--- name: token-session-kickstart description: Schedules a minimal one-token remote Haiku session to fire at the exact moment the user's 5-hour Claude token window refreshes, so the new window starts immediately. Use whenever the user mentions their session is ending soon, wants to "kickstart" a new token window, says "my 5-hour session ends in X minutes", "schedule a session refresh", "kickstart my token window", or wants to maximize usable tokens per hour by pre-starting the next window. --- Schedule a one-token remote Haiku session to fire at the exact moment the 5-hour Claude usage window resets. ## Why this matters Claude Pro/Max plans use a 5-hour rolling token window. By firing a minimal task the instant the window resets, the new clock starts immediately. When the user resumes real work, they're already partway into the next cycle — maximizing tokens available in any given burst. ## Step 1: Get the window end time The reset time is server-side only — no local file or API exposes it. Try these sources in order: ### Path A — Chrome MCP (preferred, fully automatic) Check if `mcp__claude-in-chrome__*` tools are available (look for them in connected MCPs or load via ToolSearch). If connected: 1. Navigate to `https://claude.ai/settings/usage` 2. Read the "Current session — Resets in X min" value from the page 3. Compute `fire_at`: ```bash python3 -c " from datetime import datetime, timezone, timedelta mins = # extracted from page fire_at = datetime.now(tz=timezone.utc) + timedelta(minutes=mins, seconds=30) print(f'fire_at: {fire_at.strftime(\"%Y-%m-%dT%H:%M:%SZ\")}') print(f'Local: {fire_at.astimezone().strftime(\"%I:%M %p %Z\")}') " ``` 4. Confirm with the user: "Detected reset in X min (fire_at = HH:MM local) — does that match your dashboard?" Wait for explicit confirmation before proceeding. If the Chrome extension is closed/disabled, the tool call will fail — fall through to Path B. ### Path B — Ask the user Say: "Chrome MCP isn't available. Check https://claude.ai/settings/usage → 'Resets in X min' and tell me the value." Once they provide it, compute `fire_at`: ```bash python3 -c " from datetime import datetime, timezone, timedelta mins = # user-provided fire_at = datetime.now(tz=timezone.utc) + timedelta(minutes=mins, seconds=30) print(f'fire_at: {fire_at.strftime(\"%Y-%m-%dT%H:%M:%SZ\")}') print(f'Local: {fire_at.astimezone().strftime(\"%I:%M %p %Z\")}') " ``` **Never use the JSONL fallback** — it only sees Claude Code sessions and will be hours off if Claude.ai chat was used. Never assume confirmation; always wait for the user to explicitly confirm the time before scheduling. ## Step 2: Load RemoteTrigger ``` ToolSearch: { "query": "select:RemoteTrigger", "max_results": 1 } ``` ## Step 3: Create the kickstart routine Use this body (generate a fresh lowercase v4 UUID for `events[].data.uuid`): ```json { "name": "[util] token-kickstart HH:MM ET", "run_once_at": "YYYY-MM-DDTHH:MM:SSZ", "enabled": true, "job_config": { "ccr": { "environment_id": "env_01ThgKwovKTpvNexvtqNnE3S", "session_context": { "model": "claude-haiku-4-5-20251001", "sources": [], "allowed_tools": [] }, "events": [ { "data": { "uuid": "", "session_id": "", "type": "user", "parent_tool_use_id": null, "message": {"content": "ok", "role": "user"} } } ] } } } ``` Note the returned `trigger.id` — you'll need it in Step 4. **Naming convention:** always prefix with `[util]` so it's visually distinct from real work routines when reviewing at claude.ai/code/routines. Format: `[util] token-kickstart HH:MM ET`. ## Step 4: Create the cleanup routine Compute `cleanup_at` = `fire_at` + 5 minutes. Then create a second routine whose CCR job deletes the kickstart routine by ID: ```bash python3 -c " from datetime import datetime, timezone, timedelta fire_at = datetime.strptime('', '%Y-%m-%dT%H:%M:%SZ').replace(tzinfo=timezone.utc) cleanup_at = fire_at + timedelta(minutes=5) print(cleanup_at.strftime('%Y-%m-%dT%H:%M:%SZ')) " ``` Create this routine (replace `` with the ID from Step 3, and generate a fresh UUID): ```json { "name": "[util] token-kickstart-cleanup HH:MM ET", "run_once_at": "", "enabled": true, "job_config": { "ccr": { "environment_id": "env_01ThgKwovKTpvNexvtqNnE3S", "session_context": { "model": "claude-haiku-4-5-20251001", "sources": [], "allowed_tools": ["RemoteTrigger"] }, "events": [ { "data": { "uuid": "", "session_id": "", "type": "user", "parent_tool_use_id": null, "message": { "content": "Delete the RemoteTrigger routine with ID , then delete the cleanup routine with ID (your own ID).", "role": "user" } } } ] } } } ``` Note: `` is the ID returned from THIS create call — set it in the message before creating. So compute cleanup_trigger_id after creation and update the message, OR just instruct the CCR job to delete any `[util] token-kickstart*` named routines. **Simpler message alternative** (avoids the circular ID problem): ``` "Delete all RemoteTrigger routines whose names start with '[util] token-kickstart'. Use RemoteTrigger list action to find them, then delete each one." ``` ## Step 5: Confirm Output both URLs: - Kickstart: `https://claude.ai/code/routines/{kickstart_id}` - Cleanup: `https://claude.ai/code/routines/{cleanup_id}` ## Notes - Model must be `claude-haiku-4-5-20251001` — cheapest, minimum token burn. - `sources: []` and `allowed_tools: []` for the kickstart; cleanup needs `["RemoteTrigger"]`. - The 30-second buffer in `fire_at` ensures the window has fully reset before the task fires. - If the detected `time remaining` is under 3 minutes, warn the user — scheduling overhead may miss the window. - The cleanup routine fires 5 minutes after the kickstart and self-deletes both routines via the `[util] token-kickstart*` name pattern. - Environment ID `env_01ThgKwovKTpvNexvtqNnE3S` is the Default cloud environment for this account.