@code-yeongyu/senpi-codemode 2026.7.26 → 2026.7.28-3

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.
package/CHANGELOG.md CHANGED
@@ -12,6 +12,51 @@
12
12
 
13
13
  ### Removed
14
14
 
15
+ ## [2026.7.28-3] - 2026-07-28
16
+
17
+ ### Breaking Changes
18
+
19
+ ### Added
20
+
21
+ - Add nested tool-call widgets that render the real call shape of tools invoked from eval cells, with truthful status, duration, and sanitized previews ([#444](https://github.com/code-yeongyu/senpi/pull/444)).
22
+
23
+ ### Changed
24
+
25
+ ### Fixed
26
+
27
+ ### Removed
28
+
29
+ ## [2026.7.28-2] - 2026-07-28
30
+
31
+ ### Breaking Changes
32
+
33
+ ### Added
34
+
35
+ ### Changed
36
+
37
+ ### Fixed
38
+
39
+ - Start a fresh eval cell when a caller reuses the ID of a terminal cell, preventing completed or failed results from being replayed as though new code had executed ([#439](https://github.com/code-yeongyu/senpi/pull/439)).
40
+ - Omit the eval `took` duration when timing metadata is unavailable, avoiding misleading zero-duration status output for detached or restored cell results ([#439](https://github.com/code-yeongyu/senpi/pull/439)).
41
+
42
+ ### Removed
43
+
44
+ ## [2026.7.28] - 2026-07-28
45
+
46
+ ### Breaking Changes
47
+
48
+ ### Added
49
+
50
+ - Add `tool_schema()` and return parameter schemas from failed eval tool calls so cells can inspect and self-correct tool invocations ([#407](https://github.com/code-yeongyu/senpi/pull/407)).
51
+
52
+ ### Changed
53
+
54
+ - Allow eval cells and extensions to activate named searchable tools lazily on the calling surface without globally widening the active tool set ([#408](https://github.com/code-yeongyu/senpi/pull/408)).
55
+
56
+ ### Fixed
57
+
58
+ ### Removed
59
+
15
60
  ## [2026.7.26] - 2026-07-26
16
61
 
17
62
  ### Breaking Changes
package/README.md CHANGED
@@ -101,6 +101,7 @@ options object and asynchronous helpers are `await`-able.
101
101
  | `write(path, content)` | Creates parent directories and writes text. `local://` paths persist in the session artifact root. |
102
102
  | `env(key?, value?)` | Reads all kernel environment values, one value, or sets one value. |
103
103
  | `tool.<name>(args)` | Invokes an active Senpi tool through the normal `pi.executeTool` pipeline. |
104
+ | `tool_schema(name?)` | Returns a tool's parameter schema without calling it; omit `name` to list tool names. |
104
105
  | `completion(prompt, model?, system?, schema?)` | Requests a one-shot host completion; `schema` asks the host to parse structured output. |
105
106
  | `agent(prompt, ...)` | Delegates to the configured active `taskTools.task` tool. Supports background handles and structured JSON results. |
106
107
  | `output(ids, format?, offset?, limit?)` | Delegates transcript retrieval to the configured active `taskTools.output` tool. |
@@ -108,6 +109,11 @@ options object and asynchronous helpers are `await`-able.
108
109
  | `pipeline(items, ...stages)` | Applies stages left to right with a barrier between stages. |
109
110
  | `log(message)` / `phase(title)` | Emits progress text and starts a status phase. |
110
111
 
112
+ When a `tool.<name>()` call fails argument validation, the error delivered back
113
+ into the cell carries the tool's expected parameters, so the cell can correct the
114
+ arguments and retry instead of falling back to one-at-a-time tool calls.
115
+ `tool_schema()` exposes the same catalog up front.
116
+
111
117
  `agent()` is available only when the configured task tool is active in the
112
118
  session. `output()` similarly requires the configured task-output tool and
113
119
  returns immediately: a running task reports its current status, while completed
@@ -179,3 +185,18 @@ Direct real-surface QA drivers live in `scripts/qa-*.ts`: kernel cells
179
185
  (`qa-py-cell.ts`, `qa-js-cell.ts`, `qa-rb-cell.ts`, `qa-jl-cell.ts`), end-to-end
180
186
  extension execution (`qa-e2e-eval.ts`), and renderer output
181
187
  (`qa-render-dump.ts`).
188
+
189
+ ### Nested tool-call widgets
190
+
191
+ When an eval cell invokes `tool.<name>(...)`, the result panel can render a
192
+ nested widget for the invoked tool. The widget captures bounded args, duration,
193
+ and a sanitized 160 code points result preview; the rendering path is
194
+ always-on and does not depend on any toggle or session flag.
195
+
196
+ The capture budget is fixed at 30 enriched calls per cell, with a 4096-character
197
+ serialized args budget. Previews are capped at 160 code points, and collapsed
198
+ widgets stay within the 8 lines collapsed widget budget.
199
+
200
+ Entries without args — including old sessions, reserved/completion rows, and
201
+ calls past the cap — render as plain rows. Edit renders a fallback row by
202
+ design, even when its args are present.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@code-yeongyu/senpi-codemode",
3
- "version": "2026.7.26",
3
+ "version": "2026.7.28-3",
4
4
  "description": "Source-only senpi extension package for codemode evaluation tools",
5
5
  "type": "module",
6
6
  "main": "./src/index.ts",
@@ -30,14 +30,14 @@
30
30
  },
31
31
  "dependencies": {
32
32
  "@babel/parser": "7.29.7",
33
- "@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@^2026.7.26",
33
+ "@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@^2026.7.28-3",
34
34
  "typebox": "1.1.38"
35
35
  },
36
36
  "peerDependencies": {
37
37
  "@code-yeongyu/senpi": "*"
38
38
  },
39
39
  "devDependencies": {
40
- "@code-yeongyu/senpi": "2026.7.26"
40
+ "@code-yeongyu/senpi": "2026.7.28-3"
41
41
  },
42
42
  "keywords": [
43
43
  "senpi",
@@ -3,6 +3,8 @@
3
3
  export const RESERVED_AGENT_TOOL = "__agent__" as const;
4
4
  /** ADAPTATION: senpi delegates output through a reserved kernel-side tool name. */
5
5
  export const RESERVED_OUTPUT_TOOL = "__output__" as const;
6
+ /** ADAPTATION: senpi resolves tool parameter schemas through a reserved kernel-side tool name. */
7
+ export const RESERVED_SCHEMA_TOOL = "__schema__" as const;
6
8
  /** Canonical oh-my-pi eval-timeout pause operation. */
7
9
  export const TIMEOUT_PAUSE_OP = "timeout-pause" as const;
8
10
  /** Canonical oh-my-pi eval-timeout resume operation. */
@@ -0,0 +1,56 @@
1
+ import type { AgentToolResult } from "@code-yeongyu/senpi";
2
+ import { RESERVED_AGENT_TOOL, RESERVED_OUTPUT_TOOL, RESERVED_SCHEMA_TOOL } from "../bridge/reserved.ts";
3
+ import type { EvalStatusEvent, ExecuteTool } from "../tool/types.ts";
4
+ import { type AgentExecuteTool, runEvalAgent } from "./agent-bridge.ts";
5
+ import { type MarshalledToolResult, type OutputExecuteTool, runEvalOutput } from "./output-bridge.ts";
6
+ import { type EvalSchemaToolInfo, runEvalSchema } from "./schema-bridge.ts";
7
+
8
+ export interface ReservedDispatchContext {
9
+ readonly callId: string;
10
+ readonly args: unknown;
11
+ readonly executeTool: AgentExecuteTool & OutputExecuteTool & ExecuteTool;
12
+ readonly taskToolName: string;
13
+ readonly taskOutputToolName: string;
14
+ readonly listTools: (() => readonly EvalSchemaToolInfo[]) | undefined;
15
+ readonly signal: AbortSignal | undefined;
16
+ readonly emitStatus: (event: EvalStatusEvent) => void;
17
+ readonly marshalToolResult: (result: AgentToolResult<unknown>) => MarshalledToolResult;
18
+ }
19
+
20
+ class SchemaUnavailableError extends Error {
21
+ readonly name = "SchemaUnavailableError";
22
+
23
+ constructor() {
24
+ super("tool_schema() unavailable: this session does not expose a tool catalog");
25
+ }
26
+ }
27
+
28
+ export function isReservedToolName(toolName: string): boolean {
29
+ return toolName === RESERVED_AGENT_TOOL || toolName === RESERVED_OUTPUT_TOOL || toolName === RESERVED_SCHEMA_TOOL;
30
+ }
31
+
32
+ export async function runReservedTool(toolName: string, context: ReservedDispatchContext): Promise<unknown> {
33
+ if (toolName === RESERVED_AGENT_TOOL) {
34
+ return await runEvalAgent(context.args, {
35
+ callId: context.callId,
36
+ taskToolName: context.taskToolName,
37
+ executeTool: context.executeTool,
38
+ ...(context.signal === undefined ? {} : { signal: context.signal }),
39
+ emitStatus: context.emitStatus,
40
+ });
41
+ }
42
+ if (toolName === RESERVED_OUTPUT_TOOL) {
43
+ return await runEvalOutput(context.args, {
44
+ taskOutputToolName: context.taskOutputToolName,
45
+ executeTool: context.executeTool,
46
+ ...(context.signal === undefined ? {} : { signal: context.signal }),
47
+ marshalToolResult: context.marshalToolResult,
48
+ });
49
+ }
50
+ if (toolName === RESERVED_SCHEMA_TOOL) {
51
+ const listTools = context.listTools;
52
+ if (listTools === undefined) throw new SchemaUnavailableError();
53
+ return runEvalSchema(context.args, { listTools });
54
+ }
55
+ throw new Error(`runReservedTool received a non-reserved tool name: ${toolName}`);
56
+ }
@@ -0,0 +1,69 @@
1
+ import { type Static, Type } from "typebox";
2
+ import { Check, Errors } from "typebox/value";
3
+
4
+ const schemaArgsSchema = Type.Object(
5
+ { name: Type.Optional(Type.String({ minLength: 1 })) },
6
+ { additionalProperties: false },
7
+ );
8
+
9
+ type SchemaArgs = Static<typeof schemaArgsSchema>;
10
+
11
+ export interface EvalSchemaToolInfo {
12
+ readonly name: string;
13
+ readonly description?: string | undefined;
14
+ readonly parameters?: unknown;
15
+ }
16
+
17
+ export interface RunEvalSchemaOptions {
18
+ readonly listTools: () => readonly EvalSchemaToolInfo[];
19
+ }
20
+
21
+ export type EvalSchemaResult =
22
+ | { readonly tools: readonly string[] }
23
+ | { readonly name: string; readonly description: string | undefined; readonly parameters: unknown };
24
+
25
+ class SchemaArgumentsError extends Error {
26
+ readonly name = "SchemaArgumentsError";
27
+
28
+ constructor(summary: string) {
29
+ super(`schema() received invalid arguments: ${summary}`);
30
+ }
31
+ }
32
+
33
+ class SchemaUnknownToolError extends Error {
34
+ readonly name = "SchemaUnknownToolError";
35
+
36
+ constructor(requested: string, suggestions: readonly string[]) {
37
+ const hint = suggestions.length > 0 ? ` Did you mean: ${suggestions.join(", ")}?` : "";
38
+ super(`schema() found no tool named "${requested}".${hint}`);
39
+ }
40
+ }
41
+
42
+ export function runEvalSchema(args: unknown, options: RunEvalSchemaOptions): EvalSchemaResult {
43
+ const parsed = parseSchemaArgs(args);
44
+ const tools = options.listTools();
45
+ if (parsed.name === undefined) return { tools: tools.map((tool) => tool.name) };
46
+
47
+ const match = tools.find((tool) => tool.name === parsed.name);
48
+ if (match) return { name: match.name, description: match.description, parameters: match.parameters };
49
+ throw new SchemaUnknownToolError(parsed.name, nearestNames(parsed.name, tools));
50
+ }
51
+
52
+ function parseSchemaArgs(value: unknown): SchemaArgs {
53
+ if (Check(schemaArgsSchema, value)) return value;
54
+ const summary = Errors(schemaArgsSchema, value)
55
+ .map((error) => `${error.instancePath || "/"} ${error.message}`)
56
+ .join("; ");
57
+ throw new SchemaArgumentsError(summary || "invalid value");
58
+ }
59
+
60
+ function nearestNames(requested: string, tools: readonly EvalSchemaToolInfo[]): readonly string[] {
61
+ const needle = requested.toLowerCase();
62
+ return tools
63
+ .map((tool) => tool.name)
64
+ .filter((name) => {
65
+ const candidate = name.toLowerCase();
66
+ return candidate.includes(needle) || needle.includes(candidate);
67
+ })
68
+ .slice(0, 5);
69
+ }
@@ -0,0 +1,140 @@
1
+ const SCHEMA_HINT_MAX_CHARS = 1_200;
2
+ const SCHEMA_HINT_MAX_PROPERTIES = 30;
3
+ const DESCRIPTION_MAX_CHARS = 80;
4
+ const ENUM_MAX_VALUES = 8;
5
+ const NESTED_MAX_DEPTH = 2;
6
+ const HINT_HEADER = "Expected parameters:";
7
+
8
+ interface SchemaLike {
9
+ readonly type?: unknown;
10
+ readonly properties?: unknown;
11
+ readonly required?: unknown;
12
+ readonly items?: unknown;
13
+ readonly description?: unknown;
14
+ readonly enum?: unknown;
15
+ readonly const?: unknown;
16
+ readonly oneOf?: unknown;
17
+ readonly anyOf?: unknown;
18
+ readonly allOf?: unknown;
19
+ }
20
+
21
+ export function appendSchemaHint(message: string, toolName: string, schema: unknown): string {
22
+ const rendered = renderSchemaHint(toolName, schema);
23
+ if (rendered === undefined) return message;
24
+ return `${message}\n\n${HINT_HEADER}\n${rendered}`;
25
+ }
26
+
27
+ export function renderSchemaHint(toolName: string, schema: unknown): string | undefined {
28
+ if (!isSchemaLike(schema)) return undefined;
29
+ const lines = renderObjectLines(schema, 0, 1);
30
+ if (lines.length === 0) return undefined;
31
+ return fitLines(lines, `[truncated; call tool_schema(${JSON.stringify(toolName)}) for the full schema]`);
32
+ }
33
+
34
+ function renderObjectLines(schema: SchemaLike, indent: number, depth: number): readonly string[] {
35
+ const required = stringArray(schema.required);
36
+ const entries = propertyEntries(schema.properties);
37
+ if (entries.length === 0 && required.length === 0) return [];
38
+
39
+ const ordered = [
40
+ ...entries.filter(([name]) => required.includes(name)),
41
+ ...entries.filter(([name]) => !required.includes(name)),
42
+ ];
43
+ const pad = " ".repeat(indent + 1);
44
+ const lines: string[] = [];
45
+ if (indent === 0 && required.length > 0) lines.push(`required: ${required.join(", ")}`);
46
+
47
+ const shown = ordered.slice(0, SCHEMA_HINT_MAX_PROPERTIES);
48
+ for (const [name, value] of shown) {
49
+ lines.push(`${pad}${name}${required.includes(name) ? "" : "?"}: ${describeProperty(value)}`);
50
+ lines.push(...nestedLines(value, indent, depth));
51
+ }
52
+ const hidden = ordered.length - shown.length;
53
+ if (hidden > 0) lines.push(`${pad}… ${hidden} more propert${hidden === 1 ? "y" : "ies"}`);
54
+ return lines;
55
+ }
56
+
57
+ function nestedLines(value: unknown, indent: number, depth: number): readonly string[] {
58
+ if (depth >= NESTED_MAX_DEPTH || !isSchemaLike(value)) return [];
59
+ const nested = isSchemaLike(value.items) ? value.items : value;
60
+ if (propertyEntries(nested.properties).length === 0) return [];
61
+ return renderObjectLines(nested, indent + 1, depth + 1);
62
+ }
63
+
64
+ function describeProperty(value: unknown): string {
65
+ if (!isSchemaLike(value)) return "unknown";
66
+ const parts = [typeLabel(value)];
67
+ const description = firstLine(value.description);
68
+ if (description !== undefined) parts.push(`— ${truncate(description, DESCRIPTION_MAX_CHARS)}`);
69
+ return parts.join(" ");
70
+ }
71
+
72
+ function typeLabel(schema: SchemaLike): string {
73
+ if (Object.hasOwn(schema, "const")) return JSON.stringify(schema.const);
74
+ const enumValues = schema.enum;
75
+ if (Array.isArray(enumValues) && enumValues.length > 0) return unionLabel(enumValues.map(literal));
76
+ const branches = schema.oneOf ?? schema.anyOf;
77
+ if (Array.isArray(branches) && branches.length > 0) return unionLabel(branches.map(branchLabel));
78
+ if (Array.isArray(schema.allOf) && schema.allOf.length > 0) return branchLabel(schema.allOf[0]);
79
+
80
+ const type = schema.type;
81
+ const base = typeof type === "string" ? type : Array.isArray(type) ? type.filter(isString).join("|") : "unknown";
82
+ if (base !== "array") return base;
83
+ return isSchemaLike(schema.items) ? `array<${typeLabel(schema.items)}>` : "array";
84
+ }
85
+
86
+ function unionLabel(values: readonly string[]): string {
87
+ const shown = values.slice(0, ENUM_MAX_VALUES).join(" | ");
88
+ return values.length > ENUM_MAX_VALUES ? `${shown} | … (${values.length - ENUM_MAX_VALUES} more)` : shown;
89
+ }
90
+
91
+ function branchLabel(value: unknown): string {
92
+ return isSchemaLike(value) ? typeLabel(value) : "unknown";
93
+ }
94
+
95
+ function literal(value: unknown): string {
96
+ return JSON.stringify(value) ?? String(value);
97
+ }
98
+
99
+ function fitLines(lines: readonly string[], marker: string): string {
100
+ const budget = SCHEMA_HINT_MAX_CHARS - HINT_HEADER.length - 1;
101
+ const kept: string[] = [];
102
+ let used = 0;
103
+ for (const line of lines) {
104
+ const next = used + line.length + (kept.length === 0 ? 0 : 1);
105
+ if (next > budget - marker.length - 1) {
106
+ kept.push(marker);
107
+ return kept.join("\n");
108
+ }
109
+ kept.push(line);
110
+ used = next;
111
+ }
112
+ return kept.join("\n");
113
+ }
114
+
115
+ function propertyEntries(properties: unknown): readonly (readonly [string, unknown])[] {
116
+ if (typeof properties !== "object" || properties === null || Array.isArray(properties)) return [];
117
+ return Object.entries(properties);
118
+ }
119
+
120
+ function stringArray(value: unknown): readonly string[] {
121
+ return Array.isArray(value) ? value.filter(isString) : [];
122
+ }
123
+
124
+ function firstLine(value: unknown): string | undefined {
125
+ if (typeof value !== "string") return undefined;
126
+ const line = value.split("\n", 1)[0]?.trim();
127
+ return line === undefined || line.length === 0 ? undefined : line;
128
+ }
129
+
130
+ function truncate(value: string, maxChars: number): string {
131
+ return value.length <= maxChars ? value : `${value.slice(0, maxChars)}…`;
132
+ }
133
+
134
+ function isSchemaLike(value: unknown): value is SchemaLike {
135
+ return typeof value === "object" && value !== null && !Array.isArray(value);
136
+ }
137
+
138
+ function isString(value: unknown): value is string {
139
+ return typeof value === "string";
140
+ }
@@ -87,8 +87,10 @@ export async function createRuntime(
87
87
  }
88
88
 
89
89
  export function createExecuteTool(pi: CodemodeRuntimeAPI, activeTools?: ReadonlySet<string>): AgentExecuteTool {
90
+ // A cell names tools directly and cannot run tool_search first, so eval opts in to
91
+ // lazy activation. Eligibility still belongs to the extension that registered the tool.
90
92
  const executeTool: AgentExecuteTool = (toolName, params, executeOptions) =>
91
- pi.executeTool(toolName, params, executeOptions);
93
+ pi.executeTool(toolName, params, { ...executeOptions, activateInactiveTool: true });
92
94
  return Object.assign(executeTool, {
93
95
  isToolAvailable: (name: string): boolean => activeTools?.has(name) ?? pi.getActiveTools().includes(name),
94
96
  });
package/src/index.ts CHANGED
@@ -1,6 +1,7 @@
1
1
  import * as os from "node:os";
2
2
  import type { ExtensionContext } from "@code-yeongyu/senpi";
3
3
  import type { AgentExecuteTool } from "./bridges/agent-bridge.ts";
4
+ import type { EvalSchemaToolInfo } from "./bridges/schema-bridge.ts";
4
5
  import { type CompletionRequest, type CompletionResult, createCompletionHandler } from "./completion/handler.ts";
5
6
  import { defaultCodemodeSettings } from "./config/settings.ts";
6
7
  import { EvalNotifier } from "./extension/eval-notifier.ts";
@@ -33,6 +34,7 @@ export interface CodemodeExtensionAPI {
33
34
  on(event: CodemodeEvent, handler: (event: unknown, ctx: ExtensionContext) => Promise<void> | void): void;
34
35
  executeTool: AgentExecuteTool;
35
36
  getActiveTools(): string[];
37
+ getAllTools(): readonly EvalSchemaToolInfo[];
36
38
  sendUserMessage(content: string, options?: { deliverAs?: "steer" | "followUp" }): void;
37
39
  }
38
40
 
@@ -67,6 +69,7 @@ export default function senpiCodemode(pi: CodemodeExtensionAPI, options: SenpiCo
67
69
  kernelManager: manager,
68
70
  cellTimeoutSeconds: runtime.settings.cellTimeoutSeconds,
69
71
  executeTool: runtime.executeTool,
72
+ listTools: () => pi.getAllTools(),
70
73
  complete,
71
74
  settings: runtime.settings,
72
75
  artifactsDir: runtime.artifactsDir,
@@ -95,6 +98,7 @@ export default function senpiCodemode(pi: CodemodeExtensionAPI, options: SenpiCo
95
98
  kernelManager: manager,
96
99
  cellTimeoutSeconds: defaultCodemodeSettings.cellTimeoutSeconds,
97
100
  executeTool: createExecuteTool(pi),
101
+ listTools: () => pi.getAllTools(),
98
102
  complete,
99
103
  settings: defaultCodemodeSettings,
100
104
  cellManager: new EvalDetachedCellManager({ notifier }),
@@ -194,6 +194,12 @@ function completion(prompt::AbstractString; model="default", system=nothing, sch
194
194
  get(response, "text", response)
195
195
  end
196
196
 
197
+ function tool_schema(name=nothing)
198
+ arguments = Dict{String, Any}()
199
+ name !== nothing && (arguments["name"] = string(name))
200
+ senpi_with_bridge_timeout_pause(() -> senpi_call_tool("__schema__", arguments))
201
+ end
202
+
197
203
  function output(ids...; format="raw", offset=nothing, limit=nothing)
198
204
  isempty(ids) && error("At least one output ID is required")
199
205
  format in ("raw", "tail") || error("output() format must be 'raw' or 'tail'")
@@ -4,6 +4,7 @@ import type { BridgeConnectionConfig } from "../../bridge/protocol.ts";
4
4
  import {
5
5
  RESERVED_AGENT_TOOL,
6
6
  RESERVED_OUTPUT_TOOL,
7
+ RESERVED_SCHEMA_TOOL,
7
8
  TIMEOUT_PAUSE_OP,
8
9
  TIMEOUT_RESUME_OP,
9
10
  } from "../../bridge/reserved.ts";
@@ -33,6 +34,7 @@ type RuntimeModuleContext = {
33
34
  readonly cwdUrl: string;
34
35
  readonly localRootUrls: Readonly<Record<string, string>>;
35
36
  readonly reservedAgentTool: string;
37
+ readonly reservedSchemaTool: string;
36
38
  readonly reservedOutputTool: string;
37
39
  readonly timeoutPauseOp: string;
38
40
  readonly timeoutResumeOp: string;
@@ -55,6 +57,7 @@ function runtimeContext(options: LocalModuleLoaderOptions): RuntimeModuleContext
55
57
  localRootUrls: roots,
56
58
  reservedAgentTool: RESERVED_AGENT_TOOL,
57
59
  reservedOutputTool: RESERVED_OUTPUT_TOOL,
60
+ reservedSchemaTool: RESERVED_SCHEMA_TOOL,
58
61
  timeoutPauseOp: TIMEOUT_PAUSE_OP,
59
62
  timeoutResumeOp: TIMEOUT_RESUME_OP,
60
63
  };
@@ -66,6 +69,7 @@ function loaderPrelude(context: RuntimeModuleContext): string {
66
69
  `globalThis.__senpi_module_context__ = ${serialized};`,
67
70
  "globalThis.__senpi_reserved_agent_tool__ = globalThis.__senpi_module_context__.reservedAgentTool;",
68
71
  "globalThis.__senpi_reserved_output_tool__ = globalThis.__senpi_module_context__.reservedOutputTool;",
72
+ "globalThis.__senpi_reserved_schema_tool__ = globalThis.__senpi_module_context__.reservedSchemaTool;",
69
73
  "globalThis.__senpi_timeout_pause_op__ = globalThis.__senpi_module_context__.timeoutPauseOp;",
70
74
  "globalThis.__senpi_timeout_resume_op__ = globalThis.__senpi_module_context__.timeoutResumeOp;",
71
75
  "globalThis.__senpi_import__ = async (source, options) => {",
@@ -50,6 +50,7 @@ export class JsWorkerRuntime {
50
50
  globalThis.read = async (path, options, ...rest) => await this.#read(path, helperOptions("read", options, rest));
51
51
  globalThis.write = async (path, content) => await this.#write(path, content);
52
52
  globalThis.output = async (...args) => await this.#output(args);
53
+ globalThis.tool_schema = async name => await this.#toolSchema(name);
53
54
  globalThis.agent = async (prompt, options, ...rest) => await this.#agent(prompt, options, rest);
54
55
  globalThis.parallel = async thunks => await this.#parallel(thunks);
55
56
  globalThis.pipeline = async (items, ...stages) => await this.#pipeline(items, stages);
@@ -250,6 +251,11 @@ export class JsWorkerRuntime {
250
251
  return node;
251
252
  }
252
253
 
254
+ async #toolSchema(name) {
255
+ const args = name === undefined || name === null ? {} : { name: String(name) };
256
+ return await this.#callTool(reservedTool("__senpi_reserved_schema_tool__", "tool_schema"), args);
257
+ }
258
+
253
259
  async #callTool(toolName, args) {
254
260
  const hooks = this.#hooks;
255
261
  if (!hooks) throw new Error("tool call outside active JS cell");
@@ -37,6 +37,7 @@ EMIT_LOCK = Lock()
37
37
  # Mirrors src/bridge/reserved.ts; this standalone subprocess asset cannot import TypeScript.
38
38
  RESERVED_AGENT_TOOL = "__agent__"
39
39
  RESERVED_OUTPUT_TOOL = "__output__"
40
+ RESERVED_SCHEMA_TOOL = "__schema__"
40
41
  TIMEOUT_PAUSE_OP = "timeout-pause"
41
42
  TIMEOUT_RESUME_OP = "timeout-resume"
42
43
 
@@ -374,6 +375,14 @@ def completion(
374
375
  return response.get("text", response)
375
376
 
376
377
 
378
+ def tool_schema(name: str | None = None) -> Any:
379
+ args: dict[str, Any] = {} if name is None else {"name": name}
380
+ return bridge_post(
381
+ "/call",
382
+ {"callId": f"py-{uuid.uuid4()}", "toolName": RESERVED_SCHEMA_TOOL, "args": args},
383
+ )
384
+
385
+
377
386
  def output(
378
387
  *ids: str,
379
388
  format: str = "raw",
@@ -844,6 +853,7 @@ USER_NS.update(
844
853
  "completion": completion,
845
854
  "agent": agent,
846
855
  "output": output,
856
+ "tool_schema": tool_schema,
847
857
  "__senpi_magic": _magic,
848
858
  "__senpi_magic_cell": _magic_cell,
849
859
  "__senpi_shell": _shell,
@@ -5,6 +5,7 @@ require "uri"
5
5
 
6
6
  SENPI_RESERVED_AGENT_TOOL = "__agent__"
7
7
  SENPI_RESERVED_OUTPUT_TOOL = "__output__"
8
+ SENPI_RESERVED_SCHEMA_TOOL = "__schema__"
8
9
  SENPI_INTERNAL_URL = Regexp.new("\\A([a-z][a-z0-9+.\\-]*)://(.*)\\z", Regexp::IGNORECASE)
9
10
 
10
11
  def __senpi_status_enabled?
@@ -187,6 +188,11 @@ def completion(prompt, model: "default", system: nil, schema: nil, **kwargs)
187
188
  result.fetch("text", result)
188
189
  end
189
190
 
191
+ def tool_schema(name = nil)
192
+ args = name.nil? ? {} : { "name" => name.to_s }
193
+ __senpi_call_tool(SENPI_RESERVED_SCHEMA_TOOL, args)
194
+ end
195
+
190
196
  def output(*ids, format: "raw", offset: nil, limit: nil)
191
197
  raise ArgumentError, "At least one output ID is required" if ids.empty?
192
198
  raise ArgumentError, "output() format must be 'raw' or 'tail'" unless ["raw", "tail"].include?(format)
@@ -154,6 +154,11 @@ env(key?=None, value?=None) → str | None | dict
154
154
  Task/agent output by id. Reads immediately: running tasks return their status; \`format\` selects full (\`"raw"\`) or trailing (\`"tail"\`) output.
155
155
  {{/if}}tool.<name>(args) → unknown
156
156
  Invoke any session tool; \`args\` = its parameter object.
157
+ tool_schema(name?) → dict
158
+ Parameter schema of a tool without calling it; omit \`name\` to list tool names.
159
+ Use it before calling a tool you have not called before — a failed call also
160
+ returns the expected parameters, so fix the args and retry in the next cell
161
+ instead of abandoning eval.
157
162
  completion(prompt, model?="default", system?=None, schema?=None) → str | dict
158
163
  Oneshot, stateless (no history/tools). \`model\`: \`"smol"\` fast | \`"default"\` session | \`"slow"\` most capable. \`schema\` (JSON-Schema) → structured output, parsed object.
159
164
  {{#if spawns}}agent(prompt, agent?="{{spawnDefaultAgent}}", model?=None, label?=None, schema?=None, handle?=False) → str | dict
@@ -0,0 +1,93 @@
1
+ import { type AgentToolResult, sanitizeTerminalLabel } from "@code-yeongyu/senpi";
2
+
3
+ export const MAX_ENRICHED_TOOL_CALLS = 30;
4
+
5
+ const MAX_ARGUMENT_STRING_CODE_POINTS = 512;
6
+ const MAX_ARGUMENT_ENTRIES = 32;
7
+ const MAX_ARGUMENT_DEPTH = 6;
8
+ const MAX_SERIALIZED_ARGUMENT_LENGTH = 4096;
9
+ const MAX_RESULT_PREVIEW_CODE_POINTS = 160;
10
+
11
+ type BoundedValue = {
12
+ readonly value: unknown;
13
+ readonly truncated: boolean;
14
+ };
15
+
16
+ export function capCodePoints(text: string, max: number): string {
17
+ let end = 0;
18
+ let count = 0;
19
+ while (count < max && end < text.length) {
20
+ const firstCodeUnit = text.charCodeAt(end);
21
+ const secondCodeUnit = text.charCodeAt(end + 1);
22
+ const isSurrogatePair =
23
+ firstCodeUnit >= 0xd800 && firstCodeUnit <= 0xdbff && secondCodeUnit >= 0xdc00 && secondCodeUnit <= 0xdfff;
24
+ end += isSurrogatePair ? 2 : 1;
25
+ count += 1;
26
+ }
27
+ return end === text.length ? text : `${text.slice(0, end)}…`;
28
+ }
29
+
30
+ function boundValue(value: unknown, depth: number, ancestors: WeakSet<object>): BoundedValue {
31
+ if (typeof value === "string") {
32
+ const capped = capCodePoints(value, MAX_ARGUMENT_STRING_CODE_POINTS);
33
+ return { value: capped, truncated: capped !== value };
34
+ }
35
+
36
+ if (value === null || typeof value !== "object") return { value, truncated: false };
37
+ if (depth >= MAX_ARGUMENT_DEPTH) return { value: "…", truncated: true };
38
+ if (ancestors.has(value)) throw new Error("cyclic tool-call arguments");
39
+
40
+ ancestors.add(value);
41
+ try {
42
+ if (Array.isArray(value)) {
43
+ const retainedLength = Math.min(value.length, MAX_ARGUMENT_ENTRIES);
44
+ const clone: unknown[] = [];
45
+ let truncated = value.length > MAX_ARGUMENT_ENTRIES;
46
+ for (let index = 0; index < retainedLength; index += 1) {
47
+ const bounded = boundValue(value[index], depth + 1, ancestors);
48
+ clone.push(bounded.value);
49
+ truncated ||= bounded.truncated;
50
+ }
51
+ return { value: clone, truncated };
52
+ }
53
+
54
+ const entries: [string, unknown][] = [];
55
+ let truncated = false;
56
+ let count = 0;
57
+ for (const [key, nestedValue] of Object.entries(value)) {
58
+ if (count >= MAX_ARGUMENT_ENTRIES) {
59
+ truncated = true;
60
+ break;
61
+ }
62
+ const bounded = boundValue(nestedValue, depth + 1, ancestors);
63
+ entries.push([key, bounded.value]);
64
+ truncated ||= bounded.truncated;
65
+ count += 1;
66
+ }
67
+ return { value: Object.fromEntries(entries), truncated };
68
+ } finally {
69
+ ancestors.delete(value);
70
+ }
71
+ }
72
+
73
+ export function boundToolCallArgs(args: unknown): { args: unknown; truncated: boolean } {
74
+ try {
75
+ const bounded = boundValue(args, 0, new WeakSet<object>());
76
+ const serialized = JSON.stringify(bounded.value);
77
+ if (typeof serialized !== "string" || serialized.length > MAX_SERIALIZED_ARGUMENT_LENGTH) {
78
+ return { args: undefined, truncated: true };
79
+ }
80
+ return { args: bounded.value, truncated: bounded.truncated };
81
+ } catch {
82
+ return { args: undefined, truncated: true };
83
+ }
84
+ }
85
+
86
+ export function toolCallResultPreview(result: AgentToolResult<unknown>): string | undefined {
87
+ for (const part of result.content) {
88
+ if (part.type !== "text") continue;
89
+ const preview = sanitizeTerminalLabel(part.text).replace(/\s+/gu, " ").trim();
90
+ return preview.length === 0 ? undefined : capCodePoints(preview, MAX_RESULT_PREVIEW_CODE_POINTS);
91
+ }
92
+ return undefined;
93
+ }
@@ -1,11 +1,18 @@
1
- import type { AgentToolResult, AgentToolUpdateCallback, ExtensionContext } from "@code-yeongyu/senpi";
1
+ import {
2
+ type AgentToolResult,
3
+ type AgentToolUpdateCallback,
4
+ type ExtensionContext,
5
+ sanitizeTerminalLabel,
6
+ } from "@code-yeongyu/senpi";
2
7
  import type { KernelToHostMessage } from "../bridge/protocol.ts";
3
- import { RESERVED_AGENT_TOOL, RESERVED_OUTPUT_TOOL } from "../bridge/reserved.ts";
4
- import { type AgentExecuteTool, runEvalAgent } from "../bridges/agent-bridge.ts";
5
- import { runEvalOutput } from "../bridges/output-bridge.ts";
8
+ import type { AgentExecuteTool } from "../bridges/agent-bridge.ts";
9
+ import { isReservedToolName, runReservedTool } from "../bridges/reserved-dispatch.ts";
10
+ import type { EvalSchemaToolInfo } from "../bridges/schema-bridge.ts";
11
+ import { appendSchemaHint } from "../bridges/schema-hint.ts";
6
12
  import type { CompletionRequest, CompletionResult } from "../completion/handler.ts";
7
13
  import { handleCompletionToolCall } from "../completion/tool-bridge.ts";
8
14
  import type { ResolvedCodemodeSettings } from "../config/settings.ts";
15
+ import { boundToolCallArgs, capCodePoints, MAX_ENRICHED_TOOL_CALLS, toolCallResultPreview } from "./call-capture.ts";
9
16
  import {
10
17
  type EvalImageResizer,
11
18
  EvalOutputCollector,
@@ -16,7 +23,19 @@ import {
16
23
  import { upsertStatusEvent } from "./status-events.ts";
17
24
  import type { EvalKernel, EvalStatusEvent, EvalToolDetails, EvalToolInput } from "./types.ts";
18
25
 
19
- type ResolvedToolReply = { readonly value: unknown; readonly toolCallOk: boolean };
26
+ type ResolvedToolReply = {
27
+ readonly value: unknown;
28
+ readonly toolCallOk: boolean;
29
+ readonly resultPreview?: string;
30
+ readonly errorText?: string;
31
+ };
32
+
33
+ type ToolCallEnrichment = {
34
+ readonly callId: string;
35
+ readonly args: unknown;
36
+ readonly startedAt: number;
37
+ readonly argsTruncated?: true;
38
+ };
20
39
 
21
40
  const LIVE_OUTPUT_PREVIEW_LINES = 8;
22
41
 
@@ -36,6 +55,7 @@ export interface CellState {
36
55
 
37
56
  export interface CellBridgeRuntime {
38
57
  readonly executeTool: AgentExecuteTool;
58
+ readonly listTools?: () => readonly EvalSchemaToolInfo[];
39
59
  readonly settings: ResolvedCodemodeSettings;
40
60
  readonly complete?: (request: CompletionRequest, ctx: ExtensionContext) => Promise<CompletionResult>;
41
61
  readonly ctx: ExtensionContext;
@@ -151,25 +171,17 @@ export class CellHandler {
151
171
  });
152
172
  return;
153
173
  }
154
- if (message.toolName === RESERVED_AGENT_TOOL) {
174
+ if (isReservedToolName(message.toolName)) {
155
175
  await this.#deliverToolReply(message, async () => ({
156
- value: await runEvalAgent(message.args, {
176
+ value: await runReservedTool(message.toolName, {
157
177
  callId: message.callId,
158
- taskToolName: this.#runtime.settings.taskTools.task,
178
+ args: message.args,
159
179
  executeTool: this.#runtime.executeTool,
160
- signal: this.#state.signal,
161
- emitStatus: (event) => this.#recordStatus(event),
162
- }),
163
- toolCallOk: true,
164
- }));
165
- return;
166
- }
167
- if (message.toolName === RESERVED_OUTPUT_TOOL) {
168
- await this.#deliverToolReply(message, async () => ({
169
- value: await runEvalOutput(message.args, {
180
+ taskToolName: this.#runtime.settings.taskTools.task,
170
181
  taskOutputToolName: this.#runtime.settings.taskTools.output,
171
- executeTool: this.#runtime.executeTool,
182
+ listTools: this.#runtime.listTools,
172
183
  signal: this.#state.signal,
184
+ emitStatus: (event) => this.#recordStatus(event),
173
185
  marshalToolResult,
174
186
  }),
175
187
  toolCallOk: true,
@@ -193,25 +205,62 @@ export class CellHandler {
193
205
  this.#emitUpdate(false);
194
206
  return;
195
207
  }
196
- await this.#deliverToolReply(message, async () => {
197
- const result = await this.#runtime.executeTool(message.toolName, message.args, { signal: this.#state.signal });
198
- return { value: marshalToolResult(result), toolCallOk: !toolResultIsError(result) };
199
- });
208
+ const capturedArgs = boundToolCallArgs(message.args);
209
+ const startedAt = Date.now();
210
+ await this.#deliverToolReply(
211
+ message,
212
+ async () => {
213
+ const result = await this.#runtime.executeTool(message.toolName, message.args, {
214
+ signal: this.#state.signal,
215
+ });
216
+ const toolCallOk = !toolResultIsError(result);
217
+ if (toolCallOk) {
218
+ const resultPreview = toolCallResultPreview(result);
219
+ return {
220
+ value: marshalToolResult(result),
221
+ toolCallOk,
222
+ ...(resultPreview === undefined ? {} : { resultPreview }),
223
+ };
224
+ }
225
+ let errorText: string | undefined;
226
+ for (const part of result.content) {
227
+ if (part.type !== "text") continue;
228
+ errorText = capCodePoints(sanitizeTerminalLabel(part.text), 512);
229
+ break;
230
+ }
231
+ return {
232
+ value: marshalToolResult(result),
233
+ toolCallOk,
234
+ ...(errorText === undefined ? {} : { errorText }),
235
+ };
236
+ },
237
+ {
238
+ callId: message.callId,
239
+ args: capturedArgs.args,
240
+ startedAt,
241
+ ...(capturedArgs.truncated ? { argsTruncated: true } : {}),
242
+ },
243
+ );
200
244
  }
201
245
 
202
246
  async #deliverToolReply(
203
247
  message: Extract<KernelToHostMessage, { type: "tool-call" }>,
204
248
  resolve: () => Promise<ResolvedToolReply>,
249
+ enrich?: ToolCallEnrichment,
205
250
  ): Promise<void> {
206
251
  try {
207
252
  const reply = await resolve();
208
253
  if (!this.#state.active) return;
209
- this.#state.toolCalls.push({ name: message.toolName, ok: reply.toolCallOk });
254
+ this.#pushToolCall(message.toolName, reply.toolCallOk, enrich, reply.resultPreview, reply.errorText);
210
255
  this.#kernel.deliverToolReply({ type: "tool-reply", callId: message.callId, ok: true, value: reply.value });
211
256
  } catch (error) {
212
257
  if (!this.#state.active) return;
213
- const text = error instanceof Error ? error.message : String(error);
214
- this.#state.toolCalls.push({ name: message.toolName, ok: false, error: text });
258
+ const text = appendSchemaHint(
259
+ error instanceof Error ? error.message : String(error),
260
+ message.toolName,
261
+ this.#toolParameters(message.toolName),
262
+ );
263
+ this.#pushToolCall(message.toolName, false, enrich, undefined, text);
215
264
  this.#kernel.deliverToolReply({
216
265
  type: "tool-reply",
217
266
  callId: message.callId,
@@ -222,6 +271,33 @@ export class CellHandler {
222
271
  this.#emitUpdate(false);
223
272
  }
224
273
 
274
+ #pushToolCall(
275
+ name: string,
276
+ ok: boolean,
277
+ enrich: ToolCallEnrichment | undefined,
278
+ resultPreview: string | undefined,
279
+ error: string | undefined,
280
+ ): void {
281
+ const summary = { name, ok, ...(error === undefined ? {} : { error }) };
282
+ const enrichedCount = this.#state.toolCalls.filter((toolCall) => toolCall.callId !== undefined).length;
283
+ if (enrich === undefined || enrichedCount >= MAX_ENRICHED_TOOL_CALLS) {
284
+ this.#state.toolCalls.push(summary);
285
+ return;
286
+ }
287
+ this.#state.toolCalls.push({
288
+ ...summary,
289
+ callId: enrich.callId,
290
+ args: enrich.args,
291
+ durationMs: Date.now() - enrich.startedAt,
292
+ ...(enrich.argsTruncated === true ? { argsTruncated: true } : {}),
293
+ ...(resultPreview === undefined ? {} : { resultPreview }),
294
+ });
295
+ }
296
+
297
+ #toolParameters(toolName: string): unknown {
298
+ return this.#runtime.listTools?.().find((tool) => tool.name === toolName)?.parameters;
299
+ }
300
+
225
301
  #recordStatus(event: EvalStatusEvent): void {
226
302
  if (!this.#runtime.settings.statusEvents) return;
227
303
  upsertStatusEvent(this.#state.statusEvents, event);
@@ -68,7 +68,10 @@ export class EvalDetachedCellManager {
68
68
 
69
69
  create(cellId: string, input: EvalToolInput): ManagedCell {
70
70
  const existing = this.#cells.get(cellId);
71
- if (existing !== undefined) throw new Error(`Eval cell ${cellId} is already managed`);
71
+ if (existing !== undefined) {
72
+ if (existing.state === "running" || existing.state === "detached") throw activeCellReuseError(existing);
73
+ this.#cells.delete(cellId);
74
+ }
72
75
  const spillPath =
73
76
  this.#artifactsDir === undefined
74
77
  ? undefined
@@ -232,6 +235,12 @@ export class EvalDetachedCellManager {
232
235
  }
233
236
  }
234
237
 
238
+ function activeCellReuseError(cell: ManagedCell): Error {
239
+ return new Error(
240
+ `Eval cell ${cell.cellId} from a previous call is still ${cell.state} in the ${cell.input.language} kernel. Use eval({ action: "peek", cell_id: "${cell.cellId}" }) to read it or eval({ action: "stop", cell_id: "${cell.cellId}" }) to end it before its id can be reused.`,
241
+ );
242
+ }
243
+
235
244
  function allowsTransition(from: EvalDetachedCellState, to: EvalDetachedCellState): boolean {
236
245
  if (from === "running") return to === "detached" || to === "completed" || to === "failed" || to === "cancelled";
237
246
  if (from === "detached") return to === "completed" || to === "failed" || to === "cancelled";
@@ -1,6 +1,7 @@
1
1
  import { randomUUID } from "node:crypto";
2
2
  import { join } from "node:path";
3
3
  import type { AgentToolResult, AgentToolUpdateCallback, ExtensionContext, ToolDefinition } from "@code-yeongyu/senpi";
4
+ import type { EvalSchemaToolInfo } from "../bridges/schema-bridge.ts";
4
5
  import type { CompletionRequest, CompletionResult } from "../completion/handler.ts";
5
6
  import { defaultCodemodeSettings, type ResolvedCodemodeSettings } from "../config/settings.ts";
6
7
  import type { EvalExecutionTracker } from "../extension/session-manager.ts";
@@ -37,6 +38,7 @@ export interface CreateEvalToolOptions {
37
38
  readonly kernelManager: EvalKernelManager;
38
39
  readonly cellTimeoutSeconds: number;
39
40
  readonly executeTool: ExecuteTool;
41
+ readonly listTools?: () => readonly EvalSchemaToolInfo[];
40
42
  readonly complete?: (request: CompletionRequest, ctx: ExtensionContext) => Promise<CompletionResult>;
41
43
  readonly settings?: ResolvedCodemodeSettings;
42
44
  readonly artifactsDir?: string;
@@ -351,6 +353,7 @@ async function executeCell(
351
353
  execution.setKernel(kernel);
352
354
  handler = new CellHandler(kernel, state, {
353
355
  executeTool: options.executeTool,
356
+ ...(options.listTools === undefined ? {} : { listTools: options.listTools }),
354
357
  settings: options.settings ?? defaultCodemodeSettings,
355
358
  ...(options.complete === undefined ? {} : { complete: options.complete }),
356
359
  ctx: bridgeContext,
@@ -19,6 +19,7 @@ import {
19
19
  JSON_TREE_SCALAR_LEN_EXPANDED,
20
20
  renderJsonTreeLines,
21
21
  } from "./json-tree.ts";
22
+ import { codePointPrefix, formatDuration, renderToolCallWidget } from "./tool-widgets.ts";
22
23
  import type {
23
24
  EvalCellResult,
24
25
  EvalInputSchema,
@@ -116,18 +117,6 @@ function appendLines(target: string[], source: readonly string[]): void {
116
117
  for (const line of source) target.push(line);
117
118
  }
118
119
 
119
- function codePointPrefix(text: string, maxCodePoints: number): string {
120
- let end = 0;
121
- for (let count = 0; count < maxCodePoints && end < text.length; count += 1) {
122
- const firstCodeUnit = text.charCodeAt(end);
123
- const secondCodeUnit = text.charCodeAt(end + 1);
124
- const isSurrogatePair =
125
- firstCodeUnit >= 0xd800 && firstCodeUnit <= 0xdbff && secondCodeUnit >= 0xdc00 && secondCodeUnit <= 0xdfff;
126
- end += isSurrogatePair ? 2 : 1;
127
- }
128
- return text.slice(0, end);
129
- }
130
-
131
120
  function renderAllVisualLines(text: string, width: number): string[] {
132
121
  return truncateToVisualLines(text, Number.POSITIVE_INFINITY, width).visualLines.map((line) => line.trimEnd());
133
122
  }
@@ -232,18 +221,6 @@ function highlightedCode(code: string, language: EvalLanguage, theme: Theme | un
232
221
  return (theme === undefined ? lines.map((line) => line.replace(/\u001b\[[0-9;]*m/gu, "")) : lines).join("\n");
233
222
  }
234
223
 
235
- function formatDuration(milliseconds: number): string {
236
- const totalSeconds = Math.floor(Math.max(0, milliseconds) / 1_000);
237
- if (totalSeconds < 1) return "<1s";
238
- const seconds = totalSeconds % 60;
239
- const totalMinutes = Math.floor(totalSeconds / 60);
240
- if (totalMinutes < 1) return `${seconds}s`;
241
- const minutes = totalMinutes % 60;
242
- const hours = Math.floor(totalMinutes / 60);
243
- if (hours < 1) return seconds === 0 ? `${minutes}m` : `${minutes}m ${seconds}s`;
244
- return minutes === 0 ? `${hours}h` : `${hours}h ${minutes}m`;
245
- }
246
-
247
224
  function spinner(frame: number | undefined): string {
248
225
  return SPINNER_FRAMES.at((frame ?? 0) % SPINNER_FRAMES.length) ?? SPINNER_FRAMES[0];
249
226
  }
@@ -695,6 +672,46 @@ function toolCallRows(details: EvalToolDetails | undefined): ToolCallRow[] {
695
672
  });
696
673
  }
697
674
 
675
+ function nestedToolCallBlock(
676
+ details: EvalToolDetails | undefined,
677
+ theme: Theme | undefined,
678
+ cwd: string,
679
+ expanded: boolean,
680
+ ): RenderBlock | undefined {
681
+ const toolCalls = details?.toolCalls;
682
+ if (theme === undefined || toolCalls === undefined || !toolCalls.some((call) => call.args !== undefined)) {
683
+ return undefined;
684
+ }
685
+ const legacyRows = toolCallRows(details);
686
+ return {
687
+ kind: "dynamic",
688
+ render: (width) => {
689
+ const retainedCalls = expanded ? toolCalls : toolCalls.slice(-TOOL_CALL_PREVIEW_COUNT);
690
+ const retainedRows = expanded ? legacyRows : legacyRows.slice(-TOOL_CALL_PREVIEW_COUNT);
691
+ const skippedCount = toolCalls.length - retainedCalls.length;
692
+ const toolCallNoun = skippedCount === 1 ? "call" : "calls";
693
+ const lines =
694
+ expanded || skippedCount === 0
695
+ ? []
696
+ : renderAllVisualLines(style(theme, "muted", `${skippedCount} earlier tool ${toolCallNoun}`), width);
697
+ const legacyBlock: Extract<RenderBlock, { kind: "toolCalls" }> = {
698
+ kind: "toolCalls",
699
+ calls: retainedRows,
700
+ expanded,
701
+ theme,
702
+ };
703
+ for (const [index, call] of retainedCalls.entries()) {
704
+ if (call.args !== undefined) {
705
+ appendLines(lines, renderToolCallWidget(call, { cwd, theme, expanded, width }));
706
+ } else {
707
+ appendLines(lines, renderToolCall(retainedRows[index], legacyBlock, width));
708
+ }
709
+ }
710
+ return lines;
711
+ },
712
+ };
713
+ }
714
+
698
715
  function resultStatus(
699
716
  details: EvalToolDetails | undefined,
700
717
  options: ToolRenderResultOptions,
@@ -732,7 +749,7 @@ function resultMetadata(
732
749
  ): RenderBlock[] {
733
750
  const metadata: string[] = [];
734
751
  if (details?.phase) metadata.push(`phase ${details.phase}`);
735
- if (!options.isPartial && details) metadata.push(`took ${details.durationMs}ms`);
752
+ if (!options.isPartial && typeof details?.durationMs === "number") metadata.push(`took ${details.durationMs}ms`);
736
753
  if (metadata.length === 0) return [];
737
754
  return [{ kind: "text", text: style(theme, "muted", metadata.join(" | ")) }];
738
755
  }
@@ -824,7 +841,9 @@ export function renderEvalResult(
824
841
  },
825
842
  ];
826
843
  const calls = toolCallRows(details);
827
- if (calls.length > 0) blocks.push({ kind: "blank" }, { kind: "toolCalls", calls, expanded, theme });
844
+ const nestedCalls = nestedToolCallBlock(details, theme, context.cwd, expanded);
845
+ if (calls.length > 0)
846
+ blocks.push({ kind: "blank" }, nestedCalls ?? { kind: "toolCalls", calls, expanded, theme });
828
847
  component.setBlocks(blocks);
829
848
  return component;
830
849
  }
@@ -888,7 +907,8 @@ export function renderEvalResult(
888
907
  );
889
908
  }
890
909
  const calls = toolCallRows(details);
891
- if (calls.length > 0) blocks.push({ kind: "blank" }, { kind: "toolCalls", calls, expanded, theme });
910
+ const nestedCalls = nestedToolCallBlock(details, theme, context.cwd, expanded);
911
+ if (calls.length > 0) blocks.push({ kind: "blank" }, nestedCalls ?? { kind: "toolCalls", calls, expanded, theme });
892
912
  if (details?.notice !== undefined)
893
913
  blocks.push({ kind: "blank" }, { kind: "text", text: style(theme, "dim", details.notice) });
894
914
  const warning = formatTruncationWarning(details?.meta) ?? (details?.truncated ? "[eval output truncated]" : null);
@@ -0,0 +1,259 @@
1
+ import {
2
+ createBashToolDefinition,
3
+ createFindToolDefinition,
4
+ createGrepToolDefinition,
5
+ createLsToolDefinition,
6
+ createReadToolDefinition,
7
+ createWriteToolDefinition,
8
+ sanitizeTerminalLabel,
9
+ type Theme,
10
+ type ThemeColor,
11
+ type ToolDefinition,
12
+ truncateToVisualLines,
13
+ } from "@code-yeongyu/senpi";
14
+ import type { TSchema } from "typebox";
15
+ import { Check } from "typebox/value";
16
+ import type { EvalToolCallSummary } from "./types.ts";
17
+
18
+ export interface WidgetOptions {
19
+ readonly cwd: string;
20
+ readonly theme: Theme | undefined;
21
+ readonly expanded: boolean;
22
+ readonly width: number;
23
+ }
24
+
25
+ type AnyToolDef<TParams extends TSchema, TDetails, TState> = ToolDefinition<TParams, TDetails, TState>;
26
+ type RenderCallOf<TParams extends TSchema, TDetails, TState> = NonNullable<
27
+ AnyToolDef<TParams, TDetails, TState>["renderCall"]
28
+ >;
29
+ type ToolRenderContext<TParams extends TSchema, TDetails, TState> = Parameters<
30
+ RenderCallOf<TParams, TDetails, TState>
31
+ >[2];
32
+ type Component<TParams extends TSchema, TDetails, TState> = ReturnType<RenderCallOf<TParams, TDetails, TState>>;
33
+
34
+ type CoreWidgetName = "bash" | "read" | "write" | "grep" | "find" | "ls";
35
+ type CoreWidgetRenderer = (summary: EvalToolCallSummary, options: WidgetOptions) => string[] | undefined;
36
+ type CoreWidgets = Readonly<Record<CoreWidgetName, CoreWidgetRenderer>>;
37
+
38
+ const MIN_RENDERER_WIDTH = 20;
39
+ const MAX_RENDERER_LINES = 5;
40
+ const MAX_FALLBACK_ARGUMENT_CODE_POINTS = 120;
41
+ const MAX_COLLAPSED_ERROR_CODE_POINTS = 512;
42
+ const MAX_COLLAPSED_ERROR_LINES = 4;
43
+ const MAX_PREVIEW_LINES = 2;
44
+ const MAX_COLLAPSED_WIDGET_LINES = 8;
45
+ const TOOL_ERROR_OMISSION_MARKER = "[tool error omitted]";
46
+ const WIDGET_TRUNCATION_MARKER = "… (widget truncated)";
47
+ const TERMINAL_ESCAPE_PATTERN =
48
+ /(?:\u001B\][\s\S]*?(?:\u0007|\u001B\\|\u009C))|[\u001B\u009B][[\]()#;?]*(?:\d{1,4}(?:[;:]\d{0,4})*)?[\dA-PR-TZcf-nq-uy=><~]/g;
49
+ const TERMINAL_CONTROL_PATTERN = /[\u0000-\u0009\u000b-\u001f\u007f-\u009f]+/g;
50
+ const coreWidgetsByCwd = new Map<string, CoreWidgets>();
51
+
52
+ export function codePointPrefix(text: string, maxCodePoints: number): string {
53
+ let end = 0;
54
+ for (let count = 0; count < maxCodePoints && end < text.length; count += 1) {
55
+ const firstCodeUnit = text.charCodeAt(end);
56
+ const secondCodeUnit = text.charCodeAt(end + 1);
57
+ const isSurrogatePair =
58
+ firstCodeUnit >= 0xd800 && firstCodeUnit <= 0xdbff && secondCodeUnit >= 0xdc00 && secondCodeUnit <= 0xdfff;
59
+ end += isSurrogatePair ? 2 : 1;
60
+ }
61
+ return text.slice(0, end);
62
+ }
63
+
64
+ export function formatDuration(milliseconds: number): string {
65
+ const totalSeconds = Math.floor(Math.max(0, milliseconds) / 1_000);
66
+ if (totalSeconds < 1) return "<1s";
67
+ const seconds = totalSeconds % 60;
68
+ const totalMinutes = Math.floor(totalSeconds / 60);
69
+ if (totalMinutes < 1) return `${seconds}s`;
70
+ const minutes = totalMinutes % 60;
71
+ const hours = Math.floor(totalMinutes / 60);
72
+ if (hours < 1) return seconds === 0 ? `${minutes}m` : `${minutes}m ${seconds}s`;
73
+ return minutes === 0 ? `${hours}h` : `${hours}h ${minutes}m`;
74
+ }
75
+
76
+ function style(theme: Theme | undefined, color: ThemeColor, text: string): string {
77
+ return theme === undefined ? text : theme.fg(color, text);
78
+ }
79
+
80
+ function renderAllVisualLines(text: string, width: number): string[] {
81
+ return truncateToVisualLines(text, Number.POSITIVE_INFINITY, width).visualLines.map((line) => line.trimEnd());
82
+ }
83
+
84
+ function renderFirstVisualLines(text: string, maxLines: number, width: number): string[] {
85
+ const truncated = truncateToVisualLines(text, maxLines, width);
86
+ if (truncated.skippedCount === 0) return truncated.visualLines.map((line) => line.trimEnd());
87
+ return renderAllVisualLines(text, width).slice(0, maxLines);
88
+ }
89
+
90
+ function indent(lines: readonly string[]): string[] {
91
+ return lines.map((line) => ` ${line}`);
92
+ }
93
+
94
+ function renderWith<TParams extends TSchema, TDetails, TState>(
95
+ definition: ToolDefinition<TParams, TDetails, TState>,
96
+ summary: EvalToolCallSummary,
97
+ options: WidgetOptions,
98
+ initialState: TState,
99
+ ): string[] | undefined {
100
+ let preparedArgs: unknown = summary.args;
101
+ try {
102
+ if (definition.prepareArguments !== undefined) preparedArgs = definition.prepareArguments(preparedArgs);
103
+ if (!Check(definition.parameters, preparedArgs)) return undefined;
104
+ } catch {
105
+ return undefined;
106
+ }
107
+
108
+ const theme = options.theme;
109
+ const innerWidth = options.width - 2;
110
+ const renderCall = definition.renderCall;
111
+ if (theme === undefined || innerWidth < MIN_RENDERER_WIDTH || renderCall === undefined) return undefined;
112
+
113
+ const context: ToolRenderContext<TParams, TDetails, TState> = {
114
+ args: preparedArgs,
115
+ toolCallId: summary.callId ?? summary.name,
116
+ invalidate: () => {},
117
+ lastComponent: undefined,
118
+ state: initialState,
119
+ cwd: options.cwd,
120
+ executionStarted: true,
121
+ argsComplete: true,
122
+ isPartial: false,
123
+ expanded: options.expanded,
124
+ showImages: false,
125
+ isError: !summary.ok,
126
+ hasResult: false,
127
+ };
128
+
129
+ try {
130
+ const component: Component<TParams, TDetails, TState> = renderCall(preparedArgs, theme, context);
131
+ const renderedLines = component.render(innerWidth).map((line) => line.trimEnd());
132
+ const retainedLines = renderedLines.slice(0, MAX_RENDERER_LINES);
133
+ if (renderedLines.length > retainedLines.length) {
134
+ retainedLines.push(style(theme, "muted", WIDGET_TRUNCATION_MARKER));
135
+ }
136
+ return indent(retainedLines);
137
+ } catch {
138
+ return undefined;
139
+ }
140
+ }
141
+
142
+ function createCoreWidgets(cwd: string): CoreWidgets {
143
+ const bashDefinition = createBashToolDefinition(cwd);
144
+ const readDefinition = createReadToolDefinition(cwd);
145
+ const writeDefinition = createWriteToolDefinition(cwd);
146
+ const grepDefinition = createGrepToolDefinition(cwd);
147
+ const findDefinition = createFindToolDefinition(cwd);
148
+ const lsDefinition = createLsToolDefinition(cwd);
149
+ const widgets: CoreWidgets = {
150
+ bash: (summary, options) =>
151
+ renderWith(bashDefinition, summary, options, {
152
+ startedAt: undefined,
153
+ endedAt: undefined,
154
+ interval: undefined,
155
+ }),
156
+ read: (summary, options) => renderWith(readDefinition, summary, options, {}),
157
+ write: (summary, options) => renderWith(writeDefinition, summary, options, {}),
158
+ grep: (summary, options) => renderWith(grepDefinition, summary, options, {}),
159
+ find: (summary, options) => renderWith(findDefinition, summary, options, {}),
160
+ ls: (summary, options) => renderWith(lsDefinition, summary, options, {}),
161
+ };
162
+ return widgets;
163
+ }
164
+
165
+ function coreWidgetsFor(cwd: string): CoreWidgets {
166
+ const cached = coreWidgetsByCwd.get(cwd);
167
+ if (cached !== undefined) return cached;
168
+ const widgets = createCoreWidgets(cwd);
169
+ coreWidgetsByCwd.set(cwd, widgets);
170
+ return widgets;
171
+ }
172
+
173
+ function isCoreWidgetName(name: string): name is CoreWidgetName {
174
+ switch (name) {
175
+ case "bash":
176
+ case "read":
177
+ case "write":
178
+ case "grep":
179
+ case "find":
180
+ case "ls":
181
+ return true;
182
+ default:
183
+ return false;
184
+ }
185
+ }
186
+
187
+ function compactArguments(args: unknown): string {
188
+ if (args === undefined) return "";
189
+ try {
190
+ const serialized = JSON.stringify(args);
191
+ if (serialized === undefined) return "";
192
+ return codePointPrefix(sanitizeTerminalLabel(serialized), MAX_FALLBACK_ARGUMENT_CODE_POINTS);
193
+ } catch {
194
+ return "[unserializable args]";
195
+ }
196
+ }
197
+
198
+ function renderFallback(summary: EvalToolCallSummary, width: number): string[] {
199
+ const name = sanitizeTerminalLabel(summary.name);
200
+ const fallback = `tool.${name}(${compactArguments(summary.args)})`;
201
+ return indent(renderAllVisualLines(fallback, width));
202
+ }
203
+
204
+ function renderStatus(summary: EvalToolCallSummary, options: WidgetOptions, width: number): string[] {
205
+ const icon = style(options.theme, summary.ok ? "success" : "error", summary.ok ? "✓" : "✗");
206
+ const parts = [icon];
207
+ if (summary.durationMs !== undefined) parts.push(formatDuration(summary.durationMs));
208
+ if (summary.argsTruncated) parts.push("(args truncated)");
209
+ return indent(renderFirstVisualLines(parts.join(" "), 1, width));
210
+ }
211
+
212
+ function sanitizeTerminalControl(text: string): string {
213
+ return text.replace(TERMINAL_ESCAPE_PATTERN, "").replace(TERMINAL_CONTROL_PATTERN, " ");
214
+ }
215
+
216
+ function renderPreview(preview: string, options: WidgetOptions, width: number): string[] {
217
+ const sanitized = sanitizeTerminalControl(preview);
218
+ if (sanitized.length === 0) return [];
219
+ return indent(renderFirstVisualLines(style(options.theme, "muted", sanitized), MAX_PREVIEW_LINES, width));
220
+ }
221
+
222
+ function renderCollapsedError(error: string, options: WidgetOptions, width: number): string[] {
223
+ const guardedError = codePointPrefix(error, MAX_COLLAPSED_ERROR_CODE_POINTS);
224
+ const guardedLines = renderAllVisualLines(style(options.theme, "error", guardedError), width);
225
+ if (guardedError.length === error.length && guardedLines.length <= MAX_COLLAPSED_ERROR_LINES) {
226
+ return indent(guardedLines);
227
+ }
228
+
229
+ const markerLines = renderAllVisualLines(style(options.theme, "muted", TOOL_ERROR_OMISSION_MARKER), width);
230
+ const errorBudget = Math.max(0, MAX_COLLAPSED_ERROR_LINES - markerLines.length);
231
+ const lines = guardedLines.slice(0, errorBudget);
232
+ lines.push(...markerLines.slice(0, MAX_COLLAPSED_ERROR_LINES - lines.length));
233
+ return indent(lines);
234
+ }
235
+
236
+ function renderError(error: string, options: WidgetOptions, width: number): string[] {
237
+ const sanitized = sanitizeTerminalControl(error);
238
+ if (options.expanded) return indent(renderAllVisualLines(style(options.theme, "error", sanitized), width));
239
+ return renderCollapsedError(sanitized, options, width);
240
+ }
241
+
242
+ /**
243
+ * The theme-less fallback is QA-only. Production's gate lives in render.ts:
244
+ * nestedToolCallBlock routes theme-less renders to legacy rows before this function is called.
245
+ */
246
+ export function renderToolCallWidget(summary: EvalToolCallSummary, options: WidgetOptions): string[] {
247
+ const innerWidth = Math.max(1, options.width - 2);
248
+ const coreLines = isCoreWidgetName(summary.name)
249
+ ? coreWidgetsFor(options.cwd)[summary.name](summary, options)
250
+ : undefined;
251
+ const lines = coreLines ?? renderFallback(summary, innerWidth);
252
+ lines.push(...renderStatus(summary, options, innerWidth));
253
+ if (summary.ok && summary.resultPreview !== undefined) {
254
+ lines.push(...renderPreview(summary.resultPreview, options, innerWidth));
255
+ } else if (!summary.ok && summary.error !== undefined) {
256
+ lines.push(...renderError(summary.error, options, innerWidth));
257
+ }
258
+ return options.expanded ? lines : lines.slice(0, MAX_COLLAPSED_WIDGET_LINES);
259
+ }
package/src/tool/types.ts CHANGED
@@ -110,13 +110,18 @@ export interface EvalKernelManager {
110
110
  export type ExecuteTool = (
111
111
  toolName: string,
112
112
  params: unknown,
113
- options?: { signal?: AbortSignal; onUpdate?: AgentToolUpdateCallback<unknown> },
113
+ options?: { signal?: AbortSignal; onUpdate?: AgentToolUpdateCallback<unknown>; activateInactiveTool?: boolean },
114
114
  ) => Promise<AgentToolResult<unknown>>;
115
115
 
116
116
  export interface EvalToolCallSummary {
117
117
  readonly name: string;
118
118
  readonly ok: boolean;
119
119
  readonly error?: string;
120
+ readonly callId?: string;
121
+ readonly args?: unknown;
122
+ readonly argsTruncated?: boolean;
123
+ readonly durationMs?: number;
124
+ readonly resultPreview?: string;
120
125
  }
121
126
 
122
127
  export type EvalStatusEvent = { readonly op: string } & Readonly<Record<string, unknown>>;