Add abode terminal hub: tmux-backed Claude Code sessions for the June 15 billing split

[brendanlong]

Why

On June 15 the Agent SDK and claude -p move to a separate, API-rate credit pool; only the interactive Claude Code TUI stays on subscription limits. This PR adds a terminal-first way to run clawed-abode sessions that keeps usage in the interactive pool, designed for SSH/Termux use from a phone.

What

pnpm abode (or bin/abode) starts a terminal hub that shares the existing database, settings, and workspace layout with the web app:

• Session list with tmux liveness (●/○), attach, stop, stop & archive (with confirm), archived list
• Attach = full-screen handover to the interactive Claude TUI in tmux (status bar off, mouse on). F12 or Ctrl-b d detaches back to the hub, which asks: keep running in background / reattach / stop / stop & archive
• New session flow: searchable repo picker (favorites first, No Repository supported), branch picker, optional issue picker that prefills name + initial prompt (shared generateIssuePrompt), prompt editing in $EDITOR
• Settings editing for global + per-repo scopes via $EDITOR on a zod-validated JSON document — secrets decrypted for editing and re-encrypted on save
• Everything injected at launch from the DB (model resolution, layered system prompt, decrypted env vars via tmux -e, generated --mcp-config in the workspace dir) — nothing is ever written into the repo clone
• Restart-safe: sessions relaunch with claude --resume <sessionId> if Claude exited or the host rebooted (Session.id doubles as the Claude session UUID)

Refactors

• buildSystemPrompt/DEFAULT_SYSTEM_PROMPT extracted to SDK-free system-prompt.ts
• GitHub repo/branch/issue listing moved from the tRPC router into the github service (router now maps GitHubApiError → TRPCError)
• NO_REPO_SENTINEL centralized in src/lib/types.ts
• generateIssuePrompt extracted from the new-session page into src/lib/issue-prompt.ts

Testing

• Unit tests: claude argv/MCP-config/env builders, shell quoting (round-tripped through a real sh), settings-doc encrypt/decrypt round-trips, editor JSON parsing
• Integration tests: real tmux on a dedicated test socket (create/list/exact-match/kill, cwd + env injection)
• Manual E2E with a stub claude binary verified the full create→launch→archive flow: correct argv (session ID, model, layered system prompt incl. __no_repo__ custom prompt, initial prompt), cwd, DB env vars in the tmux session, and archive killing tmux + removing the workspace
• pnpm test:run (411 passed), pnpm test:integration (145 passed), typecheck and lint clean (one pre-existing warning)

Notes

• tmux ignores -e PATH=... (login shell resets it), so ABODE_CLAUDE_BIN overrides the claude binary path when needed
• TUI messages aren't persisted to the Message table; history lives in Claude Code's transcript (used by --resume) and tmux scrollback
• The web app is untouched and still works for settings or metered sessions

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a terminal-based CLI hub called 'abode' to manage interactive Claude Code sessions inside dedicated tmux instances, allowing users to list, attach, stop, and archive sessions, as well as edit global and per-repository settings using their local text editor. The feedback highlights several critical improvements: ensuring database consistency by wrapping repository settings upserts inside the transaction, preventing resume failures on errored sessions, avoiding shell-specific quoting issues in tmux by passing arguments directly, correcting session status tracking when archiving is canceled, and splitting the editor command string to handle arguments correctly.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The upsert operation is executed outside of the database transaction. If the subsequent transaction (which deletes and recreates the environment variables and MCP servers) fails, the parent repoSettings record will still be updated or created, leaving the database in an inconsistent state.

Using Prisma's interactive transaction API ensures that all dependent operations are executed atomically and rolled back together if any part fails.

Suggested change

	const settings = await prisma.repoSettings.upsert({
	where: { repoFullName },
	create: { repoFullName, ...settingsData },
	update: settingsData,
	});
	const envVarRows = doc.envVars.map((ev) => envVarDocToDb(ev, settings.id));
	const mcpServerRows = doc.mcpServers.map((server) => mcpServerDocToDb(server, settings.id));
	await prisma.$transaction([
	prisma.envVar.deleteMany({ where: { repoSettingsId: settings.id } }),
	prisma.envVar.createMany({ data: envVarRows }),
	prisma.mcpServer.deleteMany({ where: { repoSettingsId: settings.id } }),
	prisma.mcpServer.createMany({ data: mcpServerRows }),
	]);
	await prisma.$transaction(async (tx) => {
	const settings = await tx.repoSettings.upsert({
	where: { repoFullName },
	create: { repoFullName, ...settingsData },
	update: settingsData,
	});
	const envVarRows = doc.envVars.map((ev) => envVarDocToDb(ev, settings.id));
	const mcpServerRows = doc.mcpServers.map((server) => mcpServerDocToDb(server, settings.id));
	await tx.envVar.deleteMany({ where: { repoSettingsId: settings.id } });
	await tx.envVar.createMany({ data: envVarRows });
	await tx.mcpServer.deleteMany({ where: { repoSettingsId: settings.id } });
	await tx.mcpServer.createMany({ data: mcpServerRows });
	});

[gemini-code-assist]

If a session's status is 'error', it means the session failed during its initial creation (e.g., during repository cloning or workspace setup). In this scenario, no Claude session was ever successfully started, so there is no existing conversation to resume. Attempting to attach to an 'error' session will try to run claude --resume <sessionId>, which will fail.

We should only attempt to resume the session if the status is neither 'creating' nor 'error'.

Suggested change

	// Resume the existing Claude conversation unless the session never got
	// past creation (no conversation to resume yet).
	await launchClaudeInTmux(session, { resume: session.status !== 'creating' });
	// Resume the existing Claude conversation unless the session never got
	// past creation (no conversation to resume yet).
	await launchClaudeInTmux(session, { resume: session.status !== 'creating' && session.status !== 'error' });

[gemini-code-assist]

Passing the command as a single shell-quoted string forces tmux to execute it via the user's default shell (e.g., /bin/sh -c "..."). If the user's default shell is non-Bourne compatible (such as fish or tcsh), the single-quote escaping syntax used in shellQuote will cause syntax errors and fail to execute.

Passing the command arguments directly to tmux new-session as multiple arguments allows tmux to execute them directly via execvp without invoking a shell, completely bypassing shell-specific quoting issues.

Suggested change

	args.push(shellQuote(options.command));
	args.push('--', ...options.command);

[gemini-code-assist]

If Claude exited and the user selects "Archive it" but subsequently declines the confirmation prompt (i.e., archiveAfterConfirm returns false), the session is neither archived nor stopped. This leaves the session's database status as 'running' indefinitely, even though the backing tmux session is dead.

If the user declines the archive confirmation, we should fall back to stopping the session so its database status is correctly updated to 'stopped'.

Suggested change

	if (action === 'archive') {
	await archiveAfterConfirm(session);
	} else {
	await stopCliSession(session);
	}
	if (action === 'archive') {
	const archived = await archiveAfterConfirm(session);
	if (!archived) {
	await stopCliSession(session);
	}
	} else {
	await stopCliSession(session);
	}

[gemini-code-assist]

If the user's $EDITOR or $VISUAL environment variable contains arguments (for example, EDITOR="code --wait" or EDITOR="nano -w"), passing the entire string directly to spawnSync will fail with ENOENT because Node.js will look for an executable named exactly "code --wait".

Splitting the editor string by whitespace allows us to separate the executable binary from its arguments, ensuring compatibility with common editor configurations.

      const [editorBin, ...editorArgs] = editor.split(/\s+/);
      const result = spawnSync(editorBin, [...editorArgs, filePath], { stdio: 'inherit' });

[brendanlong]

Addressed all review feedback in 8ae0424:

• tmux argv (verified empirically on tmux 3.6): multi-argument commands are exec'd directly without a shell, so the command is now passed as -- ...argv and shellQuote is deleted entirely. Added an integration test that round-trips tricky arguments (quotes, $HOME, backticks, spaces) through a real tmux session.
• Repo settings save now uses an interactive transaction so the parent upsert rolls back together with the env var / MCP server replacement.
• error-status sessions no longer launch with --resume (nothing to resume).
• Declining the archive confirm after Claude exited now marks the session stopped instead of leaving it running.
• $EDITOR with arguments (code --wait, nano -w) is now split into binary + args.

The CI integration failure was a race in the new tmux test: a fast-exiting command shut down the tmux server (exit-empty) before the post-create set-option calls ran. Server option setup is now best-effort — for real sessions Claude keeps the server alive, and if the command already exited there's nothing left to configure. Ran the integration suite 5x locally to confirm it's stable.

— Claude