Files
Triple-C/ROADMAP.md
T
shadowdaoandClaude Opus 5 c0e4c87cec
Secret Scan / scan (push) Successful in 4s
Build App (Preview) / compute-version (pull_request) Successful in 3s
Secret Scan / scan (pull_request) Successful in 4s
Build App (Preview) / create-release (pull_request) Successful in 1s
Build App (Preview) / build-macos (pull_request) Successful in 2m44s
Build App (Preview) / build-linux (pull_request) Successful in 5m8s
Build App (Preview) / build-windows (pull_request) Successful in 6m28s
Build App (Preview) / prune-previews (pull_request) Successful in 2s
Build Container / build-container (pull_request) Failing after 14m49s
Give the mouse back, retire the follow controls, update Claude per session
Three things the terminal was getting wrong.

**A program that grabs the mouse and dies used to freeze the tab.** A TUI sets
DECSET ?1000/?1002/?1003; if it exits without resetting them, xterm keeps
routing clicks, drags and — under ?1003 — every pointer *move* to the PTY.
Text selection dies and escape bytes flood the prompt. The only exit was
closing the tab. `TerminalView` now reconciles a flag against
`term.modes.mouseTrackingMode` in the `term.write()` callback — the mode only
changes because the container printed a sequence, so one check per write
catches every transition with no polling — and `Ctrl+Shift+X` or a status-bar
button writes the resets back through `term.write`, never `sendInput`: the
reset belongs to xterm's parser, and a still-live TUI told about it would just
re-grab on its next repaint.

The control is in the status bar deliberately. Mouse tracking is the *normal*
state of htop, vim, lazygit and Claude Code, so a badge over the terminal
would be on screen for the whole life of those programs and would swallow
clicks aimed at their own top-right corner. `macOptionClickForcesSelection` is
also on now: xterm's force-select is Shift everywhere except macOS, where it
is Option and is gated behind that option, which defaults to false — so until
now Mac users had no way to select text while a program held the mouse.

**"Following" and "Jump to Current" are gone.** Claude Code draws on the
alternate screen, which has no scrollback, so `viewportY` always equalled
`baseY` and neither control could do anything. They did still work in bash
tabs; xterm's native follow covers that, and the per-write `scrollToBottom()`
went with them because it fought exactly that. What remains, on activate and
after a refit, now samples `viewportY >= baseY` *before* the fit, so opening
the Notes dock no longer yanks a reader to the tail.

**`claude update` runs before every Claude session, not just at container
start.** Containers here stop/start and often just keep running, so a
long-lived one never re-checked. Both copies take the same flock: the
entrypoint prints "container ready" only after its own update finishes, so
opening a tab immediately would otherwise run two updaters against the same
~/.claude/bin, with `|| echo` hiding a half-written install one line before
`exec claude` ran it.

This turns the non-Bedrock path from a bare argv into a `bash -c` wrapper, so
flags and session names are shell-interpolated now and must go through
`shell_quote_arg`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0145mQi9NZiCDrznBUEEDE4n
2026-09-08 11:02:33 -07:00

15 KiB
Raw Permalink Blame History

Triple-C Roadmap — Claude Code Feature Parity

Date: 2026-08-09 · Baseline: v0.3.0 · Claude Code reference: 2.1.226

Companion to DESIGN-REVIEW.md, which covers visual design and information architecture. This document covers which Claude Code capabilities Triple-C should surface, and why.


Guiding principle

Triple-C shows state and launches things. Claude Code edits its own config.

Triple-C's built-in MCP server management was removed in this cycle because Claude Code absorbed the capability natively (claude mcp add/list/remove, .mcp.json, /mcp). Hooks, skills, agents, plugins, output styles, and statusline are the same species: files under .claude/ with first-class Claude Code TUIs. Building GUI form editors for them means losing the same race again.

What Claude Code cannot do is what Triple-C uniquely owns: the container boundary and what persists behind it — the config volume, workspace mounts, lifecycle, the bundled scheduler, and the fleet view across many projects.


Current coverage (v0.3.0)

Triple-C sets exactly six settings.json keys, plus a sandbox block:

Key Surfaced as
tui TUI mode select — unset (Claude Code chooses), default (classic renderer), fullscreen (flicker-free alt-screen). Three distinct states, not two.
effortLevel Effort level select (low/medium/high/xhigh)
viewMode Focus mode toggle, written as "focus". Unset means the user's own verbose setting and sticky /focus choice still apply.
autoScrollEnabled Auto-scroll toggle. Claude Code's default is true, so it is the off state that writes false.
showThinkingSummaries Thinking summaries toggle (Claude Code default false)
awaySummaryEnabled Session recap toggle. Claude Code's recap is on by default, so again it is the off state that writes false.
sandbox.* Sandbox toggle (enabled, enableWeakerNestedSandbox, allowUnsandboxedCommands)

Every one of those keys is emitted on every start, with a JSON null standing for "delete this key". ~/.claude/settings.json sits on the config volume and the entrypoint merges into it, so a key merely omitted when its control goes off left the previous on-value in place forever.

Plus four env feature flags — CLAUDE_CODE_NO_FLICKER, CLAUDE_CODE_ENABLE_AWAY_SUMMARY, CLAUDE_CODE_SUBPROCESS_ENV_SCRUB, ENABLE_PROMPT_CACHING_1H — and arbitrary user-set CLAUDE_CODE_* vars via the Env Vars modal. The four are written on every container create including their off value, because docker commit bakes a container's env into the snapshot image: a value written once would otherwise ride that snapshot into every future container. That also makes them Triple-C's to own, so all four are reserved names — hand-setting one in the Env Vars modal is skipped with a warning, the same as any other triple-c.*-managed variable. CLAUDE_CODE_ENABLE_AWAY_SUMMARY is what actually enforces the recap choice — it takes precedence over awaySummaryEnabled and over the in-container /config toggle, so turning the control off sends 0 while leaving it on sends an empty value rather than 1: Triple-C's default must not overrule a /config choice it never asked about.

Also covered: per-project auth backends (Anthropic OAuth, Bedrock incl. SSO refresh, Ollama, OpenAI-compatible), user-level CLAUDE.md composition, claude update on every container start and before every Claude session launches, terminal ergonomics (OAuth URL detection, OSC 52 clipboard, image paste, file drag-drop, STT), the web terminal, and workspace backup.


Gap analysis

Committed for this cycle

# Gap Today Plan
1 Permission modes one boolean → --dangerously-skip-permissions Four-state control (Plan / Default / Accept Edits / Bypass) → --permission-mode. Verified choices on 2.1.226: acceptEdits, auto, bypassPermissions, manual, dontAsk, plan.
2 Session resume none List sessions from the config volume; [Resume] opens a terminal on claude --resume <id>.
3 Capability inventory none Read-only counts + names for skills / agents / hooks / plugins / commands / native MCP servers. Deep-link to the terminal to manage.
4 Automation triple-c-scheduler ships in every container with zero UI Task list, cron editor, run-now, logs, notification badges.
5 Container auth handoff manual code paste See "Authentication handoff" below — design decision pending.

Deliberately skipped

Status line builder · output-styles editor · hook editors · checkpoint/rewind browser · plugin marketplace browser. Each is niche, natively handled by Claude Code's own TUI, or a settings-editor trap. Surface counts and deep-link instead.

Not yet scheduled

  • Granular permissions.allow / ask / deny rules and additionalDirectories
  • Sandbox detail settings (filesystem.allowRead/allowWrite, allowedDomains, excludedCommands) — currently documented for hand-editing via SANDBOX_INSTRUCTIONS
  • Project-level .claude/settings.json vs user-level settings hierarchy
  • A model picker. Note: the only model strings in the app today are stale placeholders (anthropic.claude-sonnet-4-20250514-v1:0 in AwsSettings.tsx and ProjectCard.tsx, qwen3.5:27b, gpt-4o / gemini-pro / etc.). These are free-text placeholders, not dropdowns, but they should be refreshed to current model identifiers regardless.
  • The container's settings.json merge is shallow (jq -s '.[0] * .[1]'), so a user-authored nested block such as sandbox.filesystem.allowWrite is replaced wholesale on every container start. Worth deepening to * recursive merge.

Authentication handoff

Goal: stop making users hand-copy an auth code into every container.

Constraint discovered during research: claude login's callback server uses an ephemeral port and its redirect URI is not configurable for the main login flow (--callback-port and oauth.callbackPort apply to MCP server OAuth only). So a design that pre-assigns each container a fixed callback port and routes to it cannot work as stated — there is no fixed port to route.

There is also a known container gotcha: on Linux, Node resolves localhost to IPv6 first, so the callback server may bind [::1]:PORT only and be unreachable over IPv4 (anthropics/claude-code#44844).

Two viable options:

Option A — long-lived token injection (simple)

claude setup-token (verified present on 2.1.226: "Set up a long-lived authentication token (requires Claude subscription)") returns a ~1-year OAuth token. Triple-C runs it in a running container, stores the token in the OS keychain via the existing secure.rs, and injects CLAUDE_CODE_OAUTH_TOKEN into every container on the Anthropic backend.

Correction to an earlier assumption in this document. setup-token does not start a loopback callback listener, so it does not need the Auth Bridge. Verified by running it under a pty: its redirect_uri is Anthropic-hosted (https://platform.claude.com/oauth/code/callback), the user copies a code off that page, and the CLI blocks at a Paste code here if prompted > prompt on stdin. A stdin path is therefore mandatory — the flow cannot complete without one.

  • No routing, no ports, no proxy.
  • One auth event covers every project.
  • Cost: small. Reuses existing keychain and env-injection plumbing.
  • Limits: token is subscription-scoped and expires annually; per the docs a setup-token token cannot drive Remote Control sessions or claude.ai connector fetches.

Change detection uses a random rotation id in the triple-c.claude-token-version label, not a hash of the token. Labels are readable by anything that can run docker inspect, so a hash would be an offline verification oracle — given a candidate token you could confirm it. A presence boolean would instead miss rotations and silently leave containers on a stale token.

Option B — the Auth Bridge (general loopback-callback bridge)

Option A only solves Claude Code. The same problem affects every CLI that authenticates by starting a temporary loopback listener and opening a browser at a URL that redirects back to it — Concourse fly login (random loopback port serving /auth/callback), aws sso login, and many others. Inside a container the host browser cannot reach that listener, so login stalls.

Because the ports are ephemeral and unconfigurable, nothing can be pre-assigned. The bridge discovers listeners instead:

  1. While enabled for a running project, poll the container for loopback TCP listeners by reading /proc/net/tcp and /proc/net/tcp6 over docker exec — no dependency on ss/netstat/lsof, which aren't guaranteed in the image.
  2. For each newly-appeared loopback listener, bind the same port on the host's 127.0.0.1 (never 0.0.0.0 — that would expose container internals to the LAN).
  3. Proxy each accepted connection into the container over the Docker API via socat - TCP:127.0.0.1:<port> (socat already ships in the image), reusing the existing attached-exec streaming in docker/exec.rs. Going through the Docker API rather than a container IP keeps this working on Docker Desktop, where container IPs are not routable from the host.
  4. Fall back to TCP6:[::1]:<port> when the listener appeared only on IPv6 — on Linux, Node resolves localhost to IPv6 first, so claude login frequently binds ::1 only (anthropics/claude-code#44844).
  5. Tear down when the listener vanishes, the container stops, the bridge is disabled, or the app exits. Ports already covered by the project's explicit port mappings are skipped; host-side conflicts are reported rather than silently swallowed.

Opt-in per project (auth_bridge_enabled, default off), since it makes container-internal loopback services reachable from the host.

Plan: ship A for Claude Code specifically — it removes the pain for the common case at a fraction of the cost — and B as the general mechanism covering every other CLI. They compose: A means most users never trigger a browser login at all; B catches AWS SSO, Concourse, and anything else that needs a real callback.


Sequencing

Phase 0 — done. Remove MCP (frontend, backend, entrypoint, docs) with a self-healing migration for containers created against the old per-project Docker network.

Phase 1 — foundations. Permission modes end-to-end (including the scheduler bug fix below). Read-only introspection backend: sessions, capabilities, scheduler.

Phase 2 — Tier-1 polish. Focus rings, contrast fixes, real buttons, inline start/stop progress, status labels, onboarding welcome screen, shared accessible <Modal>.

Phase 3 — Project Home. Move project config out of the sidebar card into a tabbed main-area view (Overview / Sessions / Automation / Config), dissolving the modal pile and splitting the 1,257-line ProjectCard.

Phase 4 — authentication handoff. Option A, then evaluate B.

Phase 5 — Library. Global skills/agents/commands with per-project enable, synced into the config volume by the entrypoint. Generalizes the pattern the MCP tab was reaching for.


Bugs found during this review

  1. Scheduled tasks ignore the project's permission setting. container/triple-c-task-runner:69 runs claude -p "$PROMPT" --dangerously-skip-permissions unconditionally, regardless of the project's Full Permissions toggle. Being fixed as part of Phase 1.

  2. Docs claim Reset preserves credentials; it does not. rebuild_project_container calls remove_project_volumes, which deletes both triple-c-home-{id} (holding ~/.claude.json) and triple-c-claude-config-{id} (holding ~/.claude). README.md, HOW-TO-USE.md, and CLAUDE.md all still state that OAuth tokens survive a Reset. Pre-existing; not yet corrected.

  3. An invalid cron expression silently unscheduled every task. Found while adding task creation to the Automation tab, and the most serious bug in this review. triple-c-scheduler never validated --schedule, and rebuild_crontab regenerates the entire crontab and pipes it to crontab, which rejects the whole file if any single line is malformed — with the error discarded by 2>/dev/null || true. So one bad schedule silently unscheduled every other task in the container, reporting success. Reproduced directly. This mattered because the global CLAUDE.md instructs Claude to use this CLI, so Claude itself could trigger it. Fixed at the root: add now validates the expression and exits non-zero, and rebuild_crontab reports a rejected crontab instead of swallowing it. The Rust add_scheduled_task command validates independently.

  4. Reset was destructive with no confirmation. It deletes both volumes — the login, installed skills, all session transcripts — from a single unconfirmed click, while the comparably destructive Remove already confirmed. Now gated by a dialog that names each loss. Fixed.

  5. Cancelling authentication did not cancel. Fixed — see the handoff section above.

  6. Stale model placeholders — see "Not yet scheduled" above.

  7. Silent save failures. Project config saves on blur; failures went only to console.error. Fixed in Phase 3 — useProjectSave now renders a Saved / Saving / Save failed indicator and raises a toast.


Known gaps left by Phase 23

  • Editing a scheduled task changes its id. triple-c-scheduler has no edit subcommand, and hand-editing its JSON behind its back would desync the crontab, so edit is implemented as add-then-remove. The add runs first, so a rejected edit leaves the original intact. The task gets a new id and its older logs stay under the old one; the editor says so before saving.
  • open_terminal_session takes no command argument. "Resume session" and "Manage in terminal" therefore open a bash tab and type the command after a fixed prompt delay. It works, but it is timing-dependent and will misfire on a slow container start. The fix is a command: Option<String> parameter on the Tauri command so the exec launches the process directly.
  • Uptime is observed, not reported. get_container_info returns a status enum with no start time, so Project Home records "running since" when the app sees the transition. A container already running when the app launches shows ● Running with no elapsed time. Surfacing Docker's State.StartedAt would fix it.
  • lucide-react was not adopted (DESIGN-REVIEW Tier-1 #9) — no package-registry access in the build environment used for this cycle. The existing inline SVGs and text glyphs remain.
  • The tab strip stayed in the TopBar rather than moving onto the terminal panel's top edge. DESIGN-REVIEW §A6 asks for the move but its own §B2 layout diagram puts the tabs in the TopBar; the diagram won. Worth revisiting.
  • Ctrl+Shift+W, not Ctrl+W, closes a tab. Plain Ctrl+W is readline's kill-word, used constantly inside the terminal this app is built around; intercepting it globally would break word-erase in every shell.