pi-harness-delegate 0.2.2 → 0.3.0

package/extensions/harnesses/opencode.ts

@@ -9,14 +9,19 @@ import type {
   StreamedResult,
 } from './types.ts';
 
+// Schema verified against opencode 1.18.16 — see tests/fixtures/opencode.jsonl.
+// `opencode run` in this version has no `--permission` or `--add-dir` flag at all (confirmed via
+// `opencode run --help`) — permission tiers map onto built-in agents instead (`opencode agent
+// list`: "plan" is the read-only-oriented primary agent, "build" is the full-access one).
+
 const execFileAsync = promisify(execFile);
 function isRecord(v: unknown): v is Record<string, unknown> {
   return typeof v === 'object' && v !== null && !Array.isArray(v);
 }
-const PERMISSION_MAP: Record<NormalizedPermission, string> = {
-  readonly: 'read-only',
-  edit: 'allow-edit',
-  danger: 'danger',
+const AGENT_MAP: Record<NormalizedPermission, string> = {
+  readonly: 'plan',
+  edit: 'build',
+  danger: 'build',
 };
 function extractOpencodeText(o: Record<string, unknown>): string | undefined {
   if (typeof o.text === 'string') return o.text;
@@ -28,6 +33,23 @@ function extractOpencodeText(o: Record<string, unknown>): string | undefined {
   if (isRecord(o.message) && typeof o.message.text === 'string') return o.message.text;
   return undefined;
 }
+
+interface OpencodeHarnessState {
+  sessionId?: string;
+  costAccum?: number;
+  inputAccum?: number;
+  outputAccum?: number;
+  cacheReadAccum?: number;
+  cacheWriteAccum?: number;
+  stepCount?: number;
+}
+
+function harnessState(state: ParseState): OpencodeHarnessState {
+  const s = (state._harness ?? {}) as OpencodeHarnessState;
+  state._harness = s as unknown as Record<string, unknown>;
+  return s;
+}
+
 export function parseOpencodeLine(line: string, state: ParseState): ParseOutcome {
   let o: unknown;
   try {
@@ -40,31 +62,33 @@ export function parseOpencodeLine(line: string, state: ParseState): ParseOutcome
   const activities: ParseOutcome['activities'] = [];
   let streamedText: string | undefined;
   const typeStr = typeof o.type === 'string' ? o.type : '';
-  // latch sessionID from step_start
-  if (typeStr === 'step_start' && typeof o.sessionID === 'string') {
-    (state as unknown as Record<string, unknown>)._harness = {
-      ...(((state as unknown as Record<string, unknown>)._harness as Record<string, unknown>) ?? {}),
-      sessionId: o.sessionID,
-    };
+  const hs = harnessState(state);
+  // latch sessionID — real events carry it top-level (not just nested under part)
+  if (typeof o.sessionID === 'string' && !hs.sessionId) hs.sessionId = o.sessionID;
+  if (isRecord(o.part) && typeof (o.part as Record<string, unknown>).sessionID === 'string' && !hs.sessionId) {
+    hs.sessionId = (o.part as Record<string, unknown>).sessionID as string;
   }
-  if (
-    isRecord(o.part) &&
-    typeof (o.part as Record<string, unknown>).sessionID === 'string' &&
-    !((state as unknown as Record<string, unknown>)._harness as Record<string, unknown> | undefined)?.sessionId
-  ) {
-    (state as unknown as Record<string, unknown>)._harness = {
-      ...(((state as unknown as Record<string, unknown>)._harness as Record<string, unknown>) ?? {}),
-      sessionId: (o.part as Record<string, unknown>).sessionID as string,
-    };
-  }
-  if (typeStr.includes('tool') || o.type === 'tool_use' || o.type === 'tool_result') {
+  // real tool schema: one combined `tool_use` event per call (already resolved by the time it's
+  // emitted — no separate start/pending line observed), correlated by part.callID.
+  if (typeStr === 'tool_use' && isRecord(o.part)) {
+    const part = o.part as Record<string, unknown>;
+    const name = typeof part.tool === 'string' ? part.tool : 'tool';
+    const id = typeof part.callID === 'string' ? part.callID : undefined;
+    const toolState = isRecord(part.state) ? (part.state as Record<string, unknown>) : null;
+    const input = toolState && isRecord(toolState.input) ? (toolState.input as Record<string, unknown>) : {};
+    const isError = Boolean(toolState?.error) || toolState?.status === 'error';
+    activities.push({ kind: 'tool_start', name });
+    activities.push({ kind: 'tool_input', name, input, id });
+    activities.push({ kind: 'tool_result', isError, id });
+  } else if (typeStr.includes('tool') || o.type === 'tool_result') {
+    // fallback for older/other event shapes that don't match the combined tool_use schema above
     const name =
       typeof o.name === 'string'
         ? o.name
         : typeof (o as Record<string, unknown>).tool === 'string'
           ? ((o as Record<string, unknown>).tool as string)
           : 'tool';
-    if (typeStr.includes('start') || o.type === 'tool_use') {
+    if (typeStr.includes('start')) {
       activities.push({ kind: 'tool_start', name });
       if (isRecord(o.input)) activities.push({ kind: 'tool_input', name, input: o.input as Record<string, unknown> });
     } else if (typeStr.includes('result') || typeStr.includes('completed') || o.type === 'tool_result')
@@ -72,54 +96,34 @@ export function parseOpencodeLine(line: string, state: ParseState): ParseOutcome
   }
   if (typeStr.includes('thinking') || o.type === 'reasoning') activities.push({ kind: 'thinking', chars: 10 });
   const directText = extractOpencodeText(o);
-  if (directText && !typeStr.includes('tool') && !typeStr.includes('thinking')) streamedText = directText;
+  if (directText && typeStr !== 'tool_use' && !typeStr.includes('tool') && !typeStr.includes('thinking'))
+    streamedText = directText;
   if (o.type === 'text' && isRecord(o.part) && typeof o.part.text === 'string') streamedText = o.part.text;
   if (o.type === 'result' || o.type === 'completed' || o.type === 'done' || o.type === 'step_finish') {
     const part = isRecord(o.part) ? (o.part as Record<string, unknown>) : null;
-    const usageRaw = isRecord(o.usage)
-      ? o.usage
-      : part && isRecord((part as Record<string, unknown>).tokens)
-        ? ((part as Record<string, unknown>).tokens as Record<string, unknown>)
-        : isRecord(o.part) && isRecord((o.part as Record<string, unknown>).tokens)
-          ? ((o.part as Record<string, unknown>).tokens as Record<string, unknown>)
-          : null;
-    const usage = isRecord(usageRaw) ? usageRaw : null;
-    const tokensInput = isRecord(usage)
-      ? typeof usage.input === 'number'
-        ? usage.input
-        : typeof usage.input_tokens === 'number'
-          ? usage.input_tokens
-          : 0
-      : 0;
-    const tokensOutput = isRecord(usage)
-      ? typeof usage.output === 'number'
-        ? usage.output
-        : typeof usage.output_tokens === 'number'
-          ? usage.output_tokens
-          : 0
-      : 0;
-    const cost =
-      typeof o.total_cost_usd === 'number'
-        ? o.total_cost_usd
-        : part && typeof (part as Record<string, unknown>).cost === 'number'
-          ? ((part as Record<string, unknown>).cost as number)
-          : typeof o.cost === 'number'
-            ? (o.cost as number)
-            : 0;
-    const latched = ((state as unknown as Record<string, unknown>)._harness as Record<string, unknown> | undefined)
-      ?.sessionId as string | undefined;
+    // step_finish reports the cost/tokens of that one step, not a running total — accumulate
+    // across every step_finish seen so far so the final result reports a genuine session total.
+    if (typeStr === 'step_finish' && part) {
+      const tokens = isRecord(part.tokens) ? (part.tokens as Record<string, unknown>) : null;
+      const cache = tokens && isRecord(tokens.cache) ? (tokens.cache as Record<string, unknown>) : null;
+      hs.costAccum = (hs.costAccum ?? 0) + (typeof part.cost === 'number' ? part.cost : 0);
+      hs.inputAccum = (hs.inputAccum ?? 0) + (tokens && typeof tokens.input === 'number' ? tokens.input : 0);
+      hs.outputAccum = (hs.outputAccum ?? 0) + (tokens && typeof tokens.output === 'number' ? tokens.output : 0);
+      hs.cacheReadAccum = (hs.cacheReadAccum ?? 0) + (cache && typeof cache.read === 'number' ? cache.read : 0);
+      hs.cacheWriteAccum = (hs.cacheWriteAccum ?? 0) + (cache && typeof cache.write === 'number' ? cache.write : 0);
+      hs.stepCount = (hs.stepCount ?? 0) + 1;
+    }
+    const measured = (hs.stepCount ?? 0) > 0;
     const sessionId =
       typeof o.session_id === 'string'
         ? o.session_id
         : typeof o.sessionID === 'string'
           ? o.sessionID
-          : typeof (o as Record<string, unknown>).sessionID === 'string'
-            ? ((o as Record<string, unknown>).sessionID as string)
-            : part && typeof (part as Record<string, unknown>).sessionID === 'string'
-              ? ((part as Record<string, unknown>).sessionID as string)
-              : typeof o.id === 'string'
-                ? o.id
-                : (latched ?? null);
+          : part && typeof (part as Record<string, unknown>).sessionID === 'string'
+            ? ((part as Record<string, unknown>).sessionID as string)
+            : typeof o.id === 'string'
+              ? o.id
+              : (hs.sessionId ?? null);
     const result: StreamedResult = {
       result:
         typeof o.result === 'string'
@@ -128,8 +132,12 @@ export function parseOpencodeLine(line: string, state: ParseState): ParseOutcome
             ? o.output
             : state.streamedText + (streamedText ?? ''),
       isError: o.is_error === true,
-      numTurns: typeof o.num_turns === 'number' ? o.num_turns : 0,
-      totalCostUsd: cost,
+      numTurns: typeof o.num_turns === 'number' ? o.num_turns : measured ? (hs.stepCount as number) : null,
+      totalCostUsd: measured
+        ? (hs.costAccum as number)
+        : typeof o.total_cost_usd === 'number'
+          ? o.total_cost_usd
+          : null,
       sessionId,
       stopReason:
         typeof o.stop_reason === 'string'
@@ -144,8 +152,13 @@ export function parseOpencodeLine(line: string, state: ParseState): ParseOutcome
       model: typeof o.model === 'string' ? o.model : null,
       contextWindow: null,
       maxOutputTokens: null,
-      usage: usage
-        ? { inputTokens: tokensInput, outputTokens: tokensOutput, cacheCreationInputTokens: 0, cacheReadInputTokens: 0 }
+      usage: measured
+        ? {
+            inputTokens: hs.inputAccum ?? 0,
+            outputTokens: hs.outputAccum ?? 0,
+            cacheCreationInputTokens: hs.cacheWriteAccum ?? 0,
+            cacheReadInputTokens: hs.cacheReadAccum ?? 0,
+          }
         : null,
     };
     return { activities, streamedText, result };
@@ -165,11 +178,11 @@ export const opencodeHarness: Harness = {
     }
   },
   buildArgs(opts: BuildArgsOpts): string[] {
-    const perm = opts.nativePermission ?? PERMISSION_MAP[opts.permission] ?? 'allow-edit';
-    const args = ['run', '--format', 'json', opts.prompt, '--permission', perm];
+    const agent = opts.nativePermission ?? AGENT_MAP[opts.permission] ?? 'build';
+    const args = ['run', opts.prompt, '--format', 'json', '--agent', agent];
+    if (opts.permission === 'danger' && !opts.nativePermission) args.push('--auto');
     if (opts.model) args.push('--model', opts.model);
     if (opts.resumeSessionId) args.push('--session', opts.resumeSessionId);
-    for (const dir of opts.addDirs ?? []) args.push('--add-dir', dir);
     return args;
   },
   parseLine(line: string, state: ParseState): ParseOutcome {
@@ -178,13 +191,12 @@ export const opencodeHarness: Harness = {
   extractResult(state: ParseState): StreamedResult | null {
     if (state.result) return state.result;
     if (state.streamedText.trim().length > 0) {
-      const latched = ((state as unknown as Record<string, unknown>)._harness as Record<string, unknown> | undefined)
-        ?.sessionId as string | undefined;
+      const latched = (state._harness as OpencodeHarnessState | undefined)?.sessionId;
       return {
         result: state.streamedText,
         isError: false,
-        numTurns: 1,
-        totalCostUsd: 0,
+        numTurns: null,
+        totalCostUsd: null,
         sessionId: latched ?? null,
         stopReason: null,
         permissionDenials: [],
@@ -199,5 +211,5 @@ export const opencodeHarness: Harness = {
     }
     return null;
   },
-  permissionMap: { readonly: ['read-only'], edit: ['allow-edit'], danger: ['danger'] },
+  permissionMap: { readonly: ['plan'], edit: ['build'], danger: ['build', '--auto'] },
 };

package/extensions/harnesses/types.ts

@@ -12,8 +12,10 @@ export interface StreamedUsage {
 export interface StreamedResult {
   result: string;
   isError: boolean;
-  numTurns: number;
-  totalCostUsd: number;
+  /** null when the harness's payload doesn't report a turn count — distinct from a measured 0. */
+  numTurns: number | null;
+  /** null when the harness's payload doesn't report cost — distinct from a measured $0. */
+  totalCostUsd: number | null;
   sessionId: string | null;
   stopReason: string | null;
   permissionDenials: unknown[];
@@ -28,8 +30,8 @@ export interface StreamedResult {
 
 export type ActivityEvent =
   | { kind: 'tool_start'; name: string }
-  | { kind: 'tool_input'; name: string; input: Record<string, unknown> }
-  | { kind: 'tool_result'; isError: boolean }
+  | { kind: 'tool_input'; name: string; input: Record<string, unknown>; id?: string }
+  | { kind: 'tool_result'; isError: boolean; id?: string }
   | { kind: 'thinking'; chars: number };
 
 export interface ParseState {