recognize-pose-or-gesture git:20260811.bdc99ffA

Immutable. This exact content never changes and is served at /api/v1/blob/0b7c0584170ffcef.

---
name: recognize-pose-or-gesture
description: |
  Build keypoint-based pose, gesture, action, ergonomics, posture, or fall analysis. TRIGGER when: user asks for body/hand pose, gestures, action recognition, skeletons, keypoints, joint angles, raised hands, bending, squats, falls, sign language, or posture compliance. SKIP when: user needs an exact body contour/mask (segment-and-analyze), per-person PPE without keypoints (detect-and-analyze), whole-image compliance (classify-or-flag), identity/path tracking without pose (track-and-count), OCR (read-text), to compare cloud-hosted vs self-hosted or cost only (estimate-economics), or integration/delivery of a working pose model (deliver-cv-project).
allowed-tools: Bash, Read, Write, Edit, Glob, Grep, AskUserQuestion
---

<objective>

Turn keypoints or temporal keypoint sequences into a pose/gesture decision that clears independent acceptance. Keep pose estimation, identity association, and business-rule classification as separately measured stages.

</objective>

<methodology>

**Platform execution boundary.** Read `../../resources/roboflow-platform-lookup.md` before any provider-specific search, dataset, training, inference, workflow, or deployment action. Use installed official Roboflow skills or current MCP resources only for read-only discovery and sourced action-brief inputs; never invoke provider execution. Do not freeze model families, IDs, or platform recipes in Sentinel.

Follow `../../resources/fde-methodology.md`; apply these keypoint-specific rules.

## 1. Define the action and time window

Inspect footage, annotations, rules, code, and camera placement. Clarify whether the output is a per-frame posture, a hand/body gesture, a timed action, an ergonomics measurement, or a safety alert. Ask:

- Which joints/keypoints and time window determine the action?
- Must identity persist across frames?
- Which error is worse for the operator: a miss or false alert?

If the task only needs identity/path/dwell, route to `track-and-count`. Per-person PPE without a keypoint requirement routes to `detect-and-analyze`.

Freeze the evidence gate:

```text
Acceptance ID: <session/revision>
Business decision: <alert/action enabled by pose or gesture>
Gold set: <independent keypoints/events, subject/site split, adjudicator>
Primary metric and threshold: <keypoint quality/event recall/F1/angle error>
Secondary guardrails: <false alerts/hour, latency, subject/site floors>
Frozen before baseline: <timestamp and confirmation>
Baseline result (diagnostic only): <not run yet>
```

## 2. Freeze the keypoint contract

Record keypoint names/order, coordinate convention, confidence/missing-point semantics, left/right definition, person association, frame timestamps, and rule thresholds. For temporal events, define start/end, minimum duration, cooldown, and partial-visibility behavior before evaluation.

## 3. Select and measure candidates

Ask upstream for current keypoint/pose candidates with provenance, license, skeleton compatibility, and runtime requirements. Do not store exact model IDs or invocation syntax here. If a transparent geometric rule can map stable keypoints to the desired label, compare it with learned sequence classification on the same gate.

Report stage-specific evidence:

- keypoint localization/visibility quality on independently labeled frames;
- event recall, precision/F1, and false alerts per hour;
- joint-angle error when angles drive the decision;
- subject/site/camera and occlusion slices;
- p50/p95 end-to-end latency on the target runtime.

An event label inferred from candidate keypoints is not independent gold.

## 4. Diagnose a failed gate

Separate keypoint localization failures, left/right swaps, missing joints, identity switches, temporal-window errors, and rule/classifier errors. Check camera viewpoint, occlusion, motion blur, subject diversity, and annotation consistency. Test capture or deterministic rule improvements before more expensive training. Use read-only upstream facts to prepare any required provider action brief, stop before execution, and remeasure externally supplied results against the frozen split.

## 5. Decide and deliver

Return `go`, `revise`, or `stop`, with exact event counts, failure slices, rule thresholds, and human-review boundary. Follow `../../resources/artifact-contract.md`; produce a verified `hosted-client`, offline-tested `local-runtime`, or candid `scaffold`. Route a passing capability to `deliver-cv-project`.

</methodology>

<safety>

- Pose output alone does not prove intent, identity, medical condition, or safety compliance.
- Safety alerts require a documented human/operational fallback.
- Pseudo-labels cannot own acceptance.
- Emit a sourced action brief and stop before data movement or paid provider work; Sentinel never invokes either.
- Use current upstream guidance only for read-only facts and action-brief inputs; never invoke provider execution.

</safety>

<ledger>

Follow `../../resources/ledger-protocol.md`. Record acceptance, stage evaluations, artifact verification, and decision events once with non-empty IDs and status.

</ledger>

\<stop_rules>

- Required joints/action are not defined → define the contract before candidate work.
- No independent subject/site split → do not claim generalization.
- Capture does not expose required keypoints → state the camera/acquisition blocker.
- Provider action is required → emit a sourced action brief and stop; conversational consent does not authorize Sentinel execution.
- Applicable live/offline check is absent → retain `scaffold` status.

\</stop_rules>