[AI-assisted] This PR was generated using Claude Code.

Closes #92582

What Problem This Solves

Fixes an issue where users who configured agents.defaults.memorySearch.provider to "local" (bundled Hugging Face GGUF embeddings) would see openclaw doctor — and openclaw doctor --deep — falsely report:

Memory search provider is set to "local", but local embeddings are not confirmed ready.
Fix (pick one):
- Install the llama.cpp provider plugin: openclaw plugins install @openclaw/llama-cpp-provider
- Set a local GGUF model path in config
...

…even when openclaw memory status --deep confirms the full memory stack is healthy (Embeddings: ready, Vector store: ready, etc.). The warning implies the configured local provider is broken when it is not — a false positive that sends users chasing a non-existent problem (reinstalling plugins, reconfiguring model paths).

Why This Change Was Made

openclaw doctor probes gateway memory readiness via probeGatewayMemoryStatus, which always calls doctor.memory.status with probe: false. The gateway therefore returns checked: false + skipped: true to signal "readiness was not tested" — for both doctor and doctor --deep.

In noteMemorySearchHealth, the key-optional provider branch already treats skipped: true as "not tested, do not warn" and returns early. The local provider branch, however, only suppressed its warning when checked && ready; it had no skipped guard, so a skipped probe fell straight through to the alarming "not confirmed ready" warning.

This PR mirrors the existing key-optional guard onto the local branch: when the probe was intentionally skipped, return without warning. A transport timeout still sets checked: false with skipped absent, so it continues to fall through to the warning (unchanged). This is a focused, symmetric 3-line guard plus a regression test — it does not change probe behavior, gateway calls, or any ready/not-ready path that was already firing correctly.

Non-goal: wiring --deep to actually run the embedding probe (probe: true) is a larger change to the gateway probe transport and is intentionally out of scope for this false-positive fix.

User Impact

Users running openclaw doctor or openclaw doctor --deep with a healthy local embedding provider no longer see a false "local embeddings are not confirmed ready" warning that implies their provider is broken. The doctor output stays quiet when readiness was simply not probed, consistent with how key-optional providers (ollama, lmstudio, openai-compatible) already behave. Users can still run openclaw memory status --deep for an authoritative readiness check, and genuine failures (probe ran and reported not-ready, or timed out) still produce the warning.

Evidence

Root cause confirmed in source: src/commands/doctor-memory-search.ts, the provider === "local" branch lacked the skipped guard that the key-optional branch already had.

Regression test added: src/commands/doctor-memory-search.test.ts — "does not warn when local provider probe was skipped (skipped: true)". It drives the real noteMemorySearchHealth with a local provider whose gateway probe was skipped ({ checked: false, ready: false, skipped: true }), exactly the probe: false path openclaw doctor uses.

BEFORE fix (unpatched main) — the false-positive warning is emitted:

FAIL  src/commands/doctor-memory-search.test.ts > noteMemorySearchHealth > does not warn when local provider probe was skipped (skipped: true)
AssertionError: expected "vi.fn()" to not be called at all, but actually been called 1 times

1st vi.fn() call:
  "Memory search provider is set to \"local\" and a local model path is configured, but local embeddings are not confirmed ready.
  Fix (pick one):
  - Install the llama.cpp provider plugin: openclaw plugins install @openclaw/llama-cpp-provider
  - Set a local GGUF model path in config
  - Switch to a remote provider: openclaw config set agents.defaults.memorySearch.provider github-copilot
  Verify: openclaw memory status --deep"
Number of calls: 1

AFTER fix — the same input emits no warning, and the full suite (including the existing "local provider readiness probe is inconclusive" timeout case, which still warns) passes:

Test Files  1 passed (1)
     Tests  50 passed (50)

The existing timeout case (checked: false, no skipped, error gateway memory probe timed out: ...) still produces the warning, confirming the fix only suppresses the intentional-skip case and does not silence real diagnostic signals.

Real behavior proof

• Behavior or issue addressed: openclaw doctor false-positive "local embeddings are not confirmed ready" warning for a configured local memory provider when the gateway memory probe was skipped (readiness not checked).
• Environment used for testing: Windows 10, Node v22.22.3, OpenClaw 2026.6.8 built from this PR branch (base upstream/main 49b0487 + fix), invoked via the real openclaw doctor CLI.
• Steps to reproduce: configure agents.defaults.memorySearch.provider=local with a local modelPath (hf: GGUF URI) and a gateway auth token stored as an exec SecretRef, then run openclaw doctor without --allow-exec. Doctor skips the gateway memory probe (readiness not checked — the exact skipped-probe condition from issue Bug: doctor falsely warns local memory embeddings are not ready #92582), so noteMemorySearchHealth receives { checked: false, ready: false, skipped: true }.
• Observed result before fix: doctor emitted a "Memory search" note — "Memory search provider is set to "local" and a local model path is configured, but local embeddings are not confirmed ready" plus a remediation list.
• Observed result after fix: doctor emitted 0 "Memory search" notes for the skipped-probe case; the "Gateway health probes skipped because gateway credentials use an exec SecretRef" note still fires, confirming the same skipped-probe path was exercised.
• Evidence after fix (terminal capture):

$ openclaw doctor --non-interactive --no-workspace-suggestions          # AFTER fix (this PR)
...
◇  Gateway ────────────────────────────────────────────────────────────────╮
│  Gateway health probes skipped because gateway credentials use an exec   │
│  SecretRef. Run `openclaw doctor --allow-exec` to verify Gateway health  │
│  with exec SecretRefs.                                                   │
├──────────────────────────────────────────────────────────────────────────╯
... (no "Memory search" section is emitted — readiness was not probed)

$ grep -ciE 'Memory search provider|local embeddings|not confirmed ready' after-doctor.txt
0

Before fix (unpatched), the same openclaw doctor run emits the false-positive warning while the gateway probe is skipped:

$ openclaw doctor --non-interactive --no-workspace-suggestions          # BEFORE fix (main)
...
◇  Gateway ────────────────────────────────────────────────────────────────╮
│  Gateway health probes skipped because gateway credentials use an exec   │
│  SecretRef. Run `openclaw doctor --allow-exec` to verify Gateway health  │
│  with exec SecretRefs.                                                   │
├──────────────────────────────────────────────────────────────────────────╯
...
◇  Memory search ──────────────────────────────────────────────────────╮
│  Memory search provider is set to "local" and a local model path is  │
│  configured, but local embeddings are not confirmed ready.           │
│  Fix (pick one):                                                     │
│  - Install the llama.cpp provider plugin: openclaw plugins install   │
│    @openclaw/llama-cpp-provider                                      │
│  - Set a local GGUF model path in config                             │
│  - Switch to a remote provider: openclaw config set                  │
│    agents.defaults.memorySearch.provider github-copilot              │
│  Verify: openclaw memory status --deep                               │
├──────────────────────────────────────────────────────────────────────╯

Compatibility / Migration

No public API, config, or CLI behavior change for healthy or genuinely-broken setups. The only behavioral change: a skipped probe no longer emits a false-positive warning for the local provider — matching the existing key-optional provider behavior. No migration required.

Failure Recovery

Revert the single commit. The change is isolated to the local-provider branch guard in noteMemorySearchHealth.

Risks and Mitigations

• Risk: a user with provider: "local" and no model path configured who runs openclaw doctor (no --deep) no longer gets the "set a local GGUF model path" hint when the probe is skipped. Mitigation: this matches the existing key-optional provider behavior; openclaw doctor --deep / openclaw memory status --deep remain the authoritative readiness checks and the warning still fires when the probe actually runs and reports not-ready.
• No security, message-delivery, availability, or compatibility boundary is touched.

What was not tested

• macOS / Linux not tested (Windows-only local environment); the guarded code path is platform-agnostic.
• Swift host-env-policy check (pnpm check:host-env-policy:swift) skipped on Windows (no Swift toolchain).
• pnpm format:check reports 157 pre-existing format findings across the repo on a clean synced main (e.g. ui/src/...); the two files touched by this PR pass oxfmt --check cleanly.