fix(llm): keep request generations coherent

This commit is contained in:
Tianyi Cui
2026-07-31 01:27:12 +08:00
parent afb05b4049
commit a9e2489db8
26 changed files with 237 additions and 142 deletions

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-29-request-level-llm-config-credentials.md
2026-07-29-request-level-llm-config-credentials.md: ec00b52bdbe8f00d334618f3e5347974a3928e67
2026-07-29-request-level-llm-config-credentials.zh.md: 29835b9fb320e6b31cb49a632ff0e56d88fb3f44
2026-07-29-request-level-llm-config-credentials.md: d84de8b74eb88720d782dfa20e193e6d180843dc
2026-07-29-request-level-llm-config-credentials.zh.md: 67b2d2e7cf952a7cb385e7c9dd67bd3a05651beb

View File

@@ -12,11 +12,11 @@ The [settings seam](2026-07-28-user-settings-seam.md) shipped without a producti
## Decision
**Per-request resolution, not fiber rebuilds.** The adapters take an options thunk and a per-stream credential resolver instead of rebuilding their fibers. Connection, credential, and request-transport facts are read for the operation, while an in-flight stream keeps the facts it started with. A missing key is a request-time `MISSING_CREDENTIAL` failure while the route remains registered. Provider routes and their retry policies are composition-fixed instead of triggering registration swaps.
**Per-request resolution, not fiber rebuilds.** The adapters take an options thunk and a per-stream credential resolver instead of rebuilding their fibers. Connection, credential, and request-transport facts are read for the stream, while an in-flight stream keeps the facts it started with. Model catalog/capability, context, reasoning-default, provider-route, and retry-policy facts are composition-fixed. A missing key is a request-time `MISSING_CREDENTIAL` failure while the route remains registered.
**Secrets are references, values live behind `ctx.credentials`.** Configuration can carry `apiKeyEnv: DEEPSEEK_API_KEY`; the read-only credential seam resolves it per operation. `credentials-local` checks the live process environment first, then parses `$DSH_HOME/.env` on demand, with no cache or mutation surface. Resolution order in the adapters is a non-empty literal `apiKey` first, then the seam, then — only without a mounted seam — the named raw environment variable.
**Per-plugin namespaces, schema ≡ `Config`.** Each adapter registers its own namespace (`llm-deepseek`, `llm-pi-ai`) with its plugin `Config` schema and `cordis.yml` entry as the composition `base`. `resolveAdapterOptions` and `resolveProfiles` remain the explicit validation steps, and a bad live snapshot keeps the last good request facts while a bad entry config fails load. pi-ai's `providers` is a non-empty dict keyed by its composition-owned routes; the user layer may override request facts for those routes but cannot add or remove them.
**Per-plugin namespaces, schema ≡ `Config`.** Each adapter registers its own namespace (`llm-deepseek`, `llm-pi-ai`) with its plugin `Config` schema and `cordis.yml` entry as the composition `base`. `resolveAdapterOptions` and `resolveProfiles` remain the explicit validation steps. A live snapshot that changes a fixed fact or fails another bound keeps the whole last-good generation, while a bad entry config fails load. pi-ai's `providers` is a non-empty dict keyed by its composition-owned routes; the user layer may override only live request facts for those routes.
## Alternatives considered

View File

@@ -12,11 +12,11 @@ Status: implemented
## 决策
**按请求解析,而非重建 fiber。**适配器接收 options thunk 与按流调用的凭据解析器,不再重建其 fiber。连接、凭据与请求传输事实在操作期间读取,进行中的流则保持其起始事实。密钥缺失会在请求时以 `MISSING_CREDENTIAL` 失败,同时路由保持注册。提供方路由及其重试策略由组合固定,不触发注册替换。
**按请求解析,而非重建 fiber。**适配器接收 options thunk 与按流调用的凭据解析器,不再重建其 fiber。连接、凭据与请求传输事实按流读取,进行中的流则保持其起始事实。模型 catalog能力、上下文、推理reasoning默认值、提供方路由与重试策略由组合固定。密钥缺失会在请求时以 `MISSING_CREDENTIAL` 失败,同时路由保持注册。
**机密是引用,值藏在 `ctx.credentials` 背后。**配置可以携带 `apiKeyEnv: DEEPSEEK_API_KEY`;只读凭据 seam 按操作解析它。`credentials-local` 先检查活跃进程环境,再按需解析 `$DSH_HOME/.env`,既不缓存,也不提供变更接口。适配器内的解析顺序为:非空的字面 `apiKey` 优先,然后是 seam最后仅在未挂载 seam 时读取点名的原始环境变量。
**按插件划分 namespaceschema ≡ `Config`。**每个适配器注册自己的 namespace`llm-deepseek``llm-pi-ai`),采用其插件 `Config` schema并以 `cordis.yml` 配置项为组合 `base``resolveAdapterOptions``resolveProfiles` 仍是显式校验步骤;错误的存活快照会保留最后可用的请求事实,错误的 entry 配置则会加载失败。pi-ai 的 `providers` 是以组合所拥有路由为键的非空字典;用户层可以覆盖这些路由的请求事实,但不能新增或移除路由
**按插件划分 namespaceschema ≡ `Config`。**每个适配器注册自己的 namespace`llm-deepseek``llm-pi-ai`),采用其插件 `Config` schema并以 `cordis.yml` 配置项为组合 `base``resolveAdapterOptions``resolveProfiles` 仍是显式校验步骤。存活快照若更改固定事实或违反其他约束,会整代沿用最后可用设置;错误的 entry 配置则会加载失败。pi-ai 的 `providers` 是以组合所拥有路由为键的非空字典;用户层只能覆盖这些路由的实时请求事实。
## 曾考虑的替代方案

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-credential-boundaries-and-atomic-registration.md
2026-07-30-credential-boundaries-and-atomic-registration.md: 09beda90d4789f951f663a3f9df794d74db2c11e
2026-07-30-credential-boundaries-and-atomic-registration.zh.md: 84f3826ba8a8b89fd5dd616164ac573151b8203c
2026-07-30-credential-boundaries-and-atomic-registration.md: 2a808048bccc619a311e61340b5aeb0f3113030d
2026-07-30-credential-boundaries-and-atomic-registration.zh.md: 4c2147ea0f0d30624500aeda92fb02932a5b0960

View File

@@ -1,4 +1,4 @@
# Agent Note: credential boundaries, whole-snapshot requests, and atomic route registration
# Agent Note: credential boundaries and whole-generation LLM requests
Status: implemented
@@ -10,7 +10,7 @@ English | [中文](2026-07-30-credential-boundaries-and-atomic-registration.zh.m
Review found the credential path leaking across boundaries it had drawn. The shipped surfaces hoisted `$DSH_HOME/.env` into `process.env` before cordis booted, so on the next run `credentials-local` classified every key it had stored itself as a read-only ambient launch override: `describe()` reported `source: 'env'` with `writable: false`, `set`/`unset` rejected as shadowed, and a key stored from the web page or TUI became unrotatable and undeletable while the adapter kept using the value captured at launch. The store's own write path repeated the settings-local defects that same review round fixed (two independent chains, whole-file render from a stale cache), plus editor bugs of its own: a physical line inside another key's quoted multi-line value read as an assignment, CRLF endings degraded to LF, a multi-line entry reported `writable: true` while `set` always threw, and `credentials/updated` was emitted bare after the commit, so one broken observer made a durable write look failed. On the read side, the file's `0600` mode stops other OS users but not the model, whose bash and filesystem tools run as the same user.
Two request-path defects sat beside them. DeepSeek's per-request resolution kept connection facts in a last-good snapshot but re-read the literal `apiKey` from the raw configuration, so a settings generation the resolver rejected could still put its key on the previous generation's endpoint. pi-ai handed the SDK `undefined` when a configured `apiKeyEnv` resolved to nothing, letting pi-ai's own environment discovery authenticate with an unrelated provider key — another tenant, silently billed. And its route swap disposed the old registration before creating the new one: a route another adapter owned dropped every existing route, after which the facts cache could equal the registry's, so restoring the working configuration never re-applied.
Request-path defects sat beside them. DeepSeek's per-request resolution kept connection facts in a last-good snapshot but re-read the literal `apiKey` from the raw configuration, so a settings generation the resolver rejected could still put its key on the previous generation's endpoint. Both adapters also read their full settings snapshot during asynchronous model-capability resolution and again at stream dispatch; a change between those reads could pair one generation's reasoning/capability facts with the next generation's endpoint and key. pi-ai handed the SDK `undefined` when a configured `apiKeyEnv` resolved to nothing, letting pi-ai's own environment discovery authenticate with an unrelated provider key — another tenant, silently billed.
## Decision
@@ -18,7 +18,7 @@ Two request-path defects sat beside them. DeepSeek's per-request resolution kept
**The stored credential has no boundary against the model, and the READMEs say so.** `0600` under a `0700` directory stops other OS users; the model's bash and filesystem tools run as that same user, and the shipped default confines nothing. What the harness does hold to is narrower and stated as exactly that: no surface hoists the document into `process.env`, and the model is never handed a resolved path to it, so reaching the value takes a deliberate read of a path it was not given. An OS-keychain provider — a store the model's processes cannot read at all — is recorded as the real answer rather than implied by a partial one.
**One request, one generation.** DeepSeek's resolved snapshot carries the credential facts (literal key and reference) beside the endpoint, and `resolveApiKey` receives that snapshot instead of re-reading configuration. A rejected generation now contributes nothing at all. pi-ai defers to provider-native discovery only for a profile naming no credential; a configured reference that misses fails with `MISSING_CREDENTIAL` naming the route and the reference. The boot-time credential probe is deleted: it could run before the credentials service mounted and reported every failure as a missing key, while the first request already gives the accurate error.
**One request cannot straddle settings generations.** Model catalog/capability, context, reasoning-default, provider-route, and retry-policy facts are captured from composition. Only connection, credential, and request-transport facts resolve live, once at stream dispatch. A live snapshot that changes any fixed fact is rejected whole, so its endpoint and key cannot combine with capability facts resolved before the change. DeepSeek's accepted snapshot carries the literal key and reference beside the endpoint, and `resolveApiKey` receives that snapshot instead of re-reading configuration. pi-ai defers to provider-native discovery only for a profile naming no credential; a configured reference that misses fails with `MISSING_CREDENTIAL` naming the route and reference.
**Provider routes are composition-owned.** `registerAdapter` binds one non-empty route set to its calling fiber and returns a disposer. Settings cannot create or remove routes or change their captured retry policy, so the registry needs no replacement lifecycle and a bad settings snapshot leaves the composition registration untouched.
@@ -33,4 +33,4 @@ Two request-path defects sat beside them. DeepSeek's per-request resolution kept
## Consequences
The local provider performs a direct environment-then-dotenv read for each resolution; mutation, description, writer locking, and change events are absent. `LlmAdapter` registrants receive an ordinary disposer, and `DeepSeekConnectionOptions` carries credential facts with its endpoint so one rejected settings generation cannot contribute only a key. An OS-keychain provider remains the path to isolating secrets from same-user model tools.
The local provider performs a direct environment-then-dotenv read for each resolution; mutation, description, writer locking, and change events are absent. `LlmAdapter` registrants receive an ordinary disposer. Each adapter captures model/capability defaults from composition and resolves one live connection snapshot at dispatch, so a rejected settings generation cannot contribute only an endpoint or key. An OS-keychain provider remains the path to isolating secrets from same-user model tools.

View File

@@ -1,10 +1,10 @@
# Agent Note: 凭据边界、按整份快照发起的请求与原子路由注册
# Agent Note: 凭据边界与 LLM大语言模型请求的同代一致性
Status: implemented
[English](2026-07-30-credential-boundaries-and-atomic-registration.md) | 中文
> 范围:对[请求级 LLM(大语言模型)配置 seam](2026-07-29-request-level-llm-config-credentials.md)的存储与请求边界修正。后续的[只读凭据与静态路由](../simplification/2026-07-31-read-only-credentials-and-static-llm-routes.md)决策移除了凭据写入、共享原子写入器与可变注册;本 note 负责保留至今的机密边界与整次请求同代规则。
> 范围:对[请求级 LLM 配置 seam](2026-07-29-request-level-llm-config-credentials.md)的存储与请求边界修正。后续的[只读凭据与静态路由](../simplification/2026-07-31-read-only-credentials-and-static-llm-routes.md)决策移除了凭据写入、共享原子写入器与可变注册;本 note 负责保留至今的机密边界与整次请求同代规则。
## 问题
@@ -14,7 +14,7 @@ Status: implemented
在读取一侧,文件的 `0600` 权限挡得住其他 OS 用户,却挡不住模型:它的 bash 与文件系统工具就以同一个用户身份运行。
与之并排的还有两个请求路径缺陷。DeepSeek 的按请求解析把连接事实保存在最后可用快照里,却仍从原始配置重新读取字面 `apiKey`,于是被 resolver 拒绝的那一代设置,照样能把自己的密钥送到上一代的端点上。配置了 `apiKeyEnv` 却解析不到值时pi-ai 会把 `undefined` 交给 SDK让 pi-ai 自己的环境发现拿一个毫不相干的提供方密钥完成鉴权——那是另一个租户,账单还悄悄记在它头上。而且它的路由替换是先释放旧注册、再创建新注册:只要有一条路由已被别的适配器占有,现有路由就会被全部丢掉,此后事实缓存可能与注册表中的事实相等,于是把配置改回可用状态也不会重新生效。
与之并排的还有请求路径缺陷。DeepSeek 的按请求解析把连接事实保存在最后可用快照里,却仍从原始配置重新读取字面 `apiKey`,于是被 resolver 拒绝的那一代设置,照样能把自己的密钥送到上一代的端点上。两个适配器还会在异步模型能力解析期间读取一次完整 settings 快照并在流派发时再次读取若两次读取之间发生变更一代的推理reasoning能力事实就可能与下一代的端点和密钥拼接在一起。配置了 `apiKeyEnv` 却解析不到值时pi-ai 会把 `undefined` 交给 SDK让 pi-ai 自己的环境发现拿一个毫不相干的提供方密钥完成鉴权——那是另一个租户,账单还悄悄记在它头上。
## 决策
@@ -22,7 +22,7 @@ Status: implemented
**存下的凭据对模型没有边界,而 README 就是这么写的。**`0700` 目录下的 `0600` 挡得住其他 OS 用户;模型的 bash 与文件系统工具正是以同一用户身份运行而已交付的默认值不约束任何东西。harness 真正守住的更窄,也就照这个宽度写下来:没有任何一个面会把该文档提升进 `process.env`模型也从不会拿到它的解析后路径因此要拿到这个值需要刻意去读一条并未交给它的路径。OS 钥匙串keychain提供方——一个模型的进程根本读不到的存储——被记录为真正的答案而不是靠一个残缺的方案去暗示它。
**一次请求,一代设置。**DeepSeek 解析出的快照在端点旁一并携带凭据事实(字面密钥与引用`resolveApiKey` 接收这份快照,而不再重新读取配置。被拒绝的那一代如今完全不再贡献任何东西。只有当一个 profile 完全没有点名凭据时pi-ai 才交给提供方原生的发现流程;配置了引用却解析不到,就以 `MISSING_CREDENTIAL` 失败,并点名路由与引用。启动时的凭据探测被删除:它可能在凭据服务挂载之前就运行,并把每一种失败都报成密钥缺失,而第一次请求本就会给出准确的错误。
**一次请求不得跨 settings 代取值。**模型 catalog能力、上下文、推理默认值、提供方路由与重试策略均从组合中捕获。只有连接、凭据与请求传输事实实时解析并且只在流派发时解析一次。存活快照若更改任何固定事实就会整代被拒绝因此它的端点与密钥无法同变更前解析出的能力事实组合。DeepSeek 已接受的快照在端点旁一并携带字面密钥与引用,`resolveApiKey` 接收这份快照,而不再重新读取配置。只有当一个 profile 完全没有点名凭据时pi-ai 才交给提供方原生的发现流程;配置了引用却解析不到,就以 `MISSING_CREDENTIAL` 失败,并点名路由与引用。
**提供方路由归组合所有。**`registerAdapter` 把一组非空路由绑定到调用方 fiber并返回释放器。settings 无法创建或移除路由,也无法更改注册时捕获的重试策略,因此注册表无需替换生命周期,错误的 settings 快照也不会影响组合注册。
@@ -37,4 +37,4 @@ Status: implemented
## 后果
本地提供方每次解析都会依次直接读取环境与 dotenv修改、描述、写入锁和变更事件均不存在。`LlmAdapter` 注册方收到普通释放器`DeepSeekConnectionOptions` 将凭据事实与端点一同携带,因此一代被拒绝的 settings 不可能只贡献密钥。OS 钥匙串提供方仍是将机密与同一用户身份下的模型工具隔离的实现路径。
本地提供方每次解析都会依次直接读取环境与 dotenv修改、描述、写入锁和变更事件均不存在。`LlmAdapter` 注册方收到普通释放器。每个适配器从组合中捕获模型/能力默认值,并在派发时解析一份实时连接快照,因此被拒绝的 settings 不可能只贡献端点或密钥。OS 钥匙串提供方仍是将机密与同一用户身份下的模型工具隔离的实现路径。

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/simplification/2026-07-31-read-only-credentials-and-static-llm-routes.md
2026-07-31-read-only-credentials-and-static-llm-routes.md: 2ebbea28c2dadeb9482c483247249ed4054749e4
2026-07-31-read-only-credentials-and-static-llm-routes.zh.md: ad21d93bf73ca0063eb97da8e610aa2814ee1726
2026-07-31-read-only-credentials-and-static-llm-routes.md: 3e0d4ceb612379d2ea29068605e0f7c60f1019a4
2026-07-31-read-only-credentials-and-static-llm-routes.zh.md: eca740c6d58d67a11270dba427efbb3b70852d2c

View File

@@ -14,7 +14,7 @@ That speculative closure accounted for much of the feature's runtime and test gr
`ctx.credentials` exposes only branded `CredentialRef` construction and `resolve(ref): Promise<string | undefined>`. `credentials-local` reads the named process environment value, then parses its dotenv file on demand. It owns no mutation, description, event, watcher, cache, editor, or writer lifecycle; externally changing either source is visible to the next resolution.
LLM provider routes and their retry policies are composition-owned. `registerAdapter()` returns a disposer rather than a mutable registration handle. DeepSeek always owns its one route, and pi-ai requires a non-empty configured route map; settings may change request-level facts for those existing routes but cannot create, remove, or retune registrations. The shared CLI composition therefore does not mount an empty pi-ai adapter.
LLM provider routes, model/capability metadata, context limits, reasoning defaults, and retry policies are composition-owned. `registerAdapter()` returns a disposer rather than a mutable registration handle. DeepSeek always owns its one route, and pi-ai requires a non-empty configured route map; settings may change only connection, credential, and request-transport facts for those existing routes. A settings generation that changes a fixed fact is rejected whole. The shared CLI composition therefore does not mount an empty pi-ai adapter.
The optional-settings helper only switches a consumer's source thunk between its composition entry and a live settings scope. Consumers read committed values through that thunk, so the helper needs no update watcher, derived-state callback, or teardown-state mirror. `settings-local` keeps its write protocol private instead of publishing a utility for a second writer that no longer exists.

View File

@@ -14,7 +14,7 @@ Status: implemented
`ctx.credentials` 只暴露品牌化 `CredentialRef` 的构造,以及 `resolve(ref): Promise<string | undefined>``credentials-local` 先读取点名的进程环境值,再按需解析其 dotenv 文件。它不拥有修改、描述、事件、watcher、缓存、编辑器或写入器生命周期从外部更改任一来源都会在下一次解析时生效。
LLM 提供方路由及其重试策略归组合所有。`registerAdapter()` 返回释放器而非可变注册句柄。DeepSeek 始终拥有自身唯一的路由pi-ai 则要求配置一份非空路由映射settings 可以更改这些现有路由的请求级事实,但不能创建、移除或重新调整注册。因此,共享 CLI命令行界面组合不会挂载空的 pi-ai 适配器。
LLM 提供方路由、模型能力元数据、上下文限制、推理reasoning默认值与重试策略归组合所有。`registerAdapter()` 返回释放器而非可变注册句柄。DeepSeek 始终拥有自身唯一的路由pi-ai 则要求配置一份非空路由映射settings 只能更改这些现有路由的连接、凭据与请求传输事实。更改固定事实的 settings 代会整代被拒绝。因此,共享 CLI命令行界面组合不会挂载空的 pi-ai 适配器。
可选 settings 辅助工具只在组合配置项与存活 settings scope 之间切换消费方的来源 thunk。消费方经该 thunk 读取已提交值,因此辅助工具不需要更新 watcher、派生状态回调或拆卸状态镜像。`settings-local` 将自身的写入协议保留为私有实现,不再为一个已不存在的第二写入方公开工具。