[Bug] Manual /compact on a resumed session summarizes through the session's stale routed provider, with zero cache reuse

[ProdaZhang]

Summary

compaction-basic resolves its summarization target as

// packages/compaction/compaction-basic/src/summarizer.ts
const latest = agent.session.requestHeader()?.config
...
const target = configured ?? latest ?? agentTarget

That ranks a persisted, arbitrarily old request header above the agent's currently configured provider. The rationale for preferring it is prefix-cache reuse, stated right above the call:

Run the default cache-reusing ctx.llm.stream() summarization call: replay the conversation prefix, then append the compaction instruction as the final user message so the provider's warm prefix cache is reused.

That premise does not hold for a manual /compact on a resumed session. There is no turn in flight, the header can be days old, and the provider's cache is long gone — so the call inherits the old provider's billing without any of the reuse it was inherited for.

What happened

A session was last active on 2026-08-14; its final request/header recorded provider: deepseek-official. The provider in use changed afterwards. On 2026-08-20 the session was reopened and /compact was run before any new turn:

command/run         {"name":"compact","source":{"kind":"user"}}
compaction/start    {"turn": null}
compaction/summary  {"provider":"deepseek-official","model":"deepseek-v4-flash",
                     "llmStreamCall":true,
                     "usage":{"inputTokens":375320,"outputTokens":3270,
                              "cacheReadTokens":0,"reasoningTokens":46}}
command/done        "Compacted 419 history items (~237097 tokens)."

Two fields carry the whole report.

turn: null — there is no request in flight, so there is no current route at all. session.requestHeader() can only return the six-day-old one. Preferring "the latest routed header" degenerates to "whatever provider this session happened to end on, whenever that was".

cacheReadTokens: 0 — the reuse that justified inheriting that provider never happened. 375,320 input tokens, none of them cached, billed to a provider the account had moved off. The next real turn, 70 seconds later, logged provider: ark.

The first signal that any of this had happened was a billing dashboard the following hour. Nothing in the session, the command result, or the UI mentions which provider a compaction used.

The cache miss was predictable, not unlucky

Inheriting the routed header is a bet that the provider still holds the conversation's prefix. That bet is a good one in the common case — a compaction fired while the session is active replays a prefix the provider served minutes ago, and the reuse is real. It is what the design is for.

But whether the bet can pay is knowable before the call, from data the harness already has: the age of the header it is about to inherit. A session reopened after days is past any provider's prefix-cache TTL, so the reuse is not merely unlikely, it is arithmetically impossible — and inheriting the route then buys nothing while still deciding who gets billed. turn: null is the same signal from the other side: no request is in flight, so there is no warm prefix to be continuing from in the first place.

So this is not "the cache sometimes misses". It is a bet placed in the one situation where it is already known to be lost.

Why this is not #1944

#1944 reports that the summarizer inherits too little of the routed header — it drops reasoningEffort and adds an explicit maxTokens, splitting the cache key — and proposes inheriting the header config wholesale. That treats "the last routed header" as the correct source of truth and makes the inheritance more complete.

In the resumed-session case that cuts the wrong way: the whole of a stale header would be carried into the summarization call rather than two fields of it.

Both readings can be right at once. Inherit the header faithfully when it is current; don't inherit it at all when it plainly isn't.

Suggested fixes

1. Gate latest on freshness, and fall through to agentTarget when it fails. Two cheap tests, both from data already at hand:

• turn: null at compaction/start — no request in flight, so there is no warm prefix to continue from, and the rationale for latest cannot apply by construction.
• The routed header's age exceeds a configured bound. The header is persisted with the session, so its age is known before the call; past the provider's prefix-cache TTL the reuse is impossible rather than uncertain. This one also covers the automatic-compaction path, where turn is not null but the session may still have been idle for days.

When either fails, agentTarget is the better default: it is at least the provider the user believes they are on, and nothing is given up, because the reuse that latest was preferred for was not available anyway.

2. Say where the summary went. command/done currently reads Compacted 419 history items (~237097 tokens). and names no provider. Appending via <provider>/<model> would surface this at the moment it happens rather than on a bill. This one is cheap and worth having on its own merits: today a /compact can spend real money on a provider the user believes is inactive, and nothing says so.

Notes

• summarizationProvider / summarizationModel in the plugin config outrank latest, so pinning them is a workaround.
• The precedence is deliberate and pinned by tests/compaction-basic.spec.ts — "resolves the latest routed provider/model before the AgentOptions pair". That test appends the header inside the same session with reason: 'initial', i.e. a warm route. No test covers a header that is stale relative to the agent's current configuration.
• I cannot tell from the session log what agent.options.provider was at the instant /compact ran, so I am not claiming a specific correct provider was ignored — only that a six-day-old header was preferred with turn: null and returned cacheReadTokens: 0.
• Observed on 0.1.0-rc.8.

[zoahdev]

你这个观察和 #3542 是同一块代码的不同入口：#3542 是 compaction 启动后被 caller 中止（5-15s 稳定超时），你这里是 compaction-basic 的 summarization target 解析优先走 persisted request header 而不是当前 agent 配置，导致 resumed session 永远用旧路由/旧 provider，零缓存命中。\n\n我们核验过 #1944：summarizeWithLlm 只继承 provider/model，强加 maxTokens 8192，不继承 routed header config——compaction 请求与正常回合参数不一致，provider 侧前缀缓存必然 miss（~122k tokens 重算）。补丁分支 fix/compaction-inherit-header-config（基于 main 99f6f02，124/124 测试全绿）已就绪，改整包继承 header config。\n\n你这个发现（summarizer.ts 用 session.requestHeader()?.config 优先于 configured ?? agentTarget）是另一个层面：即使 header config 继承修好了，persisted header 本身可能比 agent 当前配置旧。建议在 configured 之后、agentTarget 之前加一个 currentRoute 从 agent.options 取，排在 anything stale 前面。

[argszero]

zoahdev 的 currentRoute 建议方向对，补一个家族视角 + 一个实现细节，让这条链彻底闭环：

1. 这是"persisted snapshot vs live state"根因族的第 3 个成员

这个 bug 与 subagent 模型路由继承（#3552，父会话切 provider 后子代理继承冻结的 agent.options → 计费错误账号）同根：都是"从持久化/创建时快照取值，而不是从当前 live 状态取值"。三处对照：

• subagent：resolveChildAgentOptions（child-agent.ts:73）读 parent.options（创建时冻结）→ 应读 live WebModelSelectionRef.current（含 pending）
• compaction：summarizer.ts:128 读 session.requestHeader()?.config（上次路由，可能六天前）→ 应读 agent.options（= agentTarget，用户当前配置）或 live route
• 上一轮我修的 subagent 继承 stale 的 agent.options.provider，忽略 Web 模型选择，导致子代理用错 provider 计费 #3552 还有第三层：子代理 ctx 根本没有 selection 槽（api-proxy 只给 Web 会话装 installSelection）——compaction 没有这层问题（它直接拿 agent 参数），但说明"快照优先"在这两个场景里都是刻意设计过、然后被 live 状态演进打破的。

共同教训："上次成功路由"作为默认值是合理的 warm-cache 启发，但它一旦可能跨天/跨会话存活，就必须有 freshness 守卫。requestHeader() 在活跃会话里是"当前路由"，在 resumed 会话里是"历史路由"——同一个 API 在两个语义之间摇摆，这才是根。

2. 修复分级建议（比"加 currentRoute"再多一步）

• 最小（zoahdev 已提）：configured ?? currentRoute ?? agentTarget，其中 currentRoute 从 agent.options 取——注意这要求 agent.options 在会话中途切换时被更新（subagent 继承 stale 的 agent.options.provider，忽略 Web 模型选择，导致子代理用错 provider 计费 #3552 的根因之一就是它不更新）。若 options 也冻结，currentRoute 需来自 live selection（同 subagent 继承 stale 的 agent.options.provider，忽略 Web 模型选择，导致子代理用错 provider 计费 #3552 的修复）。
• 建议追加（新鲜度守卫）：latest 只有在"新鲜"时才优先——例如 requestHeader() 的 at 距今 < 某阈值（如 24h），否则跳过它直接走 agentTarget。这正好覆盖你证据里的场景（6 天前的 header）+ zoahdev 提的自动 compaction 路径（turn 非 null 但会话闲置数天）。阈值语义直白："跨天不信任"。
• 说人话（你的建议 2 完全赞成）：command/done 输出 via <provider>/<model>——今天一个 /compact 可以在用户以为不活跃的 provider 上花真钱而无人知晓，这行字是最便宜的修复。

3. 测试盲区确认（你点到的）

tests/compaction-basic.spec.ts 的用例都在同一会话内先路由再 compact（warm header），所以 latest 优先永远"正确"。要补的回归是：resumed 会话 + 冷 header（或 header 缺失）+ 当前 agent.options 不同 时，target 必须落到 agentTarget（或 currentRoute）。这个用例同时锁住 freshness 守卫和优先级修正。

家族状态：subagent（#3552）+ compaction（#3565）= "snapshot vs live" 2 报告；修复形态共享"注入 live source"原则。已建家族记忆跟踪。