@poncho-ai/sdk 1.10.0 → 1.12.0

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/sdk@1.10.0 build /home/runner/work/poncho-ai/poncho-ai/packages/sdk
+> @poncho-ai/sdk@1.12.0 build /home/runner/work/poncho-ai/poncho-ai/packages/sdk
 > tsup src/index.ts --format esm --dts
 
 CLI Building entry: src/index.ts
@@ -8,7 +8,7 @@
 CLI Target: es2022
 ESM Build start
 ESM dist/index.js 17.24 KB
-ESM ⚡️ Build success in 21ms
+ESM ⚡️ Build success in 20ms
 DTS Build start
-DTS ⚡️ Build success in 1318ms
-DTS dist/index.d.ts 28.29 KB
+DTS ⚡️ Build success in 1257ms
+DTS dist/index.d.ts 29.41 KB

package/CHANGELOG.md

@@ -1,5 +1,68 @@
 # @poncho-ai/sdk
 
+## 1.12.0
+
+### Minor Changes
+
+- [#118](https://github.com/cesr/poncho-ai/pull/118) [`e8df464`](https://github.com/cesr/poncho-ai/commit/e8df4649618cba0b408a6c143f923f0dcb2046c8) Thanks [@cesr](https://github.com/cesr)! - harness: 1h static system-prompt cache breakpoint + per-run cache kill-switch
+
+  Two related changes to Anthropic prompt caching:
+
+  **1-hour static system-prompt breakpoint.** The harness now splits the
+  assembled system prompt into a static portion (agent body + skill
+  context + browser/fs/isolate context — stable across many turns and
+  jobs within an hour) and a dynamic tail (memory, todos, time). On
+  Anthropic models, these are sent as two `role: "system"` messages with
+  `cacheControl: { ttl: "1h" }` on the static block. The existing 5-min
+  tail breakpoint on the last user/assistant/tool message is retained.
+
+  This lets later turns and job runs read ~95% of the system prompt at
+  0.1× (cache read) instead of paying 1× whenever the 5-min tail cache
+  has expired — the previous setup only cached for 5 minutes via the
+  tail breakpoint. Within-user cross-conversation and interactive-vs-job
+  all share the static cache.
+
+  **Per-run cache kill-switch.** Added `RunInput.disablePromptCache?:
+boolean` (also exposed on `RunConversationTurnOpts.disablePromptCache`,
+  forwarded into `runInput`). When set, the harness skips the 5-min tail
+  breakpoint for that run. The 1-hour static breakpoint is still
+  applied — the run still benefits from reading the shared static cache,
+  just doesn't write a new tail entry that won't be read before TTL.
+
+  Intended for one-shot programmatic invocations (cron-fired jobs,
+  subagent dispatch) where no follow-up turn is coming within the 5-min
+  TTL window, so the 1.25× write surcharge would be pure waste.
+
+  Non-Anthropic providers fall through to the previous single concatenated
+  `system:` string with no cache control — those providers auto-cache.
+
+  Internal: `isAnthropicModel` is now exported from `prompt-cache.ts`
+  for reuse at the streamText site.
+
+## 1.11.0
+
+### Minor Changes
+
+- [`1adaae2`](https://github.com/cesr/poncho-ai/commit/1adaae2d4cc55800f01d602f2a7d6ecc65031443) Thanks [@cesr](https://github.com/cesr)! - harness: device-dispatch mode for tools that execute on a connected client
+
+  Tools can now be marked `dispatch: "device"` on `loadedConfig.tools`. When
+  the model calls such a tool the dispatcher pauses the run, emits a new
+  `tool:device:required` event, and checkpoints with the new
+  `kind: "device"` discriminator on `pendingApprovals` — same plumbing as
+  the approval flow, different trigger and different resume payload.
+  Consumers (e.g. PonchOS for iOS device tools) drive the external
+  execution and feed the result back via `continueFromToolResult`.
+
+  Approval can be combined: `{access: "approval", dispatch: "device"}`
+  yields the approval card first, then on resume falls through to the
+  device-required event. The wire vocabulary for approvals
+  (`approvalId` etc.) is unchanged; the `pendingApprovals` column /
+  field name stays.
+
+  `ToolAccess` is broadened to accept both the legacy string `"approval"`
+  and the new `{access?, dispatch?}` object form. Existing configs keep
+  working unchanged.
+
 ## 1.10.0
 
 ### Minor Changes

package/dist/index.d.ts

@@ -747,6 +747,13 @@ interface RunInput {
     conversationId?: string;
     /** When true, ignores PONCHO_MAX_DURATION soft deadline (used for background subagent runs). */
     disableSoftDeadline?: boolean;
+    /**
+     * When true, skip the Anthropic prompt-cache breakpoint for this run.
+     * Use for one-shot runs with no follow-up turn coming (e.g. cron-fired
+     * jobs) — the 1.25× write surcharge is pure waste when no later read
+     * will hit the cache before the 5-min TTL expires.
+     */
+    disablePromptCache?: boolean;
     /** Scope this run to a specific tenant. */
     tenantId?: string;
 }
@@ -858,6 +865,32 @@ type AgentEvent = {
         name: string;
         input: Record<string, unknown>;
     }>;
+} | {
+    /**
+     * Tool wants to execute on a connected client device (e.g. iOS).
+     * The consumer of the harness is responsible for routing this event
+     * to the appropriate WebSocket and POSTing the tool's result back via
+     * `resumeRunFromCheckpoint`. Carries the same envelope as the
+     * approval-required event; `requestId` plays the role of `approvalId`.
+     */
+    type: "tool:device:required";
+    tool: string;
+    input: unknown;
+    requestId: string;
+} | {
+    type: "tool:device:checkpoint";
+    approvals: Array<{
+        approvalId: string;
+        tool: string;
+        toolCallId: string;
+        input: Record<string, unknown>;
+    }>;
+    checkpointMessages: Message[];
+    pendingToolCalls: Array<{
+        id: string;
+        name: string;
+        input: Record<string, unknown>;
+    }>;
 } | {
     type: "browser:frame";
     data: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/sdk",
-  "version": "1.10.0",
+  "version": "1.12.0",
   "description": "Core types and utilities for building Poncho skills",
   "repository": {
     "type": "git",

package/src/index.ts

@@ -128,6 +128,13 @@ export interface RunInput {
   conversationId?: string;
   /** When true, ignores PONCHO_MAX_DURATION soft deadline (used for background subagent runs). */
   disableSoftDeadline?: boolean;
+  /**
+   * When true, skip the Anthropic prompt-cache breakpoint for this run.
+   * Use for one-shot runs with no follow-up turn coming (e.g. cron-fired
+   * jobs) — the 1.25× write surcharge is pure waste when no later read
+   * will hit the cache before the 5-min TTL expires.
+   */
+  disablePromptCache?: boolean;
   /** Scope this run to a specific tenant. */
   tenantId?: string;
 }
@@ -196,6 +203,30 @@ export type AgentEvent =
       checkpointMessages: Message[];
       pendingToolCalls: Array<{ id: string; name: string; input: Record<string, unknown> }>;
     }
+  | {
+      /**
+       * Tool wants to execute on a connected client device (e.g. iOS).
+       * The consumer of the harness is responsible for routing this event
+       * to the appropriate WebSocket and POSTing the tool's result back via
+       * `resumeRunFromCheckpoint`. Carries the same envelope as the
+       * approval-required event; `requestId` plays the role of `approvalId`.
+       */
+      type: "tool:device:required";
+      tool: string;
+      input: unknown;
+      requestId: string;
+    }
+  | {
+      type: "tool:device:checkpoint";
+      approvals: Array<{
+        approvalId: string;
+        tool: string;
+        toolCallId: string;
+        input: Record<string, unknown>;
+      }>;
+      checkpointMessages: Message[];
+      pendingToolCalls: Array<{ id: string; name: string; input: Record<string, unknown> }>;
+    }
   | { type: "browser:frame"; data: string; width: number; height: number }
   | {
       type: "browser:status";