- PRTS 卫戍协议：盟约 下半/PRTS盟约记录, 阿戈尔 ※: 「该付与来源获得所有标记单位的基础攻击力（最终加算）和阻挡数」. New
additive mod key atkFinal (PRTS 游戏数据基础 A_f = F_t[(A + D_p)(1 + D_t) + F_p]):
ATK = ((base + ΣatkFlat) × (1 + ΣatkPct) + ΣatkFinal) × ΠatkMul. The devour buff carries atkFinal instead of atkFlat:
1000 base ATK, +100 % skill, +2000 devoured → 4000 (was 6000). Every other mod key keeps its meaning.
- 「标记按付与顺序触发【吞噬】效果，目标首次被击倒后解除自身被付与但还未触发的【吞噬】效果」 cancels only the knocked-out
target's marks: a marker knocked out earlier in the pass now resolves its own marks, the kill credited to it
(「击杀来源始终为对应标记的付与来源」). A → B → C with B devoured first: C takes B's 5000 and the kill is B's. The 3
revive slots, the 联防 front-tile devour and the down-at-start rules (DESIGN §24.3 / §24.4) are unchanged.
- Point 5 (the 流失's own source is the target itself) stays open: the hooks still see the marker (DESIGN §24.7).
- Tests: bonds_core — new 最终加算 test (4000; a 直接加算 still scaled; ×0.5 scales the whole), new A → B → C test,
the dead-marker test rewritten (its mark now lands); feedback3-egir-down, feedback4-egir-unite, feedback3-ulpia-move
read atkFinal; match/bosshp recomputes ATK with the 最终加算.
- Docs: DESIGN §24.7 (and the §24 intro, §5.2 / §5.3, the §24.4 atkFlat lines marked), SIM.md §3, research 02 §3.6
and its JSON spec; data/bonds.json regenerated with build-data --offline — only egirShip.spec.algorithm[2] / [3]
changed.
- Golden: bond-egirShip-high (the devourers 深巡 / 幽灵鲨 / 斯卡蒂 no longer scale the devoured ATK by their ATK +%; the
fight ends at 89.03 s instead of 74.03 s), coop2-ABYSS-10-boosted (from round 8 ai_0 has 3 more 阿戈尔 and 奥术
layers — 海霓's first-kill trait — and both bots' boss damage moves). The marker rule moves no scenario.
Battle.leak emits its own 'leak' event, which no sound handled, so every leak was silent; the client now plays battle.ON_ENEMY_REACHED_EXIT (b_ui_alarmenter) with the bank's own one-at-a-time rule, leaks closer than the sound's length sharing one alarm. The manifest gains only audio.sfx.battle.leak. Contributed by @Convey123.
Deploy, skill and result lines come from the operators' official voice slots; a newer line takes the channel over (stale callbacks of the replaced line no longer release it), and the result line is spoken by an operator of the player's own battle. Contributed by @Convey123.
The owner's decision of 2026-10-05 (GitHub #140 comment 4, #105): PRTS 「战斗开始时依次吞噬身前一格干员」 has no
own-side limit. The devour's front-tile lookup took only the player's own operators.
- devour opAt: the operator on the front tile whoever owns it — a living one, else one lying there since the 联防
start (carry.down, forced out). The marker gains its base ATK and block count exactly as in a solo battle; a mark
on a standing teammate's operator resolves (5000 物理流失 less DEF, credited to the marker, the knock-out is its
owner's: their revive slots), a mark on a down one resolves nothing (it stays forced out, §19.3 kept).
- The chain goes on through a marked 阿戈尔 by S.isMember: for own operators exactly the members, as before; a
teammate's by its own bonds ([ASSUMED]).
- 联防 still adds no layers (layerGainsEnabled false, unchanged).
- On the current maps the two 联防 helpers stand on board cols 3-9 and 11-17 with the road column 10 between them,
so a teammate's operator is never in front at the battle start; the tests place one on col 10. The own-board case
(food down since the own combat) is verified through unite.js's real carry: 联防 gives exactly the own combat's
atkFlat / blockCnt.
- research 02 §3.6 and its JSON spec; data/bonds.json regenerated offline (only egirShip.spec.assumed[1] changed).
- Tests: test/content/feedback4-egir-unite.test.js (teammate's living and down operator, solo parity of ATK / block /
物理流失, no layers, the chain through a teammate's 阿戈尔 and her owner's 5-tier slot, the own-board case through
uniteBattleOpts). The teammate cases fail on the previous code.
Golden: no scenario moved (no corpus field puts a teammate's operator in front of an 阿戈尔).
The owner's decision of 2026-10-05, following players' reports (GitHub #105 comment 2, #140 comment 3 and a report
the owner quoted): 0.1.3 fixed the 3 holders at battle start by position, so uneaten front members held slots they
never used and whatever the food chain ate only 3 阿戈尔 stood. Now:
- A member operator's first knock-out ('killed') of the battle takes a slot while fewer than 3 are spent, the
battle-start devour included; it redeploys at once (free) where it lies. A member never knocked out takes none, so
survivors = uneaten + 3. Later knock-outs never take one; a first knock-out after the 3 are gone stays down.
- One slot at most per member however many 阿戈尔 mark it: the devour still cancels a target's pending marks once
it is knocked out. The report that each devour of the same operator wastes a revive was checked and not
reproduced on 0.1.3 (the extra marks were already cancelled); the fixed holders were what players saw.
- An operator's own save uses no slot: every kit / item / band saver is a fatal hook (斯卡蒂's DRE-Y, M3茧甲,
坚固维式重锤, 埃芒加德) and runs before the death hook, so the unit is not knocked out; the 0.1.3 rule that an
in-place save spends the holder's revive is gone ([ASSUMED] for 埃芒加德, a band, treated like M3茧甲).
- 联防: a member entering down is not knocked out (no slot) until it stands up and falls again. 埃芒加德's own
"first 3 knock-downs" rule is a separate hook and is unchanged.
- research 02 §3.6 and its JSON spec; data/bonds.json regenerated offline (only egirShip.spec.tiers[2].effect
changed).
- Tests (bonds_core): knock-out order wherever members stand; the devour chain with 6 members (survivors = 2 + 3,
normal and 联防); a member marked three times spends one slot; 斯卡蒂's DRE-Y, M3茧甲 and 埃芒加德 saves use none;
联防 / 调和 / mirrored boss player. All five fail on 0.1.3's code.
Golden: no scenario moved (no corpus battle has a 5-阿戈尔 knock-out whose slot differs; bond-egirShip-high runs at
999 layers and nobody falls to the devour).
The 联防 phase plays m_bat_corrosion (intro + loop), the track of the official escaped stages, instead of carrying on with the round's battle music; without the downloaded track it falls back to the battle music. Asset plan, manifest entry (audio.bgm.unite), audio.js and tests. Contributed by @Convey123.
The trait line is on every hookmaster and is not trusted. Placement is the loadout at place time: golden 歌蕾蒂娅 carrying uniequip_003_glady may use a ranged tile; everyone else stays on the ground.
§22.2: seen / official / cause / now / cost / players notice / [ASSUMED] / tests for GitHub issue #51. Research 02
§3.18's [ASSUMED] landing line (md and json) now states the rule; `node tools/build-data.mjs --offline` changed only
data/bonds.json raidShip.spec.assumed[0].
The owner's decision (2026-10-03, after community feedback): 深巡 S2 行动能力剥夺, 雷蛇 S2 反击电弧, 号角 S2 暴风号令
and S3 终极防线, 灰毫 S1 攻击力强化·γ型 and S2 专注轰击 take the basic strategy (DEFAULT) instead of the 下半 TANK class
row's TAKE_DAMAGE. Every other MANUAL 重装 skill, 深巡 S1 included, still waits for a hit.
- tools/build-data.mjs TRIGGER_DEVIATIONS: a small table keyed by the normal chess id and the skill id (the elite
follows through baseId), never by the skill id alone - 灰毫 S1 is the generic skcom_atk_up[3]. resolveTrigger applies
it last; rawRule keeps the official TAKE_DAMAGE; validateAll fails the build when an entry no longer meets a
TAKE_DAMAGE row. data/chess.json rebuilt with --offline: exactly the 12 records' rule changes.
- PR #12's kit lines (trigger 'DEFAULT' on 深巡 / 雷蛇 S2) stay and agree with the data; their comments and tests now
give this reason instead of the community summary.
- Tests: test/sim/feedback1-tank-triggers.test.js (the data, kit = data, a real battle on 战场#01(下半) in which each of
the six casts with an enemy in range before any hit and each of the other 46 MANUAL 重装 records casts on the first
hit); data.test.js trigger expectations; the 灰毫 / 号角 kit tests follow; docs-consistency.
- Docs: DESIGN §21.29 (new), §5.6, §19.2, §20.2, §21.20 (settled row), §21.23; DATA, SIM, PLAYING, CHANGELOG 0.1.1;
research 03 §1.4 drops the arknights.wiki.gg "offensive / defensive" summary and records the activity_table history
(上半 had no TANK row, 下半 added it) and this deviation; research 00-INDEX.
Conflicts resolved keeping both sides:
- effectsMeta.js header: the QA residual's "every [holder, item] pair, taken before the first item runs" and "an item
moved during it runs at most once", with the cell branch's wording of what 突变细胞 does mid-walk.
- DESIGN §6.6: the band pick that skips a strategy tied to a switched-off bond and the deployment of the operator a
transformation gains / the bench shed that spares it.
- DESIGN §21.25: PR #2's bench rule credited (the cell branch) plus the QA residual's snapshot-timing sentence.
- PLAYING: the new 突变细胞 sentence with the residual's hammer sentence (revives, borrowed hammers, the lend window).
- docs-consistency: both new tests (batch 6 QA residuals; 突变细胞 after the WA merge).
The band -> bond derivation moves out of GameData into shared/bandBonds.js,
which tools/build-data.mjs runs per band to write bands.json bondIds
(regenerated with --offline; 9 bands are tied, the build checks the ids).
GameData.bandBondIds reads the field, so the server and the client share
one source. DATA.md documents the field; the test checks the helper on
synthetic bands and that the data matches it.
Official gameplay footage (bilibili BV1vzyVBuEN9 at about 8:24, BV1Qkw1zMEoR at about 7:25) shows the
next prep with the carrier's tile empty, one more deployment left, and the returned equipment plus the
new tier+1 operator on the bench, deployed by hand afterwards. That is the PRTS 备注 read as written
("生效时，原干员销毁，获得一名高一阶的随机初始干员"): a destroy followed by a gain. The closed PR #2 made
the same claim from a playtest.
- PlayerState.transformChess: the carrier is destroyed wherever it stands (board tile freed, summons
removed, copies back to the pool); its equipment, the cell included, returns to the hand first
(overflow temp, auto-merge); then the new operator is gained through acquireChess like any gained
operator (hand, overflow temp, the no-room rule of every gain; no summon card until it is deployed).
The returnItems option is gone (the cell's built-in was its only caller).
- A merge the gain completes merges like any gained copy: the elite takes a consumed deployed copy's
tile, else the hand. _mergeChess loses fromKey / fromDir, so the destroyed carrier's tile never
counts (PR #2 had kept that path; not adopted); audit.js drops its exception for it.
- builtinMeta, ctx.transform, items/meta.js and bot.js comments follow. Bots and AI 托管 deploy the
gained operator in their normal lineup step (no bot change needed; a real co-op round pins it).
- ITEM_RULES note and implFormula for the item card; data/items.json rebuilt with build-data --offline.
- Tests: facing, merge, playtest6-elite-to-board and items no longer pin the kept tile; feedback1-meta
covers a deployed and a bench carrier, hand full -> temp, hand and temp full, a merge with and
without a deployed copy (never the carrier's tile), 商业包装方案's count on the returned item, and bots /
AI 托管 redeploying; feedback1-cell checks the new note.
Merges the 7 WB commits after the first WB merge (d1f8a49..2cb84bd): R3 one of 10 fixed
sets (9 seen), R9 one of 6 boss groups picked by the matches they came in, R11 one of the
7 seen lists of 7; R11 families 14 / 4 / 4; the 机密商店 and the 战术决策 can offer the
same card twice (R11 shop: two VI, at least one V and one 盟约之币). data/choices.json
rebuilt with tools/build-data.mjs --offline (identical to the branch's). The tactic bot
test uses a real seat (the merged bot scores tactic cards on the player's bonds).
DESIGN §21.2 / §7 / §21.20 and CHANGELOG describe the 22-match model.
The user: "战术决策也按官方改成可以重复吧". The 4 official 战术决策 (R11 of
matches 7, 9, 18 and 20) show six cards each; match 7 offers 补给 twice.
- Every 战术决策 card is drawn on its own (with replacement), so the same
card can come twice. choices.js tacticDraftCards / tacticCard.
- R11 (绝境 / 终极, the round of the screenshots; choices.json tacticDraft,
build-data TACTIC_DRAFT): only the ally cards, weighted 1 + the official
cards each showed on. All 24 official cards are ally cards (no 排斥 /
责罚 / 裁决, no terrain card: about 0.04 % for the previous draw) and
列装 / 财富 / 补给 / 整备 / 升华 made 12 of them (uniform: about 4.6).
Other rounds (标准 / 险境, no screenshot) keep every card, uniform, with
replacement too. The weights and the R11 terrain exclusion are assumed.
- The header is the official "进行协同调整，做好迎战准备。".
- Identical cards are separate cards: own index, own object; a pick
applies its card once and leaves the twin pickable; two picks of a team
card stack. Match / bot / overlay already go by index.
- Tests: official drafts vs data, generated R11 / solo / earlier rounds,
a seeded draft with 补给 twice, co-op and solo E2E picks of twins, the
bot, the overlay, and a headless-Chrome render (game-mock variant=tactic
is now the official match 7 draft). Tests that drive given tactic cards
lay them out explicitly. The 奥术 Final Assault test takes the first of
a few seeds whose bots pair two 奥术 players (the draws moved seed 12).
The 7 seen R11 lists stay picked uniformly, assumed because that invents no list. By the blocks of
effectChoiceInfoDict, R11 would fire bounty_hunter_8..15 (beside artifact_paid_4/5 and hardbuff_select),
so the block order is no longer cited as support for '1..7'. 7, 8 or 9 events fit the 14 drafts about
equally; with 8, an unseen list would come in about 1 R11 bounty draft in 8 (open question).
- R11 悬赏决策: the list is one of the 7 seen lists, uniform (slots 7, taken as the events
bounty_hunter_1..7, one block of effectChoiceInfoDict). No list is built from nothing any more;
the rule only completes the 4 lists seen with 6 cards, so 12_8 / 16_2 / 16_6 come in about
1 R11 bounty draft in 15 (14 drafts without one: about 39 %).
- R9: the groups hold 9, 9, 9, 9, 8 and 6 cards (comments, docs, test); match 4's base-only
group may be a draft of another group (open). 58 -> 59 bounty drafts in the build-data note.
- 机密商店: shopDraft.rounds [11]. The official composition (VI, VI, V, 盟约之币 + 2) applies at
R11 (绝境 / 终极, where the screenshots are); the earlier 标准 / 险境 shops keep the previous draw,
any tier I-VI per card with replacement (repeats still possible, about 26 %).
- Tests: R11 slots / block order / never-shown rate, R9 group sizes, the non-R11 shop draw.
Bounty drafts: every 悬赏决策 event is a fixed card list and the draft shows
6 different cards of it, drawn with weight 1 + the official drafts of its
group the card showed in.
- R3: 9 of the 10 fixed sets seen (two new: 20_1 + 20_5, and one with 17_6);
3 x I + 2 x II + 1 x III in 22 of 22. The unseen 10th set is built by a
rule whose III is 10_6 or 20_6, the only two-battle cards no draft showed.
- R9: 6 groups of 9 cards, one per bossInitial event. The 鼠王 group came in
14 of 22 matches (every grey-board match among them), so groups are picked
by the matches they came in. Named bosses no longer always come together.
- R11: lists of 7 cards, the draft leaves out one (matches 1 / 6 / 21 / 22
leave out four different cards of one list); 7 of 15 lists seen, the rest
and the 4 partial lists' 7th cards built by rule. Families 14 / 4 / 4.
机密商店 (the user: "机密商店按官方改成可以重复吧"): choices.json shopDraft,
six slots drawn with replacement - VI, VI, V, 盟约之币 and twice
V / IV / III / 盟约之币, items weighted by the official cards they showed on.
A repeated item now comes in about 61 % of shops (official 3 of 4). Two
identical cards stay two cards (picks by index): E2E, overlay and a
headless-Chrome test (game-mock ?variant=shop, the official match 8 shop).
data/assets.json pointed 灼热源石虫 / 炽焰源石虫 at /assets/local/ files that only exist
where tools/local-extract ran, with no fallback: without the extraction the
asset tests failed, setup re-ran fetch-assets (rewriting the tracked manifest)
and the slugs drew as icon diamonds.
- The web model stays the 源石虫 alias (spine + spineAliasOf). The official
model is enemies[id].spineLocal: file names in the data/local-assets.json
group spine/enemy/<id> plus the parsed metadata, from the committed
tools/assets/local-enemy-spines.json (fetch-assets --local-spines refreshes
it from the extracted files, read only). The manifest is byte-identical with
or without the extraction and has no /assets/local/ URL.
- Client: assets.js spineEntry uses spineLocal only when the local manifest
lists every file (createFieldView awaits assets.local() with the manifest);
UnitView loads the web model when the local one fails.
- extract.py writes the atlas with size: / pma: true itself (the same text as
atlas.mjs normalizeAtlas), so the extraction is usable without a fetch pass.
- setup notes when an older extraction lacks the enemy models.
sganggs