markers = [];
+        // (v0.33.0, #56) Same for named wars: a legacy game has none declared
+        // (any pre-v0.33 `war` relation stays a bare/implicit relation — see
+        // syncWarRelations, which never force-creates a War entity).
+        if (state.wars === undefined) state.wars = [];
         // (v0.28, #46) Rows saved before per-region stats existed are migrated
         // ONCE at load time (and never re-migrated afterwards — the field is
         // then present): each owned region is seeded with a copy of its owner's

server/engine.ts

@@ -15,6 +15,7 @@ import type {
   RegionStats,
   RoundItem,
   StatKey,
+  War,
 } from '../shared/types';
 import { EVENT_TYPES, RELATION_STATUSES } from '../shared/types';
 import {
@@ -23,12 +24,15 @@ import {
   clampStat,
   deriveNationStats,
   fnv1a32,
+  HYBRID_OP_TYPES,
   isDateLike,
   isMarkerType,
   makeEvent,
   markerId,
   MAX_MARKERS,
   nationForPrompt,
+  type HybridOp,
+  type HybridOpType,
   type MarkerInfo,
   type MarkerSpec,
   type NationOp,
@@ -36,6 +40,9 @@ import {
   type Takeover,
   type TurnEventSpec,
   type TurnResult,
+  type WarDeclareOp,
+  type WarEndOp,
+  type WarInfo,
 } from '../shared/protocol';
 import {
   DEFAULT_MAX_TOKENS,
@@ -51,6 +58,12 @@ const VALID_EVENT_TYPES = new Set<string>(EVENT_TYPES);
 const VALID_RELATIONS = new Set<string>(RELATION_STATUSES);
 const VALID_OUTCOMES = new Set<string>(['success', 'failure', 'partial', 'neutral']);
 const VALID_STATS = new Set<string>(['military', 'economy', 'stability', 'population']);
+const VALID_HYBRID_TYPES = new Set<string>(HYBRID_OP_TYPES);
+
+/** True when `t` is one of the seven known hybrid op types (v0.34.0, #57). */
+export function isHybridOpType(t: unknown): t is HybridOpType {
+  return typeof t === 'string' && VALID_HYBRID_TYPES.has(t);
+}
 
 // --- The LLM's getNeighbours tool (v0.21, #16 P2: contiguous region transfers) ---
 
@@ -201,6 +214,8 @@ export function buildSystemPrompt(preset: Preset, withNeighboursTool = false): s
     '    {"kind":"merge","nationRefs":["<id>","<id>"],"survivingRef":"<id of the surviving nation>","into":"<new name>"}',
     '  ],',
     '  "takeovers": [{"nationRef":"<nation id or name>","newRegimeName":"<name of the new regime>"}],',
+    '  "wars": {"declare":[{"sides":["<nationId>","<nationId>"],"aims":["<short aim>"]}],"end":[{"id":"<war id>","outcome":"<armistice|capitulation>","terms":"<short terms>"}]},',
+    '  "hybrid": {"ops":[{"type":"<spy|sabotage|assassinate|cyber|economic|information|proxy>","from":"<nation id or name>","target":"<nation id or name, or region key>","detail":"<short detail>","outcome":"<success|exposed>"}]},',
     '  "markers": {"add":[{"type":"<city|factory|army|port|capital|fleet|rocket|gathering>","name":"<short name>","regionKey":"<region key>"}],"remove":["<marker id>"]}',
     '}',
     '',
@@ -219,6 +234,8 @@ export function buildSystemPrompt(preset: Preset, withNeighboursTool = false): s
     '- Split: add {"kind":"split","nationRef":"<nation id or name>","parts":[{"name":"<new name>","territory":["<region key>"]}, ...]} when a state breaks up (dissolution, secession). Use at least 2 parts and assign each of the nation\'s region keys to at most one part — any key you leave unassigned goes to the first part.',
     '- Merge: add {"kind":"merge","nationRefs":["<id or name>","<id or name>"],"survivingRef":"<id or name of the surviving nation>","into":"<new name>"} when states reunify or one is absorbed. Use at least 2 input nations and "survivingRef" must name one of them — the survivor keeps its id and its stats.',
     '- Takeover (automatic, NOT player-confirmed): add {"takeovers":[{"nationRef":"<nation id or name>","newRegimeName":"<name of the new regime>"}]} when a low-stability nation (e.g. stability < 30) hosting a "gathering" marker is overthrown in a coup or revolution. This renames the nation to the new regime (e.g. "France" → "Vichy Regime"). Also emit a narrative "event" (political or social, outcome "success") and a stability/economy shock via "stateChanges". A takeover is an automatic in-world regime change — use "nationOps" rename only for a diplomatic/political name change that the player must confirm.',
+    '- Wars ("wars") are named conflicts, each with a stable id. To DECLARE one, add {"declare":[{"sides":["<id>","<id>"],"aims":[...]}]} with at least 2 distinct belligerents: while it is active the engine AUTOMATICALLY sets both directions of every warring pair\'s relation to "war" and drags in the mutual-defense allies of each belligerent, so do NOT also emit a "war" relation for those pairs (you can never "make peace" between them without ending the war). To END an active war, add {"end":[{"id":"<war id from the user message\'s "wars" field>","outcome":"<armistice|capitulation>"}]} — only active wars can be ended, and when one ends its belligerent pairs fall back to "hostile". Propose at most 1-2 declarations per turn and narrate each declaration and each ending in "events".',
+    '- Covert / hybrid ops ("hybrid.ops") are non-kinetic operations — spies, sabotage, assassination attempts, cyberattacks, economic pressure, disinformation, and proxy support. Unlike territory they NEVER move borders: a successful op applies a small, deterministic effect (a small dip in the target\'s economy / stability / military for sabotage / assassinate / cyber / economic, or a relation nudge toward hostile for information / proxy), while an "exposed" op backfires (no effect, and the actor\'s relation with the target sours by one step toward hostile). "target" may be a nation id / name or a region key (a region key hits that region\'s sole owner). Keep each detail short, propose at most a few per turn, and narrate each op in "events".',
     '- Prefer 1-4 events per turn; keep each text 1-3 sentences.',
     '- Keep your internal reasoning brief (a few sentences) and go straight to the JSON result; do not over-deliberate or re-derive facts already present in the state.',
     '- Keep every stat value between 0 and 100.',
@@ -263,6 +280,21 @@ export function buildUserPrompt(
    * Omitted entirely when undefined.
    */
   markers?: MarkerInfo[],
+  /**
+   * (v0.33.0, #56) The game's currently ACTIVE wars, as the model sees them:
+   * stable id, sides and aims. Emitted verbatim (even when empty) so the
+   * model can reference a war's id in `TurnResult.wars.end` to end it.
+   * Omitted entirely when undefined.
+   */
+  wars?: WarInfo[],
+  /**
+   * (v0.34.0, #57) The player's current relations to every other nation
+   * (`nationId -> RelationStatus`). Emitted verbatim (even when empty) so the
+   * model sees the relation ladder it can nudge with `information` / `proxy`
+   * hybrid ops and knows the current state before an `exposed` backfire.
+   * Omitted entirely when undefined.
+   */
+  relations?: Record<string, RelationStatus>,
 ): string {
   const player = state.nations[state.playerNationId];
   const playerName = player?.name ?? state.playerNationId;
@@ -375,6 +407,8 @@ export function buildUserPrompt(
   if (takeoverRisks !== undefined) payload.takeoverRisks = takeoverRisks;
   if (noMansLand !== undefined) payload.unowned = noMansLand;
   if (markers !== undefined) payload.markers = markers;
+  if (wars !== undefined) payload.wars = wars;
+  if (relations !== undefined) payload.relations = relations;
   return JSON.stringify(payload, null, 2);
 }
 
@@ -513,6 +547,8 @@ export function validateTurnResult(
     unowned: [],
     nationOps: [],
     takeovers: [],
+    wars: { declare: [], end: [] },
+    hybrid: { ops: [] },
     markers: { add: [], remove: [] },
   };
 
@@ -607,6 +643,16 @@ export function validateTurnResult(
   // by applyTurnResult (automatic, not player-confirmed).
   result.takeovers = validateTakeovers(o.takeovers);
 
+  // (v0.33.0, #56) War declare/end ops: creating and ending named Wars.
+  // Applied directly by applyTurnResult (after takeovers, before markers).
+  result.wars = validateWarOps(o.wars);
+
+  // (v0.34.0, #57) Covert hybrid ops (spies, sabotage, assassination, cyber,
+  // economic, disinformation, proxy). Applied directly by applyTurnResult
+  // (after wars, before markers) — a deterministic per-type effect or an
+  // exposed backfire.
+  result.hybrid = { ops: validateHybridOps(o.hybrid) };
+
   // (v0.27, #44) Map-marker ops: new placements (validated here) and
   // removals by stable id. Applied by applyTurnResult after the territory
   // moves (a new marker's owner is the region's post-turn owner).
@@ -736,13 +782,397 @@ export function validateTakeovers(raw: unknown): Takeover[] {
   return out;
 }
 
+/**
+ * (v0.33.0, #56) Validate the raw `wars` object of a turn into a safe
+ * `{ declare, end }` shape. Defensive like the rest of the parser — never
+ * throws. `declare`: each entry must be an object whose `sides` is an array
+ * of non-empty strings (trimmed, capped at 200 chars each, deduped) that
+ * yields at least TWO distinct sides — anything else (a one-sided or
+ * empty-sided "war") is dropped; `aims`, when an array, is kept only as
+ * non-empty strings (capped at 200 chars, at most 5). `end`: each entry must
+ * be an object with a non-empty `id` (capped at 200 chars) and an `outcome`
+ * of exactly `armistice` or `capitulation` (anything else is dropped);
+ * `terms`, when a non-empty string, is capped at 400 chars. A non-object
+ * input yields `{ declare: [], end: [] }`.
+ */
+export function validateWarOps(raw: unknown): { declare: WarDeclareOp[]; end: WarEndOp[] } {
+  const out: { declare: WarDeclareOp[]; end: WarEndOp[] } = { declare: [], end: [] };
+  if (!raw || typeof raw !== 'object' || Array.isArray(raw)) return out;
+  const o = raw as Record<string, unknown>;
+  if (Array.isArray(o.declare)) {
+    for (const e of o.declare) {
+      if (!e || typeof e !== 'object') continue;
+      const d = e as Record<string, unknown>;
+      const sides = (Array.isArray(d.sides) ? d.sides : [])
+        .map((s) => (typeof s === 'string' ? s.trim().slice(0, 200) : ''))
+        .filter((s) => s !== '')
+        // dedupe case-insensitively? No — sides are nation ids: keep exact
+        // strings, drop exact duplicates only (resolution is later).
+        .filter((s, i, arr) => arr.indexOf(s) === i);
+      if (sides.length < 2) continue; // a war needs at least two belligerents
+      const aims = Array.isArray(d.aims)
+        ? (d.aims as unknown[])
+            .map((a) => (typeof a === 'string' ? a.trim().slice(0, 200) : ''))
+            .filter((a) => a !== '')
+            .slice(0, 5)
+        : undefined;
+      out.declare.push({ sides, ...(aims && aims.length > 0 ? { aims } : {}) });
+    }
+  }
+  if (Array.isArray(o.end)) {
+    for (const e of o.end) {
+      if (!e || typeof e !== 'object') continue;
+      const w = e as Record<string, unknown>;
+      const id = typeof w.id === 'string' ? w.id.trim().slice(0, 200) : '';
+      if (!id) continue;
+      if (w.outcome !== 'armistice' && w.outcome !== 'capitulation') continue;
+      const terms = typeof w.terms === 'string' ? w.terms.trim().slice(0, 400) : '';
+      out.end.push({ id, outcome: w.outcome, ...(terms ? { terms } : {}) });
+    }
+  }
+  return out;
+}
+
+/**
+ * (v0.34.0, #57) Validate the raw `hybrid` field of a turn into a safe
+ * `{ ops }` shape. Defensive like the rest of the parser — never throws. The
+ * raw value may be the object `{ ops: [...] }` (the TurnResult shape) or a bare
+ * array (be lenient about either). Each op needs a known `type` (one of
+ * {@link HYBRID_OP_TYPES}), a non-empty `from` (the acting nation, capped at
+ * 200 chars) and a non-empty `target` (a nation id/name or a region key, capped
+ * at 200 chars); an `outcome` of exactly `success` or `exposed` is kept (the
+ * default, `success`, is stored as absent to keep the shape minimal — the
+ * engine treats a missing outcome as success); a non-empty `detail` is capped
+ * at 300 chars. Resolution of `from` / `target` to concrete nation ids or
+ * region owners happens at apply time (`applyHybridOp`), not here.
+ */
+export function validateHybridOps(raw: unknown): HybridOp[] {
+  if (!Array.isArray(raw)) {
+    // Lenient: accept the `{ ops: [...] }` object shape (or a non-array → []).
+    if (raw && typeof raw === 'object' && !Array.isArray(raw)) {
+      const inner = (raw as Record<string, unknown>).ops;
+      if (!Array.isArray(inner)) return [];
+      return validateHybridOpsList(inner);
+    }
+    return [];
+  }
+  return validateHybridOpsList(raw);
+}
+
+function validateHybridOpsList(raw: unknown[]): HybridOp[] {
+  const out: HybridOp[] = [];
+  for (const e of raw) {
+    if (!e || typeof e !== 'object') continue;
+    const o = e as Record<string, unknown>;
+    if (!isHybridOpType(o.type)) continue;
+    const from = str(o.from, 200);
+    const target = str(o.target, 200);
+    if (!from || !target) continue;
+    const detail = str(o.detail, 300);
+    const exposed = o.outcome === 'exposed'; // 'success' / missing → success
+    out.push({
+      type: o.type,
+      from,
+      target,
+      ...(detail ? { detail } : {}),
+      ...(exposed ? { outcome: 'exposed' as const } : {}),
+    });
+  }
+  return out;
+}
+
+/**
+ * (v0.33.0, #56) The UNORDERED pairs of belligerents of a set of ACTIVE
+ * wars: a war with sides [a, b, c] contributes the pairs {a,b}, {a,c},
+ * {b,c} (any two distinct belligerents of the same active war are at war —
+ * see the War type's semantics). Pure; exported for the unit tests.
+ */
+export function activeWarPairs(wars: War[]): Set<string> {
+  const pairs = new Set<string>();
+  const key = (a: string, b: string) => (a < b ? `${a}\u0000${b}` : `${b}\u0000${a}`);
+  for (const w of wars) {
+    if (w.status !== 'active') continue;
+    for (let i = 0; i < w.sides.length; i++) {
+      for (let j = i + 1; j < w.sides.length; j++) {
+        pairs.add(key(w.sides[i], w.sides[j]));
+      }
+    }
+  }
+  return pairs;
+}
+
+/** True when nations `a` and `b` are both belligerents of ≥1 active war. */
+export function atWar(wars: War[], a: string, b: string): boolean {
+  const key = a < b ? `${a}\u0000${b}` : `${b}\u0000${a}`;
+  return activeWarPairs(wars).has(key);
+}
+
+/**
+ * (v0.33.0, #56) Derive the `war` relation from the set of ACTIVE wars,
+ * MUTATING the (cloned) `next` state in place (the caller passes a clone —
+ * applyTurnResult does) and never touching `prev`. For every unordered pair
+ * of distinct existing nations:
+ *
+ *  - the pair is on the sides of ≥1 ACTIVE war in `next` → BOTH directions
+ *    of the pair's relation are set to `war` (the engine owns this relation
+ *    while the war is active — a manual `relations` entry to a lesser status
+ *    on a warring pair is overridden, so the two can never "make peace"
+ *    without the war being ended);
+ *  - the pair is NOT at war in `next` but WAS at war in `prev` (one of the
+ *    wars that tied them ended this turn) AND the current relation is
+ *    `war` → both directions become `hostile` (a war is the strongest
+ *    hostility — when it stops, the nations stay hostile, not suddenly
+ *    neutral);
+ *  - otherwise (not at war now, and the relation is not `war`, or the pair
+ *    was never tied by an active war) → untouched: the LLM owns
+ *    ally/neutral/rival/hostile freely, and a **bare** `war` relation that
+ *    no active war ever covered (a preset leftover) stays as-is — an
+ *    implicit, unnamed war, never force-created into a War entity and never
+ *    clobbered by the sync.
+ *
+ * Never throws.
+ */
+export function syncWarRelations(prev: GameState, next: GameState): void {
+  const prevAtWar = activeWarPairs(prev.wars ?? []);
+  const nextAtWar = activeWarPairs(next.wars ?? []);
+  const ids = Object.keys(next.nations);
+  for (let i = 0; i < ids.length; i++) {
+    for (let j = i + 1; j < ids.length; j++) {
+      const a = ids[i];
+      const b = ids[j];
+      const key = a < b ? `${a}\u0000${b}` : `${b}\u0000${a}`;
+      const na = next.nations[a];
+      const nb = next.nations[b];
+      if (nextAtWar.has(key)) {
+        na.relations[b] = 'war';
+        nb.relations[a] = 'war';
+      } else if (prevAtWar.has(key) && (na.relations[b] === 'war' || nb.relations[a] === 'war')) {
+        na.relations[b] = 'hostile';
+        nb.relations[a] = 'hostile';
+      }
+    }
+  }
+}
+
+// --- Hybrid war (v0.34.0, #57): covert, non-kinetic operations -------------
+//
+// The engine's DETERMINISTIC effects for the seven hybrid op types. There is no
+// hidden randomness: the LLM declares an intended `outcome` (success / exposed)
+// and the engine acts only on that. Each effect is small and documented:
+//
+//   spy         informational only — no stat, no relation (pure intel)
+//   sabotage    a small ECONOMY dip on the target
+//   assassinate a small STABILITY dip on the target (absent the named-leaders
+//               feature, #61, a leader's removal sours the polity)
+//   cyber       a small MILITARY dip on the target
+//   economic    a small ECONOMY dip on the target
+//   information a relation nudge toward hostile (cap hostile) — disinformation
+//   proxy       a light relation nudge toward hostile (cap rival) — proxy war
+//
+// A successful stat op dips the target's `regionStats` (the single named region
+// when the target was a region key, otherwise every region the target owns) and
+// re-derives the target nation's aggregate stats so the effect persists. A
+// successful information/proxy and any exposed op sour the actor's relation
+// with the target by one step on the hostile ladder.
+
+/** The deterministic per-type stat effect of a hybrid op (v0.34.0, #57). */
+export const HYBRID_STAT_EFFECT: Record<HybridOpType, { stat: StatKey; delta: number } | null> = {
+  spy: null, // informational only
+  sabotage: { stat: 'economy', delta: -3 },
+  assassinate: { stat: 'stability', delta: -5 },
+  cyber: { stat: 'military', delta: -3 },
+  economic: { stat: 'economy', delta: -5 },
+  information: null, // relation nudge only
+  proxy: null, // relation nudge only
+};
+
+/**
+ * The hostile-relations ladder, weakest to strongest (the `cap` in
+ * {@link shiftRelationTowardHostile} is a member of this array). A missing
+ * relation is treated as `neutral`; `war` is never touched (a war is its own
+ * entity — a covert op neither fabricates nor ends one).
+ */
+const HOSTILE_ORDER: RelationStatus[] = ['ally', 'neutral', 'rival', 'hostile'];
+
+/**
+ * (v0.34.0, #57) Nudge the bilateral relation between nations `a` and `b`
+ * ONE step toward hostile (on the {@link HOSTILE_ORDER} ladder), applied
+ * symmetrically to both directions, but never past the given `cap`. A missing
+ * relation starts at `neutral`; a `war` relation is never touched (a covert op
+ * neither fabricates nor downgrades a war). Used for the `information` /
+ * `proxy` nudge and the `exposed` backfire. Mutates `state` in place (the
+ * caller passes a clone — applyTurnResult does). Never throws.
+ */
+export function shiftRelationTowardHostile(
+  state: GameState,
+  a: string,
+  b: string,
+  cap: RelationStatus,
+): void {
+  const na = state.nations[a];
+  const nb = state.nations[b];
+  if (!na || !nb) return;
+  const capIdx = HOSTILE_ORDER.indexOf(cap);
+  if (capIdx === -1) return; // unknown cap: no-op (defensive)
+  const step = (cur: RelationStatus | undefined): RelationStatus | null => {
+    if (cur === 'war') return null; // never touch a war relation
+    const s = cur ?? 'neutral'; // a missing relation is neutral
+    const i = HOSTILE_ORDER.indexOf(s);
+    if (i === -1) return 'neutral'; // defensive: unknown status
+    if (i >= capIdx) return s; // at or past the cap: no change
+    return HOSTILE_ORDER[i + 1];
+  };
+  const stepA = step(na.relations[b]);
+  if (stepA !== null) na.relations[b] = stepA;
+  const stepB = step(nb.relations[a]);
+  if (stepB !== null) nb.relations[a] = stepB;
+}
+
+/**
+ * (v0.34.0, #57) The display name of a hybrid op's target for the auto
+ * backfire event: a nation target shows the nation's name; a region target
+ * shows the region's display name (or its raw key when no resolver is given).
+ */
+function hybridTargetDisplayName(
+  target: { nation: Nation; regionKey?: string },
+  regionNameOf?: (key: string) => string | undefined,
+): string {
+  if (target.regionKey !== undefined) return regionNameOf?.(target.regionKey) ?? target.regionKey;
+  return target.nation.name;
+}
+
+/**
+ * (v0.34.0, #57) Resolve a hybrid op's `target` (a nation id/name OR a region
+ * key) to the concrete nation it affects. A nation ref resolves via
+ * {@link resolveNationRef}; a region key resolves to its SOLE owner (a region
+ * with zero or 2+ owners can't be a valid target). Returns the affected nation
+ * plus the specific region key (when the target was a region) so a stat op can
+ * dip just that region. `validKeys` (when supplied) gates the region-key path;
+ * a ref that is neither a known nation nor a valid region key returns `null`.
+ * Pure; never throws.
+ */
+function resolveHybridTarget(
+  state: GameState,
+  target: string,
+  validKeys?: ReadonlySet<string>,
+): { nation: Nation; regionKey?: string } | null {
+  const id = resolveNationRef(state, target);
+  if (id !== null) {
+    const n = state.nations[id];
+    return n ? { nation: n } : null;
+  }
+  const key = (target ?? '').trim();
+  if (!key) return null;
+  if (validKeys && !validKeys.has(key)) return null; // unknown region key
+  const owners = Object.values(state.nations).filter((n) => n.territory.includes(key));
+  if (owners.length !== 1) return null; // 0 or 2+ owners: not a targetable region
+  return { nation: owners[0], regionKey: key };
+}
+
+/**
+ * (v0.34.0, #57) Apply ONE hybrid op to the (cloned) `next` state, mutating it
+ * in place and never throwing. Resolution order:
+ *   1. `from` must resolve to a nation (id first, then a unique name); an
+ *      unknown / ambiguous actor drops the op.
+ *   2. `target` must resolve to a concrete nation (a nation ref, or a region
+ *      key's sole owner); an unresolvable target drops the op.
+ *   3. A self-target (`from` and the target nation are the same) is dropped.
+ * Then, per type:
+ *   - an `exposed` op BACKFIRES: no intended effect, the actor's relation with
+ *     the target sours by one step (cap hostile), and an auto diplomatic/
+ *     failure event is emitted (the engine's own narrative, mirroring the
+ *     "repelled" auto-event).
+ *   - a successful stat op (sabotage / assassinate / cyber / economic) dips the
+ *     target's `regionStats` by its per-type delta — the single named region
+ *     when the target was a region key, otherwise every region the target
+ *     owns — then re-derives the target nation's aggregate stats (the same
+ *     0-region / no-entry guards as the main recompute), so the effect
+ *     persists to the next turn.
+ *   - a successful `information` / `proxy` sours the actor's relation with the
+ *     target by one step (information caps at hostile, proxy at rival).
+ *   - `spy` is informational only (no stat, no relation) — a no-op except for
+ *     whatever the LLM narrates in `events`.
+ */
+function applyHybridOp(
+  next: GameState,
+  op: HybridOp,
+  geoOf?: (key: string) => { area: number; population: number } | undefined,
+  validKeys?: ReadonlySet<string>,
+  regionNameOf?: (key: string) => string | undefined,
+  actingTurn?: number,
+): void {
+  const turn = actingTurn ?? next.turn;
+  const from = resolveNationRef(next, op.from);
+  if (from === null) return; // unknown / ambiguous actor: drop
+  const target = resolveHybridTarget(next, op.target, validKeys);
+  if (target === null) return; // unresolvable target: drop
+  if (from === target.nation.id) return; // self-target: drop
+
+  if (op.outcome === 'exposed') {
+    // Backfire: no intended effect; the relation sours by one step (cap
+    // hostile) and an auto diplomatic/failure event is emitted.
+    shiftRelationTowardHostile(next, from, target.nation.id, 'hostile');
+    const actor = next.nations[from];
+    next.events.push(
+      makeEvent(
+        {
+          nation: from,
+          type: 'diplomatic',
+          title: 'Covert operation exposed',
+          text: `${actor?.name ?? from}'s ${op.type} operation against ${hybridTargetDisplayName(
+            target,
+            regionNameOf,
+          )} was exposed; relations sour in response.`,
+          outcome: 'failure',
+        },
+        turn,
+        next.date,
+      ),
+    );
+    return;
+  }
+
+  const eff = HYBRID_STAT_EFFECT[op.type];
+  if (eff) {
+    // A successful stat op: dip the target's region(s), then re-derive.
+    if (target.regionKey !== undefined) {
+      const rs = next.regionStats[target.regionKey];
+      if (rs) rs[eff.stat] = clampStat(rs[eff.stat] + eff.delta);
+    } else {
+      for (const key of target.nation.territory) {
+        const rs = next.regionStats[key];
+        if (!rs) continue;
+        rs[eff.stat] = clampStat(rs[eff.stat] + eff.delta);
+      }
+    }
+    // Re-derive the target nation's aggregate stats. The main recompute ran
+    // earlier this turn, so without this the dip would be washed out next
+    // turn. The same 0-region / no-entry guards as the main recompute apply.
+    if (target.nation.territory.length > 0) {
+      if (target.nation.territory.some((k) => next.regionStats[k])) {
+        target.nation.stats = deriveNationStats(
+          target.nation.territory,
+          next.regionStats,
+          geoOf,
+        );
+      }
+    }
+    return;
+  }
+
+  // information / proxy: a relation nudge (spy is a pure no-op here).
+  if (op.type === 'information') shiftRelationTowardHostile(next, from, target.nation.id, 'hostile');
+  else if (op.type === 'proxy') shiftRelationTowardHostile(next, from, target.nation.id, 'rival');
+}
+
 /**
  * A turn is "meaningful" if it produced at least an event, a stat change, a
  * no-man's-land release (v0.22 #16 P3: a release is a real world change even
- * when the model under-narrates it), or a nation-op proposal (v0.23 #16 P4a:
+ * when the model under-narrates it), a nation-op proposal (v0.23 #16 P4a:
  * a proposal is a real pending-world change the player will resolve, so it
  * must survive the retry/nudge path instead of being replaced by the generic
- * "unclear" result).
+ * "unclear" result), a takeover (v0.32.0 #54), or a war declared or ended
+ * (v0.33.0 #56: both are real world changes).
  */
 export function isMeaningful(result: TurnResult): boolean {
   return (
@@ -754,6 +1184,12 @@ export function isMeaningful(result: TurnResult): boolean {
     result.nationOps.length > 0 ||
     // (v0.32.0, #54) a takeover is a real world change (regime rename).
     result.takeovers.length > 0 ||
+    // (v0.33.0, #56) a war declared or ended is a real world change.
+    result.wars.declare.length > 0 ||
+    result.wars.end.length > 0 ||
+    // (v0.34.0, #57) a hybrid op (a successful one or an exposed backfire) is
+    // a real world change (a stat dip or a soured relation).
+    result.hybrid.ops.length > 0 ||
     // (v0.27, #44) a marker placed or removed this turn is a real world change.
     result.markers.add.length > 0 ||
     result.markers.remove.length > 0
@@ -1030,6 +1466,86 @@ export function applyTurnResult(
     n.name = t.newRegimeName;
   }
 
+  // --- Wars (v0.33.0, #56: war as a first-class state) ---
+  //
+  // Wars are named entities (`GameState.wars`), not just a `war` relation.
+  // Two kinds of change, applied in this order against the (cloned) `next`:
+  //
+  //   (1) DECLARE: each validated declare is resolved to concrete nation ids
+  //       (`resolveNationRef`, so the LLM may use an id OR a unique name) and
+  //       becomes a new active War with a fresh id (`w1`, `w2`, ...), stamped
+  //       with the acting turn and the turn's new date. Mutually-defensive
+  //       allies of the belligerents are DRAGGED IN: any nation that is an
+  //       `ally` of a belligerent (per the current, post-relations relations)
+  //       and is not already a side joins the war's `sides`. A declare with
+  //       fewer than two DISTINCT resolved sides is dropped; a declare whose
+  //       sides are all already contained in an existing ACTIVE war is a
+  //       duplicate and is skipped (the existing war stands).
+  //   (2) END: each validated end targets a war BY ID and sets its status to
+  //       the outcome (`armistice` / `capitulation`). Only ACTIVE wars are
+  //       affected — an unknown id or an already-ended war is a no-op.
+  //       (A war may be declared and ended in the SAME turn: declare runs
+  //       first, so the just-declared id is a valid end target.)
+  //   (3) RELATION SYNC: `syncWarRelations` derives the `war` relation from
+  //       the ACTIVE wars (see its doc) — the load-bearing step that keeps
+  //       the relation table honest with the war entities.
+  //
+  // Nothing here throws: every unresolvable / invalid op is dropped.
+  {
+    const takenWarIds = new Set(next.wars.map((w) => w.id));
+    // (1) DECLARE
+    for (const d of result.wars.declare) {
+      const resolved: string[] = [];
+      for (const ref of d.sides) {
+        const id = resolveNationRef(next, ref);
+        if (id !== null && !resolved.includes(id)) resolved.push(id);
+      }
+      if (resolved.length < 2) continue; // not enough distinct belligerents
+      // Duplicate: an active war already contains all of these sides.
+      const dup = next.wars.some(
+        (w) => w.status === 'active' && resolved.every((s) => w.sides.includes(s)),
+      );
+      if (dup) continue;
+      const sides = [...resolved];
+      // Mutual-defense drag-in: each belligerent's allies join the war.
+      for (const s of resolved) {
+        const rels = next.nations[s]?.relations ?? {};
+        for (const ally of Object.keys(rels).filter((x) => rels[x] === 'ally').sort()) {
+          if (!sides.includes(ally)) sides.push(ally);
+        }
+      }
+      const id = freshWarId(takenWarIds);
+      takenWarIds.add(id);
+      next.wars.push({
+        id,
+        sides,
+        declaredTurn: actingTurn,
+        declaredDate: next.date,
+        ...(d.aims && d.aims.length > 0 ? { aims: d.aims } : {}),
+        status: 'active',
+      });
+    }
+    // (2) END (by id; active only)
+    for (const e of result.wars.end) {
+      const w = next.wars.find((w) => w.id === e.id && w.status === 'active');
+      if (!w) continue; // unknown id / already ended: no-op
+      w.status = e.outcome;
+    }
+    // (3) RELATION SYNC (prev = the state this turn started from)
+    syncWarRelations(state, next);
+  }
+
+  // --- Hybrid war (v0.34.0, #57): covert, non-kinetic operations ---
+  //
+  // Each op's deterministic per-type effect is applied to the post-wars state
+  // (a stat dip on the target's regions, or a relation nudge toward hostile).
+  // An `exposed` op backfires instead. They never move borders, so they run
+  // after the territory changes and before the markers (a marker's owner is
+  // recomputed from the final ownership, unaffected by a stat/relation dip).
+  for (const op of result.hybrid.ops) {
+    applyHybridOp(next, op, geoOf, validKeys, regionNameOf, actingTurn);
+  }
+
   // --- Map markers (v0.27, #44) ---
   //
   // Markers are anchored to a region key and are processed against the FINAL
@@ -1215,6 +1731,19 @@ export function freshNationId(taken: ReadonlySet<string>): string {
   }
 }
 
+/**
+ * Generate a fresh, unused war id for a newly declared {@link War}
+ * (v0.33.0, #56): the first free of `w1`, `w2`, ... The caller adds each id
+ * it creates to `taken` as it goes, so ids stay unique within a turn and are
+ * never reused from an ended war.
+ */
+export function freshWarId(taken: ReadonlySet<string>): string {
+  for (let i = 1; ; i++) {
+    const id = `w${i}`;
+    if (!taken.has(id)) return id;
+  }
+}
+
 /**
  * The baseline stats a freshly created nation starts with (split parts,
  * v0.24 #16 P4b) — a neutral middle-of-the-scale profile, deliberately NOT
@@ -1505,7 +2034,31 @@ export async function runTurn(opts: {
     regionKey: m.regionKey,
     owner: m.ownerNationId,
   }));
-  const user = buildUserPrompt(state, items, regionReference, frontier, state.unowned ?? [], markers);
+  // (v0.33.0, #56) the model always sees the currently ACTIVE wars (empty list
+  // on a fresh game) so it can reference a war's id in `wars.end` to end it;
+  // ended wars stay in GameState.wars as history but are not surfaced.
+  const wars = (state.wars ?? [])
+    .filter((w) => w.status === 'active')
+    .map((w) => ({
+      id: w.id,
+      sides: w.sides,
+      ...(w.aims && w.aims.length > 0 ? { aims: w.aims } : {}),
+    }));
+  // (v0.34.0, #57) the player's current relations to every other nation, so the
+  // model sees the ladder its `information` / `proxy` hybrid ops nudge (and the
+  // starting point before an `exposed` backfire). Always emitted ({} when the
+  // player has no relations yet).
+  const relations = state.nations[state.playerNationId]?.relations ?? {};
+  const user = buildUserPrompt(
+    state,
+    items,
+    regionReference,
+    frontier,
+    state.unowned ?? [],
+    markers,
+    wars,
+    relations,
+  );
   const nudge =
     'Your previous output was invalid. Respond with ONLY the JSON object described above for this batch — no prose, no fences.';
 
@@ -1558,6 +2111,8 @@ export async function runTurn(opts: {
       unowned: [],
       nationOps: [],
       takeovers: [],
+      wars: { declare: [], end: [] },
+      hybrid: { ops: [] },
       markers: { add: [], remove: [] },
     };
   }

server/index.ts

@@ -178,6 +178,10 @@ app.post('/api/games', (req: Request, res) => {
     // real region keys — each region starts as a copy of its owner's initial
     // stats, so the derived aggregate equals the nation's stats on turn 1.
     regionStats: {},
+    // (v0.33.0, #56) No wars on a fresh game; the LLM declares them over time
+    // (any preset-level `war` relation stays a bare/implicit one, not an
+    // entity — see syncWarRelations).
+    wars: [],
   };
   // v0.19 (#16 P1): expand each nation's ISO_A3 country list to the region keys
   // of all its sub-regions, so ownership (and the choropleth) is at region

server/map.test.ts

@@ -52,6 +52,8 @@ const state: GameState = {
   markers: [],
   // (v0.28, #46) the map is ownership-only; the region-stats table is unused here.
   regionStats: {},
+  // (v0.33.0, #56) the map is ownership-only; wars are unused here.
+  wars: [],
 };
 
 // The RegionMap built from the fixture: the authoritative region key → owner source.

server/protocol.test.ts

@@ -41,6 +41,7 @@ function mkState(
     pendingNationOps: [],
     markers: [],
     regionStats: {} as GameState['regionStats'],
+    wars: [],
   };
   if (regionStats !== undefined) base.regionStats = regionStats as GameState['regionStats'];
   else delete (base as { regionStats?: unknown }).regionStats; // simulate a legacy row

server/regions.test.ts

@@ -110,6 +110,7 @@ describe('regions (region-key loader, v0.19 #16 P1)', () => {
       markers: [],
       // (v0.28, #46) unused here (allTerritoryKeys is ownership-only).
       regionStats: {},
+      wars: [],
     };
     // 16 GBR sub-regions + 42 FRA sub-regions, no overlap between the two.
     expect(allTerritoryKeys(state).size).toBe(58);

server/stub.ts

@@ -3,7 +3,7 @@
 // speaks the OpenAI chat-completions wire shape, so it plugs in as a Transport.
 import { createLlmClient, type ChatMessage, type Transport } from './llm';
 import { addDays, clampStat, type TerritoryChange, type TurnEventSpec, type TurnResult } from '../shared/protocol';
-import type { RoundItem } from '../shared/types';
+import type { RelationStatus, RoundItem } from '../shared/types';
 
 export const ADVISOR_MARKER = '<<ADVISOR>>';
 
@@ -38,6 +38,20 @@ export interface StubPayload {
    * path is covered by unit tests against applyTurnResult, not by the stub.
    */
   markers?: Array<{ id: string; type: string; name: string; regionKey: string; owner: string | null }>;
+  /**
+   * (v0.33.0, #56) Currently ACTIVE wars (`{ id, sides, aims? }`), emitted by
+   * the engine's user prompt. The stub ignores this field — it never declares
+   * or ends a war (see deterministicTurn), so the war path is covered by unit
+   * tests against applyTurnResult, not by the stub.
+   */
+  wars?: Array<{ id: string; sides: string[]; aims?: string[] }>;
+  /**
+   * (v0.34.0, #57) The player's current relations to every other nation
+   * (`nationId -> RelationStatus`), emitted by the engine's user prompt. The
+   * stub ignores this field — it never runs a hybrid op (see
+   * deterministicTurn), so the hybrid path is covered by unit tests only.
+   */
+  relations?: Record<string, RelationStatus>;
 }
 
 function truncate(s: string, n: number): string {
@@ -81,6 +95,14 @@ function truncate(s: string, n: number): string {
  * (v0.32.0, #54) The stub likewise NEVER performs a non-military takeover
  * (it always returns `takeovers: []`) — the regime-rename path is covered by
  * unit tests only, keeping the e2e assertions untouched.
+ *
+ * (v0.33.0, #56) The stub likewise NEVER declares or ends a war (it always
+ * returns `wars: { declare: [], end: [] }`) — the war path is covered by unit
+ * tests only, keeping the e2e assertions untouched.
+ *
+ * (v0.34.0, #57) The stub likewise NEVER runs a hybrid / covert op (it always
+ * returns `hybrid: { ops: [] }`) — the hybrid-war path is covered by unit
+ * tests only, keeping the e2e assertions untouched.
  */
 export function deterministicTurn(p: StubPayload): TurnResult {
   const state = p.state;
@@ -148,6 +170,8 @@ export function deterministicTurn(p: StubPayload): TurnResult {
     unowned: [], // the stub never creates no man's land (see doc comment)
     nationOps: [], // the stub never proposes nation ops (see doc comment)
     takeovers: [], // the stub never performs a takeover (see doc comment)
+    wars: { declare: [], end: [] }, // the stub never declares/ends wars (see doc comment)
+    hybrid: { ops: [] }, // the stub never runs hybrid/covert ops (see doc comment)
     markers: { add: [], remove: [] }, // the stub never touches markers (see doc comment)
   };
 }

shared/protocol.ts

@@ -51,6 +51,50 @@ export interface TerritoryChange {
   lose: string[];
 }
 
+// --- Wars (v0.33.0, #56): first-class war entities --------------------------
+//
+// A war is an explicit world entity (see types.War), not just a boolean
+// relation: the LLM DECLARES one (it becomes an active War with an engine-
+// assigned stable id) and later ENDS one by id. The engine derives the
+// symmetric `war` relation from the set of ACTIVE wars on every turn (two
+// nations are in a `war` relation iff they are both belligerents of at least
+// one active war) and downgrades it to `hostile` when the war ends — so the
+// relation can never drift apart from the war state.
+
+/**
+ * (v0.33.0, #56) Declare a new war this turn. `sides` names every
+ * belligerent (≥2 distinct nation ids — each resolved by `resolveNationRef`,
+ * so a unique name also works); any two distinct sides end the turn in a
+ * `war` relation. `aims` are optional short war aims for the narrative. The
+ * engine assigns the stable id, the declared turn/date, drags in the
+ * mutual-defense allies of each belligerent, and drops the op when fewer than
+ * two sides resolve (never throws).
+ */
+export interface WarDeclareOp {
+  /** The belligerent nations (≥2 distinct ids or unique names). */
+  sides: string[];
+  /** Optional short war aims (narrative only). */
+  aims?: string[];
+}
+
+/**
+ * (v0.33.0, #56) End an existing ACTIVE war this turn, by its stable id
+ * (the model sees the current active wars in the user payload's `wars`
+ * field). `outcome` is `armistice` (a negotiated stop) or `capitulation`
+ * (one side surrenders); the war keeps its id/sides and stays in
+ * `GameState.wars` as history with that final status. `terms` are optional
+ * short peace terms (narrative only — the model narrates the details in
+ * `events`). An unknown id, or a war that is already ended, is a no-op
+ * (never throws).
+ */
+export interface WarEndOp {
+  /** The stable id of the active war to end (from the user payload). */
+  id: string;
+  outcome: 'armistice' | 'capitulation';
+  /** Optional short peace terms (narrative only). */
+  terms?: string;
+}
+
 // --- Nation-management ops (v0.23 P4a / v0.24 P4b, #16: the hybrid "LLM
 // proposes, player confirms" model) ------------------------------------------
 //
@@ -189,6 +233,19 @@ export interface MarkerInfo {
   owner: string | null;
 }
 
+/**
+ * (v0.33.0, #56) One ACTIVE war as the user prompt presents it to the model
+ * (the `wars` field): its stable `id` (the model echoes it in
+ * `TurnResult.wars.end` to end it), its belligerent `sides`, and optional
+ * short `aims`. Ended wars are NOT included — they stay in `GameState.wars`
+ * as history, but the model only needs to address the active ones.
+ */
+export interface WarInfo {
+  id: string;
+  sides: string[];
+  aims?: string[];
+}
+
 /**
  * The per-type presentation of a marker: its display label (tooltip /
  * dossier), its color, and its glyph character (a symbol-layer text label —
@@ -248,6 +305,64 @@ export function isMarkerType(t: unknown): t is MarkerType {
   return typeof t === 'string' && (MARKER_TYPES as string[]).includes(t);
 }
 
+// --- Hybrid war (v0.34.0, #57): non-kinetic operations ---------------------
+//
+// The LLM can run covert, non-kinetic operations each turn: spies, sabotage,
+// assassination attempts, cyberattacks, economic pressure, disinformation, and
+// proxy support. Unlike kinetic conquest (territory), these never move borders
+// — they dip a target's stats or sour a relation. The LLM declares an intended
+// outcome (`success` or `exposed`); the engine is fully deterministic (no
+// hidden randomness) and acts only on that declaration: a successful op applies
+// its small, per-type effect, and an "exposed" op backfires (no effect, the
+// actor's relation with the target sours by one step, and an auto
+// diplomatic/failure event is emitted).
+
+/** The seven covert operation types a hybrid op can be (v0.34.0, #57). */
+export type HybridOpType =
+  | 'spy'
+  | 'sabotage'
+  | 'assassinate'
+  | 'cyber'
+  | 'economic'
+  | 'information'
+  | 'proxy';
+
+/** All hybrid op types, in canonical order (the LLM contract + validator). */
+export const HYBRID_OP_TYPES: HybridOpType[] = [
+  'spy',
+  'sabotage',
+  'assassinate',
+  'cyber',
+  'economic',
+  'information',
+  'proxy',
+];
+
+/**
+ * (v0.34.0, #57) One covert (non-kinetic) operation the LLM may run in a turn
+ * (`TurnResult.hybrid.ops`). `from` is the acting nation (its id or a unique
+ * name — resolved by the engine's `resolveNationRef`). `target` is EITHER a
+ * nation (id or unique name) or a specific region key; a region target hits
+ * that region's sole owner (a region with zero or 2+ owners can't be
+ * targeted). `detail` is a short narrative (the LLM may also narrate the op in
+ * `events`). `outcome` is the LLM's *intended* result: `success` (the default)
+ * applies the op's deterministic effect, while `exposed` backfires it — the
+ * effect is NOT applied, the actor's relation with the target sours by one
+ * step, and an auto diplomatic/failure event is emitted.
+ */
+export interface HybridOp {
+  /** The operation type (see {@link HybridOpType}). */
+  type: HybridOpType;
+  /** The acting nation's id or current name (resolved by `resolveNationRef`). */
+  from: string;
+  /** The target: a nation's id/name, or a region key (hits that region's owner). */
+  target: string;
+  /** Short narrative detail (optional). */
+  detail?: string;
+  /** The LLM's intended result; `exposed` backfires the op. Defaults to success. */
+  outcome?: 'success' | 'exposed';
+}
+
 /** The exact JSON object the world-simulator LLM must return for a turn. */
 export interface TurnResult {
   date: string;
@@ -297,6 +412,33 @@ export interface TurnResult {
    * (default `[]`).
    */
   takeovers: Takeover[];
+  /**
+   * (v0.33.0, #56) First-class war ops this turn. `declare` starts new wars
+   * (≥2 distinct belligerents; the engine assigns the stable id / declared
+   * turn / date, drags in the mutual-defense allies of each belligerent, and
+   * derives the `war` relation between every pair of sides); `end` closes an
+   * existing ACTIVE war by id (`armistice` = a negotiated stop,
+   * `capitulation` = one side surrenders), which downgrades the relation the
+   * war had derived back to `hostile`. Both sub-fields are required and
+   * default `[]`; a turn with war activity is meaningful (it survives the
+   * retry/nudge path). The model sees the current active wars (id + sides +
+   * aims) in the user payload's `wars` field, so it can `end` by id.
+   */
+  wars: { declare: WarDeclareOp[]; end: WarEndOp[] };
+  /**
+   * (v0.34.0, #57) Covert, non-kinetic ("hybrid") operations this turn — spies,
+   * sabotage, assassination attempts, cyberattacks, economic pressure,
+   * disinformation, and proxy support. Unlike `nationOps` (player-confirmed
+   * proposals) these are applied directly by the turn, like territory / state
+   * changes: each op's deterministic per-type effect is applied by
+   * `applyTurnResult` (a small stat dip on the target's regions, or a
+   * relation nudge toward hostile for `information` / `proxy`); an `exposed`
+   * op backfires (no effect, the actor's relation with the target sours by
+   * one step, and an auto diplomatic/failure event is emitted). They never
+   * move borders — sole ownership is untouched. Required (default
+   * `{ ops: [] }`).
+   */
+  hybrid: { ops: HybridOp[] };
   /**
    * (v0.27, #44) Map-marker ops this turn. `add` places new named markers
    * (the engine positions each at its region's centroid + a small

shared/scenario.ts

@@ -453,6 +453,9 @@ export function worldDraftToGameState(
     // Seed AFTER the territories hold real region keys: each owned region gets
     // a copy of its owner's stats, so the derived aggregate equals them.
     regionStats: seedRegionStats(nations),
+    // (v0.33.0, #56) No wars on a fresh (dynamic) game; any preset-level `war`
+    // relation stays a bare/implicit one, not an entity (see syncWarRelations).
+    wars: [],
   };
   return state;
 }

shared/types.ts

@@ -143,6 +143,44 @@ export interface MapMarker {
   placedTurn: number;
 }
 
+/**
+ * (v0.33.0, #56) A first-class **war** — a named, ongoing armed conflict
+ * between two or more nations. Unlike the boolean `war` relation (which a
+ * preset or a bare `relations` entry can carry without any structure behind
+ * it), a `War` is an explicit world entity the engine tracks: the LLM
+ * declares it (a turn's `wars.declare`), the engine assigns it a stable id
+ * (`w1`, `w2`, …), the turn and in-game date it was declared, and it stays in
+ * `GameState.wars` for the rest of the game as history after it ends.
+ *
+ * **Semantics of `sides`:** the flat list of every belligerent in the war
+ * (≥2 distinct nation ids). Any two distinct members of an active war's
+ * sides are at war with each other — the engine derives the symmetric `war`
+ * relation for every such pair on every turn (see the engine's
+ * `syncWarRelations`), and downgrades the relation to `hostile` when the war
+ * ends. A mutual-defense ally of a belligerent is dragged into the same list
+ * on declaration (the engine does this automatically).
+ */
+export interface War {
+  /** Stable id (`w1`, `w2`, …) assigned by the engine on declaration. */
+  id: string;
+  /** The belligerent nations (≥2 distinct ids; see the note above). */
+  sides: string[];
+  /** The turn number the war was declared on. */
+  declaredTurn: number;
+  /** The in-game date the war took effect on (`YYYY-MM-DD`). */
+  declaredDate?: string;
+  /** Optional short war aims the LLM names at declaration. */
+  aims?: string[];
+  /**
+   * Lifecycle: `active` until a turn ends the war, then the end outcome —
+   * `armistice` (a negotiated stop) or `capitulation` (one side surrenders) —
+   * or the generic `ended`. Ended wars stay in `GameState.wars` as history.
+   */
+  status: 'active' | 'armistice' | 'capitulation' | 'ended';
+  /** Optional region keys the war's front line runs through (reserved). */
+  front?: string[];
+}
+
 /** A narrative event the simulation emitted on some turn. */
 export interface EventEntry {
   id: string;
@@ -251,6 +289,20 @@ export interface GameState {
    * — additive, no data wipe.
    */
   regionStats: Record<string, RegionStats>;
+  /**
+   * (v0.33.0, #56) First-class war entities (see {@link War}): declared by a
+   * turn's `wars.declare` (the engine assigns the id, declared turn/date, and
+   * drags in mutual-defense allies) and closed by `wars.end` (armistice /
+   * capitulation). The engine DERIVES the `war` relation from the ACTIVE
+   * wars on every turn (two nations are in a `war` relation iff they are both
+   * belligerents of ≥1 active war), so the boolean relation and the war
+   * entities can never drift apart. A bare preset `war` relation with no
+   * active war behind it stays as-is (an implicit, unnamed war — never
+   * force-created into an entity). Always `[]` on a fresh game; rows saved
+   * before v0.33 load with this defaulted to `[]` (see server/db.ts) —
+   * additive, no data wipe.
+   */
+  wars: War[];
 }
 
 /**