@junghanacs/entwurf 0.12.9 → 0.13.0

package/scripts/check-acp-provider-surface.ts

@@ -1,5 +1,6 @@
 // Deterministic gate for the ACP provider registration surface (S0 loader/fence,
-// updated S2c when the real backend replaced the fail-loud stub).
+// updated S2c when the real backend replaced the fail-loud stub, and again in
+// 0.13 when cortex joined claude on the adapter rail).
 //
 // Three layers:
 //   (1 lib)     loads the REAL lib modules and asserts the curated Claude
@@ -8,10 +9,13 @@
 //               imports resolve to real emitted `.js`), imports the compiled
 //               acp-provider.js, and drives its default export against a fake pi
 //               that captures registerProvider — real execution capture of the
-//               actual entry, idempotency included, and that streamSimple is the
-//               real streamShellAcp backend (by name, NOT invoked — that spawns);
+//               actual entry, idempotency included, the EXACT curated model set
+//               BOTH adapters contribute, and that streamSimple is the real
+//               streamShellAcp backend (by name, NOT invoked — that spawns);
 //   (3 source)  an auxiliary source-shape lock on acp-provider.ts.
 //
+// [QK:*] labels mark the claims kill-qualified by scripts/mutants/acp-cortex.json.
+//
 // Layer 2 is the GPT-reviewed resolution to a fence tension: acp-provider.ts
 // imports its lib with `.js` suffixes (the root/jiti runtime convention), which
 // plain `node --experimental-strip-types` cannot resolve to `.ts`. Rather than
@@ -22,7 +26,7 @@
 
 import { strict as assert } from "node:assert";
 import { execFileSync } from "node:child_process";
-import { readFileSync, rmSync } from "node:fs";
+import { readFileSync, rmdirSync, rmSync } from "node:fs";
 import { resolve } from "node:path";
 import { pathToFileURL } from "node:url";
 import {
@@ -30,6 +34,8 @@ import {
 	curatedClaudeModels,
 	ENTWURF_ACP_NO_AUTH_SENTINEL,
 	PROVIDER_ID,
+	SUPPORTED_ANTHROPIC_MODEL_IDS,
+	SUPPORTED_CORTEX_MODEL_IDS,
 } from "../pi-extensions/lib/acp/models.ts";
 
 // ---------------------------------------------------------------------------
@@ -120,6 +126,32 @@ try {
 	for (const want of ["claude-sonnet-5", CURATED_ANCHOR_MODEL_ID]) {
 		assert.ok(capIds.includes(want), `entry model surface missing ${want} (got: ${capIds.join(", ") || "none"})`);
 	}
+
+	// The REAL entry must register the curated rows of EVERY adapter on the rail —
+	// claude's 2 unprefixed ids AND cortex's 4 `cortex-` ids. This is the JOIN
+	// claim, distinct from the per-backend data claims: check-acp-cortex pins the
+	// 4-row cortex curation in models.ts, but nothing there proves those rows
+	// actually reach `pi.registerProvider`. An entry that filtered `cortex-` out
+	// of `allCuratedModels()` would leave models.ts and check-acp-cortex fully
+	// green while the operator's model list silently lost the whole backend.
+	//
+	// Set comparison (sorted, exact) — not `includes`: a substring/anchor probe
+	// cannot see a row that VANISHED, and the cortex ids contain the claude ids as
+	// substrings. Expected is derived from the two source id constants so there is
+	// one SSOT; the literal 6 below is the independent floor that catches a
+	// curation list mutated to empty (which would move expected and captured
+	// together).
+	const expectedIds = [...SUPPORTED_ANTHROPIC_MODEL_IDS, ...SUPPORTED_CORTEX_MODEL_IDS].slice().sort();
+	assert.equal(
+		expectedIds.length,
+		6,
+		`curated id constants must total 6 rows (claude 2 + cortex 4) — got ${expectedIds.length}: ${expectedIds.join(", ")}`,
+	);
+	assert.deepEqual(
+		capIds.slice().sort(),
+		expectedIds,
+		`[QK:CORTEX-PROVIDER-SIX-ROW-SURFACE] compiled entry must register the EXACT curated set of both adapters (claude 2 + cortex 4) — got: ${capIds.join(", ") || "none"}`,
+	);
 	assert.equal(typeof cap.cfg.streamSimple, "function", "entry streamSimple must be a function");
 	// Regression guard: the entry must wire the REAL backend (streamShellAcp),
 	// not the removed S0 stub. We check by name rather than invoking — invoking
@@ -131,6 +163,17 @@ try {
 	);
 } finally {
 	rmSync(TMP_EMIT, { recursive: true, force: true });
+	try {
+		// The emit created `.tmp-verify/` as TMP_EMIT's parent. This gate now carries
+		// a mutant lane, so it runs inside check-gate-qualification's PURITY-checked
+		// snapshot, whose tree manifest walks ignored paths too — a leftover empty
+		// parent dir reads as IMPURE drift. Remove it when empty; a concurrent
+		// sibling gate's emit keeps it alive and the rmdir just fails.
+		// (Same repair as check-acp-cortex / check-acp-session-reuse.)
+		rmdirSync(".tmp-verify");
+	} catch {
+		// non-empty or already gone — fine either way
+	}
 }
 
 // ---------------------------------------------------------------------------
@@ -154,6 +197,7 @@ assert.ok(!/backend-stub/.test(entrySrc), "entry must not import the removed S0
 
 console.log(
 	`[check-acp-provider-surface] ok — compiled entry registers ${PROVIDER_ID} once (idempotent) with the no-auth ` +
-		`sentinel + curated Claude surface (sonnet + ${CURATED_ANCHOR_MODEL_ID}) + real streamShellAcp backend (S2c); ` +
-		`lib surface verified live (${models.length} model(s))`,
+		`sentinel + the EXACT curated set of both adapters (claude ${SUPPORTED_ANTHROPIC_MODEL_IDS.length} incl. ` +
+		`${CURATED_ANCHOR_MODEL_ID} + cortex ${SUPPORTED_CORTEX_MODEL_IDS.length}) + real streamShellAcp backend (S2c); ` +
+		`claude lib surface verified live (${models.length} model(s))`,
 );

package/scripts/check-acp-sdk-surface.ts

@@ -4,12 +4,23 @@
 // the peer-resolution that makes the Claude ACP adapter satisfiable:
 //
 //   @agentclientprotocol/sdk              1.3.0    wire SDK (acp-bridge import source)
-//   @agentclientprotocol/claude-agent-acp 0.61.0   Claude adapter (spawn binary)
+//   @agentclientprotocol/claude-agent-acp 0.62.0   Claude adapter (spawn binary)
 //   @anthropic-ai/sdk                     0.100.1  peer-resolution pin ONLY (see below)
 //
+// 2026-07-27 bump 0.61.0 → 0.62.0: the adapter's own `dist/` is byte-identical
+// across the two releases; 0.62.0 moves only its declared deps
+// (claude-agent-sdk 0.3.217 → 0.3.219; its devDep anthropic sdk, which does not
+// reach us). So this bump is a dependency refresh that carries NO adapter-code
+// change — and specifically NOT a fix for the bundled-MCP readiness race
+// (ROADMAP 🔴): 0.62.0 adds no explicit readiness fence. Effective runtime
+// behavior is NOT claimed identical: the transitive SDK moved, and MCP startup
+// lives inside it, so timing may differ — what is established is only that no
+// new fence was added.
+//
 // The anthropic SDK is NOT an API client / auth surface here. It is a direct
-// dep solely to satisfy @anthropic-ai/claude-agent-sdk@0.3.217's peer floor
-// (>=0.93.0); drop it and the tree resolves a stale 0.91.1 so the peer goes
+// dep solely to satisfy @anthropic-ai/claude-agent-sdk@0.3.219's peer floor
+// (>=0.93.0 — re-measured at 0.3.219, unchanged from 0.3.217, so the 0.100.1
+// pin stands rather than rising mechanically); drop it and the tree resolves a stale 0.91.1 so the peer goes
 // unmet — a failure that would only surface at the first raw turn. The
 // lockfile proves the same shape. Source-level import / API-client
 // instantiation / credential use stays forbidden — asserted in layer (4).
@@ -43,7 +54,7 @@ const pkg = JSON.parse(read("package.json")) as { dependencies?: Record<string,
 const deps = pkg.dependencies ?? {};
 const PINS: Record<string, string> = {
 	"@agentclientprotocol/sdk": "1.3.0",
-	"@agentclientprotocol/claude-agent-acp": "0.61.0",
+	"@agentclientprotocol/claude-agent-acp": "0.62.0",
 	[ANTHROPIC_SDK]: "0.100.1",
 };
 for (const [name, ver] of Object.entries(PINS)) {
@@ -63,13 +74,13 @@ for (const [name, ver] of Object.entries(PINS)) {
 const lock = read("pnpm-lock.yaml");
 assert.match(
 	lock,
-	/@agentclientprotocol\/claude-agent-acp@0\.61\.0\(@anthropic-ai\/sdk@0\.100\.1/,
-	"pnpm-lock: claude-agent-acp@0.61.0 must peer-resolve @anthropic-ai/sdk@0.100.1 (peer-pin), not the stale 0.91.1",
+	/@agentclientprotocol\/claude-agent-acp@0\.62\.0\(@anthropic-ai\/sdk@0\.100\.1/,
+	"pnpm-lock: claude-agent-acp@0.62.0 must peer-resolve @anthropic-ai/sdk@0.100.1 (peer-pin), not the stale 0.91.1",
 );
 assert.match(
 	lock,
-	/@anthropic-ai\/claude-agent-sdk@0\.3\.217\(@anthropic-ai\/sdk@0\.100\.1/,
-	"pnpm-lock: claude-agent-sdk@0.3.217 must peer-resolve @anthropic-ai/sdk@0.100.1 — else its >=0.93.0 peer floor is unmet",
+	/@anthropic-ai\/claude-agent-sdk@0\.3\.219\(@anthropic-ai\/sdk@0\.100\.1/,
+	"pnpm-lock: claude-agent-sdk@0.3.219 must peer-resolve @anthropic-ai/sdk@0.100.1 — else its >=0.93.0 peer floor is unmet",
 );
 
 // ---------------------------------------------------------------------------
@@ -108,8 +119,8 @@ const casEntry = adapterRequire.resolve("@anthropic-ai/claude-agent-sdk");
 const casInfo = pkgInfoFromEntry(casEntry);
 assert.equal(
 	casInfo.version,
-	"0.3.217",
-	`@anthropic-ai/claude-agent-sdk must runtime-resolve to 0.3.217 from the adapter context (got ${casInfo.version})`,
+	"0.3.219",
+	`@anthropic-ai/claude-agent-sdk must runtime-resolve to 0.3.219 from the adapter context (got ${casInfo.version})`,
 );
 const casRequire = createRequire(resolve(casInfo.dir, "package.json"));
 const sdkEntry = casRequire.resolve(ANTHROPIC_SDK);
@@ -152,7 +163,7 @@ const mcpEntry = casRequire.resolve("@modelcontextprotocol/sdk/server/index.js")
 const mcpInfo = pkgInfoFromEntry(mcpEntry);
 assert.ok(
 	mcpInfo.version.startsWith("1.29."),
-	`@modelcontextprotocol/sdk must runtime-resolve to 1.29.x from the claude-agent-sdk context (got ${mcpInfo.version}) — claude-agent-sdk 0.3.217 declares a ^1.29.0 peer`,
+	`@modelcontextprotocol/sdk must runtime-resolve to 1.29.x from the claude-agent-sdk context (got ${mcpInfo.version}) — claude-agent-sdk 0.3.219 declares a ^1.29.0 peer`,
 );
 
 // ---------------------------------------------------------------------------

package/scripts/check-acp-session-reuse.ts

@@ -26,7 +26,7 @@
 
 import { strict as assert } from "node:assert";
 import { execFileSync } from "node:child_process";
-import { copyFileSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { copyFileSync, mkdirSync, mkdtempSync, readFileSync, rmdirSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { resolve } from "node:path";
 import { pathToFileURL } from "node:url";
@@ -735,8 +735,71 @@ try {
 			"showToolNotifications is display-only — never perturbs the bridge config signature (no spurious rebuild)",
 		);
 	}
+
+	// ----------------------------------------------------------------------
+	// Section I — overlay scope wiring (P0-1, GPT review 2026-07-29): the turn
+	// loop hands adapter.ensureOverlay the AUTHORITATIVE resolveSessionKey value
+	// (opts.sessionId → PI_SESSION_ID → cwd). A session-scoped overlay (cortex)
+	// scopes its dir on this key, so two same-process/cwd sessions with
+	// different opts.sessionId must reach the adapter as two DIFFERENT keys —
+	// an ambient re-derivation (PI_SESSION_ID/cwd only) would alias them onto
+	// one overlay/mcp.json. Captured by wrapping the routed claudeAdapter's
+	// ensureOverlay on the compiled module (claude ignores the params, so a
+	// no-op stub is behavior-safe for these two turns).
+	// ----------------------------------------------------------------------
+	{
+		const captured: string[] = [];
+		const originalEnsureOverlay = adapterMod.claudeAdapter.ensureOverlay;
+		adapterMod.claudeAdapter.ensureOverlay = (p: { sessionKey: string }) => {
+			captured.push(p.sessionKey);
+			return { envOverrides: {} };
+		};
+		const savedPiSessionId = process.env.PI_SESSION_ID;
+		delete process.env.PI_SESSION_ID;
+		try {
+			const h = makeHarness(recordDir);
+			await collect(
+				backend.streamAcpTurn(
+					sonnet,
+					{ messages: [{ role: "user", content: "hi NONCE-I1", timestamp: 0 }] },
+					{ sessionId: "gate-I-one" },
+					h.deps,
+				) as Stream,
+			);
+			await collect(
+				backend.streamAcpTurn(
+					sonnet,
+					{ messages: [{ role: "user", content: "hi NONCE-I2", timestamp: 0 }] },
+					{ sessionId: "gate-I-two" },
+					h.deps,
+				) as Stream,
+			);
+			assert.deepEqual(
+				captured,
+				["pi:gate-I-one", "pi:gate-I-two"],
+				"the turn loop must pass adapter.ensureOverlay the AUTHORITATIVE resolveSessionKey value (opts.sessionId " +
+					"first) — two sessions in one process/cwd must arrive as two DIFFERENT overlay scope keys " +
+					"[QK:ACP-OVERLAY-SESSIONKEY-WIRED]",
+			);
+		} finally {
+			adapterMod.claudeAdapter.ensureOverlay = originalEnsureOverlay;
+			if (savedPiSessionId === undefined) delete process.env.PI_SESSION_ID;
+			else process.env.PI_SESSION_ID = savedPiSessionId;
+		}
+	}
 } finally {
 	rmSync(TMP_EMIT, { recursive: true, force: true });
+	try {
+		// The emit created `.tmp-verify/` as TMP_EMIT's parent. This gate now runs
+		// inside check-gate-qualification's PURITY-checked snapshot (it is the gate
+		// for the ACP-OVERLAY-SESSIONKEY-WIRED mutant), and the snapshot tree
+		// manifest walks ignored paths too — a leftover empty parent dir reads as
+		// IMPURE drift. Remove it when empty; a concurrent sibling gate's emit
+		// keeps it alive and the rmdir just fails.
+		rmdirSync(".tmp-verify");
+	} catch {
+		// non-empty or already gone — fine either way
+	}
 	rmSync(recordDir, { recursive: true, force: true });
 }