Files
deepseek-harness/packages/acp/src/codec.ts
T
Tianyi Cui fb9636db44 feat(acp): ACP bridge — drive the coding agent from an editor over JSON-RPC stdio
Implements the RFC 010 MVP: a new `@deepseek-ai/dsh-acp` package bridges the
harness agent to the Agent Client Protocol (JSON-RPC 2.0 over newline-delimited
stdio), so Zed and other ACP editors can drive the coding agent — streaming
render, tool-call display, and resumable sessions via `session/load`.

- packages/acp: AgentSideConnection wiring; initialize/newSession/loadSession/
  prompt/cancel; a total TurnEndReason→StopReason codec; settle-once with a
  fallback chain (agent/turn-end → logged turn/end → idle); single-session
  guard; cwd-must-equal-launch-dir validation; load replays from the persisted
  event log (assistant/chunk→agent_message_chunk, tool/call/result→tool_call*).
- agent: add Agent.whenIdle() quiescence signal to the interface; LoopAgent
  implements it (resolves on the first running→idle/disposed transition). The
  bridge awaits it on disposal so teardown reaches quiescence, not just abort.
- examples: extract the shared provider/tool core into examples/base.yml;
  coding-agent nest-includes it; new examples/acp-agent serves the agent over
  ACP with JSONL persistence and no stdout logger (stdout is the protocol).
- Permission gate deferred (TODO(rfc010-permission-gate)): tools run with the
  executor's full authority; only the Agent→sessionId ownership seam is laid
  down. Cancel is best-effort for a not-yet-started queued turn
  (TODO(rfc010-cancel-prestep)). RFC 010 stays `proposed`.
- Docs: package README + Zed snippet; client-driver cookbook section; root and
  packages layout/commands; RFC 010 implementation-status note.

48 bridge tests + whenIdle coverage; 100% per-file coverage; e2e boots the
example as a subprocess and verifies a written file on disk (key-gated, with a
no-key stdout-purity check).
2026-06-16 18:44:31 +08:00

104 lines
4.7 KiB
TypeScript

/**
* Pure translation between harness vocabulary and ACP wire types. No I/O, no
* Cordis context — every function here is total and unit-testable in isolation.
* Keeping the mapping pure is deliberate: the SDK rejects an unknown
* `stopReason`, so the {@link turnEndToStopReason} total function (with its
* exhaustive test over every `TurnEndReason` kind) is the guard that a turn
* always settles to a legal wire value.
*
* @module @deepseek-ai/dsh-acp/codec
*/
import type { ContentBlock } from '@deepseek-ai/dsh-llm'
import type { TurnEndReason } from '@deepseek-ai/dsh-session'
import type { ContentBlock as AcpContentBlock, StopReason } from '@agentclientprotocol/sdk'
/**
* Map a harness {@link TurnEndReason} to the ACP `StopReason` wire enum.
*
* The mapping is total over the kinds the loop actually produces today
* (`completed`/`aborted`/`error`/`disposed`/`max-tokens`). `TurnEndReason` is
* merge-extensible, so an unknown future kind falls through to `end_turn` —
* the safest default (the turn DID end; we just lack a more specific wire
* reason) — rather than throwing into the SDK, which would reject an unknown
* `stopReason` and break the prompt RPC. When a new kind gains a dedicated ACP
* reason (e.g. a future `refusal` → `refusal`), add an explicit case here.
*
* - `completed` → `end_turn` (the model chose to stop)
* - `max-tokens` → `max_tokens` (cut off at the output-token ceiling)
* - `aborted` → `cancelled` (an `agent.abort()`, e.g. from `session/cancel`)
* - `error` → `end_turn` (defensive fallback only: the bridge REJECTS the
* `session/prompt` RPC on an error turn BEFORE calling this, so
* a client sees a JSON-RPC error, not a stop reason — see
* `rejectPrompt` in index.ts. This case keeps the function total
* for any non-bridge caller / property test.)
* - `disposed` → `cancelled` (the agent was torn down mid-turn — closest to a
* cancellation from the client's perspective)
*/
export function turnEndToStopReason(reason: TurnEndReason): StopReason {
switch (reason.kind) {
case 'completed':
return 'end_turn'
case 'max-tokens':
return 'max_tokens'
case 'aborted':
return 'cancelled'
case 'disposed':
return 'cancelled'
case 'error':
return 'end_turn'
// Merge-extensible: an unknown future TurnEndReason kind still has to
// produce a legal wire value (the SDK rejects unknown stopReason), so
// default to end_turn rather than assertNever. Add an explicit case when a
// new kind gains a dedicated ACP reason.
default:
return 'end_turn'
}
}
/**
* Translate a harness {@link ContentBlock} from a prompt into ACP content for
* replay, or `undefined` for block kinds the bridge does not surface to the
* client as message content. Today only `text` maps (text-only
* `promptCapabilities`); `reasoning` is surfaced via `agent_thought_chunk`
* streaming rather than as a message block, and `tool-call`/`tool-result`/
* `image` are handled by the tool-call update path or not advertised.
*/
export function harnessBlockToAcpContent(block: ContentBlock): AcpContentBlock | undefined {
switch (block.type) {
case 'text':
return { type: 'text', text: block.text }
// reasoning → streamed as agent_thought_chunk, not a message block
// tool-call / tool-result → the tool_call / tool_call_update path
// image → not advertised (text-only promptCapabilities)
default:
return undefined
}
}
/**
* Extract plain text from an ACP prompt's content blocks, concatenating every
* `text` block. Non-text blocks are ignored here; the caller rejects a prompt
* carrying image/audio per the advertised text-only capabilities BEFORE
* calling this, so dropping them here only affects `resource`/`resource_link`
* (which carry no inline text to forward in the MVP).
*/
export function acpPromptToText(prompt: readonly AcpContentBlock[]): string {
return prompt
.filter((block): block is AcpContentBlock & { type: 'text'; text: string } => block.type === 'text')
.map(block => block.text)
.join('')
}
/**
* Whether an ACP prompt contains any content the text-only bridge cannot
* accept — i.e. ANY non-`text` block (image, audio, `resource`, `resource_link`,
* …). The caller rejects such a prompt up front rather than silently dropping
* the unsupported parts: a prompt like `[text, resource_link]` carries context
* the model would otherwise never see, so running it text-only would be silent
* data loss. When richer block kinds are supported, narrow this.
*/
export function promptHasUnsupportedContent(prompt: readonly AcpContentBlock[]): boolean {
return prompt.some(block => block.type !== 'text')
}