vision-delivery

# Sentinel / vision-delivery — full agent context

`vision-delivery` ships the `sentinel` plugin for business-first CV delivery in Codex and Claude Code. Start from the operational outcome, not a model. The plugin translates plain language into a CV output type and evidence plan, measures a baseline before recommending training, stops with an action brief before paid/data-moving/state-changing provider actions, requests inspectable artifacts, and frames economics and the next human decision.

## Non-negotiable boundary

Sentinel is not a model, sandbox, hard authorization system, autonomous production owner, domain expert, legal reviewer, or safety certifier. It is designed to lower the methodology burden. The user still owns:

1. authority and allowed purpose for data and accounts,
2. the business outcome,
3. representative examples,
4. consequences of misses and false alarms,
5. final acceptance and operational governance.

No independent novice study establishes “no knowledge required.” Use “designed to lower the CV methodology burden.”

## Support tiers

- **Historical evidence:** a result was recorded under its original conditions; it is not current support or independent reproduction evidence.
- **Guided:** a workflow exists but equivalent live end-to-end evidence is missing.
- **Delegated upstream:** current product facts must come from official Roboflow sources.
- **Expert required:** Sentinel may structure evidence but cannot independently validate or approve the work.

Detection/counting has one historical private B1 result but remains guided. Classification, tracking, OCR, segmentation, pose/gesture, pipelines, delivery, and economics are guided. Regulated or safety-critical work, people surveillance, medical use, physical measurement, production streams/edge systems, governance, and production sign-off are expert-required.

## Roboflow relationship

Sentinel owns task framing, success gates, baseline-first decisions, proof requests, provenance interpretation, and economics. Roboflow owns current MCP tool semantics, model families/IDs, Workflows, platform navigation, plans, billing, and pricing.

Lookup order for platform truth:

1. installed official `roboflow/computer-vision-skills`,
2. exposed `roboflow://skills/...` MCP resources,
3. official current documentation or a request to install the official skills,
4. otherwise mark guidance unverified.

Do not copy platform recipes into Sentinel. Two full plugins may duplicate a `roboflow` MCP server definition on hosts that do not deduplicate configuration.

## Invocation

Codex:

```bash
codex plugin marketplace add https://github.com/Borda/vision-delivery
codex plugin add sentinel@sentinel
```

Claude Code:

```bash
claude plugin marketplace add Borda/vision-delivery
claude plugin install sentinel@sentinel
```

No credential environment variable is required for plugin installation. The v0.4 release candidate passed local clean-home marketplace simulations. Manual public-GitHub marketplace installation was recorded for v0.3 on Codex and Claude Code; the v0.4 public-install path remains unverified. When the host or account requires it, using the MCP capability requests hosted authorization; an existing authorized session may need no prompt. Live Roboflow account authorization and service behavior remain upstream checks. Use `claude plugin validate .` and `claude --plugin-dir .` only for development from a checkout. Cursor support is unverified. Credentials must never appear in chat or tracked files.

## Workflow

1. Confirm data/account authority, allowed purpose, and human ownership.
2. Read only authorized project context and representative samples.
3. Translate the output into boxes, labels, text, masks, tracks, keypoints, or a typed pipeline.
4. Define metric, threshold, eval slice, failure consequence, and mode before model selection.
5. Measure an existing/pretrained baseline.
6. Investigate misses and try the cheapest justified improvement.
7. Explain and confirm credit-spending/state-changing actions.
8. Request local proof artifacts and execute/inspect them.
9. Compare observations with the gate and preserve failures/unknowns.
10. Estimate economics and hand final acceptance to the human owners.

Paid-action consent is a prose instruction, not a hard block. Use host approvals, least-privilege account authorization or sessions, account budgets, and non-production targets. A separately generated standalone client may require a provider key; follow current provider guidance and scope it to the minimum permissions.

## Artifacts and ledger

Routes may request `eval_definition.md`, scripts, JSONL outputs, model/workflow identifiers, and `.vision-delivery/ledger.jsonl`. Requested or generated artifacts are not self-validating. Inspect dependencies and data flow, run syntax checks, execute a representative fixture, and compare observed outputs with the committed gate.

The shared Codex and Claude hook can record selected host-reported Roboflow outcomes and deduplicate stable event IDs. Codex plugin hooks require explicit trust review after installation or changes. On malformed hook input, the hook may write only a fixed diagnostic code and timestamp under host-managed plugin data; it never writes the payload, credentials, or raw error text. Coverage remains best-effort. A ledger row is not complete telemetry, authorization proof, or an independent receipt from Roboflow. Identifiers and notes can still be sensitive.

## Evidence register

- One pre-v0.2 Claude Sonnet trigger run: 143 labeled prompts, precision 0.94, recall 0.85, 4 false positives, 11 false negatives. Two direct specialist routes increased tolerant positive coverage to 65/74 (0.88). The run excludes delivery/setup routes and is not current-route, Codex, or task-completion evidence.
- Synthetic process A/B: 16 mocked runs at N=1 per cell; one supported cell, six mixed/parity, one loss. Developer-machine contamination and live capability confirmation remain unresolved.
- B1: historical private aerial vehicle fixture, 11 test images; recorded post-train mAP@50 84.6%. No controlled plain arm, public independent reproduction, post-train count MAE, verified conveyor transfer, or current endpoint verification.
- B2-B5: specifications only. Do not turn proposed models or thresholds into measured claims.
- No novice usability study, broad modality study, production reliability study, or security certification exists.

## Sensitive uses

Before faces, plates, people tracking, forms, medical imagery, worker monitoring, minors, or location-linked media, require authority/purpose, minimization/retention, access control, representative and subgroup evaluation, error consequences, named human review/appeal, and applicable legal/privacy/security/domain review.

Do not use Sentinel alone for medical, employment, law-enforcement, access-control, or physical-safety decisions. Physical units require calibrated capture and domain validation.

## Source map

- Documentation: https://borda.github.io/vision-delivery/
- Scope/evidence: https://borda.github.io/vision-delivery/support-and-scope/
- Workflow: https://borda.github.io/vision-delivery/workflow/
- Roboflow boundary: https://borda.github.io/vision-delivery/roboflow-skills/
- Trust: https://borda.github.io/vision-delivery/trust/
- Benchmarks: https://borda.github.io/vision-delivery/benchmarks/
- Support: https://github.com/Borda/vision-delivery/blob/main/.github/SUPPORT.md
- Compatibility: https://borda.github.io/vision-delivery/compatibility/
- Release policy: https://borda.github.io/vision-delivery/release-policy/
- Security: https://github.com/Borda/vision-delivery/blob/main/.github/SECURITY.md
- Official Roboflow skills: https://github.com/roboflow/computer-vision-skills