shadow-testandClaude Opus 5
dd2894cc60
Two independent sets of fixes.
## Disk: stop the growth, no UI this round
The dangling-snapshot sweep was already correct and was never the leak. The
leak is that every `docker commit` **stacks** a layer and nothing compacts one:
a file deleted after it has been committed becomes a whiteout, not free bytes.
24 conditions trigger recreation+commit, so changing one settings field costs a
multi-gigabyte layer for the life of the project. One project was measured with
14 stacked commit layers, ~5.1 GB above its base.
* **Scrub the writable layer before every commit** (`docker/container.rs`,
`SNAPSHOT_SCRUB_PATHS` / `scrub_writable_layer`). The one moment those bytes
are still free to drop is before the commit that captures them. Measured on
one container's 4.48 GB pending layer: 3.0 GB of agent scratchpad under
`/tmp/claude-*`, the terminal drag-drop staging area (256 MiB per file, with
no `rm` for it anywhere in the repo), a PNG per pasted image, and the apt
lists/cache/logs that `browser_view/install.rs` and `triple-c-playwright-heal`
leave behind with no `apt-get clean`. A hardcoded list, never a heuristic:
`/workspace/{mount_name}` is a host bind mount and nothing here may reach one,
and the three `/tmp` globs cannot select the read-only `.host-ca`/`.host-aws`
mounts. Failure is a log line — a scrub must never block a snapshot.
* **Cap container logs** (`capped_log_config`). There was no `LogConfig`
anywhere, so containers ran on the daemon's unbounded `json-file` default.
Deliberately *not* wired into `container_needs_recreation`: participating
would recreate every project once, and a recreation costs a commit, which is
the thing being fixed. Picked up on the next natural recreation.
* **Make superseded base images sweepable** (`container/Dockerfile`). It carried
no `LABEL` at all, so `orphan_sweep_filters`' `dangling` + `triple-c.managed`
pair provably could not match one — ~11.9 GB observed stranded. Stamping
`triple-c.managed=true` is the whole fix; the sweep needed no change.
`create_container` writes the new `triple-c.base` key explicitly empty, or
Docker's label inheritance plus `docker commit` would make every snapshot
claim to be a base image. `force: false` stays, and now says why.
* **Sweep at startup** (`lib.rs`), not only after recreation: probes first
(a probe pins an image the unforced sweep then refuses), pins second, sweep
last. `sweep_orphaned_snapshots_logged` exists because all three callers threw
the report away — `reclaimed_bytes`, `failed` and `unavailable` included.
* **Reap migration leftovers.** `rollback_migration` retagged and orphaned the
migrated snapshot with no sweep. Stale `pre-migration-*` pins are now
age-reaped by scanning the tag pattern rather than trusting the state file —
`migration_store::load` reports an unparseable record as absent, which
stranded a 4-12 GB pin nothing could name again; `load` now moves a corrupt
record aside so `has_record` is trustworthy. A pin whose migration is still
awaiting confirmation is never reaped at any age. The probe container's
removal was a plain statement after an await, so a dropped future (an app quit
mid-migration) leaked a container pinning a multi-gigabyte image; it is a
`Drop` guard now, with `reap_probe_containers` for the case where the process
itself dies.
* **Prune scheduler logs.** `remove` deleted a task's JSON but never its log
directory, and the task runner appended uncapped `claude -p` output.
* **Fix the delete copy.** It said "the container, config volume, and stored
credentials"; it removes *both* volumes and the snapshot image.
No prune UI, and no unfiltered `prune_images`/`prune_volumes` anywhere — the
daemon is shared with the user's unrelated work.
## Claude Code settings: two invented keys, one inverted default, one sticky bug
Verified against code.claude.com/docs/en/settings-reference.md and env-vars.md.
* `effort` -> **`effortLevel`**, the key Claude Code actually reads; the old one
was written and silently ignored. `xhigh` added to the dropdown.
* `focusMode` -> **`viewMode: "focus"`**. `focusMode` was invented. The real key
does exactly what the existing UI hint already described.
* **Session recap was inverted.** Claude Code's recap is on by default, so
`CLAUDE_CODE_ENABLE_AWAY_SUMMARY=1`-when-enabled was a no-op and the control
could never turn the recap *off*. The field is renamed to
`session_recap_disabled` rather than reused: reusing the name with the
opposite meaning would have read every stored `enable_session_recap: false` —
which is every project that never touched the control — as "the user turned
this off".
* **The stickiness, which is the important one.** Keys were emitted only when
non-default, and the entrypoint *merges* into a settings.json on a persisted
volume, so switching a setting off omitted its key, the merge preserved the
stale on-value, and the setting stayed on until a destructive Reset. The fix
already existed in the same file — the sandbox block is emitted
unconditionally for exactly this reason — and is now applied to all five keys.
A key whose neutral state is *unset* (`tui`, `effortLevel`, `viewMode`,
`awaySummaryEnabled`) is emitted as JSON `null` and the entrypoint deletes it,
because a stand-in value is not neutral: `tui: "default"` pins the classic
renderer where unset lets Claude Code choose, and `viewMode: "default"`
overrides the user's own sticky `/focus` choice.
* The same stickiness existed, unnoticed, in the **env vars**: `docker commit`
bakes container env into the snapshot image, so a `=1` written once rode it
forever. All four are now emitted on every create, extracted into
`claude_code_env_vars` and unit tested. Two use an empty value for "off"
rather than `0`, because they outrank a setting the user can change from
inside their own container and Triple-C's default must not overrule a
`/config` choice it never asked about.
* TUI mode is now a genuine three-way choice (automatic / classic / fullscreen),
which the always-emitted key makes both necessary and possible.
`merge_claude_code_settings` is untouched by choice: a project-level OFF still
cannot override a globally-ON setting.
Tests: 364 frontend (+5), 308 Rust (+23), covering the scrub path list and
script, log rotation, pin reaping, and that toggling a setting off actually
clears a previously-set ON value.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc
266 lines
15 KiB
Markdown
266 lines
15 KiB
Markdown
# 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](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, 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](https://github.com/anthropics/claude-code/issues/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](https://github.com/anthropics/claude-code/issues/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 2–3
|
||
|
||
- **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.
|