fix: add missing items field to two array schemas (Gemini Pro strict JSON Schema compat)

[DmitryBMsk]

Problem

Two array properties in src/core/ are declared with type: 'array' but no items field, which strict JSON Schema validators reject. Most notably, Gemini Pro's tool-call schema validator throws:

LLM request rejected: Invalid schema for function 'gbrain__extract_facts':
In context=('properties', 'entity_hints'), array schema missing items.

Repro

1. Run gbrain serve as a stdio MCP child of any host that uses Gemini Pro as its chat model (OpenClaw 2026.5.4 with mcp.servers.<name>.command config is one example; ChatGPT Tool calls and other strict-schema LLM gateways behave the same).
2. The daemon enumerates tools and forwards schemas to the LLM.
3. The whole gbrain MCP surface (60 tools) becomes unavailable for that session because the LLM API rejects the request before any tool can be called.

End-to-end repro on my live OpenClaw deployment: bundle-mcp logs show failed to start server "gbrain": McpError: MCP error -32000: Connection closed, and the Telegram bot returns the Invalid schema for function 'gbrain__extract_facts' message verbatim. After the patch is applied locally and the stdio child is restarted, the bot enumerates all 60 gbrain__* tools and successfully calls e.g. gbrain__get_stats returning real numbers from Supabase.

Fix

Add the missing items field to both schemas. Behavior unchanged — these are JSON Schema metadata, not runtime contracts.

src/core/operations.ts — extract_facts.params.entity_hints

The description already says "canonical entity slugs" and the handler casts the value via p.entity_hints as string[]. Schema now matches:

-    entity_hints: { type: 'array', description: 'Existing canonical entity slugs ...' },
+    entity_hints: { type: 'array', items: { type: 'string' }, description: 'Existing canonical entity slugs ...' },

src/core/resolvers/builtin/x-api/handle-to-tweet.ts — outputSchema.candidates

The TypeScript interface XTweetCandidate already defines the per-element shape; the JSON Schema now spells the same fields out for any consumer that runs a JSON Schema validator on the resolver output:

-      candidates: { type: 'array' },
+      candidates: {
+        type: 'array',
+        items: {
+          type: 'object',
+          properties: {
+            tweet_id: { type: 'string' },
+            text: { type: 'string' },
+            created_at: { type: 'string', format: 'date-time' },
+            score: { type: 'number' },
+            url: { type: 'string', format: 'uri' },
+          },
+          required: ['tweet_id', 'text', 'created_at', 'score', 'url'],
+        },
+      },

Why not a typedef-driven generator?

Could be a future improvement (run the same TS interface through a to-json-schema step) — out of scope for this fix. Both schemas now match the TypeScript types that already exist alongside them.

Test plan

• Reproduced the gbrain__extract_facts rejection on a live OpenClaw 2026.5.4 + Gemini Pro deployment (private OCI setup).
• Applied the patch, restarted the OpenClaw stdio child, confirmed bundle-mcp no longer errors and all 60 gbrain__* tools are surfaced into the LLM tool inventory.
• Confirmed gbrain__get_stats and gbrain__search round-trip successfully end-to-end (host → MCP → Supabase) after the patch.
• Repo-level test suite — I trust the existing CI to catch regressions; both diffs are pure JSON Schema metadata, no code path changes.

Notes

• I grepped for other type: 'array' declarations without items in src/; only these two showed up in master at the time of writing (commit 17b190e).
• Other MCP libraries may be more lenient — Claude Sonnet 4.6 accepts the unfixed schema, which is likely why this slipped through. Gemini Pro / Google's strict validator catches it.

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews