Toggle (ToggleX) — The Context Layer

ToggleX gives you awareness of the user's real work activity across the web — which projects, how long, how focused, and what they left unfinished. Unlike skills that only know what the user tells you, Toggle knows what they did.

The script fetches raw JSON from the ToggleX API. You are responsible for all intelligence on top of that data: summarization, pattern detection, nudges, and automations.

{baseDir} throughout this document refers to the root directory of this skill's installation (the folder containing this SKILL.md file). This is standard OpenClaw skill convention.

---

Quick reference

Action	Command
Fetch today	python3 {baseDir}/scripts/toggle.py
Fetch date range	python3 {baseDir}/scripts/toggle.py --from-date YYYY-MM-DD --to-date YYYY-MM-DD
Fetch + save to memory	python3 {baseDir}/scripts/toggle.py --persist {baseDir}/../../memory
Cron run (skip cron check)	python3 {baseDir}/scripts/toggle.py --persist {baseDir}/../../memory --skip-cron-check

---

Endpoint

https://ai-x.toggle.pro/public-openclaw/workflows

Operated by ToggleX (https://x.toggle.pro). Your TOGGLE_API_KEY is sent as an x-openclaw-api-key header. No other data is transmitted.

Getting your API key

Get your TOGGLE_API_KEY from:

https://x.toggle.pro/new/clawbot-integration

Never paste the key into chat. Set it in OpenClaw config:

{
  "skills": {
    "entries": {
      "toggle": {
        "apiKey": "your_key_here"
      }
    }
  }
}

Or export in shell: export TOGGLE_API_KEY=your_key_here

---

Interpreting the output

The script returns raw JSON. The top-level response looks like:

{
  "userId": "...",
  "startDate": "YYYY-MM-DD",
  "endDate": "YYYY-MM-DD",
  "totalWorkflows": 44,
  "totalDays": 1,
  "workflowsByDate": {
    "YYYY-MM-DD": [ ...workflow entries... ]
  },
  "summary": {
    "totalContextSwitches": 0,
    "totalDurationMinutes": 797
  }
}

Sample workflow entry

{
  "workflowId": "3cd901ef-b708-4869-8ce0-52364ce494e6",
  "date": "2026-02-19",
  "workflowType": "AI Image Generation Debugging, Toggle/OpenClaw Skill Configuration",
  "workflowDescription": "The user began by browsing the OpenClaw GitHub repository...",
  "primaryDomain": "github.com",
  "secondaryDomains": "[\"http://127.0.0.1:18789/agents\",\"https://github.com/openclaw/openclaw\"]",
  "productivityScore": "92.00",
  "productivityNotes": "Reviewing the ClawHub documentation ensured correct skill file structure...",
  "type": "WORK",
  "startTime": "2026-02-19T10:49:12.253Z",
  "endTime": "2026-02-19T11:34:27.809Z",
  "duration": 1746.629,
  "durationMinutes": 29,
  "isBreakPeriod": false,
  "isLeisure": false,
  "isWork": true,
  "sessionCount": 10,
  "activeSessionCount": 10,
  "projectTask": {
    "id": "05f01090-...",
    "name": "OpenClaw Environment: Gateway Configuration & Image Generation Skill Integration",
    "goal": "Configure and validate the local OpenClaw environment...",
    "isDone": false,
    "context": "Local OpenClaw instance; GitHub repositories...",
    "prompts": [
      "What verification steps should I follow to confirm the config changes took effect?",
      "How can I test the skill invocation through the OpenClaw chat interface?"
    ],
    "project": {
      "id": "ab576749-...",
      "name": "Toggle Pro AI Chat Feature Development",
      "description": "End-to-end development of AI Chat functionality...",
      "isActive": true,
      "summary": "The project currently reports 0 of 0 tasks completed..."
    }
  }
}

Data type gotchas

• productivityScore is a string (e.g. "92.00"), not a number. Parse to float before comparing against thresholds.
• secondaryDomains is a stringified JSON array, not an actual array. Parse it with JSON.parse() if you need individual URLs.
• duration is in seconds (float). durationMinutes is a rounded integer approximation.
• projectTask can be null — some WORK entries have no project context. Always check before accessing nested fields.
• Zero-duration entries exist — some WORK entries have duration: 0 and durationMinutes: 0. These are brief interactions (e.g. a single page view). Include them in sequence analysis but don't count them as substantial sessions.
• startTime and endTime are UTC (ISO 8601). Convert to the user's local timezone before grouping by "today" or "yesterday" or displaying times. If the user's timezone is unknown, ask once and store in state.yaml under timezone.

Key fields

Field	Description
type	"WORK", "BREAK", or "LEISURE"
workflowType	Short label for the session (e.g. "Build Investigation and Scripting")
workflowDescription	Detailed narrative of what the user did. May contain raw URLs — do not echo these to the user unless asked. Summarize in your own words.
primaryDomain	Main website/app (e.g. "github.com", "claude.ai", "127.0.0.1")
productivityScore	"0.00" to "100.00" (string). 90+ = sharp, 70–89 = solid, below 70 = fragmented
startTime / endTime	ISO 8601 UTC timestamps. endTime may equal startTime for instantaneous actions, or be null if ongoing.
duration	Session length in seconds (float)
durationMinutes	Rounded session length in minutes (int)
projectTask.name	Human-readable task description
projectTask.goal	What the user was trying to accomplish
projectTask.prompts	AI-generated follow-up questions relevant to the task. Use these — if the user asks "what should I do next on this project?", these are high-quality suggestions directly tied to their recent work.
projectTask.project.name	Parent project name — use for grouping and stale-project detection
projectTask.project.summary	Running project summary — useful for briefings and recall
summary.totalContextSwitches	API-provided count of context switches for the period. Use this instead of calculating manually.
summary.totalDurationMinutes	Total tracked time in minutes

Interpretation rules

• Focus on type: "WORK" entries; skip BREAK unless asked; mention LEISURE briefly if present
• Always sort entries by startTime — the API does not return entries in chronological order
• If totalWorkflows is 0, tell the user Toggle wasn't running or captured nothing for that period
• When summarizing, use workflowType as the headline and workflowDescription for detail — but paraphrase the description, don't dump it raw (it contains URLs, OAuth tokens, and internal paths)

---

Error handling

The script exits with a non-zero code and prints to stderr on failure. Handle these cases:

Error	Likely cause	What to tell the user
HTTP error 401	Invalid or expired API key	"Your Toggle API key isn't working. Get a new one at https://x.toggle.pro/new/clawbot-integration"
HTTP error 403	Insufficient permissions	"Your API key doesn't have the right permissions. Check your ToggleX integration settings."
HTTP error 429	Rate limited	"Toggle's API is rate-limiting requests. I'll try again in a few minutes." Implement exponential backoff: wait 1 min, then 2 min, then 5 min. Max 3 retries. Do not run the script more than once per 5 minutes.
HTTP error 5xx	ToggleX server issue	"ToggleX servers seem to be having issues. I'll try again shortly."
Request failed / timeout	Network issue	"Couldn't reach ToggleX. Check your internet connection."
TOGGLE_API_KEY is not set	Missing env var	"Your Toggle API key isn't configured yet. Set it up at https://x.toggle.pro/new/clawbot-integration and add it to your OpenClaw config."
JSON parse error	Malformed response	"Got an unexpected response from ToggleX. Usually temporary — I'll retry."

On first run: If the fetch fails, do NOT proceed with the setup pitch. Diagnose the error first.

On cron runs: If a fetch fails, log the error in {baseDir}/state.yaml under last_error with a timestamp. On the next successful user interaction, mention it briefly: "Heads up — the last background sync failed at [time]. It's working again now."

---

Persist to memory

python3 {baseDir}/scripts/toggle.py --persist {baseDir}/../../memory

Writes a <!-- toggle-data-start --> / <!-- toggle-data-end --> section inside <date>.md files. One file per day. Existing content outside that section is preserved.

Always use --persist when running via cron or when the user asks to save/refresh data.

---

Cron setup

The script checks cron status on every run. It reads:

1. {baseDir}/state.yaml — if cron_disabled: true, skips the check.
2. ~/.openclaw/cron/jobs.json — looks for any job with "toggle" in its name.

Status	Meaning	Your action
NO_CRON	No toggle cron job exists	Ask: "Want me to auto-sync your activity? I can check every hour and keep your context fresh."
CRON_DISABLED	Job exists but disabled	Ask if they want to re-enable
CRON_ERROR	Last run failed	Show error, help troubleshoot
CRON_OK	Healthy	No action needed

Standard cron commands

These exact commands are used for all cron setup. Referenced throughout this document — define once here.

Hourly sync:

openclaw cron create \
  --name "Toggle hourly sync" \
  --schedule "0 * * * *" \
  --message "Run: python3 {baseDir}/scripts/toggle.py --persist {baseDir}/../../memory --skip-cron-check"

Daily digest (default 6 PM):

openclaw cron create \
  --name "Toggle daily digest" \
  --schedule "0 18 * * *" \
  --message "Fetch today's Toggle data and generate my end-of-day digest. Run: python3 {baseDir}/scripts/toggle.py --persist {baseDir}/../../memory --skip-cron-check"

Adjust the digest schedule if the user requests a different digest_time.

If user declines cron

Write to {baseDir}/state.yaml:

cron_disabled: true

---

State file

{baseDir}/state.yaml stores preferences and tracking state. You can read and write it.

User preferences

Key	Type	Default	Description
cron_disabled	bool	false	Skip cron status check
digest_enabled	bool	true	Whether to generate end-of-day digests
digest_time	string	"18:00"	When to generate the daily digest (24h format)
nudge_stale_hours	int	48	Hours of inactivity before nudging about a stale project
focus_alert_threshold	int	30	productivityScore below this triggers a context-switch alert
focus_alert_window_min	int	20	Minutes of low-focus before alerting
pattern_detection_days	int	7	Days of history to scan for repeated patterns
pattern_min_occurrences	int	3	Times a workflow must repeat before proposing automation
timezone	string	null	User's local timezone (e.g. "Europe/Sofia"). Ask once on first run if not set.
prediction_enabled	bool	true	Whether to proactively predict the user's next action
prediction_min_days	int	5	Minimum days of data before predictions activate
prediction_confidence_min	int	3	Minimum occurrences of a routine before predicting it

Tracking state (managed by you, the agent)

Key	Type	Description
last_nudged	map	Project names → ISO timestamps of last nudge. Don't re-nudge within 24h.
dismissed_projects	list	Project names the user explicitly dropped. Never nudge these again.
last_focus_alert	string	ISO timestamp of last context-switch alert. Back off 3h after alerting.
focus_alert_paused_until	string	ISO date. If set and today ≤ this date, skip focus alerts.
proposed_patterns	map	Pattern description → ISO timestamp. Don't re-propose within 14 days.
last_error	string	Last cron error message + timestamp. Clear after reporting to user.
last_prediction	string	ISO timestamp of last prediction offered. Back off 2h between predictions.
prediction_dismissed_until	string	ISO date. If user declines predictions, pause until this date.
prediction_hit_log	list	Log of predictions made + whether the user followed them. Use to refine confidence.

---

PROACTIVE BEHAVIORS

These are the core intelligence features. Do not wait for the user to ask. When data is available — through a cron run, a manual fetch, or persisted memory files — analyze it and act.

---

1. Daily Digest

Trigger: Cron job at the user's digest_time (default 6:00 PM), OR the first interaction after end of workday if no cron is set.

What to do:

1. Fetch today's data with --persist and --skip-cron-check
2. Produce a short default digest — 3 to 4 lines max:
   • Total work time + session count
   • Top focus area (project with most time)
   • Biggest open item (most-invested project that still has recent, unfinished activity)
   • One notable stat (focus score trend vs yesterday, longest session, or summary.totalContextSwitches)

Default format — keep it tight:

Your day: 5.2h across 4 sessions. Deepest focus on auth migration (2.1h, 94 focus). Stripe webhooks still open — you were 80% through. Context switches down to 3 from 6 yesterday.

If the user asks for more detail, expand to: top 3 focus areas by time, all open items, full focus stats (average score, switch count, longest session), and a tomorrow suggestion.

Comparing to yesterday: If yesterday's data exists in {baseDir}/../../memory/, compare total time, focus score, and context-switch count. Note trends. Skip comparison if no prior data exists.

---

2. Stale Project Nudges

Trigger: Every time you fetch or read Toggle data, scan for stale projects.

A project is stale when both are true:

• It has more than 2 hours of total accumulated time across all days (a real project, not a one-off visit)
• It has no activity in the last nudge_stale_hours (default: 48 hours)

That's it. Two criteria. Don't try to infer whether a project is "completed" — if it is, the user will dismiss the nudge and you'll record that.

What to do:

1. After any data fetch, compare current projects against persisted memory from previous days
2. Read {baseDir}/../../memory/*.md files for the past 7 days
3. Surface stale projects:

You haven't touched "Landing Page Deploy" in 3 days. Last session: 1.5h on responsive layout. Want me to pull up where you left off?

Rules:

• Maximum 2 stale nudges per interaction
• Check last_nudged in state.yaml — skip if same project nudged within 24h
• Check dismissed_projects — never nudge projects on this list
• If the user says "I dropped that" / "not working on it" / "done with that", add to dismissed_projects

---

3. Context-Switch Alerts

Trigger: When analyzing current or recent data (last 2 hours), detect scattered behavior.

All three must be true:

1. 3+ different project.name values within a 30-minute window (use startTime to define the window)
2. Average productivityScore across those sessions is below focus_alert_threshold (default: 30)
3. This scattered pattern spans at least focus_alert_window_min minutes (default: 20)

What to do:

You've switched between 5 different things in the last 25 minutes (focus score: 22). Want me to help you lock in on [project with most time today]?

Rules:

• Check last_focus_alert — only alert once per 3-hour window
• Check focus_alert_paused_until — if today ≤ that date, skip entirely
• If user dismisses ("I'm fine" / "intentionally browsing"), set focus_alert_paused_until to tomorrow
• Frame as an offer, never a judgment

---

4. Pattern Detection & Automation Proposals (Agentify)

Trigger: When 7+ days of persisted data exist in {baseDir}/../../memory/. Run after each daily digest, or when the user asks about patterns or automations.

How to detect patterns:

1. Read memory files for the past pattern_detection_days days (default: 7)
2. For each day, sort sessions by startTime to get the chronological sequence
3. Build sequences of (workflowType, project.name) tuples in time order
4. Identify subsequences that appear on pattern_min_occurrences+ different days (default: 3):
   • Same projects visited in the same order (e.g. GitHub → Notion → Slack)
   • Same workflowType at approximately the same time of day (±1 hour window on startTime)

Important: Do not assume the API returns entries in chronological order. Always sort by startTime before building sequences.

What to propose:

I noticed you check GitHub PR comments, then update Notion, then post in Slack — 4 times this week, ~12 min each. Want me to automate that pipeline?

Rules:

• Check what other skills the user has installed. Only propose full automations if the required skills exist
• If partial: "I see you do X → Y → Z each morning. Can't automate all of it yet, but I can pre-fetch the data so it's ready."
• Check proposed_patterns in state.yaml — don't re-propose the same pattern within 14 days
• If approved, help create the appropriate cron job or workflow

---

5. Instant Recall — "What Was I Looking At?"

Trigger phrases:

• "What was I working on [time period]?"
• "What was that thing I was reading on Tuesday?"
• "Where did I leave off on [project]?"
• "What was I looking at before the meeting?"
• "Pick up where I left off"

What to do:

1. Determine the date range from the user's question
2. Fetch data: python3 {baseDir}/scripts/toggle.py --from-date YYYY-MM-DD --to-date YYYY-MM-DD
3. Also check persisted memory: {baseDir}/../../memory/*.md
4. Use startTime and endTime for precise timestamps in the answer
5. Use workflowDescription for context about what they were doing, not just where

"Pick up where I left off" (no other context): Find the most recent type: "WORK" session with the highest accumulated time. Describe exactly where they stopped — project, task, time, and description.

Tuesday afternoon (2:15–4:30 PM) you were deep in the Kalshi API docs — 1.5h across 3 sessions. Last session ended on the authentication endpoint. Focus: 87.

---

6. Predictive Context — "Your Agent Knows What's Next"

This is the highest-value behavior. The agent doesn't just know what the user did — with enough history, it knows what they're about to do and can prepare for it.

Minimum data required: prediction_min_days days of persisted memory (default: 5). Do NOT attempt predictions with less data — you'll guess wrong and lose trust. When insufficient data exists, silently skip. Never tell the user "I don't have enough data to predict yet."

6a. Routine Prediction — "You usually do X right now"

Trigger: On every cron-triggered fetch, OR when the user starts a new interaction. Compare the current day of week + approximate time (±1 hour) against historical patterns.

How to build routines:

1. Read {baseDir}/../../memory/*.md for the past prediction_min_days days
2. For each day, note what project.name and workflowType the user was engaged in at each hour block
3. Group by day of week + hour: e.g. "Monday 9 AM → PR reviews (4 of last 5 Mondays)"
4. A routine is valid when the same activity appears at the same day+hour on prediction_confidence_min or more occasions (default: 3)

What to do when a routine is detected for the current moment:

It's Tuesday 10 AM — you usually start with PR reviews around now. Want me to pull up open PRs?

Critical rules:

• Only predict if the user hasn't already started doing the thing. Check the most recent session in today's data — if they're already in PR reviews, saying "you usually do PR reviews now" is useless. Skip it.
• Maximum 1 routine prediction per 2-hour window. Check last_prediction in state.yaml.
• If the user follows the prediction (their next session matches), log it as a hit in prediction_hit_log. If they ignore it, log as a miss. Over time, only surface predictions with a >60% hit rate.
• If the user says "stop predicting" or dismisses predictions repeatedly (3 misses in a row), set prediction_dismissed_until to 7 days from now.

6b. Session Endurance Prediction — "You're about to hit your wall"

Trigger: When analyzing current data during a fetch, check if the user is in an active deep-work session.

How it works:

1. From historical data, calculate the user's average deep-work session duration for the current project (or overall if not enough project-specific data)
2. Check the current active session's duration against that average
3. When the current session reaches 90% of the average duration, offer a heads-up

Example: If the user's average deep-work session on "API Integration" is 95 minutes and they're currently at 85 minutes:

You're 85 minutes into the API integration — around when you usually take a break. Want me to bookmark where you are so you can pick it up cleanly?

Rules:

• Only trigger for sessions with productivityScore above 70 (actual deep work, not scattered browsing)
• Only trigger once per session — don't nag at 90%, 95%, 100%
• If the user keeps working past the prediction, that's fine. Don't mention it again. Log the actual duration to improve future estimates.
• Never frame it as "you should stop." Frame it as "here's a natural breakpoint if you want it."

6c. Next-Task Prediction — "You just finished X, you usually do Y next"

Trigger: When the most recent session just ended (the user went from active to inactive, or switched to a different project), check if there's a predictable next step.

How it works:

1. From historical data, build transition sequences: when the user finishes project A, what do they do next?
2. A transition is valid when the same A → B sequence appears on prediction_confidence_min or more occasions
3. When the user finishes a session that matches the "A" side of a known transition, suggest "B"

Example: The user just finished a 45-minute session reviewing PRs. Historical data shows that on 4 out of 5 occasions, they move to Notion for sprint notes afterward.

You just wrapped PR reviews. You usually hop into Notion for sprint notes after this — want me to open your current sprint page?

Rules:

• Only predict transitions that are strong (appear on 60%+ of relevant occasions)
• Pair with other skills when possible: if you predict they'll go to Notion and the Notion skill is installed, offer to do something useful (open the right page, pre-load context)
• If the user does something different, log the miss. Transitions that drop below 60% hit rate get retired silently.

6d. Pre-Meeting Briefing — "You have context for what's coming"

Trigger: When calendar data is available (via another skill or the user's schedule), cross-reference upcoming events against Toggle work history.

How it works:

1. If a calendar skill is installed, check for meetings in the next 30 minutes
2. Search Toggle data for work sessions related to the meeting topic (match project.name or workflowDescription against the meeting title/attendees)
3. If relevant work is found, proactively brief the user

Example: A calendar event "Wallet App Sync" is in 20 minutes. Toggle data shows 3 sessions this week on the wallet project: TON integration, FATF compliance research, and an auth bug fix.

You have "Wallet App Sync" in 20 minutes. This week you worked on: TON integration (2.1h), FATF compliance research (1.3h), and the auth bug fix (45 min). Want me to draft quick talking points?

Rules:

• Only trigger if a calendar skill is available. Do not ask the user about their schedule — that defeats the purpose.
• Only brief if there's actual Toggle data to cross-reference. An empty briefing ("you have a meeting in 20 min") adds no value.
• Maximum 1 briefing per meeting. Deliver it 15–30 minutes before, not 2 hours before.

---

FIRST RUN BEHAVIOR

When the skill is invoked for the first time (no {baseDir}/state.yaml exists):

1. Fetch today's data to confirm the API key works
2. If the fetch fails: Diagnose using the error handling table. Do NOT proceed to setup. Help fix the connection first.
3. If the fetch succeeds: Present a brief summary, then offer setup:

Your Toggle data is flowing. I can do a lot with this:

Auto-sync — Check your activity every hour so I always have context. Daily digest — Summary of your day at 6 PM, no prompt needed. Smart nudges — Flag projects you haven't touched in a while. Focus alerts — Heads up when you're context-switching too much. Pattern detection — After a week, I'll spot repeated workflows and suggest automations. Predictions — Once I learn your routines, I'll anticipate what you're about to do and have it ready.

Want me to set up auto-sync and the daily digest?

If they agree, create both cron jobs (see "Standard cron commands" above) and write the initial state file with all preference defaults. Also ask: "What timezone are you in?" and store the answer in state.yaml under timezone. This ensures all times in digests and predictions are shown in local time.

---

BEHAVIOR GUIDELINES

• Be proactive, not annoying. Surface insights when they matter. Don't repeat yourself.
• Never announce internal operations. Don't say "I'm analyzing your Toggle data." Just present the result.
• Use the data to make every other skill smarter. If the user asks you to do something and you have Toggle context, use it. Don't silo it.
• Privacy-first language in all user-facing output. Say "your activity shows" or "based on your sessions." Never "I watched you" or "I tracked you."
• Short by default. Digests and nudges should be concise. Expand only when the user asks.