Drive Automation Session

Drive from Intent

Overview

Drive a Kobiton device from a natural-language intent. The caller provides:

1. An intent string (e.g. "open Settings, enable Bluetooth, then go back to home").
2. A UDID for a device the caller has already reserved.
3. An app reference (kobiton-store:vXXXXX) or a browser name for web sessions.

The skill opens an automation-type Appium session (with appium:newCommandTimeout: 1800 so the session survives pauses while the model asks for guidance), runs an observe-decide-act loop, and returns the session id. The returned session is consumable by the Kobiton MCP tools getSession, getSessionArtifacts, and saveTestCase exactly like any other automation session.

Tool naming. This doc refers to Kobiton MCP tools by their bare names (getSession, reserveDevice, terminateSession, …). The exact registered name depends on how the host loaded the MCP server (e.g. Claude Code as a plugin exposes mcp__plugin_automate_kobiton__getSession; a repo-local .mcp.json or claude mcp add exposes mcp__kobiton__getSession; other hosts differ). Models fuzzy-match on the bare name, so use it and let the host resolve the prefix.

This skill complements — and does NOT replace — run-interactive-session. They serve different session types:

Prerequisites

Runs on any OS, but needs two things from the host: a persistent local filesystem — the observe-decide-act loop passes state between turns through iter-N.*.json files, so that file handoff is the loop — and ~/.kobiton/.credentials, which scripts/appium.js reads directly (see below). A host with a filesystem but no way to run /automate:setup still can't finish; route those users to create-test-run. Node.js 18+; no native binary (scripts/appium.js uses node:https). See the Skill compatibility matrix in CLAUDE.md.

• Credentials available. scripts/appium.js reads ~/.kobiton/.credentials (written by /automate:setup) directly on every invocation and stops with no-credentials if it's missing — this file is required, not a fallback, and the script never calls the MCP getCredential tool. Reading it directly is deliberate: credentials never pass through argv, env, or the host transcript.
• A device. Either the user provides one (UDID, deviceName, platformVersion) or the skill helps pick + reserve one (see Step 0 below).
• An app (for app testing) — a Kobiton-store reference like kobiton-store:vXXXXX, an .apk / .ipa to upload, or a browser name (for web testing). The skill helps locate or upload it in Step 0.

Step 0: Device + app selection (ask before picking)

Before reserving anything, the skill clarifies what the user wants. Do NOT auto-pick a device unless the user's intent unambiguously implies a platform AND the user did not constrain the choice.

Device

If the user didn't specify a device in their intent:

1. Ask: "Which device or platform? (e.g., 'a Pixel running Android 14', 'an iPhone 15 Pro', or pick any available Android)." Wait for the response.
2. If they name a class (e.g., "any Pixel", "any iOS phone"), call listDevices with the relevant platform filter and present the top 3-5 options with name + OS version + availability.
3. If they name a specific device, verify availability via getDeviceStatus before reserving.
4. Reserve via reserveDevice. Capture the UDID, deviceName, platformVersion, automationName for Step 2's caps render.

If the user's intent strongly implies a platform (e.g., "test the iOS App Store flow", "open Chrome on Android"), it's OK to auto-pick — but state which device you're picking and why in one line before reserving, so the user can redirect.

App

For an app testing session:

• If the user referenced a build (@build.apk, "the latest staging build", kobiton-store:vXXXXX), use it. Verify the store reference exists via listApps if uncertain.
• If they didn't reference one, ask: "Which app should the session install? (path to a local .apk/.ipa, an existing kobiton-store:vXXXXX reference, or none if the intent is browser-only)."
• Upload local files via uploadAppToStore → confirmAppUpload → poll getAppParsingStatus until the state is terminal (OK = ready; stop on a FAILURE_* value). confirmAppUpload only kicks off async parsing on the backend — it does NOT guarantee the app is ready, so the session can fail to install if you skip the poll.

For a web testing session, skip the app step — --testingType web --browserName chrome|safari is enough.

Live view

If the user didn't say how they want to observe the session, ask before creating it:

"Open the device live view to watch the agent drive (foreground), or run the session in the background?"

Two options:

• Open in foreground — after the session is created, build the device-only URL <portal>/devices/launch?id=<deviceId>&view=device-only (same shape as run-automation-suite Step 5; the <deviceId> is the one captured during Step 0 reservation) and launch it via the chromeless launcher chain (see Step 3b for the chromeless-vs-default-browser branching). The URL is also surfaced as text so the user can click manually if the launcher can't open Chrome and no default-browser fallback applies.
• Run in background — proceed straight to the loop. The user can still inspect artifacts (XML/PNG/video) after the cycle ends.

Remember the choice; act on it in Step 3 right after the session is created. Do NOT auto-open without asking — some users prefer silent runs, especially on shared screens.

Inputs

Steps

1. Verify credentials are available

appium.js reads ~/.kobiton/.credentials directly on every invocation. If the file is missing, tell the user to run /automate:setup and stop:

if [ ! -s ~/.kobiton/.credentials ]; then
  echo "Credentials not found. Run /automate:setup first." >&2
  exit 1
fi

2. Render desired capabilities

Reuse run-automation-suite's renderer. The --newCommandTimeout 1800 flag is the key piece for this skill.

SKILL_DIR=$(dirname "$0")   # this skill's directory
RENDER=$(realpath "$SKILL_DIR/../run-automation-suite/scripts/render-capabilities.js")

# Unique temp file — two concurrent sessions (two terminals / conversations)
# must not clobber each other's rendered caps before Step 3 reads them.
CAPS_TMP=$(mktemp "${TMPDIR:-/tmp}/drive-automation-session-caps.XXXXXX.json")

node "$RENDER" \
  --platformName "$PLATFORM_NAME" \
  --udid "$UDID" \
  --deviceName "$DEVICE_NAME" \
  --platformVersion "$PLATFORM_VERSION" \
  --automationName "$AUTOMATION_NAME" \
  --app "$APP" \
  --testingType "${TESTING_TYPE:-app}" \
  --newCommandTimeout 1800 \
  --scriptlessCapture \
  > "$CAPS_TMP"

appium.js reads ~/.kobiton/.credentials on each invocation. No flags, no env vars.

3. Create the session

SESSION_ID=$(node "$SKILL_DIR/scripts/appium.js" \
  --method POST --url /session \
  --req-body "@$CAPS_TMP" \
  | jq -r '.value.sessionId // .sessionId')

SESSION_DIR=".kobiton/sessions/$SESSION_ID"
mkdir -p "$SESSION_DIR"
mv "$CAPS_TMP" "$SESSION_DIR/caps.json"
printf '%s session=%s started intent=%q\n' "$(date -u +%FT%TZ)" "$SESSION_ID" "$INTENT" > "$SESSION_DIR/session.log"

trap 'cleanup' EXIT INT TERM
cleanup() {
  [ -z "$SESSION_ID" ] && return 0
  # DELETE /wd/hub/session/<id> ends the WebDriver session cleanly; Kobiton
  # records state=COMPLETE. appium.js already treats 404 as idempotent success
  # (the session was already gone). appium.js always exits 0 by policy, so we
  # check stderr instead of $? to detect a real failure.
  delete_err=$(node "$SKILL_DIR/scripts/appium.js" \
    --method DELETE --url "/session/$SESSION_ID" 2>&1 >/dev/null)
  if [ -z "$delete_err" ]; then
    printf '%s session=%s end-via-trap status=COMPLETE\n' "$(date -u +%FT%TZ)" "$SESSION_ID" >> "$SESSION_DIR/session.log"
  else
    printf '%s session=%s end-via-trap delete-failed (session will time out per appium:newCommandTimeout) err=%s\n' "$(date -u +%FT%TZ)" "$SESSION_ID" "$delete_err" >> "$SESSION_DIR/session.log"
  fi
}

3b. Act on the live-view choice from Step 0

Build the device-only URL from the device id captured during Step 0 reservation and the portal base URL (e.g., https://portal.kobiton.com) — the AI host has the portal value from earlier MCP responses in the conversation context. Same URL shape as run-automation-suite Step 5.

LIVE_VIEW_URL="<portal>/devices/launch?id=<deviceId>&view=device-only"
printf 'Live view: %s\n' "$LIVE_VIEW_URL"

(The post-run replay URL is <portal>/sessions/<kobiton-session-id>/explorer — a different page, NOT the live view. Surface that one only after the session completes if the user wants to review the recording.)

If the user chose background in Step 0, stop here — the URL is logged and the loop proceeds. They can paste it into a browser later if they get curious.

If the user chose foreground, reuse run-automation-suite's launcher chain instead of re-implementing it. The pattern (chromeless Chrome for users who haven't expressed a preference or prefer Chrome; default-browser fallback otherwise) is identical to run-automation-suite Step 5.

Launch via the chromeless helper (default — gated on saved browser preference). Check auto memory for a saved browser preference:

• No preference saved, OR preference is Google Chrome → invoke the chromeless launcher below.
• Preference is Safari, Firefox, or Default browser → skip the launcher; go straight to the default-browser fallback table.

Pick width and height from the reserved device's form factor (the device name from Step 0):

Swap width and height if getSession / rendered caps report orientation=LANDSCAPE — default is portrait.

Invoke per host OS. Run in the background (Claude Code: Bash tool with run_in_background: true; other hosts: append & and disown) so the resize-polling loop doesn't block iteration 1. The launcher's stdout/stderr will surface later when it completes; the URL printed above is the user's fallback if the launcher silently fails.

Launcher exit codes (surface in the background-task completion event):

• 0 — Chrome was launched (resize may or may not have succeeded; the warning is informational).
• 2 — Chrome / Chromium is not installed. The user already has $LIVE_VIEW_URL from the printf above; they can paste it into their default browser, or you can offer the default-browser fallback table below as a follow-up.
• 64 — usage error in the launcher invocation. Surface to the user.

Default-browser fallback (when the saved preference is Safari/Firefox/Default, or chromeless exited 2). Ask the user once which browser to use (save to auto memory):

On Linux, fall back to xdg-open "$LIVE_VIEW_URL" (browser selection isn't supported).

If neither path is available, the URL is already on stdout from the printf above — the user can copy-paste it.

4. Per-turn iteration pattern (three branches)

There is no Bash while. The skill is turn-based: each turn, you (the AI host) increment ITER and pick exactly one of three branches:

Each branch consumes one ITER. The script always exits 0; the host detects failures by checking whether iter-N.error.json was written.

export ITER=$((ITER + 1))

# Pick ONE of the three. Credentials inherit from env (Step 1).

# Branch A: observe — captures BOTH iter-N.xml and iter-N.png by default.
#   Native overlays / OS dialogs only show in the PNG (the webview source
#   XML doesn't see them). --xml-only skips the screenshot to save tokens
#   when you trust the XML is complete; --png-only skips the source when
#   you're verifying layout / image rendering only.
node "$SKILL_DIR/scripts/appium.js" screen \
  --session-id "$SESSION_ID" --session-dir "$SESSION_DIR"
# Stdout: {"hash":"<sha256>","xmlBytes":N,"pngBytes":M}
# Track the hash across turns in your conversation context — repetition is a
# signal, never a forced stop. See references/loop-discipline.md "Stuck patterns".

# Branch B: act — execute an Appium call. Writes iter-N.request.json + either
#   iter-N.response.json (success) or iter-N.error.json (any failure). The
#   error file contains the raw Appium body or {error, message} for usage
#   errors; the host reads it to decide what to do next turn.
node "$SKILL_DIR/scripts/appium.js" <YOUR-ARGV> --session-dir "$SESSION_DIR"
# Examples of <YOUR-ARGV>:
#   --method POST --url /session/$SESSION_ID/element --req-body '{"using":"xpath","value":"//Button"}'
#   actions --session-id $SESSION_ID --type swipe --from-x 540 --from-y 1800 --to-x 540 --to-y 600
#   touch-perform --session-id $SESSION_ID --steps @/tmp/steps.json
#   --method POST --url /session/$SESSION_ID/element/$EL_ID/clear --req-body '{}'

# Branch C: end the cycle.
node "$SKILL_DIR/scripts/appium.js" control \
  --done --reason "Settings reached and Bluetooth toggled" \
  --session-dir "$SESSION_DIR"
# OR --blocked --reason "Two modals stacked; cannot dismiss"

# ---- Per-turn bookkeeping (runs after the chosen branch) ----

# Log failures so session.log has a timeline. appium.js writes iter-N.error.json
# on any failure — Appium HTTP error, network blip, usage error like missing
# flag or malformed JSON. The host reads the file next turn and re-plans.
ITER_PAD=$(printf '%03d' "$ITER")
if [ -f "$SESSION_DIR/iter-$ITER_PAD.error.json" ]; then
  printf 'iter=%d error\n' "$ITER" >> "$SESSION_DIR/session.log"
fi

# Iteration ceiling — safety net only. Not a stuck-detection mechanism.
if [ "$ITER" -ge "${MAX_ITERS:-100}" ]; then
  printf 'iter=%d max-iters-reached limit=%d\n' "$ITER" "${MAX_ITERS:-100}" >> "$SESSION_DIR/session.log"
  exit 0
fi

Branch decision guide

What to pick next turn depends on what just happened:

• Successful act → next turn screen (the screen probably changed).
• screen just ran → next turn act (you have a fresh observation).
• act returned no such element / invalid selector / bad-input → next turn act again with a corrected call. The screen didn't change (the action didn't fire), so the most recent iter-K.xml is still valid; re-read it from disk if you need to, but don't burn an ITER on a fresh screen.
• act returned stale element reference → next turn screen (the element id is from a prior state; you need a fresh observation to get current element ids).
• act returned HTTP 5xx / network timeout → next turn act again with the same call (retry). If it fails twice, control --blocked.
• Goal reached → control --done.
• Genuinely stuck (per the patterns in loop-discipline.md "Stuck patterns") → control --blocked.

The host is responsible for tracking which iter-K.xml represents the current screen state. After a successful act, the previous iter-K.xml is stale until the next screen. After a failed act, the previous iter-K.xml is still current.

Building act calls from the XML

When the chosen branch is act, the argv body is constructed from the most recent iter-K.xml — selectors come from attributes you can see in the XML, not from guesses. See references/endpoint-reference.md "Building Appium calls from the observed XML" for the find-element → element-id workflow, selector-strategy preference order (accessibility id > id > relative xpath > css selector > class name), and the coordinates-from-bounds fallback for gestures with no element target.

The full loop discipline (artifact paths, error feedback, termination conditions) lives in references/loop-discipline.md. The endpoint catalog (allowlisted vs not, helpers vs generic) lives in references/endpoint-reference.md.

5. Cleanup

The Bash trap (set up in Step 3) issues DELETE /wd/hub/session/$SESSION_ID on exit (loop end, error, user stop). That's the clean way to end the session — Kobiton records the session state as COMPLETE, which is what saveTestCase and analytics expect.

Do NOT call the terminateSession MCP tool here. It ends the session via Kobiton's platform-side kill path, which marks the session state as TERMINATED — distinct from COMPLETE and treated as an abnormal exit by the recording pipeline. Reserve terminateSession for the genuinely-stuck case where the WebDriver DELETE is unreachable (network down, Appium hub unresponsive) AND the user explicitly asks to force-kill the session.

If the DELETE somehow fails silently, the session will time out on the Kobiton side per appium:newCommandTimeout (30 min, set in Step 2). Better to wait for that timeout than to mark the session TERMINATED.

6. Return

Emit the session id back to the caller. The session is now consumable by:

• getSession({sessionId}) — session metadata, device info, results.
• getSessionArtifacts({sessionId}) — video, logs, screenshots, reports.
• saveTestCase({sessionId}) — persist as a re-runnable test case (sibling feature, out of scope here).

Endpoint reference

See references/endpoint-reference.md for the full Appium endpoint catalog — allowlisted (capturable by saveTestCase) vs not, helpers (actions, touch-perform) vs generic mode. The AI host picks endpoints from that table when emitting iter-N.call.json.

Stuck patterns — host decides when to pause

The script does not enforce blocker thresholds. The AI host tracks the hash from screen across turns in its conversation context and decides when to emit control --blocked. See references/loop-discipline.md "Stuck patterns" for concrete examples — same-call repetition, screen oscillation (A→B→A→B), credentials prompt, CAPTCHA, lazy load (use a no-op observe to wait), network spinner.

The only hard programmatic stop is MAX_ITERS=100 (override per session), which is a safety net against runaway cycles — not a stuck-detection mechanism.

Errors

Notes

• The kobiton:aiToolName capability is auto-detected by render-capabilities.js (Claude / Codex / Copilot / Gemini). No flag needed.
• Artifacts live under .kobiton/sessions/<session-id>/ — workspace-relative, not /tmp (consistent with run-interactive-session).
• The skill writes nothing to tools/*.yaml and does not add any MCP tool. It only consumes existing MCP tools (reserveDevice, getSession, terminateSession, etc.).