@code-yeongyu/senpi 2026.9.3-3 → 2026.9.4-2

package/node_modules/@code-yeongyu/senpi-codemode/src/prompt/eval-prompt.ts

@@ -64,91 +64,48 @@ export function evalEmphasisStyle(modelId: string | undefined): EvalEmphasisStyl
 
 type ContextValue = string | boolean;
 type Context = Readonly<Record<string, ContextValue>>;
-type EvalPromptExample = {
-	readonly caption: string;
-	readonly language: keyof EnabledLanguages;
-	readonly summary: string;
-	readonly code: string;
-};
-
-// senpi ToolDefinition has no examples field, so description embeds the examples.
-// ADAPTATION: payloads diverge from omp's json-config chain to teach batch read,
-// comprehension filtering, and parallel tool.<name> fan-out while keeping the
-// three-cell reuse narrative.
-const REUSE_CHAIN_EXAMPLES = [
-	{
-		caption: "First call — set up once",
-		language: "js",
-		summary: "Count all TypeScript source files under src/ excluding tests",
-		code: "import { readdir } from 'node:fs/promises'\nimport { extname } from 'node:path'\nconst files = (await readdir('src', { recursive: true })).filter(f => extname(f) === '.ts' && !f.includes('test'))\nprint(files.length)",
-	},
-	{
-		caption: "Second call — reuse `files`, fan out session tools in parallel",
-		language: "js",
-		summary: "Grep legacyClient per directory in one cell",
-		code: "const dirs = [...new Set(files.map(f => f.split('/')[0]))]\nconst hits = await Promise.all(dirs.map(d => tool.grep({ pattern: 'legacyClient', path: d })))\ndisplay(hits.map(h => h.matches?.length ?? 0))",
-	},
-	{
-		caption: "JS kernel is busy with a detached cell — continue in py",
-		language: "py",
-		summary: "Aggregate legacyClient hits while JS is busy",
-		code: "from pathlib import Path\nprint(sum('legacyClient' in read(p) for p in Path('src').rglob('*.ts')))",
-	},
-] as const satisfies readonly EvalPromptExample[];
-
 const EVAL_PROMPT_TEMPLATE = `Run one step of code in a persistent kernel.
 
 <instruction>
-**One eval call = one cell = one logical step.** State persists per language across separate eval calls and tool calls{{#if spawns}}, and \`task\` subagents{{/if}} — define helpers, datasets, and clients in one call, then later calls reuse them directly.
-
-Work incrementally: imports in one call, define in the next, test, then use — each its own eval call. Re-run setup ONLY after \`reset\`, a kernel crash, or a \`NameError\`/\`ReferenceError\` proving the state is gone.
+**One eval call = one cell = one logical step.** Top-level names persist per language across eval calls{{#if spawns}}, tool calls and \`task\` subagents{{else}} and tool calls{{/if}}: define helpers and clients once and reuse them instead of re-importing or re-reading. Rebuild state only after \`reset\`, a kernel restart, or a \`NameError\`/\`ReferenceError\`, and check a sentinel variable first so a re-run cannot duplicate side effects.
 
 {{#if styleClaude}}<eval_first_batching>
 \`eval\` is your default execution surface: if a step needs more than one tool call, write ONE cell that performs the whole step — never issue the calls one at a time.
 - Enumerate every lookup the step needs, then run all independent ones simultaneously with \`parallel(thunks)\` inside the cell; keep calls sequential only when one result feeds the next.
 - Write real code around the calls: loop or comprehend over file sets with \`read()\`/stdlib, branch per case, and wrap risky calls in try/except so one failure degrades only its item — recover or retry inside the cell, keep the batch alive.
 - Post-process \`tool.<name>()\` results programmatically — filter, join, aggregate — and return distilled facts, not raw dumps.
-{{#if monitor}}- Start long-running work (build, test run, deploy, or watch) through \`tool.monitor({ command, filter })\`, putting the decisive-line filter inside the same cell, then keep working until its event wakes the turn; sleeping, timed retries, and repeated status reads are not waiting.{{/if}}
+{{#if monitor}}- Start long-running work (build, test run, deploy, or watch) through \`tool.monitor({ command, filter })\`, putting the decisive-line filter inside the same cell, then keep working until its event wakes the turn.{{/if}}
 </eval_first_batching>{{/if}}{{#if styleGpt}}<gpt_eval_dialect>
 GPT eval: compose multi-tool work inside one cell with \`tool.<name>(args)\` and \`parallel(thunks)\`; do not split a planned step into serial tool calls.
-- Long pure-compute cells detach on timeout and notify on completion. Do not poll or re-run them; use \`eval({ action: "peek"|"stop", cell_id })\` only to inspect or stop a detached cell.
+- Long cells detach on timeout and notify on completion; do not poll or re-run them.
 - Filter, join, and aggregate tool results in the cell; return only decision-relevant facts.
-{{#if monitor}}- For long-running build, test run, deploy, or watch work, start \`tool.monitor({ command, filter })\` with the decisive-line filter in the same cell; keep working while its event wakes the turn; sleeping, timed retries, and repeated status reads are not waiting.{{/if}}
+{{#if monitor}}- For long-running build, test run, deploy, or watch work, start \`tool.monitor({ command, filter })\` with the decisive-line filter in the same cell; keep working until its event wakes the turn.{{/if}}
 </gpt_eval_dialect>{{/if}}{{#if styleCodex}}Route multi-call steps through eval: one cell per step, independent lookups dispatched together via \`parallel(thunks)\`; keep work sequential only when one result determines the next action.
 - Loop or comprehend over file sets with \`read()\`/stdlib instead of reading files one call at a time; post-process \`tool.<name>()\` results programmatically — filter, join, aggregate.
 - Wrap failable calls in try/except inside the cell; a failed item degrades only itself. After two distinct failed strategies for the same fact, fall back to direct tool calls.
 - Reduce large results in-kernel to the facts the task needs before returning.
-{{#if monitor}}- Long-running build/test/deploy/watch work: start \`tool.monitor({ command, filter })\` with the decisive-line filter inside the same cell, then continue working; its event wakes the turn — sleeping, timed retries, and repeated status reads are not waiting.{{/if}}{{/if}}{{#if styleKimi}}**EVAL IS YOUR SUPERPOWER — MAKE IT YOUR DEFAULT WAY TO ACT.** Before any step, think: "how do I execute this WHOLE step in ONE parallelized cell?" — then write that ONE cell.
+{{#if monitor}}- Long-running build/test/deploy/watch work: start \`tool.monitor({ command, filter })\` with the decisive-line filter inside the same cell, then continue working until its event wakes the turn.{{/if}}{{/if}}{{#if styleKimi}}**EVAL IS YOUR SUPERPOWER — MAKE IT YOUR DEFAULT WAY TO ACT.** Before any step, think: "how do I execute this WHOLE step in ONE parallelized cell?" — then write that ONE cell.
 - **BATCH EVERYTHING AT ONCE:** enumerate EVERY independent lookup the step needs and dispatch them ALL simultaneously with \`parallel(thunks)\` in that cell; keep calls sequential only when one result feeds the next.
 - **WRITE REAL CODE, NOT CALL CHAINS:** loop or comprehend over file sets with \`read()\`/stdlib, post-process \`tool.<name>()\` results programmatically, and put try/except around each risky call so the rest of the batch completes.
 - **DISTILL IN-KERNEL:** filter, join, and aggregate \`tool.<name>()\` results in code, then return ONLY the distilled facts.
-{{#if monitor}}- **DO start long-running build, test run, deploy, or watch work with \`tool.monitor({ command, filter })\`, put the decisive-line filter INSIDE THE SAME CELL, and KEEP WORKING until its event wakes the turn; sleeping, timed retries, and repeated status reads are not waiting.**{{/if}}{{/if}}{{#if styleDefault}}**EVAL IS YOUR PRIMARY EXECUTION SURFACE.** Any step that needs MORE THAN ONE tool call MUST be written as ONE cell — NEVER as a chain of single tool calls.
+{{#if monitor}}- **DO start long-running build, test run, deploy, or watch work with \`tool.monitor({ command, filter })\`, put the decisive-line filter INSIDE THE SAME CELL, and KEEP WORKING until its event wakes the turn.**{{/if}}{{/if}}{{#if styleDefault}}**EVAL IS YOUR PRIMARY EXECUTION SURFACE.** Any step that needs MORE THAN ONE tool call MUST be written as ONE cell — NEVER as a chain of single tool calls.
 - **PLAN THE WHOLE STEP, THEN BATCH IT.** Enumerate every read/search/lookup the step needs and dispatch ALL independent ones through \`parallel(thunks)\` in one cell.
 - **WRITE REAL CODE, NOT CALL LISTS.** Loop or comprehend over file sets with \`read()\`/stdlib, branch \`if\`/\`else\` per case, post-process \`tool.<name>()\` results programmatically, and wrap EVERY risky call in try/except so ONE failure NEVER kills the batch.
 - **DISTILL IN-KERNEL.** Filter, join, diff, and aggregate in code before returning; return facts, NOT dumps.
-{{#if monitor}}- **LONG-RUNNING build, test run, deploy, or watch work MUST start with \`tool.monitor({ command, filter })\`, with the decisive-line filter INSIDE THE SAME CELL; KEEP WORKING until its event wakes the turn — SLEEPING, TIMED RETRIES, AND REPEATED STATUS READS ARE NOT WAITING.**{{/if}}{{/if}}
+{{#if monitor}}- **LONG-RUNNING build, test run, deploy, or watch work MUST start with \`tool.monitor({ command, filter })\`, with the decisive-line filter INSIDE THE SAME CELL; KEEP WORKING until its event wakes the turn.**{{/if}}{{/if}}
 {{#if hostLine}}
 Host: {{hostLine}} — cells execute here. Size \`parallel(thunks)\` pools to its cores; \`tool.<name>()\` shell commands must fit this platform, even when the code you are writing targets another machine.
 {{/if}}
 
-Fields:
-
-- \`language\` — {{#if py}}\`"py"\` IPython kernel{{/if}}{{#ifAll py js}}, {{/ifAll}}{{#if js}}\`"js"\` persistent JavaScript VM{{/if}}{{#if rb}}{{#ifAny py js}}, {{/ifAny}}\`"rb"\` persistent Ruby kernel{{/if}}{{#if jl}}{{#ifAny py js rb}}, {{/ifAny}}\`"jl"\` persistent Julia kernel{{/if}}.
-- \`code\` — cell body, verbatim. Newlines/quotes JSON-encoded; no fences, no headers.
-- \`summary\` (REQUIRED for run) — ONE line in the USER'S conversational language stating WHAT this cell does and FOR WHAT PURPOSE (e.g. Korean conversation -> "src 전체에서 legacyClient 사용처 집계"); shown in the TUI while the cell runs; >80 chars is force-truncated.
-- \`timeout\` (optional) — seconds. Raise only for heavy compute or long{{#if spawns}} non-agent{{/if}} tool calls.
-- \`on_timeout\` (optional) — \`"detach"\` keeps pure computation running in interactive sessions (the default); \`"error"\` interrupts for deadline-sensitive work and is the print/json default.
-- Every cell is killed at a wall-clock hard limit (default 1800s) that survives detach and is never paused by tool calls; a larger explicit \`timeout\` raises it, and a killed cell notifies you that it hit the limit.
-- \`reset\` (optional) — wipe this language's kernel first.{{#ifAll py js}} Per-language: a \`py\` reset never touches the JS VM.{{/ifAll}}
-- \`action\` (optional) — defaults to \`"run"\`. A detached cell returns its id: use \`eval({ action: "peek", cell_id })\` for buffered output/state or \`eval({ action: "stop", cell_id })\` to cancel it.
+\`language\`: {{#if py}}\`"py"\` IPython kernel{{/if}}{{#ifAll py js}}, {{/ifAll}}{{#if js}}\`"js"\` persistent JavaScript VM{{/if}}{{#if rb}}{{#ifAny py js}}, {{/ifAny}}\`"rb"\` persistent Ruby kernel{{/if}}{{#if jl}}{{#ifAny py js rb}}, {{/ifAny}}\`"jl"\` persistent Julia kernel{{/if}}.
 
-A detached cell keeps its language kernel busy while it finishes; another language can continue. Do not re-run a detached cell: the same-language busy error names its cell id and output tail. Completion arrives as one notification with the final value/error and buffered output. Stopping a cell interrupts its kernel; the stop result states whether kernel state survived or the kernel was restarted and its variables lost.
+A cell that outlives the foreground window detaches: it keeps its language kernel busy (another language can continue) and completes as one notification with its value or error and buffered output. Do not re-run a detached cell; read or cancel it with \`eval({ action: "peek", cell_id })\` / \`eval({ action: "stop", cell_id })\`.
 
-{{#if py}}Live event loop: use top-level \`await\` directly; \`asyncio.run(…)\` raises "cannot be called from a running event loop".{{/if}}
+{{#if py}}Python runs on a live event loop: use top-level \`await\`; \`asyncio.run(…)\` raises.{{/if}}
 {{#if js}}{{#if jsBun}}JS runs in-process on Bun {{jsVersion}}: top-level \`await\`/\`return\` work; \`Bun.*\` builtins available.{{#if bunSkillPath}} MUST READ the bun-1-4 skill at {{bunSkillPath}} before your first js cell — its builtins replace the npm packages you would otherwise install.{{/if}}{{else}}JS runs under Node.js worker: top-level \`await\`/\`return\` work; \`fetch\`/\`Buffer\` available.{{/if}}{{/if}}
 {{#if rb}}Ruby: synchronous; helper options are keyword args{{#if spawns}} (e.g. \`output("id", limit: 2)\`){{/if}}; the last expression auto-displays unless it is \`nil\`, an assignment, or a definition (like IRB).{{/if}}
 {{#if jl}}Julia: synchronous; helper options are standard keyword args{{#if spawns}} (e.g. \`output("id", limit=2)\`){{/if}}; the last expression auto-displays unless it is an assignment or a definition (like the Julia REPL).{{/if}}
-On error, fix and re-run only the failing step. State usually survives a normal error, but a timeout or stop may have restarted the kernel — its message says which. Before rebuilding state, check a sentinel (a variable you defined earlier); only re-establish what is actually gone, since blind re-runs duplicate side effects.
+On error, fix and re-run only the failing step; a normal error keeps state, while a timeout or stop message says whether the kernel restarted.
 </instruction>
 
 <prelude>
@@ -163,25 +120,21 @@ read(path, offset?=1, limit?=None) → str
 write(path, content) → str
     Write file (creates parents) → resolved path. \`local://…\` persists across turns/subagents.
 env(key?=None, value?=None) → str | None | dict
-    No args → full env dict; one → value of \`key\`; two → set \`key=value\`, return value.
+    No args → full env dict; one → value; two → set \`key=value\`.
 {{#if spawns}}output(*ids, format?="raw", offset?=None, limit?=None) → str | dict | list[dict]
-    Task/agent output by id. Reads immediately: running tasks return their status; \`format\` selects full (\`"raw"\`) or trailing (\`"tail"\`) output.
+    Task/agent output by id. Reads immediately: running tasks return their status; \`format\` \`"raw"\` = full, \`"tail"\` = trailing.
 {{/if}}tool.<name>(args) → unknown
     Invoke any session tool; \`args\` = its parameter object.
 tool_schema(name?) → dict
-    Parameter schema of a tool without calling it; omit \`name\` to list tool names.
-    Use it before calling a tool you have not called before — a failed call also
-    returns the expected parameters, so fix the args and retry in the next cell
-    instead of abandoning eval.
+    Parameter schema of a tool (omit \`name\` to list tool names); a failed \`tool.<name>()\` call also returns the expected parameters.
 completion(prompt, model?="default", system?=None, schema?=None) → str | dict
-    Oneshot, stateless (no history/tools). \`model\`: \`"smol"\` fast | \`"default"\` session | \`"slow"\` most capable. \`schema\` (JSON-Schema) → structured output, parsed object.
+    Oneshot, stateless. \`model\`: \`"smol"\` fast | \`"default"\` session | \`"slow"\` most capable. \`schema\` (JSON-Schema) → parsed structured output.
 {{#if spawns}}agent(prompt, agent?="{{spawnDefaultAgent}}", model?=None, label?=None, schema?=None, handle?=False) → str | dict
-    Run a subagent → final output. \`agent\` picks another discovered agent; omit it to use \`{{spawnDefaultAgent}}\`. \`schema\` as in completion(). Background via \`local://\` files named in the prompt. \`handle\` → workflow node dict { text, output, handle: \`agent://<id>\`, id, agent } (parsed under \`data\` when \`schema\` set).
-{{#if js}}    JS: options are ONE trailing object — agent(prompt, { agent, schema, handle }).
-{{/if}}{{/if}}parallel(thunks) → list
-    Thunks through a bounded pool (wide as a \`task\` batch — don't pre-shrink), input order kept; returns when all finish, a throwing thunk propagates.
+    Run a subagent → final output. \`agent\` picks a discovered agent. \`schema\` as in completion(). \`handle\` → workflow node { text, output, handle: \`agent://<id>\`, id, agent } (parsed under \`data\` with \`schema\`).
+{{/if}}parallel(thunks) → list
+    Thunks through a bounded pool (as wide as a \`task\` batch), input order kept; a throwing thunk propagates.
 pipeline(items, ...stages) → list
-    Map items through one-arg stages left-to-right, barrier between stages; stage 1 gets the item, later stages the previous result.
+    Map items through one-arg stages with a barrier between stages; each stage receives the previous stage's result.
 log(message) → None
     Progress line above the status tree.
 phase(title) → None
@@ -190,18 +143,10 @@ phase(title) → None
 </prelude>
 {{#if spawns}}
 <workflow>
-Define the workflow spec IN CODE: partition the work into logically distinct steps, one node per step, then wire them as acyclic waves — never hand-author the graph as a single opaque call.
-- **Name nodes.** Capture each \`agent(…, {{#if py}}handle=True{{/if}}{{#if js}}{ handle: true }{{/if}}{{#if jl}}handle=true{{/if}})\` result; carries \`handle\` (\`agent://<id>\`) + \`output\`.
-- **Wire edges by reference.** Put an upstream node's \`handle\`/\`output\` in the dependent stage's prompt — large transcript never re-inlined. Bulk: \`write("local://<name>.md", …)\`, pass the URI.
-- **\`pipeline(items, *stages)\` = staged waves**, barrier between stages (every item clears stage N before any enters N+1). **\`parallel(thunks)\` = one wave** of independent nodes.
-- **Isolate failure.** A raising node re-raises the lowest-index error, aborts its wave; wrap risky nodes in try/except so a failure degrades only its dependent subtree, independent branches finish.
-- **Acyclic only.** A node never waits on its own descendant.
+Multi-agent work is an acyclic graph in code: one \`agent(…)\` node per step with its handle option ({{#if py}}\`handle=True\`{{/if}}{{#ifAll py js}} / {{/ifAll}}{{#if js}}\`{ handle: true }\`{{/if}}{{#if jl}}{{#ifAny py js}} / {{/ifAny}}\`handle=true\`{{/if}}), \`parallel(thunks)\` for independent nodes, \`pipeline(items, *stages)\` for staged waves. Pass an upstream node's \`handle\` or \`output\` (or a \`write("local://…")\` URI for bulk text) into dependents instead of re-inlining transcripts, and wrap risky nodes in try/except so a failure aborts only its subtree.
 </workflow>
 {{/if}}
-
-<critical>
-Prior top-level names (\`data\`, \`sessions\`, helpers, imports) survive into the next eval call — reuse them; NEVER re-import, re-require, or re-declare a helper. Re-read a file only if it may have changed since the last read.
-</critical>`;
+`;
 
 export function buildEvalPrompt(
 	enabled: EnabledLanguages,
@@ -230,20 +175,9 @@ export function buildEvalPrompt(
 		jsVersion: options.jsRuntime?.version ?? "",
 		bunSkillPath: options.bunSkillPath ?? "",
 	};
-	const examples = REUSE_CHAIN_EXAMPLES.filter((example) => enabled[example.language])
-		.map((example) => {
-			const call = { language: example.language, summary: example.summary, code: example.code };
-			return `### ${example.caption}\n\`\`\`json\n${JSON.stringify(call, null, 2)}\n\`\`\``;
-		})
-		.join("\n\n");
-	const description = [
-		renderTemplate(EVAL_PROMPT_TEMPLATE, context)
-			.replace(/\n{3,}/g, "\n\n")
-			.trim(),
-		examples === "" ? "" : `<examples>\n${examples}\n</examples>`,
-	]
-		.filter((part) => part !== "")
-		.join("\n\n");
+	const description = renderTemplate(EVAL_PROMPT_TEMPLATE, context)
+		.replace(/\n{3,}/g, "\n\n")
+		.trim();
 	return {
 		description,
 		promptSnippet: "Run one incremental code cell in a persistent language kernel.",

package/node_modules/@code-yeongyu/senpi-codemode/src/tool/cell-execution.ts

@@ -17,6 +17,12 @@ export interface CellExecutionOptions {
 	readonly callerSignal: AbortSignal;
 	readonly cellId: string;
 	readonly timeoutMs: number;
+	/**
+	 * Caps how long a host-bridge pause may suspend the idle watchdog. When the cell will detach on
+	 * timeout this is set to the foreground window so a bridge-parked cell still frees the turn at the
+	 * window; left undefined (error mode) it keeps the idle-timeout default grace.
+	 */
+	readonly maxPauseGraceMs?: number;
 	readonly timeoutFactory: EvalTimeoutFactory;
 	readonly onTimeout: (error: Error) => void;
 	readonly onAbort: (error: Error) => void;
@@ -46,6 +52,7 @@ export class CellExecution {
 		this.#watchdog = options.timeoutFactory.create({
 			cellId: options.cellId,
 			timeoutMs: options.timeoutMs,
+			...(options.maxPauseGraceMs === undefined ? {} : { maxPauseGraceMs: options.maxPauseGraceMs }),
 			onTimeout: ({ error }) => options.onTimeout(error),
 		});
 		this.#callerSignal.addEventListener("abort", this.#handleCallerAbort, {

package/node_modules/@code-yeongyu/senpi-codemode/src/tool/eval-tool-options.ts

@@ -21,6 +21,12 @@ export interface CreateEvalToolOptions {
 	readonly enabledLanguages: EnabledEvalLanguages;
 	readonly kernelManager: EvalKernelManager;
 	readonly cellTimeoutSeconds: number;
+	/**
+	 * Longest an interactive (detach-behavior) call blocks the agent loop before the cell detaches,
+	 * capping the `timeout` detach budget. Defaults to {@link DEFAULT_FOREGROUND_WINDOW_SECONDS}.
+	 * Does not affect `on_timeout: "error"` calls or the wall-clock hard limit.
+	 */
+	readonly foregroundWindowSeconds?: number;
 	/** Wall-clock kill deadline applied to every cell; only used when this factory creates its own manager. */
 	readonly hardLimitSeconds?: number;
 	readonly executeTool: ExecuteTool;

package/node_modules/@code-yeongyu/senpi-codemode/src/tool/eval-tool.ts

@@ -1,7 +1,7 @@
 import { randomUUID } from "node:crypto";
 import { join } from "node:path";
 import type { AgentToolResult, ExtensionContext, ToolDefinition } from "@code-yeongyu/senpi";
-import { defaultCodemodeSettings } from "../config/settings.ts";
+import { DEFAULT_FOREGROUND_WINDOW_SECONDS, defaultCodemodeSettings } from "../config/settings.ts";
 import { buildEvalPrompt } from "../prompt/eval-prompt.ts";
 import { TIMEOUT_PAUSE_OP, TIMEOUT_RESUME_OP } from "../timeouts/bridge-timeout.ts";
 import { abortError, CellExecution, defaultTimeoutFactory } from "./cell-execution.ts";
@@ -101,8 +101,19 @@ async function runEvalCell(
 	invocation: EvalCellInvocation,
 ): Promise<AgentToolResult<EvalToolDetails>> {
 	if (invocation.signal.aborted) throw abortError(invocation.signal.reason);
-	const timeoutMs = Math.floor((invocation.input.timeout ?? options.cellTimeoutSeconds) * 1_000);
 	const timeoutBehavior = evalTimeoutBehavior(invocation.input, invocation.ctx);
+	const requestedTimeoutMs = Math.floor((invocation.input.timeout ?? options.cellTimeoutSeconds) * 1_000);
+	// The `timeout` (and its `cellTimeoutSeconds` default) is the detach budget for interactive calls.
+	// Cap it at the foreground window so a large `timeout` — whose real purpose is to raise the
+	// wall-clock hard limit (see EvalDetachedCellManager) — frees the turn at the window instead of
+	// blocking the agent loop for its full duration. `on_timeout: "error"` (and print/json) keep the
+	// unclamped deadline, since there the cell is killed rather than detached.
+	const foregroundWindowMs = (options.foregroundWindowSeconds ?? DEFAULT_FOREGROUND_WINDOW_SECONDS) * 1_000;
+	const timeoutMs =
+		timeoutBehavior === "detach" ? Math.min(requestedTimeoutMs, foregroundWindowMs) : requestedTimeoutMs;
+	// A cell that pauses its watchdog for a host bridge call would otherwise wait the full pause grace
+	// (~10 min) before detaching; cap the grace at the foreground window too so the detach guarantee
+	// holds for bridge-parked cells. Error mode keeps the default grace (its timeout is the deadline).
 	const bridgeAbortController = new AbortController();
 	const cellSignal = AbortSignal.any([invocation.signal, bridgeAbortController.signal]);
 	const bridgeContext: ExtensionContext = { ...invocation.ctx, signal: cellSignal };
@@ -131,6 +142,7 @@ async function runEvalCell(
 		callerSignal: invocation.signal,
 		cellId: invocation.cellId,
 		timeoutMs,
+		...(timeoutBehavior === "detach" ? { maxPauseGraceMs: foregroundWindowMs } : {}),
 		timeoutFactory: options.timeoutFactory ?? defaultTimeoutFactory,
 		onTimeout: (error) => {
 			if (timeoutBehavior === "detach" && cellManager.detach(cell)) {

package/node_modules/@code-yeongyu/senpi-codemode/src/tool/types.ts

@@ -13,6 +13,12 @@ export function enabledLanguageList(enabled: EnabledEvalLanguages): EvalLanguage
 
 export const EVAL_SUMMARY_MAX_LENGTH = 80;
 
+const TIMEOUT_FIELD_DESCRIPTION =
+	"Seconds the cell may block the turn before it detaches, and the amount by which it raises the wall-clock hard limit. In interactive sessions the detach point is capped at the foreground window (default 60s), so a large value frees the turn at the window while the cell keeps running; on_timeout:'error' (and print/json) keep the full value as the uncapped deadline.";
+
+const ON_TIMEOUT_FIELD_DESCRIPTION =
+	"Timeout behavior. Interactive sessions detach by default (at the foreground window); print/json sessions error by default. 'error' uses the full timeout as an uncapped deadline.";
+
 export interface EvalToolInput {
 	readonly language: EvalLanguage;
 	readonly code: string;
@@ -47,10 +53,10 @@ const fullEvalInputSchema = Type.Object({
 				"REQUIRED for run. ONE line in the USER'S conversational language (Korean conversation -> Korean summary) stating WHAT this cell does and FOR WHAT PURPOSE; shown in the TUI while the cell runs. Longer values are force-truncated to 80 chars.",
 		}),
 	),
-	timeout: Type.Optional(Type.Number({ minimum: 1, description: "Timeout in seconds." })),
+	timeout: Type.Optional(Type.Number({ minimum: 1, description: TIMEOUT_FIELD_DESCRIPTION })),
 	on_timeout: Type.Optional(
 		Type.Union([Type.Literal("detach"), Type.Literal("error")], {
-			description: "Timeout behavior. Interactive sessions detach by default; print/json sessions error by default.",
+			description: ON_TIMEOUT_FIELD_DESCRIPTION,
 		}),
 	),
 	reset: Type.Optional(Type.Boolean({ description: "Reset this language kernel before running." })),
@@ -83,11 +89,10 @@ export function createEvalInputSchema(enabled: EnabledEvalLanguages): EvalInputS
 						"REQUIRED for run. ONE line in the USER'S conversational language (Korean conversation -> Korean summary) stating WHAT this cell does and FOR WHAT PURPOSE; shown in the TUI while the cell runs. Longer values are force-truncated to 80 chars.",
 				}),
 			),
-			timeout: Type.Optional(Type.Number({ minimum: 1, description: "Timeout in seconds." })),
+			timeout: Type.Optional(Type.Number({ minimum: 1, description: TIMEOUT_FIELD_DESCRIPTION })),
 			on_timeout: Type.Optional(
 				Type.Union([Type.Literal("detach"), Type.Literal("error")], {
-					description:
-						"Timeout behavior. Interactive sessions detach by default; print/json sessions error by default.",
+					description: ON_TIMEOUT_FIELD_DESCRIPTION,
 				}),
 			),
 			reset: Type.Optional(Type.Boolean({ description: "Reset this language kernel before running." })),

package/node_modules/@earendil-works/pi-agent-core/dist/agent-loop.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent-loop.d.ts","sourceRoot":"","sources":["../src/agent-loop.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EAIN,KAAK,OAAO,EAEZ,WAAW,EAIX,MAAM,uBAAuB,CAAC;AAU/B,OAAO,KAAK,EACX,YAAY,EACZ,UAAU,EACV,eAAe,EAEf,YAAY,EACZ,SAAS,EACT,aAAa,EAEb,QAAQ,EACR,MAAM,YAAY,CAAC;AAEpB,MAAM,MAAM,cAAc,GAAG,CAAC,KAAK,EAAE,UAAU,KAAK,OAAO,CAAC,IAAI,CAAC,GAAG,IAAI,CAAC;AAEzE;;;GAGG;AACH,wBAAgB,SAAS,CACxB,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,WAAW,CAAC,UAAU,EAAE,YAAY,EAAE,CAAC,CAiBzC;AAED;;;;;;;GAOG;AACH,wBAAgB,iBAAiB,CAChC,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,WAAW,CAAC,UAAU,EAAE,YAAY,EAAE,CAAC,CAwBzC;AAED,wBAAsB,YAAY,CACjC,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,IAAI,EAAE,cAAc,EACpB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,OAAO,CAAC,YAAY,EAAE,CAAC,CAgBzB;AAED,wBAAsB,oBAAoB,CACzC,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,IAAI,EAAE,cAAc,EACpB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,OAAO,CAAC,YAAY,EAAE,CAAC,CAiBzB;AAmBD,qBAAa,uBAAwB,SAAQ,KAAK;gBACrC,SAAS,EAAE,MAAM;CAM7B;AAqPD,uGAAuG;AACvG,wBAAsB,oBAAoB,CACzC,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE,cAAc,GAAG,kBAAkB,CAAC,EAClE,MAAM,CAAC,EAAE,WAAW,GAClB,OAAO,CAAC,OAAO,CAAC,CAgBlB;AA+lBD,MAAM,WAAW,qBAAqB;IACrC,QAAQ,EAAE,aAAa,CAAC;IACxB,IAAI,EAAE,SAAS,CAAC;IAChB,IAAI,EAAE,OAAO,CAAC;CACd;AAED,wBAAgB,6BAA6B,CAAC,IAAI,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,GAAG,aAAa,CAYrG;AAED,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,GAAG,qBAAqB,CAOpG"}
+{"version":3,"file":"agent-loop.d.ts","sourceRoot":"","sources":["../src/agent-loop.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EAIN,KAAK,OAAO,EAEZ,WAAW,EAIX,MAAM,uBAAuB,CAAC;AAU/B,OAAO,KAAK,EACX,YAAY,EACZ,UAAU,EACV,eAAe,EAEf,YAAY,EACZ,SAAS,EACT,aAAa,EAEb,QAAQ,EACR,MAAM,YAAY,CAAC;AAEpB,MAAM,MAAM,cAAc,GAAG,CAAC,KAAK,EAAE,UAAU,KAAK,OAAO,CAAC,IAAI,CAAC,GAAG,IAAI,CAAC;AAEzE;;;GAGG;AACH,wBAAgB,SAAS,CACxB,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,WAAW,CAAC,UAAU,EAAE,YAAY,EAAE,CAAC,CAiBzC;AAED;;;;;;;GAOG;AACH,wBAAgB,iBAAiB,CAChC,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,WAAW,CAAC,UAAU,EAAE,YAAY,EAAE,CAAC,CAwBzC;AAED,wBAAsB,YAAY,CACjC,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,IAAI,EAAE,cAAc,EACpB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,OAAO,CAAC,YAAY,EAAE,CAAC,CAgBzB;AAED,wBAAsB,oBAAoB,CACzC,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,eAAe,EACvB,IAAI,EAAE,cAAc,EACpB,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,QAAQ,EAAE,QAAQ,GAChB,OAAO,CAAC,YAAY,EAAE,CAAC,CAiBzB;AAmBD,qBAAa,uBAAwB,SAAQ,KAAK;gBACrC,SAAS,EAAE,MAAM;CAM7B;AA2ND,uGAAuG;AACvG,wBAAsB,oBAAoB,CACzC,OAAO,EAAE,YAAY,EACrB,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE,cAAc,GAAG,kBAAkB,CAAC,EAClE,MAAM,CAAC,EAAE,WAAW,GAClB,OAAO,CAAC,OAAO,CAAC,CAIlB;AA+mBD,MAAM,WAAW,qBAAqB;IACrC,QAAQ,EAAE,aAAa,CAAC;IACxB,IAAI,EAAE,SAAS,CAAC;IAChB,IAAI,EAAE,OAAO,CAAC;CACd;AAED,wBAAgB,6BAA6B,CAAC,IAAI,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,GAAG,aAAa,CAYrG;AAED,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,GAAG,qBAAqB,CAOpG"}

package/node_modules/@earendil-works/pi-agent-core/dist/agent-loop.js

@@ -96,9 +96,8 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
     let config = initialConfig;
     let firstTurn = true;
     let firstProviderRequest = true;
-    // Check for steering messages at start (user may have typed while waiting)
-    let pendingMessages = (await config.getSteeringMessages?.()) || [];
     let drainedTerminatingQueue;
+    let turnStartAlreadyEmitted = false;
     const refreshTerminatingQueueDrain = async () => {
         if (!drainedTerminatingQueue || !config.restorePendingMessages)
             return;
@@ -110,12 +109,17 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
             drainedTerminatingQueue = pendingMessages.length > 0 ? "followUp" : undefined;
         }
     };
+    // Check for steering messages at start (user may have typed while waiting)
+    let pendingMessages = (await config.getSteeringMessages?.()) || [];
     // Outer loop: continues when queued follow-up messages arrive after agent would stop
     while (true) {
         let hasMoreToolCalls = true;
         // Inner loop: process tool calls and steering messages
         while (hasMoreToolCalls || pendingMessages.length > 0) {
-            if (!firstTurn) {
+            if (turnStartAlreadyEmitted) {
+                turnStartAlreadyEmitted = false;
+            }
+            else if (!firstTurn) {
                 await emit({ type: "turn_start" });
             }
             else {
@@ -139,10 +143,7 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
                 }
                 pendingMessages = [];
             }
-            // Stream assistant response. Continuation-scoped overrides apply to one
-            // provider request only. Once that request emits its first event, the
-            // configured idle timeout resumes so healthy reasoning gaps are not bound
-            // by the short liveness probe.
+            // Stream assistant response
             const isInitialProviderRequest = firstProviderRequest;
             firstProviderRequest = false;
             const requestConfig = isInitialProviderRequest
@@ -152,16 +153,10 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
                     streamStartTimeoutMs: config.initialRequestStreamStartTimeoutMs ?? config.streamStartTimeoutMs,
                 }
                 : config;
-            const streamIdleTimeoutMs = isInitialProviderRequest ? config.timeoutMs : requestConfig.timeoutMs;
-            const streamed = await streamAssistantResponse(currentContext, requestConfig, signal, emit, withEmptyAssistantRecovery(requestConfig.model, streamFunction), streamIdleTimeoutMs);
+            const streamed = await streamAssistantResponse(currentContext, requestConfig, signal, emit, withEmptyAssistantRecovery(requestConfig.model, streamFunction), isInitialProviderRequest ? config.timeoutMs : requestConfig.timeoutMs);
             const message = promoteStopWithPendingToolCalls(streamed.message);
             const providerToolResults = streamed.providerToolResults;
             newMessages.push(message);
-            // Provider-resolved (Cursor exec-channel) tool results pair with
-            // already-resolved toolCall blocks in the assistant message. They are
-            // appended right after it — including on terminal error/abort paths,
-            // where dropping them would leave resolved calls unpaired and strip
-            // the interaction from every rebuilt transcript.
             const toolResults = [];
             for (const result of providerToolResults) {
                 await emit({ type: "message_start", message: result });
@@ -175,18 +170,14 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
                 await emit({ type: "agent_end", messages: newMessages });
                 return;
             }
-            // Check for tool calls. Blocks stamped `kCursorExecResolved` were
-            // already executed by Cursor's exec channel mid-stream (their results
-            // arrived via `providerToolResults`); running them here again would
-            // duplicate side-effecting tools.
+            // Check for tool calls
             const toolCalls = message.content.filter((c) => c.type === "toolCall" && !isCursorExecResolved(c));
             hasMoreToolCalls = false;
             let toolBatchTerminated = false;
             if (toolCalls.length > 0) {
-                // A native "length" stop means the output was cut off by the token limit,
-                // so every tool call in the message may carry truncated arguments. Text
-                // tool-call middleware finalizes its calls as "toolUse", leaving only
-                // native, unwrapped length responses for this message-wide safeguard.
+                // A "length" stop means the output was cut off by the token limit, so
+                // every tool call in the message may carry truncated arguments. Fail
+                // them all instead of executing potentially borked calls.
                 const executedToolBatch = message.stopReason === "length"
                     ? await failToolCallsFromTruncatedMessage(toolCalls, emit)
                     : await executeToolCalls(currentContext, message, config, signal, emit);
@@ -209,34 +200,35 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
                 context: currentContext,
                 newMessages,
             };
+            if (await config.shouldStopAfterTurn?.(nextTurnContext)) {
+                await emit({ type: "agent_end", messages: newMessages });
+                return;
+            }
             if (toolBatchTerminated) {
-                if (await config.shouldStopAfterTurn?.(nextTurnContext)) {
-                    await emit({ type: "agent_end", messages: newMessages });
-                    return;
-                }
                 pendingMessages = (await config.getSteeringMessages?.()) || [];
-                if (pendingMessages.length > 0) {
+                if (pendingMessages.length > 0)
                     drainedTerminatingQueue = "steering";
-                }
                 if (pendingMessages.length === 0) {
                     pendingMessages = (await config.getFollowUpMessages?.()) || [];
-                    if (pendingMessages.length > 0) {
+                    if (pendingMessages.length > 0)
                         drainedTerminatingQueue = "followUp";
-                    }
                 }
                 if (pendingMessages.length === 0) {
                     await emit({ type: "agent_end", messages: newMessages });
                     return;
                 }
+                // Give queue owners a boundary before preparation refreshes the drained
+                // snapshot, so a clear or replacement wins before admission.
+                await emit({ type: "turn_start" });
+                turnStartAlreadyEmitted = true;
             }
             let nextTurnSnapshot;
             try {
                 nextTurnSnapshot = await config.prepareNextTurn?.(nextTurnContext);
             }
             catch (error) {
-                if (drainedTerminatingQueue) {
+                if (drainedTerminatingQueue)
                     await config.restorePendingMessages?.(drainedTerminatingQueue, pendingMessages);
-                }
                 throw error;
             }
             if (nextTurnSnapshot) {
@@ -256,9 +248,8 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
                 };
             }
             if (signal?.aborted) {
-                if (drainedTerminatingQueue) {
+                if (drainedTerminatingQueue)
                     await config.restorePendingMessages?.(drainedTerminatingQueue, pendingMessages);
-                }
                 await emit({ type: "agent_end", messages: newMessages });
                 return;
             }
@@ -268,16 +259,7 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
                     await emit({ type: "agent_end", messages: newMessages });
                     return;
                 }
-            }
-            if (!toolBatchTerminated &&
-                (await config.shouldStopAfterTurn?.({
-                    message,
-                    toolResults,
-                    context: currentContext,
-                    newMessages,
-                }))) {
-                await emit({ type: "agent_end", messages: newMessages });
-                return;
+                drainedTerminatingQueue = undefined;
             }
             if (!toolBatchTerminated) {
                 pendingMessages = (await config.getSteeringMessages?.()) || [];
@@ -297,19 +279,10 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
 }
 /** Build the provider context using the same transform and conversion pipeline as an agent request. */
 export async function buildProviderContext(context, config, signal) {
-    // Apply context transform if configured (AgentMessage[] → AgentMessage[])
     let messages = context.messages;
-    if (config.transformContext) {
+    if (config.transformContext)
         messages = await config.transformContext(messages, signal);
-    }
-    // Convert to LLM-compatible messages (AgentMessage[] → Message[])
-    const llmMessages = await config.convertToLlm(messages);
-    // Build LLM context
-    return {
-        systemPrompt: context.systemPrompt,
-        messages: llmMessages,
-        tools: context.tools,
-    };
+    return { systemPrompt: context.systemPrompt, messages: await config.convertToLlm(messages), tools: context.tools };
 }
 /**
  * Stream an assistant response from the LLM.
@@ -695,6 +668,7 @@ async function executeToolCallsSequential(currentContext, assistantMessage, tool
 }
 async function executeToolCallsParallel(currentContext, assistantMessage, toolCalls, config, signal, emit) {
     const finalizedCalls = [];
+    const preparedCalls = [];
     let lastSequentialCall;
     let currentParallelWave = [];
     for (const toolCall of toolCalls) {
@@ -711,24 +685,44 @@ async function executeToolCallsParallel(currentContext, assistantMessage, toolCa
             : lastSequentialCall
                 ? [lastSequentialCall]
                 : [];
-        const finalizedCall = (async () => {
-            await Promise.all(dependencies);
-            const finalized = await runPreparedToolCall(currentContext, assistantMessage, preparation, config, signal, emit);
-            await emitToolExecutionEnd(finalized, emit);
-            return finalized;
-        })();
-        finalizedCalls.push(finalizedCall);
+        preparedCalls.push({ preparation, isSequential, dependencies });
         if (isSequential) {
-            lastSequentialCall = finalizedCall;
+            // Dependencies are assigned in the second phase after all preflight hooks
+            // have completed, so a later preflight abort vetoes every execution.
+            lastSequentialCall = undefined;
             currentParallelWave = [];
         }
         else {
-            currentParallelWave.push(finalizedCall);
+            currentParallelWave.push(Promise.resolve(undefined));
         }
         if (signal?.aborted) {
             break;
         }
     }
+    let previousSequential;
+    let previousWave = [];
+    for (const { preparation, isSequential } of preparedCalls) {
+        const dependencies = isSequential
+            ? [...(previousSequential ? [previousSequential] : []), ...previousWave]
+            : previousSequential
+                ? [previousSequential]
+                : [];
+        const finalizedCall = (async () => {
+            await Promise.all(dependencies);
+            const finalized = signal?.aborted
+                ? { toolCall: preparation.toolCall, result: createErrorToolResult("Operation aborted"), isError: true }
+                : await runPreparedToolCall(currentContext, assistantMessage, preparation, config, signal, emit);
+            await emitToolExecutionEnd(finalized, emit);
+            return finalized;
+        })();
+        finalizedCalls.push(finalizedCall);
+        if (isSequential) {
+            previousSequential = finalizedCall;
+            previousWave = [];
+        }
+        else
+            previousWave.push(finalizedCall);
+    }
     const orderedFinalizedCalls = await Promise.all(finalizedCalls);
     const messages = [];
     for (const finalized of orderedFinalizedCalls) {