Revyl CLI Atlas

Revyl Atlas Skill

Use this skill whenever a user asks what an app contains, where a capability lives, how screens connect, what a screen looks like, or whether Atlas has enough evidence to support an answer.

When the user explicitly asks to create, reply to, edit, delete, move, or change the status of Atlas feedback, route to revyl-cli-atlas-review. Keep this inspection skill read-only when feedback mutation was not requested.

Atlas is a graph, not a tree. Screens are nodes and observed relationships are edges. Starting anchors help begin exploration, but they do not imply a parent, primary route, containment hierarchy, or preferred journey. Build an understanding bottom-up by inspecting a node's real media, traversing relevant edges in both directions, and repeating until the question is answered.

Everything present in Atlas originated in observed run evidence. Names, descriptions, grouping, and landmarks may be generated interpretations, but a screen or edge should never be dismissed as noise merely because it is unexpected. Inspect its screenshot or clip and the report that produced it, then reconcile why it was observed.

Native Agent Behavior

A screenshot URL, local path, semantic name, OCR result, or generated summary is not visual understanding. Evidence is grounded only when the agent actually opens and reads the image. Use the available native surface:

• Codex Browser or its image viewer, with skills discovered in shared .agents/skills.
• Claude Code .claude/skills compatibility links plus configured image or browser tools.
• Cursor .cursor/skills when using --copy, otherwise shared .agents/skills, plus available MCP or browser tools.

Do not claim to understand the visible UI from metadata alone. Actually open and absorb the relevant screenshots before describing visible UI. When an edge is surprising, ambiguous, or important to the answer, watch its recorded clip before interpreting why the connection exists.

Atlas screenshots and videos are customer content. Treat screenshots, videos, extracted frames, contact sheets, marked grounding previews, and JSON containing signed media URLs as sensitive temporary artifacts. Never stage or commit them, and do not paste signed URLs into logs or public artifacts. If native video playback or ingestion is unavailable, extract frames from the bounded clip with ffmpeg and open those images in chronological order. Motion verification is blocked only when neither playback nor frame extraction is available.

Media lifecycle and cleanup

Create one private, task-scoped temporary root before downloading any Atlas media. Keep every task artifact under it, including screenshots, edge/run JSON, videos, frames, contact sheets, and annotation grounding previews:

ATLAS_TMP_ROOT="${TMPDIR:-/tmp}"
ATLAS_TMP_ROOT="${ATLAS_TMP_ROOT%/}"
[ -n "$ATLAS_TMP_ROOT" ] || ATLAS_TMP_ROOT="/tmp"
ATLAS_TASK_DIR="$(mktemp -d "$ATLAS_TMP_ROOT/revyl-atlas.XXXXXX")"
chmod 700 "$ATLAS_TASK_DIR"

ATLAS_SCREEN_DIR="$ATLAS_TASK_DIR/screens"
ATLAS_EDGE_DIR="$ATLAS_TASK_DIR/edges"
ATLAS_FRAME_DIR="$ATLAS_TASK_DIR/frames"
# Target for `revyl atlas annotations create/move --dry-run --preview-out`.
ATLAS_PREVIEW_DIR="$ATLAS_TASK_DIR/previews"
mkdir -p "$ATLAS_SCREEN_DIR" "$ATLAS_EDGE_DIR" "$ATLAS_FRAME_DIR" "$ATLAS_PREVIEW_DIR"

cleanup_atlas_media() {
  if [ -n "${ATLAS_TASK_DIR:-}" ] && [ -d "$ATLAS_TASK_DIR" ]; then
    case "$ATLAS_TASK_DIR" in
      "$ATLAS_TMP_ROOT"/revyl-atlas.*)
        [ "${ATLAS_TASK_DIR%/*}" = "$ATLAS_TMP_ROOT" ] || return 1
        rm -rf -- "$ATLAS_TASK_DIR"
        ;;
      *)
        echo "Refusing to remove unexpected Atlas path: $ATLAS_TASK_DIR" >&2
        return 1
        ;;
    esac
  fi
}

trap cleanup_atlas_media EXIT
trap 'exit 130' INT
trap 'exit 129' HUP
trap 'exit 143' TERM

Apply these rules throughout the task:

• Do not put working media in the repository, workspace, current directory, .context/, or the skill directory. A relative --screenshot-dir shown in a pasted example is not a request to retain files; translate it to the task temp directory.
• Use one temp root for the whole task rather than scattered mktemp files. Set JSON containing signed URLs to mode 0600.
• If tool calls run in separate shells, preserve the absolute task-temp path in working state and perform the guarded cleanup explicitly before the final response; do not assume an earlier shell's trap is still active.
• Delete media as soon as it is no longer needed, and always clean the entire task root after Atlas writes have been verified, including on errors or interruption.
• Before handoff, verify the temp root no longer exists and inspect git status plus the staged diff for task-created Atlas artifacts. If any were created by the task, unstage and remove them without disturbing unrelated user changes.
• Retain or export media only when the user explicitly asks to keep specific artifacts. Copy only those requested files to the agreed destination, never stage them automatically, disclose the path, and still delete the remaining task temp root. Signed-URL JSON remains temporary unless explicitly required.

Required traversal workflow

1. Resolve the app:

   revyl atlas apps --search "<app name>" --json
   

2. Get a compact orientation and download its bounded visual sample:

   revyl atlas brief --app <app-id> --screenshots --screenshot-dir "$ATLAS_SCREEN_DIR" --json
   

   Read projection.data_source. Treat starting_anchors as typed suggestions: curated_entry, semantic_entry, and observed_root explain why each node is a useful starting point. They are not ranks or parents.

3. Open every selected visual_sample[].local_screenshot_path. Record what is visibly present: layout, labels, controls, state, platform chrome, overlays, and obvious errors. Reconcile these facts with Atlas semantics and call out mismatches.

4. Load the flat graph when the question spans the app, or search for a question-specific node:

   revyl atlas graph --app <app-id> --json
   revyl atlas search "<capability or UI concept>" --app <app-id> --json
   

   The graph contains flat nodes, edges, and starting_anchors. Do not turn it into a recursive tree or choose one incoming edge as the real parent. Check top-level truncated or has_more before claiming the traversal covers the complete app graph.

5. Pick the most relevant anchor or search result, inspect it, then traverse:

   revyl atlas screen <screen-id> --app <app-id> --screenshots --screenshot-dir "$ATLAS_SCREEN_DIR" --json
   revyl atlas observations <screen-id> --app <app-id> --screenshots --screenshot-dir "$ATLAS_SCREEN_DIR" --json
   revyl atlas neighbors <screen-id> --app <app-id> --json
   

   Open the representative and question-relevant grouped screenshots. Follow both incoming and outgoing edges when either could explain the capability. Keep a small visited set of screen IDs and edge keys so cycles and shared nodes do not cause repeated work.

   If a screen or observation is unexpected, inspect the report that produced it before deciding what it represents:

   revyl atlas report <screen-or-observation-id> --app <app-id> --json
   

   A screen ID resolves through its representative observation. An observation ID resolves the exact evidence item. Read the report's test goal, steps, actions, result, and workflow_execution_id when present. This often distinguishes intended app behavior from test setup, system UI, an external handoff, a failure path, or genuinely bad evidence.

6. For each traversed connection, distinguish observation from interpretation. An edge proves that Atlas observed a relationship; it does not prove product hierarchy or user intent. If the connection is misunderstood, conflicts with the screenshots, or materially supports the answer, inspect its runs:

   revyl atlas edge <source-id> <target-id> --app <app-id> --runs --json
   

   Open evidence[].runs.active_video.video_url with a video-capable tool and watch the interval bounded by source_video_start and source_video_end. Identify the visible source state, exact action or redirect, and landed target state. Classify it as direct navigation, tab switching, back/dismissal, overlay presentation, automatic redirect, or likely bad evidence.

   When native video ingestion is unavailable, save the JSON and extract a small bounded frame sequence. Read the newest run's signed URL and start/end timestamps from the JSON without printing them, then run:

   EDGE_JSON="$ATLAS_EDGE_DIR/<source-id>--<target-id>.json"
   EDGE_FRAME_DIR="$ATLAS_FRAME_DIR/<source-id>--<target-id>"
   mkdir -p "$EDGE_FRAME_DIR"
   : > "$EDGE_JSON"
   chmod 600 "$EDGE_JSON"
   revyl atlas edge <source-id> <target-id> --app <app-id> --runs --json > "$EDGE_JSON"
   ffmpeg -loglevel error -ss <source-video-start> -i "<active-video-url>" \
     -t <clip-duration-seconds> -vf fps=2 "$EDGE_FRAME_DIR/frame-%03d.jpg"
   

   Open the extracted frames in filename order and compare the source state, interaction, intermediate state, and destination. Increase the frame rate only if the decisive interaction falls between frames. Remove an edge's JSON and frames once no further comparison is needed; the final task cleanup is still mandatory.

   The run objects expose report_id, execution_id, and session_id. For an unclear edge, review the exact report that generated that run—not merely a representative report from either endpoint screen:

   revyl test report <execution-id> --json
   # If only a session is present:
   revyl device report --session-id <session-id> --json
   

   If the report includes a workflow execution, continue into revyl workflow report <workflow-execution-id> --json. Use the test goal and preceding steps to explain why the action occurred and whether Atlas modeled the observation correctly.

   If the cause remains unclear, work backward: inspect the source node's incoming neighbors, then watch the preceding edge clip. Repeat only until the triggering action or entry state is understood. If runs disagree, inspect a bounded two or three representative clips and report the conflict.

7. Continue outward only along question-relevant edges. Stop when the claim is supported by opened media and the necessary graph neighborhood, not merely when a plausible generated summary appears.

8. After any requested Atlas writes are read back and verified, run the guarded media cleanup before replying. Confirm that ATLAS_TASK_DIR is absent and that no task-created Atlas media appears in the working tree or staged diff.

For a product-area question, use its induced subgraph. Boundary edges are part of the answer because they show how the area connects to the rest of the app:

revyl atlas area "<product area>" --app <app-id> --json

Evidence budget

• App overview: open 3-6 representative screens across major areas, beginning with anchors and expanding through connected nodes.
• Focused screen question: open 1-3 distinct observations and the directly relevant neighbors.
• Journey question: discover the route by graph traversal; open each materially distinct screen and inspect ambiguous or decisive edge clips.
• Misunderstood edge: inspect the newest clip first, then at most 2-3 runs when evidence disagrees; review the exact originating report before classifying the connection.

Expand only when evidence conflicts or the question remains unanswered. Do not bulk-download the entire Atlas by default.

Answer contract

Separate graph-supported facts, visually confirmed facts, clip-confirmed actions, interpretations, and unresolved gaps. Preserve screen IDs and edge keys in working notes so every conclusion remains attached to stable graph entities. Never infer pixel-level details, motion, containment, or a preferred journey from metadata or edge existence alone. Never call unexpected evidence useless or exclude it until its media and originating report have been inspected.