@ai-sdk/harness 0.0.0 → 1.0.0-canary.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (68) hide show
  1. package/CHANGELOG.md +24 -0
  2. package/LICENSE +13 -0
  3. package/README.md +142 -0
  4. package/agent/index.ts +17 -0
  5. package/bridge/index.ts +10 -0
  6. package/dist/agent/index.d.ts +1480 -0
  7. package/dist/agent/index.js +2554 -0
  8. package/dist/agent/index.js.map +1 -0
  9. package/dist/bridge/index.d.ts +111 -0
  10. package/dist/bridge/index.js +414 -0
  11. package/dist/bridge/index.js.map +1 -0
  12. package/dist/index.d.ts +1510 -0
  13. package/dist/index.js +15834 -0
  14. package/dist/index.js.map +1 -0
  15. package/dist/observability/index.d.ts +97 -0
  16. package/dist/observability/index.js +225 -0
  17. package/dist/observability/index.js.map +1 -0
  18. package/dist/utils/index.d.ts +196 -0
  19. package/dist/utils/index.js +327 -0
  20. package/dist/utils/index.js.map +1 -0
  21. package/package.json +104 -1
  22. package/src/agent/harness-agent-session.ts +352 -0
  23. package/src/agent/harness-agent-settings.ts +131 -0
  24. package/src/agent/harness-agent-tool-approval-continuation.ts +94 -0
  25. package/src/agent/harness-agent.ts +750 -0
  26. package/src/agent/harness-diagnostics.ts +88 -0
  27. package/src/agent/internal/bootstrap-recipe.ts +124 -0
  28. package/src/agent/internal/bridge-port-registry.ts +52 -0
  29. package/src/agent/internal/harness-stream-text-result.ts +720 -0
  30. package/src/agent/internal/permission-mode.ts +50 -0
  31. package/src/agent/internal/resolve-observability.ts +128 -0
  32. package/src/agent/internal/resume-state-validation.ts +51 -0
  33. package/src/agent/internal/run-prompt.ts +811 -0
  34. package/src/agent/internal/strip-work-dir.ts +68 -0
  35. package/src/agent/internal/to-harness-stream.ts +75 -0
  36. package/src/agent/internal/translate-stream-part.ts +221 -0
  37. package/src/agent/internal/turn-telemetry.ts +359 -0
  38. package/src/agent/prewarm.ts +46 -0
  39. package/src/bridge/index.ts +700 -0
  40. package/src/errors/harness-capability-unsupported-error.ts +41 -0
  41. package/src/errors/harness-error.ts +22 -0
  42. package/src/index.ts +3 -0
  43. package/src/observability/file-reporter.ts +209 -0
  44. package/src/observability/index.ts +13 -0
  45. package/src/observability/trace-tree-reporter.ts +122 -0
  46. package/src/utils/classify-disk-log.ts +43 -0
  47. package/src/utils/index.ts +7 -0
  48. package/src/utils/sandbox-channel.ts +453 -0
  49. package/src/v1/harness-v1-bootstrap.ts +46 -0
  50. package/src/v1/harness-v1-bridge-protocol.ts +310 -0
  51. package/src/v1/harness-v1-builtin-tool.ts +138 -0
  52. package/src/v1/harness-v1-call-warning.ts +22 -0
  53. package/src/v1/harness-v1-diagnostic.ts +66 -0
  54. package/src/v1/harness-v1-metadata.ts +13 -0
  55. package/src/v1/harness-v1-network-sandbox-session.ts +123 -0
  56. package/src/v1/harness-v1-observability.ts +20 -0
  57. package/src/v1/harness-v1-permission-mode.ts +11 -0
  58. package/src/v1/harness-v1-prompt-control.ts +41 -0
  59. package/src/v1/harness-v1-prompt.ts +11 -0
  60. package/src/v1/harness-v1-resume-state.ts +46 -0
  61. package/src/v1/harness-v1-sandbox-provider.ts +76 -0
  62. package/src/v1/harness-v1-session.ts +268 -0
  63. package/src/v1/harness-v1-skill.ts +22 -0
  64. package/src/v1/harness-v1-stream-part.ts +363 -0
  65. package/src/v1/harness-v1-tool-spec.ts +31 -0
  66. package/src/v1/harness-v1.ts +83 -0
  67. package/src/v1/index.ts +93 -0
  68. package/utils/index.ts +1 -0
@@ -0,0 +1,97 @@
1
+ import { Telemetry } from 'ai';
2
+
3
+ /** Severity of a diagnostic. */
4
+ type HarnessDebugLevel = 'error' | 'warn' | 'info' | 'debug' | 'trace';
5
+ /**
6
+ * A forwarded bridge diagnostic, normalized for host consumers.
7
+ *
8
+ * The bridge emits two raw frame kinds — captured console lines (`sandbox-log`)
9
+ * and structured events (`debug-event`). The framework normalizes both into
10
+ * this single shape before handing them to a consumer's `onLog` callback, the
11
+ * `HARNESS_DEBUG` stderr default, and observability reporters. Diagnostics are
12
+ * kept first-class and per-line — they are never folded into telemetry spans.
13
+ */
14
+ type HarnessDiagnostic = {
15
+ /**
16
+ * Severity. Structured events carry their own level; captured console lines
17
+ * map `stderr` → `'warn'` and `stdout` → `'info'`.
18
+ */
19
+ readonly level: HarnessDebugLevel;
20
+ /** Human-readable line (console capture) or message (structured event). */
21
+ readonly message: string;
22
+ /**
23
+ * Dotted subsystem. For captured console output this is
24
+ * `sandbox.log.<source>`; for structured events it is the adapter-supplied
25
+ * subsystem (e.g. `bridge.turn`).
26
+ */
27
+ readonly subsystem: string;
28
+ /** `'log'` = captured console line; `'event'` = structured `bridgeLog`. */
29
+ readonly kind: 'log' | 'event';
30
+ /** Originating sandbox source label (console capture). */
31
+ readonly source?: string;
32
+ /** Which standard stream the line came from (console capture). */
33
+ readonly stream?: 'stdout' | 'stderr';
34
+ /** Structured attributes (structured events only). */
35
+ readonly attrs?: Record<string, unknown>;
36
+ /** Error payload (structured events only). */
37
+ readonly error?: {
38
+ name?: string;
39
+ message: string;
40
+ stack?: string;
41
+ };
42
+ /** The harness session this diagnostic originated from. */
43
+ readonly sessionId?: string;
44
+ /** Host receipt time (epoch ms). */
45
+ readonly timestamp: number;
46
+ };
47
+ /**
48
+ * A telemetry integration that also wants the per-line diagnostics stream. The
49
+ * framework calls `ingestDiagnostic` for every forwarded bridge diagnostic in
50
+ * addition to driving the standard `Telemetry` span lifecycle, so a single
51
+ * reporter object (e.g. `createFileReporter`) registered in
52
+ * `telemetry.integrations` receives both spans and logs.
53
+ */
54
+ interface HarnessDiagnosticConsumer {
55
+ ingestDiagnostic?(diagnostic: HarnessDiagnostic): void;
56
+ }
57
+
58
+ /**
59
+ * A harness observability reporter that writes a unified, non-lossy
60
+ * `events.jsonl` containing **both** the telemetry span lifecycle (turn / step
61
+ * / tool) **and** the forwarded bridge diagnostics (console lines + structured
62
+ * events). It is a single object registered in `telemetry.integrations`: the
63
+ * framework drives its `Telemetry` methods for spans and calls
64
+ * `ingestDiagnostic` for diagnostics.
65
+ *
66
+ * No external collector or OTel setup required — this is the AI-SDK-idiomatic
67
+ * replacement for the original SDK's host-side artifact files.
68
+ */
69
+ interface FileReporterOptions {
70
+ /** Directory for `events.jsonl` (created if absent). */
71
+ dir: string;
72
+ /**
73
+ * Buffer a turn's records in memory and write them only if the turn produced
74
+ * an error (an `error`-level diagnostic, a failed tool, or an error finish).
75
+ * Default false (write everything).
76
+ */
77
+ failOnly?: boolean;
78
+ /** File name within `dir`. Default `events.jsonl`. */
79
+ fileName?: string;
80
+ }
81
+ type FileReporter = Telemetry & HarnessDiagnosticConsumer;
82
+ declare function createFileReporter(options: FileReporterOptions): FileReporter;
83
+
84
+ /**
85
+ * A harness observability reporter that renders an ASCII trace tree of a
86
+ * turn's span lifecycle (turn → steps → tools) to a stream at turn end. It is a
87
+ * `Telemetry` integration — register it in `telemetry.integrations`. Useful for
88
+ * zero-setup local debugging when no OTel collector is wired up; a real OTel
89
+ * backend (via `@ai-sdk/otel`) is a strict superset.
90
+ */
91
+ interface TraceTreeReporterOptions {
92
+ /** Where to write the rendered tree. Default `process.stderr.write`. */
93
+ write?: (chunk: string) => void;
94
+ }
95
+ declare function createTraceTreeReporter(options?: TraceTreeReporterOptions): Telemetry;
96
+
97
+ export { type FileReporter, type FileReporterOptions, type HarnessDiagnostic, type HarnessDiagnosticConsumer, type TraceTreeReporterOptions, createFileReporter, createTraceTreeReporter };
@@ -0,0 +1,225 @@
1
+ // src/observability/file-reporter.ts
2
+ import { appendFileSync, mkdirSync } from "fs";
3
+ function createFileReporter(options) {
4
+ var _a, _b;
5
+ const fileName = (_a = options.fileName) != null ? _a : "events.jsonl";
6
+ const path = `${options.dir}/${fileName}`;
7
+ const failOnly = (_b = options.failOnly) != null ? _b : false;
8
+ const turns = /* @__PURE__ */ new Map();
9
+ let lastOpenCallId;
10
+ let dirReady = false;
11
+ const bucketFor = (callId) => {
12
+ let bucket = turns.get(callId);
13
+ if (!bucket) {
14
+ bucket = { lines: [], errored: false };
15
+ turns.set(callId, bucket);
16
+ }
17
+ return bucket;
18
+ };
19
+ const record = (callId, rec) => {
20
+ const id = callId != null ? callId : lastOpenCallId;
21
+ if (id == null) {
22
+ flushLines([rec]);
23
+ return;
24
+ }
25
+ bucketFor(id).lines.push(rec);
26
+ };
27
+ const flushLines = (lines) => {
28
+ if (lines.length === 0) return;
29
+ try {
30
+ if (!dirReady) {
31
+ mkdirSync(options.dir, { recursive: true });
32
+ dirReady = true;
33
+ }
34
+ appendFileSync(path, lines.map((l) => JSON.stringify(l)).join("\n") + "\n");
35
+ } catch (e) {
36
+ }
37
+ };
38
+ const finishTurn = (callId) => {
39
+ const bucket = turns.get(callId);
40
+ if (!bucket) return;
41
+ turns.delete(callId);
42
+ if (lastOpenCallId === callId) lastOpenCallId = void 0;
43
+ if (failOnly && !bucket.errored) return;
44
+ flushLines(bucket.lines);
45
+ };
46
+ return {
47
+ onStart(event) {
48
+ const e = event;
49
+ lastOpenCallId = e.callId;
50
+ record(e.callId, {
51
+ ts: Date.now(),
52
+ kind: "turn-start",
53
+ callId: e.callId,
54
+ operationId: e.operationId,
55
+ provider: e.provider,
56
+ modelId: e.modelId,
57
+ // Input prompt, unless the consumer opted out via `recordInputs: false`.
58
+ ...e.recordInputs === false ? {} : { input: { messages: e.messages, instructions: e.instructions } }
59
+ });
60
+ },
61
+ onStepStart(event) {
62
+ const e = event;
63
+ record(e.callId, {
64
+ ts: Date.now(),
65
+ kind: "step-start",
66
+ callId: e.callId,
67
+ step: e.stepNumber
68
+ });
69
+ },
70
+ onToolExecutionStart(event) {
71
+ const e = event;
72
+ record(e.callId, {
73
+ ts: Date.now(),
74
+ kind: "tool-start",
75
+ callId: e.callId,
76
+ toolName: e.toolCall.toolName,
77
+ toolCallId: e.toolCall.toolCallId,
78
+ input: e.toolCall.input
79
+ });
80
+ },
81
+ onToolExecutionEnd(event) {
82
+ const e = event;
83
+ const isError = e.toolOutput.type === "error";
84
+ if (isError) bucketFor(e.callId).errored = true;
85
+ record(e.callId, {
86
+ ts: Date.now(),
87
+ kind: "tool-end",
88
+ callId: e.callId,
89
+ toolCallId: e.toolCall.toolCallId,
90
+ isError,
91
+ output: isError ? e.toolOutput.error : e.toolOutput.output
92
+ });
93
+ },
94
+ onStepFinish(event) {
95
+ const e = event;
96
+ record(e.callId, {
97
+ ts: Date.now(),
98
+ kind: "step-finish",
99
+ callId: e.callId,
100
+ usage: e.usage,
101
+ // The model's output content, unless `recordOutputs: false`.
102
+ ...e.recordOutputs === false || e.content == null ? {} : { output: e.content }
103
+ });
104
+ },
105
+ onEnd(event) {
106
+ var _a2;
107
+ const e = event;
108
+ record(e.callId, {
109
+ ts: Date.now(),
110
+ kind: "turn-finish",
111
+ callId: e.callId,
112
+ finishReason: e.finishReason,
113
+ usage: (_a2 = e.totalUsage) != null ? _a2 : e.usage
114
+ });
115
+ finishTurn(e.callId);
116
+ },
117
+ onError(error) {
118
+ if (lastOpenCallId != null) bucketFor(lastOpenCallId).errored = true;
119
+ record(lastOpenCallId, {
120
+ ts: Date.now(),
121
+ kind: "error",
122
+ error: error instanceof Error ? { name: error.name, message: error.message } : error
123
+ });
124
+ },
125
+ ingestDiagnostic(diagnostic) {
126
+ if (diagnostic.level === "error" && lastOpenCallId != null) {
127
+ bucketFor(lastOpenCallId).errored = true;
128
+ }
129
+ record(lastOpenCallId, {
130
+ ts: diagnostic.timestamp,
131
+ kind: "diagnostic",
132
+ diagnostic
133
+ });
134
+ }
135
+ };
136
+ }
137
+
138
+ // src/observability/trace-tree-reporter.ts
139
+ function createTraceTreeReporter(options = {}) {
140
+ var _a;
141
+ const write = (_a = options.write) != null ? _a : ((chunk) => void process.stderr.write(chunk));
142
+ const turns = /* @__PURE__ */ new Map();
143
+ const render = (node, depth) => {
144
+ const indent = " ".repeat(depth);
145
+ const dur = node.endMs != null ? `${Math.max(0, Math.round(node.endMs - node.startMs))}ms` : "(open)";
146
+ let out = `${indent}- ${node.label} ${dur}
147
+ `;
148
+ for (const child of node.children) out += render(child, depth + 1);
149
+ return out;
150
+ };
151
+ return {
152
+ onStart(event) {
153
+ var _a2;
154
+ const e = event;
155
+ turns.set(e.callId, {
156
+ root: {
157
+ label: `${(_a2 = e.operationId) != null ? _a2 : "turn"}${e.modelId ? ` ${e.modelId}` : ""}`,
158
+ startMs: Date.now(),
159
+ children: []
160
+ },
161
+ tools: /* @__PURE__ */ new Map()
162
+ });
163
+ },
164
+ onStepStart(event) {
165
+ var _a2;
166
+ const e = event;
167
+ const turn = turns.get(e.callId);
168
+ if (!turn) return;
169
+ const step = {
170
+ label: `step ${((_a2 = e.stepNumber) != null ? _a2 : turn.root.children.length) + 1}`,
171
+ startMs: Date.now(),
172
+ children: []
173
+ };
174
+ turn.step = step;
175
+ turn.root.children.push(step);
176
+ },
177
+ onToolExecutionStart(event) {
178
+ var _a2;
179
+ const e = event;
180
+ const turn = turns.get(e.callId);
181
+ const parent = (_a2 = turn == null ? void 0 : turn.step) != null ? _a2 : turn == null ? void 0 : turn.root;
182
+ if (!turn || !parent) return;
183
+ const node = {
184
+ label: `tool ${e.toolCall.toolName}`,
185
+ startMs: Date.now(),
186
+ children: []
187
+ };
188
+ turn.tools.set(e.toolCall.toolCallId, node);
189
+ parent.children.push(node);
190
+ },
191
+ onToolExecutionEnd(event) {
192
+ var _a2;
193
+ const e = event;
194
+ const node = (_a2 = turns.get(e.callId)) == null ? void 0 : _a2.tools.get(e.toolCall.toolCallId);
195
+ if (!node) return;
196
+ node.endMs = Date.now();
197
+ if (e.toolOutput.type === "error") node.label += " [error]";
198
+ },
199
+ onStepFinish(event) {
200
+ const e = event;
201
+ const turn = turns.get(e.callId);
202
+ if (turn == null ? void 0 : turn.step) {
203
+ turn.step.endMs = Date.now();
204
+ turn.step = void 0;
205
+ }
206
+ },
207
+ onEnd(event) {
208
+ const e = event;
209
+ const turn = turns.get(e.callId);
210
+ if (!turn) return;
211
+ turn.root.endMs = Date.now();
212
+ turns.delete(e.callId);
213
+ try {
214
+ write(`
215
+ ${render(turn.root, 0)}`);
216
+ } catch (e2) {
217
+ }
218
+ }
219
+ };
220
+ }
221
+ export {
222
+ createFileReporter,
223
+ createTraceTreeReporter
224
+ };
225
+ //# sourceMappingURL=index.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"sources":["../../src/observability/file-reporter.ts","../../src/observability/trace-tree-reporter.ts"],"sourcesContent":["import { appendFileSync, mkdirSync } from 'node:fs';\nimport type { Telemetry } from 'ai';\nimport type {\n HarnessDiagnostic,\n HarnessDiagnosticConsumer,\n} from '../agent/harness-diagnostics';\n\n/**\n * A harness observability reporter that writes a unified, non-lossy\n * `events.jsonl` containing **both** the telemetry span lifecycle (turn / step\n * / tool) **and** the forwarded bridge diagnostics (console lines + structured\n * events). It is a single object registered in `telemetry.integrations`: the\n * framework drives its `Telemetry` methods for spans and calls\n * `ingestDiagnostic` for diagnostics.\n *\n * No external collector or OTel setup required — this is the AI-SDK-idiomatic\n * replacement for the original SDK's host-side artifact files.\n */\nexport interface FileReporterOptions {\n /** Directory for `events.jsonl` (created if absent). */\n dir: string;\n /**\n * Buffer a turn's records in memory and write them only if the turn produced\n * an error (an `error`-level diagnostic, a failed tool, or an error finish).\n * Default false (write everything).\n */\n failOnly?: boolean;\n /** File name within `dir`. Default `events.jsonl`. */\n fileName?: string;\n}\n\ntype Record_ = { ts: number } & Record<string, unknown>;\n\nexport type FileReporter = Telemetry & HarnessDiagnosticConsumer;\n\nexport function createFileReporter(options: FileReporterOptions): FileReporter {\n const fileName = options.fileName ?? 'events.jsonl';\n const path = `${options.dir}/${fileName}`;\n const failOnly = options.failOnly ?? false;\n\n // Per-turn buffers, keyed by the telemetry callId. Diagnostics (which carry a\n // sessionId, not a callId) attach to the most recently started, open turn.\n const turns = new Map<string, { lines: Record_[]; errored: boolean }>();\n let lastOpenCallId: string | undefined;\n let dirReady = false;\n\n const bucketFor = (callId: string) => {\n let bucket = turns.get(callId);\n if (!bucket) {\n bucket = { lines: [], errored: false };\n turns.set(callId, bucket);\n }\n return bucket;\n };\n\n const record = (callId: string | undefined, rec: Record_): void => {\n const id = callId ?? lastOpenCallId;\n if (id == null) {\n // No active turn — write standalone (best-effort).\n flushLines([rec]);\n return;\n }\n bucketFor(id).lines.push(rec);\n };\n\n const flushLines = (lines: Record_[]): void => {\n if (lines.length === 0) return;\n try {\n if (!dirReady) {\n mkdirSync(options.dir, { recursive: true });\n dirReady = true;\n }\n appendFileSync(path, lines.map(l => JSON.stringify(l)).join('\\n') + '\\n');\n } catch {\n // Best-effort: never let observability break a turn.\n }\n };\n\n const finishTurn = (callId: string): void => {\n const bucket = turns.get(callId);\n if (!bucket) return;\n turns.delete(callId);\n if (lastOpenCallId === callId) lastOpenCallId = undefined;\n if (failOnly && !bucket.errored) return;\n flushLines(bucket.lines);\n };\n\n return {\n onStart(event) {\n const e = event as {\n callId: string;\n operationId?: string;\n modelId?: string;\n provider?: string;\n messages?: unknown;\n instructions?: unknown;\n recordInputs?: boolean;\n };\n lastOpenCallId = e.callId;\n record(e.callId, {\n ts: Date.now(),\n kind: 'turn-start',\n callId: e.callId,\n operationId: e.operationId,\n provider: e.provider,\n modelId: e.modelId,\n // Input prompt, unless the consumer opted out via `recordInputs: false`.\n ...(e.recordInputs === false\n ? {}\n : { input: { messages: e.messages, instructions: e.instructions } }),\n });\n },\n onStepStart(event) {\n const e = event as { callId: string; stepNumber?: number };\n record(e.callId, {\n ts: Date.now(),\n kind: 'step-start',\n callId: e.callId,\n step: e.stepNumber,\n });\n },\n onToolExecutionStart(event) {\n const e = event as {\n callId: string;\n toolCall: { toolName: string; toolCallId: string; input: unknown };\n };\n record(e.callId, {\n ts: Date.now(),\n kind: 'tool-start',\n callId: e.callId,\n toolName: e.toolCall.toolName,\n toolCallId: e.toolCall.toolCallId,\n input: e.toolCall.input,\n });\n },\n onToolExecutionEnd(event) {\n const e = event as {\n callId: string;\n toolCall: { toolCallId: string };\n toolOutput: { type: string; output?: unknown; error?: unknown };\n };\n const isError = e.toolOutput.type === 'error';\n if (isError) bucketFor(e.callId).errored = true;\n record(e.callId, {\n ts: Date.now(),\n kind: 'tool-end',\n callId: e.callId,\n toolCallId: e.toolCall.toolCallId,\n isError,\n output: isError ? e.toolOutput.error : e.toolOutput.output,\n });\n },\n onStepFinish(event) {\n const e = event as {\n callId: string;\n usage?: unknown;\n content?: unknown[];\n recordOutputs?: boolean;\n };\n record(e.callId, {\n ts: Date.now(),\n kind: 'step-finish',\n callId: e.callId,\n usage: e.usage,\n // The model's output content, unless `recordOutputs: false`.\n ...(e.recordOutputs === false || e.content == null\n ? {}\n : { output: e.content }),\n });\n },\n onEnd(event) {\n const e = event as {\n callId: string;\n finishReason?: unknown;\n usage?: unknown;\n totalUsage?: unknown;\n };\n record(e.callId, {\n ts: Date.now(),\n kind: 'turn-finish',\n callId: e.callId,\n finishReason: e.finishReason,\n usage: e.totalUsage ?? e.usage,\n });\n finishTurn(e.callId);\n },\n onError(error) {\n if (lastOpenCallId != null) bucketFor(lastOpenCallId).errored = true;\n record(lastOpenCallId, {\n ts: Date.now(),\n kind: 'error',\n error:\n error instanceof Error\n ? { name: error.name, message: error.message }\n : error,\n });\n },\n ingestDiagnostic(diagnostic: HarnessDiagnostic) {\n if (diagnostic.level === 'error' && lastOpenCallId != null) {\n bucketFor(lastOpenCallId).errored = true;\n }\n record(lastOpenCallId, {\n ts: diagnostic.timestamp,\n kind: 'diagnostic',\n diagnostic,\n });\n },\n };\n}\n","import type { Telemetry } from 'ai';\n\n/**\n * A harness observability reporter that renders an ASCII trace tree of a\n * turn's span lifecycle (turn → steps → tools) to a stream at turn end. It is a\n * `Telemetry` integration — register it in `telemetry.integrations`. Useful for\n * zero-setup local debugging when no OTel collector is wired up; a real OTel\n * backend (via `@ai-sdk/otel`) is a strict superset.\n */\nexport interface TraceTreeReporterOptions {\n /** Where to write the rendered tree. Default `process.stderr.write`. */\n write?: (chunk: string) => void;\n}\n\ntype Node = {\n label: string;\n startMs: number;\n endMs?: number;\n children: Node[];\n};\n\nexport function createTraceTreeReporter(\n options: TraceTreeReporterOptions = {},\n): Telemetry {\n const write =\n options.write ?? ((chunk: string) => void process.stderr.write(chunk));\n\n type TurnState = {\n root: Node;\n step?: Node;\n tools: Map<string, Node>;\n };\n const turns = new Map<string, TurnState>();\n\n const render = (node: Node, depth: number): string => {\n const indent = ' '.repeat(depth);\n const dur =\n node.endMs != null\n ? `${Math.max(0, Math.round(node.endMs - node.startMs))}ms`\n : '(open)';\n let out = `${indent}- ${node.label} ${dur}\\n`;\n for (const child of node.children) out += render(child, depth + 1);\n return out;\n };\n\n return {\n onStart(event) {\n const e = event as {\n callId: string;\n operationId?: string;\n modelId?: string;\n };\n turns.set(e.callId, {\n root: {\n label: `${e.operationId ?? 'turn'}${e.modelId ? ` ${e.modelId}` : ''}`,\n startMs: Date.now(),\n children: [],\n },\n tools: new Map(),\n });\n },\n onStepStart(event) {\n const e = event as { callId: string; stepNumber?: number };\n const turn = turns.get(e.callId);\n if (!turn) return;\n const step: Node = {\n label: `step ${(e.stepNumber ?? turn.root.children.length) + 1}`,\n startMs: Date.now(),\n children: [],\n };\n turn.step = step;\n turn.root.children.push(step);\n },\n onToolExecutionStart(event) {\n const e = event as {\n callId: string;\n toolCall: { toolName: string; toolCallId: string };\n };\n const turn = turns.get(e.callId);\n const parent = turn?.step ?? turn?.root;\n if (!turn || !parent) return;\n const node: Node = {\n label: `tool ${e.toolCall.toolName}`,\n startMs: Date.now(),\n children: [],\n };\n turn.tools.set(e.toolCall.toolCallId, node);\n parent.children.push(node);\n },\n onToolExecutionEnd(event) {\n const e = event as {\n callId: string;\n toolCall: { toolCallId: string };\n toolOutput: { type: string };\n };\n const node = turns.get(e.callId)?.tools.get(e.toolCall.toolCallId);\n if (!node) return;\n node.endMs = Date.now();\n if (e.toolOutput.type === 'error') node.label += ' [error]';\n },\n onStepFinish(event) {\n const e = event as { callId: string };\n const turn = turns.get(e.callId);\n if (turn?.step) {\n turn.step.endMs = Date.now();\n turn.step = undefined;\n }\n },\n onEnd(event) {\n const e = event as { callId: string };\n const turn = turns.get(e.callId);\n if (!turn) return;\n turn.root.endMs = Date.now();\n turns.delete(e.callId);\n try {\n write(`\\n${render(turn.root, 0)}`);\n } catch {\n // Never let rendering break the turn.\n }\n },\n };\n}\n"],"mappings":";AAAA,SAAS,gBAAgB,iBAAiB;AAmCnC,SAAS,mBAAmB,SAA4C;AAnC/E;AAoCE,QAAM,YAAW,aAAQ,aAAR,YAAoB;AACrC,QAAM,OAAO,GAAG,QAAQ,GAAG,IAAI,QAAQ;AACvC,QAAM,YAAW,aAAQ,aAAR,YAAoB;AAIrC,QAAM,QAAQ,oBAAI,IAAoD;AACtE,MAAI;AACJ,MAAI,WAAW;AAEf,QAAM,YAAY,CAAC,WAAmB;AACpC,QAAI,SAAS,MAAM,IAAI,MAAM;AAC7B,QAAI,CAAC,QAAQ;AACX,eAAS,EAAE,OAAO,CAAC,GAAG,SAAS,MAAM;AACrC,YAAM,IAAI,QAAQ,MAAM;AAAA,IAC1B;AACA,WAAO;AAAA,EACT;AAEA,QAAM,SAAS,CAAC,QAA4B,QAAuB;AACjE,UAAM,KAAK,0BAAU;AACrB,QAAI,MAAM,MAAM;AAEd,iBAAW,CAAC,GAAG,CAAC;AAChB;AAAA,IACF;AACA,cAAU,EAAE,EAAE,MAAM,KAAK,GAAG;AAAA,EAC9B;AAEA,QAAM,aAAa,CAAC,UAA2B;AAC7C,QAAI,MAAM,WAAW,EAAG;AACxB,QAAI;AACF,UAAI,CAAC,UAAU;AACb,kBAAU,QAAQ,KAAK,EAAE,WAAW,KAAK,CAAC;AAC1C,mBAAW;AAAA,MACb;AACA,qBAAe,MAAM,MAAM,IAAI,OAAK,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,IAAI,IAAI;AAAA,IAC1E,SAAQ;AAAA,IAER;AAAA,EACF;AAEA,QAAM,aAAa,CAAC,WAAyB;AAC3C,UAAM,SAAS,MAAM,IAAI,MAAM;AAC/B,QAAI,CAAC,OAAQ;AACb,UAAM,OAAO,MAAM;AACnB,QAAI,mBAAmB,OAAQ,kBAAiB;AAChD,QAAI,YAAY,CAAC,OAAO,QAAS;AACjC,eAAW,OAAO,KAAK;AAAA,EACzB;AAEA,SAAO;AAAA,IACL,QAAQ,OAAO;AACb,YAAM,IAAI;AASV,uBAAiB,EAAE;AACnB,aAAO,EAAE,QAAQ;AAAA,QACf,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,QAAQ,EAAE;AAAA,QACV,aAAa,EAAE;AAAA,QACf,UAAU,EAAE;AAAA,QACZ,SAAS,EAAE;AAAA;AAAA,QAEX,GAAI,EAAE,iBAAiB,QACnB,CAAC,IACD,EAAE,OAAO,EAAE,UAAU,EAAE,UAAU,cAAc,EAAE,aAAa,EAAE;AAAA,MACtE,CAAC;AAAA,IACH;AAAA,IACA,YAAY,OAAO;AACjB,YAAM,IAAI;AACV,aAAO,EAAE,QAAQ;AAAA,QACf,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,QAAQ,EAAE;AAAA,QACV,MAAM,EAAE;AAAA,MACV,CAAC;AAAA,IACH;AAAA,IACA,qBAAqB,OAAO;AAC1B,YAAM,IAAI;AAIV,aAAO,EAAE,QAAQ;AAAA,QACf,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,QAAQ,EAAE;AAAA,QACV,UAAU,EAAE,SAAS;AAAA,QACrB,YAAY,EAAE,SAAS;AAAA,QACvB,OAAO,EAAE,SAAS;AAAA,MACpB,CAAC;AAAA,IACH;AAAA,IACA,mBAAmB,OAAO;AACxB,YAAM,IAAI;AAKV,YAAM,UAAU,EAAE,WAAW,SAAS;AACtC,UAAI,QAAS,WAAU,EAAE,MAAM,EAAE,UAAU;AAC3C,aAAO,EAAE,QAAQ;AAAA,QACf,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,QAAQ,EAAE;AAAA,QACV,YAAY,EAAE,SAAS;AAAA,QACvB;AAAA,QACA,QAAQ,UAAU,EAAE,WAAW,QAAQ,EAAE,WAAW;AAAA,MACtD,CAAC;AAAA,IACH;AAAA,IACA,aAAa,OAAO;AAClB,YAAM,IAAI;AAMV,aAAO,EAAE,QAAQ;AAAA,QACf,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,QAAQ,EAAE;AAAA,QACV,OAAO,EAAE;AAAA;AAAA,QAET,GAAI,EAAE,kBAAkB,SAAS,EAAE,WAAW,OAC1C,CAAC,IACD,EAAE,QAAQ,EAAE,QAAQ;AAAA,MAC1B,CAAC;AAAA,IACH;AAAA,IACA,MAAM,OAAO;AA1KjB,UAAAA;AA2KM,YAAM,IAAI;AAMV,aAAO,EAAE,QAAQ;AAAA,QACf,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,QAAQ,EAAE;AAAA,QACV,cAAc,EAAE;AAAA,QAChB,QAAOA,MAAA,EAAE,eAAF,OAAAA,MAAgB,EAAE;AAAA,MAC3B,CAAC;AACD,iBAAW,EAAE,MAAM;AAAA,IACrB;AAAA,IACA,QAAQ,OAAO;AACb,UAAI,kBAAkB,KAAM,WAAU,cAAc,EAAE,UAAU;AAChE,aAAO,gBAAgB;AAAA,QACrB,IAAI,KAAK,IAAI;AAAA,QACb,MAAM;AAAA,QACN,OACE,iBAAiB,QACb,EAAE,MAAM,MAAM,MAAM,SAAS,MAAM,QAAQ,IAC3C;AAAA,MACR,CAAC;AAAA,IACH;AAAA,IACA,iBAAiB,YAA+B;AAC9C,UAAI,WAAW,UAAU,WAAW,kBAAkB,MAAM;AAC1D,kBAAU,cAAc,EAAE,UAAU;AAAA,MACtC;AACA,aAAO,gBAAgB;AAAA,QACrB,IAAI,WAAW;AAAA,QACf,MAAM;AAAA,QACN;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;;;AC3LO,SAAS,wBACd,UAAoC,CAAC,GAC1B;AAvBb;AAwBE,QAAM,SACJ,aAAQ,UAAR,aAAkB,CAAC,UAAkB,KAAK,QAAQ,OAAO,MAAM,KAAK;AAOtE,QAAM,QAAQ,oBAAI,IAAuB;AAEzC,QAAM,SAAS,CAAC,MAAY,UAA0B;AACpD,UAAM,SAAS,KAAK,OAAO,KAAK;AAChC,UAAM,MACJ,KAAK,SAAS,OACV,GAAG,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,QAAQ,KAAK,OAAO,CAAC,CAAC,OACrD;AACN,QAAI,MAAM,GAAG,MAAM,KAAK,KAAK,KAAK,IAAI,GAAG;AAAA;AACzC,eAAW,SAAS,KAAK,SAAU,QAAO,OAAO,OAAO,QAAQ,CAAC;AACjE,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,QAAQ,OAAO;AA9CnB,UAAAC;AA+CM,YAAM,IAAI;AAKV,YAAM,IAAI,EAAE,QAAQ;AAAA,QAClB,MAAM;AAAA,UACJ,OAAO,IAAGA,MAAA,EAAE,gBAAF,OAAAA,MAAiB,MAAM,GAAG,EAAE,UAAU,IAAI,EAAE,OAAO,KAAK,EAAE;AAAA,UACpE,SAAS,KAAK,IAAI;AAAA,UAClB,UAAU,CAAC;AAAA,QACb;AAAA,QACA,OAAO,oBAAI,IAAI;AAAA,MACjB,CAAC;AAAA,IACH;AAAA,IACA,YAAY,OAAO;AA7DvB,UAAAA;AA8DM,YAAM,IAAI;AACV,YAAM,OAAO,MAAM,IAAI,EAAE,MAAM;AAC/B,UAAI,CAAC,KAAM;AACX,YAAM,OAAa;AAAA,QACjB,OAAO,UAASA,MAAA,EAAE,eAAF,OAAAA,MAAgB,KAAK,KAAK,SAAS,UAAU,CAAC;AAAA,QAC9D,SAAS,KAAK,IAAI;AAAA,QAClB,UAAU,CAAC;AAAA,MACb;AACA,WAAK,OAAO;AACZ,WAAK,KAAK,SAAS,KAAK,IAAI;AAAA,IAC9B;AAAA,IACA,qBAAqB,OAAO;AAzEhC,UAAAA;AA0EM,YAAM,IAAI;AAIV,YAAM,OAAO,MAAM,IAAI,EAAE,MAAM;AAC/B,YAAM,UAASA,MAAA,6BAAM,SAAN,OAAAA,MAAc,6BAAM;AACnC,UAAI,CAAC,QAAQ,CAAC,OAAQ;AACtB,YAAM,OAAa;AAAA,QACjB,OAAO,QAAQ,EAAE,SAAS,QAAQ;AAAA,QAClC,SAAS,KAAK,IAAI;AAAA,QAClB,UAAU,CAAC;AAAA,MACb;AACA,WAAK,MAAM,IAAI,EAAE,SAAS,YAAY,IAAI;AAC1C,aAAO,SAAS,KAAK,IAAI;AAAA,IAC3B;AAAA,IACA,mBAAmB,OAAO;AAzF9B,UAAAA;AA0FM,YAAM,IAAI;AAKV,YAAM,QAAOA,MAAA,MAAM,IAAI,EAAE,MAAM,MAAlB,gBAAAA,IAAqB,MAAM,IAAI,EAAE,SAAS;AACvD,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,KAAK,IAAI;AACtB,UAAI,EAAE,WAAW,SAAS,QAAS,MAAK,SAAS;AAAA,IACnD;AAAA,IACA,aAAa,OAAO;AAClB,YAAM,IAAI;AACV,YAAM,OAAO,MAAM,IAAI,EAAE,MAAM;AAC/B,UAAI,6BAAM,MAAM;AACd,aAAK,KAAK,QAAQ,KAAK,IAAI;AAC3B,aAAK,OAAO;AAAA,MACd;AAAA,IACF;AAAA,IACA,MAAM,OAAO;AACX,YAAM,IAAI;AACV,YAAM,OAAO,MAAM,IAAI,EAAE,MAAM;AAC/B,UAAI,CAAC,KAAM;AACX,WAAK,KAAK,QAAQ,KAAK,IAAI;AAC3B,YAAM,OAAO,EAAE,MAAM;AACrB,UAAI;AACF,cAAM;AAAA,EAAK,OAAO,KAAK,MAAM,CAAC,CAAC,EAAE;AAAA,MACnC,SAAQC,IAAA;AAAA,MAER;AAAA,IACF;AAAA,EACF;AACF;","names":["_a","_a","e"]}
@@ -0,0 +1,196 @@
1
+ import { FlexibleSchema } from '@ai-sdk/provider-utils';
2
+ import { WebSocket } from 'ws';
3
+
4
+ /**
5
+ * Diagnostic event surfaced by {@link SandboxChannel} during its connection
6
+ * lifecycle. Silent unless a consumer wires `onDebug`. Reconnects are otherwise
7
+ * invisible — the channel reconnects transparently and the in-flight turn keeps
8
+ * streaming.
9
+ */
10
+ type SandboxChannelDebugEvent = {
11
+ event: 'reconnect-attempt';
12
+ attempt: number;
13
+ lastSeenEventId: number;
14
+ } | {
15
+ event: 'reconnected';
16
+ attempt: number;
17
+ lastSeenEventId: number;
18
+ } | {
19
+ event: 'reconnect-failed';
20
+ attempts: number;
21
+ lastSeenEventId: number;
22
+ cause: unknown;
23
+ };
24
+ interface SandboxChannelReconnectOptions {
25
+ /** Give up reconnecting after this many milliseconds. Default 30_000. */
26
+ readonly maxElapsedMs?: number;
27
+ /** First backoff delay. Default 50. */
28
+ readonly initialDelayMs?: number;
29
+ /** Backoff ceiling. Default 2_000. */
30
+ readonly maxDelayMs?: number;
31
+ }
32
+ interface SandboxChannelOptions<TOut> {
33
+ /**
34
+ * Open a fresh WebSocket to the bridge and resolve once it is ready to carry
35
+ * frames (i.e. after any adapter-specific handshake such as Claude Code's
36
+ * `bridge-hello`). Called once by {@link SandboxChannel.open} and again on
37
+ * every transient reconnect. Must reject if the connection cannot be
38
+ * established.
39
+ */
40
+ connect: () => Promise<WebSocket>;
41
+ /** Schema validating inbound (bridge → host) frames. */
42
+ outboundSchema: FlexibleSchema<TOut>;
43
+ reconnect?: SandboxChannelReconnectOptions;
44
+ onDebug?: (event: SandboxChannelDebugEvent) => void;
45
+ /**
46
+ * Sink for forwarded bridge diagnostics — `sandbox-log` (captured
47
+ * console lines) and `debug-event` (structured) frames. When set, these
48
+ * frame types are routed here instead of the per-type listener dispatch, so
49
+ * they never reach the consumer's stream. Typed to the diagnostic members of
50
+ * `TOut`, so it is a no-op union for channels whose protocol has none.
51
+ */
52
+ onDiagnostic?: (event: Extract<TOut, {
53
+ type: 'sandbox-log' | 'debug-event';
54
+ }>) => void;
55
+ /**
56
+ * Seed the host-side cursor before the first connect. Pass the
57
+ * `lastSeenEventId` persisted from a prior process so the bridge replays only
58
+ * events past it when this channel opens with `{ resume: true }` — the
59
+ * cross-process attach handshake. Defaults to `0` (fresh session).
60
+ */
61
+ initialLastSeenEventId?: number;
62
+ }
63
+ type EventTypeOf<TOut extends {
64
+ type: string;
65
+ }> = TOut['type'];
66
+ type Listener<TOut extends {
67
+ type: string;
68
+ }, T extends EventTypeOf<TOut>> = (event: Extract<TOut, {
69
+ type: T;
70
+ }>) => void;
71
+ /**
72
+ * Host-side typed wrapper around the bridge WebSocket connection.
73
+ *
74
+ * Buffers inbound messages until a listener for their type is registered, so
75
+ * callers that subscribe asynchronously do not miss early frames. Inbound
76
+ * dispatch is serialised through a promise chain so a `close` event that
77
+ * arrives on the same microtask as the final `finish` message does not fire
78
+ * close handlers until the message has been dispatched.
79
+ *
80
+ * Survives transient disconnects. The bridge keeps running and
81
+ * accumulates events in an in-memory log keyed by a monotonic `seq`; on an
82
+ * unexpected socket drop this channel re-invokes `connect`, re-wires the new
83
+ * socket, and asks the bridge to replay everything past `lastSeenEventId`. The
84
+ * in-flight turn never observes the blip — `onClose` fires only after a
85
+ * host-initiated close or once the reconnect budget is exhausted.
86
+ */
87
+ declare class SandboxChannel<TOut extends {
88
+ type: string;
89
+ }, TIn extends {
90
+ type: string;
91
+ } = {
92
+ type: string;
93
+ }> {
94
+ private readonly listeners;
95
+ private readonly buffered;
96
+ private readonly onCloseHandlers;
97
+ private readonly connectThunk;
98
+ private readonly outboundSchema;
99
+ private readonly onDebug;
100
+ private readonly onDiagnostic;
101
+ private readonly maxElapsedMs;
102
+ private readonly initialDelayMs;
103
+ private readonly maxDelayMs;
104
+ private ws;
105
+ private connected;
106
+ /** Host has begun teardown; suppresses reconnect so a bridge-side close finalises. */
107
+ private closing;
108
+ /**
109
+ * Host has gracefully suspended (slice boundary). Inbound frames are ignored
110
+ * from this point so the cursor stops advancing exactly at the last delivered
111
+ * event — the bridge keeps the turn running and the not-yet-delivered tail is
112
+ * replayed to the next process on `resume`.
113
+ */
114
+ private suspended;
115
+ /** Channel is fully torn down; `send` throws and `onClose` has fired. */
116
+ private terminal;
117
+ private _lastSeenEventId;
118
+ private readonly pendingSends;
119
+ private dispatchChain;
120
+ constructor(options: SandboxChannelOptions<TOut>);
121
+ /**
122
+ * Highest bridge event `seq` this channel has observed. Persist it (e.g. via
123
+ * the adapter's resume handle) so a future process can seed
124
+ * {@link SandboxChannelOptions.initialLastSeenEventId} and attach.
125
+ */
126
+ get lastSeenEventId(): number;
127
+ /**
128
+ * Establish the initial connection. A single attempt — startup failures
129
+ * reject so the caller can fail `doStart` cleanly. Reconnect retries apply
130
+ * only to drops after a successful open.
131
+ *
132
+ * Pass `{ resume: true }` to attach to a bridge that is already mid-session:
133
+ * after the socket opens, the channel sends `{ type: 'resume', lastSeenEventId }`
134
+ * so the bridge replays everything past the seeded cursor. This is the
135
+ * cross-process attach handshake — identical to what a transient reconnect
136
+ * does, but triggered by the initial open from a new process.
137
+ */
138
+ open(opts?: {
139
+ resume?: boolean;
140
+ }): Promise<void>;
141
+ on<T extends EventTypeOf<TOut>>(type: T, listener: Listener<TOut, T>): () => void;
142
+ onClose(handler: (code: number, reason: string) => void): void;
143
+ send(message: TIn): void;
144
+ /**
145
+ * Mark that the host is tearing the session down. The next socket close is
146
+ * then treated as terminal rather than triggering a reconnect. Call before
147
+ * sending a `shutdown` / `detach` message whose ack the bridge follows with a
148
+ * socket close.
149
+ */
150
+ beginClose(): void;
151
+ close(): void;
152
+ /**
153
+ * Gracefully suspend at a slice boundary: stop processing inbound frames
154
+ * (so the cursor freezes at the last delivered event), drain any frames
155
+ * already queued for dispatch, then close the socket and finalise with the
156
+ * close reason `'suspended'`. Resolves with the final `lastSeenEventId`.
157
+ *
158
+ * The bridge keeps the in-flight turn running (a host socket close never
159
+ * aborts it) and accumulates events past the cursor for the next process to
160
+ * `resume`. Unlike {@link close}, the consumer's active turn is wound down
161
+ * cleanly — adapters distinguish a suspend from an unexpected drop via the
162
+ * `'suspended'` close reason and resolve `done` successfully.
163
+ */
164
+ suspend(): Promise<number>;
165
+ isClosed(): boolean;
166
+ private wire;
167
+ private reconnectLoop;
168
+ private rawSend;
169
+ private flushPending;
170
+ private enqueue;
171
+ private handleIncoming;
172
+ private dispatch;
173
+ private finalizeClose;
174
+ }
175
+
176
+ /**
177
+ * Recovery rung selected from an on-disk bridge event log when attach is not
178
+ * possible (the bridge process is gone): `'replay'` when the log holds a
179
+ * complete turn the host can resume from its cursor, `'rerun'` otherwise.
180
+ */
181
+ type DiskLogRecoveryMode = 'replay' | 'rerun';
182
+ /**
183
+ * Decide whether a respawned bridge can `replay` a turn from its persisted
184
+ * `event-log.ndjson`, or must `rerun` it from scratch.
185
+ *
186
+ * A turn is replayable only when its log ends in a terminal `finish` event —
187
+ * a log that is missing, empty, or ends mid-turn means the bridge died before
188
+ * completing the turn, so there is no coherent tail to deliver and the runtime
189
+ * must re-run it (continuing its own thread from the sandbox snapshot).
190
+ *
191
+ * @param eventLog Raw contents of `event-log.ndjson` (newline-delimited JSON),
192
+ * or `null`/`undefined`/empty when the file is absent.
193
+ */
194
+ declare function classifyDiskLog(eventLog: string | null | undefined): Promise<DiskLogRecoveryMode>;
195
+
196
+ export { type DiskLogRecoveryMode, SandboxChannel, type SandboxChannelDebugEvent, type SandboxChannelOptions, type SandboxChannelReconnectOptions, classifyDiskLog };