sendblue-notify – skill Agent & Tự động hóa cho agent AI

Sendblue Notify

Overview

Outbound, fire-and-forget notifications from a local Claude Code session, script, or scheduled job to the user's phone via Sendblue. This is the "walk away from the terminal" pattern: kick off something long, get an iMessage when it lands. This skill owns when to notify and what to say. Actual sending goes through [[sendblue-cli]]. Hook wiring (so notifications fire automatically) goes through [[update-config]].

When to Use This Skill

• Use when the user says "text me when X is done", "ping my phone", "notify me on completion", "let me know when the build/deploy/migration finishes", or "send me an iMessage when…".
• Use when the user asks to wire a hook that texts on agent stop, /loop iteration, or /schedule completion.
• Use for a long-running task only when the user has explicitly requested phone notification.
• Do not use for short, interactive tasks where the user is watching the terminal — the notify is noise.

Prerequisites

The CLI must be installed and authenticated:

npx @sendblue/cli whoami        # confirms creds
# or, if first run:
npx @sendblue/cli setup

The user's phone number must be a verified contact on the account. On the free plan, the contact has to text the Sendblue number once before outbound sends work — confirm with sendblue contacts before relying on notify in an unattended workflow.

Cache the destination number once per project rather than re-asking. A NOTIFY_NUMBER env var or a one-line .notify-number file is fine; defer storage strategy to whatever the surrounding project already does.

Operational boundaries

These are imported documentation examples; no CLI, daemon, server or hook runtime

is bundled. Verify the pinned upstream version, account, billing, current limits and

actual command/API schema before use. Never print credentials (whoami/show-keys

may expose them) to agent output or logs. Authorization for drafting is not sending.

Do not circumvent provider anti-abuse limits by distributing sends across lines.

How It Works

Step 1: Decide whether notify is appropriate

Notify is for long, unattended work — not chatter. Good triggers:

• Agent turns over ~2 minutes (build, large refactor, migration, dataset crunch).
• /loop and /schedule jobs that produce a discrete result.
• CI / deploy completion when watched from the terminal.
• Multi-step playbooks where the user has gone heads-down on something else.

Bad triggers (do not silently wire these):

• Every Stop event, regardless of duration — produces spam, trains the user to ignore.
• Read-only or sub-second commands.
• Anything inside a tight loop.

If the user asks for "notify me when done" on a short task, do the obvious one-shot inline send (Example 1) and do not install a global hook.

Step 2: Pick a delivery pattern

• One-shot inline send — default for a single ad-hoc task. No config changes.
• Stop hook — opt-in, project-scoped, for sessions the user explicitly wants on automatic notify. Always gate by duration.
• End-of-/loop or /schedule ping — append the send to the routine's body.

Step 3: Compose the notification copy

• One line, under ~140 chars — fits in the lock-screen preview.
• Lead with outcome — ✅/❌, "done", "failed", "needs review".
• Include something actionable — branch name, error tail, PR number, duration.
• No emojis the user didn't ask for beyond a single status glyph.
• No agent self-narration ("I have completed the task as requested" — just say what happened).

Examples

Example 1: One-shot inline send

For a single task, append the send to the command. This is the default — no config changes, no surprise behavior later.

Branch on the task's exit status with an if/else. Do not use a task && send-success || send-failure chain: if the task succeeds but the success-send itself returns non-zero, the || fires the failure message — so the user sees ❌ even though the task completed. The if/else keeps the outcome tied solely to the task.

if long_running_thing; then
  npx @sendblue/cli send +15551234567 "✅ done: $(date +%H:%M)"
else
  npx @sendblue/cli send +15551234567 "❌ failed: $(date +%H:%M)"
fi

Or, when the result is interesting, include a one-line summary:

if run-migration >migration.log 2>&1; then
  npx @sendblue/cli send +15551234567 "migration succeeded; review the local log"
else
  npx @sendblue/cli send +15551234567 "migration failed; review the local log"
fi

Example 2: Claude Code Stop hook (opt-in, scoped)

Register a Stop hook in .claude/settings.json (project-scoped) — never in global settings unless asked. Defer the actual file edit to [[update-config]]. The hook command itself should:

1. Run cheaply (it fires on *every* Stop).

2. Gate on duration — skip sends for turns under a threshold (e.g. 90s).

3. Never fail the parent — pipe to || true so a notify error doesn't surface as a hook failure.

[ "$CLAUDE_TURN_DURATION_SECONDS" -ge 90 ] && \
  npx @sendblue/cli send "$NOTIFY_NUMBER" "turn done in ${CLAUDE_TURN_DURATION_SECONDS}s" || true

(Adjust the env var names to whatever the hook contract actually provides — verify against the current Claude Code hooks reference before writing the config; the harness owns those names, not this skill.)

Show the proposed hook config to the user and get confirmation before invoking [[update-config]]. Automated outbound messages are a footgun if the threshold is wrong.

Example 3: End-of-/loop or /schedule ping

/loop 10m "check deploy; npx @sendblue/cli send +15551234567 \"deploy: \$(deploy-status)\""

For /schedule, the routine itself can shell out at the end. Same copy rules apply.

Composing with textme

If the user has @textme installed (njerschow/textme — daemon that lets you *text Claude* from your phone), notify is still useful and not redundant. They run in opposite directions:

• textme: phone → Claude (user initiates from the phone).
• sendblue-notify: local Claude → phone (Claude initiates from a local session).

You can install both: textme on a server for inbound, notify as a local Stop-hook for outbound. Different problems, same Sendblue account.

Best Practices

• ✅ Default to one-shot inline sends for single tasks. Only escalate to a hook when the user asks for automatic notify.
• ✅ Gate hooks by duration. A 90s threshold is a sensible starting default.
• ✅ Show the hook config before installing it and get explicit user confirmation.
• ✅ Store the destination number per-user (env var or gitignored file), not in committed config.
• ❌ Don't install global hooks unless the user explicitly asks. Project-scoped is the default.
• ❌ Don't let a failed notify fail the parent. Trail with || true.

Limitations

• Notify is outbound-only. For "text Claude from the phone" use the @textme skill instead.
• On the free Sendblue plan, the destination phone must have texted the Sendblue number at least once before outbound succeeds. Verify with sendblue contacts before relying on notify in an unattended workflow.
• This skill does not own credentials, account setup, or the hook config file format. Those belong to [[sendblue-cli]] and [[update-config]] respectively.

Security & Safety Notes

• Lock-screen previews leak. Anyone holding the phone can read notification copy. Do not embed secrets, customer data, full error stacks, or auth tokens. Link to a log, dashboard, or PR instead.
• Confirm before sending or wiring hooks. Preview the destination, message template, trigger, and duration gate; wait for explicit user confirmation before running sendblue send or editing hook config.
• Automated outbound is a footgun. A misconfigured Stop hook can fire dozens of messages a minute. Always gate by duration and prove the threshold in a dry run before committing.
• Per-user numbers. The destination phone number is a personal identifier — keep it in user-local config (env var, gitignored file), not in committed repo files or CI logs.
• Free-plan verification is silent. If the destination contact hasn't texted in once, sends return an API error but the user just sees "no text arrived". Confirm verification before wiring an unattended hook.

Common Pitfalls

• Spam from over-eager Stop hooks. Always gate by duration. A user who gets pinged every 4 seconds will rip the hook out within an hour.
• Hardcoding the destination number in committed files. Use an env var or gitignored file; the number is per-user, not per-repo.
• Letting a failed notify fail the parent. Always trail with || true in hooks; surface the failure in logs, not by aborting the agent turn.
• Free-plan contact gotcha. If the destination contact hasn't texted in once on a free-plan account, the send silently fails for the user's purposes. Verify with sendblue contacts before

… (nội dung đầy đủ có trong gói tải về)