Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(llm): bind reasoning resolution to adapter lifecycle

Browse Source
This commit is contained in:
Yichen Jiang
2026-07-25 22:59:45 +08:00
parent 478376acaf
commit baea5018e5
30 changed files with 429 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
.agents/notes/implemented/architecture/2026-07-24-adapter-owned-reasoning-effort-capabilities.i18n.yaml
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
2026-07-24-adapter-owned-reasoning-effort-capabilities.md: 806e808ab18d617003f757200f0cdee5853476f5
2026-07-24-adapter-owned-reasoning-effort-capabilities.zh.md: 9be8b6f4f7bcbe063ff90492ac44b9590860f67f
2026-07-24-adapter-owned-reasoning-effort-capabilities.md: 04aebf5fa61d896ac16b29578a39c9b0e6e38cf4
2026-07-24-adapter-owned-reasoning-effort-capabilities.zh.md: a1c89417562276fd70ac5e5c1be17c9e639eb3ed

4
.agents/notes/implemented/architecture/2026-07-24-adapter-owned-reasoning-effort-capabilities.md
View File

@@ -12,7 +12,7 @@ Reasoning strength was adapter configuration only, so a conversation could not d
`dsh-llm` represents a reasoning effort as the opaque branded `ReasoningEffortId`. An adapter's `resolveModelReasoning(provider, model)` returns a non-empty ordered list of ids with display metadata and may name one configured default. The core validates metadata, requires an explicit or configured effort to appear exactly in that list, and never clamps or aliases a value.
`LlmCallConfig` and `GenerateOptions` carry the optional effort. The agent loop resolves the post-`agent/request` config before writing `request/header`, so defaults and dynamic changes are model-visible only after becoming durable facts. A route with no registered adapter retains its proposed config so an `llm/stream` middleware can own and short-circuit it; terminal dispatch still rejects an unhandled route. A resumed loop retains the logged effort only when its initial provider/model route is unchanged; a route change discards the previous model's opaque id. The terminal `LlmService` adapter boundary repeats resolution for direct calls that do not pass through the loop.
`LlmCallConfig` and `GenerateOptions` carry the optional effort. The agent loop prepares the post-`agent/request` config under the active turn signal before writing `request/header`, so defaults and dynamic changes are model-visible only after becoming durable facts. The prepared call retains the exact adapter registration across asynchronous capability resolution, durable header logging, and dispatch; direct `LlmService.stream()` calls likewise capture their final registration before awaiting resolution. A route with no registered adapter retains its proposed config so an `llm/stream` middleware can own and short-circuit it; terminal dispatch still rejects an unhandled route. A resumed loop retains the logged effort only when its initial provider/model route is unchanged; a route change discards the previous model's opaque id.
The native DeepSeek adapter advertises `high` and `max`, defaults to configured effort or `high`, and exposes no effort capability while thinking is disabled. The pi-ai adapter derives each exact model's list from `getSupportedThinkingLevels()`, excludes `off`, preserves an absent profile default as a provider default, and leaves provider wire-value mapping inside pi-ai.
@@ -30,4 +30,4 @@ The native DeepSeek adapter advertises `high` and `max`, defaults to configured
Clients can query one exact route and render the adapter's order and names without knowing a global enum. Adapter configuration remains the deployment-default owner, while `agent/request` can replace the effective effort on each step. Invalid metadata fails with `INVALID_MODEL_REASONING`, and unsupported explicit or configured values fail with `UNSUPPORTED_REASONING_EFFORT` before provider I/O.
The capability query is asynchronous and exact-model resolution may fail for adapters backed by authoritative catalogs. Keyless service, adapter, loop, session, and request-header tests pin validation, defaulting, dynamic changes, logging, and resume behavior; runnable snapshots pin the resolved effort in real assembled request headers, while key-gated adapter tests exercise provider serialization.
The capability query is asynchronous and exact-model resolution may fail for adapters backed by authoritative catalogs. Its optional signal is the caller's cancellation boundary; an asynchronous adapter must settle promptly after abort so loop disposal can reach quiescence. Keyless service, adapter, loop, session, and request-header tests pin validation, defaulting, dynamic changes, logging, resume behavior, HMR registration ownership, and cancellation; runnable snapshots pin the resolved effort in real assembled request headers, while key-gated adapter tests exercise provider serialization.

4
.agents/notes/implemented/architecture/2026-07-24-adapter-owned-reasoning-effort-capabilities.zh.md
View File

@@ -12,7 +12,7 @@ Status: implemented
`dsh-llm` 使用不透明的品牌类型 `ReasoningEffortId` 表示推理强度。适配器的 `resolveModelReasoning(provider, model)` 返回非空的有序 ID 列表及其展示元数据，并可指定一个由配置确定的默认值。核心会校验元数据，要求显式指定或配置指定的推理强度与列表中的某个 ID 完全一致，且绝不自动调整或为值提供别名。
`LlmCallConfig` 和 `GenerateOptions` 携带可选的推理强度。agent loop（智能体循环）在 `agent/request` 处理完成后、写入 `request/header` 前解析配置，因此默认值和动态变更只有成为持久化事实后才对模型可见。没有已注册适配器的路由会保留原定配置，使 `llm/stream` 中间件可以接管并短路该请求；若仍未得到处理，最终分发会拒绝该路由。恢复后的主循环仅在初始提供方/模型路由未变时保留日志中记录的推理强度；如果路由发生变化，则丢弃上一模型的不透明 ID。最终的 `LlmService` 适配器边界会再次执行解析，以覆盖未经过主循环的直接调用。
`LlmCallConfig` 和 `GenerateOptions` 携带可选的推理强度。agent loop（智能体循环）在活跃轮次信号的控制下准备 `agent/request` 处理完成后的配置，再写入 `request/header`，因此默认值和动态变更只有成为持久化事实后才对模型可见。准备完成的调用在异步能力解析、请求头持久记录和分派全程保留同一项确切的适配器注册；直接调用 `LlmService.stream()` 时，也会在等待解析前捕获最终的适配器注册。没有已注册适配器的路由会保留原定配置，使 `llm/stream` 中间件可以接管并短路该请求；若仍未得到处理，最终分发会拒绝该路由。恢复后的主循环仅在初始提供方/模型路由未变时保留日志中记录的推理强度；如果路由发生变化，则丢弃上一模型的不透明 ID。
原生 DeepSeek 适配器声明 `high` 和 `max`，默认使用配置指定的推理强度，若未配置则使用 `high`；禁用思考时不暴露推理强度能力。pi-ai 适配器通过 `getSupportedThinkingLevels()` 按具体模型推导等级列表，排除 `off`，在 profile 未指定默认值时保留提供方默认行为，并将提供方协议值的映射留在 pi-ai 内部。
@@ -30,4 +30,4 @@ Status: implemented
客户端可以查询一条确切路由，并按适配器给出的顺序和名称渲染等级，而无需了解全局枚举。适配器配置仍负责提供部署默认值，`agent/request` 则可以在每个步骤替换实际生效的推理强度。元数据无效时抛出 `INVALID_MODEL_REASONING`；显式指定或配置指定的值不受支持时，会在提供方 I/O 前抛出 `UNSUPPORTED_REASONING_EFFORT`。
能力查询采用异步方式；对于由权威目录支持的适配器，确切模型解析可能失败。无密钥的服务、适配器、主循环、会话和请求头测试为校验、默认值解析、动态变更、日志记录和恢复行为提供回归保障；可运行快照锁定实际组装请求头中的已解析推理强度，仅在有密钥时运行的适配器测试则覆盖提供方序列化。
能力查询采用异步方式；对于由权威目录支持的适配器，确切模型解析可能失败。可选信号构成调用方的取消边界；异步适配器必须在信号中止后迅速完成结算，使主循环的资源释放达到完全停稳。无密钥的服务、适配器、主循环、会话和请求头测试为校验、默认值解析、动态变更、日志记录、恢复行为、HMR（热模块替换）期间的注册所有权和取消提供回归保障；可运行快照锁定实际组装请求头中的已解析推理强度，仅在有密钥时运行的适配器测试则覆盖提供方序列化。