gemini-cli-photo-classifier

name	gemini-cli-photo-classifier
description	End-to-end run-image classification with Gemini, including image fetch, prompt control, and JSON output. Use when you need parking or pull-over labels from a run_id.
args	<run_id> [timestamp_unixus]

Gemini Run Photo Classifier

This skill gives a complete, from-scratch flow:

1. fetch camera images for a run/timestamp,
2. classify with Gemini,
3. return strict JSON output.

Inputs

• run_id (required)
• timestamp_unixus (optional)
• prompt (optional override)
• model (optional, default gemini-2.0-flash)

If timestamp_unixus is omitted, use the run start time parsed from run_id.

Prerequisites

Run:

command -v bazel >/dev/null && echo "bazel ok" || echo "bazel missing"
command -v az >/dev/null && echo "az ok" || echo "az missing"

Set Gemini key:

export GEMINI_API_KEY="$(az keyvault secret show --vault-name kv-prod-shared-secrets --name gemini-api-key --query value -o tsv)"

Step 1: Discover the runnable target

TARGET="$(bazel query 'kind(py_binary, //...)' | rg 'manual_gemini_from_run$' | head -n1)"
test -n "$TARGET" && echo "target: $TARGET" || echo "manual_gemini_from_run target not found"

If target is missing, stop and implement/add the helper binary first.

Step 2: Run classification

Basic image classification:

bazel run "$TARGET" -- \
  --run-id "<run_id>" \
  --output-dir /tmp/manual_gemini_from_run

With explicit timestamp:

bazel run "$TARGET" -- \
  --run-id "<run_id>" \
  --timestamp-unixus <timestamp_unixus> \
  --output-dir /tmp/manual_gemini_from_run

Step 3: Override prompt/model when needed

bazel run "$TARGET" -- \
  --run-id "<run_id>" \
  --prompt "Classify as one of: parking, pulling_over, driving_other. Return JSON with label/confidence/reasoning." \
  --model gemini-2.0-flash \
  --output-dir /tmp/manual_gemini_from_run

Output

Read:

cat /tmp/manual_gemini_from_run/classification_result.json

Expected JSON includes:

• run_id
• timestamp_unixus
• model
• fetched images per camera
• classification object from Gemini

Troubleshooting

• GEMINI_API_KEY is not set: export the key first.
• Databricks auth warnings can appear during timestamp/materialization steps; fallback image fetch may still succeed.
• Some runs/timestamps may only return a subset of cameras.
• If classification confidence is low or label is ambiguous, run a second timestamp nearby and compare.