@gajae-code/coding-agent 0.14.2 → 0.15.0

package/src/tools/subagent.ts

@@ -4,7 +4,7 @@ import { formatDuration, logger, prompt } from "@gajae-code/utils";
 import * as z from "zod/v4";
 import { type AsyncJob, AsyncJobManager, jobElapsedMs, type SubagentRecord } from "../async";
 import subagentDescription from "../prompts/tools/subagent.md" with { type: "text" };
-import type { AgentProgress, AgentSource, TaskToolDetails } from "../task/types";
+import type { AgentProgress, AgentSource, LocalErrorSummary } from "../task/types";
 import { Ellipsis, truncateToWidth } from "../tui";
 import type { ToolSession } from "./index";
 import { replaceTabs } from "./render-utils";
@@ -85,6 +85,8 @@ export interface SubagentSnapshot {
 	errorText?: string;
 	/** Safe setup failure cause retained from the executor launch path. */
 	setupFailureSummary?: string;
+	/** Safe, bounded summary of a terminal local (non-provider) failure (e.g. local_buffer_overflow). */
+	localErrorSummary?: LocalErrorSummary;
 	resultPreview?: string;
 	outputRef?: string;
 	truncated?: boolean;
@@ -92,8 +94,8 @@ export interface SubagentSnapshot {
 	steerMessage?: string;
 	steerState?: "queued" | "resume_queued" | "resume_started";
 	steerPauseRequested?: boolean;
-	/** Live streaming progress for the awaited subagent (await panel only; UI detail). */
-	progress?: AgentProgress;
+	/** Bounded live progress approved for the await panel and public tool details. */
+	progress?: SubagentLiveProgress;
 	/** True when a live in-session progress producer exists for this subagent. */
 	liveProgressAvailable?: boolean;
 	/** Model the subagent actually runs on (after any auth fallback). */
@@ -106,6 +108,63 @@ export interface SubagentSnapshot {
 	fastMode?: boolean;
 }
 
+/**
+ * Public await-panel progress. This is deliberately not `AgentProgress`: raw
+ * progress contains model deltas, tool arguments, arbitrary output, and nested
+ * task payloads that must never enter tool-result, ACP, or telemetry envelopes.
+ */
+export interface SubagentLiveProgress {
+	id: string;
+	status: AgentProgress["status"];
+	currentTool?: string;
+	recentTool?: string;
+	recentOutputSummary?: { lineCount: number };
+	fastMode?: boolean;
+	retryState?: {
+		attempt: number;
+		maxAttempts: number;
+		unbounded?: boolean;
+		kind: NonNullable<AgentProgress["retryState"]>["kind"];
+		provider?: string;
+		lastProviderProgressAtMs?: number;
+		delayMs: number;
+		startedAtMs: number;
+	};
+	retryFailure?: { attempt: number };
+}
+
+function toSubagentLiveProgress(progress: AgentProgress): SubagentLiveProgress {
+	return {
+		id: progress.id,
+		status: progress.status,
+		...(progress.currentTool ? { currentTool: progress.currentTool } : {}),
+		...(progress.currentTool === undefined && progress.recentTools[0]
+			? { recentTool: progress.recentTools[0].tool }
+			: {}),
+		...(progress.recentOutput.length > 0
+			? { recentOutputSummary: { lineCount: Math.min(progress.recentOutput.length, 6) } }
+			: {}),
+		...(progress.fastMode ? { fastMode: true } : {}),
+		...(progress.retryState
+			? {
+					retryState: {
+						attempt: progress.retryState.attempt,
+						maxAttempts: progress.retryState.maxAttempts,
+						...(progress.retryState.unbounded ? { unbounded: true } : {}),
+						kind: progress.retryState.kind,
+						...(progress.retryState.provider ? { provider: progress.retryState.provider } : {}),
+						...(progress.retryState.lastProviderProgressAtMs !== undefined
+							? { lastProviderProgressAtMs: progress.retryState.lastProviderProgressAtMs }
+							: {}),
+						delayMs: progress.retryState.delayMs,
+						startedAtMs: progress.retryState.startedAtMs,
+					},
+				}
+			: {}),
+		...(progress.retryFailure ? { retryFailure: { attempt: progress.retryFailure.attempt } } : {}),
+	};
+}
+
 export type SubagentAwaitOutcome = "completed" | "timed_out" | "interrupted";
 
 const AWAIT_INTERRUPTED_GUIDANCE =
@@ -723,6 +782,22 @@ export class SubagentTool implements AgentTool<typeof subagentSchema, SubagentTo
 			if (snapshot.description) lines.push(`Description: ${snapshot.description}`);
 			if (snapshot.outputRef) lines.push(`Output: ${snapshot.outputRef}`);
 			if (snapshot.setupFailureSummary) lines.push(`Setup failure: ${snapshot.setupFailureSummary}`);
+			if (snapshot.localErrorSummary) {
+				const local = snapshot.localErrorSummary;
+				lines.push(`Local failure (${local.kind}): ${local.summary}`);
+				// Guidance is kind-conditional: an overflow is a local staging
+				// limit that reproduces on re-issue; a snapshot failure is a
+				// serialization defect, which re-issuing does NOT reproduce.
+				if (local.kind === "local_buffer_overflow") {
+					lines.push(
+						"This is a local gjc staging-buffer limit, not a provider or context-window failure; the same request reproduces it.",
+					);
+				} else if (local.kind === "local_snapshot_failure") {
+					lines.push(
+						"This is a local gjc event-serialization defect, not a provider failure; re-issuing is safe to retry.",
+					);
+				}
+			}
 			if (snapshot.assignment) lines.push("Assignment:", "```", snapshot.assignment, "```");
 			if (snapshot.steerMessage) {
 				lines.push(`Steer (${snapshot.steerState ?? "queued"}):`, "```", snapshot.steerMessage, "```");
@@ -770,7 +845,7 @@ export class SubagentTool implements AgentTool<typeof subagentSchema, SubagentTo
 		const progress = manager.getSubagentProgress(record.subagentId);
 		return {
 			liveProgressAvailable: true,
-			...(progress ? { progress } : {}),
+			...(progress ? { progress: toSubagentLiveProgress(progress) } : {}),
 		};
 	}
 
@@ -853,6 +928,14 @@ export class SubagentTool implements AgentTool<typeof subagentSchema, SubagentTo
 			...(job.setupFailureSummary
 				? { setupFailureSummary: sanitizeText(job.setupFailureSummary, RECEIPT_PREVIEW_WIDTH) }
 				: {}),
+			...(job.localErrorSummary
+				? {
+						localErrorSummary: {
+							kind: job.localErrorSummary.kind,
+							summary: sanitizeText(job.localErrorSummary.summary, RECEIPT_PREVIEW_WIDTH),
+						},
+					}
+				: {}),
 			...(outputRef ? { outputRef } : {}),
 			...(runningTimeoutGuidance ? { guidance: runningTimeoutGuidance } : {}),
 			...this.#modelFields(record),
@@ -1063,36 +1146,15 @@ function canonicalizeSnapshotForSignature(snapshot: SubagentSnapshot): unknown {
 	};
 }
 
-function canonicalizeProgressForSignature(progress: AgentProgress): unknown {
+function canonicalizeProgressForSignature(progress: SubagentLiveProgress): unknown {
 	return {
 		id: progress.id,
-		agent: progress.agent,
-		agentSource: progress.agentSource,
 		status: progress.status,
-		task: progress.task,
-		assignment: progress.assignment ?? null,
-		description: progress.description ?? null,
-		lastIntent: progress.lastIntent ?? null,
 		currentTool: progress.currentTool ?? null,
-		currentToolArgs: progress.currentToolArgs ?? null,
-		// currentToolStartMs intentionally excluded (only drives elapsed rendering).
-		recentTools: progress.recentTools.map(tool => ({ tool: tool.tool, args: tool.args })),
-		recentOutput: progress.recentOutput,
-		toolCount: progress.toolCount,
-		tokens: progress.tokens,
-		contextTokens: progress.contextTokens ?? null,
-		contextWindow: progress.contextWindow ?? null,
-		cost: progress.cost,
-		modelOverride: progress.modelOverride ?? null,
-		modelSubstitutionWarning: progress.modelSubstitutionWarning ?? null,
-		// The nested task panel renders this, so it must reach the signature or a
-		// fast-mode-only change would render differently while comparing byte-identical,
-		// suppressing the very update that introduces the glyph.
+		recentTool: progress.recentTool ?? null,
+		recentOutputSummary: progress.recentOutputSummary ?? null,
 		fastMode: progress.fastMode ?? false,
-		// durationMs intentionally excluded (time-derived).
-		extractedToolData: progress.extractedToolData
-			? canonicalizeExtractedToolDataForSignature(progress.extractedToolData)
-			: null,
+		retryFailure: progress.retryFailure ?? null,
 		retryState: progress.retryState
 			? {
 					attempt: progress.retryState.attempt,
@@ -1101,52 +1163,7 @@ function canonicalizeProgressForSignature(progress: AgentProgress): unknown {
 					kind: progress.retryState.kind,
 					provider: progress.retryState.provider ?? null,
 					delayMs: progress.retryState.delayMs,
-					errorMessage: progress.retryState.errorMessage,
-					// startedAtMs intentionally excluded (drives countdown only).
 				}
 			: null,
-		retryFailure: progress.retryFailure ?? null,
-		inflightTaskDetails: progress.inflightTaskDetails
-			? canonicalizeTaskDetailsForSignature(progress.inflightTaskDetails)
-			: null,
-	};
-}
-
-/**
- * Nested `task` data (`extractedToolData.task` and `inflightTaskDetails`) is the
- * one place the await signature reaches into a live, ticking structure: nested
- * `AgentProgress` carries the same time-derived fields excluded above, and
- * `TaskToolDetails` adds `totalDurationMs` / per-result `durationMs`. Signing it
- * wholesale would defeat idle gating whenever an awaited subagent is itself inside
- * a live `task` call, so these helpers canonicalize the rendered, non-time subset
- * recursively (mutually recursive with `canonicalizeProgressForSignature`).
- */
-function canonicalizeExtractedToolDataForSignature(data: Record<string, unknown[]>): Record<string, unknown> {
-	const out: Record<string, unknown> = {};
-	for (const key of Object.keys(data)) {
-		// Only the `task` key holds time-ticking `TaskToolDetails`; other handler
-		// data (yield/report_finding/generic) is stable and passes through as-is.
-		out[key] = key === "task" ? (data[key] as TaskToolDetails[]).map(canonicalizeTaskDetailsForSignature) : data[key];
-	}
-	return out;
-}
-
-function canonicalizeTaskDetailsForSignature(details: TaskToolDetails): unknown {
-	// `extractedToolData` is an untyped boundary (`Record<string, unknown[]>`), so
-	// guard each field instead of trusting the `TaskToolDetails` cast.
-	return {
-		// totalDurationMs intentionally excluded (time-derived).
-		results: Array.isArray(details.results) ? details.results.map(canonicalizeTaskResultForSignature) : null,
-		progress: Array.isArray(details.progress) ? details.progress.map(canonicalizeProgressForSignature) : null,
-		async: details.async
-			? { state: details.async.state, jobId: details.async.jobId, type: details.async.type }
-			: null,
 	};
 }
-
-function canonicalizeTaskResultForSignature(result: TaskToolDetails["results"][number]): unknown {
-	// Completed results do not tick, but drop `durationMs` so the only time-derived
-	// field in the receipt can never reintroduce idle churn.
-	const { durationMs: _durationMs, ...rest } = result;
-	return rest;
-}

package/src/tools/tool-catalog.generated.ts

@@ -681,6 +681,35 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 		"summary": "Execute Python or JavaScript code in an in-process eval backend",
 		"concurrency": "exclusive"
 	},
+	"python": {
+		"name": "python",
+		"label": "Python",
+		"description": "Execute Python in a persistent per-session REPL kernel.\n\nVariables, imports, and loaded data persist across calls in the current GJC session. The kernel runs in the session working directory. Each `execute` call is appended to a JSONL transcript under `.gjc/_session-{sessionid}/ipykernels/`; display artifacts are stored under `.gjc/_session-{sessionid}/ipykernels/artifacts/`.\n\n## Actions\n\n- `execute` (default) — run `code` in the persistent REPL. Requires `code`.\n- `clear` — dispose this session's Python kernel. The next `execute` starts a fresh kernel with no retained state.\n\n## Use\n\nUse this tool for stateful Python work. It is distinct from `eval`: each has a separate kernel and owner, so state is not shared between them.\n",
+		"parameters": {
+			"type": "object",
+			"properties": {
+				"action": {
+					"default": "execute",
+					"description": "\"execute\" runs `code` in the persistent per-session REPL and is the default. \"clear\" disposes this session's kernel; the next execute starts a fresh kernel.",
+					"type": "string",
+					"enum": [
+						"execute",
+						"clear"
+					]
+				},
+				"code": {
+					"description": "Python source to execute when action is \"execute\" (required then, ignored for \"clear\").",
+					"type": "string"
+				}
+			},
+			"additionalProperties": false
+		},
+		"strict": true,
+		"deferrable": true,
+		"loadMode": "discoverable",
+		"summary": "Execute Python in a persistent per-session REPL kernel (every call is appended to the session transcript)",
+		"concurrency": "exclusive"
+	},
 	"calc": {
 		"name": "calc",
 		"label": "Calc",
@@ -2058,6 +2087,15 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 								"type": "string",
 								"description": "per-task instructions; self-contained"
 							},
+							"tier": {
+								"description": "Advisory unless autorouting is enabled; omitted routes as balanced.",
+								"type": "string",
+								"enum": [
+									"fast",
+									"balanced",
+									"strong"
+								]
+							},
 							"executionMode": {
 								"description": "typed executor mode: default keeps ordinary executor behavior; ultragoal-red-team injects the Ultragoal QA/red-team prompt fragment. Prefer this over free-form assignment text (#2698).",
 								"type": "string",
@@ -2708,7 +2746,7 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 	"skill_discovery": {
 		"name": "skill_discovery",
 		"label": "SkillDiscovery",
-		"description": "Discover project and user runtime skills without loading full skill content.\n\n<instruction>\n- Searches canonical GJC skill locations in precedence order: project `.gjc/skills` (ancestors from cwd to repo root, closest first), then user locations under the home directory: canonical `<config>/agent/skills`, configured legacy `<config>/skills`, and historical legacy `.gjc/skills`. `<config>` is the home-relative directory name from `GJC_CONFIG_DIR`, then `PI_CONFIG_DIR`, then `.gjc`; even an absolute-looking configured name is joined beneath `<home>`. Project scope shadows user scope; within a scope, earlier locations above win. Bundled GJC workflow skills (`deep-interview`, `ralplan`, `team`, `ultragoal`) are always available and cannot be replaced by filesystem skills.\n- Returns thin metadata only: name, description, source scope, path, and use conditions when present.\n- Claude Code (`.claude/skills`) and Codex (`.codex/skills`) layouts are explicit import sources into `.gjc`, never invokable candidates. They are not returned as candidates; instead, each convention skill found in a trusted scope is reported in `diagnostics` with the exact copy command that enables it (copy into `.gjc/skills`), so a skill placed in a documented convention location is discoverable in a normal session without being silently loaded.\n- Discovery is on by default in a normal session. When zero candidates are returned because discovery config is disabled (`skills.enabled` master switch, or `skills.trustProjectSkills` / `skills.trustUserSkills` scope trust), the result carries a `notice` explaining which setting blocked the search — an empty result without a `notice` means the searched scopes genuinely contain no matching skills.\n- When skills were scanned but not advertised (protected-name collision with a bundled workflow skill, include/ignore/disable policy filters, invalid frontmatter, shadowing), the result carries a bounded `diagnostics` list explaining why.\n- To load a selected skill's full `SKILL.md`, invoke it through the existing `skill` tool with the exact `name` returned here.\n</instruction>\n\nInput:\n- `query` (optional): words to match against skill name, description, source, or use conditions.\n- `source` (optional): `all`, `project`, or `user`.\n- `limit` (optional): maximum results, 1-50.",
+		"description": "Discover project and user runtime skills without loading full skill content.\n\n<instruction>\n- Searches canonical GJC skill locations in precedence order: project `.gjc/skills` (ancestors from cwd to repo root, closest first), then user locations under the home directory: canonical `<config>/agent/skills`, configured legacy `<config>/skills`, and historical legacy `.gjc/skills`. `<config>` is the home-relative directory name from `GJC_CONFIG_DIR`, then `PI_CONFIG_DIR`, then `.gjc`; even an absolute-looking configured name is joined beneath `<home>`. Project scope shadows user scope; within a scope, earlier locations above win. Bundled GJC workflow skills (`autoresearch`, `deep-interview`, `ralplan`, `ultragoal`) are always available and cannot be replaced by filesystem skills.\n- Returns thin metadata only: name, description, source scope, path, and use conditions when present.\n- Claude Code (`.claude/skills`) and Codex (`.codex/skills`) layouts are explicit import sources into `.gjc`, never invokable candidates. They are not returned as candidates; instead, each convention skill found in a trusted scope is reported in `diagnostics` with the exact copy command that enables it (copy into `.gjc/skills`), so a skill placed in a documented convention location is discoverable in a normal session without being silently loaded.\n- Discovery is on by default in a normal session. When zero candidates are returned because discovery config is disabled (`skills.enabled` master switch, or `skills.trustProjectSkills` / `skills.trustUserSkills` scope trust), the result carries a `notice` explaining which setting blocked the search — an empty result without a `notice` means the searched scopes genuinely contain no matching skills.\n- When skills were scanned but not advertised (protected-name collision with a bundled workflow skill, include/ignore/disable policy filters, invalid frontmatter, shadowing), the result carries a bounded `diagnostics` list explaining why.\n- To load a selected skill's full `SKILL.md`, invoke it through the existing `skill` tool with the exact `name` returned here.\n</instruction>\n\nInput:\n- `query` (optional): words to match against skill name, description, source, or use conditions.\n- `source` (optional): `all`, `project`, or `user`.\n- `limit` (optional): maximum results, 1-50.",
 		"parameters": {
 			"type": "object",
 			"properties": {
@@ -2799,7 +2837,7 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 	"skill": {
 		"name": "skill",
 		"label": "Skill",
-		"description": "Invoke another available skill in the current turn.\n\n<conditions>\n- A SKILL document instructs you to chain into another skill on completion (e.g. ralplan → ultragoal)\n- You finished one skill's workflow and the next step requires another skill's full prompt context\n</conditions>\n\n<instruction>\n- `name` is the skill name as it appears in `/skill:<name>` (e.g. `ralplan`, `ultragoal`, `team`, `deep-interview`)\n- `args` is the free-form argument string the skill would receive after `/skill:<name>` on the command line\n- The tool loads the callee's SKILL.md into the current turn and handles native workflow caller→callee state handoff when the caller is one of the built-in GJC workflows.\n- The chain is refused while a native workflow caller is still active. If your current skill is one of `deep-interview`, `ralplan`, `ultragoal`, or `team` and has not yet reached a terminal phase, prepare it first with `gjc state <skill> write --input '{\"current_phase\":\"handoff\"}' --json`; no other handoff command is needed. Runtime project/user skills do not use `gjc state <skill>`.\n- Call once per chain step. To chain `A → B → C`, A calls `skill(B)`; B's next agent turn calls `skill(C)`.\n</instruction>\n\n<critical>\n- Do NOT use this tool to \"remind yourself\" of a skill you're already running. The current SKILL.md is already in your context.\n- Do NOT chain into the same skill recursively. If a skill's flow needs another iteration, follow its in-document instructions.\n- `name` MUST be one concrete skill name, NOT a glob or wildcard. Passing `*`, `?`, or a pattern like `git-*` is rejected immediately — the `--skills '*'` launch filter is unrelated to this tool's `name`.\n- The chained skill's planning/execution-boundary rules still apply. Chaining does not grant execution approval.\n</critical>\n\n<examples>\n# Hand off from ralplan to ultragoal after an approved plan\n{\"name\": \"ultragoal\", \"args\": \"track execution of .gjc/plans/ralplan/<run-id>/pending-approval.md\"}\n\n# Trigger deep-interview with no arguments\n{\"name\": \"deep-interview\"}\n</examples>",
+		"description": "Invoke another available skill in the current turn.\n\n<conditions>\n- A SKILL document instructs you to chain into another skill on completion (e.g. ralplan → ultragoal)\n- You finished one skill's workflow and the next step requires another skill's full prompt context\n</conditions>\n\n<instruction>\n- `name` is the skill name as it appears in `/skill:<name>` (e.g. `ralplan`, `ultragoal`, `autoresearch`, `deep-interview`)\n- `args` is the free-form argument string the skill would receive after `/skill:<name>` on the command line\n- The tool loads the callee's SKILL.md into the current turn and handles native workflow caller→callee state handoff when the caller is one of the built-in GJC workflows.\n- The chain is refused while a native workflow caller is still mid-flight. `autoresearch` chains from any of its phases (`intake`/`research`/`verdict`) — a research mission is always handoff-ready. `deep-interview` chains once its final spec is persisted (phase `handoff`), and `ralplan` chains from `final` or `handoff`. Only a mid-flight `ralplan` or `ultragoal` needs preparation first: `gjc state <skill> write --input '{\"current_phase\":\"handoff\"}' --json`; no other handoff command is needed. Runtime project/user skills do not use `gjc state <skill>`.\n- Call once per chain step. To chain `A → B → C`, A calls `skill(B)`; B's next agent turn calls `skill(C)`.\n</instruction>\n\n<critical>\n- Do NOT use this tool to \"remind yourself\" of a skill you're already running. The current SKILL.md is already in your context.\n- Do NOT chain into the same skill recursively. If a skill's flow needs another iteration, follow its in-document instructions.\n- `name` MUST be one concrete skill name, NOT a glob or wildcard. Passing `*`, `?`, or a pattern like `git-*` is rejected immediately — the `--skills '*'` launch filter is unrelated to this tool's `name`.\n- The chained skill's planning/execution-boundary rules still apply. Chaining does not grant execution approval.\n</critical>\n\n<examples>\n# Hand off from ralplan to ultragoal after an approved plan\n{\"name\": \"ultragoal\", \"args\": \"track execution of .gjc/plans/ralplan/<run-id>/pending-approval.md\"}\n\n# Trigger deep-interview with no arguments\n{\"name\": \"deep-interview\"}\n</examples>",
 		"parameters": {
 			"type": "object",
 			"properties": {
@@ -2856,6 +2894,30 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 		"loadMode": "essential",
 		"intent": "omit"
 	},
+	"move_session": {
+		"name": "move_session",
+		"label": "Move Session",
+		"description": "Rescope the session to a narrower working directory.\n\nUse this only when the session's working directory is a broad launcher root (for example a\nmulti-repo workspace like `~/Projects`) and the task has clearly converged on one subdirectory\nor repository: after this call, every later turn resolves relative paths and the bash default\ncwd from the new directory, and project-scoped plugins/capabilities reload for it.\n\n- `path` must be an existing directory; relative paths resolve against the current session cwd.\n  The canonical target must be strictly inside the current session directory — moves to a\n  parent, a sibling project, or an unrelated absolute path are refused.\n- A session can be moved this way at most once, and never while another move is running; a\n  rejected call does not consume the move. Use it once the target repo is identified — not\n  speculatively — because the session file and caches move with the session.\n- This tool is unavailable in subagent sessions and restricted profiles; ask the top-level\n  session to rescope instead.",
+		"parameters": {
+			"type": "object",
+			"properties": {
+				"path": {
+					"type": "string",
+					"description": "target directory: absolute, or relative to the current session cwd"
+				}
+			},
+			"required": [
+				"path"
+			],
+			"additionalProperties": false
+		},
+		"strict": true,
+		"deferrable": false,
+		"loadMode": "essential",
+		"nonAbortable": true,
+		"concurrency": "exclusive",
+		"intent": "omit"
+	},
 	"yield": {
 		"name": "yield",
 		"label": "Submit Result",

package/src/tools/write.ts

@@ -1,6 +1,5 @@
 import { Database } from "bun:sqlite";
 import * as fs from "node:fs/promises";
-import * as path from "node:path";
 import type { AgentTool, AgentToolContext, AgentToolResult, AgentToolUpdateCallback } from "@gajae-code/agent-core";
 import type { Component } from "@gajae-code/tui";
 import { Text } from "@gajae-code/tui";
@@ -17,6 +16,7 @@ import type { ToolSession } from "../sdk";
 import { Ellipsis, Hasher, type RenderCache, renderStatusLine, truncateToWidth } from "../tui";
 import { resolveFileDisplayMode } from "../utils/file-display-mode";
 import { parseArchivePathCandidates } from "./archive-reader";
+import { formatFileWriteError, writeFileAtomically } from "./atomic-file-write";
 import { assertEditableFile } from "./auto-generated-guard";
 import {
 	type ConflictEntry,
@@ -249,11 +249,6 @@ export class WriteTool implements AgentTool<typeof writeSchema, WriteToolDetails
 	): Promise<AgentToolResult<WriteToolDetails>> {
 		const isZip = resolvedArchivePath.absolutePath.toLowerCase().endsWith(".zip");
 
-		const parentDir = path.dirname(resolvedArchivePath.absolutePath);
-		if (parentDir && parentDir !== ".") {
-			await fs.mkdir(parentDir, { recursive: true });
-		}
-
 		if (isZip) {
 			const zipEntries: Record<string, Uint8Array> = {};
 
@@ -275,9 +270,9 @@ export class WriteTool implements AgentTool<typeof writeSchema, WriteToolDetails
 			try {
 				const { zipSync } = await loadFflate();
 				const zipBuffer = zipSync(zipEntries);
-				await Bun.write(resolvedArchivePath.absolutePath, zipBuffer);
+				await writeFileAtomically(resolvedArchivePath.absolutePath, zipBuffer);
 			} catch (error) {
-				throw new ToolError(error instanceof Error ? error.message : String(error));
+				throw new ToolError(formatFileWriteError(error, resolvedArchivePath.absolutePath, { destUnchanged: true }));
 			}
 		} else {
 			const archiveEntries: Record<string, string | File> = {};
@@ -304,9 +299,11 @@ export class WriteTool implements AgentTool<typeof writeSchema, WriteToolDetails
 			archiveEntries[resolvedArchivePath.archiveSubPath] = content;
 
 			try {
-				await Bun.Archive.write(resolvedArchivePath.absolutePath, archiveEntries);
+				const archive = new Bun.Archive(archiveEntries);
+				const archiveBytes = await archive.bytes();
+				await writeFileAtomically(resolvedArchivePath.absolutePath, archiveBytes);
 			} catch (error) {
-				throw new ToolError(error instanceof Error ? error.message : String(error));
+				throw new ToolError(formatFileWriteError(error, resolvedArchivePath.absolutePath, { destUnchanged: true }));
 			}
 		}
 
@@ -748,9 +745,10 @@ export class WriteTool implements AgentTool<typeof writeSchema, WriteToolDetails
 				try {
 					await bridgePromise;
 				} catch (error) {
-					throw new ToolError(error instanceof Error ? error.message : String(error));
+					throw new ToolError(formatFileWriteError(error, absolutePath));
 				}
 				invalidateFsScanAfterWrite(absolutePath);
+				this.session.fileReadCache?.invalidate(absolutePath);
 				const displayPath = formatPathRelativeToCwd(absolutePath, this.session.cwd);
 				let resultText = `Successfully wrote ${cleanContent.length} bytes to ${displayPath}`;
 				if (stripped) {
@@ -759,8 +757,14 @@ export class WriteTool implements AgentTool<typeof writeSchema, WriteToolDetails
 				return { content: [{ type: "text", text: resultText }], details: {} };
 			}
 
-			const diagnostics = await this.#writethrough(absolutePath, cleanContent, signal, undefined, batchRequest);
+			let diagnostics: FileDiagnosticsResult | undefined;
+			try {
+				diagnostics = await this.#writethrough(absolutePath, cleanContent, signal, undefined, batchRequest);
+			} catch (error) {
+				throw new ToolError(formatFileWriteError(error, absolutePath));
+			}
 			invalidateFsScanAfterWrite(absolutePath);
+			this.session.fileReadCache?.invalidate(absolutePath);
 
 			const displayPath = formatPathRelativeToCwd(absolutePath, this.session.cwd);
 			let resultText = `Successfully wrote ${cleanContent.length} bytes to ${displayPath}`;

package/src/utils/xml-escape.ts

@@ -0,0 +1,14 @@
+/** Escape a string for interpolation inside a double- or single-quoted XML attribute. */
+export function escapeXmlAttribute(input: string): string {
+	if (!/[&<>"']/.test(input)) return input;
+	let output = "";
+	for (const char of input) {
+		if (char === "&") output += "&amp;";
+		else if (char === "<") output += "&lt;";
+		else if (char === ">") output += "&gt;";
+		else if (char === '"') output += "&quot;";
+		else if (char === "'") output += "&apos;";
+		else output += char;
+	}
+	return output;
+}

package/src/web/search/providers/codex.ts

@@ -9,7 +9,7 @@
 import * as os from "node:os";
 import { type AuthStorage, getBundledModels } from "@gajae-code/ai/core";
 import { decodeJwt } from "@gajae-code/ai/utils/oauth/openai-codex";
-import { $env, readSseJson } from "@gajae-code/utils";
+import { $env, readSseJson, sanitizeHeaderComponent } from "@gajae-code/utils";
 import packageJson from "../../../../package.json" with { type: "json" };
 import type { SearchResponse, SearchSource } from "../../../web/search/types";
 import { SearchProviderError } from "../../../web/search/types";
@@ -158,7 +158,7 @@ function buildCodexHeaders(accessToken: string, accountId: string): Record<strin
 		"chatgpt-account-id": accountId,
 		"OpenAI-Beta": "responses=experimental",
 		originator: "pi",
-		"User-Agent": `pi/${packageJson.version} (${os.platform()} ${os.release()}; ${os.arch()})`,
+		"User-Agent": `pi/${packageJson.version} (${sanitizeHeaderComponent(os.platform())} ${sanitizeHeaderComponent(os.release())}; ${sanitizeHeaderComponent(os.arch())})`,
 		Accept: "text/event-stream",
 		"Content-Type": "application/json",
 	};

package/src/workflow/workflow-intent-diff.ts

@@ -5,8 +5,8 @@ import {
 } from "./workflow-intent-report";
 
 export const WORKFLOW_INTENT_DIFF_CUSTOM_TYPE = "workflow-intent-diff";
-export const WORKFLOW_INTENT_ROUTES = ["direct", "deep-interview", "ralplan", "ultragoal", "team"] as const;
-export const WORKFLOW_ESCALATION_ROUTES = ["deep-interview", "ralplan", "ultragoal", "team"] as const;
+export const WORKFLOW_INTENT_ROUTES = ["direct", "deep-interview", "ralplan", "ultragoal", "autoresearch"] as const;
+export const WORKFLOW_ESCALATION_ROUTES = ["deep-interview", "ralplan", "ultragoal", "autoresearch"] as const;
 
 export type WorkflowIntentRoute = (typeof WORKFLOW_INTENT_ROUTES)[number];
 export type WorkflowEscalationRoute = (typeof WORKFLOW_ESCALATION_ROUTES)[number];
@@ -77,11 +77,12 @@ const ARCHITECTURE_SEQUENCE_PATTERNS = [
 	/\bdestructive\b/i,
 ] as const;
 
-const TEAM_PATTERNS = [
-	/\buse (?:a )?team\b/i,
-	/\bcoordinated (?:persistent )?workers\b/i,
-	/\bpersistent workers\b/i,
-	/\bworker coordination\b/i,
+const AUTORESEARCH_PATTERNS = [
+	/\buse (?:a )?autoresearch\b/i,
+	/\b(?:research|investigate|experiment|benchmark|measure)\b[^.!?\n]{0,60}\b(?:mission|finding|evidence|verdict|dataset|against|compare|hypothesis)\b/i,
+	/\b(?:mission|finding|evidence|verdict|dataset|against|compare|hypothesis)\b[^.!?\n]{0,60}\b(?:research|investigate|experiment|benchmark|measure)\b/i,
+	/\brun (?:a |the )?(?:bounded )?research (?:mission|task|investigation)\b/i,
+	/\b(?:benchmark|experiment)\b[^.!?\n]{0,40}\b(?:and|then|to)\b[^.!?\n]{0,40}\b(?:report|measure|compare)\b/i,
 ] as const;
 
 const ROOT_CAUSE_TRIGGERS = [
@@ -172,12 +173,12 @@ function classifyRoute(text: string): RouteMatch {
 		};
 	}
 
-	if (matchesAny(text, TEAM_PATTERNS)) {
+	if (matchesAny(text, AUTORESEARCH_PATTERNS)) {
 		return {
-			route: "team",
-			reason: "coordinated persistent workers requested",
-			recommendedInvocation: "/skill:team",
-			triggers: ["coordinated workers"],
+			route: "autoresearch",
+			reason: "goal-directed research mission requested",
+			recommendedInvocation: "/skill:autoresearch",
+			triggers: ["research mission"],
 		};
 	}
 

package/src/workflow/workflow-intent-report-consumer.ts

@@ -15,7 +15,7 @@ export interface WorkflowRouteCounts {
 	readonly "deep-interview": number;
 	readonly ralplan: number;
 	readonly ultragoal: number;
-	readonly team: number;
+	readonly autoresearch: number;
 }
 
 export interface WorkflowIntentReportEntry {
@@ -66,7 +66,7 @@ function emptyRouteCounts(): WorkflowRouteCounts {
 		"deep-interview": 0,
 		ralplan: 0,
 		ultragoal: 0,
-		team: 0,
+		autoresearch: 0,
 	};
 }
 
@@ -87,8 +87,8 @@ function incrementRoute(counts: WorkflowRouteCounts, route: WorkflowIntentRoute)
 			return { ...counts, ralplan: counts.ralplan + 1 };
 		case "ultragoal":
 			return { ...counts, ultragoal: counts.ultragoal + 1 };
-		case "team":
-			return { ...counts, team: counts.team + 1 };
+		case "autoresearch":
+			return { ...counts, autoresearch: counts.autoresearch + 1 };
 	}
 }
 
@@ -102,8 +102,8 @@ function routeCount(counts: WorkflowRouteCounts, route: WorkflowIntentRoute): nu
 			return counts.ralplan;
 		case "ultragoal":
 			return counts.ultragoal;
-		case "team":
-			return counts.team;
+		case "autoresearch":
+			return counts.autoresearch;
 	}
 }
 

package/dist/types/autoresearch/helpers.d.ts

@@ -1,24 +0,0 @@
-import type { ASIData, MetricDirection, NumericMetricMap } from "./types";
-export declare const METRIC_LINE_PREFIX = "METRIC";
-export declare const ASI_LINE_PREFIX = "ASI";
-export declare const EXPERIMENT_MAX_LINES = 10;
-export declare const EXPERIMENT_MAX_BYTES: number;
-export declare function parseMetricLines(output: string): Map<string, number>;
-export declare function parseAsiLines(output: string): ASIData | null;
-export declare function mergeAsi(base: ASIData | null, override: ASIData | undefined): ASIData | undefined;
-export declare function commas(value: number): string;
-export declare function fmtNum(value: number, decimals?: number): string;
-export declare function formatNum(value: number | null, unit: string): string;
-export declare function formatElapsed(milliseconds: number): string;
-export declare function killTree(pid: number, signal?: NodeJS.Signals | number): void;
-export declare function isBetter(current: number, best: number, direction: MetricDirection): boolean;
-export declare function inferMetricUnitFromName(name: string): string;
-export declare function normalizePathSpec(value: string): string;
-export declare function pathMatchesSpec(pathValue: string, specValue: string): boolean;
-export declare function dedupeStrings(values: readonly string[]): string[];
-export declare function ensureNumericMetricMap(value: NumericMetricMap | undefined): NumericMetricMap;
-export declare function sanitizeAsi(value: {
-    [key: string]: unknown;
-} | undefined): ASIData | undefined;
-export declare function tryGitStatus(cwd: string): Promise<string>;
-export declare function tryGitPrefix(cwd: string): Promise<string>;

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

@@ -1,2 +0,0 @@
-import type { ExtensionFactory } from "../extensibility/extensions";
-export declare const createAutoresearchExtension: ExtensionFactory;

package/dist/types/autoresearch/state.d.ts

@@ -1,17 +0,0 @@
-import type { SessionEntry } from "../session/session-manager";
-import type { RunRow, SessionRow } from "./storage";
-import type { AutoresearchRuntime, ExperimentResult, ExperimentState, MetricDef, MetricDirection, NumericMetricMap, ReconstructedControlState, RuntimeStore } from "./types";
-export declare function createExperimentState(): ExperimentState;
-export declare function createSessionRuntime(): AutoresearchRuntime;
-export declare function cloneExperimentState(state: ExperimentState): ExperimentState;
-export declare function currentResults(results: ExperimentResult[], segment: number): ExperimentResult[];
-export declare function findBaselineResult(results: ExperimentResult[], segment: number): ExperimentResult | null;
-export declare function findBaselineMetric(results: ExperimentResult[], segment: number): number | null;
-export declare function findBestKeptMetric(results: ExperimentResult[], segment: number, direction: MetricDirection): number | null;
-export declare function findBaselineRunNumber(results: ExperimentResult[], segment: number): number | null;
-export declare function findBaselineSecondary(results: ExperimentResult[], segment: number, knownMetrics: MetricDef[]): NumericMetricMap;
-export declare function sortedMedian(values: number[]): number;
-export declare function computeConfidence(results: ExperimentResult[], segment: number, direction: MetricDirection): number | null;
-export declare function buildExperimentState(session: SessionRow, loggedRuns: RunRow[]): ExperimentState;
-export declare function reconstructControlState(entries: SessionEntry[]): ReconstructedControlState;
-export declare function createRuntimeStore(): RuntimeStore;