Context breakdown retains only system prompts
上下文构成只保留系统提示词
`contextBreakdown` kept one `{ seq, heuristicTokens, system }` entry per retained surface node so that any replacement could subtract the exact price of its range and re-select the last nonempty system prompt. Its checkpoint therefore grew with the number of messages in the model context. In a local corpus of 836 Sessions, 151,122 entries made the unit's rows 7.13 MB of the projection cache and up to 103 KB for one S
English
Problem
contextBreakdown kept one { seq, heuristicTokens, system } entry per retained surface node so that any replacement could subtract the exact price of its range and re-select the last nonempty system prompt. Its checkpoint therefore grew with the number of messages in the model context. In a local corpus of 836 Sessions, 151,122 entries made the unit's rows 7.13 MB of the projection cache and up to 103 KB for one Session, and every cached Session listing validated all of them. Long tool-heavy runs below the compaction threshold kept about 2,000 entries each.
Decision
breakdown-projection.ts retains the surviving system/message nodes in surface order with their heuristic prices, one running surface total, and at most one armed shadow-price claim. State version 6 discards version-5 rows; other projection keys keep their rows.
- An append adds its price; a
system/messageappend also joins the end of the system list. compaction/summaryandcompaction/prunearm a claim fromshadowedRange,shadowedTokenCount, and the retained system prompts named byshadowedSeqs. The next event consumes or expires it. A replacement over the claimed range subtracts the claim price and drops those prompts; a replacement naming another range, or putting asystem/messagein place of a metered range, throws.- An unmetered replacement of one retained system prompt subtracts that prompt's own price. A
system/messagereplacement keeps its position; any other replacement removes it. - Any other unmetered replacement folds at zero delta, the rule
contextPressurealready applies.
systemTokens is the last retained prompt with a positive price, and messageTokens is the surface total minus it; both are floored at zero. shadowedSeqs is the authoritative set of shadowed nodes, so membership decides which prompts leave without locating non-system endpoints in surface order.
Verification
Folding all 836 stored logs with the version-5 and version-6 definitions produced identical breakdowns after every event. The logs held 2,206 metered replacements, 106 system-prompt rewrites, and no other replacement. Retained state fell from 7.13 MB to 110 KB, with at most five surviving system prompts in one Session.
Alternatives considered
Keep per-node entries with a faster validator. A hand-written node check cut listing validation from 75 ms to 4.5 ms but kept 7.13 MB of rows, an O(surface) array copy per surface event, and checkpoint writes proportional to the message count.
Columnar per-node entries. Parallel arrays shrink rows about 5.7 times but keep O(surface) state for information the shadow-price events already carry.
Exact pricing of unmetered non-system replacements. That requires per-node prices. Current producers meter every non-system replacement, and released formats from v0 carry the shadow-price fields, so the remaining case is a log that breaks producer adjacency; it drifts like contextPressure.
Consequences
Checkpoint size and per-event cost follow the number of surviving system prompts rather than messages, and cached listings no longer validate per-message entries. The figures stay exact for metered replacements and system-prompt rewrites, and their sum equals the measure() heuristic total under those producers. An unmetered non-system replacement leaves the figures off by that replacement's price difference. A future range replacement that inserts a system prompt is rejected until the fold can order it among the retained prompts. This supersedes the positional-entry choice for this unit in the in-history system-prompt replacement note; measure() keeps positional nodes for route pricing.
中文
Problem
contextBreakdown 曾为每个保留的 surface 节点存一条 { seq, heuristicTokens, system }，以便任意替换都能减去其范围的精确价格，并重新选出最后一个非空系统提示词。因此它的检查点随模型上下文中的消息数增长。在一个包含 836 个 Session 的本地语料中，151,122 条条目让该单元的行占用 projection cache 的 7.13 MB，单个 Session 最多 103 KB，而每次缓存会话列表都要校验全部条目。低于压缩阈值、工具调用密集的长时间运行各自保留约 2,000 条条目。
Decision
breakdown-projection.ts 按 surface 顺序保留存活的 system/message 节点及其启发式价格、一个 surface 运行总量，以及至多一个已武装的影子价 claim。状态版本 6 丢弃版本 5 的行；其他 projection key 保留各自的行。
- 追加加上其价格；
system/message追加还会加入系统列表末尾。 compaction/summary和compaction/prune依据shadowedRange、shadowedTokenCount以及shadowedSeqs中列出的已保留系统提示词武装 claim。下一个事件消费或使其失效。覆盖所 claim 范围的替换减去 claim 价格并移除这些提示词；指向其他范围的替换，或以system/message取代已计量范围的替换，会抛出错误。- 对单个已保留系统提示词的未计量替换减去该提示词自身的价格。
system/message替换保持其位置；其他替换将其移除。 - 其他未计量替换以零增量折叠，这与
contextPressure已采用的规则相同。
systemTokens 是最后一个价格为正的保留提示词，messageTokens 是 surface 总量减去它；两者都不低于零。shadowedSeqs 是被遮蔽节点的权威集合，因此由成员关系决定哪些提示词离开，无需在 surface 顺序中定位非系统端点。
Verification
用版本 5 和版本 6 的定义折叠全部 836 份已存日志，每个事件之后的构成数字都完全相同。这些日志包含 2,206 次已计量替换、106 次系统提示词改写，没有其他替换。保留状态从 7.13 MB 降到 110 KB，单个 Session 最多有五个存活的系统提示词。
Alternatives considered
保留逐节点条目，换用更快的校验器。 手写节点检查把列表校验从 75 ms 降到 4.5 ms，但仍保留 7.13 MB 的行、每个 surface 事件一次 O(surface) 数组复制，以及与消息数成正比的检查点写入。
列式逐节点条目。 并行数组让行缩小约 5.7 倍，但仍为影子价事件已经携带的信息保留 O(surface) 状态。
精确定价未计量的非系统替换。 这需要逐节点价格。当前生产方会计量每一次非系统替换，自 v0 起的已发布格式都携带影子价字段，因此剩余情况只是破坏生产方相邻性的日志；它像 contextPressure 一样产生偏差。
Consequences
检查点大小和每个事件的成本取决于存活系统提示词的数量而不是消息数，缓存列表也不再校验逐消息条目。对已计量替换和系统提示词改写，数字保持精确，其和在这些生产方下等于 measure() 的启发式总量。未计量的非系统替换会让数字偏离该替换的价格差。未来插入系统提示词的范围替换会被拒绝，直到折叠能确定它在已保留提示词中的顺序。这取代了 历史内系统提示词替换 说明中针对该单元的位置条目选择；measure() 仍为路由定价保留位置节点。