@poncho-ai/sdk 1.11.0 → 1.13.0

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/sdk@1.11.0 build /home/runner/work/poncho-ai/poncho-ai/packages/sdk
+> @poncho-ai/sdk@1.13.0 build /home/runner/work/poncho-ai/poncho-ai/packages/sdk
 > tsup src/index.ts --format esm --dts
 
 CLI Building entry: src/index.ts
@@ -10,5 +10,5 @@
 ESM dist/index.js 17.24 KB
 ESM ⚡️ Build success in 21ms
 DTS Build start
-DTS ⚡️ Build success in 1385ms
-DTS dist/index.d.ts 29.09 KB
+DTS ⚡️ Build success in 1513ms
+DTS dist/index.d.ts 29.81 KB

package/CHANGELOG.md

@@ -1,5 +1,54 @@
 # @poncho-ai/sdk
 
+## 1.13.0
+
+### Minor Changes
+
+- [`773f113`](https://github.com/cesr/poncho-ai/commit/773f11309e2410d6c5e17af0fde17425953105f2) Thanks [@cesr](https://github.com/cesr)! - harness: add a per-run `suppressTelemetry` flag so one harness can serve both telemetry-on and telemetry-off runs.
+
+  Telemetry was effectively an instance-level property: whether the OTLP exporter is attached is decided at construction, so a host that wants telemetry-off runs (e.g. incognito) had to build and maintain a _second_ harness instance with no exporter — duplicating all per-harness provisioning (tool registration, subagent manager, etc.) and risking drift between the two.
+
+  `RunInput.suppressTelemetry` lets a single harness — built once, with the exporter attached — emit nothing for a given run: the `invoke_agent` root span, the `execute_tool` spans, and the AI-SDK spans are all gated on `!input.suppressTelemetry`. Hosts can now keep one harness per user and pass `suppressTelemetry: true` per run instead of routing to a parallel exporter-less instance.
+
+## 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

package/dist/index.d.ts

@@ -747,8 +747,23 @@ 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;
+    /**
+     * When true, emit no telemetry for this run — no `invoke_agent` /
+     * `execute_tool` spans and no AI-SDK spans, even on a harness built with an
+     * OTLP exporter attached. Lets a single harness serve both telemetry-on and
+     * telemetry-off (e.g. incognito) runs, instead of needing a separate
+     * exporter-less harness instance per mode.
+     */
+    suppressTelemetry?: boolean;
 }
 interface TokenUsage {
     input: number;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/sdk",
-  "version": "1.11.0",
+  "version": "1.13.0",
   "description": "Core types and utilities for building Poncho skills",
   "repository": {
     "type": "git",

package/src/index.ts

@@ -128,8 +128,23 @@ 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;
+  /**
+   * When true, emit no telemetry for this run — no `invoke_agent` /
+   * `execute_tool` spans and no AI-SDK spans, even on a harness built with an
+   * OTLP exporter attached. Lets a single harness serve both telemetry-on and
+   * telemetry-off (e.g. incognito) runs, instead of needing a separate
+   * exporter-less harness instance per mode.
+   */
+  suppressTelemetry?: boolean;
 }
 
 export interface TokenUsage {