battle-log-analysis

name	battle-log-analysis
description	Produces a round-by-round (回合) analysis of a finished battle from a battle_log.txt (the OCR'd 战报详情 log), grounded in web/public/game-data/formula.md (the damage formula) and the skill descriptions in web/public/game-data/database.json. For each round it surfaces every hero's 减少伤害/增加伤害 (区间A/B/C multipliers), healing, and damage above a threshold, identifies the turning points (who killed whom and via which skill), explains the win/loss/draw cause, and can suggest a concise team/formation adjustment. Triggered when the user asks to analyse / 复盘 / 分析 a battle log and points at a battle_log.txt location.
allowed-tools	["open_files","expand_code_chunks","grep","bash"]

Battle Log Analysis (战报复盘)

Use this skill after a battle has been fought and the user wants a mechanics-grounded round-by-round (逐回合) breakdown of what happened: for every hero, their 减少伤害 (damage reduction) and 增加伤害 (damage dealt) in formula terms, their healing, and the damage they dealt / took (only entries above a threshold), plus the turning points and the final win/loss/draw cause.

This skill analyses an actual battle outcome from its log and can optionally suggest a concise team/formation adjustment grounded in the same formula rules.

Trigger: user asks to "分析/复盘这场战报", "analyse the battle log", "逐回合分析", "give me round-by-round 减伤/增伤/治疗/伤害", etc., and points at a battle log (see input below).

Required inputs (ask if missing)

1. Battle log location — the user specifies it. It is a battle_log.txt, normally under study-battle-report/battles/<id>/battle_log.txt (produced by the battle-screenshots-to-log skill). Accept either a full path or a battle <id> (then read study-battle-report/battles/<id>/battle_log.txt). If multiple battles exist and none is specified, list them (ls study-battle-report/battles/) and ask which one.
2. Threshold for "damage worth listing" (default 150) — single damage events below this are summarised, not enumerated. Healing default threshold 50. Use defaults silently unless the user overrides.

Do NOT invent a log. If the file is missing, say so.

Level assumption

The log already reflects the actual in-battle numbers (50级/满级 values), so use the numbers as they appear in the log. Only fall back to the database's range values when explaining why a number is what it is.

Data sources

Read all three:

<battle_log.txt>        # the actual event log to analyse (user-specified)
web/public/game-data/formula.md         # the canonical damage formula (区间A/B/C, 同向乘法稀释, 异向线性相减, 主属性对位)
web/public/game-data/database.json    # skills[*].desc/type/prob/tier (explain each skill's mechanic), heroes[*], buffs, debuffs, bonds

Always re-read web/public/game-data/formula.md and look up every skill that fires in the log in database.json so the explanation matches the real mechanic, not memory.

Core procedure

1. Load the log. open_files / expand_code_chunks the full battle_log.txt. Identify:
   • the two sides and their 3 heroes each ([我方:…] blue / [敌方:…] red),
   • each side's 阵型 and faction (群/蜀/…) from the pre-battle 强化 block,
   • the result line at the end (胜利/失败/平局).
2. Read web/public/game-data/formula.md and internalise the bucket model:
   • 区间A = attacker 「造成X伤害」 (linear within a tag; separate tags = separate multiplicative regions; 100% cap).
   • 区间B = 「使敌方造成X伤害降低」 (mirror of A).
   • 区间C = 「受到X伤害」 (same-direction multiplicative dilution M = 1 − Π(1−mᵢ); opposite-direction linear R = E − M; 通用/兵刃/谋略 sub-tags are separate multiplicative regions).
   • 主属性对位: 兵刃 = 攻方武力 vs 守方统率; 谋略 = 攻方智力 vs 守方(智力+统率)/2.
3. Look up every firing skill in database.json (a small python pass over the skill names that appear in 【…】 brackets). Keep each skill's desc, type, prob, tier handy so you can name the mechanic behind each log line (e.g. 焰燎江天 = 协防 + 迟滞 + 52% 谋伤无视40%减伤, 每层迟滞 +7%).
4. Parse the pre-battle block (before 第一回合) into a per-hero starting 增伤/减伤 table, classifying each 【造成伤害】提升 / 【受到伤害】降低 / 兵种加成 / 阵型 / 栋梁 line into 区间A or 区间C.
5. Walk the log round by round. A round starts at 第N回合 (or the ·第N回合 / 第N回合 markers). For each round collect, per hero:
   • 减伤 (区间B/C): every 【受到X伤害】降低/提升, 抵御 (此次伤害减少…%), 规避 (成功规避), 协防 (为…承担伤害 + 减少…%), 造成伤害降低…% (defender 被动), 迟滞/技穷 stacks, 兵动若神 受伤-X%, etc. Note which are multiplicative-dilution (same tag) vs separate regions vs flat 抵御/规避.
   • 增伤 (区间A): every 【造成X伤害】提升, 蓄势待发 stacks, 机变无穷 谋伤 stacks, 攻心, 玉玺 伤害提升, 胜敌益强 武/智 growth, 空城计 谋伤, etc.
   • healing (恢复了兵力N) above the heal threshold, grouped by source skill (惩前毖后 / 皇思淑仁 / 攻心 / 清风驱疾 …).
   • damage (损失了兵力N / 由于…效果,损失了兵力N) above the damage threshold, attributed to source skill and target.
   • any HP milestone / death (兵力为0无法再战).
6. Identify turning points. Mark the round where a key unit dies, where an AOE burst (e.g. 僭号天子玉玺 兵刃AOE, 兵动若神满军令AOE) lands, or where a defensive engine breaks (规避/抵御/协防 exhausted, 迟滞被驱散). Tie each to the formula (e.g. "玉玺是兵刃AOE，绕过了以空城计谋略减伤为主的防御，专杀低统率单位").
7. Explain the result. Summarise why it was 胜/负/平: compare both sides' sustained 增伤 (区间A), 减伤/规避 (区间B/C), and healing throughput; name the single most decisive event.
8. (Optional) Team adjustment. If the user asks "how to fix" or the loss/draw is clearly structural, name the empty formula sub-bucket or exploited weakness (e.g. "缺兵刃减伤/低统率前排被玉玺秒"), and propose one concrete formation or skill swap using ONLY heroes/skills the user has, justified by the formula (区间C 单一大数值优于多个小数值; 统率墙半克谋略; etc.).

Counting / tallying helpers

For questions like "清风驱散了多少次迟滞" or "周瑜江天总共打了多少", use small grep/bash passes over the log and distinguish mechanism from coincidence:

• A 负面状态 「迟滞」效果已消失 line is only a driver-dispel if it is caused by a dispel skill (e.g. 清风驱疾) firing on the same actor/round; 迟滞 also expires naturally after its 持续2回合 — do NOT count natural expiry as a dispel. Check the surrounding lines for 发动战法【清风驱疾】 / 执行来自【清风驱疾】 before attributing.
• A skill 因几率未发动 did NOT fire — exclude it from counts.
• When tallying a damage skill's total, sum the 损失了兵力N lines whose cause clause names that skill (由于…【skill】…效果), per round and overall.

Always show the line numbers you counted from so the tally is auditable.

Output format

Use this shape (Chinese, concise, decision-oriented). Use <cite> line-citations into the log where possible.

# 战报分析：<我方阵容> vs <敌方阵容>（<结果>）

## 阵容与角色定位
- 我方（<faction> · <阵型>）：<3 heroes + 一句话定位>
- 敌方（<faction> · <阵型>）：<3 heroes + 一句话定位>

## 战前布署（行动前）
| 角色 | 增伤（区间A） | 减伤/受伤（区间C等） |
（each hero's pre-battle multipliers, classified into buckets）

## 第N回合
**减伤（区间B/C）** — per relevant hero, with dilution/规避/抵御/协防 noted
**增伤（区间A）** — per relevant hero
**伤害（>阈值，按来源）** — skill → target: -N, grouped by source skill
**治疗（>阈值）**
（+ 转折点/death notes inline）
... repeat per round ...

## 总结：为什么是<结果>
| 维度 | 我方 | 敌方 |
| 持续输出(区间A) | … | … |
| 减伤/规避(区间B/C) | … | … |
| 治疗 | … | … |
| 致命点 | … | … |
- 核心判断: 2-4 条，点名最决定性的事件。

## （可选）配队/阵型调整建议
- 指出被利用的空槽/弱点 + 用现有战法的具体改法，按公式说明收益。

Omit empty sections. Don't enumerate sub-threshold chip damage — summarise it.

Important

• Do NOT invent heroes, skills, mechanics, or numbers — heroes/skills/damage come from the log; mechanics come from database.json; stacking rules from web/public/game-data/formula.md.
• Preserve exact Chinese names from the log / database.json.
• Attribute every listed damage/heal to its source skill and target; when the log line wraps, read the surrounding lines to recover the full event.
• Distinguish mechanism vs natural expiry / 几率未发动 when counting (especially dispels like 清风驱疾 and stack-based effects like 迟滞).
• Tie turning points to the formula: which 区间 was bypassed (e.g. 兵刃AOE vs 谋略减伤), which engine ran out (规避/抵御/协防/治疗), which 统率墙 mattered.
• Always re-read web/public/game-data/formula.md; quote the relevant bucket rule when it changes a conclusion.
• Keep it concise and decision-oriented — a round-by-round attack/defense/heal breakdown + turning points + cause, not a raw dump of the log.
• When the user wants a fix, keep it concise: one formula-grounded adjustment rather than a full squad rebuild.