Skip to content

FAQ

What is Sentinel?

Sentinel is the plugin shipped by vision-delivery. It turns an operational image/video request into a CV task, success gate, cheapest baseline, inspectable proof plan, and economics decision. It is a workflow layer, not a model.

Is it for people who do not know computer vision?

It is designed for them: a user can describe the outcome and consequences in plain language while Sentinel proposes the methodology. The user must still have authority to use the data, representative examples, and a human owner. No novice outcome study has yet proven an “almost no knowledge” claim.

What evidence exists?

Little: one historical routing run, a mocked process comparison, and one small private detection result. Support & Evidence states every claim with its numbers and limits.

Does it guarantee accuracy or production readiness?

No. Accuracy depends on the domain, data, model, threshold, and operating environment. A careful plain agent or human with the same tools can reach comparable metrics. Sentinel's value is making evidence-first process easier to follow.

Does it replace Roboflow MCP or official Roboflow skills?

No. Sentinel uses MCP for read-only discovery and, on Claude Code, for paid or state-changing actions that pass its pre-action gate. It owns task framing, evaluation, proof artifacts, and economics. Current MCP semantics, model IDs, Workflows, plans, pricing, and platform navigation belong to Roboflow's official skills and documentation. Read Roboflow Skills Integration.

Which hosts are supported?

Claude Code and Codex, installed from the public GitHub repository. They are not equal. Claude Code gets the methodology and the spend gate. Codex is the methodology tier: the locked target, the baseline ladder, and the scorer are the same there (no recorded live run on either host yet), but no pre-action hook exists, so there is no gate. Other hosts are outside the supported surface. See the compatibility policy.

How do I authenticate to Roboflow?

Plugin installation does not require a credential environment variable. Codex metadata requests authorization on use; verify the actual host-managed sign-in, account, and scope in the live host. Credentials for a generated standalone hosted client, if current upstream guidance requires them, are a separate deployment concern and must not be pasted into chat or committed.

Does it spend credits automatically?

Not on Claude Code without a recorded brief and your approval. On Claude Code, a Sentinel PreToolUse hook holds each paid, data-moving, or state-changing Roboflow MCP call for a recorded action brief and your approval at the host permission prompt; single-image inference is approved once, as a call budget. After that first approval, later calls inside the budget follow your own host permission rules. On Codex there is no gate: Sentinel is instructed to write the brief and stop, and nothing in the plugin enforces it. The gate does not cap spend and does not see calls made outside MCP. Trust and Safety describes the gate and what it does not cover.

What access does it request?

Routes can request filesystem reads/writes/edits, search, shell commands, and Roboflow MCP operations. Codex and Claude Code hooks can append a local activity ledger. Review host permission prompts and, in Codex, review and trust changed plugin hooks with /hooks before they run.

Is the ledger proof that an action succeeded?

No. New hook records can classify host-reported outcomes and deduplicate event IDs, but coverage is best-effort. The ledger is not complete telemetry, an authorization receipt, or an independent Roboflow confirmation.

Are generated scripts ready to run?

Not by default. Inspect dependencies and data flow, run syntax checks, execute a representative fixture, and compare outputs with the agreed gate. An instruction to write a script or a chat statement that it exists is not verification.

Can it process people, plates, forms, or medical images?

Only after authority, purpose, minimization, retention, representative evaluation, human review, and relevant legal/security/domain checks are established. Do not use Sentinel as the sole decision-maker for consequential outcomes.

What should the first prompt contain?

State the input, desired output, representative sample availability, and consequence of a wrong result:

Count pallets crossing this line from these 60 frames. Report an hourly total.
A miss is worse than a duplicate. I am authorized to use this footage.

Where do I get help?

Use the repository support policy. Keep public reports free of credentials and private or sensitive data.