fix(mcp): support draft-2020-12 tool schemas (wire in Firecrawl)

[AvivK5498]

Problem

Firecrawl's v2 MCP tools declare their input schema as JSON Schema draft 2020-12. Mastra's tool-input validator can't resolve that meta-schema, so it rejects every call locally before it leaves the machine:

Tool input validation failed for firecrawl_firecrawl_scrape.
- root: Schema validation error: no schema with key or ref "https://json-schema.org/draft/2020-12/schema"

The tools connect and appear in the agent's toolset, but every call fails. (brave-search is unaffected — its schema is draft-07.)

Fix

• Schema override — replace the input schema of the two whitelisted firecrawl tools (scrape, search) with clean Zod schemas. The override mutates the Tool instance in place rather than spreading a copy: Tool.execute validates against this.inputSchema at call time and is an arrow function lexically bound to the instance, so a spread copy never takes effect (found this the hard way — validation passed but execute still failed until the mutation).
• Env-var MCP urls — expandEnvVars now applies to HTTP MCP server urls, so the Firecrawl key lives in .env as ${FIRECRAWL_MCP_URL} instead of being hardcoded in the git-tracked mcp-servers.yaml.
• Register firecrawl in mcp-servers.yaml, whitelisted to scrape + search.

Setup

Consumers add to .env:

FIRECRAWL_MCP_URL=https://mcp.firecrawl.dev/<your-key>/v2/mcp

and opt an agent in via its mcpServers: [firecrawl] setting.

Verification

Probed the real code path (initMCPClient → getMCPTools → tool execute): firecrawl_scrape on https://example.com returns real page markdown end-to-end, no validation error.

🤖 Generated with Claude Code

[ak-prbot]

The workspace only has the skills/ dir mounted; the repo files aren't accessible, so I'm reviewing from the diff in context.

Verdict

No blockers, but a few real risks around secret-in-URL logging, opaque failure on a missing env var, and an in-place mutation that's untested and brittle.

Warnings

• src/agent/mcp-client.ts:87 — new URL(expandEnvVars(cfg.url)) will throw (or silently produce a broken URL) if FIRECRAWL_MCP_URL is unset, and the error surfaces at MCP init rather than pointing at the missing var — guard with an explicit check that throws a clear "FIRECRAWL_MCP_URL not set" message.
• src/agent/mcp-client.ts:87 — the Firecrawl API key is embedded in the URL; if the MCPClient config or logger ever prints url, the key leaks into logs — confirm nothing downstream logs the resolved URL, or strip/redact it before passing to the client.
• src/agent/mcp-client.ts:158 — the in-place mutation (tool as { inputSchema?: unknown }).inputSchema = ... is a load-bearing workaround with no test and a hardcoded comment about Mastra internals; if Mastra re-reads the schema from a different source or freezes the instance, this silently regresses to the original failure — add a regression test that asserts tool.inputSchema is the Zod schema after init for the two whitelisted ids.
• src/agent/mcp-client.ts:157 — override is keyed by firecrawl_firecrawl_scrape / firecrawl_firecrawl_search, but the actual tool name produced by getMCPTools depends on the server's reported tool names and the serverName_toolName join convention; if Firecrawl renames or the join differs, the override silently no-ops and calls fail again — log a warning when SCHEMA_OVERRIDES[name] exists but assignment is skipped, or assert the expected names at init.

Nits

• mcp-servers.yaml:6 — url:/tools: are indented 7 spaces while the sibling brave block uses 8 for its children; align for consistency (valid YAML either way).