Publish the first six shell.online blog posts: a Codex run, a deadline with a stop control, session summaries, who answers Claude Code's prompts, file sharing without symlink escapes, and a background Haiku 5.5 job

[artemiia]

What changes for the user?

The blog gets its first six posts written for shell.online itself, all by Artemii Amelin. This pull request stays open as the daily posts accumulate until it is merged; each post is its own commit.

1. How to Watch a Long Codex Run From Your Phone and Share It With a Teammate (2026-10-03)

A step-by-step guide: install shell, start Codex through it, open the session on a phone, give a teammate a --read-only link, keep track of several runs with shell list, shell ls and the 0.25.0 session summaries, then rotate the password. The hook is OpenAI's GPT-6.1 Sol release on 29 September.

2. How to Put a Deadline and a Stop Button on an Unattended Coding Agent (2026-10-05)

A step-by-step guide: start an agent with --auto-close so the run has a hard deadline, watch it from a phone, and stop it three ways (Escape or Ctrl-C from the phone, the app's Stop button, shell kill on the host), with a section on what none of that undoes. The hook is OpenAI cancelling the GPT-6.1 Astra release over acting without permission and misreporting its actions.

3. How to See What Each Claude Code Session Is Doing From Your Phone (2026-10-06)

A step-by-step guide: start named Claude Code sessions with shell --name, find a session's account ID with shell ls --json, turn on the 0.25.0 session summaries with shell permissions <ID> --summaries=true, and read them in app.shell.online by long-pressing a row. It explains how shell binds a session to its Claude Code transcript (--session-id plus a SessionStart hook), what the enclave path does for other commands, and what summaries do not do. The hook is Claude Code 2.1.290 (published 5 October), which lets claude attach and claude logs find a background session by part of its name.

4. How to Choose Who Can Answer Claude Code's Prompts in a Shared Terminal (2026-10-07)

A step-by-step guide to deciding who can answer a Claude Code permission prompt when the session runs under shell: you alone from a phone (an interactive session nobody else has the password for), teammates watching while you answer at the host (--read-only plus shell attach), or several typists. It explains the relay's one-typist rule (a 1,800 ms input lease, "Guest N is typing…", local host before browser before MCP), a watcher agent on an observe-only grant (shell mcp grant <ID> "prompt watcher" observe 900), taking input back with shell password rotate, and a callout that "Yes, and don't ask again" saves a rule to .claude/settings.local.json that rotation does not remove. The hook is Anthropic's launch of Claude Code mods on 1 October, which can approve or deny permission requests and add confirmation steps.

5. How to Share a Coding Agent's Files Without Letting a Symlink Widen Access (2026-10-08)

A step-by-step security guide to shell --files and --files-root: pick the smallest root when the session starts, resolve it with pwd -P (a root given as a symlink is followed when it is opened), what viewers see in the Files panel (256 entries per listing, 16 MiB previews, 128 MiB downloads, pull-driven 16 KiB chunks, eight requests at once, read-only links still read opted-in files), and what happens when an agent plants ln -s ~/.ssh keys inside the root (left out of listings, refused on open by os.Root). A callout warns that dotfiles such as .env and .git/config are not hidden. The "does not do" section covers hard links (which pass), the fact that the root does not sandbox the agent, mount points, and that viewers cannot write. The hook is brig advisory GHSA-wp6x-29qx-fpr7 (26 September, fixed in brig 0.3.0) and Endor Labs' write-up of it (28 September), in which an agent's symlink made the sandbox mount a host directory read-write.

6. How to Run a Claude Haiku 5.5 Job in the Background and Step Back In (2026-10-09)

A step-by-step guide: start Claude Code on Haiku 5.5 with shell --name "..." claude --model claude-haiku-5-5, find it again with shell list, attach from a local terminal with shell attach <prefix> (screen replayed on attach, the attached window owns the grid, one local terminal at a time), detach with Ctrl-X then D (shell waits 750 ms for D and forwards any other Ctrl-X sequence, so Claude Code's Ctrl+X Ctrl+E/S/K shortcuts still work; Ctrl-] is the legacy key), check the run from a phone, and stop it with shell kill or --auto-close. A callout warns that Ctrl-C goes to Claude Code and a second press at an idle prompt exits it, which ends the session. It compares shell with Claude Code's own claude --bg / claude attach, and the "does not do" section covers attaching to an existing process, local-only attach, sleep, cost (Haiku 5.5's tokenizer counts about 30% more tokens) and crash recovery. The hook is Anthropic's release of Claude Haiku 5.5 on 7 October and Claude Code 2.1.293, which made it the default Haiku model the same day.

For each post:

• blog/posts.json gains one entry. None has a canonical, so unlike the 23 imported posts these are canonical on shell.online and are listed in /blog/sitemap.xml.
• blog/posts/<slug>.html is the body, in the format blog/README.md describes.
• public/blog/banners/<slug>.webp is the cover, a 1200×630 drawn illustration. Because it is not an SVG, the post also uses it as its social image.

The branch also carries a merge of main (8876e0e) so it stays current.

Verification

Run on 2026-10-09 on the branch head (435e2f0), which is main at 8876e0e plus the six posts:

• npx vitest run tests/blog.test.ts → 1 file, 15 tests passed.
• node scripts/test-landing-seo.mjs → "Primary landing SEO, attribution, and use-case checks passed."
• npm run build:web → exit 0 (only the existing chunk-size warning).
• House lint on post 6 → OK: 2,074 words including code blocks, 25 links (14 outside across 8 sites, 3 to pilotprotocol.network).
• Link check on post 6 → 25 links: 25 ok, 0 blocked, 0 dead.
• Every command and behavior in post 6 was checked against cmd/shell/help.go (start, list, attach, kill, password, e2ee, platforms and reference topics), cmd/shell/attach_unix.go (detach filter, 750 ms timeout, title handling, grid following), cmd/shell/sessions_unix.go (handleAttach, the "another local terminal is attached" refusal, resize only from an attached client), cmd/shell/terminal_grid.go and cmd/shell/session_unix.go (local grid limit set on attach and cleared on detach), cmd/shell/sessions.go (prefix rules and the shell list columns) and cmd/shell/session_output.go (the session card). cmd/shell is unchanged between v0.25.0 and main, so nothing described is unreleased. Claude Code flags and shortcuts were checked against the Claude Code CLI reference, interactive-mode page and CHANGELOG (2.1.293, 2.1.295); Haiku 5.5 figures against Anthropic's announcement and migration guide. Behavior was read from source, not run on a live session.

Posts 1 to 5 were checked when they were added (3, 5, 6, 7 and 8 October) as described in their commits and earlier revisions of this description: blog test, landing SEO check and build passed, and their link checks reported no dead links. Today's blog test and build above cover all six posts; the earlier posts' link checks were not repeated today.

Not done: no screenshots attached.

Access and privacy

No change to authorization, terminal input, passwords, encryption, logging or stored content. The links, password fragments, session IDs and session names shown in the covers and examples are invented.

• No real share links, credentials, private terminal content or deployment secrets are included.
• User-facing docs and the changelog are updated, or I explain why no update is needed.

No changelog entry: these are articles, not product changes, and the changelog already records the blog itself.

Rollout

Website only. No CLI or host update, no migration, no config change. The posts are live after the next site deploy.

🤖 Generated with Claude Code