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

docs: replace vague provenance prose with recorded facts

Browse Source
This commit is contained in:
Turtle
2026-08-09 15:35:02 +08:00
parent 8c124f84b6
commit 9704749b01
380 changed files with 946 additions and 874 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
docs/subsystems/llm-streaming.md
View File

@@ -31,10 +31,10 @@ The block interfaces (full fields in source): `TextBlock` (`text`), `ReasoningBl
Source: [`packages/llm/llm/src/message.ts`](../../packages/llm/llm/src/message.ts)
A `Message` is one identified, immutable role/source/content value. Model-produced assistant messages carry provider/model ownership and optional adapter-private replay metadata in their source:
A `Message` is one identified, immutable role/source/content value. Model-produced assistant messages name the provider and model that produced them and carry optional adapter-private replay data in their source:
```ts type-equiv
/** Provider ownership and adapter-private replay data for an assistant message. */
/** Provider/model identity and adapter-private replay data for an assistant message. */
interface AssistantProvenance {
/** Provider route that produced the message. */
provider: string
@@ -58,7 +58,7 @@ interface Message {
readonly role: 'system' | 'user' | 'assistant'
/** Exact model-facing blocks. */
readonly content: ContentBlock[]
/** Required producer provenance. */
/** Required source fields supplied by the producer. */
readonly source: MessageSource
}
```
@@ -78,12 +78,12 @@ interface MessageSourceMap {
}
```
Provenance and shape are two independent axes. `kind` answers *who produced this*; the optional `form` a producer mixes in answers *what shape of information it is*, so several producers may share one presentation and one producer may emit more than one shape over a session. The vocabulary is semantic and grows one value at a time; an absent or unrecognized value is the documented default, presented as opaque content:
Producer identity and content shape are independent. `kind` answers *who produced this*; the optional `form` a producer mixes in answers *what shape of information it is*, so several producers may share one presentation and one producer may emit more than one shape over a session. The vocabulary is semantic and grows one value at a time; an absent or unrecognized value is the documented default, presented as opaque content:
```ts type-equiv
/**
* What SHAPE of information a producer-supplied context carries, declared by
* the producer beside its provenance.
* the producer beside the source fields it supplied.
*
* `MessageSource.kind` answers *who produced this*; `form` answers *what kind
* of thing it is*, and the two axes are deliberately independent — several
@@ -210,7 +210,7 @@ Every adapter MUST obey these, and every consumer may rely on them:
- **Context overflow has one canonical code.** Both DeepSeek adapters classify explicit provider detail through `isContextWindowExceededError()` and surface `CONTEXT_WINDOW_EXCEEDED`, whether the failure arrives as a thrown HTTP `LlmError` or an in-band finish error. Consumers route on the code, never provider text.
- **An empty completion is a retryable error, not a silent success.** Both adapters map a terminal `stop` finish that carried no content blocks to `finish {kind:'error'}` with the canonical `EMPTY_RESPONSE` code, and `dsh-llm-retry` retries it by default; see [empty model responses are retryable](../../.agents/notes/implemented/bug-fix/2026-07-24-empty-model-response-is-retryable.md).
- **Every provider HTTP request carries the app-attribution header.** Adapters send `attributionHeaders()` (below) - the `User-Agent` baseline - and prove it with a wire-level test (mock server asserting the received header, or the library's header hook for a library-backed adapter).
- **Replay state is adapter-owned.** A successful `finish` may carry lossless-JSON state needed to reconstruct a native provider response. The loop stores it with the assembled assistant message. On a later request, `LlmService` passes the state only when the historical provider and target provider are currently registered to the exact same adapter instance. That adapter validates the state and owns any cross-model or cross-provider conversion; other adapters receive the provider-neutral content and provenance without the private state.
- **Replay state is adapter-owned.** A successful `finish` may carry lossless-JSON state needed to reconstruct a native provider response. The loop stores it with the assembled assistant message. On a later request, `LlmService` passes the state only when the historical provider and target provider are currently registered to the exact same adapter instance. That adapter validates the state and owns any cross-model or cross-provider conversion; other adapters receive the provider-neutral content plus provider/model fields without the private state.
This contract is pinned down by two deliberately independent implementations: `dsh-llm-deepseek` (direct fetch, SSE framing via `eventsource-parser`) and `dsh-llm-pi-ai` (a generic multi-provider adapter through `@earendil-works/pi-ai`). The library-backed adapter exercises the finish-chunk error path, while transport-boundary tests prove each idle watchdog stops its actual request.
@@ -264,7 +264,7 @@ interface TokenUsage {
## `BlockAssembler`
`BlockAssembler` ([`packages/llm/llm/src/assembler.ts`](../../packages/llm/llm/src/assembler.ts)) is the single shared implementation that folds a `StreamChunk` stream back into `ContentBlock`s, usage, finish reason, and replay state. The loop logs the raw chunks while feeding the same chunks through an assembler, then stores the assembled assistant content with its provider/model provenance. A consumer that needs the assembled result without re-implementing the fold uses this.
`BlockAssembler` ([`packages/llm/llm/src/assembler.ts`](../../packages/llm/llm/src/assembler.ts)) is the single shared implementation that folds a `StreamChunk` stream back into `ContentBlock`s, usage, finish reason, and replay state. The loop logs the raw chunks while feeding the same chunks through an assembler, then stores the assembled assistant content with the provider and model that produced it. A consumer that needs the assembled result without re-implementing the fold uses this.
```ts public-api
/**
@@ -585,9 +585,9 @@ interface LlmDiscoveredModel {
### The request envelope: `LlmCallConfig` and the logged header
The loop builds each request from logged state. `EpochHeader` records call config, adapter-default provenance, rendered prompt, and authoritative returned tool order (configured by `toolOrder`, or lexicographic when unset) through full `request/header` snapshots. Together with derived history, this makes the request reconstructable from the session log. See [session.md](session.md#the-request-header-event-requestheader) and the [reconstructability Agent Note](../../.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.md).
The loop builds each request from logged state. `EpochHeader` records call config, marks the fields supplied by adapter defaults, and records the rendered prompt and authoritative returned tool order (configured by `toolOrder`, or lexicographic when unset) through full `request/header` snapshots. Together with derived history, this makes the request reconstructable from the session log. See [session.md](session.md#the-request-header-event-requestheader) and the [reconstructability Agent Note](../../.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.md).
`agent/request` receives a frozen call-config seed and may return a replacement to switch provider, model, reasoning effort, or sampling. Before the waterfall, the loop removes values marked as adapter defaults so exact-model preparation materializes the selected route's current values; unmarked explicit settings remain in the proposal. After the waterfall, preparation rejects unsupported explicit effort ids without clamping and logs the effective config plus provenance under the turn signal. The prepared call keeps one adapter registration through dispatch. Requests reaching `llm/stream` are deep-frozen, so mutation throws, and carry a process-local loop identity so observers do not confuse separately logged frozen auxiliary calls with conversation requests.
`agent/request` receives a frozen call-config seed and may return a replacement to switch provider, model, reasoning effort, or sampling. Before the waterfall, the loop removes values marked as adapter defaults so exact-model preparation materializes the selected route's current values; unmarked explicit settings remain in the proposal. After the waterfall, preparation rejects unsupported explicit effort ids without clamping and logs the effective config plus the fields supplied by adapter defaults under the turn signal. The prepared call keeps one adapter registration through dispatch. Requests reaching `llm/stream` are deep-frozen, so mutation throws, and carry a process-local loop identity so observers do not confuse separately logged frozen auxiliary calls with conversation requests.
On the wire, a loop-built request reads the `system` slot (the rendered prompt assembly) followed by the derived history — the boundary snapshot, whose tail is the newest `user/message` on a turn's first step and the previous step's tool results on later steps. The dev invariant recomputes exactly this equation against every loop-built request.