Local build & real-E2E verification report (merge reference)

Verified at head 0a6bece5 (merge-base 35c91864) in an isolated detached worktree with a real npm ci and the actual built CLI on macOS. Before every drive, the gate implementation inside the built artifact was grep-verified (return !this.getBareMode() on the base build vs return this.enableManagedAutoMemory && !this.getBareMode() on the PR build), so each side of the A/B ran the code it claims to.

Verdict: LGTM — the fix works exactly as described end-to-end, no regression found. Two behavioral notes below for maintainer awareness (neither is a blocker).

1. Wire-level A/B — system prompt actually sent to the model

Drove the built CLI (node packages/cli/dist/index.js -p "hello", isolated $HOME, pre-seeded settings) against a local OpenAI-compatible capture server and parsed the system message of every captured POST /chat/completions body:

Project QWEN.md content survived in all scenarios (marker string retained), matching the PR's focused test at the wire level. The ~5.2 KB delta is with an empty memory store; populated stores (cf. s5's 6,379-char block with a team index) are consistent with the 7–9 KB reported in #6936.

2. TUI surface A/B — real built TUI via pty

With the setting off, the merge-base build still offered /dream and /forget in the slash-command menu (their operations would have failed — the surface/ops mismatch). On the PR build both are hidden, and a control run with default settings keeps them:

3. Tests & typecheck (isolated worktree, real install)

• core src/config/config.test.ts: 396/396 (includes the new regression test + the flipped expectation) — matches the PR author's run.
• cli gated surfaces (rememberCommand, BuiltinCommandLoader, MemoryDialog): 34/34.
• core src/memory/ full suite: 419 passed, 5 failed — all 5 are pre-existing local-machine artifacts, proven by re-running the same files at merge-base in the same worktree (fails identically: 4 failed | 3 passed): refresh.test.ts ×4 (macOS-only /var → /private/var tmpdir-symlink artifact; those tests mock isManagedMemoryAvailable, so they never touch the changed gate) and memoryLifecycle.integration.test.ts ×1 (real ~/.qwen/memories floods the recall prompt budget; the test lacks HOME isolation). Linux CI is unaffected by either class.
• tsc --noEmit (core): clean.

4. Call-site audit & behavioral notes

All 8 production call sites of isManagedMemoryAvailable() were reviewed: system-prompt assembly (core/config.ts:2839), recall prefetch (core/client.ts:1992 — already double-gated with getManagedAutoMemoryEnabled(), no change), managed remember (memory/remember.ts:154), refresh-after-write (memory/refresh.ts:156), /remember prompt choice (rememberCommand.ts:46), MemoryDialog.tsx:177, BuiltinCommandLoader.ts:150 (/dream, /forget), and 4 ACP daemon methods (acpAgent.ts:6001/6022/6113/6190). The ACP surfaces flip consistently: workspaceMemoryRememberAvailability reports available: false and remember/forget/dream throw managed_memory_unavailable, which consumers already handle via the availability endpoint.

Two things worth knowing when merging:

1. Team memory rides on the same gate (s4 vs s5 above). The entire team-memory block — prompt section, index rebuild, git sync — is nested inside if (isManagedMemoryAvailable()) in refreshHierarchicalMemory, so enableManagedAutoMemory: false now also suppresses team memory even with memory.enableTeamMemory: true (verified with the env override forced on). Defensible — team memory is a tier of managed memory and its recall was already ops-gated — but the two settings are nominally independent, so flagging it.
2. Safe mode now effectively strips the block too. The CLI wiring (packages/cli/src/config/config.ts:2188) already passes enableManagedAutoMemory: false under safe mode, so after this PR safe mode also loses the prompt block, /dream, /forget and the ACP surfaces (previously availability stayed true in safe mode while ops were off). This closes the same mismatch class for safe mode — the PR's "avoids broadening safe-mode semantics" claim is accurate at the core gate, and the CLI wiring is what carries it over.

Also for the record: the CI precheck flag (prompt_injection:system_prompt) appears to fire because the diff touches system-prompt assembly; the 2-file diff contains no injected instructions.

Environment

macOS (Darwin 24.6), isolated detached worktree at 0a6bece5, real npm ci (prepare build), capture server + pty drivers, per-scenario isolated $HOME.

@qwen-code /triage

Maintainer E2E Verification (macOS)

Ran a real CLI E2E test on macOS to verify this PR before merging.

Test Method

Built the PR branch (npm run build && npm run bundle), then ran the bundled CLI against a mock OpenAI server with --openai-logging to capture the actual system prompt sent to the API. Compared two runs:

1. Default (memory.enableManagedAutoMemory unset → true)
2. Disabled (memory.enableManagedAutoMemory: false in user settings.json)

Results

Context savings: 4,711 chars (~1,178 tokens) per request when disabled.

Unit Tests

✓ src/config/config.test.ts (396 tests) — all passed

The 4 PR-specific tests (including the new refreshHierarchicalMemory should omit managed auto-memory prompt when disabled and the updated returns false when enableManagedAutoMemory is false) all pass.

Screenshot

Verdict

✅ PASS — The one-line change in isManagedMemoryAvailable() correctly gates the # auto memory system prompt block. When enableManagedAutoMemory is false, the block is fully removed from the system prompt, saving ~4.7 KB of context per request. Project memory (QWEN.md/AGENTS.md) is still loaded normally.

✓ src/config/config.test.ts (396 tests) — 全部通过