fix: resolve 25 pre-publish blockers

- postinstall.mjs: fix alias package detection
- migrate-legacy-plugin-entry: dedupe + regression tests
- task_system: default consistency across runtime paths
- task() contract: consistent tool behavior
- runtime model selection, tool cap, stale-task cancellation
- recovery sanitization, context-limit gating
- Ralph semantic DONE hardening, Atlas fallback persistence
- native-skill description/content, skill path traversal guard
- publish workflow: platform awaited via reusable workflow job
- release: version edits reapplied before commit/tag
- JSONC plugin migration: top-level plugin key safety
- cold-cache: user fallback models skip disconnected providers
- docs/version/release framing updates

Verified: bun test (4599 pass), tsc --noEmit clean, bun run build clean

src/features/background-agent/manager.test.ts

@@ -3312,6 +3312,9 @@ describe("BackgroundManager.checkAndInterruptStaleTasks", () => {
         prompt: async () => ({}),
         promptAsync: async () => ({}),
         abort: async () => ({}),
+        get: async () => {
+          throw new Error("missing")
+        },
       },
     }
     const manager = new BackgroundManager({ client, directory: tmpdir() } as unknown as PluginInput, { staleTimeoutMs: 180_000 })
@@ -3348,6 +3351,9 @@ describe("BackgroundManager.checkAndInterruptStaleTasks", () => {
         prompt: async () => ({}),
         promptAsync: async () => ({}),
         abort: async () => ({}),
+        get: async () => {
+          throw new Error("missing")
+        },
       },
     }
     const manager = new BackgroundManager({ client, directory: tmpdir() } as unknown as PluginInput, { staleTimeoutMs: 180_000 })
@@ -3437,6 +3443,7 @@ describe("BackgroundManager.checkAndInterruptStaleTasks", () => {
       status: "running",
       startedAt: new Date(Date.now() - 15 * 60 * 1000),
       progress: undefined,
+      consecutiveMissedPolls: 2,
     }
 
     getTaskMap(manager).set(task.id, task)
@@ -3471,6 +3478,7 @@ describe("BackgroundManager.checkAndInterruptStaleTasks", () => {
       status: "running",
       startedAt: new Date(Date.now() - 15 * 60 * 1000),
       progress: undefined,
+      consecutiveMissedPolls: 2,
     }
 
     getTaskMap(manager).set(task.id, task)

src/features/background-agent/task-poller.test.ts

@@ -8,6 +8,7 @@ describe("checkAndInterruptStaleTasks", () => {
   const mockClient = {
     session: {
       abort: mock(() => Promise.resolve()),
+      get: mock(() => Promise.resolve({ data: { id: "ses-1" } })),
     },
   }
   const mockConcurrencyManager = {
@@ -35,6 +36,11 @@ describe("checkAndInterruptStaleTasks", () => {
   beforeEach(() => {
     fixedTime = Date.now()
     spyOn(globalThis.Date, "now").mockReturnValue(fixedTime)
+    mockClient.session.abort.mockClear()
+    mockClient.session.get.mockReset()
+    mockClient.session.get.mockResolvedValue({ data: { id: "ses-1" } })
+    mockConcurrencyManager.release.mockClear()
+    mockNotify.mockClear()
   })
 
   afterEach(() => {
@@ -288,6 +294,59 @@ describe("checkAndInterruptStaleTasks", () => {
     expect(task.status).toBe("running")
   })
 
+  it("should NOT cancel healthy task on first missing status poll", async () => {
+    //#given — one missing poll should not be enough to declare the session gone
+    const task = createRunningTask({
+      startedAt: new Date(Date.now() - 300_000),
+      progress: {
+        toolCalls: 1,
+        lastUpdate: new Date(Date.now() - 120_000),
+      },
+    })
+
+    //#when
+    await checkAndInterruptStaleTasks({
+      tasks: [task],
+      client: mockClient as never,
+      config: { staleTimeoutMs: 180_000, sessionGoneTimeoutMs: 60_000 },
+      concurrencyManager: mockConcurrencyManager as never,
+      notifyParentSession: mockNotify,
+      sessionStatuses: {},
+    })
+
+    //#then
+    expect(task.status).toBe("running")
+    expect(task.consecutiveMissedPolls).toBe(1)
+    expect(mockClient.session.get).not.toHaveBeenCalled()
+  })
+
+  it("should NOT cancel task when session.get confirms the session still exists", async () => {
+    //#given — repeated missing polls but direct lookup still succeeds
+    const task = createRunningTask({
+      startedAt: new Date(Date.now() - 300_000),
+      progress: {
+        toolCalls: 1,
+        lastUpdate: new Date(Date.now() - 120_000),
+      },
+      consecutiveMissedPolls: 2,
+    })
+
+    //#when
+    await checkAndInterruptStaleTasks({
+      tasks: [task],
+      client: mockClient as never,
+      config: { staleTimeoutMs: 180_000, sessionGoneTimeoutMs: 60_000 },
+      concurrencyManager: mockConcurrencyManager as never,
+      notifyParentSession: mockNotify,
+      sessionStatuses: {},
+    })
+
+    //#then
+    expect(task.status).toBe("running")
+    expect(task.consecutiveMissedPolls).toBe(0)
+    expect(mockClient.session.get).toHaveBeenCalledWith({ path: { id: "ses-1" } })
+  })
+
   it("should use session-gone timeout when session is missing from status map (with progress)", async () => {
     //#given — lastUpdate 2min ago, session completely gone from status
     const task = createRunningTask({
@@ -296,8 +355,11 @@ describe("checkAndInterruptStaleTasks", () => {
         toolCalls: 1,
         lastUpdate: new Date(Date.now() - 120_000),
       },
+      consecutiveMissedPolls: 2,
     })
 
+    mockClient.session.get.mockRejectedValue(new Error("missing"))
+
     //#when — empty sessionStatuses (session gone), sessionGoneTimeoutMs = 60s
     await checkAndInterruptStaleTasks({
       tasks: [task],
@@ -318,8 +380,11 @@ describe("checkAndInterruptStaleTasks", () => {
     const task = createRunningTask({
       startedAt: new Date(Date.now() - 120_000),
       progress: undefined,
+      consecutiveMissedPolls: 2,
     })
 
+    mockClient.session.get.mockRejectedValue(new Error("missing"))
+
     //#when — session gone, sessionGoneTimeoutMs = 60s
     await checkAndInterruptStaleTasks({
       tasks: [task],
@@ -343,8 +408,11 @@ describe("checkAndInterruptStaleTasks", () => {
         toolCalls: 1,
         lastUpdate: new Date(Date.now() - 120_000),
       },
+      consecutiveMissedPolls: 2,
     })
 
+    mockClient.session.get.mockRejectedValue(new Error("missing"))
+
     //#when — session is idle (present in map), staleTimeoutMs = 180s
     await checkAndInterruptStaleTasks({
       tasks: [task],
@@ -367,8 +435,11 @@ describe("checkAndInterruptStaleTasks", () => {
         toolCalls: 1,
         lastUpdate: new Date(Date.now() - 120_000),
       },
+      consecutiveMissedPolls: 2,
     })
 
+    mockClient.session.get.mockRejectedValue(new Error("missing"))
+
     //#when — no config (default sessionGoneTimeoutMs = 60_000)
     await checkAndInterruptStaleTasks({
       tasks: [task],

src/features/background-agent/task-poller.ts

@@ -16,6 +16,8 @@ import {
 import { removeTaskToastTracking } from "./remove-task-toast-tracking"
 
 import { isActiveSessionStatus } from "./session-status-classifier"
+
+const MIN_SESSION_GONE_POLLS = 3
 const TERMINAL_TASK_STATUSES = new Set<BackgroundTask["status"]>([
   "completed",
   "error",
@@ -97,6 +99,15 @@ export function pruneStaleTasksAndNotifications(args: {
 
 export type SessionStatusMap = Record<string, { type: string }>
 
+async function verifySessionExists(client: OpencodeClient, sessionID: string): Promise<boolean> {
+  try {
+    const result = await client.session.get({ path: { id: sessionID } })
+    return !!result.data
+  } catch {
+    return false
+  }
+}
+
 export async function checkAndInterruptStaleTasks(args: {
   tasks: Iterable<BackgroundTask>
   client: OpencodeClient
@@ -130,14 +141,28 @@ export async function checkAndInterruptStaleTasks(args: {
 
     const sessionStatus = sessionStatuses?.[sessionID]?.type
     const sessionIsRunning = sessionStatus !== undefined && isActiveSessionStatus(sessionStatus)
-    const sessionGone = sessionStatuses !== undefined && sessionStatus === undefined
+    const sessionMissing = sessionStatuses !== undefined && sessionStatus === undefined
     const runtime = now - startedAt.getTime()
 
+    if (sessionMissing) {
+      task.consecutiveMissedPolls = (task.consecutiveMissedPolls ?? 0) + 1
+    } else if (sessionStatuses !== undefined) {
+      task.consecutiveMissedPolls = 0
+    }
+
+    const sessionGone = sessionMissing && (task.consecutiveMissedPolls ?? 0) >= MIN_SESSION_GONE_POLLS
+
     if (!task.progress?.lastUpdate) {
       if (sessionIsRunning) continue
+      if (sessionMissing && !sessionGone) continue
       const effectiveTimeout = sessionGone ? sessionGoneTimeoutMs : messageStalenessMs
       if (runtime <= effectiveTimeout) continue
 
+      if (sessionGone && await verifySessionExists(client, sessionID)) {
+        task.consecutiveMissedPolls = 0
+        continue
+      }
+
       const staleMinutes = Math.round(runtime / 60000)
       const reason = sessionGone ? "session gone from status registry" : "no activity"
       task.status = "cancelled"
@@ -171,11 +196,16 @@ export async function checkAndInterruptStaleTasks(args: {
     if (timeSinceLastUpdate <= effectiveStaleTimeout) continue
     if (task.status !== "running") continue
 
-     const staleMinutes = Math.round(timeSinceLastUpdate / 60000)
-     const reason = sessionGone ? "session gone from status registry" : "no activity"
-     task.status = "cancelled"
-     task.error = `Stale timeout (${reason} for ${staleMinutes}min). This is a FINAL cancellation - do NOT create a replacement task. If the timeout is too short, increase 'background_task.${sessionGone ? "sessionGoneTimeoutMs" : "staleTimeoutMs"}' in .opencode/oh-my-opencode.json.`
-     task.completedAt = new Date()
+    if (sessionGone && await verifySessionExists(client, sessionID)) {
+      task.consecutiveMissedPolls = 0
+      continue
+    }
+
+    const staleMinutes = Math.round(timeSinceLastUpdate / 60000)
+    const reason = sessionGone ? "session gone from status registry" : "no activity"
+    task.status = "cancelled"
+    task.error = `Stale timeout (${reason} for ${staleMinutes}min). This is a FINAL cancellation - do NOT create a replacement task. If the timeout is too short, increase 'background_task.${sessionGone ? "sessionGoneTimeoutMs" : "staleTimeoutMs"}' in .opencode/oh-my-opencode.json.`
+    task.completedAt = new Date()
 
     if (task.concurrencyKey) {
       concurrencyManager.release(task.concurrencyKey)

src/features/background-agent/types.ts

@@ -66,6 +66,8 @@ export interface BackgroundTask {
   lastMsgCount?: number
   /** Number of consecutive polls with stable message count */
   stablePolls?: number
+  /** Number of consecutive polls where session was missing from status map */
+  consecutiveMissedPolls?: number
 }
 
 export interface LaunchInput {

src/features/opencode-skill-loader/loaded-skill-template-extractor.ts

@@ -3,9 +3,13 @@ import { parseFrontmatter } from "../../shared/frontmatter"
 import type { LoadedSkill } from "./types"
 
 export function extractSkillTemplate(skill: LoadedSkill): string {
-	if (skill.path) {
-		const content = readFileSync(skill.path, "utf-8")
-		const { body } = parseFrontmatter(content)
+  if (skill.scope === "config" && skill.definition.template) {
+    return skill.definition.template
+  }
+
+  if (skill.path) {
+    const content = readFileSync(skill.path, "utf-8")
+    const { body } = parseFrontmatter(content)
 		return body.trim()
 	}
 	return skill.definition.template || ""

src/hooks/anthropic-context-window-limit-recovery/message-builder.test.ts

@@ -0,0 +1,105 @@
+import { afterAll, beforeEach, describe, expect, mock, test } from "bun:test"
+
+const replaceEmptyTextPartsAsync = mock(() => Promise.resolve(false))
+const injectTextPartAsync = mock(() => Promise.resolve(false))
+const findMessagesWithEmptyTextPartsFromSDK = mock(() => Promise.resolve([] as string[]))
+
+mock.module("../../shared", () => ({
+  normalizeSDKResponse: (response: { data?: unknown[] }) => response.data ?? [],
+}))
+
+mock.module("../../shared/logger", () => ({
+  log: () => {},
+}))
+
+mock.module("../../shared/opencode-storage-detection", () => ({
+  isSqliteBackend: () => true,
+}))
+
+mock.module("../session-recovery/storage", () => ({
+  findEmptyMessages: () => [],
+  findMessagesWithEmptyTextParts: () => [],
+  injectTextPart: () => false,
+  replaceEmptyTextParts: () => false,
+}))
+
+mock.module("../session-recovery/storage/empty-text", () => ({
+  replaceEmptyTextPartsAsync,
+  findMessagesWithEmptyTextPartsFromSDK,
+}))
+
+mock.module("../session-recovery/storage/text-part-injector", () => ({
+  injectTextPartAsync,
+}))
+
+async function importFreshMessageBuilder(): Promise<typeof import("./message-builder")> {
+  return import(`./message-builder?test=${Date.now()}-${Math.random()}`)
+}
+
+afterAll(() => {
+  mock.restore()
+})
+
+describe("sanitizeEmptyMessagesBeforeSummarize", () => {
+  beforeEach(() => {
+    replaceEmptyTextPartsAsync.mockReset()
+    replaceEmptyTextPartsAsync.mockResolvedValue(false)
+    injectTextPartAsync.mockReset()
+    injectTextPartAsync.mockResolvedValue(false)
+    findMessagesWithEmptyTextPartsFromSDK.mockReset()
+    findMessagesWithEmptyTextPartsFromSDK.mockResolvedValue([])
+  })
+
+  test("#given sqlite message with tool content and empty text part #when sanitizing #then it fixes the mixed-content message", async () => {
+    const { sanitizeEmptyMessagesBeforeSummarize, PLACEHOLDER_TEXT } = await importFreshMessageBuilder()
+    const client = {
+      session: {
+        messages: mock(() => Promise.resolve({
+          data: [
+            {
+              info: { id: "msg-1" },
+              parts: [
+                { type: "tool_result", text: "done" },
+                { type: "text", text: "" },
+              ],
+            },
+          ],
+        })),
+      },
+    } as never
+    findMessagesWithEmptyTextPartsFromSDK.mockResolvedValue(["msg-1"])
+    replaceEmptyTextPartsAsync.mockResolvedValue(true)
+
+    const fixedCount = await sanitizeEmptyMessagesBeforeSummarize("ses-1", client)
+
+    expect(fixedCount).toBe(1)
+    expect(replaceEmptyTextPartsAsync).toHaveBeenCalledWith(client, "ses-1", "msg-1", PLACEHOLDER_TEXT)
+    expect(injectTextPartAsync).not.toHaveBeenCalled()
+  })
+
+  test("#given sqlite message with mixed content and failed replacement #when sanitizing #then it injects the placeholder text part", async () => {
+    const { sanitizeEmptyMessagesBeforeSummarize, PLACEHOLDER_TEXT } = await importFreshMessageBuilder()
+    const client = {
+      session: {
+        messages: mock(() => Promise.resolve({
+          data: [
+            {
+              info: { id: "msg-2" },
+              parts: [
+                { type: "tool_use", text: "call" },
+                { type: "text", text: "" },
+              ],
+            },
+          ],
+        })),
+      },
+    } as never
+    findMessagesWithEmptyTextPartsFromSDK.mockResolvedValue(["msg-2"])
+    injectTextPartAsync.mockResolvedValue(true)
+
+    const fixedCount = await sanitizeEmptyMessagesBeforeSummarize("ses-2", client)
+
+    expect(fixedCount).toBe(1)
+    expect(injectTextPartAsync).toHaveBeenCalledWith(client, "ses-2", "msg-2", PLACEHOLDER_TEXT)
+  })
+})

src/hooks/anthropic-context-window-limit-recovery/message-builder.ts

@@ -8,7 +8,7 @@ import {
   injectTextPart,
   replaceEmptyTextParts,
 } from "../session-recovery/storage"
-import { replaceEmptyTextPartsAsync } from "../session-recovery/storage/empty-text"
+import { findMessagesWithEmptyTextPartsFromSDK, replaceEmptyTextPartsAsync } from "../session-recovery/storage/empty-text"
 import { injectTextPartAsync } from "../session-recovery/storage/text-part-injector"
 import type { Client } from "./client"
 
@@ -86,12 +86,14 @@ export async function sanitizeEmptyMessagesBeforeSummarize(
 ): Promise<number> {
   if (client && isSqliteBackend()) {
     const emptyMessageIds = await findEmptyMessageIdsFromSDK(client, sessionID)
-    if (emptyMessageIds.length === 0) {
+    const emptyTextPartIds = await findMessagesWithEmptyTextPartsFromSDK(client, sessionID)
+    const allIds = [...new Set([...emptyMessageIds, ...emptyTextPartIds])]
+    if (allIds.length === 0) {
       return 0
     }
 
     let fixedCount = 0
-    for (const messageID of emptyMessageIds) {
+    for (const messageID of allIds) {
       const replaced = await replaceEmptyTextPartsAsync(client, sessionID, messageID, PLACEHOLDER_TEXT)
       if (replaced) {
         fixedCount++
@@ -107,7 +109,7 @@ export async function sanitizeEmptyMessagesBeforeSummarize(
       log("[auto-compact] pre-summarize sanitization fixed empty messages", {
         sessionID,
         fixedCount,
-        totalEmpty: emptyMessageIds.length,
+        totalEmpty: allIds.length,
       })
     }
 

src/hooks/atlas/boulder-continuation-injector.ts

@@ -1,8 +1,9 @@
 import type { PluginInput } from "@opencode-ai/plugin"
 import type { BackgroundManager } from "../../features/background-agent"
+import { isAgentRegistered } from "../../features/claude-code-session-state"
 import { log } from "../../shared/logger"
 import { createInternalAgentTextPart, resolveInheritedPromptTools } from "../../shared"
-import { getAgentDisplayName } from "../../shared/agent-display-names"
+import { getAgentConfigKey } from "../../shared/agent-display-names"
 import { HOOK_NAME } from "./hook-name"
 import { BOULDER_CONTINUATION_PROMPT } from "./system-reminder-templates"
 import { resolveRecentPromptContextForSession } from "./recent-model-resolver"
@@ -48,24 +49,33 @@ export async function injectBoulderContinuation(input: {
   const preferredSessionContext = preferredTaskSessionId
     ? `\n\n[Preferred reuse session for current top-level plan task${preferredTaskTitle ? `: ${preferredTaskTitle}` : ""}: ${preferredTaskSessionId}]`
     : ""
-  const prompt =
-    BOULDER_CONTINUATION_PROMPT.replace(/{PLAN_NAME}/g, planName) +
-    `\n\n[Status: ${total - remaining}/${total} completed, ${remaining} remaining]` +
-    preferredSessionContext +
-    worktreeContext
+	const prompt =
+		BOULDER_CONTINUATION_PROMPT.replace(/{PLAN_NAME}/g, planName) +
+		`\n\n[Status: ${total - remaining}/${total} completed, ${remaining} remaining]` +
+		preferredSessionContext +
+		worktreeContext
+	const continuationAgent = agent ?? (isAgentRegistered("atlas") ? "atlas" : undefined)
 
-  try {
-    log(`[${HOOK_NAME}] Injecting boulder continuation`, { sessionID, planName, remaining })
+	if (!continuationAgent || !isAgentRegistered(continuationAgent)) {
+		log(`[${HOOK_NAME}] Skipped injection: continuation agent unavailable`, {
+			sessionID,
+			agent: continuationAgent ?? agent ?? "unknown",
+		})
+		return
+	}
+
+	try {
+		log(`[${HOOK_NAME}] Injecting boulder continuation`, { sessionID, planName, remaining })
 
     const promptContext = await resolveRecentPromptContextForSession(ctx, sessionID)
     const inheritedTools = resolveInheritedPromptTools(sessionID, promptContext.tools)
 
-    await ctx.client.session.promptAsync({
-      path: { id: sessionID },
-      body: {
-        agent: getAgentDisplayName(agent ?? "atlas"),
-        ...(promptContext.model !== undefined ? { model: promptContext.model } : {}),
-        ...(inheritedTools ? { tools: inheritedTools } : {}),
+		await ctx.client.session.promptAsync({
+			path: { id: sessionID },
+			body: {
+				agent: getAgentConfigKey(continuationAgent),
+				...(promptContext.model !== undefined ? { model: promptContext.model } : {}),
+				...(inheritedTools ? { tools: inheritedTools } : {}),
         parts: [createInternalAgentTextPart(prompt)],
       },
       query: { directory: ctx.directory },

src/hooks/atlas/compaction-agent-filter.test.ts

@@ -6,7 +6,7 @@ import { join } from "node:path"
 import { randomUUID } from "node:crypto"
 
 import { clearBoulderState, writeBoulderState } from "../../features/boulder-state"
-import { _resetForTesting } from "../../features/claude-code-session-state"
+import { _resetForTesting, registerAgentName } from "../../features/claude-code-session-state"
 import type { BoulderState } from "../../features/boulder-state"
 
 const TEST_STORAGE_ROOT = join(tmpdir(), `atlas-compaction-storage-${randomUUID()}`)
@@ -66,6 +66,8 @@ describe("atlas hook compaction agent filtering", () => {
     mkdirSync(testDirectory, { recursive: true })
     clearBoulderState(testDirectory)
     _resetForTesting()
+    registerAgentName("atlas")
+    registerAgentName("sisyphus")
   })
 
   afterEach(() => {

src/hooks/atlas/idle-event-lineage.test.ts

@@ -6,7 +6,7 @@ import { tmpdir } from "node:os"
 import { join } from "node:path"
 import { clearBoulderState, readBoulderState, writeBoulderState } from "../../features/boulder-state"
 import type { BoulderState } from "../../features/boulder-state"
-import { _resetForTesting, setSessionAgent, subagentSessions } from "../../features/claude-code-session-state"
+import { _resetForTesting, registerAgentName, setSessionAgent, subagentSessions } from "../../features/claude-code-session-state"
 
 const { createAtlasHook } = await import("./index")
 
@@ -64,6 +64,8 @@ describe("atlas hook idle-event session lineage", () => {
     promptCalls = []
     clearBoulderState(testDirectory)
     _resetForTesting()
+    registerAgentName("atlas")
+    registerAgentName("sisyphus")
     subagentSessions.clear()
   })
 

src/hooks/atlas/idle-event.ts

@@ -5,7 +5,7 @@ import {
   readBoulderState,
   readCurrentTopLevelTask,
 } from "../../features/boulder-state"
-import { getSessionAgent, subagentSessions } from "../../features/claude-code-session-state"
+import { getSessionAgent, isAgentRegistered, subagentSessions } from "../../features/claude-code-session-state"
 import { getAgentConfigKey } from "../../shared/agent-display-names"
 import { log } from "../../shared/logger"
 import { injectBoulderContinuation } from "./boulder-continuation-injector"
@@ -141,7 +141,15 @@ export async function handleAtlasSessionIdle(input: {
   if (subagentSessions.has(sessionID)) {
     const sessionAgent = getSessionAgent(sessionID)
     const agentKey = getAgentConfigKey(sessionAgent ?? "")
-    const requiredAgentKey = getAgentConfigKey(boulderState.agent ?? "atlas")
+    const requiredAgentName = boulderState.agent ?? (isAgentRegistered("atlas") ? "atlas" : undefined)
+    if (!requiredAgentName || !isAgentRegistered(requiredAgentName)) {
+      log(`[${HOOK_NAME}] Skipped: boulder agent is unavailable for continuation`, {
+        sessionID,
+        requiredAgent: boulderState.agent ?? "unknown",
+      })
+      return
+    }
+    const requiredAgentKey = getAgentConfigKey(requiredAgentName)
     const agentMatches =
       agentKey === requiredAgentKey ||
       (requiredAgentKey === getAgentConfigKey("atlas") && agentKey === getAgentConfigKey("sisyphus"))
@@ -149,10 +157,10 @@ export async function handleAtlasSessionIdle(input: {
       log(`[${HOOK_NAME}] Skipped: subagent agent does not match boulder agent`, {
         sessionID,
         agent: sessionAgent ?? "unknown",
-        requiredAgent: boulderState.agent ?? "atlas",
-      })
-      return
-    }
+          requiredAgent: requiredAgentName,
+        })
+        return
+      }
   }
 
   const sessionState = getState(sessionID)

src/hooks/atlas/index.test.ts

@@ -9,7 +9,7 @@ import {
   readBoulderState,
 } from "../../features/boulder-state"
 import type { BoulderState } from "../../features/boulder-state"
-import { _resetForTesting, subagentSessions, updateSessionAgent } from "../../features/claude-code-session-state"
+import { _resetForTesting, registerAgentName, subagentSessions, updateSessionAgent } from "../../features/claude-code-session-state"
 import type { PendingTaskRef } from "./types"
 
 const TEST_STORAGE_ROOT = join(tmpdir(), `atlas-message-storage-${randomUUID()}`)
@@ -90,6 +90,9 @@ describe("atlas hook", () => {
   }
 
   beforeEach(() => {
+    _resetForTesting()
+    registerAgentName("atlas")
+    registerAgentName("sisyphus")
     TEST_DIR = join(tmpdir(), `atlas-test-${randomUUID()}`)
     SISYPHUS_DIR = join(TEST_DIR, ".sisyphus")
     if (!existsSync(TEST_DIR)) {
@@ -102,6 +105,7 @@ describe("atlas hook", () => {
   })
 
   afterEach(() => {
+    _resetForTesting()
     clearBoulderState(TEST_DIR)
     if (existsSync(TEST_DIR)) {
       rmSync(TEST_DIR, { recursive: true, force: true })
@@ -1182,9 +1186,11 @@ session_id: ses_untrusted_999
 
      beforeEach(() => {
        _resetForTesting()
-       subagentSessions.clear()
-       setupMessageStorage(MAIN_SESSION_ID, "atlas")
-     })
+       registerAgentName("atlas")
+       registerAgentName("sisyphus")
+        subagentSessions.clear()
+        setupMessageStorage(MAIN_SESSION_ID, "atlas")
+      })
 
     afterEach(() => {
       cleanupMessageStorage(MAIN_SESSION_ID)

src/hooks/legacy-plugin-toast/auto-migrate.test.ts

@@ -2,7 +2,9 @@ import { afterEach, beforeEach, describe, expect, it } from "bun:test"
 import { mkdirSync, readFileSync, rmSync, writeFileSync } from "node:fs"
 import { tmpdir } from "node:os"
 import { join } from "node:path"
-import { autoMigrateLegacyPluginEntry } from "./auto-migrate"
+async function importFreshAutoMigrateModule(): Promise<typeof import("./auto-migrate")> {
+  return import(`./auto-migrate?test=${Date.now()}-${Math.random()}`)
+}
 
 describe("autoMigrateLegacyPluginEntry", () => {
   let testConfigDir = ""
@@ -17,13 +19,15 @@ describe("autoMigrateLegacyPluginEntry", () => {
   })
 
   describe("#given opencode.json has a bare legacy plugin entry", () => {
-    it("#then replaces oh-my-opencode with oh-my-openagent", () => {
+    it("#then replaces oh-my-opencode with oh-my-openagent", async () => {
       // given
       writeFileSync(
         join(testConfigDir, "opencode.json"),
         JSON.stringify({ plugin: ["oh-my-opencode"] }, null, 2) + "\n",
       )
 
+      const { autoMigrateLegacyPluginEntry } = await importFreshAutoMigrateModule()
+
       // when
       const result = autoMigrateLegacyPluginEntry(testConfigDir)
 
@@ -37,13 +41,15 @@ describe("autoMigrateLegacyPluginEntry", () => {
   })
 
   describe("#given opencode.json has a version-pinned legacy entry", () => {
-    it("#then preserves the version suffix", () => {
+    it("#then preserves the version suffix", async () => {
       // given
       writeFileSync(
         join(testConfigDir, "opencode.json"),
         JSON.stringify({ plugin: ["oh-my-opencode@3.10.0"] }, null, 2) + "\n",
       )
 
+      const { autoMigrateLegacyPluginEntry } = await importFreshAutoMigrateModule()
+
       // when
       const result = autoMigrateLegacyPluginEntry(testConfigDir)
 
@@ -57,13 +63,15 @@ describe("autoMigrateLegacyPluginEntry", () => {
   })
 
   describe("#given both canonical and legacy entries exist", () => {
-    it("#then removes legacy entry and keeps canonical", () => {
+    it("#then removes legacy entry and keeps canonical", async () => {
       // given
       writeFileSync(
         join(testConfigDir, "opencode.json"),
         JSON.stringify({ plugin: ["oh-my-openagent", "oh-my-opencode"] }, null, 2) + "\n",
       )
 
+      const { autoMigrateLegacyPluginEntry } = await importFreshAutoMigrateModule()
+
       // when
       const result = autoMigrateLegacyPluginEntry(testConfigDir)
 
@@ -75,8 +83,9 @@ describe("autoMigrateLegacyPluginEntry", () => {
   })
 
   describe("#given no config file exists", () => {
-    it("#then returns migrated false", () => {
+    it("#then returns migrated false", async () => {
       // given - empty dir
+      const { autoMigrateLegacyPluginEntry } = await importFreshAutoMigrateModule()
 
       // when
       const result = autoMigrateLegacyPluginEntry(testConfigDir)
@@ -88,13 +97,15 @@ describe("autoMigrateLegacyPluginEntry", () => {
   })
 
   describe("#given opencode.jsonc has comments and a legacy entry", () => {
-    it("#then preserves comments and replaces entry", () => {
+    it("#then preserves comments and replaces entry", async () => {
       // given
       writeFileSync(
         join(testConfigDir, "opencode.jsonc"),
         '{\n  // my config\n  "plugin": ["oh-my-opencode"]\n}\n',
       )
 
+      const { autoMigrateLegacyPluginEntry } = await importFreshAutoMigrateModule()
+
       // when
       const result = autoMigrateLegacyPluginEntry(testConfigDir)
 
@@ -108,11 +119,13 @@ describe("autoMigrateLegacyPluginEntry", () => {
   })
 
   describe("#given only canonical entry exists", () => {
-    it("#then returns migrated false and leaves file untouched", () => {
+    it("#then returns migrated false and leaves file untouched", async () => {
       // given
       const original = JSON.stringify({ plugin: ["oh-my-openagent"] }, null, 2) + "\n"
       writeFileSync(join(testConfigDir, "opencode.json"), original)
 
+      const { autoMigrateLegacyPluginEntry } = await importFreshAutoMigrateModule()
+
       // when
       const result = autoMigrateLegacyPluginEntry(testConfigDir)
 

src/hooks/preemptive-compaction.test.ts

@@ -669,4 +669,43 @@ describe("preemptive-compaction", () => {
 
     expect(ctx.client.session.summarize).toHaveBeenCalled()
   })
+
+  it("should ignore stale cached Anthropic limits for older models", async () => {
+    const modelContextLimitsCache = new Map<string, number>()
+    modelContextLimitsCache.set("anthropic/claude-sonnet-4-5", 500000)
+
+    const hook = createPreemptiveCompactionHook(ctx as never, {} as never, {
+      anthropicContext1MEnabled: false,
+      modelContextLimitsCache,
+    })
+    const sessionID = "ses_old_anthropic_limit"
+
+    await hook.event({
+      event: {
+        type: "message.updated",
+        properties: {
+          info: {
+            role: "assistant",
+            sessionID,
+            providerID: "anthropic",
+            modelID: "claude-sonnet-4-5",
+            finish: true,
+            tokens: {
+              input: 170000,
+              output: 0,
+              reasoning: 0,
+              cache: { read: 10000, write: 0 },
+            },
+          },
+        },
+      },
+    })
+
+    await hook["tool.execute.after"](
+      { tool: "bash", sessionID, callID: "call_1" },
+      { title: "", output: "test", metadata: null }
+    )
+
+    expect(ctx.client.session.summarize).toHaveBeenCalled()
+  })
 })

src/hooks/ralph-loop/completion-promise-detector.test.ts

@@ -186,7 +186,7 @@ describe("detectCompletionInSessionMessages", () => {
   })
 
   describe("#given semantic completion patterns", () => {
-    test("#when agent says 'task is complete' #then should detect semantic completion", async () => {
+    test("#when agent says 'task is complete' without explicit promise #then should NOT detect completion", async () => {
       // #given
       const messages: SessionMessage[] = [
         {
@@ -205,10 +205,10 @@ describe("detectCompletionInSessionMessages", () => {
       })
 
       // #then
-      expect(detected).toBe(true)
+      expect(detected).toBe(false)
     })
 
-    test("#when agent says 'all items are done' #then should detect semantic completion", async () => {
+    test("#when agent says 'all items are done' without explicit promise #then should NOT detect completion", async () => {
       // #given
       const messages: SessionMessage[] = [
         {
@@ -227,10 +227,10 @@ describe("detectCompletionInSessionMessages", () => {
       })
 
       // #then
-      expect(detected).toBe(true)
+      expect(detected).toBe(false)
     })
 
-    test("#when agent says 'nothing left to do' #then should detect semantic completion", async () => {
+    test("#when agent says 'nothing left to do' without explicit promise #then should NOT detect completion", async () => {
       // #given
       const messages: SessionMessage[] = [
         {
@@ -249,10 +249,10 @@ describe("detectCompletionInSessionMessages", () => {
       })
 
       // #then
-      expect(detected).toBe(true)
+      expect(detected).toBe(false)
     })
 
-    test("#when agent says 'successfully completed all' #then should detect semantic completion", async () => {
+    test("#when agent says 'successfully completed all' without explicit promise #then should NOT detect completion", async () => {
       // #given
       const messages: SessionMessage[] = [
         {
@@ -271,7 +271,7 @@ describe("detectCompletionInSessionMessages", () => {
       })
 
       // #then
-      expect(detected).toBe(true)
+      expect(detected).toBe(false)
     })
 
     test("#when promise is VERIFIED #then semantic completion should NOT trigger", async () => {
@@ -295,6 +295,75 @@ describe("detectCompletionInSessionMessages", () => {
       // #then
       expect(detected).toBe(false)
     })
+
+    test("#when completion text appears inside a quote #then should NOT detect completion", async () => {
+      // #given
+      const messages: SessionMessage[] = [
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: 'The user wrote: "the task is complete". I am still working.' }],
+        },
+      ]
+      const ctx = createPluginInput(messages)
+
+      // #when
+      const detected = await detectCompletionInSessionMessages(ctx, {
+        sessionID: "session-quoted",
+        promise: "DONE",
+        apiTimeoutMs: 1000,
+        directory: "/tmp",
+      })
+
+      // #then
+      expect(detected).toBe(false)
+    })
+
+    test("#when tool_result says all items are complete #then should NOT detect completion", async () => {
+      // #given
+      const messages: SessionMessage[] = [
+        {
+          info: { role: "assistant" },
+          parts: [
+            { type: "tool_result", text: "Background agent report: all items are complete." },
+            { type: "text", text: "Still validating the final behavior." },
+          ],
+        },
+      ]
+      const ctx = createPluginInput(messages)
+
+      // #when
+      const detected = await detectCompletionInSessionMessages(ctx, {
+        sessionID: "session-tool-result-semantic",
+        promise: "DONE",
+        apiTimeoutMs: 1000,
+        directory: "/tmp",
+      })
+
+      // #then
+      expect(detected).toBe(false)
+    })
+
+    test("#when assistant says complete but not actually done #then should NOT detect completion", async () => {
+      // #given
+      const messages: SessionMessage[] = [
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "The implementation looks complete, but I still need to run the tests." }],
+        },
+      ]
+      const ctx = createPluginInput(messages)
+
+      // #when
+      const detected = await detectCompletionInSessionMessages(ctx, {
+        sessionID: "session-not-actually-done",
+        promise: "DONE",
+        apiTimeoutMs: 1000,
+        directory: "/tmp",
+      })
+
+      // #then
+      expect(detected).toBe(false)
+    })
   })
 })
 

src/hooks/ralph-loop/completion-promise-detector.ts

@@ -39,6 +39,8 @@ const SEMANTIC_COMPLETION_PATTERNS = [
 	/\bnothing\s+(?:left|more|remaining)\s+to\s+(?:do|implement|fix)\b/i,
 ]
 
+const SEMANTIC_DONE_FALLBACK_ENABLED = false
+
 export function detectSemanticCompletion(text: string): boolean {
 	return SEMANTIC_COMPLETION_PATTERNS.some((pattern) => pattern.test(text))
 }
@@ -65,9 +67,8 @@ export function detectCompletionInTranscript(
 				const entryText = extractTranscriptEntryText(entry)
 				if (!entryText) continue
 				if (pattern.test(entryText)) return true
-				// Fallback: semantic completion only for DONE promise and assistant entries
 				const isAssistantEntry = entry.type === "assistant" || entry.type === "text"
-				if (promise === "DONE" && isAssistantEntry && detectSemanticCompletion(entryText)) {
+				if (SEMANTIC_DONE_FALLBACK_ENABLED && promise === "DONE" && isAssistantEntry && detectSemanticCompletion(entryText)) {
 					log("[ralph-loop] WARNING: Semantic completion detected in transcript (agent used natural language instead of <promise>DONE</promise>)")
 					return true
 				}
@@ -135,8 +136,7 @@ export async function detectCompletionInSessionMessages(
 				return true
 			}
 
-			// Fallback: semantic completion only for DONE promise
-			if (options.promise === "DONE" && detectSemanticCompletion(responseText)) {
+			if (SEMANTIC_DONE_FALLBACK_ENABLED && options.promise === "DONE" && detectSemanticCompletion(responseText)) {
 				log("[ralph-loop] WARNING: Semantic completion detected (agent used natural language instead of <promise>DONE</promise>)", {
 					sessionID: options.sessionID,
 				})

src/hooks/session-recovery/storage/readers-from-sdk.test.ts

@@ -1,7 +1,13 @@
 import { describe, expect, it } from "bun:test"
-import { readMessagesFromSDK, readPartsFromSDK } from "../storage"
-import { readMessages } from "./messages-reader"
-import { readParts } from "./parts-reader"
+async function importFreshReaders() {
+  const token = `${Date.now()}-${Math.random()}`
+  const [{ readMessagesFromSDK, readMessages }, { readPartsFromSDK, readParts }] = await Promise.all([
+    import(`./messages-reader?test=${token}`),
+    import(`./parts-reader?test=${token}`),
+  ])
+
+  return { readMessagesFromSDK, readPartsFromSDK, readMessages, readParts }
+}
 
 function createMockClient(handlers: {
   messages?: (sessionID: string) => unknown[]
@@ -28,6 +34,7 @@ function createMockClient(handlers: {
 describe("session-recovery storage SDK readers", () => {
   it("readPartsFromSDK returns empty array when fetch fails", async () => {
     //#given a client that throws on request
+    const { readPartsFromSDK } = await importFreshReaders()
     const client = createMockClient({}) as Parameters<typeof readPartsFromSDK>[0]
 
     //#when readPartsFromSDK is called
@@ -39,6 +46,7 @@ describe("session-recovery storage SDK readers", () => {
 
   it("readPartsFromSDK returns stored parts from SDK response", async () => {
     //#given a client that returns a message with parts
+    const { readPartsFromSDK } = await importFreshReaders()
     const sessionID = "ses_test"
     const messageID = "msg_test"
     const storedParts = [
@@ -58,6 +66,7 @@ describe("session-recovery storage SDK readers", () => {
 
   it("readMessagesFromSDK normalizes and sorts messages", async () => {
     //#given a client that returns messages list
+    const { readMessagesFromSDK } = await importFreshReaders()
     const sessionID = "ses_test"
     const client = createMockClient({
       messages: () => [
@@ -78,8 +87,9 @@ describe("session-recovery storage SDK readers", () => {
     ])
   })
 
-  it("readParts returns empty array for nonexistent message", () => {
+  it("readParts returns empty array for nonexistent message", async () => {
     //#given a message ID that has no stored parts
+    const { readParts } = await importFreshReaders()
     //#when readParts is called
     const parts = readParts("msg_nonexistent")
 
@@ -87,8 +97,9 @@ describe("session-recovery storage SDK readers", () => {
     expect(parts).toEqual([])
   })
 
-  it("readMessages returns empty array for nonexistent session", () => {
+  it("readMessages returns empty array for nonexistent session", async () => {
     //#given a session ID that has no stored messages
+    const { readMessages } = await importFreshReaders()
     //#when readMessages is called
     const messages = readMessages("ses_nonexistent")
 

src/hooks/start-work/index.test.ts

@@ -12,7 +12,6 @@ import {
 import type { BoulderState } from "../../features/boulder-state"
 import * as sessionState from "../../features/claude-code-session-state"
 import * as worktreeDetector from "./worktree-detector"
-import * as worktreeDetector from "./worktree-detector"
 
 describe("start-work hook", () => {
   let testDir: string
@@ -26,6 +25,9 @@ describe("start-work hook", () => {
   }
 
   beforeEach(() => {
+    sessionState._resetForTesting()
+    sessionState.registerAgentName("atlas")
+    sessionState.registerAgentName("sisyphus")
     testDir = join(tmpdir(), `start-work-test-${randomUUID()}`)
     sisyphusDir = join(testDir, ".sisyphus")
     if (!existsSync(testDir)) {
@@ -38,6 +40,7 @@ describe("start-work hook", () => {
   })
 
   afterEach(() => {
+    sessionState._resetForTesting()
     clearBoulderState(testDir)
     if (existsSync(testDir)) {
       rmSync(testDir, { recursive: true, force: true })
@@ -409,7 +412,7 @@ describe("start-work hook", () => {
       // given
       const hook = createStartWorkHook(createMockPluginInput())
       const output = {
-        message: {},
+        message: {} as Record<string, unknown>,
         parts: [{ type: "text", text: "<session-context></session-context>" }],
       }
 
@@ -422,6 +425,29 @@ describe("start-work hook", () => {
       // then
       expect(output.message.agent).toBe("Atlas (Plan Executor)")
     })
+
+    test("should keep the current agent when Atlas is unavailable", async () => {
+      // given
+      sessionState._resetForTesting()
+      sessionState.registerAgentName("sisyphus")
+      sessionState.updateSessionAgent("ses-prometheus-to-sisyphus", "sisyphus")
+
+      const hook = createStartWorkHook(createMockPluginInput())
+      const output = {
+        message: {} as Record<string, unknown>,
+        parts: [{ type: "text", text: "<session-context></session-context>" }],
+      }
+
+      // when
+      await hook["chat.message"](
+        { sessionID: "ses-prometheus-to-sisyphus" },
+        output
+      )
+
+      // then
+      expect(output.message.agent).toBe("Sisyphus (Ultraworker)")
+      expect(sessionState.getSessionAgent("ses-prometheus-to-sisyphus")).toBe("sisyphus")
+    })
   })
 
   describe("worktree support", () => {

src/hooks/start-work/start-work-hook.ts

@@ -12,7 +12,7 @@ import {
 } from "../../features/boulder-state"
 import { log } from "../../shared/logger"
 import { getAgentDisplayName } from "../../shared/agent-display-names"
-import { updateSessionAgent, isAgentRegistered } from "../../features/claude-code-session-state"
+import { getSessionAgent, isAgentRegistered, updateSessionAgent } from "../../features/claude-code-session-state"
 import { detectWorktreePath } from "./worktree-detector"
 import { parseUserRequest } from "./parse-user-request"
 
@@ -80,14 +80,13 @@ export function createStartWorkHook(ctx: PluginInput) {
       if (!promptText.includes("<session-context>")) return
 
       log(`[${HOOK_NAME}] Processing start-work command`, { sessionID: input.sessionID })
-      const atlasDisplayName = getAgentDisplayName("atlas")
-      if (isAgentRegistered("atlas") || isAgentRegistered(atlasDisplayName)) {
-        updateSessionAgent(input.sessionID, "atlas")
-        if (output.message) {
-          output.message["agent"] = atlasDisplayName
-        }
-      } else {
-        log(`[${HOOK_NAME}] Atlas agent not available, continuing with current agent`, { sessionID: input.sessionID })
+      const activeAgent = isAgentRegistered("atlas")
+        ? "atlas"
+        : getSessionAgent(input.sessionID) ?? "sisyphus"
+      const activeAgentDisplayName = getAgentDisplayName(activeAgent)
+      updateSessionAgent(input.sessionID, activeAgent)
+      if (output.message) {
+        output.message["agent"] = activeAgentDisplayName
       }
 
       const existingState = readBoulderState(ctx.directory)
@@ -116,7 +115,7 @@ The requested plan "${getPlanName(matchedPlan)}" has been completed.
 All ${progress.total} tasks are done. Create a new plan with: /plan "your task"`
           } else {
             if (existingState) clearBoulderState(ctx.directory)
-            const newState = createBoulderState(matchedPlan, sessionId, "atlas", worktreePath)
+            const newState = createBoulderState(matchedPlan, sessionId, activeAgent, worktreePath)
             writeBoulderState(ctx.directory, newState)
 
             contextInfo = `
@@ -223,7 +222,7 @@ All ${plans.length} plan(s) are complete. Create a new plan with: /plan "your ta
         } else if (incompletePlans.length === 1) {
           const planPath = incompletePlans[0]
           const progress = getPlanProgress(planPath)
-          const newState = createBoulderState(planPath, sessionId, "atlas", worktreePath)
+          const newState = createBoulderState(planPath, sessionId, activeAgent, worktreePath)
           writeBoulderState(ctx.directory, newState)
 
           contextInfo += `

src/hooks/tasks-todowrite-disabler/hook.ts

@@ -9,7 +9,7 @@ export interface TasksTodowriteDisablerConfig {
 export function createTasksTodowriteDisablerHook(
   config: TasksTodowriteDisablerConfig,
 ) {
-  const isTaskSystemEnabled = config.experimental?.task_system ?? false;
+  const isTaskSystemEnabled = config.experimental?.task_system ?? true;
 
   return {
     "tool.execute.before": async (

src/hooks/tasks-todowrite-disabler/index.test.ts

@@ -59,7 +59,7 @@ describe("tasks-todowrite-disabler", () => {
     })
   })
 
-  describe("when experimental.task_system is disabled or undefined", () => {
+  describe("when experimental.task_system is disabled", () => {
     test("should not block TodoWrite when flag is false", async () => {
       // given
       const hook = createTasksTodowriteDisablerHook({ experimental: { task_system: false } })
@@ -78,7 +78,7 @@ describe("tasks-todowrite-disabler", () => {
       ).resolves.toBeUndefined()
     })
 
-    test("should not block TodoWrite when experimental is undefined", async () => {
+    test("should block TodoWrite when experimental is undefined because task_system defaults to enabled", async () => {
       // given
       const hook = createTasksTodowriteDisablerHook({})
       const input = {
@@ -93,7 +93,7 @@ describe("tasks-todowrite-disabler", () => {
       // when / then
       await expect(
         hook["tool.execute.before"](input, output)
-      ).resolves.toBeUndefined()
+      ).rejects.toThrow("TodoRead/TodoWrite are DISABLED")
     })
 
     test("should not block TodoRead when flag is false", async () => {

src/plugin-config.test.ts

@@ -246,7 +246,13 @@ describe("parseConfigPartially", () => {
       const result = parseConfigPartially({});
 
       expect(result).not.toBeNull();
-      expect(Object.keys(result!).length).toBe(0);
+      expect(result).toEqual({
+        git_master: {
+          commit_footer: true,
+          include_co_authored_by: true,
+          git_env_prefix: "GIT_MASTER=1",
+        },
+      });
     });
   });
 

src/plugin-handlers/agent-config-handler.test.ts

@@ -290,7 +290,7 @@ describe("applyAgentConfig builtin override protection", () => {
     })
 
     // then
-    expect(createSisyphusJuniorAgentSpy).toHaveBeenCalledWith(undefined, "openai/gpt-5.4", false)
+    expect(createSisyphusJuniorAgentSpy).toHaveBeenCalledWith(undefined, "openai/gpt-5.4", true)
   })
 
   test("includes project and global .agents skills in builtin agent awareness", async () => {

src/plugin-handlers/agent-config-handler.ts

@@ -90,7 +90,7 @@ export async function applyAgentConfig(params: {
     params.pluginConfig.browser_automation_engine?.provider ?? "playwright";
   const currentModel = params.config.model as string | undefined;
   const disabledSkills = new Set<string>(params.pluginConfig.disabled_skills ?? []);
-  const useTaskSystem = params.pluginConfig.experimental?.task_system ?? false;
+  const useTaskSystem = params.pluginConfig.experimental?.task_system ?? true;
   const disableOmoEnv = params.pluginConfig.experimental?.disable_omo_env ?? false;
 
   const includeClaudeAgents = params.pluginConfig.claude_code?.agents ?? true;

src/plugin-handlers/config-handler.test.ts

@@ -1243,7 +1243,7 @@ describe("per-agent todowrite/todoread deny when task_system enabled", () => {
     expect(agentResult[getAgentDisplayName("hephaestus")]?.permission?.todoread).toBeUndefined()
   })
 
-  test("does not deny todowrite/todoread when task_system is undefined", async () => {
+  test("denies todowrite/todoread when task_system is undefined", async () => {
     //#given
     const createBuiltinAgentsMock = agents.createBuiltinAgents as unknown as {
       mockResolvedValue: (value: Record<string, unknown>) => void
@@ -1271,8 +1271,8 @@ describe("per-agent todowrite/todoread deny when task_system enabled", () => {
 
     //#then
     const agentResult = config.agent as Record<string, { permission?: Record<string, unknown> }>
-    expect(agentResult[getAgentDisplayName("sisyphus")]?.permission?.todowrite).toBeUndefined()
-    expect(agentResult[getAgentDisplayName("sisyphus")]?.permission?.todoread).toBeUndefined()
+    expect(agentResult[getAgentDisplayName("sisyphus")]?.permission?.todowrite).toBe("deny")
+    expect(agentResult[getAgentDisplayName("sisyphus")]?.permission?.todoread).toBe("deny")
   })
 })
 

src/plugin-handlers/tool-config-handler.test.ts

@@ -15,7 +15,7 @@ function createParams(overrides: {
   return {
     config: { tools: {}, permission: {} } as Record<string, unknown>,
     pluginConfig: {
-      experimental: { task_system: overrides.taskSystem ?? false },
+      experimental: overrides.taskSystem === undefined ? undefined : { task_system: overrides.taskSystem },
       disabled_tools: overrides.disabledTools,
     } as OhMyOpenCodeConfig,
     agentResult: agentResult as Record<string, unknown>,
@@ -216,6 +216,30 @@ describe("applyToolConfig", () => {
     })
   })
 
+  describe("#given task_system is undefined", () => {
+    describe("#when applying tool config", () => {
+      it.each([
+        "atlas",
+        "sisyphus",
+        "hephaestus",
+        "prometheus",
+        "sisyphus-junior",
+      ])("#then should deny todo tools for %s agent by default", (agentName) => {
+        const params = createParams({
+          agents: [agentName],
+        })
+
+        applyToolConfig(params)
+
+        const agent = params.agentResult[agentName] as {
+          permission: Record<string, unknown>
+        }
+        expect(agent.permission.todowrite).toBe("deny")
+        expect(agent.permission.todoread).toBe("deny")
+      })
+    })
+  })
+
   describe("#given disabled_tools includes 'question'", () => {
     let originalConfigContent: string | undefined
     let originalCliRunMode: string | undefined

src/plugin-handlers/tool-config-handler.ts

@@ -15,7 +15,7 @@ function getConfigQuestionPermission(): string | null {
 }
 
 function agentByKey(agentResult: Record<string, unknown>, key: string): AgentWithPermission | undefined {
-  return (agentResult[key] ?? agentResult[getAgentDisplayName(key)]) as
+  return (agentResult[getAgentDisplayName(key)] ?? agentResult[key]) as
     | AgentWithPermission
     | undefined;
 }
@@ -25,7 +25,8 @@ export function applyToolConfig(params: {
   pluginConfig: OhMyOpenCodeConfig;
   agentResult: Record<string, unknown>;
 }): void {
-  const denyTodoTools = params.pluginConfig.experimental?.task_system
+  const taskSystemEnabled = params.pluginConfig.experimental?.task_system ?? true
+  const denyTodoTools = taskSystemEnabled
     ? { todowrite: "deny", todoread: "deny" }
     : {}
 
@@ -40,7 +41,7 @@ export function applyToolConfig(params: {
     LspCodeActionResolve: false,
     "task_*": false,
     teammate: false,
-    ...(params.pluginConfig.experimental?.task_system
+    ...(taskSystemEnabled
       ? { todowrite: false, todoread: false }
       : {}),
     ...(skillDeniedByHost

src/plugin/tool-execute-before.test.ts

@@ -1,6 +1,7 @@
 const { describe, expect, test } = require("bun:test")
 const { createToolExecuteBeforeHandler } = require("./tool-execute-before")
 const { createToolRegistry } = require("./tool-registry")
+const { builtinTools } = require("../tools")
 
 describe("createToolExecuteBeforeHandler", () => {
   test("does not execute subagent question blocker hook for question tool", async () => {
@@ -268,6 +269,44 @@ describe("createToolRegistry", () => {
       })
     })
   })
+
+  describe("#given max_tools is lower than or equal to builtin tool count", () => {
+    describe("#when creating the tool registry", () => {
+      test("#then it trims to the exact configured cap", () => {
+        const result = createToolRegistry(
+          createRegistryInput({
+            experimental: { max_tools: Object.keys(builtinTools).length },
+          }),
+        )
+
+        expect(Object.keys(result.filteredTools)).toHaveLength(Object.keys(builtinTools).length)
+      })
+    })
+  })
+
+  describe("#given max_tools is set below the full plugin tool count", () => {
+    describe("#when creating the tool registry", () => {
+      test("#then it enforces the exact cap deterministically", () => {
+        const result = createToolRegistry(
+          createRegistryInput({
+            experimental: { max_tools: 10 },
+          }),
+        )
+
+        expect(Object.keys(result.filteredTools)).toHaveLength(10)
+      })
+
+      test("#then it keeps the task tool when lower-priority tools can satisfy the cap", () => {
+        const result = createToolRegistry(
+          createRegistryInput({
+            experimental: { max_tools: 10 },
+          }),
+        )
+
+        expect(result.filteredTools.task).toBeDefined()
+      })
+    })
+  })
 })
 
 export {}

src/plugin/tool-registry.ts

@@ -40,6 +40,63 @@ export type ToolRegistryResult = {
   taskSystemEnabled: boolean
 }
 
+const LOW_PRIORITY_TOOL_ORDER = [
+  "session_list",
+  "session_read",
+  "session_search",
+  "session_info",
+  "interactive_bash",
+  "look_at",
+  "call_omo_agent",
+  "task_create",
+  "task_get",
+  "task_list",
+  "task_update",
+  "background_output",
+  "background_cancel",
+  "hashline_edit",
+  "ast_grep_replace",
+  "ast_grep_search",
+  "glob",
+  "grep",
+  "skill_mcp",
+  "skill",
+  "task",
+  "lsp_rename",
+  "lsp_prepare_rename",
+  "lsp_find_references",
+  "lsp_goto_definition",
+  "lsp_symbols",
+  "lsp_diagnostics",
+] as const
+
+function trimToolsToCap(filteredTools: ToolsRecord, maxTools: number): void {
+  const toolNames = Object.keys(filteredTools)
+  if (toolNames.length <= maxTools) return
+
+  const removableToolNames = [
+    ...LOW_PRIORITY_TOOL_ORDER.filter((toolName) => toolNames.includes(toolName)),
+    ...toolNames
+      .filter((toolName) => !LOW_PRIORITY_TOOL_ORDER.includes(toolName as (typeof LOW_PRIORITY_TOOL_ORDER)[number]))
+      .sort(),
+  ]
+
+  let currentCount = toolNames.length
+  let removed = 0
+
+  for (const toolName of removableToolNames) {
+    if (currentCount <= maxTools) break
+    if (!filteredTools[toolName]) continue
+    delete filteredTools[toolName]
+    currentCount -= 1
+    removed += 1
+  }
+
+  log(
+    `[tool-registry] Trimmed ${removed} tools to satisfy max_tools=${maxTools}. Final plugin tool count=${currentCount}.`,
+  )
+}
+
 export function createToolRegistry(args: {
   ctx: PluginContext
   pluginConfig: OhMyOpenCodeConfig
@@ -158,29 +215,7 @@ export function createToolRegistry(args: {
 
   const maxTools = pluginConfig.experimental?.max_tools
   if (maxTools) {
-    const estimatedBuiltinTools = 20
-    const pluginToolBudget = maxTools - estimatedBuiltinTools
-    const toolEntries = Object.entries(filteredTools)
-    if (pluginToolBudget > 0 && toolEntries.length > pluginToolBudget) {
-      const excess = toolEntries.length - pluginToolBudget
-      log(`[tool-registry] Tool count (${toolEntries.length} plugin + ~${estimatedBuiltinTools} builtin = ~${toolEntries.length + estimatedBuiltinTools}) exceeds max_tools=${maxTools}. Trimming ${excess} lower-priority tools.`)
-      const lowPriorityTools = [
-        "session_list", "session_read", "session_search", "session_info",
-        "call_omo_agent", "interactive_bash", "look_at",
-        "task_create", "task_get", "task_list", "task_update",
-      ]
-      let removed = 0
-      for (const toolName of lowPriorityTools) {
-        if (removed >= excess) break
-        if (filteredTools[toolName]) {
-          delete filteredTools[toolName]
-          removed += 1
-        }
-      }
-      if (removed < excess) {
-        log(`[tool-registry] WARNING: Could not trim enough tools. ${toolEntries.length - removed} plugin tools remain.`)
-      }
-    }
+    trimToolsToCap(filteredTools, maxTools)
   }
 
   return {

src/shared/context-limit-resolver.test.ts

@@ -45,7 +45,7 @@ describe("resolveActualContextLimit", () => {
     expect(actualLimit).toBe(1_000_000)
   })
 
-  it("returns cached limit for Anthropic models when modelContextLimitsCache has entry", () => {
+  it("returns default 200K for older Anthropic models when 1M mode is disabled", () => {
     // given
     delete process.env[ANTHROPIC_CONTEXT_ENV_KEY]
     delete process.env[VERTEX_CONTEXT_ENV_KEY]
@@ -59,7 +59,7 @@ describe("resolveActualContextLimit", () => {
     })
 
     // then
-    expect(actualLimit).toBe(500_000)
+    expect(actualLimit).toBe(200_000)
   })
 
   it("returns default 200K for Anthropic models without cached limit and 1M mode disabled", () => {
@@ -126,6 +126,40 @@ describe("resolveActualContextLimit", () => {
     expect(actualLimit).toBe(1_000_000)
   })
 
+  it("supports Anthropic 4.6 high-variant model IDs without widening older models", () => {
+    // given
+    delete process.env[ANTHROPIC_CONTEXT_ENV_KEY]
+    delete process.env[VERTEX_CONTEXT_ENV_KEY]
+    const modelContextLimitsCache = new Map<string, number>()
+    modelContextLimitsCache.set("anthropic/claude-sonnet-4-6-high", 500_000)
+
+    // when
+    const actualLimit = resolveActualContextLimit("anthropic", "claude-sonnet-4-6-high", {
+      anthropicContext1MEnabled: false,
+      modelContextLimitsCache,
+    })
+
+    // then
+    expect(actualLimit).toBe(500_000)
+  })
+
+  it("ignores stale cached limits for older Anthropic models with suffixed IDs", () => {
+    // given
+    delete process.env[ANTHROPIC_CONTEXT_ENV_KEY]
+    delete process.env[VERTEX_CONTEXT_ENV_KEY]
+    const modelContextLimitsCache = new Map<string, number>()
+    modelContextLimitsCache.set("anthropic/claude-sonnet-4-5-high", 500_000)
+
+    // when
+    const actualLimit = resolveActualContextLimit("anthropic", "claude-sonnet-4-5-high", {
+      anthropicContext1MEnabled: false,
+      modelContextLimitsCache,
+    })
+
+    // then
+    expect(actualLimit).toBe(200_000)
+  })
+
   it("returns null for non-Anthropic providers without a cached limit", () => {
     // given
     delete process.env[ANTHROPIC_CONTEXT_ENV_KEY]

src/shared/context-limit-resolver.ts

@@ -19,6 +19,10 @@ function getAnthropicActualLimit(modelCacheState?: ContextLimitModelCacheState):
     : DEFAULT_ANTHROPIC_ACTUAL_LIMIT
 }
 
+function supportsCachedAnthropicLimit(modelID: string): boolean {
+  return /^claude-(opus|sonnet)-4(?:-|\.)6(?:-high)?$/.test(modelID)
+}
+
 export function resolveActualContextLimit(
   providerID: string,
   modelID: string,
@@ -29,7 +33,7 @@ export function resolveActualContextLimit(
     if (explicit1M === 1_000_000) return explicit1M
 
     const cachedLimit = modelCacheState?.modelContextLimitsCache?.get(`${providerID}/${modelID}`)
-    if (cachedLimit) return cachedLimit
+    if (cachedLimit && supportsCachedAnthropicLimit(modelID)) return cachedLimit
 
     return DEFAULT_ANTHROPIC_ACTUAL_LIMIT
   }

src/shared/migrate-legacy-plugin-entry.test.ts

@@ -2,7 +2,10 @@ import { afterEach, beforeEach, describe, expect, it } from "bun:test"
 import { mkdirSync, readFileSync, rmSync, writeFileSync } from "node:fs"
 import { tmpdir } from "node:os"
 import { join } from "node:path"
-import { migrateLegacyPluginEntry } from "./migrate-legacy-plugin-entry"
+
+async function importFreshMigrationModule(): Promise<typeof import("./migrate-legacy-plugin-entry")> {
+  return import(`./migrate-legacy-plugin-entry?test=${Date.now()}-${Math.random()}`)
+}
 
 describe("migrateLegacyPluginEntry", () => {
   let testDir = ""
@@ -18,9 +21,10 @@ describe("migrateLegacyPluginEntry", () => {
 
   describe("#given opencode.json contains oh-my-opencode plugin entry", () => {
     describe("#when migrating the config", () => {
-      it("#then replaces oh-my-opencode with oh-my-openagent", () => {
+      it("#then replaces oh-my-opencode with oh-my-openagent", async () => {
         const configPath = join(testDir, "opencode.json")
         writeFileSync(configPath, JSON.stringify({ plugin: ["oh-my-opencode@latest"] }, null, 2))
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
 
         const result = migrateLegacyPluginEntry(configPath)
 
@@ -34,9 +38,10 @@ describe("migrateLegacyPluginEntry", () => {
 
   describe("#given opencode.json contains bare oh-my-opencode entry", () => {
     describe("#when migrating the config", () => {
-      it("#then replaces with oh-my-openagent", () => {
+      it("#then replaces with oh-my-openagent", async () => {
         const configPath = join(testDir, "opencode.json")
         writeFileSync(configPath, JSON.stringify({ plugin: ["oh-my-opencode"] }, null, 2))
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
 
         const result = migrateLegacyPluginEntry(configPath)
 
@@ -50,9 +55,10 @@ describe("migrateLegacyPluginEntry", () => {
 
   describe("#given opencode.json contains pinned oh-my-opencode version", () => {
     describe("#when migrating the config", () => {
-      it("#then preserves the version pin", () => {
+      it("#then preserves the version pin", async () => {
         const configPath = join(testDir, "opencode.json")
         writeFileSync(configPath, JSON.stringify({ plugin: ["oh-my-opencode@3.11.0"] }, null, 2))
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
 
         const result = migrateLegacyPluginEntry(configPath)
 
@@ -65,10 +71,11 @@ describe("migrateLegacyPluginEntry", () => {
 
   describe("#given opencode.json already uses oh-my-openagent", () => {
     describe("#when checking for migration", () => {
-      it("#then returns false and does not modify the file", () => {
+      it("#then returns false and does not modify the file", async () => {
         const configPath = join(testDir, "opencode.json")
         const original = JSON.stringify({ plugin: ["oh-my-openagent@latest"] }, null, 2)
         writeFileSync(configPath, original)
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
 
         const result = migrateLegacyPluginEntry(configPath)
 
@@ -78,9 +85,89 @@ describe("migrateLegacyPluginEntry", () => {
     })
   })
 
+  describe("#given plugin entries contain both canonical and legacy values", () => {
+    describe("#when migrating the config", () => {
+      it("#then removes the legacy entry instead of duplicating the canonical one", async () => {
+        const configPath = join(testDir, "opencode.json")
+        writeFileSync(configPath, JSON.stringify({ plugin: ["oh-my-openagent", "oh-my-opencode"] }, null, 2))
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
+
+        const result = migrateLegacyPluginEntry(configPath)
+
+        expect(result).toBe(true)
+        const saved = JSON.parse(readFileSync(configPath, "utf-8")) as { plugin: string[] }
+        expect(saved.plugin).toEqual(["oh-my-openagent"])
+      })
+    })
+  })
+
+  describe("#given unrelated strings contain the legacy package name", () => {
+    describe("#when migrating the config", () => {
+      it("#then rewrites only plugin entries and preserves unrelated fields", async () => {
+        const configPath = join(testDir, "opencode.json")
+        writeFileSync(
+          configPath,
+          JSON.stringify(
+            {
+              plugin: ["oh-my-opencode"],
+              notes: "keep oh-my-opencode in this text field",
+              paths: ["/tmp/oh-my-opencode/cache"],
+            },
+            null,
+            2,
+          ),
+        )
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
+
+        const result = migrateLegacyPluginEntry(configPath)
+
+        expect(result).toBe(true)
+        const saved = JSON.parse(readFileSync(configPath, "utf-8")) as {
+          plugin: string[]
+          notes: string
+          paths: string[]
+        }
+        expect(saved.plugin).toEqual(["oh-my-openagent"])
+        expect(saved.notes).toBe("keep oh-my-opencode in this text field")
+        expect(saved.paths).toEqual(["/tmp/oh-my-opencode/cache"])
+      })
+    })
+  })
+
+  describe("#given opencode.jsonc contains a nested plugin key before the top-level plugin array", () => {
+    describe("#when migrating the config", () => {
+      it("#then rewrites only the top-level plugin array", async () => {
+        const configPath = join(testDir, "opencode.jsonc")
+        writeFileSync(
+          configPath,
+          `{
+  "nested": {
+    "plugin": ["oh-my-opencode"]
+  },
+  "plugin": ["oh-my-opencode@latest"]
+}
+`,
+        )
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
+
+        const result = migrateLegacyPluginEntry(configPath)
+
+        expect(result).toBe(true)
+        const content = readFileSync(configPath, "utf-8")
+        expect(content).toContain(`"nested": {
+    "plugin": ["oh-my-opencode"]
+  }`)
+        expect(content).toContain(`"plugin": [
+    "oh-my-openagent@latest"
+  ]`)
+      })
+    })
+  })
+
   describe("#given config file does not exist", () => {
     describe("#when attempting migration", () => {
-      it("#then returns false", () => {
+      it("#then returns false", async () => {
+        const { migrateLegacyPluginEntry } = await importFreshMigrationModule()
         const result = migrateLegacyPluginEntry(join(testDir, "nonexistent.json"))
 
         expect(result).toBe(false)

src/shared/migrate-legacy-plugin-entry.ts

@@ -1,8 +1,54 @@
 import { existsSync, readFileSync, writeFileSync } from "node:fs"
+import { applyEdits, modify } from "jsonc-parser"
 
+import { parseJsoncSafe } from "./jsonc-parser"
 import { log } from "./logger"
 import { LEGACY_PLUGIN_NAME, PLUGIN_NAME } from "./plugin-identity"
 
+interface OpenCodeConfig {
+  plugin?: string[]
+}
+
+function isLegacyEntry(entry: string): boolean {
+  return entry === LEGACY_PLUGIN_NAME || entry.startsWith(`${LEGACY_PLUGIN_NAME}@`)
+}
+
+function isCanonicalEntry(entry: string): boolean {
+  return entry === PLUGIN_NAME || entry.startsWith(`${PLUGIN_NAME}@`)
+}
+
+function toCanonicalEntry(entry: string): string {
+  if (entry === LEGACY_PLUGIN_NAME) return PLUGIN_NAME
+  if (entry.startsWith(`${LEGACY_PLUGIN_NAME}@`)) {
+    return `${PLUGIN_NAME}${entry.slice(LEGACY_PLUGIN_NAME.length)}`
+  }
+  return entry
+}
+
+function normalizePluginEntries(entries: string[]): string[] {
+  const hasCanonical = entries.some(isCanonicalEntry)
+
+  if (hasCanonical) {
+    return entries.filter((entry) => !isLegacyEntry(entry))
+  }
+
+  return entries.map((entry) => (isLegacyEntry(entry) ? toCanonicalEntry(entry) : entry))
+}
+
+function updateJsoncPluginArray(content: string, pluginEntries: string[]): string | null {
+  const edits = modify(content, ["plugin"], pluginEntries, {
+    formattingOptions: {
+      insertSpaces: true,
+      tabSize: 2,
+      eol: "\n",
+    },
+    getInsertionIndex: () => 0,
+  })
+
+  if (edits.length === 0) return null
+  return applyEdits(content, edits)
+}
+
 export function migrateLegacyPluginEntry(configPath: string): boolean {
   if (!existsSync(configPath)) return false
 
@@ -10,8 +56,15 @@ export function migrateLegacyPluginEntry(configPath: string): boolean {
     const content = readFileSync(configPath, "utf-8")
     if (!content.includes(LEGACY_PLUGIN_NAME)) return false
 
-    const updated = content.replaceAll(LEGACY_PLUGIN_NAME, PLUGIN_NAME)
-    if (updated === content) return false
+    const parseResult = parseJsoncSafe<OpenCodeConfig>(content)
+    const pluginEntries = parseResult.data?.plugin
+    if (!pluginEntries || !pluginEntries.some(isLegacyEntry)) return false
+
+    const updatedPluginEntries = normalizePluginEntries(pluginEntries)
+    const updated = configPath.endsWith(".jsonc")
+      ? updateJsoncPluginArray(content, updatedPluginEntries)
+      : JSON.stringify({ ...(parseResult.data as OpenCodeConfig), plugin: updatedPluginEntries }, null, 2) + "\n"
+    if (!updated || updated === content) return false
 
     writeFileSync(configPath, updated, "utf-8")
     log("[migrateLegacyPluginEntry] Auto-migrated opencode.json plugin entry", {

src/shared/skill-path-resolver.test.ts

@@ -125,4 +125,28 @@ describe("resolveSkillPathReferences", () => {
 		//#then
 		expect(result).toBe("/skills/frontend/scripts/search.py")
 	})
+
+	it("does not resolve traversal paths that escape the base directory", () => {
+		//#given
+		const content = "Read @data/../../../../etc/passwd before running"
+		const basePath = "/skills/frontend"
+
+		//#when
+		const result = resolveSkillPathReferences(content, basePath)
+
+		//#then
+		expect(result).toBe("Read @data/../../../../etc/passwd before running")
+	})
+
+	it("does not resolve directory traversal with trailing slash", () => {
+		//#given
+		const content = "Inspect @data/../../../secret/"
+		const basePath = "/skills/frontend"
+
+		//#when
+		const result = resolveSkillPathReferences(content, basePath)
+
+		//#then
+		expect(result).toBe("Inspect @data/../../../secret/")
+	})
 })

src/shared/skill-path-resolver.ts

@@ -1,4 +1,4 @@
-import { join } from "path"
+import { isAbsolute, relative, resolve, sep } from "node:path"
 
 function looksLikeFilePath(path: string): boolean {
 	if (path.endsWith("/")) return true
@@ -6,22 +6,21 @@ function looksLikeFilePath(path: string): boolean {
 	return /\.[a-zA-Z0-9]+$/.test(lastSegment)
 }
 
-/**
- * Resolves @path references in skill content to absolute paths.
- *
- * Matches @references that contain at least one slash (e.g., @scripts/search.py, @data/)
- * to avoid false positives with decorators (@param), JSDoc tags (@ts-ignore), etc.
- * Also skips npm scoped packages (@scope/package) by requiring a file extension or trailing slash.
- *
- * Email addresses are excluded since they have alphanumeric characters before @.
- */
 export function resolveSkillPathReferences(content: string, basePath: string): string {
 	const normalizedBase = basePath.endsWith("/") ? basePath.slice(0, -1) : basePath
 	return content.replace(
 		/(?<![a-zA-Z0-9])@([a-zA-Z0-9_-]+\/[a-zA-Z0-9_.\-\/]*)/g,
 		(match, relativePath: string) => {
 			if (!looksLikeFilePath(relativePath)) return match
-			return join(normalizedBase, relativePath)
+			const resolvedPath = resolve(normalizedBase, relativePath)
+			const relativePathFromBase = relative(normalizedBase, resolvedPath)
+			if (relativePathFromBase.startsWith("..") || isAbsolute(relativePathFromBase)) {
+				return match
+			}
+			if (relativePath.endsWith("/") && !resolvedPath.endsWith(sep)) {
+				return `${resolvedPath}/`
+			}
+			return resolvedPath
 		}
 	)
 }

src/tools/delegate-task/model-selection.test.ts

@@ -76,7 +76,9 @@ describe("resolveModelForDelegateTask", () => {
 		})
 
 		describe("#when availableModels is empty (cache exists but empty)", () => {
-			test("#then falls through to category default model (existing behavior)", () => {
+			test("#then keeps the category default when its provider is connected", () => {
+				const readConnectedProvidersSpy = spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["anthropic"])
+
 				const result = resolveModelForDelegateTask({
 					categoryDefaultModel: "anthropic/claude-sonnet-4-6",
 					fallbackChain: [
@@ -87,6 +89,40 @@ describe("resolveModelForDelegateTask", () => {
 				})
 
 				expect(result).toEqual({ model: "anthropic/claude-sonnet-4-6" })
+				readConnectedProvidersSpy.mockRestore()
+			})
+
+			test("#then skips a disconnected category default and resolves via a connected fallback", () => {
+				const readConnectedProvidersSpy = spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["openai"])
+
+				const result = resolveModelForDelegateTask({
+					categoryDefaultModel: "anthropic/claude-sonnet-4-6",
+					fallbackChain: [
+						{ providers: ["openai"], model: "gpt-5.4", variant: "high" },
+					],
+					availableModels: new Set(),
+					systemDefaultModel: "anthropic/claude-sonnet-4-6",
+				})
+
+				expect(result).toEqual({
+					model: "openai/gpt-5.4",
+					variant: "high",
+					fallbackEntry: { providers: ["openai"], model: "gpt-5.4", variant: "high" },
+					matchedFallback: true,
+				})
+				readConnectedProvidersSpy.mockRestore()
+			})
+
+			test("#then skips disconnected user fallback models and keeps the first connected fallback", () => {
+				const readConnectedProvidersSpy = spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["openai"])
+
+				const result = resolveModelForDelegateTask({
+					userFallbackModels: ["anthropic/claude-sonnet-4-6", "openai/gpt-5.4"],
+					availableModels: new Set(),
+				})
+
+				expect(result).toEqual({ model: "openai/gpt-5.4", matchedFallback: true })
+				readConnectedProvidersSpy.mockRestore()
 			})
 		})
 
@@ -225,16 +261,23 @@ describe("resolveModelForDelegateTask", () => {
 		})
 
 		describe("#when availableModels is empty", () => {
-			test("#then falls through to existing resolution (cache partially ready)", () => {
+			test("#then uses connected providers to avoid disconnected category defaults", () => {
+				const readConnectedProvidersSpy = spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["openai"])
+
 				const result = resolveModelForDelegateTask({
 					categoryDefaultModel: "anthropic/claude-sonnet-4-6",
 					fallbackChain: [
-						{ providers: ["anthropic"], model: "claude-sonnet-4-6" },
+						{ providers: ["openai"], model: "gpt-5.4" },
 					],
 					availableModels: new Set(),
 				})
 
-				expect(result).toBeDefined()
+				expect(result).toEqual({
+					model: "openai/gpt-5.4",
+					fallbackEntry: { providers: ["openai"], model: "gpt-5.4" },
+					matchedFallback: true,
+				})
+				readConnectedProvidersSpy.mockRestore()
 			})
 		})
 	})

src/tools/delegate-task/model-selection.ts

@@ -59,6 +59,8 @@ export function resolveModelForDelegateTask(input: {
     return { model: userModel }
   }
 
+  const connectedProviders = input.availableModels.size === 0 ? readConnectedProvidersCache() : null
+
   // Before provider cache is created (first run), skip model resolution entirely.
   // OpenCode will use its system default model when no model is specified in the prompt.
   if (input.availableModels.size === 0 && !hasProviderModelsCache() && !hasConnectedProvidersCache()) {
@@ -77,7 +79,15 @@ export function resolveModelForDelegateTask(input: {
     }
 
     if (input.availableModels.size === 0) {
-      return { model: categoryDefault }
+      const categoryProvider = categoryDefault.includes("/") ? categoryDefault.split("/")[0] : undefined
+      if (!connectedProviders || !categoryProvider || connectedProviders.includes(categoryProvider)) {
+        return { model: categoryDefault }
+      }
+
+      log("[resolveModelForDelegateTask] skipping disconnected category default on cold cache", {
+        categoryDefault,
+        connectedProviders,
+      })
     }
 
     const parts = categoryDefault.split("/")
@@ -95,9 +105,19 @@ export function resolveModelForDelegateTask(input: {
   const userFallbackModels = input.userFallbackModels
   if (userFallbackModels && userFallbackModels.length > 0) {
     if (input.availableModels.size === 0) {
-      const first = userFallbackModels[0] ? parseUserFallbackModel(userFallbackModels[0]) : undefined
-      if (first) {
-        return { model: first.baseModel, variant: first.variant, matchedFallback: true }
+      for (const fallbackModel of userFallbackModels) {
+        const parsedFallback = parseUserFallbackModel(fallbackModel)
+        if (!parsedFallback) continue
+
+        if (
+          connectedProviders &&
+          parsedFallback.providerHint &&
+          !parsedFallback.providerHint.some((provider) => connectedProviders.includes(provider))
+        ) {
+          continue
+        }
+
+        return { model: parsedFallback.baseModel, variant: parsedFallback.variant, matchedFallback: true }
       }
     } else {
       for (const fallbackModel of userFallbackModels) {
@@ -115,7 +135,6 @@ export function resolveModelForDelegateTask(input: {
   const fallbackChain = input.fallbackChain
   if (fallbackChain && fallbackChain.length > 0) {
     if (input.availableModels.size === 0) {
-      const connectedProviders = readConnectedProvidersCache()
       if (connectedProviders) {
         const connectedSet = new Set(connectedProviders)
         for (const entry of fallbackChain) {

src/tools/delegate-task/tools.ts

@@ -76,13 +76,13 @@ export function createDelegateTask(options: DelegateTaskToolOptions): ToolDefini
   - category: For task delegation (uses Sisyphus-Junior with category-optimized model)
   - subagent_type: For direct agent invocation (explore, librarian, oracle, etc.)
   
-  **DO NOT provide both.** If category is provided, subagent_type is ignored.
+  **DO NOT provide both.** category and subagent_type are mutually exclusive.
   
   - load_skills: ALWAYS REQUIRED. Pass [] if no skills needed, or ["skill-1", "skill-2"] for category tasks.
   - category: Use predefined category → Spawns Sisyphus-Junior with category config
     Available categories:
   ${categoryList}
-  - subagent_type: Use specific agent directly (explore, librarian, oracle, metis, momus)
+  - subagent_type: Use a specific callable non-primary agent directly (for example: explore, librarian, oracle, metis, momus)
   - run_in_background: REQUIRED. true=async (returns task_id), false=sync (waits). Use background=true ONLY for parallel exploration with 5+ independent queries.
   - session_id: Existing Task session to continue (from previous task output). Continues agent with FULL CONTEXT PRESERVED - saves tokens, maintains continuity.
   - command: The command that triggered this task (optional, for slash command tracking).
@@ -102,7 +102,7 @@ export function createDelegateTask(options: DelegateTaskToolOptions): ToolDefini
       prompt: tool.schema.string().describe("Full detailed prompt for the agent"),
       run_in_background: tool.schema.boolean().describe("REQUIRED. true=async (returns task_id), false=sync (waits). Use false for task delegation, true ONLY for parallel exploration."),
       category: tool.schema.string().optional().describe(`REQUIRED if subagent_type not provided. Do NOT provide both category and subagent_type.`),
-      subagent_type: tool.schema.string().optional().describe("REQUIRED if category not provided. Do NOT provide both category and subagent_type. Valid values: explore, librarian, oracle, metis, momus"),
+      subagent_type: tool.schema.string().optional().describe("REQUIRED if category not provided. Do NOT provide both category and subagent_type. Must be a callable non-primary agent name returned by app.agents()."),
       session_id: tool.schema.string().optional().describe("Existing Task session to continue"),
       command: tool.schema.string().optional().describe("The command that triggered this task"),
     },
@@ -115,7 +115,7 @@ export function createDelegateTask(options: DelegateTaskToolOptions): ToolDefini
           `  - You provided: category="${args.category}", subagent_type="${args.subagent_type}"\n` +
           `  - Use category for task delegation (e.g., category="${categoryExamples.split(", ")[0]}")\n` +
           `  - Use subagent_type for direct agent invocation (e.g., subagent_type="explore")\n` +
-          `  - Valid subagent_type values: explore, librarian, oracle, metis, momus`
+          `  - subagent_type must be a callable non-primary agent name returned by app.agents()`
         )
       }
       if (args.category) {

src/tools/skill/tools.test.ts

@@ -616,6 +616,33 @@ describe("skill tool - browserProvider forwarding", () => {
 })
 
 describe("skill tool - nativeSkills integration", () => {
+  it("includes native skills in the description even when skills are pre-seeded", async () => {
+    //#given
+    const tool = createSkillTool({
+      skills: [createMockSkill("seeded-skill")],
+      nativeSkills: {
+        async all() {
+          return [{
+            name: "native-visible-skill",
+            description: "Native skill exposed from config",
+            location: "/external/skills/native-visible-skill/SKILL.md",
+            content: "Native visible skill body",
+          }]
+        },
+        async get() { return undefined },
+        async dirs() { return [] },
+      },
+    })
+
+    //#when
+    expect(tool.description).toContain("seeded-skill")
+    await tool.execute({ name: "native-visible-skill" }, mockContext)
+
+    //#then
+    expect(tool.description).toContain("seeded-skill")
+    expect(tool.description).toContain("native-visible-skill")
+  })
+
   it("merges native skills exposed by PluginInput.skills.all()", async () => {
     //#given
     const tool = createSkillTool({
@@ -639,6 +666,6 @@ describe("skill tool - nativeSkills integration", () => {
 
     //#then
     expect(result).toContain("external-plugin-skill")
-    expect(result).toContain("Test skill body content")
+    expect(result).toContain("External plugin skill body")
   })
 })

src/tools/skill/tools.ts

@@ -105,6 +105,11 @@ async function extractSkillBody(skill: LoadedSkill): Promise<string> {
     return templateMatch ? templateMatch[1].trim() : fullTemplate
   }
 
+  if (skill.scope === "config" && skill.definition.template) {
+    const templateMatch = skill.definition.template.match(/<skill-instruction>([\s\S]*?)<\/skill-instruction>/)
+    return templateMatch ? templateMatch[1].trim() : skill.definition.template
+  }
+
   if (skill.path) {
     return extractSkillTemplate(skill)
   }
@@ -235,11 +240,13 @@ export function createSkillTool(options: SkillLoadOptions = {}): ToolDefinition
     return cachedDescription
   }
 
-  // Eagerly build description when callers pre-provide skills/commands.
   if (options.skills !== undefined) {
     const skillInfos = options.skills.map(loadedSkillToInfo)
     const commandsForDescription = options.commands ?? []
     cachedDescription = formatCombinedDescription(skillInfos, commandsForDescription)
+    if (options.nativeSkills) {
+      void buildDescription()
+    }
   } else if (options.commands !== undefined) {
     cachedDescription = formatCombinedDescription([], options.commands)
   } else {
@@ -248,6 +255,9 @@ export function createSkillTool(options: SkillLoadOptions = {}): ToolDefinition
 
   return tool({
     get description() {
+      if (cachedDescription === null) {
+        void buildDescription()
+      }
       return cachedDescription ?? TOOL_DESCRIPTION_PREFIX
     },
     args: {
@@ -259,8 +269,8 @@ export function createSkillTool(options: SkillLoadOptions = {}): ToolDefinition
     },
     async execute(args: SkillArgs, ctx?: { agent?: string }) {
       const skills = await getSkills()
-      cachedDescription = null
       const commands = getCommands()
+      cachedDescription = formatCombinedDescription(skills.map(loadedSkillToInfo), commands)
 
       const requestedName = args.name.replace(/^\//, "")