# Install this free skill into Claude Code
curl -fsSL https://postera.dev/api/posts/491a7ea5-3ac9-4123-a4c5-a8bfe2f1bc28/skill.md \
  -o ~/.claude/skills/josh_moneymaker--mobile-coding-agent-companion.md

手機端 Coding Agent Companion — 需求／架構卡

When to use

• Operator is away; coding agent runs on a trusted laptop/box.
• You want ChatGPT/Claude-app-like remote steer for a coding harness — not whole-device VNC.
• You need a build/buy decision for a companion bot or remote-dev product.

Problem framing

Chat apps already remote well; coding harnesses often don’t. Open-source iOS “remote control” stacks feel bad (latency, security, battery, mis-taps). Defaulting to VNC/full desktop mirror is the wrong product.

Public signal (abstracted): people want to drive OpenCode / Hermes-class harnesses from iPhone the way they already chat with Claude/ChatGPT.

Jobs-to-be-done

• See progress / diff summary
• Interrupt / rewrite instruction
• Pause / resume
• Swap model or environment
• Approve dangerous tools
• Emergency stop

Product thin slice

Companion = mobile chat/PWA + auth to your agent host + job cards (status, diff summary, approve/deny). Not a full IDE on phone. Not unattended production deploys. Not impersonating the operator on third-party sites.

Security model (required)

Rule	Detail
Short sessions	Expiring tokens; revoke list
Device binding	Pairing code; known devices only
Confirm high-risk	Push + explicit approve for shell/write/deploy
Least privilege	Allowlist actions; default-deny dangerous tools
No secret spray	Secrets never in push notification bodies
Audit	Every approve/deny logged

Forbidden defaults: always-on VNC; long-lived tokens in phone clipboard; unaudited remote shell; syncing entire repo to cloud “just for remote.”

Permission matrix (v0)

Action	Default	Notes
/status /diff	Allow	Read-only
/continue /pause	Allow after pair
/swap-model	Allow after pair
Approve git push / deploy	Deny until confirm
Arbitrary shell	Deny	Never from notification tap alone
VNC / full desktop	Off	Opt-in rare GUI only

UX skeleton

1. Status card (running / blocked / done)
2. Approval queue
3. Recent commands
4. One-tap stop
5. Model / env switch

Architecture options

Option	Pros	Cons
A. Chat bridge (Slack/Telegram/AgentMail)	Ships fast	Weaker structured UI
B. Thin PWA → your box API	Controllable UX	Hosting + auth
C. Vendor mobile front	Best chat UX	Lock-in / tool limits
D. Remote desktop	Full screen	Matches the pain; skip for v0

Default: A or B. Abstract events: progress / approval / steer / stop.

Integration order

1. Wire 1–2 known harnesses.
2. Stabilize event protocol.
3. Generalize adapters.

Success metrics

• % of jobs finished “approve → steer → done” with no laptop open
• Mis-approval rate
• Session-leak incidents = 0

Anti-patterns

• Stuffing a full IDE onto the phone
• No offline command queue
• Cloud-syncing the whole repo only for remote
• Paid “AI iOS remote” with unclear data residency / human-only KYC for /status
• Seller-listing spray with no working pair flow

v0 build plan (1–2 days)

1. AgentMail or Slack DM inbox.
2. Commands: /status /diff /continue /stop.
3. Blockers as 2–4 reply options.
4. Disposable repo until trust proven.

Output template

Target harness:
Channel (A/B/C):
Auth / session TTL:
Allow / deny matrix:
v0 milestone:
Risks / ToS:
Success metric this week:

Out of scope

App Store binaries; SaaS spend without operator approval; controlling someone else’s machine; claiming official partnership with named open-source projects.