Grok shared tools: local validation

Validated on 2026-10-22 using an isolated Temurin JDK 17 and Leiningen. No production configuration or live credentials were inspected or changed.

Automated evidence

• lein midje :all: 742 checks passed, zero failures or errors.
• CLOVERAGE_VERSION=1.1.2 lein with-profile +test cloverage --codecov --runner midje: 742 checks passed.
• Final runs include upstream GitHub directory-decoding fix 8138dc5.
• Local HTTP Responses and MCP servers exercised mixed native/client tools, encrypted output replay, call_id continuation, MCP validation and budgets, uncertain outcomes, safe final turns, redaction, and connection cleanup.
• The real interpreter routed both ai --provider grok --toolset ... and grok --toolset ... through shared authorization. Disallowed users and models were rejected before an MCP connection opened.
• Chat Completions compatibility, configuration/environment normalization, provider/model selection, and legacy absent-config fallback passed.
• --no-tools disables native search and MCP. Failed/incomplete Responses results produce safe errors rather than partial answers.
• Hook checks preserve direct grok agent ... and gi ... routing. Prompts containing agent after flags stay on the shared Grok path.
• Discord tests cover success/failure replies, late progress creation, creation and deletion failures, and deletion waiting for an in-flight edit.

Review follow-up

Eight review lenses produced three confirmed P2 findings. All were applied inline: anchor legacy-agent routing; delete late-created Discord status messages; exercise active progress edits during cleanup. Routing and late-creation regressions failed before their fixes. The edit test uses synchronization gates rather than arbitrary test sleeps. No confirmed review findings remain open.

Not established by local tests

• Live xAI support/account access for the chosen model and configured native tools, service-side billing, or encrypted replay behavior in production.
• Authenticated Cloudflare MCP compatibility or actual token permissions. Read-only access requires narrowly scoped Read permissions; allowing the Cloudflare execute tool alone does not make a token read-only.
• Live Discord transport, rate limits, thread permissions, or bot deployment.
• Container startup: the local Docker daemon was unavailable. JVM tests ran directly; the temporary Java installation did not change global setup.

Before deployment, configure the Responses provider and a restricted named MCP toolset using AI.md, then qualify native search, an authorized Cloudflare read, an unauthorized-toolset attempt, --no-tools, and Discord progress cleanup. Existing gi image generation still requires xai.key; it does not use the shared text-provider configuration.