@schoolai/shipyard 3.19.0 → 3.20.0-rc.20260723.0

package/dist/capability-detector-worker.js

@@ -3,7 +3,7 @@ import {
   detectCapabilities,
   refreshMcpServers,
   registerBuiltInProfiles
-} from "./chunk-LVQLRQKF.js";
+} from "./chunk-OBYRPZRA.js";
 import "./chunk-TUQTEWK3.js";
 import {
   getRepoMetadata
@@ -15,9 +15,9 @@ import {
 import "./chunk-RR6V6SNM.js";
 import "./chunk-ESCSN2NN.js";
 import "./chunk-ZFKJAYAN.js";
-import "./chunk-FASB7KFF.js";
-import "./chunk-6TK5KO76.js";
-import "./chunk-HD66ZG4O.js";
+import "./chunk-LYUR7G6B.js";
+import "./chunk-OC75AUKJ.js";
+import "./chunk-TJ4TMMMB.js";
 import "./chunk-EHQITHQX.js";
 import "./chunk-X3MULCV5.js";
 import {

package/dist/{chunk-KMMF5R3M.js → chunk-FMHVDAJK.js}

@@ -18,8 +18,8 @@ function getDaemonVersion() {
   return cached;
 }
 function readDaemonVersion() {
-  if ("3.19.0".length > 0) {
-    return "3.19.0";
+  if ("3.20.0".length > 0) {
+    return "3.20.0";
   }
   try {
     const here = dirname(fileURLToPath(import.meta.url));
@@ -48,4 +48,4 @@ function readDaemonVersion() {
 export {
   getDaemonVersion
 };
-//# sourceMappingURL=chunk-KMMF5R3M.js.map
+//# sourceMappingURL=chunk-FMHVDAJK.js.map

package/dist/{chunk-FASB7KFF.js → chunk-LYUR7G6B.js}

@@ -2,7 +2,7 @@
 import {
   isImageMimeType,
   parseAssetUri
-} from "./chunk-6TK5KO76.js";
+} from "./chunk-OC75AUKJ.js";
 import {
   assertNever
 } from "./chunk-X3MULCV5.js";
@@ -588,4 +588,4 @@ export {
   toRecord,
   buildCursorUserPrompt
 };
-//# sourceMappingURL=chunk-FASB7KFF.js.map
+//# sourceMappingURL=chunk-LYUR7G6B.js.map

package/dist/{chunk-NRXFTDBH.js → chunk-MD32Y6U4.js}

@@ -15,7 +15,7 @@ import {
   DevicePollResponseSchema,
   DeviceStartResponseSchema,
   ROUTES
-} from "./chunk-HD66ZG4O.js";
+} from "./chunk-TJ4TMMMB.js";
 import {
   isDevMode,
   validateEnv
@@ -223,4 +223,4 @@ export {
   ensureAuthenticated,
   loginCommand
 };
-//# sourceMappingURL=chunk-NRXFTDBH.js.map
+//# sourceMappingURL=chunk-MD32Y6U4.js.map

package/dist/{chunk-LVQLRQKF.js → chunk-OBYRPZRA.js}

@@ -38,7 +38,7 @@ import {
   formatTaskFeedback,
   toRecord,
   toSdkContent
-} from "./chunk-FASB7KFF.js";
+} from "./chunk-LYUR7G6B.js";
 import {
   AGENT_IDS,
   AGENT_SYSTEM_CLAUDE_CODE,
@@ -50,14 +50,14 @@ import {
   REASONING_EFFORTS,
   inflateToolResultContent,
   isImageMimeType
-} from "./chunk-6TK5KO76.js";
+} from "./chunk-OC75AUKJ.js";
 import {
   AnthropicAuthStatusSchema,
   CodexAuthStatusSchema,
   CursorAuthStatusSchema,
   PermissionModeSchema,
   TESTED_CODEX_VERSION
-} from "./chunk-HD66ZG4O.js";
+} from "./chunk-TJ4TMMMB.js";
 import {
   logger
 } from "./chunk-X53VIV2J.js";
@@ -3353,7 +3353,7 @@ var codexProfile = {
       return null;
     },
     async applyLive(subprocess, target) {
-      if (target.kind !== "dropLastTurns") return { handled: false };
+      if (target.kind !== "dropLastTurns" || target.numTurns <= 0) return { handled: false };
       await subprocess.rollback(target.numTurns);
       return { handled: true };
     }
@@ -3701,7 +3701,11 @@ var codexProfile = {
    * codex-subprocess-spawn-helpers.ts; add the rendered section there.
    */
   subagentBackend: {
-    /** Codex's `spawn_agent` MCP tool launches a sub-thread. */
+    /**
+     * Codex's native `spawn_agent` tool launches a sub-thread. It is built
+     * into the Codex runtime — NOT a Shipyard harness MCP tool (#4784
+     * showed the bare word "MCP" here misleads; say which surface owns it).
+     */
     spawnTool: "spawn_agent",
     /**
      * Progress lands as collab-tool-call updates on the parent thread —
@@ -3855,7 +3859,7 @@ var codexProfile = {
    */
   systemPromptFragments: {
     backgroundExecution: "Codex does not support background execution. Long-running shell commands block the current turn \u2014 keep them short or break work into smaller steps so the user sees progress.",
-    planMode: "In plan mode Codex runs read-only \u2014 file edits and shell writes are blocked. Use this turn to research the codebase and propose a plan in your final message; Shipyard captures that message as the plan and surfaces it for review. Do not call apply_patch or destructive shell commands until the user approves and exits plan mode.",
+    planMode: "In plan mode Codex runs under a read-only filesystem sandbox with network enabled \u2014 apply_patch and mutating shell commands are blocked, while shell reads and network research are allowed; update_plan is also blocked in plan mode. Publish your plan by wrapping it in a <proposed_plan>...</proposed_plan> block in your streamed message \u2014 Shipyard extracts the block and surfaces it for review; text outside the block is chat, not plan. Do not mutate files until reviewers approve.",
     todoGuidance: "Use the update_plan tool to keep the user-visible task list in sync with your work. When starting, include any items already visible in <task-list> using their exact subjects \u2014 then add your own steps below them. Don't paraphrase template items or you'll create duplicate sidebar entries. Update statuses (pending \u2192 in_progress \u2192 completed) as you progress; call update_plan whenever you start, finish, or reorder a step.",
     toolReferences: {
       subagentSpawn: "spawn_agent",
@@ -3884,7 +3888,33 @@ var codexProfile = {
    * system-prompt assembly path (~2500 chars / 4 ≈ 625 tokens as of 2026-05-25).
    */
   systemPromptOverheadTokens: 625,
-  stallTimeoutMs: null,
+  /**
+   * Stall watchdog (#4921). Codex previously set every window to null, so no
+   * watchdog ever armed — an overnight Codex hang had no recovery path. The
+   * watchdog machinery is agent-neutral (`resolveStallBudgets` reads these
+   * fields; the pure FSM resets on `stream_activity`), so declaring the windows
+   * here is what turns it on. Codex liveness is emitted per turn frame from
+   * `codex-turn-event-dispatch.ts` (`onLiveness` → `stream_activity`), so the
+   * timer measures TRUE silence (a genuine hang), not turn cadence.
+   *
+   * 180 s mid-run silence — the agent-operation timeout floor; matches Cursor.
+   */
+  stallTimeoutMs: 18e4,
+  /**
+   * 5 min from spawn to the first liveness frame. Covers Codex's `app-server`
+   * spawn + `thread/start` handshake with generous headroom. Because Codex
+   * emits `stream_activity` (not `step_progress`), the cold-start latch is
+   * never retired — this window stays in force all session, which is fine: it
+   * still catches a multi-hour hang far below the "overnight" failure it fixes.
+   */
+  firstStepStallTimeoutMs: 3e5,
+  /**
+   * Per-tool extended window stays disabled for v1 — Codex emits no
+   * `tool_execution_started` / `tool_execution_settled` edges, so there is
+   * nothing to latch. A silent long-running shell command (no output frames
+   * for >300 s) is the known v1 gap; adding the tool-execution edges is the
+   * follow-up, mirroring Cursor's #4229 fix.
+   */
   toolExecutionStallTimeoutMs: null
 };
 
@@ -5410,8 +5440,8 @@ var cursorProfile = {
    */
   systemPromptFragments: {
     backgroundExecution: "Cursor does not support background execution. Long-running Shell commands block the current turn \u2014 keep them short or break work into smaller steps so the user sees progress.",
-    planMode: 'In plan mode Cursor runs read-only \u2014 `write`, `edit`, `delete`, `shell`, and `task` are blocked at the hook layer. Research the codebase, then propose your plan by calling `createPlan({ plan: "..." })` with the full plan as markdown \u2014 that tool call is how Shipyard captures the plan, so a plan written only as a chat message will not be captured. Wait for the user to approve before attempting writes.',
-    todoGuidance: 'Use the shipyard_update_plan MCP tool to keep the user-visible task list in sync with your work. When starting, include any items already visible in <task-list> with their visible IDs (for example, use id: "1" for #1). Keep that same ID if you rename a template item, and omit id only for brand-new steps. Update statuses (pending \u2192 in_progress \u2192 completed) as you progress; call shipyard_update_plan whenever you start, finish, or reorder a step. Use the task tool to spawn subagents for parallel work \u2014 that is a separate concern.',
+    planMode: 'In plan mode Cursor runs read-only \u2014 `Write`, `Edit`, `MultiEdit`, `Shell`, and `applyPatch` are denied at the hook layer, and the native `task` subagent tool must not be called (blocked by instruction, not the hook). Research the codebase, then propose your plan by calling `createPlan({ plan: "..." })` with the full plan as markdown \u2014 that tool call is how Shipyard captures the plan, so a plan written only as a chat message will not be captured. Wait for the user to approve before attempting writes.',
+    todoGuidance: 'Use the shipyard_update_plan MCP tool to keep the user-visible task list in sync with your work. When starting, include any items already visible in <task-list> with their visible IDs (for example, use id: "1" for #1). Keep that same ID if you rename a template item, and omit id only for brand-new steps. Update statuses (pending \u2192 in_progress \u2192 completed) as you progress; call shipyard_update_plan whenever you start, finish, or reorder a step. Use the native `task` tool to spawn subagents for parallel work \u2014 that is a separate concern (and distinct from the Shipyard `task` MCP tool, which manages sibling tasks).',
     toolReferences: {
       subagentSpawn: "task",
       shellExec: "Shell",
@@ -8251,4 +8281,4 @@ export {
   refreshInstalledAgents,
   refreshCodexAuthSlice
 };
-//# sourceMappingURL=chunk-LVQLRQKF.js.map
+//# sourceMappingURL=chunk-OBYRPZRA.js.map