dsh-agy-link 0.4.20 → 0.4.22

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## 0.4.22 (2026-08-27)
+
+- **Fixed: `registration.adapter.prepareCall is not a function` on new DSH hosts (本轮运行失败 UNKNOWN).**
+  - **Field report**: after upgrading the DSH host (manjh's source checkout), every turn failed immediately with `registration.adapter.prepareCall is not a function` — category UNKNOWN in the GUI, no reply at all.
+  - **Root cause (interface drift)**: dsh-llm >= 0.1.1-rc.2 added `LlmAdapter.prepareCall(provider, model, signal)` and `LlmRuntime.prepareCall()` now calls `registration.adapter.prepareCall(...)` **unconditionally** (both the prepared-call path and the direct `adapterStream` path). The plugin's devDependency pinned `@deepseek-ai/dsh-llm ^0.1.0-rc.6` (base class without `prepareCall`); when the plugin's adapter class resolved an older dsh-llm copy than the host runtime, `registration.adapter.prepareCall` was `undefined` and every turn threw a bare TypeError.
+  - **Fix**: `AgyAdapter` now implements `prepareCall` explicitly, returning `{ model: await this.resolveModel(...), stream: options => this.stream(options) }` — the exact one-generation capability-bound handle the new runtime expects (same shape as the base default, but present regardless of which dsh-llm copy the plugin resolves). Backward compatible: runtimes < 0.1.1-rc.2 never call it. The method is declared structurally (no `override` / new-type import) so the plugin keeps typechecking against dsh-llm `^0.1.0-rc.6` and the repo's `npm ci` peer resolution stays intact; the runtime contract matches the 0.1.1-rc.2 `PreparedAdapterCall` shape.
+  - Regression test pins the contract end-to-end: `prepareCall` returns resolved model metadata (id/provider/defaultMaxTokens) plus a stream that runs a full fake turn.
+
+## 0.4.21 (2026-08-27)
+
+- **Fixed: Silent `agy exited with code 1` Hid the Real Cause (发送消息无回复、只见 exit 1).**
+  - **Field report**: a user session log showed every turn failing with the bare message `agy exited with code 1` (code PROCESS_EXIT, empty stderr, zero tokens) — each attempt ran ~7–10s, failed once, retried once, failed again, and the UI never surfaced WHY (no reply text at all).
+  - **Root cause of the blindness (verified live against agy 1.1.22)**: agy reports its failure as a `result` envelope on **stdout** (`{"event":"result",...,"status":"ERROR","error":"<human-readable reason>"}`) and exits 1 with EMPTY stderr — e.g. an invalid model/effort pairing fails instantly with `--model gemini-3.7-flash requires --effort (available: low, medium, high)`. The adapter's non-zero-exit branch dropped that envelope entirely: rate-limit-shaped errors were still classified (the classifier already read `lastResultError`), but any OTHER failure degraded to the bare exit-code line with no cause.
+  - **Fix**: the PROCESS_EXIT failure message now prefers the stdout envelope's error text (falling back to the stderr tail). Same error code and retry policy; users finally see agy's own reason, e.g. `agy exited with code 1: upstream request failed while generating` or `…: --model gemini-3.7-flash requires --effort (available: low, medium, high)`.
+  - **Diagnostics unaffected**: `/agy doctor` (`~/.dsh/agy-link/diagnostics/doctor-*.md`) still carries the full redacted raw stdout tail for deeper incidents.
+  - Regression test pins the exact silent-failure shape (stdout envelope + exit 1 + empty stderr) via a new `exit-error` fake-agy mode.
+
 ## 0.4.20 (2026-08-27)
 
 - **Fixed: Primary Account Quota Fetched With a STALE Token (主账号刷新/同步拿的值不对).**

package/dist/index.js

@@ -2127,6 +2127,25 @@ var AgyAdapter = class extends LlmAdapter {
 		}
 		return resolved;
 	}
+	/**
+	* Bind exact model metadata and dispatch to ONE adapter generation.
+	* Required by dsh-llm >= 0.1.1-rc.2: LlmRuntime.prepareCall calls
+	* registration.adapter.prepareCall(provider, model, signal) unconditionally,
+	* so an adapter compiled against the old base class (no prepareCall) crashed
+	* with "registration.adapter.prepareCall is not a function" on new hosts.
+	* Implemented explicitly here so both old and new runtimes work: the old
+	* runtime never calls it, the new runtime gets the capability-bound handle.
+	* Declared without `override`/imported types on purpose: the plugin still
+	* typechecks against dsh-llm ^0.1.0-rc.6 (no prepareCall in the base), while
+	* the runtime contract matches the 0.1.1-rc.2 PreparedAdapterCall shape
+	* { model: LlmResolvedModelInfo; stream(options): AsyncIterable<StreamChunk> }.
+	*/
+	async prepareCall(provider, model, signal) {
+		return {
+			model: await this.resolveModel(provider, model, signal),
+			stream: (options) => this.stream(options)
+		};
+	}
 	/** Build the agy argv for one call. Exported for tests. */
 	buildArgs(opts) {
 		const args = [
@@ -2430,12 +2449,14 @@ var AgyAdapter = class extends LlmAdapter {
 					message: "Google Antigravity quota / rate limit reached: " + bestMsg
 				};
 			} else if (!consumable) {
-				if (outcome.code !== 0) failure = {
-					kind: "error",
-					code: Err.PROCESS_EXIT,
-					message: "agy exited with code " + outcome.code + (outcome.stderrTail !== "" ? ": " + brief(outcome.stderrTail) : "")
-				};
-				else if (parser.stats.lastResultError) failure = {
+				if (outcome.code !== 0) {
+					const detail = parser.stats.lastResultError ?? (outcome.stderrTail !== "" ? brief(outcome.stderrTail) : "");
+					failure = {
+						kind: "error",
+						code: Err.PROCESS_EXIT,
+						message: "agy exited with code " + outcome.code + (detail !== "" ? ": " + detail : "")
+					};
+				} else if (parser.stats.lastResultError) failure = {
 					kind: "error",
 					code: Err.AGY_ERROR,
 					message: "agy reported an error: " + parser.stats.lastResultError

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-agy-link",
-  "version": "0.4.20",
+  "version": "0.4.22",
   "description": "Google Antigravity (agy CLI) models for DeepSeek Harness — stream Gemini/Claude/GPT-OSS subscriptions into DSH with thinking, tool activity, token usage and in-GUI Google OAuth login.",
   "type": "module",
   "license": "MIT",