fix: respect framework-level memorySearch.enabled switch in autoRecall

[eisen0419]

Summary

Fixes #58 — autoRecall now respects the framework-level memorySearch.enabled switch.

Previously, when agents.defaults.memorySearch.enabled = false in OpenClaw config, the plugin still executed auto-recall because the before_agent_start handler did not check the framework-level setting.

Changes

• Added resolveFrameworkMemorySearchEnabled(cfg, agentId) helper that reads memorySearch.enabled from api.config, checking agent-specific override first (agents.list[].memorySearch.enabled), then falling back to agents.defaults.memorySearch.enabled
• Added early return in before_agent_start handler when framework memorySearch.enabled = false, with a debug log message
• Moved agentId resolution earlier in the handler to support the check (no behavioral change for the existing code path)

Precedence

agents.defaults	agent-specific	Result
false	not set	skip recall
false	true	allow recall
true	false	skip recall
not set	not set	allow recall (unchanged)

Test plan

• New standalone test file test/issue-58-framework-memory-search.test.mjs with 3 cases:
  • defaults.memorySearch.enabled=false → skips recall, 0 embedding requests
  • defaults=false + agent override true → allows recall
  • defaults=true + agent override false → skips recall
• Existing test/config-session-strategy-migration.test.mjs still passes

[AliceLJY]

Review: fix: respect framework-level memorySearch.enabled switch in autoRecall

Verdict: Fix-then-merge — same blocking item as #148.

---

✅ What's working

• Priority chain is correct: agent-level override → agents.defaults → undefined (allow recall). Existing users with no memorySearch config see zero behavior change
• Early return placement is right — framework permission check happens before shouldSkipRetrieval(), not after
• Test infrastructure is solid: real HTTP embedding server mock makes the assertions trustworthy (no false passes from over-mocking)
• All 3 test cases pass, including the agent-override-wins scenarios
• agentId move to top of handler is a clean refactor; resolveHookAgentId is pure, no side effects

🔴 Blocking (same issue as #148)

New test file is not wired into npm test. issue-58-framework-memory-search.test.mjs is missing from package.json scripts — CI never runs it.

Fix: append to the test script in package.json:

&& node --test test/issue-58-framework-memory-search.test.mjs

🟡 Suggested before merge

Missing backward-compat test — the most common real-world scenario (user upgrades without adding memorySearch to config) has no test coverage. Adding one case would close the gap:

test("issue #58: no memorySearch config → recall proceeds normally", async () => {
  const outcome = await runAutoRecallScenario({}); // empty OpenClaw config
  assert.ok(outcome.requests > 0);
});

⚪ Non-blocking

• NODE_PATH hardcodes /opt/homebrew — harmless in Docker (jiti resolves without it), but worth making portable
• catch { // ignore } in resolveFrameworkMemorySearchEnabled is intentionally fail-open; a short comment would help future readers

---

Merge order note

This should merge before #123 (fresh-session continuity fix) since both touch the autoRecall skip logic. #148 can go in either order — different files.

Good fix. The two-level config lookup matches how OpenClaw models other agent settings (e.g. resolveAgentPrimaryModelRef follows the same pattern).