lark-remote 0.1.13 → 0.2.1

package/dist/runner/kimi/acp/translator.d.ts

@@ -2,8 +2,11 @@
  * KimiAcpTranslator: translates ACP session/update notifications and
  * session/request_permission server requests into AgentEvents.
  *
- * Per-turn stateful: accumulates text/thinking deltas, tracks tool calls,
- * and produces turn_started events with operationKind.
+ * Extends BaseAcpTranslator (shared envelope dispatch, prompt-response
+ * stopReason mapping, tool_call normalization, usage occupancy). Kimi's
+ * divergence from opencode is the content channel: per-turn stateful —
+ * accumulates text/thinking deltas and emits turn_diff snapshots (NOT
+ * assistant deltas — see handleAgentMessageChunk), plus question elicitation.
  *
  * Wire envelope (R1, source: kimi-code packages/acp-server/src/events-map.ts,
  * verified live 2026-08-15/16 against kimi 0.36.0):
@@ -13,74 +16,28 @@
  * Event mapping (design doc §4.2):
  *   agent_message_chunk   → turn_diff text snapshot (content.text accumulation)
  *   agent_thought_chunk   → turn_diff reasoning snapshot (content.text accumulation)
- *   tool_call             → assistant/tool_use (rawInput object passthrough,
- *                           string → JSON.parse fallback)
- *   tool_call_update      → user/tool_result (status:'failed' → is_error:true)
- *   plan / available_commands_update / session_info_update /
- *   current_mode_update / config_option_update → discard
- *   usage_update          → live context occupancy {used, size} →
- *                           {total_tokens, context_limit}; NO input/output
- *                           split exists on the wire — cumulative token stats
- *                           fall back to wire.jsonl usage.record (dual path)
+ *   tool_call / tool_call_update / usage_update / plan / control-plane noise →
+ *                         base class (BaseAcpTranslator)
  *   prompt stopReason:'end_turn'   → result success
  *   prompt stopReason:'cancelled'  → result interrupted (独立终态)
  *   prompt other stopReason/error  → result error
  *   turn_started (self-produced)   → operationKind 'turn'/'compaction'
- *   session/request_permission     → approval_requested
+ *   session/request_permission     → approval_requested（含 question 桥）
+ *   elicitation/create             → question approval event
  */
-import type { AgentEvent } from '../../types.js';
-import type { ApprovalRequestedEvent } from '../../types.js';
-/** Self-produced turn_started event (no wire equivalent — produced on prompt start). */
-export interface AcpTurnStartedEvent {
-    type: 'turn_started';
-    threadId: string;
-    turnId: string;
-    operationKind: 'turn' | 'compaction';
-    timestamp?: string;
-}
-/**
- * Live usage snapshot from usage_update. The wire only carries context
- * occupancy (`used`) and the model context window (`size`) — there is NO
- * input/output token split (events-map.ts:505-516). input_tokens/output_tokens
- * stay undefined so the bridge falls back to wire.jsonl usage.record for
- * cumulative token stats (R1 dual path).
- */
-export interface AcpLiveUsage {
-    total_tokens?: number;
-    context_limit?: number;
-    input_tokens?: number;
-    output_tokens?: number;
-}
-/** Usage update event for live token display. */
-export interface AcpUsageEvent {
-    type: 'usage';
-    usage: AcpLiveUsage;
-    timestamp?: string;
-}
-export type AcpTranslatorEvent = AgentEvent | AcpTurnStartedEvent | AcpUsageEvent | ApprovalRequestedEvent;
-export declare class KimiAcpTranslator {
+import { type RequestPermissionParams, type AgentMessageChunkEvent, type AgentThoughtChunkEvent } from '../../common/acp/protocol-types.js';
+import { BaseAcpTranslator, type AcpTranslatorEvent } from '../../common/acp/base-translator.js';
+export type { AcpTurnStartedEvent, AcpLiveUsage, AcpUsageEvent, AcpTranslatorEvent, } from '../../common/acp/base-translator.js';
+export declare class KimiAcpTranslator extends BaseAcpTranslator {
+    protected readonly logTag = "kimi-acp-translator";
     /** Accumulated text per tool-call or message (keyed by a logical item id). */
     private textByItem;
     /** Accumulated thinking per item. */
     private thinkingByItem;
-    /** Pending tool calls keyed by toolCallId (for matching tool_call_update). */
-    private pendingToolCalls;
-    /** Current operation kind (turn vs compact). */
-    private operationKind;
-    /** Live usage snapshot from usage_update events (context occupancy only). */
-    private liveUsage;
-    /** Whether any usage_update has been seen this turn (no update → no usage). */
-    private hasLiveUsage;
-    /** Current turn id, set by produceTurnStarted; carried on turn_diff events. */
-    private currentTurnId;
-    /**
-     * Handle a notification from the ACP server and return translated events.
-     */
-    handleNotification(method: string, params: unknown): AcpTranslatorEvent[];
-    /**
-     * Handle a server request (reverse RPC from the ACP server).
-     */
-    handleServerRequest(id: number | string, method: string, params: unknown): AcpTranslatorEvent[];
+    protected handleAgentMessageChunk(event: AgentMessageChunkEvent, sessionId: string): AcpTranslatorEvent[];
+    protected handleAgentThoughtChunk(event: AgentThoughtChunkEvent, sessionId: string): AcpTranslatorEvent[];
+    protected handleRequestPermission(requestId: number | string, params: RequestPermissionParams): AcpTranslatorEvent[];
+    protected handleOtherServerRequest(requestId: number | string, method: string, params: unknown): AcpTranslatorEvent[];
     /**
      * Kimi elicitation form：requestedSchema.properties 按顺序解析为 UserQuestion[]
      * （type:array → 多选；oneOf/anyOf const → 选项 label；title → header??question）。
@@ -89,49 +46,10 @@ export declare class KimiAcpTranslator {
      * form 会丢弃非声明选项值 → isOther=false（卡片隐藏自定义答案输入）。
      */
     private handleElicitationCreate;
-    /**
-     * Produce a turn_started event. Called by the runner when session/prompt
-     * is sent (no wire notification for this — self-produced).
-     */
-    produceTurnStarted(sessionId: string, turnId: string): AcpTurnStartedEvent;
-    /**
-     * Translate a prompt response into a result event. Carries the live usage
-     * snapshot only when a usage_update was seen (otherwise undefined, so the
-     * bridge falls back to wire.jsonl usage.record readback).
-     */
-    handlePromptResponse(sessionId: string, result: {
-        stopReason: string;
-    }): AgentEvent;
-    /**
-     * Produce an error result event (for JSON-RPC errors during prompt).
-     */
-    produceErrorResult(sessionId: string, errorMessage: string): AgentEvent;
-    /**
-     * Set the operation kind (turn vs compact).
-     */
-    setOperationKind(kind: 'turn' | 'compact'): void;
-    private handleAgentMessageChunk;
-    private handleAgentThoughtChunk;
-    private handleToolCall;
-    private handleToolCallUpdate;
-    private handleUsageUpdate;
-    private handleRequestPermission;
     /**
      * request_permission 兜底桥（elicitation/create 失败或旧版 kimi）：
      * 选项为 q{n}_opt_{i}（allow_once，label 直取 name）+ q{n}_skip（reject_once，
      * 跳过语义）。单题单选、无 Other（kimi 桥丢弃非声明值）。
      */
     private handlePermissionQuestion;
-    /**
-     * Derive the approval decision list from the server's offered options.
-     * accept/decline/cancel are universal; acceptForSession requires an
-     * always-class option (kimi `approve_always`, opencode `allow_always`).
-     */
-    /**
-     * ResultEvent.usage requires input_tokens/output_tokens: number, but the
-     * ACP wire has no such split — the live snapshot is context occupancy only.
-     * Cast is deliberate: consumers must tolerate undefined input/output (the
-     * bridge's kimi path already falls back to wire.jsonl readback).
-     */
-    private liveUsageSnapshot;
 }

package/dist/runner/kimi/acp/translator.js

@@ -2,8 +2,11 @@
  * KimiAcpTranslator: translates ACP session/update notifications and
  * session/request_permission server requests into AgentEvents.
  *
- * Per-turn stateful: accumulates text/thinking deltas, tracks tool calls,
- * and produces turn_started events with operationKind.
+ * Extends BaseAcpTranslator (shared envelope dispatch, prompt-response
+ * stopReason mapping, tool_call normalization, usage occupancy). Kimi's
+ * divergence from opencode is the content channel: per-turn stateful —
+ * accumulates text/thinking deltas and emits turn_diff snapshots (NOT
+ * assistant deltas — see handleAgentMessageChunk), plus question elicitation.
  *
  * Wire envelope (R1, source: kimi-code packages/acp-server/src/events-map.ts,
  * verified live 2026-08-15/16 against kimi 0.36.0):
@@ -13,25 +16,20 @@
  * Event mapping (design doc §4.2):
  *   agent_message_chunk   → turn_diff text snapshot (content.text accumulation)
  *   agent_thought_chunk   → turn_diff reasoning snapshot (content.text accumulation)
- *   tool_call             → assistant/tool_use (rawInput object passthrough,
- *                           string → JSON.parse fallback)
- *   tool_call_update      → user/tool_result (status:'failed' → is_error:true)
- *   plan / available_commands_update / session_info_update /
- *   current_mode_update / config_option_update → discard
- *   usage_update          → live context occupancy {used, size} →
- *                           {total_tokens, context_limit}; NO input/output
- *                           split exists on the wire — cumulative token stats
- *                           fall back to wire.jsonl usage.record (dual path)
+ *   tool_call / tool_call_update / usage_update / plan / control-plane noise →
+ *                         base class (BaseAcpTranslator)
  *   prompt stopReason:'end_turn'   → result success
  *   prompt stopReason:'cancelled'  → result interrupted (独立终态)
  *   prompt other stopReason/error  → result error
  *   turn_started (self-produced)   → operationKind 'turn'/'compaction'
- *   session/request_permission     → approval_requested
+ *   session/request_permission     → approval_requested（含 question 桥）
+ *   elicitation/create             → question approval event
  */
 import { makeQuestionApprovalEvent } from '../../question-common.js';
-import { NotificationMethod, ServerRequestMethod, SessionEventType, } from '../../common/acp/protocol-types.js';
+import { ServerRequestMethod, } from '../../common/acp/protocol-types.js';
 import { deriveAcpAvailableDecisions, truncateWithEllipsis, } from '../../common/acp/protocol-helpers.js';
 import { getLogger } from '../../../logger/index.js';
+import { BaseAcpTranslator } from '../../common/acp/base-translator.js';
 /**
  * Fixed snapshot item ids: ACP has exactly one text stream and one thinking
  * stream per turn (no itemId on the wire), so turn_diff snapshots use these
@@ -42,153 +40,14 @@ const THINKING_ITEM_ID = 'thinking';
 // =============================================================================
 // Translator
 // =============================================================================
-export class KimiAcpTranslator {
+export class KimiAcpTranslator extends BaseAcpTranslator {
+    logTag = 'kimi-acp-translator';
     /** Accumulated text per tool-call or message (keyed by a logical item id). */
     textByItem = new Map();
     /** Accumulated thinking per item. */
     thinkingByItem = new Map();
-    /** Pending tool calls keyed by toolCallId (for matching tool_call_update). */
-    pendingToolCalls = new Map();
-    /** Current operation kind (turn vs compact). */
-    operationKind = 'turn';
-    /** Live usage snapshot from usage_update events (context occupancy only). */
-    liveUsage = {};
-    /** Whether any usage_update has been seen this turn (no update → no usage). */
-    hasLiveUsage = false;
-    /** Current turn id, set by produceTurnStarted; carried on turn_diff events. */
-    currentTurnId = '';
-    /**
-     * Handle a notification from the ACP server and return translated events.
-     */
-    handleNotification(method, params) {
-        if (method !== NotificationMethod.SESSION_UPDATE) {
-            return [];
-        }
-        // R1: real envelope is {sessionId, update: {sessionUpdate: ...}}.
-        const notif = params;
-        const update = notif.update;
-        switch (update.sessionUpdate) {
-            case SessionEventType.AGENT_MESSAGE_CHUNK:
-                return this.handleAgentMessageChunk(update, notif.sessionId);
-            case SessionEventType.AGENT_THOUGHT_CHUNK:
-                return this.handleAgentThoughtChunk(update, notif.sessionId);
-            case SessionEventType.TOOL_CALL:
-                return this.handleToolCall(update, notif.sessionId);
-            case SessionEventType.TOOL_CALL_UPDATE:
-                return this.handleToolCallUpdate(update, notif.sessionId);
-            case SessionEventType.PLAN:
-                getLogger().debug(`[kimi-acp-translator] discarding plan event`);
-                return [];
-            case SessionEventType.AVAILABLE_COMMANDS_UPDATE:
-            case SessionEventType.SESSION_INFO_UPDATE:
-            case SessionEventType.CURRENT_MODE_UPDATE:
-            case SessionEventType.CONFIG_OPTION_UPDATE:
-                // Control-plane noise (command list, mode/config echoes) — not content.
-                return [];
-            case SessionEventType.USAGE_UPDATE:
-                return this.handleUsageUpdate(update);
-            default:
-                return [];
-        }
-    }
-    /**
-     * Handle a server request (reverse RPC from the ACP server).
-     */
-    handleServerRequest(id, method, params) {
-        if (method === ServerRequestMethod.REQUEST_PERMISSION) {
-            return this.handleRequestPermission(id, params);
-        }
-        if (method === ServerRequestMethod.ELICITATION_CREATE) {
-            return this.handleElicitationCreate(id, params);
-        }
-        return [];
-    }
-    /**
-     * Kimi elicitation form：requestedSchema.properties 按顺序解析为 UserQuestion[]
-     * （type:array → 多选；oneOf/anyOf const → 选项 label；title → header??question）。
-     * 题面完整文本只在 form 级 message 里合并出现（逐字段 title 可能只是短 header），
-     * 因此 message 作为卡片概要行（view.intro）透传。
-     * form 会丢弃非声明选项值 → isOther=false（卡片隐藏自定义答案输入）。
-     */
-    handleElicitationCreate(requestId, params) {
-        const properties = params.requestedSchema?.properties ?? {};
-        const questions = [];
-        for (const [key, prop] of Object.entries(properties)) {
-            const q = elicitationPropertyToQuestion(prop);
-            if (!q) {
-                getLogger().warn(`[kimi-acp-translator] elicitation property skipped: key=${key}`);
-                continue;
-            }
-            questions.push(q);
-        }
-        if (questions.length === 0) {
-            // 无法解析 → 交还 runner 自动响应（不悬挂服务端）。
-            return [];
-        }
-        const message = params.message?.trim();
-        const single = questions.length === 1;
-        const intro = message && (!single || message !== questions[0].question) ? message : undefined;
-        const event = makeQuestionApprovalEvent(requestId, questions, params.sessionId);
-        if (intro)
-            event.view.intro = intro;
-        return [event];
-    }
-    /**
-     * Produce a turn_started event. Called by the runner when session/prompt
-     * is sent (no wire notification for this — self-produced).
-     */
-    produceTurnStarted(sessionId, turnId) {
-        this.currentTurnId = turnId;
-        return {
-            type: 'turn_started',
-            threadId: sessionId,
-            turnId,
-            operationKind: this.operationKind === 'compact' ? 'compaction' : 'turn',
-            timestamp: new Date().toISOString(),
-        };
-    }
-    /**
-     * Translate a prompt response into a result event. Carries the live usage
-     * snapshot only when a usage_update was seen (otherwise undefined, so the
-     * bridge falls back to wire.jsonl usage.record readback).
-     */
-    handlePromptResponse(sessionId, result) {
-        // §4.2: cancelled is independent terminal state, must NOT merge into error
-        const subtype = result.stopReason === 'end_turn'
-            ? 'success'
-            : result.stopReason === 'cancelled'
-                ? 'interrupted'
-                : 'error';
-        return {
-            type: 'result',
-            subtype,
-            session_id: sessionId,
-            ...(this.hasLiveUsage ? { usage: this.liveUsageSnapshot() } : {}),
-            ...(subtype === 'error'
-                ? { errorMessage: `Prompt ended with stopReason: ${result.stopReason}` }
-                : {}),
-        };
-    }
-    /**
-     * Produce an error result event (for JSON-RPC errors during prompt).
-     */
-    produceErrorResult(sessionId, errorMessage) {
-        return {
-            type: 'result',
-            subtype: 'error',
-            session_id: sessionId,
-            errorMessage,
-            ...(this.hasLiveUsage ? { usage: this.liveUsageSnapshot() } : {}),
-        };
-    }
-    /**
-     * Set the operation kind (turn vs compact).
-     */
-    setOperationKind(kind) {
-        this.operationKind = kind;
-    }
     // =========================================================================
-    // Private notification handlers
+    // Content channel: wire deltas → turn_diff snapshots
     // =========================================================================
     handleAgentMessageChunk(event, sessionId) {
         // Use a stable key for the current text stream. Kimi ACP doesn't
@@ -214,6 +73,10 @@ export class KimiAcpTranslator {
                 threadId: sessionId,
                 turnId: this.currentTurnId,
                 timestamp: new Date().toISOString(),
+                // 卡片块锚点时间刷新为本次 diff 时间：'text' 是整轮累积流，工具间隙
+                // 仍会续写增量；首见时间（如 08:39）会让标题与 move-to-end 后的位置
+                // 看起来“乱序”。这里显式声明 diff 时间即块的最新写入时间。
+                refreshTimestamp: true,
             },
         ];
     }
@@ -233,80 +96,12 @@ export class KimiAcpTranslator {
                 threadId: sessionId,
                 turnId: this.currentTurnId,
                 timestamp: new Date().toISOString(),
+                refreshTimestamp: true,
             },
         ];
     }
-    handleToolCall(event, _sessionId) {
-        // rawInput arrives as an object on the wire; older kimi versions sent a
-        // JSON string — parse defensively in that case. The acp-server's
-        // lazy-create tool_call (first args delta, events-map.ts
-        // toolCallLazyCreateToSessionUpdate) carries NO rawInput field at all —
-        // normalize to an empty object so the card renders an empty args summary
-        // instead of crashing on undefined input (2026-08-17 live TypeError).
-        let input = event.rawInput ?? {};
-        if (typeof event.rawInput === 'string') {
-            try {
-                input = JSON.parse(event.rawInput);
-            }
-            catch {
-                // Keep raw string as-is
-            }
-        }
-        this.pendingToolCalls.set(event.toolCallId, {
-            id: event.toolCallId,
-            name: event.title,
-            input,
-        });
-        return [
-            {
-                type: 'assistant',
-                message: {
-                    content: [
-                        {
-                            type: 'tool_use',
-                            id: event.toolCallId,
-                            name: event.title,
-                            input,
-                        },
-                    ],
-                },
-                timestamp: new Date().toISOString(),
-            },
-        ];
-    }
-    handleToolCallUpdate(event, _sessionId) {
-        // §4.2: status:'failed' → is_error:true (fixes CLI mode's is_error always false)
-        const isError = event.status === 'failed';
-        const content = event.rawOutput ?? '';
-        return [
-            {
-                type: 'user',
-                message: {
-                    content: [
-                        {
-                            type: 'tool_result',
-                            tool_use_id: event.toolCallId,
-                            content,
-                            is_error: isError,
-                        },
-                    ],
-                },
-                timestamp: new Date().toISOString(),
-            },
-        ];
-    }
-    handleUsageUpdate(event) {
-        // R1: usage_update is {sessionUpdate:'usage_update', used, size}
-        // (events-map.ts:505-516) — context occupancy, not cumulative tokens.
-        this.liveUsage = {
-            total_tokens: event.used,
-            context_limit: event.size,
-        };
-        this.hasLiveUsage = true;
-        return [];
-    }
     // =========================================================================
-    // Private approval handlers
+    // Approval + question elicitation
     // =========================================================================
     handleRequestPermission(requestId, params) {
         // §5.4 升级：question elicitation 识别（isQuestion 标记 / AskUserQuestion
@@ -349,6 +144,42 @@ export class KimiAcpTranslator {
             },
         ];
     }
+    handleOtherServerRequest(requestId, method, params) {
+        if (method !== ServerRequestMethod.ELICITATION_CREATE) {
+            return [];
+        }
+        return this.handleElicitationCreate(requestId, params);
+    }
+    /**
+     * Kimi elicitation form：requestedSchema.properties 按顺序解析为 UserQuestion[]
+     * （type:array → 多选；oneOf/anyOf const → 选项 label；title → header??question）。
+     * 题面完整文本只在 form 级 message 里合并出现（逐字段 title 可能只是短 header），
+     * 因此 message 作为卡片概要行（view.intro）透传。
+     * form 会丢弃非声明选项值 → isOther=false（卡片隐藏自定义答案输入）。
+     */
+    handleElicitationCreate(requestId, params) {
+        const properties = params.requestedSchema?.properties ?? {};
+        const questions = [];
+        for (const [key, prop] of Object.entries(properties)) {
+            const q = elicitationPropertyToQuestion(prop);
+            if (!q) {
+                getLogger().warn(`[kimi-acp-translator] elicitation property skipped: key=${key}`);
+                continue;
+            }
+            questions.push(q);
+        }
+        if (questions.length === 0) {
+            // 无法解析 → 交还 runner 自动响应（不悬挂服务端）。
+            return [];
+        }
+        const message = params.message?.trim();
+        const single = questions.length === 1;
+        const intro = message && (!single || message !== questions[0].question) ? message : undefined;
+        const event = makeQuestionApprovalEvent(requestId, questions, params.sessionId);
+        if (intro)
+            event.view.intro = intro;
+        return [event];
+    }
     /**
      * request_permission 兜底桥（elicitation/create 失败或旧版 kimi）：
      * 选项为 q{n}_opt_{i}（allow_once，label 直取 name）+ q{n}_skip（reject_once，
@@ -370,23 +201,6 @@ export class KimiAcpTranslator {
         ];
         return [makeQuestionApprovalEvent(requestId, questions, params.sessionId)];
     }
-    /**
-     * Derive the approval decision list from the server's offered options.
-     * accept/decline/cancel are universal; acceptForSession requires an
-     * always-class option (kimi `approve_always`, opencode `allow_always`).
-     */
-    // =========================================================================
-    // Private helpers
-    // =========================================================================
-    /**
-     * ResultEvent.usage requires input_tokens/output_tokens: number, but the
-     * ACP wire has no such split — the live snapshot is context occupancy only.
-     * Cast is deliberate: consumers must tolerate undefined input/output (the
-     * bridge's kimi path already falls back to wire.jsonl readback).
-     */
-    liveUsageSnapshot() {
-        return { ...this.liveUsage };
-    }
 }
 // =============================================================================
 // Helpers

package/dist/runner/opencode/acp/runner.d.ts

@@ -7,10 +7,8 @@
  * → synthetic init → session/prompt → consume session/update notifications
  * until prompt settles → result event.
  *
- * Structurally aligned with KimiAcpRunner via the shared ConnectionBasedRunner
- * base (same notification queue + waitResolve pattern, same forceFinish/
- * stopRequested semantics, same turn idle timeout, same ConnectionLostError
- * retry). Differences from kimi:
+ * Turn/connection/compact/approval orchestration lives in BaseAcpRunner;
+ * opencode-specific parts:
  * - session/set_mode applies the configured mode (build/plan; default 'build'
  *   keeps manual approvals flowing); opencode sends no set_mode notification,
  *   so the client refreshes the local mode view itself
@@ -18,11 +16,15 @@
  *   service.ts:555-571 summarize completes before the prompt response)
  * - text/thinking go to the assistant incremental channel (opencode wire
  *   deltas), not turn_diff snapshots
+ * - OpencodeLogErrorMonitor tails opencode's own log for LLM stream errors
+ *   the ACP protocol never forwards
  */
-import type { AgentKind, AgentSessionReader, AgentEvent, AgentStatusInfo, SpawnOptions } from '../../types.js';
+import type { AgentKind, AgentSessionReader, AgentStatusInfo } from '../../types.js';
 import { JsonRpcClient } from '../../common/jsonrpc/client.js';
-import { type OpencodeAcpTranslatorEvent } from './translator.js';
-import { ConnectionBasedRunner } from '../../common/connection-based-runner.js';
+import { OpencodeAcpTranslator } from './translator.js';
+import { type SessionNewResult, type SessionResumeResult, type SessionPromptResult } from '../../common/acp/protocol-types.js';
+import { type AcpPendingApproval } from '../../common/acp/protocol-helpers.js';
+import { BaseAcpRunner } from '../../common/acp/base-acp-runner.js';
 export interface OpencodeAcpRunnerOptions {
     kind: AgentKind;
     sessionReader: AgentSessionReader;
@@ -48,11 +50,7 @@ export interface OpencodeAcpRunnerOptions {
     /** Tail poll interval for the opencode log error monitor. Defaults to 2000ms. */
     errorMonitorPollIntervalMs?: number;
 }
-export declare class OpencodeAcpRunner extends ConnectionBasedRunner<JsonRpcClient, OpencodeAcpTranslatorEvent> {
-    private connectionManager;
-    private currentTranslator;
-    private activeSessionId;
-    private model?;
+export declare class OpencodeAcpRunner extends BaseAcpRunner<OpencodeAcpTranslator> {
     /** Configured session mode applied via session/set_mode (§P5). */
     private configuredMode;
     /** Current session mode id, parsed from session/new|resume configOptions. */
@@ -62,22 +60,12 @@ export declare class OpencodeAcpRunner extends ConnectionBasedRunner<JsonRpcClie
     /** Tails opencode's own log to surface LLM stream errors the ACP protocol
      *  never forwards (quota/limit errors back off for ~1h → turn would hang). */
     private readonly errorMonitor;
-    /** Pending approval requests: requestId → kind + view + options. */
-    private pendingApprovals;
-    /** Tracker for the in-flight prompt request — needed for cancellation. */
-    private promptSettled;
-    /** Whether the prompt request has been fired (distinguishes "stop before
-     *  prompt sent" from "stop while prompt in-flight" — the latter sends
-     *  session/cancel immediately). */
-    private promptSent;
     constructor(opts: OpencodeAcpRunnerOptions);
     protected get logTag(): string;
     protected get turnTimeoutErrorMessage(): string;
     protected get turnInterruptedErrorMessage(): string;
-    protected currentSessionId(): string | null;
+    protected get connectionClosedMessage(): string;
     protected shouldDeferStop(): boolean;
-    protected cancelCurrentTurn(): Promise<void>;
-    protected clearTurnState(): void;
     /**
      * Start watching opencode's own log for LLM stream errors of the active
      * session. opencode ACP does not surface retry/error state to the client
@@ -88,23 +76,14 @@ export declare class OpencodeAcpRunner extends ConnectionBasedRunner<JsonRpcClie
      * until the idle watchdog.
      */
     private startErrorMonitor;
-    protected releaseConnection(cwd: string): Promise<void>;
-    protected notifyIdle(cwd: string): void;
-    protected disposeConnections(): Promise<void>;
+    protected createTranslator(): OpencodeAcpTranslator;
+    protected applyTurnSettings(client: JsonRpcClient, sessionId: string): Promise<void>;
     /**
-     * Run a compact operation on the current session.
-     *
-     * opencode special-cases the `/compact` text prompt server-side
-     * (service.ts:555-571): it runs session.summarize synchronously and the
-     * prompt response settles only after compaction completes — no background
-     * record polling needed (unlike kimi).
-     */
-    runCompact(_message: string, opts: SpawnOptions): AsyncGenerator<AgentEvent>;
-    /**
-     * Respond to an approval server request. `response` is the bridge
-     * ApprovalAction (`{ action: 'accept' | 'decline' | 'cancel' }`).
+     * opencode /compact is synchronous server-side — standard prompt-settle
+     * translation (no wire.jsonl polling, unlike kimi).
      */
-    respondApproval(requestId: number | string, response: unknown): Promise<void>;
+    protected onCompactPromptFired(promptPromise: Promise<SessionPromptResult>, translator: OpencodeAcpTranslator, sessionId: string, _opts: unknown, _baselineTotalCount: number): void;
+    protected buildApprovalOutcome(action: string, pending: AcpPendingApproval, _response: unknown): unknown;
     getStatusInfo(): AgentStatusInfo;
     /**
      * Hot-apply a mode change to the live ACP session (§P5).
@@ -121,18 +100,15 @@ export declare class OpencodeAcpRunner extends ConnectionBasedRunner<JsonRpcClie
     updateApprovalMode(settings: {
         mode?: 'build' | 'plan';
     }): Promise<void>;
-    /**
-     * Acquire the connection and set up the turn:
-     * session/new (new session) or session/resume (existing sessionId)
-     * → session/prompt.
-     */
-    protected setupTurn(message: string, opts: SpawnOptions): Promise<void>;
     /**
      * Track the current session mode and model from a session/new|resume
      * result's configOptions (opencode buildModeSelectOption/buildModelSelectOption:
      * {id, category, currentValue}).
      */
     private trackConfigOptions;
-    private handleNotification;
-    private handleServerRequest;
+    protected onTurnSessionReady(sessionId: string): void;
+    protected onSessionEstablished(result: SessionResumeResult | SessionNewResult): void;
+    protected onClearTurnState(): void;
+    protected handleNotification(method: string, params: unknown): void;
+    protected handleServerRequest(id: number | string, method: string, params: unknown): void;
 }