@gajae-code/coding-agent 0.12.0 → 0.12.2

package/dist/types/task/receipt.d.ts

@@ -1,4 +1,4 @@
-import { type SingleResult, type TaskToolDetails } from "./types";
+import { type ReviewFindingsArtifactRef, type SingleResult, type TaskToolDetails } from "./types";
 export interface TaskRoi {
     tokens: number;
     contextTokens?: number;
@@ -29,6 +29,7 @@ export interface TaskResultReceipt {
     contextWindow?: number;
     modelOverride?: string | string[];
     modelSubstitutionWarning?: SingleResult["modelSubstitutionWarning"];
+    fastMode?: boolean;
     usage?: SingleResult["usage"];
     cost?: number;
     usageCostBreakdownComplete?: true;
@@ -37,6 +38,9 @@ export interface TaskResultReceipt {
         attempt: number;
         errorSummary: string;
     };
+    setupFailure?: {
+        summary: string;
+    };
     errorSummary?: string;
     abortSummary?: string;
     preview: string;
@@ -46,6 +50,8 @@ export interface TaskResultReceipt {
         sizeBytes: number;
         lineCount: number;
         sha256?: string;
+        /** Output remains readable for the parent session lifetime (and same-session descendants). */
+        durability?: "session";
     };
     outputUnavailable?: boolean;
     review?: {
@@ -55,6 +61,8 @@ export interface TaskResultReceipt {
             severity?: string;
             summary: string;
         }>;
+        /** Canonical full-fidelity findings; inline summaries are display-only. */
+        findingsRef?: ReviewFindingsArtifactRef;
     };
     extractedToolCounts?: Record<string, number>;
     forkContext?: SingleResult["forkContext"];

package/dist/types/task/types.d.ts

@@ -368,6 +368,13 @@ export interface ReviewFinding {
     line_start: number;
     line_end: number;
 }
+/** Durable full-fidelity review findings artifact associated with a task result. */
+export interface ReviewFindingsArtifactRef {
+    uri: `artifact://${string}`;
+    sizeBytes: number;
+    sha256: string;
+    findingCount: number;
+}
 /** Review summary submitted by the reviewer agent */
 export interface ReviewSummary {
     overall_correctness: "correct" | "incorrect";
@@ -439,6 +446,8 @@ export interface AgentProgress {
     durationMs: number;
     modelOverride?: string | string[];
     modelSubstitutionWarning?: ModelSubstitutionWarning;
+    /** Whether the resolved subagent model runs under the effective fast service tier. */
+    fastMode?: boolean;
     /** Data extracted by registered subprocess tool handlers (keyed by tool name) */
     extractedToolData?: Record<string, unknown[]>;
     /**
@@ -469,6 +478,8 @@ export interface AgentProgress {
         attempt: number;
         errorMessage: string;
     };
+    /** Safe diagnostic retained in terminal progress when setup fails before the first LLM request. */
+    setupFailure?: SetupFailureSummary;
     /**
      * Snapshot of the most recent `task` tool call's in-flight `TaskToolDetails`,
      * captured from `tool_execution_update`. Lets the parent UI surface live
@@ -478,6 +489,12 @@ export interface AgentProgress {
      */
     inflightTaskDetails?: TaskToolDetails;
 }
+/** Bounded diagnostic retained when subagent setup fails before an LLM request starts. */
+export interface SetupFailureSummary {
+    summary: string;
+}
+/** Create a bounded diagnostic that preserves a setup failure's cause without exposing credentials or local paths. */
+export declare function createSetupFailureSummary(error: unknown): SetupFailureSummary;
 /** Result from a single agent execution */
 export interface SingleResult {
     index: number;
@@ -501,7 +518,11 @@ export interface SingleResult {
     contextWindow?: number;
     modelOverride?: string | string[];
     modelSubstitutionWarning?: ModelSubstitutionWarning;
+    /** Whether the resolved subagent model ran under the effective fast service tier. */
+    fastMode?: boolean;
     error?: string;
+    /** Safe diagnostic for a failure before the subagent sent its first LLM request. */
+    setupFailure?: SetupFailureSummary;
     aborted?: boolean;
     abortReason?: string;
     paused?: boolean;
@@ -521,6 +542,8 @@ export interface SingleResult {
     producedChanges?: boolean;
     /** Data extracted by registered subprocess tool handlers (keyed by tool name) */
     extractedToolData?: Record<string, unknown[]>;
+    /** Full wrapper-owned review evidence, kept separate from caller completion data. */
+    reviewFindingsRef?: ReviewFindingsArtifactRef;
     /**
      * Terminal retry failure, when the subagent exited because the auto-retry
      * loop gave up (retry-after exceeded the cap, or all attempts exhausted).

package/dist/types/tools/ask-answer-registry.d.ts

@@ -8,9 +8,9 @@
  */
 import type { WorkflowGateEmitter } from "../modes/shared/agent-wire/workflow-gate-broker";
 import type { AskAnswerSource } from "./index";
-/** Register `source` for `sessionId`. Returns a disposer that clears it. */
+/** Register `source` for `sessionId`. Returns a disposer that restores the prior source. */
 export declare function registerAskAnswerSource(sessionId: string, source: AskAnswerSource): () => void;
-/** The answer source for `sessionId`, if one is registered. */
+/** The highest-priority answer source for `sessionId`, if one is registered. */
 export declare function getAskAnswerSource(sessionId: string): AskAnswerSource | undefined;
 /** Publish a session's current workflow-gate emitter after mode initialization. */
 export declare function notifyWorkflowGateEmitterChanged(sessionId: string, emitter: WorkflowGateEmitter | undefined): void;

package/dist/types/tools/bash-interactive.d.ts

@@ -1,5 +1,5 @@
 import type { AgentToolContext } from "@gajae-code/agent-core";
-import { type OutputSummary } from "../session/streaming-output";
+import { type OutputSummary, type TerminalArtifactPublisher } from "../session/streaming-output";
 export interface BashInteractiveResult extends OutputSummary {
     exitCode: number | undefined;
     cancelled: boolean;
@@ -13,4 +13,7 @@ export declare function runInteractiveBashPty(ui: NonNullable<AgentToolContext["
     env?: Record<string, string>;
     artifactPath?: string;
     artifactId?: string;
+    artifactPublisher?: TerminalArtifactPublisher;
+    spillThreshold?: number;
+    headBytes?: number;
 }): Promise<BashInteractiveResult>;

package/dist/types/tools/bash.d.ts

@@ -1,12 +1,14 @@
 import type { AgentTool, AgentToolContext, AgentToolResult, AgentToolUpdateCallback } from "@gajae-code/agent-core";
 import type { Component } from "@gajae-code/tui";
 import * as z from "zod/v4";
+import { type BashArtifactSaveResult } from "../exec/bash-executor";
 import type { RenderResultOptions } from "../extensibility/custom-tools/types";
 import { type Theme } from "../modes/theme/theme";
 import type { ToolSession } from ".";
 import { type OutputMeta } from "./output-meta";
 export declare const BASH_DEFAULT_PREVIEW_LINES = 10;
-export declare function saveBashOriginalArtifactForTests(session: ToolSession, originalText: string): Promise<string | undefined>;
+export type BashOriginalArtifactSaveResult = BashArtifactSaveResult;
+export declare function saveBashOriginalArtifactForTests(session: ToolSession, originalText: string, onResult?: (result: BashOriginalArtifactSaveResult) => void): Promise<string | undefined>;
 declare const bashSchemaBase: z.ZodObject<{
     command: z.ZodString;
     env: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;

package/dist/types/tools/image-gen.d.ts

@@ -134,7 +134,11 @@ export declare function setConfiguredImageModel(config: ImageProviderConfig | nu
 export declare function getConfiguredImageModel(): ImageProviderConfig | null;
 /** Resolve the effective image model for a configured provider */
 export declare function resolveImageModel(provider: string, modelOverride: string | null): string;
+/** Test seam: the Google image key fallback as resolved from trusted env. */
+export declare function googleImageApiKeyFromEnvForTest(): string | undefined;
 export declare function isOpenAIHostedImageModel(model: Model | undefined): model is Model;
+/** Test seam: the image-request base URL as resolved from trusted env. */
+export declare function getOpenAIImageBaseUrlForTest(model: Model, authCredentialType?: "api_key" | "oauth"): string;
 export declare const imageGenTool: CustomTool<typeof imageGenSchema, ImageGenToolDetails>;
 export declare function getImageGenTools(modelRegistry?: ModelRegistry, activeModel?: Model): Promise<Array<CustomTool<typeof imageGenSchema, ImageGenToolDetails>>>;
 export declare function getImageGenToolsWithRegistry(modelRegistry: ModelRegistry, activeModel?: Model): Promise<Array<CustomTool<typeof imageGenSchema, ImageGenToolDetails>>>;

package/dist/types/tools/index.d.ts

@@ -215,8 +215,12 @@ export interface ToolSession {
     strictToolNames?: boolean;
     /** Get artifacts directory for artifact:// URLs */
     getArtifactsDir?: () => string | null;
+    /** Additional artifacts directories explicitly authorized for this session's tree (parent/child/sibling reads). Derived only from the explicitly adopted/shared `ArtifactManager`; never enumerates unrelated live sessions. */
+    getAuthorizedArtifactsDirs?: () => readonly string[];
     /** Get the ArtifactManager backing this session (shared across parent + subagents). */
     getArtifactManager?: () => ArtifactManager | null;
+    /** Register teardown work owned by the current logical session. */
+    registerSessionCleanup?: (cleanup: () => Promise<void> | void) => () => void;
     /** Allocate a new artifact path and ID for session-scoped truncated output. */
     allocateOutputArtifact?: (toolType: string) => Promise<{
         id?: string;
@@ -240,6 +244,8 @@ export interface ToolSession {
     settings: Settings;
     /** Live service-tier intent of the parent session, inherited by `inherit` subagents. */
     serviceTier?: ServiceTier;
+    /** Whether the effective subagent tier grants fast mode for a resolved provider. */
+    isFastForSubagentProvider?: (provider?: string) => boolean;
     /** Plan mode state (if active) */
     getPlanModeState?: () => PlanModeState | undefined;
     /** Goal mode state (if active or paused) */

package/dist/types/tools/output-meta.d.ts

@@ -46,8 +46,12 @@ export interface TruncationMeta {
     elidedBytes?: number;
     /** Lines elided from the middle. */
     elidedLines?: number;
-    /** Artifact ID if full output was saved */
+    /** Artifact ID when output was persisted; completeness is tracked separately below. */
     artifactId?: string;
+    /** Bytes omitted from an artifact after its hard storage cap was reached. */
+    artifactTruncatedBytes?: number;
+    /** Bounded diagnostic when artifact writer creation, write, or finalization failed. */
+    artifactFailureDiagnostic?: string;
     /** Next offset for pagination (head truncation only) */
     nextOffset?: number;
 }
@@ -145,7 +149,7 @@ export declare class OutputMetaBuilder {
         maxBytes?: number;
         rangeBase?: "file" | "window";
     }): this;
-    /** Add truncation info from OutputSummary. No-op if not truncated. */
+    /** Add truncation info from OutputSummary. No-op if not truncated or artifact evidence is absent. */
     truncationFromSummary(summary: OutputSummary, options: TruncationSummaryOptions): this;
     /** Add truncation info from truncated output text. No-op if truncation not detected. */
     truncationFromText(text: string, options: TruncationTextOptions): this;
@@ -178,6 +182,10 @@ export declare class OutputMetaBuilder {
 /** Create a new OutputMetaBuilder. */
 export declare function outputMeta(): OutputMetaBuilder;
 export declare function formatFullOutputReference(artifactId: string): string;
+/**
+ * Format an artifact reference without claiming completeness when storage was hard-capped.
+ */
+export declare function formatArtifactReference(artifactId: string, artifactTruncatedBytes?: number): string;
 export declare function formatTruncationMetaNotice(truncation: TruncationMeta): string;
 /**
  * Format styled artifact reference with warning color and brackets.
@@ -211,6 +219,19 @@ export declare function stripOutputNotice(text: string, meta: OutputMeta | undef
  * middle elision with the same per-user configuration.
  */
 export declare function resolveOutputSinkHeadBytes(s: Settings | undefined): number;
+export declare const BASH_DEFAULT_OUTPUT_TAIL_BYTES = 1024;
+/**
+ * Bash uses a deliberately small tail window to nudge callers toward focused
+ * commands and dedicated search tools. An explicitly configured shared tail
+ * budget still wins.
+ */
+export declare function resolveBashOutputSinkTailBytes(s: Settings): number;
+/**
+ * Bash keeps only the tail unless the user explicitly opts into the shared
+ * head-retention setting. Schema defaults still apply to other streaming
+ * tools without silently turning Bash back into middle-elision mode.
+ */
+export declare function resolveBashOutputSinkHeadBytes(s: Settings): number;
 /**
  * Resolve the per-line column cap from session settings. Shared by streaming
  * executors (bash/python/ssh/eval via OutputSink) and the `read` tool's

package/dist/types/tools/path-utils.d.ts

@@ -112,6 +112,7 @@ export interface ToolScopeOptions {
     rawPaths: string[];
     cwd: string;
     getArtifactsDir?: () => string | null;
+    getAuthorizedArtifactsDirs?: () => readonly string[];
     /** Verb used in the "Cannot {action} internal URL without a backing file: …" message. */
     internalUrlAction: string;
     /** Collect absolute paths flagged immutable by their internal-URL handler. */

package/dist/types/tools/read.d.ts

@@ -84,12 +84,6 @@ export interface ReadToolDetails {
     conflictCount?: number;
 }
 type ReadParams = ReadToolInput;
-/**
- * Read tool implementation.
- *
- * Reads files with support for images, converted documents (via markit), and text.
- * Directories return a formatted listing with modification times.
- */
 export declare class ReadTool implements AgentTool<typeof readSchema, ReadToolDetails> {
     #private;
     private readonly session;

package/dist/types/tools/review.d.ts

@@ -50,14 +50,12 @@ export interface SubmitReviewDetails {
 }
 export type { ReportFindingDetails };
 /**
- * Coerce a tool-side `ReportFindingDetails` into the cross-boundary
- * `ReviewFinding` shape consumed by the reviewer agent's JTD output schema.
+ * Convert tool-side `ReportFindingDetails` into the numeric-priority
+ * `ReviewFinding` shape used by caller-owned reviewer payloads.
  *
  * The `report_finding` tool exposes `priority` as a string enum (`"P0".."P3"`)
- * for ergonomics, but the bundled reviewer schema (and every custom review
- * agent that mirrors it) declares `priority: number`. Without this coercion
- * the auto-populated `findings[]` fails JTD validation and every review run
- * that surfaces a finding is rejected with `findings.0.priority: expected
- * number, received string`.
+ * for ergonomics, while reviewer schemas may explicitly declare numeric
+ * priorities. This pure conversion does not add findings to completion data;
+ * callers that include `findings` own both that field and its output schema.
  */
 export declare function toReviewFinding(details: ReportFindingDetails): ReviewFinding;

package/dist/types/tools/subagent.d.ts

@@ -38,6 +38,8 @@ export interface SubagentSnapshot {
     durationMs: number;
     resultText?: string;
     errorText?: string;
+    /** Safe setup failure cause retained from the executor launch path. */
+    setupFailureSummary?: string;
     resultPreview?: string;
     outputRef?: string;
     truncated?: boolean;

package/dist/types/utils/gh.d.ts

@@ -0,0 +1,13 @@
+export interface GhResult {
+    exitCode: number;
+    stdout: string;
+    stderr: string;
+    timedOut?: boolean;
+}
+export type RunGh = (args: string[], options?: {
+    timeoutMs?: number;
+}) => Promise<GhResult>;
+/** Run GitHub CLI without allowing it to consume the parent terminal's input. */
+export declare function runGhDefault(args: string[], options?: {
+    timeoutMs?: number;
+}): Promise<GhResult>;

package/dist/types/utils/image-resize.d.ts

@@ -5,6 +5,7 @@ export interface ImageResizeOptions {
     maxBytes?: number;
     jpegQuality?: number;
     excludeWebP?: boolean;
+    forceReencode?: boolean;
 }
 export interface ResizedImage {
     buffer: Uint8Array;

package/dist/types/web/search/providers/kimi.d.ts

@@ -16,6 +16,8 @@ export interface KimiSearchParams {
     authStorage: AuthStorage;
     sessionId?: string;
 }
+/** Test seam: the Kimi search base URL as resolved from trusted env. */
+export declare function resolveKimiSearchBaseUrlForTest(): string;
 /** Execute Kimi web search. */
 export declare function searchKimi(params: KimiSearchParams): Promise<SearchResponse>;
 /** Search provider for Kimi web search. */

package/dist/types/web/search/providers/searxng.d.ts

@@ -28,6 +28,13 @@ import type { AuthStorage } from "@gajae-code/ai";
 import type { SearchResponse } from "../../../web/search/types";
 import type { SearchParams } from "./base";
 import { SearchProvider } from "./base";
+/** Test seam: the SearXNG endpoint and auth material as resolved from trusted env. */
+export declare function resolveSearxngConfigForTest(): {
+    endpoint: string | null;
+    token: string | null;
+    basicUsername: string | null;
+    basicPassword: string | null;
+};
 /** Execute SearXNG web search. */
 export declare function searchSearXNG(params: {
     query: string;

package/dist/types/web/search/providers/xai.d.ts

@@ -32,6 +32,8 @@ export interface XaiSearchParams {
     authStorage: AuthStorage;
     sessionId?: string;
 }
+/** Test seam: the xAI search base URL as resolved from trusted env. */
+export declare function resolveXaiSearchBaseUrlForTest(): string;
 export declare function buildXaiRequestBody(params: {
     query: string;
     systemPrompt: string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@gajae-code/coding-agent",
-	"version": "0.12.0",
+	"version": "0.12.2",
 	"description": "Gajae Code CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://gajae-code.com",
 	"author": "Yeachan-Heo and Gajae Code Contributors",
@@ -53,13 +53,13 @@
 		"@agentclientprotocol/sdk": "1.3.0",
 		"@babel/parser": "^7.29.3",
 		"@mozilla/readability": "^0.6.0",
-		"@gajae-code/stats": "0.12.0",
-		"@gajae-code/agent-core": "0.12.0",
-		"@gajae-code/ai": "0.12.0",
-		"@gajae-code/bridge-client": "0.12.0",
-		"@gajae-code/natives": "0.12.0",
-		"@gajae-code/tui": "0.12.0",
-		"@gajae-code/utils": "0.12.0",
+		"@gajae-code/stats": "0.12.2",
+		"@gajae-code/agent-core": "0.12.2",
+		"@gajae-code/ai": "0.12.2",
+		"@gajae-code/bridge-client": "0.12.2",
+		"@gajae-code/natives": "0.12.2",
+		"@gajae-code/tui": "0.12.2",
+		"@gajae-code/utils": "0.12.2",
 		"@puppeteer/browsers": "^2.13.0",
 		"@types/turndown": "5.0.6",
 		"@xterm/headless": "^6.0.0",
@@ -327,6 +327,8 @@
 			"types": "./dist/types/export/html/*.d.ts",
 			"import": "./src/export/html/*.ts"
 		},
+		"./extensibility/gjc-plugins/installer": null,
+		"./extensibility/gjc-plugins/registry": null,
 		"./extensibility/*": {
 			"types": "./dist/types/extensibility/*.d.ts",
 			"import": "./src/extensibility/*.ts"

package/scripts/acp-conformance-agent.ts

@@ -0,0 +1,218 @@
+#!/usr/bin/env bun
+
+/**
+ * Credential-free ACP fixture launcher. The loopback endpoint is deliberately
+ * deterministic; the child remains the production `gjc --mode acp` surface.
+ */
+import * as fs from "node:fs/promises";
+import * as os from "node:os";
+import * as path from "node:path";
+
+const root = path.resolve(import.meta.dir, "..", "..", "..");
+const agentDir = await fs.mkdtemp(path.join(os.tmpdir(), "gjc-acp-conformance-"));
+let child: ReturnType<typeof Bun.spawn> | undefined;
+let server: ReturnType<typeof Bun.serve> | undefined;
+
+/**
+ * Deterministic replies for the pinned `acp-core-v1` corpus. Each branch mirrors one
+ * upstream case expectation; nothing here reaches the network.
+ */
+function response(prompt: string): string {
+	if (prompt.includes("inspect-prompt")) return `received blocks: {"type":"resource"}`;
+	if (prompt.includes("this-command-does-not-exist")) return "unrecognized prompt";
+	if (/^\s*read\b/m.test(prompt)) return "read README.md: acpx";
+	if (/^\s*write\b/m.test(prompt)) {
+		const target = /write\s+(\S+)/.exec(prompt)?.[1] ?? "file";
+		return `wrote ${target}`;
+	}
+	const echo = /^\s*echo\s*(.*)$/m.exec(prompt);
+	if (echo) return echo[1] || "echo";
+	return "ok";
+}
+
+function sse(value: unknown): string {
+	return `data: ${JSON.stringify(value)}\n\n`;
+}
+
+function completion(delta: Record<string, unknown>, finishReason: string | null = null): string {
+	return sse({ id: "gjc-conformance", choices: [{ delta, finish_reason: finishReason }] });
+}
+
+function toolCall(name: "read" | "write", arguments_: Record<string, string>): string {
+	return completion(
+		{
+			tool_calls: [
+				{
+					index: 0,
+					id: `conformance-${name}`,
+					type: "function",
+					function: { name, arguments: JSON.stringify(arguments_) },
+				},
+			],
+		},
+		"tool_calls",
+	);
+}
+
+function textCompletion(text: string): string {
+	return `${completion({ content: text })}${completion({}, "stop")}data: [DONE]\n\n`;
+}
+
+async function cleanup(code = 0): Promise<never> {
+	server?.stop(true);
+	if (child && child.exitCode === null) child.kill();
+	await fs.rm(agentDir, { recursive: true, force: true });
+	process.exit(code);
+}
+
+server = Bun.serve({
+	hostname: "127.0.0.1",
+	port: 0,
+	async fetch(request) {
+		const body = (await request.json().catch(() => ({}))) as {
+			messages?: Array<{ role?: string; content?: unknown }>;
+		};
+		// OpenAI-style content is either a plain string or an array of typed blocks.
+		const textOf = (content: unknown): string =>
+			typeof content === "string"
+				? content
+				: Array.isArray(content)
+					? content
+							.map(block =>
+								typeof block === "object" &&
+								block !== null &&
+								typeof (block as { text?: unknown }).text === "string"
+									? (block as { text: string }).text
+									: "",
+							)
+							.join("")
+					: "";
+		const messages = body.messages ?? [];
+		const prompt =
+			[...messages]
+				.reverse()
+				.filter(message => message.role === "user")
+				.map(message => textOf(message.content))
+				.find(text => text.length > 0) ?? "";
+		const hasToolResult = messages.some(message => message.role === "tool");
+		const isRead = /^\s*read\b/m.test(prompt);
+		const isWrite = /^\s*write\b/m.test(prompt);
+		const isLateTool = prompt.includes("late-tool 40 follow-up");
+
+		if (prompt.includes("sleep")) {
+			const requested = Number(/\bsleep\s+(\d+)/.exec(prompt)?.[1] ?? 0);
+			const delay = Math.min(Math.max(requested, 0), 30_000);
+			const stream = new ReadableStream<Uint8Array>({
+				start(controller) {
+					let closed = false;
+					const close = () => {
+						if (closed) return;
+						closed = true;
+						controller.close();
+					};
+					request.signal.addEventListener("abort", close, { once: true });
+					void Bun.sleep(delay).then(() => {
+						if (closed) return;
+						controller.enqueue(new TextEncoder().encode(textCompletion("slept")));
+						close();
+					});
+				},
+			});
+			return new Response(stream, { headers: { "content-type": "text/event-stream" } });
+		}
+
+		if ((isRead || isWrite || isLateTool) && !hasToolResult) {
+			if (isLateTool) {
+				const prefix = `${completion({ content: "preparing" })}${completion({ content: " writing now" })}${completion({ content: " before tool" })}`;
+				return new Response(
+					`${prefix}${toolCall("write", { path: ".acpx-conformance-late.txt", content: "late" })}data: [DONE]\n\n`,
+					{
+						headers: { "content-type": "text/event-stream" },
+					},
+				);
+			}
+			// Neutral pre-tool text: every corpus assertion below must come from the real
+			// tool result, never from a marker announced before the tool ran.
+			const provisional = isRead ? "checking the requested file" : "applying the requested write";
+			return new Response(
+				`${completion({ content: provisional })}${toolCall(
+					isRead ? "read" : "write",
+					isRead
+						? { path: "README.md" }
+						: {
+								path: /write\s+(\S+)/.exec(prompt)?.[1] ?? "file",
+								content: /write\s+\S+\s*(.*)/.exec(prompt)?.[1] ?? "",
+							},
+				)}data: [DONE]\n\n`,
+				{ headers: { "content-type": "text/event-stream" } },
+			);
+		}
+
+		// Derive the reply from what the tool actually returned so a broken permission
+		// bridge, filesystem bridge, or tool result cannot still satisfy the corpus.
+		const toolText = messages
+			.filter(message => message.role === "tool")
+			.map(message => textOf(message.content))
+			.join("\n");
+		const toolDenied = /rejected by user|permission denied|not permitted/i.test(toolText);
+		const reply = toolDenied
+			? `permission denied: ${toolText.slice(0, 200)}`
+			: isRead
+				? `read README.md: ${toolText.trim().slice(0, 200)}`
+				: isWrite
+					? `wrote ${/write\s+(\S+)/.exec(prompt)?.[1] ?? "file"}: ${toolText.trim().slice(0, 120)}`
+					: isLateTool
+						? `writing now: ${toolText.trim().slice(0, 120)}`
+						: response(prompt);
+		return new Response(textCompletion(reply), { headers: { "content-type": "text/event-stream" } });
+	},
+});
+
+await Bun.write(
+	path.join(agentDir, "models.json"),
+	JSON.stringify({
+		providers: {
+			conformance: {
+				api: "openai-completions",
+				baseUrl: `http://127.0.0.1:${server.port}/v1`,
+				auth: "none",
+				models: [{ id: "fixture", name: "fixture", input: ["text"], contextWindow: 1_000_000, maxTokens: 8192 }],
+			},
+		},
+	}),
+);
+
+// The pinned corpus reads `README.md` from the session cwd the runner passes on
+// `session/new`. The runner creates that directory but seeds no content, so seed it
+// here (never inside the upstream corpus) when the harness names it.
+const scratchCwd = process.env.GJC_ACP_CONFORMANCE_CWD?.trim();
+if (scratchCwd) {
+	const scratchReadme = path.join(scratchCwd, "README.md");
+	if (!(await Bun.file(scratchReadme).exists())) await Bun.write(scratchReadme, "acpx conformance workspace\n");
+}
+
+child = Bun.spawn(
+	[
+		"bun",
+		path.join(root, "packages", "coding-agent", "src", "cli.ts"),
+		"--mode",
+		"acp",
+		"--model",
+		"conformance/fixture",
+	],
+	{
+		cwd: root,
+		env: { ...process.env, GJC_CODING_AGENT_DIR: agentDir, PI_CODING_AGENT_DIR: agentDir },
+		stdin: "inherit",
+		stdout: "inherit",
+		stderr: "inherit",
+	},
+);
+process.once("SIGTERM", () => {
+	void cleanup(0);
+});
+process.once("SIGINT", () => {
+	void cleanup(0);
+});
+await child.exited;
+await cleanup(child.exitCode ?? 1);