@junghanacs/entwurf 0.30.0 → 0.30.1

package/scripts/mutants/acp-prompt-lifecycle.json

@@ -54,8 +54,8 @@
 			"claim": "REUSE-CARRIES-CHILD-DIAGNOSTICS",
 			"title": "the reuse path drops the child's stderr tail again — exactly the live sonnet failure that showed only \"ACP connection closed\"",
 			"subject": "pi-extensions/lib/acp/backend.ts",
-			"find": ["\t\t\tfinishError(err, aborted, session.stderrTail, lifecycle, session.launchObservation);"],
-			"replace": ["\t\t\tfinishError(err, aborted, undefined, lifecycle, session.launchObservation);"],
+			"find": ["\t\t\tfinishError(err, aborted, session.stderrTail, lifecycle, session.launchObservation, adapter);"],
+			"replace": ["\t\t\tfinishError(err, aborted, undefined, lifecycle, session.launchObservation, adapter);"],
 			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
 			"timeoutSeconds": 180,
 			"signature": "[QK:REUSE-CARRIES-CHILD-DIAGNOSTICS]",

package/scripts/mutants/acp-vendor-rejection.json

@@ -0,0 +1,62 @@
+{
+	"schemaVersion": 1,
+	"lane": "acp-vendor-rejection",
+	"mutants": [
+		{
+			"claim": "ACP-INCOMPLETE-TOOL-NO-REPLAY-NEW",
+			"title": "finishError stops consulting the adapter's prompt-rejection reader — the vendor's raw 'Internal error' text becomes errorMessage again and pi replays a turn whose tools already ran",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\tconst rejection = aborted ? undefined : adapter?.readPromptRejection?.(err);"],
+			"replace": ["\t\tconst rejection = aborted ? undefined : undefined;"],
+			"gate": ["bash", "run.sh", "check-acp-vendor-rejection"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:ACP-INCOMPLETE-TOOL-NO-REPLAY-NEW]",
+			"signatureSource": "pi-extensions/lib/acp/backend.vendor-rejection.test.ts"
+		},
+		{
+			"claim": "ACP-PROMPT-REJECTION-CLAUDE-ONLY",
+			"title": "the claude reader stops checking the vendor's errorKind — any -32603 carrying an errorKind (no_result included) is rewritten to the incomplete-tool verdict",
+			"subject": "pi-extensions/lib/acp/backend-adapter.ts",
+			"find": ["\tif (data.errorKind !== CLAUDE_INCOMPLETE_TOOL_CALL) return undefined;"],
+			"replace": ["\tif (data.errorKind === undefined) return undefined;"],
+			"gate": ["bash", "run.sh", "check-acp-vendor-rejection"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:ACP-PROMPT-REJECTION-CLAUDE-ONLY]",
+			"signatureSource": "pi-extensions/lib/acp/backend-adapter.test.ts"
+		},
+		{
+			"claim": "ACP-INCOMPLETE-TOOL-RAW-EVIDENCE",
+			"title": "the recognised rejection is sealed with the verdict but its raw vendor message, data and stderr are dropped instead of riding a pi diagnostic",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\t\tappendAssistantMessageDiagnostic(",
+				"\t\t\t\tstate.output,",
+				"\t\t\t\tcreateAssistantMessageDiagnostic(ACP_PROMPT_REJECTION_DIAGNOSTIC, err, {",
+				"\t\t\t\t\terrorKind: rejection.errorKind,",
+				"\t\t\t\t\tdataJson: rejection.dataJson,",
+				"\t\t\t\t\t...(lifecycle ? { lifecycle } : {}),",
+				"\t\t\t\t\t...(observed ? { launchSignal: observed } : {}),",
+				"\t\t\t\t\t...(tail ? { stderrTail: tail } : {}),",
+				"\t\t\t\t\t...(hint ? { hint } : {}),",
+				"\t\t\t\t}),",
+				"\t\t\t);"
+			],
+			"replace": ["\t\t\t// mutant: the raw rejection is not preserved"],
+			"gate": ["bash", "run.sh", "check-acp-vendor-rejection"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:ACP-INCOMPLETE-TOOL-RAW-EVIDENCE]",
+			"signatureSource": "pi-extensions/lib/acp/backend.vendor-rejection.test.ts"
+		},
+		{
+			"claim": "ACP-REUSE-COOPERATIVE-CANCEL",
+			"title": "a reused session's cooperatively cancelled turn is sealed and retained but no longer persisted — the record lags the live session",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\t\tfinishSuccess(adapter, session, promptResult);", "\t\t\tpersistRecord(session, deps);"],
+			"replace": ["\t\t\tfinishSuccess(adapter, session, promptResult);"],
+			"gate": ["bash", "run.sh", "check-acp-vendor-rejection"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:ACP-REUSE-COOPERATIVE-CANCEL]",
+			"signatureSource": "pi-extensions/lib/acp/backend.vendor-rejection.test.ts"
+		}
+	]
+}

package/scripts/mutants/herdr-plugin-build.json

@@ -68,6 +68,17 @@
 			"signature": "[QK:HPB-SEQUENTIAL-TWO-STAGE]",
 			"signatureSource": "scripts/check-herdr-plugin-build.ts"
 		},
+		{
+			"claim": "HPB-DEFAULT-LOCK-READ",
+			"title": "reading the default lock from the repository root rather than the plugin directory breaks the real build while source-injected fixture drives stay green; the production composition must consume this plugin's committed source and artifact fields",
+			"subject": "plugins/herdr/lib/build.mjs",
+			"find": ["\tconst lock = lockReader(PLUGIN_DIR);"],
+			"replace": ["\tconst lock = lockReader(CHECKOUT_ROOT);"],
+			"gate": ["bash", "run.sh", "check-herdr-plugin-build"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:HPB-DEFAULT-LOCK-READ]",
+			"signatureSource": "scripts/check-herdr-plugin-build.ts"
+		},
 		{
 			"claim": "HPB-POST-BUILD-GAP",
 			"title": "claiming Herdr's registration in this runner's own receipt states a step that has not run: herdr commits its checkout and registry only AFTER the build returns, so the next reconcile would be arguing with a lie",

package/scripts/mutants/herdr-supply.json

@@ -8,7 +8,7 @@
 			"subject": "scripts/check-herdr-sandbox.ts",
 			"find": ["const SUPPLY_MANIFEST = \"scripts/fixtures/herdr-supply.json\";"],
 			"replace": [
-				"const SUPPLY_MANIFEST = \"scripts/fixtures/herdr-supply.json\";\nconst PINNED_FALLBACK = \"0.9.1\";\nvoid PINNED_FALLBACK;"
+				"const SUPPLY_MANIFEST = \"scripts/fixtures/herdr-supply.json\";\nconst PINNED_FALLBACK = \"0.9.3\";\nvoid PINNED_FALLBACK;"
 			],
 			"gate": ["bash", "run.sh", "check-herdr-supply"],
 			"timeoutSeconds": 120,
@@ -19,8 +19,8 @@
 			"claim": "HSUP-DIGEST-SHAPE",
 			"title": "a pinned digest is truncated, so the check that is supposed to catch a swapped binary can no longer fail",
 			"subject": "scripts/fixtures/herdr-supply.json",
-			"find": ["\"sha256\": \"2a02fed16beb651ef006e1d43f048f652ca4dc58ad053cd2d44450563d5c54b7\""],
-			"replace": ["\"sha256\": \"2a02fed16beb651ef006e1d43f04\""],
+			"find": ["\"sha256\": \"18a8dc65f1c2fa485884344356dea1cfd911c6f06cf46fa78e193f4087f4dba7\""],
+			"replace": ["\"sha256\": \"18a8dc65f1c2fa485884344356de\""],
 			"gate": ["bash", "run.sh", "check-herdr-supply"],
 			"timeoutSeconds": 120,
 			"signature": "[QK:HSUP-DIGEST-SHAPE]",

package/scripts/smoke-acp-raw-turn-live.ts

@@ -3,7 +3,7 @@
 //   LIVE=1 ./run.sh smoke-acp-raw-turn-live
 //
 // What this proves (and ONLY this): the pinned Claude ACP adapter
-// (@agentclientprotocol/claude-agent-acp@0.79.0) spawns, speaks the ACP wire
+// (@agentclientprotocol/claude-agent-acp, exact pin in package.json) spawns, speaks the ACP wire
 // protocol over stdio NDJSON, and returns one real model turn. It is the
 // bytes-flow proof that the S2a dep surface is not just installable but
 // actually drivable — before any provider/overlay/streamSimple code (S2b+).

package/scripts/smoke-herdr-fresh-call-live.ts

@@ -196,10 +196,9 @@ function transcriptRecords(file: string): string[] {
 }
 
 /**
- * The BARE verb names, as the two surfaces this smoke reads actually spell them — NOT the
- * model-facing dialect. `[측정 2026-09-20, this smoke's own red fixture]` the claude MCP activity
- * log carries `Calling MCP tool: entwurf_callback` / `Tool 'entwurf_callback' completed`, and pi
- * writes `"type":"toolCall","name":"entwurf_callback"`; neither records
+ * The verb names as the CLAUDE MCP activity log spells them — NOT the model-facing dialect.
+ * `[측정 2026-09-20, this smoke's own red fixture]` that log carries `Calling MCP tool:
+ * entwurf_callback` / `Tool 'entwurf_callback' completed`, never
  * `mcp__entwurf-bridge__entwurf_callback`. The per-host dialect is owned by
  * `FRESH_CALL_CALLBACK_TOOL` / `FRESH_CALL_DELIVERY_TOOL` in `fresh-call-composition.ts` and is
  * what the FRAMING says; these two are what an OBSERVER sees, and conflating them is what this
@@ -208,6 +207,17 @@ function transcriptRecords(file: string): string[] {
 const CALLBACK_VERB = "entwurf_callback";
 const DELIVERY_VERB = "entwurf_v2";
 
+/**
+ * The callback as a Pi 1.0 TRANSCRIPT spells it. Pi's built-in MCP is the tool ingress since
+ * 0.30.0, and Pi records the qualified name the model called. The bare `entwurf_callback` this
+ * smoke read on 2026-09-20 belonged to the retired native-tool registration; a pi→pi child that
+ * called back correctly would then never match, which is a red about the oracle, not the rail.
+ * Declared here as the OBSERVED literal, not imported from the compositor, so a change to the
+ * production framing cannot quietly move this expectation with it — the same rule as
+ * `PI_SOURCE_VERBS` in `lib/codex-fresh-source-receipts.ts`. No prefix strip, no legacy OR.
+ */
+const PI_CALLBACK_TOOL = "mcp__entwurf_bridge__entwurf_callback";
+
 /**
  * The id of the CALLBACK toolCall in this record, or null. Pi writes a call and its result as two
  * records; this is one half of the join that replaces a same-record read.
@@ -238,7 +248,9 @@ function entwurfCallIdFor(record: string, nonce: string, callerGid: string): str
 	if (!Array.isArray(content)) return null;
 	for (const part of content) {
 		const call = part as { type?: unknown; name?: unknown; id?: unknown; arguments?: unknown };
-		if (call.type !== "toolCall" || call.name !== CALLBACK_VERB) continue;
+		if (call.type !== "toolCall") continue;
+		if (call.name === CALLBACK_VERB) throw new Error(`${LABEL}: retired raw Pi tool name ${CALLBACK_VERB}`);
+		if (call.name !== PI_CALLBACK_TOOL) continue;
 		// Zero-argument verb: absent, null and `{}` are all the shape it is called with. A supplied
 		// target or message is a model naming an address the env already owns — not this act.
 		const args = call.arguments as { message?: unknown; target?: unknown } | undefined | null;
@@ -444,8 +456,10 @@ async function main(): Promise<void> {
 			childBackend: "pi",
 			callerKind: "pi",
 			childKind: "pi",
-			callerModel: "openai-codex/gpt-5.6-sol",
-			childModel: "openai-codex/gpt-5.6-sol",
+			// The approved roster's lighter Pi GPT tier (agent-config MODELS.md, 2026-09-30); gpt-5.6-sol
+			// is still in the catalog but no longer on the roster. A LIVE cost choice, not a product set.
+			callerModel: "openai-codex/gpt-6-luna",
+			childModel: "openai-codex/gpt-6-luna",
 		},
 		{
 			label: "claude→claude",