100 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
shadow-test	6cc48b3266	Document the Wayland icon-cache-needs-relogin gotcha ... A user hit this after installing the new Arch/CachyOS package (triple-c#34): icon missing in the app menu, taskbar, and titlebar alike, with no error in the app's own log. Root cause has nothing to do with the app or its packaging — GNOME/KDE cache the installed-app list and resolved icons in the shell process's memory at startup, and Wayland has no equivalent to X11's soft shell-restart trick to force a live reload. Logging out and back in fixed it for them.	2026-08-27 15:53:19 -07:00
shadow-test	f2cfc0be8f	Also attach the Arch package to the matching Gitea release ... The workflow only ever uploaded to the GitHub release — the Gitea release for the same version (the plain, unsuffixed vX.Y.Z tag build-app.yml's Linux job creates, which already holds the .deb/.rpm/.AppImage) never got it, so it looked missing to anyone checking releases on Gitea instead of GitHub. New step mirrors build-app.yml's own Gitea upload step exactly: same get-or-create-by-tag, delete-existing-asset, upload-as-octet-stream shape, same REGISTRY_TOKEN secret. Verified the read side (release lookup, asset listing) against the real v0.4.16 release before writing this — resolves to the correct release id and correctly finds no existing asset yet.	2026-08-27 15:14:54 -07:00
shadow-test	99c9dd3cc2	Add an Installation section — nothing told a new user how to get the app ... HOW-TO-USE.md's Prerequisites jumped straight to Docker and a Claude Code account, assuming Triple-C was already installed; the app itself had no download/install instructions anywhere in the docs. Covers all six release assets, including the new Arch/CachyOS .pkg.tar.zst (triple-c#34) that publish-arch-package.yml now attaches to each release.	2026-08-27 15:06:43 -07:00
shadow-test	3c8296843f	Skip AUR for now — attach the built Arch package to the GitHub release ... Publishing to the AUR needs a maintainer AUR account and its SSH key registered as a secret here, neither of which exists yet. Rather than leave the workflow permanently failing at that last step, it now stops short of AUR and instead uploads the built .pkg.tar.zst to the same GitHub release it built from, as a plain downloadable asset (`pacman -U` to install). The AUR-push step is still in this file's git history if that setup happens later. Renamed publish-aur-package.yml -> publish-arch-package.yml to match. The render/validate steps are unchanged; new here is capturing the exact built package filename from inside the build container (makepkg is the only thing that actually knows it) and an upload step that follows the same create-or-reuse-release, strip-upload_url, POST-octet-stream pattern build-app.yml and backfill-releases.yml already use for GitHub assets, plus a delete-existing-asset-first step so a re-dispatch for an already-packaged version replaces rather than 422s. Verified with a real Docker run end to end: rendered a real PKGBUILD, built a real (synthetic) .deb through makepkg + namcap in an archlinux container, confirmed the container exits 0, and confirmed the exact package filename it captures (triple-c-bin-<version>-1-x86_64.pkg.tar.zst) round-trips out via docker cp intact.	2026-08-27 14:48:39 -07:00
shadow-test	6dcdeb89cb	Fix PKGBUILD render silently no-op'ing on every AUR publish run ... The "Render PKGBUILD" step's Python heredoc built its old_source match string via an f-string, escaping literal braces as `${{pkgver}}` — which put that exact four-character sequence directly in this workflow file's own YAML text. Gitea Actions scans a run: block for `${{ ... }}` and tries to evaluate whatever's inside as one of its own expressions before the shell ever sees the script; "pkgver" isn't a valid expression context, so every run has been failing that interpolation and emptying the step instead of raising anything visible there. The next step's `makepkg` then failed with "PKGBUILD does not exist" — the actual point of failure was one step earlier and unrelated to AUR credentials. Rebuilt the same match string with a "$" variable and plain concatenation so the file's own text never contains the trigger sequence. Verified by extracting the exact heredoc and running it standalone against the real PKGBUILD template — renders identically to the intended output.	2026-08-27 14:29:27 -07:00
shadow-test	97e58db3c1	Close gateway-secret desync, TOCTOU, and undisclosed custom-image gaps ... Round 4 review findings: - Disclose and warn on a custom Docker image the import would set (HIGH): it's the image every project container is created from, so an undisclosed change here was a sharper version of the redirected-base-URL problem round 3 already flagged for the model backends. - Recreate a running gateway container when an import restores a new secret with the shape unchanged (MEDIUM): reconcile_gateway's shape comparison can't see a secret-only change, so the container would otherwise keep serving old key material indefinitely. - Report keychain write failures back to the caller instead of only logging them (MEDIUM): apply_settings_import now returns SettingsImportOutcome with secret_restore_warnings so a partial restore can't read as unqualified success. - Pin a hash of the previewed file's ciphertext and refuse to apply if it changed on disk (MEDIUM): closes a TOCTOU between preview and apply. - Sanitize and cap every free-form string a preview surfaces, and move the warning boxes above the replace list in the UI (MEDIUM): an unbounded base URL or image name could otherwise push the security warnings below the scroll fold. - Validate the Docker socket path on import the same as the SSH key and CA cert paths (LOW): it was the one mounted host path validate_settings_update didn't cover. - Fix ExportedSecrets::is_empty() to treat whitespace-only as blank, like every other secret-presence check in this feature (LOW). - Authenticate the file header as AEAD associated data (LOW, defense in depth) and correct two doc comments that overstated the password not being cached.	2026-08-27 14:24:06 -07:00
shadow-test	a606e3ab20	Validate settings imports before writing secrets; disclose base URLs ... A rejected import (bad env var name, disallowed host path) used to leave keychain secrets already overwritten while the settings themselves stayed unchanged. apply_settings_import now runs update_settings's validation (extracted into validate_settings_update) before any secret write. Also from this review round: sharpened two format-version tests that previously passed against the pre-fix code too, added a direct test for split_settings_and_secrets, warned on a dormant web terminal token even when the terminal import leaves it off, matched the password-length check to the frontend's unit of measure, zeroized the export plaintext buffer, and surfaced non-blank Ollama/llama.cpp/OpenAI-compatible/gateway base URLs in the import preview so a traffic redirect isn't silent.	2026-08-27 13:13:48 -07:00
shadow-testandClaude Sonnet 5	925e51e435	Fix a real credential-leak vector a review found, plus four smaller issues ... The headline finding: WebTerminalSettings::access_token is a live bearer credential for a server that binds every interface, stored as a plain field on AppSettings — which this feature was exporting and importing wholesale as if it were as inert as a port number. A crafted export file could set web_terminal.enabled and access_token together, and importing it (with no more warning than any other setting change) would silently stand up a LAN-listening terminal server with an attacker-known token on the victim's next launch. Fixed by carving the token out into ExportedSecrets, same as the other three global secrets, with the same "only overwrite what the import actually has" treatment — except that has to be done by hand here, since this one lives inside the AppSettings blob that gets replaced wholesale rather than in the keychain. Added SettingsImportPreview:: enables_web_terminal so "this turns on a listening service" gets its own visible warning in the confirmation modal rather than hiding inside a generic "settings replaced" bullet list. Also fixed: - read_and_decrypt checked format_version only after attempting to parse the full payload, so a future version bump that isn't deserialize-compatible would fail on the shape mismatch before the version check ever ran — and serde's type-mismatch errors quote the offending value inline, which is a real leak path since the plaintext here can hold a live credential. Now probes just the version field first, and neither error path interpolates the underlying serde message into what the user sees. - apply_settings_import cleared the pending-import path before it could fail, so a rejected import (an invalid host path, anything update_settings validates) dead-ended the modal with no way back except cancelling and reopening the file picker. The path is now only cleared on success. - Secrets are restored before the settings replace runs, not after — replacing settings is what triggers reconcile_gateway, and restoring secrets afterward left a real window where a gateway recreation happened against the destination's stale keys. - The 8-character password minimum was frontend-only; export_settings now enforces it too, since that's the actual boundary a weak password has to cross. The derived key and decrypted plaintext are wrapped in zeroize::Zeroizing (already in the tree via aes-gcm). Added test coverage the review named as missing: format-version ordering, the generic-error-message guarantee, non_blank's blank-vs- absent handling, and the new web-terminal preview/warning behavior on both sides of the IPC boundary. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 12:16:43 -07:00
shadow-testandClaude Sonnet 5	722d9aeff1	Add password-encrypted settings export/import ... Closes #35. Exports the host environment — global AppSettings (already the non-secret shape persisted to settings.json) plus the global secrets that live in the OS keychain instead (the shared Claude Code OAuth login, the model gateway's provider API key and master key) — to one password-encrypted file, and restores it on another machine. Per-project settings, per-project secrets, and Docker volumes are deliberately out of scope; this is not a project backup. Designed with the user in issue #35's comments: global settings only, no docker volumes, the password is the lock/key, and the export is portable as one file. Crypto (storage/settings_crypto.rs): Argon2id derives a 256-bit key from the password (memory-hard, meaningfully resistant to GPU/ASIC brute-forcing in a way PBKDF2 at any reasonable iteration count is not), AES-256-GCM does the actual encryption. A wrong password fails GCM's authentication tag rather than producing silent garbage. Salt and nonce are random per export and stored in the clear in the file header — their job is uniqueness, not secrecy. The save/open dialogs are opened from Rust, matching the boundary file_commands.rs's pick_save_path/pick_files_to_upload already establish: a frontend-driven dialog handing Rust a host path is the exact shape of bug that produced this app's past criticals. preview_settings_import resolves the chosen import path itself and remembers it (AppState::pending_settings_import) so apply_settings_import re-reads the same file without a path crossing back over IPC. The password is re-entered rather than cached between preview and apply, so nothing here holds decrypted plaintext in memory for longer than one command's execution; the preview returned to the frontend carries counts and presence flags only, never a secret value. Import replaces settings wholesale (an import is "restore this environment"), but only writes secrets actually present in the file — an absent secret means "the source machine never had this configured," not "delete this on import." Added storage::secure::store_gateway_master_key and get_gateway_master_key (read-only, unlike get_or_create_gateway_master_key which mints one as a side effect) since neither existed and import needs to restore an exact captured value rather than mint a new random one. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 11:57:16 -07:00
shadow-testandClaude Sonnet 5	b3d07bda09	Fix real workflow bugs a review found: dead bind mount, blind error gate ... A review found the "Validate with makepkg and namcap" step's bind mount (docker run -v "$PWD/rendered:/work") would very likely fail on Gitea's own act_runner: a containerized job's $PWD isn't a path the daemon's host can resolve, so the mount would silently attach an empty directory instead of failing loudly — the same class of problem noted elsewhere for this exact environment. Switched to docker create + docker cp (in and back out) + docker start -a, the pattern already validated locally, which works regardless of where the daemon actually lives. Also found and fixed, most severe first: - The namcap error gate (`grep -q "^[a-zA-Z0-9_-]*bin E:"`) only matched one of namcap's two line shapes for reporting an error ("triple-c-bin E: ...") and missed the other ("PKGBUILD (triple-c-bin) E: ...") entirely — confirmed by reproducing both against a real namcap run. The PKGBUILD-level half of the safety net was dead. Replaced with a plain `grep -q " E: "`, confirmed to match both real shapes (and a split-package variant) and nothing else. - package()'s `ar x "Triple-C_${pkgver}_amd64.deb"` named the asset literally, defeating the whole point of the resolve step discovering the real filename from the release instead of assuming a pattern — a future Tauri bundler naming change would still break here with an opaque error. Changed to `ar x ./*_amd64.deb`, which `source=()` already guarantees matches exactly one file. - `pacman -Sy` before installing packages is the canonical Arch partial- upgrade footgun; changed to `pacman -Syu --noconfirm --needed`. - `${{ inputs.version }}` was interpolated directly into a shell step instead of routed through `env:`, unlike every other step in the file. - `git push origin master` assumes the local branch name after cloning a brand-new (not-yet-created) AUR repo's empty state is `master`, which depends on the runner's own `init.defaultBranch` if the server sends no symref. `git push origin HEAD:master` is unambiguous either way. - The private key was written with a plain redirect then chmod'd after, leaving a window where it's world-readable; now created at its final mode first via `install -m 600 /dev/null`. Added `-o IdentitiesOnly=yes` so a runner ssh-agent can't offer a different key first. - Added GH_PAT auth to the api.github.com calls, matching every other workflow in this repo, to avoid the unauthenticated 60/hour rate limit. - Fixed two comments: the `options` comment credited `!debug` for suppressing the empty debug-package directory, when it's actually `!strip` doing that (verified in a real build); and documented in the README that a hand-edit made directly in the AUR repo is silently reverted by the next dispatch, since every run renders fresh from this repo's template. All of the above re-verified with the same real end-to-end methodology as the original commit: real makepkg build, real namcap lint (clean), and the exact updated docker create/cp/start sequence run against a live container. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 11:24:31 -07:00
shadow-testandClaude Sonnet 5	e025a7441a	Add a native Arch/CachyOS package via its own AUR publish workflow ... Part of triple-c#34's third ask ("I would like to also have an Arch/CachyOS native version as well"), addressed separately from the Wayland crash fix (fix/wayland-webkit-egl-crash) since it's an unrelated feature, not a bug. packaging/arch/PKGBUILD is a "-bin" AUR package repackaging the same .deb build-app.yml already produces — no Rust/Node toolchain needed to install it, and the user gets exactly the binary the project ships and tests. Verified end to end against a real release (v0.4.14) rather than going by Tauri's generic docs: downloaded the actual .deb, ldd'd the actual binary to ground-truth `depends` (dropped `pango` and `libayatana-appindicator` from an earlier draft — the first is already pulled in transitively by gtk3, the second was never linked at all since this app has no tray icon or menu), and ran a real makepkg/namcap/pacman -U cycle. namcap caught a real issue this way (missing license file under /usr/share/licenses/triple-c-bin/), now fixed by fetching LICENSE alongside the .deb. .gitea/workflows/publish-aur-package.yml does the actual publishing: given a version (or "latest"), it finds that release's real Linux asset on GitHub, downloads it, computes real checksums, renders the PKGBUILD template, validates the result with makepkg and namcap inside a real Arch container, and pushes to AUR. workflow_dispatch only, deliberately — the same reasoning that killed sync-release.yml in triple-c#32 (releases are assembled by build-app.yml across three separate platform jobs, so there's no single automatic event that fires only once the Linux .deb this needs actually exists) applies here too. Requires a repo secret this workflow cannot set up itself: AUR_SSH_PRIVATE_KEY, from an AUR account that has already created (or been given co-maintainer access to) triple-c-bin — both one-time manual steps on aur.archlinux.org. Until that secret exists, the workflow fails loudly at the push step rather than silently doing nothing. See packaging/arch/README.md for the full maintenance flow. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 11:10:55 -07:00
shadow-testandClaude Sonnet 5	8f62949902	Correct two overclaims in the Wayland workaround's comment ... Review found: "nothing this app's UI depends on" is backwards — the terminal's @xterm/addon-webgl renderer is exactly the GPU compositing path this setting disables, it just degrades gracefully (the addon's own construction already handles WebGL being unavailable) rather than crashing. And the "not simply Wayland vs X11" justification for going unconditional doesn't hold up: WAYLAND_DISPLAY is exported into an XWayland client's environment too, so gating on it would have caught that case as well — the real reason to go unconditional is that there's no reliable heuristic for the thing that actually matters (which Mesa/driver/compositor combination is affected), not that the naive gate misses XWayland specifically. Also noted, not changed: the env var leaks to whatever the app spawns afterwards (a cold-launched default browser via xdg-open), and the "=0 re-enables it" parenthetical isn't verified against WebKitGTK's own source, so softened to say what's actually guaranteed (an already-set value is left alone) rather than assume presence-vs-boolean parsing. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 10:57:45 -07:00
shadow-testandClaude Sonnet 5	6354cb42b2	Work around WebKitGTK's EGL crash on Wayland (triple-c#34) ... Reported on CachyOS/Arch with Wayland: the app aborts immediately with "Could not create default EGL display: EGL_BAD_PARAMETER. Aborting." printed straight to stderr by WebKitGTK's own C code, before Triple-C's own logging even gets a chance to say anything useful about it. This is WebKitGTK's DMA-BUF renderer (its default accelerated-compositing path since 2.42) failing on some Mesa/driver/compositor combinations. Set WEBKIT_DISABLE_DMABUF_RENDERER=1 unconditionally on Linux before the Tauri builder runs, which is where GTK/WebKitGTK actually read it — there's no reliable way to detect the affected combination ahead of time (reports of this exact failure exist under XWayland too, not just pure Wayland sessions), and WebKitGTK's fallback compositing path costs some rendering performance this app's UI doesn't need. Left alone if a user has already set the variable themselves. Does not address the other two things filed under the same issue (links not opening on the host, and a request for a native Arch/CachyOS package) — those need more information / are a separate scope, respectively. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 10:45:24 -07:00
shadow-testandClaude Sonnet 5	049232099b	Dedupe the preview-build predicate, fix two comment inaccuracies ... Final review pass gave this a clean bill of health overall but named three small things: - get_app_version() and check_for_updates() each read option_env!("TRIPLE_C_BUILD_SUFFIX") independently with slightly different idioms — if one were ever edited alone, the About panel and the update check could silently disagree about whether this is a preview build. Extracted preview_build_suffix() as the single place that reads and classifies it. - pick_update's doc comment described the unparseable-tag case as a `-preview.<sha>` suffix; the actual tag build-app-preview.yml creates is `preview-<sha>` (no version, no dot) — already correct in the neighboring GitHubRelease::prerelease comment, just not here. - That same prerelease comment claimed defence against a preview release leaking through backfill-releases.yml, but a preview's tag already fails semver parsing on its own — this field's actual job is the case parsing can't catch: a normally-tagged release someone flags prerelease on Gitea (a hotfix candidate, an RC) that a backfill would otherwise mirror as-is. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 10:34:56 -07:00
shadow-testandClaude Sonnet 5	945883bb9d	Actually offer a preview the release it precedes, and fix two more gaps ... An Opus review of the previous commit found its headline claim didn't hold: a preview and the release it precedes compute to the identical numeric version by construction, but check_for_updates compared with a strict `>` against the bare CARGO_PKG_VERSION (never the suffixed display string), so `(0,4,13) > (0,4,13)` is false and the release was never offered. Plain semver ordering doesn't make a `-preview.<sha>` suffix sort below the same numeric release on its own here, since the comparison never sees the suffix at all. pick_update now takes is_preview_build, derived from whether TRIPLE_C_BUILD_SUFFIX was baked in, and relaxes that one comparison to `>=` — so "a release exists at my own number" reads as an update. A production build still requires strictly newer. Also: ported build-app.yml's `git tag --points-at HEAD` guard into the preview version computation. Without it, workflow_dispatch (which this workflow allows on main, not just PR builds) run on a commit a release was already cut from would compute one past that release — reintroducing "preview outranks production" through the manual-dispatch door. And corrected two comments that claimed the prerelease filter was currently a no-op: backfill-releases.yml mirrors every Gitea release to GitHub unfiltered, prerelease flag included, so it's real defence-in-depth against a dispatched backfill leaking a preview release, not a no-op. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 10:24:24 -07:00
shadow-testandClaude Sonnet 5	b71e15c2c0	Make preview versions monotonic and distinguishable from production ... build-app-preview.yml computed its patch number as `git rev-list --count <latest tag>..HEAD` — the exact formula build-app.yml itself documents as broken and replaced (#26): a distance from whichever tag sorts highest, not a counter, so it resets to zero on every release and previews went backwards (0.4.62 -> 0.4.0) the moment one landed. Ported the same "one past the highest patch already used" computation build-app.yml uses for real releases, reading the same tags (including -mac/-win suffixes), so a preview built right before a release now computes the exact number that release is about to take — semver already orders `0.4.12-preview.<sha> < 0.4.12`, so a preview user is offered the release the moment it ships instead of being silently pinned forever. The installed preview's reported version was also indistinguishable from production: the bundle's own version field strips the `-preview.<sha>` suffix before touching tauri.conf.json/Cargo.toml/package.json, since the Windows MSI's ProductVersion has no room for one. Rather than risk that (unverifiable without an actual Windows build), preview builds now bake the suffix into the binary separately via a TRIPLE_C_BUILD_SUFFIX build-time env var, and get_app_version() appends it when present — a production build sets nothing, so this is a no-op there. Also: added `prerelease` to `GitHubRelease` and filter on it in check_for_updates (currently a no-op against real data — nothing mirrored to GitHub is ever prerelease:true — but the updater is no longer structurally incapable of enforcing a channel split if one is ever made explicit). And deleted sync-release.yml: workflow_dispatch-only, reading gitea.event.release.* fields a manual dispatch never populates, so it could never have actually run; build-app.yml's inline mirror already does the same job. Refactored check_for_updates' filtering into a pure, testable pick_update helper (this file had no tests before), and added tests for it and the new get_app_version suffix handling. Fixes #32. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 10:11:35 -07:00
shadow-testandClaude Sonnet 5	61bdbc4a5b	Close the crash-window gap and exec-session leak a third review found ... A third Opus review pass confirmed round 2's fixes hold up, then found: - The pending-cleanup record `remove_project` writes is fully durable (fsync'd); the projects_store.remove() that follows it is a plain fs::write with no fsync. A crash or power loss in that window — or that store write failing outright, beyond what the previous round's in-process rollback catches — leaves a record on disk naming a project projects.json still lists as present. The very next startup retry would then delete that project's container, snapshot image, and both volumes (including the one holding the OAuth credential and every session transcript) out from under a project the user still sees in the sidebar. retry_pending_cleanup_logged now takes the ProjectsStore and refuses to touch — clearing instead — any record whose project id still exists. Also stopped swallowing the round-2 rollback's own failure. - Resolving the container through find_existing_container instead of project.container_id (round 2's stale-id fix) changed what drove close_sessions_for_container in remove_project and rebuild_project_ container: sessions are now leaked when Docker is unreachable (nothing resolves, so nothing closes, and the project record is gone a moment later) and in the stale-id race itself (sessions were opened against the container that actually exists, not the id find_existing_container bypasses). Both functions now close sessions for the stored id unconditionally, and again for the resolved id if it differs. - A pronoun-agreement bug in the no-retry removal toast ("remove them manually" for a single leftover) that was fixed one line above for verb agreement but not for the pronoun. Also closed the test gaps the review named: the pending-cleanup corrupt-record aside-move had no test, the Reset toast's leftover copy was inline and untested (extracted to lib/resetOutcome.ts, mirroring components/projects/home/removalReport.ts, with unit tests), and nothing asserted rebuild()'s success path maps outcome.project into the list rather than the whole outcome. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 09:47:45 -07:00
shadow-testandClaude Sonnet 5	439ef16f07	Fix two new bugs a second review found: stale container id, orphaned record ... A second Opus review of commit 2 found it had introduced real problems of its own rather than just polish gaps: - remove_project's "None or stale" container-id fallback only handled None. A stale id (the documented start-failure race in start_project_container_locked, where the old container is removed and the new one's id isn't persisted until after start_container succeeds) still 404'd on removal — now treated as success by commit 1's own fix — while the real container survived to block every volume removal with a 409 forever, with nothing in the pending-cleanup record ever naming it. Both remove_project and rebuild_project_container now resolve the container via find_existing_container() unconditionally, matching every other container-destroying path in the codebase, and remove_project fails closed (records a leftover rather than silently skipping) if Docker itself can't be reached to check. - remove_project could leave a pending-cleanup record for a project still live in projects.json: if the store's own save failed after the record was written, startup housekeeping would delete that project's container and volumes out from under it on the next launch. The record is now rolled back when the store write fails. - rebuild_project_container (Reset) only surfaced a leftover volume, not a leftover snapshot image — the more serious failure, since the next container is built from that image whenever it exists, silently reviving the exact system layer Reset was asked to discard. ProjectResetOutcome now carries leftover_image too, and the toast's "run docker volume rm" advice is corrected: the new container has already remounted the volume by the time the toast renders, so that command would just hit the same conflict Reset did. Also from the same pass: reworded a couple of log/toast lines that still asserted resources were "still present" when the daemon-unreachable case covered by the same code path can't actually confirm that; fixed a singular/verb mismatch in the leftover toast text; moved an unparseable pending-cleanup record aside instead of re-warning about it forever; and added a debug log when a record's recorded_at can't be parsed, so aging never silently no-ops. Pulled describeLeftovers/leftoverVerb out of ProjectHome.tsx into their own module with unit tests, and added tests for the recorded_at staleness check — the previous commit's equivalent logic had none. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 09:09:46 -07:00
shadow-testandClaude Sonnet 5	d8bb5ab262	Address review findings: durability, stale container ids, honest toasts ... An Opus review of the previous commit found several real gaps: - pending_cleanup::save used plain write-temp-then-rename, unlike migration_store's fsync'd write it claimed to mirror — a crash in that window left a truncated record that list() would skip forever, silently reproducing the exact bug this module exists to fix. Now matches migration_store's File::create/write_all/sync_all/rename/sync_dir shape, and the tests exercise the real save/list/clear functions against a temp dir instead of re-implementing their bodies inline. - remove_project and rebuild_project_container only ever looked at project.container_id, unlike every other container-destroying path in the codebase, which falls back to find_existing_container for exactly this race (a crash between creating a container and persisting its id). A miss here left a container that then blocked every subsequent volume removal with a 409, forever. Both now resolve the same way the rest of the codebase does, and record the container by its deterministic name rather than its id so a retry still has something that resolves. - remove_project's toast promised an automatic retry unconditionally, even when writing the pending-cleanup record itself failed (the one case where nothing will actually retry). ProjectRemovalReport now carries retry_scheduled, and the UI is honest about which case it's in. - remove_volumes_by_name now retries once after a short delay on a 409, since Docker releasing a volume's mount reference right after its container is removed is not always instantaneous, and this is exactly the sequence remove_project runs. - rebuild_project_container (Reset) returns ProjectResetOutcome so the UI can warn when Reset could not fully clear a project's volumes, instead of only logging it — the new container silently reuses old data otherwise, which is what Reset promises not to do. - retry_pending_cleanup_logged escalates a record's log level after it has failed for a week, since recorded_at was otherwise write-only. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 08:36:27 -07:00
shadow-testandClaude Sonnet 5	4827170715	Report and retry Docker resources remove_project could not delete ... remove_project_volumes always returned Ok(()) regardless of what actually happened, making the `if let Err(e)` guarding it at every call site dead code. remove_project then dropped the project record unconditionally, so a volume, image or container that failed to delete became permanently unreachable — confirmed against a real orphaned volume pair found in the wild (fixes #31). remove_project_volumes/remove_snapshot_image/remove_container now report what they could not remove (treating "already gone" as success rather than a leftover), remove_project surfaces this to the user via a toast, and before dropping the project record it writes a pending-cleanup record that startup housekeeping retries automatically on the next launch. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FGjXq6fqtAFHdbhk4f3PfZ	2026-08-27 08:18:41 -07:00
shadow-testandClaude Opus 5	68b73a9102	Refuse a commit that adds something shaped like a credential ... The companion to the fixture removal. A site-admin token sat in a test file for 92 commits and fourteen days on a public mirror, past five audit rounds and two independent reviews, because all of them read the code under change and this was not under change. A grep would have caught it the first day. `scripts/scan-secrets.sh` is that grep, in three rules: * vendor-prefixed credentials — `ghp_`, `github_pat_`, `glpat-`, `xox*-`, `sk-`, `AKIA`/`ASIA`, `ya29.`, `AIza`, `npm_`, `dckr_pat_`. Shape alone identifies these, so there is no context to get wrong. * `BEGIN … PRIVATE KEY` blocks. * an opaque literal assigned to a secret-shaped name — the rule that would have caught this one. The third rule needs **both** halves, and that is what makes it usable rather than another disabled check. Measured before writing it: an entropy-only rule flags 317 literals in this tree, and name-proximity alone flags four, three of which are `secure::get_project_secret(&id, "aws-secret-access-key")` — a keychain *key name* sitting next to the word `secret`. Requiring the literal itself to be hex or base64 with no word structure is what excludes those. Validated rather than asserted: * **0 false positives** across every tracked file. * **Catches the real incident** — `--range 9b2f4fe~1..9b2f4fe` is refused. * Twelve shaped cases pass and fail as intended, including a sha256 in an `assert_eq!`, a git sha in a comment and the new dummy fixture, none of which trip it. * The hook was proved to block an actual `git commit`, not just to exist. Two halves, because each covers the other's gap: * `.githooks/pre-commit`, enabled per clone by `npm run hooks`. Git will not let a repository set its own hooks path — cloning would then be enough to run its code — so this is opt-in everywhere and `--no-verify` skips it. * `Secret Scan`, which nobody can bypass. It carries **no `paths:` filter** on purpose: the leak lived in `app/**` and `build.yml` only runs for `container/**`, so a path-filtered scan would have missed the very thing it exists for. It scans the whole tracked tree rather than a range, because a wrong range fails *open* and the full pass takes 0.5s. Also fixed while here: `core.hooksPath` in this clone pointed at `/workspace/.git/hooks`, a directory that does not exist — so git hooks were disabled outright and anything dropped in `.git/hooks` would have been ignored in silence. A hook that never runs is worse than no hook, because the checklist says it is there. `--tracked` skips binaries. Feeding a blob to grep gets "binary file matches" instead of the line, so a genuine finding inside one would arrive as a sentence nobody can act on. A line ending `pragma: allowlist secret` is skipped — wordy on purpose, so it reads as a claim and leaves something greppable. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LHL9ty7arp8FHwvE77ne7y	2026-08-25 11:51:04 -07:00
shadow-testandClaude Opus 5	d09e2a2743	Stop using a live credential as a test fixture ... `the_custom_env_fingerprint_never_carries_the_value` asserted that the custom-env fingerprint does not leak a secret — using the maintainer's real Gitea token as the secret. It was committed on 2026-08-11 in 9b2f4fe, reached 92 commits, and was readable for fourteen days in the public GitHub mirror at `shadowdao/triple-c`, confirmed by fetching the raw file. The token was a **site-admin** token (`is_admin: true`, user id 1) with admin rights on every repository the account can see, not a repo-scoped one. It has been revoked; the API now answers 401. Release bundles were never affected — the literal is inside `#[cfg(test)]`, and a search of the shipped 0.4.62 AppImage finds nothing. The fixture is now an obviously fake string, and the test is unchanged otherwise. Mutation-checked against the replacement: a fingerprint that returns the raw value, one that returns the key name, and one that ignores its input are all still caught, so nothing about the test's power depended on the value being real — which was true the whole time. History is deliberately **not** rewritten. The value was public for two weeks, so rotation is the fix and the old value is now worthless; rewriting 92 commits of published history would break every clone to hide something already seen. Worth noting how this survived: five audit rounds and two independent reviews all pointed at new code, and this sat in a test file that none of them had reason to open. A high-entropy-literal check in CI would have caught it on the day. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LHL9ty7arp8FHwvE77ne7y	2026-08-25 11:29:16 -07:00
shadow-testandClaude Opus 5	eead748222	Close what two reviews found in the Files tab transfers ... Two independent reviews of 2c9482a, one for correctness and one against the threat model. Between them they found two ways to lose a file, one way for a container to choose a Windows save destination, and a rule that made every dotfile unsavable. Every finding below was demonstrated against a real container before being fixed, and the fixes are demonstrated the same way. ## The download was accepting truncated reads and refusing whole ones The `[ -f ]` bracket around the read was wrong in both directions. It missed the case that loses data. Truncation *in place* — `> file`, log rotation, `tar -x`, most build tools — leaves a regular file behind, so `dd` stopped at the new EOF and exited 0. Measured: a 600 MB source truncated mid-read delivered 34 MB, which was then renamed over the user's own earlier copy and toasted as `Saved (34.1 MB)`. Truncate-to-zero did the same and needs no adversary at all. And it failed *good* downloads. `dd` already holds the fd, and neither `rm` nor `mv` can touch an open one — the bytes are complete. But `rm` makes `[ -f ]` false, so a finished 600 MB transfer of a file a bundler happened to unlink was deleted, reporting "nothing was saved". The comment claiming the bracket caught a "mix of two files" was simply wrong; an open fd cannot be a mix. So the script now measures the file before reading it and puts the answer on stderr, and Rust checks that at least that many bytes arrived. One rule, subsuming everything the bracket was for. Verified against a real container: truncation mid-read and the FIFO race are refused; deletion, rename-over, a growing file, an empty file and an untouched 600 MB read all pass. ## On Windows the container, not the user, was naming the save destination `suggested_save_name` split on `/` only, and it feeds the save dialog's pre-filled name. Backslash is a legal Linux filename character and `validate_container_path` has no reason to object — `..\..\Users\…` is one POSIX segment. The Windows common file dialog parses its name box as a path on Save, so a container-created file called `..\..\..\Users\vic\AppData\Roaming\Microsoft\Word\STARTUP\x.dotm` put its own bytes in an auto-loading Office directory on one un-read click. `resolve_host_path` did not stop it: Word's `STARTUP` and Excel's `XLSTART` are not in the autorun denylist, which this file's own docs already concede is "losing by construction" and was never meant to be the boundary here. The name is now sanitized of every separator, the drive colon and the rest of what NTFS refuses, so it cannot be a path on any platform this ships to. ## No dotfile could be saved, and the app pre-filled the name that guaranteed it The write policy judged the leaf for hiddenness, so `/workspace/.env` was refused *after* the modal and the overwrite prompt — quoting the name the app itself had suggested. `.gitignore`, `.dockerignore`, `.eslintrc.json`, `.nvmrc`: all unsavable, while uploading them worked, so a dotfile could go in and never come out. `HostPathUse` gains a third mode. `WriteChosenName` drops the leaf check and keeps every directory rule, and only `download_container_file` uses it — the dialog is a real boundary for that caller and only that caller. `download_container_backup` still takes its path over IPC as a string and keeps the strict rule. ## Smaller, all found by the reviews * A download had no ceiling. `dd` resolves through the container's `PATH`, which its agent owns with passwordless sudo; a replacement writing forever was measured at ~6 GB/s, so one click on a file listed as 2 KB filled the host disk with no progress shown and no cancel. Bounded now by what the file measured, with slack that is absolute for small files and proportional for large ones, so an honest growing log is unaffected. * "Framed rather than verbatim" did not stop container text reaching the toast headline: `readableRefusal` matches with `includes`, and it has to, because the app's own refusals carry those markers mid-sentence. Anchoring would break them. The fix is at the injection point — container text is clipped to one 200-character line with control characters stripped — plus a `max-h-40` on the toast message, which the `detail` block always had and this half did not. 8 KB of prose in a `z-[60]` card pushed its own dismiss button off-screen. * `savingPath` was a scalar while the design deliberately allows concurrent saves. Starting a second freed the first's row mid-transfer, and whichever finished first cleared both; dismissing the second dialog was enough. It is a `Set` now. * `setUploading(false)` fired when the command settled, not when the refresh finished, so a second click landed mid-relisting. * `upload_files_to_container` checked the container directory before the picker and then used the unresolved path. A modal has no time limit. It re-checks after, which is what `docker::exec`'s doc comment already claimed. * `wait_for_exec_exit` flattened a missing exit code to `Some(0)`, which made the new `!= Some(0)` check unreachable by construction. * `normalize_host_path` did not collapse repeated separators, so the lexical system-root rule was silently absent for `C:\\Windows\…`. Not exploitable — the resolved pass catches it — but a documented layer that does nothing is a trap for the next caller. * README still carried the "no host path crosses IPC in either direction" claim the previous commit narrowed everywhere else, and HOW-TO-USE described the hidden rule without saying it applies to folders only. 480 Rust tests, 602 frontend, no new clippy warnings. Eleven mutations against the new tests, all killed — two of the first round survived and were rewritten: one because `split_whitespace` already handled the case I thought I was testing, one because I had deleted a comment rather than the behaviour. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LHL9ty7arp8FHwvE77ne7y	2026-08-25 10:42:26 -07:00
shadow-testandClaude Opus 5	2c9482a67d	Give the Files tab back its uploads and downloads ... `upload_file_to_container` and `download_container_file` existed on main before any of this work started. "Ship the Files tab container-side only" removed them and called it narrowing scope; from a user's side it was a regression they upgraded into. This restores the feature. The reason for the removal was real — four consecutive audits found their criticals in host paths crossing IPC — so the feature comes back only in the shape that removes the class rather than patching it a fifth time. The dialogs are opened by **Rust** (`pick_save_path`, `pick_files_to_upload`), not by the webview. A frontend `open()`/`save()` handing the backend a path string is exactly what failed, and the backend cannot tell such a string from one a compromised webview invented. Now the webview can ask for a picker and that is the whole of its influence: it cannot name a host path as an input. That is the shape the previous round's own notes named as the honest one if this ever returned. None of the machinery the audits condemned returns. No `link(2)` destination reservation, no placeholder rollback, no collision marker: the OS save dialog already asks about overwriting and Docker's extractor overwrites on upload the way `cp` does, so there was nothing left for it to do. Download reuses the sequence `download_container_backup` has been using unchanged — resolve, stream into a partial file beside the destination, rename last — so a failed transfer never touches the file that was already there. Upload reuses the terminal drop's hardened uploader, with the container's uid/gid resolved once per selection rather than once per file. Against a container that is actively hostile rather than merely surprising: * the read is `dd iflag=nonblock`, not `cat`. `[ -f ]` and the `open` after it are two syscalls and the container owns the filesystem in between; a loop swapping the file for a FIFO wins that race, and `cat` then blocks forever with no writer and no timeout anywhere on the path — the `invoke` never settles and a partial is left in the user's directory for good. Verified in a real container that `cat` hangs, that `iflag=nonblock` returns, and that it is byte-identical on a regular file. * the read is bracketed by a second `[ -f ]`, because non-blocking turns that hang into an empty file that would otherwise be renamed over the destination and reported as a successful save. * an *undeterminable* exit code is a failure. Backup catches this class with its `total == 0` check, which download cannot have because an empty file is a legitimate save; without a replacement, a project restarted mid-download renames a truncated partial over the user's file and reports the byte count as if it were whole. * container stderr is capped. Every other reader of container output in the tree is capped for this reason; the two streaming commands were the exception, and stdout was bounded by disk while stderr was bounded by nothing. * the script's refusals are framed rather than used verbatim, so a directory named to look like one of our own sentences cannot become the toast headline through `readableRefusal`. * the partial name is capped at NAME_MAX. A bundler's 230-character content hash is a name that fits its directory and produces a partial name that does not. Also: a non-UTF-8 dialog path is refused by name rather than silently mangled into a different path by U+FFFD substitution; both actions carry in-flight state, so a second click cannot open a second dialog and a slow save is not indistinguishable from a dead button; and the upload's completion message names the directory, since the picker is modal and the user can browse elsewhere while it is open. Not restored: drag-and-drop, in either direction. `drag:allow-start-drag` stays ungranted and `hold/disk-and-dragout` still holds that work. Two bugs the new tests caught while being written: a double-click on "Save to host…" opened the file viewer on top of the save dialog, and an N-file upload made N redundant execs to re-ask `id -u`. Docs that asserted this feature did not and must not exist are corrected — CLAUDE.md, README, HOW-TO-USE, TECHNICAL and the capability threat model. The "no host path crosses IPC" claim is deliberately narrowed to the inbound direction: paths do still travel outward inside error text, canonical ones included, and the reviewed record should not overstate. 600 frontend tests, 473 Rust, no new clippy warnings. Every new test was mutation-checked; four that survived their first mutation were rewritten, including two whose mutations turned out to be unfaithful and one that was blind to a dismissal leaving a row stuck on "Saving…". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LHL9ty7arp8FHwvE77ne7y	2026-08-25 10:00:21 -07:00
shadow-testandClaude Opus 5	88ffb4744a	Cancel the keydown on Shift+Enter, or xterm submits anyway ... The handler returned `false` from xterm's custom key handler and a comment claimed that was enough to stop the bare CR. It is not. `_keyDown` returns the instant the handler says `false` — before it sets `_keyDownHandled` and before it cancels the event — and `_keyPress` then checks that same flag, finds it false, and emits a bare CR for Enter's charCode 13. So the headline feature of this branch did the wrong thing in a real browser: Shift+Enter inserted the newline and then submitted the half-written prompt, now with a stray blank line in it. Arguably worse than before the fix. Reproduced in Chromium and confirmed against the bundled xterm 5.5.0 source. `preventDefault()` is what stops the browser firing keypress at all. Applied at both call sites — the desktop terminal and the web terminal's copy. The test could not have caught this. jsdom never synthesizes the follow-up keypress, so `expect(sent()).not.toContain("\r")` was asserting a property the environment cannot falsify — a test named for a behaviour it could not exercise. It now asserts `defaultPrevented`, which is the mechanism that actually suppresses the keypress and which jsdom can observe. Mutation-checked: removing the `preventDefault()` fails it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 20:47:01 -07:00
shadow-testandClaude Opus 5	016de8f641	Close the blockers from the fifth audit ... Docs and disclosure. HOW-TO-USE.md's settings table still described the pre-fix behaviour — and help_commands.rs fetches that file from GitHub main at runtime, ahead of the embedded copy, so it would have reached every user's Help dialog the moment this merged. The Config tab named three settings that need a base-image update; there are four, and the omitted one (Session recap) is the one that fails *without* the "won't switch off" symptom the warning teaches. Both now also state the cost nobody had written down: changing any of these recreates the container, which commits a layer. Two stale comments that told a reviewer the code was safe when it was not. compute_claude_code_settings_fingerprint still claimed the historical fingerprint is preserved so an upgrade cannot churn every container — carried over from before the widening, false since the format string changed. And capabilities/default.json, which is the reviewed threat model of record, described a "Save to host…" action this branch deletes. Security and correctness. update_settings validated env vars and nothing else, so the *global* default_ssh_key_path — the fallback for every project without an override — took `/` and read-only bind-mounted the host, which entrypoint.sh then copies into the home volume. classify_ mount_source ran canonicalize on the raw string, which resolves a relative path against Triple-C's own cwd, so `.` and `..` were accepted or refused depending on where the app was launched; the daemon then refuses the mount and the project can never start. Its test passed only because its examples did not exist under app/src-tauri. bind_mount_exclusions still derived a path from every row while project_path_mounts had learned to skip unmountable ones, so a legacy row made /workspace/<name> ordinary container content that a migration would then exclude from staging and destroy. The skip is also logged now rather than silently dropping a folder. The terminal's file-in path checked is_dir() but not file type, so a dropped FIFO blocked forever with no timeout — and it is the only route in now. The web terminal labelled sessions from a global set at request time, so two quick opens swapped them; harmless until Shift+Enter became type-dependent, at which point a mislabelled Claude session submitted a half-written prompt. Opened now carries the type. Every ~/.claude.json write goes through one atomic helper. The awsAuthRefresh branches still truncated in place — the same corruption the Shift+Enter block was fixed for twenty lines later, and its own comment said so. Demonstrated: a failed write now leaves the original byte-identical. And the registration test I added yesterday could pass while the property was false: an audit got five real unregistered commands past its exact-string attribute match, and "exactly once" was in its name but not its body. Mutation-checked against all six shapes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 18:45:11 -07:00
shadow-testandClaude Opus 5	4d1a5a2417	Remove list_sibling_containers, which nothing called ... It listed every container on the daemon — including the user's unrelated postgres, mysql and other work — and handed the summaries to the webview. It had a registration, a command, a docker-layer helper, a typed frontend wrapper and a `SiblingContainer` type, and zero call sites. An audit named it as step one of an escalation chain: enumerate the daemon's containers, then point `update_project`'s unvalidated `container_id` at one and read its files through the file-command surface. The second half of that chain is closed now, but a command that exposes the user's unrelated containers and serves no feature is surface with no upside. Found by the registration test added in the previous commit, which is the answer to "why test something the compiler already checks": the compiler is perfectly happy with a command nobody calls. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 17:15:38 -07:00
shadow-test	a323047964	Merge branch 'r4/scrub' into ship/core	2026-08-23 17:12:34 -07:00
shadow-testandClaude Opus 5	11216c45e3	Assert every command is registered, and every registration exists ... This is the shape of the bug behind the original OAuth-callback report: `set_auth_bridge_enabled` existed, worked, and had a typed frontend wrapper — with zero call sites. The switch the docs told users to flip was wired to nothing, so every login callback was refused. Both halves compiled, so nothing noticed. The reverse direction is the sharper one: a command that is registered but reachable from nowhere is still IPC surface a compromised webview can call. `list_sibling_containers`, which returns every container on the daemon including the user's unrelated work, sits in exactly that state. Mutation-checked both ways: removing a registration fails the test, restoring it passes. The first parser I wrote split the list on commas, which glued each `// Docker` style comment to the command after it and then dropped that command as a comment — silently, once per group, 17 in total. Line-based now. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 17:12:34 -07:00
shadow-testandClaude Opus 5	913aa85805	Stop the pre-commit scrub running with its guards silently absent ... Two fail-open gaps, both latent on the shipped image and both live on an ImageSource::Custom one. `--one-file-system` is the only bound on the scrub's `rm -rf`, and it was probed into `$rmopt` with `rm --one-file-system --help`. BusyBox answers that with `unrecognized option` and exits 1, so on any busybox- or toybox-derived image the variable was empty and the delete ran as root across mount boundaries — reported as a completed `Reclaimed(n)`. Measured in `busybox:latest` with a named volume one level below the match at /tmp/claude-x/inner: emptied, `###TRIPLE-C-SCRUBBED 102423`. It is a prerequisite now, probed by `rm --one-file-system -f -- ''` — the code path that matters, with the one operand no filesystem can name — and an image without it prints `###TRIPLE-C-SCRUB-UNAVAILABLE` and deletes nothing. That costs Alpine and busybox images their scrub, which is the cheaper half of the trade: declining costs disk, proceeding costs the mount. The second is that resetting `PATH` was never the whole of command lookup. `bash` builds functions out of `BASH_FUNC_<name>%%` environment variables, function lookup precedes `PATH` entirely, and `command -v` reports a function as found — so a planted `stat` passed the prerequisite probe and then answered the containment checks. Every in-shell answer is itself importable: `unset`, `command`, and — the point that settles it — `[`, `pwd` and `cd`, which are checks 0 to 2 rather than merely the tools. So the script is no longer run by the shell that read the environment. The exec's argv is a bootstrap using only reserved words, parameter expansions and command words containing a `/` (which `bash` refuses to import a function for), and it hands the script as `$1` to a second `/bin/sh` started by `env -i`. Measured on `/bin/sh -> bash` with `BASH_FUNC_stat%%` set on the container and a volume mounted at the match: the old invocation emptied it and printed 306565, the bootstrap left it intact and printed 65536. `/usr/bin/env` then `/bin/env`, because H3's lesson about hardcoded coreutils locations applies to `env` too. The comment claiming the `PATH` reset "defeats it just as completely as spelling /usr/bin/stat out" is corrected, as is the one claiming a sudo-written /usr/bin/stat does not survive a container restart — it lands in the writable layer, which is exactly what the commit this runs in front of captures. Also here: a stored project path row with an empty host_path or mount_name no longer becomes a mount. The first sends `field Source must not be empty` back for the whole create, so the project cannot start at all until the row is gone, and no amount of save-time validation reaches a record already on disk; the second mounts over /workspace itself and the daemon then creates the other rows' mount points inside the user's real folder. And in migration_commands, the deferred-reconcile claim is RAII rather than a trailing statement — a panic inside `reconcile_migration_now` stranded it for the rest of the process — and `await_release` looks before it sleeps, so a project released a moment later no longer costs a full twenty seconds. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 17:09:47 -07:00
shadow-test	e9902f0564	Merge branch 'r4/narrow' into ship/core	2026-08-23 17:08:04 -07:00
shadow-test	7488fc5b70	Merge branch 'r4/host' into ship/core	2026-08-23 17:08:04 -07:00
shadow-testandClaude Opus 5	06ccb4d818	Ship the Files tab container-side only ... Four successive audits found the same thing: host filesystem paths crossing IPC is where the criticals in this work live. The most recent one found the `link(2)` upload reservation returning success against a *directory* (linking into it, leaving permanent stray files, and via a symlink-to-directory writing outside the validated write root), failing every upload permanently on any filesystem without hard links, and the post-resolution credential check weakened from a general rule to an eleven-name denylist. Rather than fix that a fifth time, the Files tab ships as what it is good at: a browser, viewer and renamer that never touches the host. Removed: `upload_file_to_container`, `download_container_file`, and everything that existed only for them — the whole reservation (`UPLOAD_RESERVATION_SCRIPT`, `reserve_upload_destination`, the placeholder rollback, `exec_oneshot_as_within` which had no other caller), `stream_container_file_to_host`, `ChannelReader`, `save_to_host`, the download ceiling, and the collision marker with its frontend contract. On the frontend: the upload button, the pane's `onDragDropEvent` handler, both "Save to host…" affordances, `uploadPaths` / `downloadFile` / the overwrite prompt, and `OverwriteConfirmModal`. `lib/uploadErrors.ts` is now `lib/refusalText.ts` and keeps only the half that turns any backend refusal into the sentence a person reads. Kept, and not weakened: `upload_host_file_to_terminal` and `download_container_backup`. They predate this work, their hardening is a real improvement over main, and they are now the whole answer to "how do I get a file in or out" — drop it on the Terminal, or Back up container. The drop gate (`lib/dropTarget.ts`, `PaneVisibility`) is untouched. `resolve_host_path` gets the general hidden-component rule back. Round 3 replaced it with `HOST_CREDENTIAL_DIRS`, which is allow-by-omission for the rest of `$HOME`: `~/.local/bin` (write there and you own the user's next shell command), `~/.password-store`, browser profiles and `~/.pki/nssdb` were all reachable through a planted symlink with a visible name — verified against a real home directory, and all five refused now. It over-catches `.pnpm` and `~/.cache`; for two occasional callers that is the cheaper mistake, and the refusal says which folder it resolved through. Two defects fixed while in here: * A symlinked directory listed as empty. `find` defaults to `-P`, which does not follow a symlink even as the starting point, so `-mindepth 1` discarded the only match and a real directory rendered as "Empty directory" — a first-order defect now that browsing *is* the feature. `-H` follows the starting point and nothing else, so a loop is `ELOOP` rather than a walk that does not end; verified against a live container for a symlinked directory, a broken link and a loop. `find`'s errno for the loop case is now a sentence. * `finish_download`'s replace path fired on *any* rename failure with a destination present — a vanished partial, a permission error, a directory at the destination — and deleted the user's file to complete a move that could not complete. It is now fenced to Windows (where a rename onto an existing path genuinely fails) and to a partial that still exists. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 17:05:56 -07:00
shadow-testandClaude Opus 5	9472cb3c4c	Resolve host paths before mounting them, and stop reading stored data as choice ... Four fixes that share a shape: a value already on disk, or one spelled around a check, being taken at face value. `/..` bind-mounted the entire host filesystem read-write. `is_filesystem_root` was purely lexical — trim trailing separators, refuse what was left only if it was empty or a bare `C:` — and nothing in the file ever called `canonicalize`, so `/..`, `/./`, `/home/..`, `/etc/../` and `C:\..` all passed. The daemon resolves them: `docker run -v /..:/mnt/probe` mounts the host root, and the app mounts read-*write* into a container whose agent has passwordless sudo. It is the escalation `check_mount_name_stays_under_workspace` exists to close, reached through the host-path half of the mount instead of the mount-name half. `classify_mount_source` replaces it and asks the OS: `canonicalize` applies `..`, follows symlinks, and resolves 8.3 aliases and UNC spellings on Windows. A path that cannot be resolved — `projects.json` synced from another machine, a folder not created yet — falls back to a lexical collapse rather than being refused, because refusing would make such a project unsavable; the gap is bounded, since what resolution adds is a property of paths that exist. A path that names no location at all (`C:x`, a relative path) is refused rather than guessed at. Same check now guards `ssh_key_path` and `ca_cert_path`, whose read-only mounts were whole-host disclosure at /tmp/.host-ssh. Custom env var names had no charset check anywhere, so `BASH_FUNC_stat%%` — bash's wire format for an exported shell function, body in the value — reached the container environment verbatim. Latent today because the image's /bin/sh is dash, but the pre-commit scrub runs `/bin/sh -c` as root and nothing pins that. Keys are now shell identifiers, on the project and the global list both, with the same grandfathering the folder rows get: a stored key is admitted, a new or edited one is not. The blank workspace row was persisted. The comment said it was dropped on save; the code computed the filtered list and then saved the unfiltered one, so "+ Add folder" plus a blur stored `{"Target": "/workspace/", "Source": ""}` and the project could never be started or recreated again. Every save in the section now goes through one filter, and a blur that changed nothing saves nothing. Widening the five `ClaudeCodeSettings` booleans to `Option<bool>` reinterpreted every stored record. They were plain `bool`s that always serialised, so every project ever saved carries an explicit `"env_scrub": false` that nobody chose — and under the new merge that `Some(false)` beats a global `Some(true)`, where the old rule let the global win. Upgrading silently turned five settings off, "strip credentials from subprocess environments" among them. Deserialisation now goes through a shim that dates the record by the presence of the pre-widening `enable_session_recap` key and reads its `false`s as unset. The fields skip serialising when unset, so an older binary can still parse `projects.json` after a downgrade — a `null` would fail to parse and take the whole list down, since `ProjectsStore` parses all-or-nothing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 17:02:03 -07:00
shadow-testandClaude Opus 5	dd23a52b41	Fix four upgrade-path defects the coherence audit found ... The ~/.claude.json write was `printf ... > "$CLAUDE_JSON"`, which truncates before it writes. A write that fails part-way — a full home volume, which is the exact condition half this release exists to prevent — leaves the file unparseable, and it never self-heals: the next start's jq fails on the corrupt file, MERGED is empty, and the guard skips the write that would have repaired it. That file holds the OAuth account, so the failure mode is a permanently lost login, in service of a cosmetic flag that suppresses a tip. Demonstrated: old pattern loses the credential, new tmp+rename leaves the original intact. The correct pattern was already in triple-c-task-runner. The web terminal scoped its xterm key handler to Claude sessions but not its mobile input bar or its dedicated newline button, so both sent ESC+CR into `bash -l`, where readline has no binding for it. Silent no-op, and worse from a button that stays on screen looking live. Both now consult the active session's type, and the button is disabled with a reason on a shell tab. The Config tab claimed "Off overrides a global On" without qualification. True for the env-var-driven settings, false for TUI mode, Effort level and Focus mode, whose off state is *removing* a key — an older base image's entrypoint ignores the instruction to remove it. The copy now says so and points at the base-image update. HOW-TO-USE.md said there is no add-task form; AutomationTab renders a "New task" button. That file is fetched from GitHub at runtime by help_commands.rs, so the error was live in every user's Help dialog. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 16:45:20 -07:00
shadow-testandClaude Opus 5	168b61d632	Close two cross-file handoffs from the round-3 fixes ... The terminal's drop target derived the tar entry name from the host path *after* symlink resolution, so dropping ~/Downloads/latest.log — where latest.log is a symlink — landed the file in the container under the target's name. Nothing errored; the user got a name they never typed. The Files pane had the identical bug and was fixed with `host_upload_name`; the terminal now calls the same helper, so the two drop targets cannot drift. It also drops the "dropped-file" fallback, which silently renamed anything the old `file_name()` could not parse. `migration_store::load` told the user a backup existed when it had deliberately not written one. `load` runs on every reconcile, survey and reaper pass, so a persistently corrupt record reaches MAX_CORRUPT_BACKUPS within seconds; from then on the copy was skipped while the log still read "(a copy was kept at <path>)". Three outcomes are now distinct, and the "enough already" case says so rather than naming a file that is not there — that being the message someone reads immediately before going to look for their data. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 15:52:49 -07:00
shadow-test	47960e46df	Merge branch 'r3/drop' into ship/core	2026-08-23 15:48:19 -07:00
shadow-test	73dfaf5785	Merge branch 'r3/auth' into ship/core	2026-08-23 15:48:19 -07:00
shadow-test	01fd38bc4b	Merge branch 'r3/files' into ship/core	2026-08-23 15:48:19 -07:00
shadow-testandClaude Opus 5	00128f9b1a	Make the scrub work off /usr/bin, and stop a log level deciding whether it runs ... H3 — the pre-commit scrub was a silent no-op on any base image whose coreutils are not under /usr/bin. Hardening against a PATH-planted `stat` shim by naming every tool absolutely bought nothing the `PATH` reset on the first line had not already bought — the shim lives in the persisted home volume, and uid 1000 cannot write /usr/bin or /bin — and it cost the whole feature on Alpine, which Settings -> Docker -> Custom accepts. Measured on one seeded tree in a real container: the absolute-path script printed `###TRIPLE-C-SCRUBBED 0` and left every planted file in place; the PATH-resolved one reclaims 77824 bytes. The default image is unchanged at 521038. It was silent three times over, and all three are fixed: * The script now probes for all six things it needs (`command -v` for the five tools, plus the root device id reading back as a number) and, if any is missing, prints `###TRIPLE-C-SCRUB-UNAVAILABLE <what>` and no total at all. * `scrub_writable_layer` reads that marker first and returns a new `ScrubOutcome::Unavailable`, warning with what the image is missing; a genuine `Reclaimed(0)` now leaves a debug line rather than nothing. * `commit_log_suffix` renders `Reclaimed(0)` as "ran and found nothing to drop" rather than "0.00 MB dropped", which is what a scrub that could not run used to look like. Verified in real containers: the mount-at-the-match defence still holds with a home-volume `stat` shim first on PATH (the volume survives; deleting the PATH reset from the same script empties it, so the harness can tell the difference). H2 — the pre-migration scrub had been folded into `log::info!`'s argument list to satisfy `#[must_use]`. `log::info!` expands to `if Info <= max_level() { … }`, so the awaited scrub lived inside the level check, and `logging::init` tolerates `dispatch.apply()` failing — which returns before `set_max_level` and leaves the process at `Off`. In that state the scrub never ran and the layer was committed into the longest-lived snapshot the app takes. The outcome is bound first now, `logging::init` restores the level on failure and says so on stderr, and a test scans all four files for an `.await` inside any `log::*!` argument list. Also: * `reconcile_migration` deferred a held project instead of dropping it. Its only caller fires once per "Docker became available", so a project held at that instant was never revisited for the session — phase un-normalised, no resume or rollback offered, pin left `Claimed`. It now waits for the holder to let go (20s x 90, one waiter per project) and reconciles then. * The scrub's byte total counts what a partly failed `rm` removed, by re-measuring rather than dropping the whole subtree on a non-zero exit — which was exactly the `--one-file-system` case. * The scrub exec blanks `LD_PRELOAD`, `LD_AUDIT` and `LD_LIBRARY_PATH`. `LD_PRELOAD` is in none of the reserved env families, so a project's custom env var reached a root exec and injected code into every tool the scrub runs, `PATH` reset or not. Verified against Engine 29.7 that `docker exec -e` wins. * The device test is described honestly: it is a mount test under `overlay2` and not under `vfs`, where checks 1 and 2 are what still hold. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 15:47:36 -07:00
shadow-testandClaude Opus 5	39934299f9	Stop the snapshot retry being an unconfirmed revoke, and unbrick legacy configs ... Three things, all reachable from a single credential-handling round. **The "Retry snapshot cleanup" button deleted the token.** `clear_claude_token` called `secure::delete_claude_oauth_token()` unconditionally; the `mode` argument only changed the toast wording, so there was no sweep-only path on the wire. The leftover panel rendered on `leftover !== null` alone and `onAuthenticated` never cleared it, so the sequence revoke -> snapshot skipped -> re-authenticate from the button directly above -> press the retry the panel is still offering threw away the token acquired seconds earlier, announced by a message about images. The deliberate Revoke needs a confirmation modal; this needed nothing. `sweep_claude_token_snapshots` is the honest primitive: it rewrites the images and never touches the keychain. The images are the durable record, so the retry re-derives its work from Docker and needs no stored token. Re-authenticating now clears the panel, and the Authenticate button is disabled while a cleanup runs — a sweep is a per-image inspect/create/commit/rmi over the Docker socket and takes minutes. **Sweep-first left the token live for that whole window.** The keychain delete sat behind `list_images` plus the per-image loop, at bollard's 120s-per-request default, while the UI said "Revoking...". A quit or crash in there and nothing was revoked at all; worse, `has_claude_token` stayed true and `shared_claude_auth` reads the keychain at container-*create* time, so a project whose `SecretScrub` guard had already released could be started later in the same sweep and be handed a fresh copy of the credential in its env. Keychain-first now, and the comment that claimed "no window in which a scrubbed image is re-poisoned" — true of images, silent about containers — is corrected. The reorder's original justification (crash-mid-sweep recoverability) is what the sweep-only command covers. A keychain refusal no longer discards a scrub report, because nothing has been swept yet, and both remedies stay on screen: Revoke, and the image sweep, which is now offered in every state rather than only when nothing is stored. **`update_project` validating every folder list bricked existing projects.** It validated nothing until recently while `WorkspaceSection` saved `{paths}` on every blur, so `projects.json` can hold a half-filled row, a mount name with a space, a duplicate, or `/` as a host path. Any such project became entirely unsavable — every Config toggle, every permission-mode change and `useTerminal.ts`'s tab rename came back with a message about folders — and refusing the save did not unmount anything. `validate_project_paths_update` admits a row carried over verbatim from what is stored and holds a new or edited row to every rule, which keeps the escalation closed: introducing a bad value through this command is exactly what a non-carried-over row is. The `/workspace/../tmp/claude-x` chain is the one exception and runs on every row regardless, because a stored one is live data loss rather than untidy data. `ssh_key_path` and `ca_cert_path` had no check at all; a filesystem root there read-only bind-mounts the whole host at /tmp/.host-ssh. Refused on change, with the same grandfathering. Tests are mutation-checked: reverting to sweep-first fails all four new Rust ordering tests, pointing the retry back at `clear_claude_token` fails five frontend tests, and validating an update in isolation fails the legacy-data tests. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 15:42:20 -07:00
shadow-testandClaude Opus 5	c6086b0ab3	Stop the file panel refusing ordinary paths, and hanging on a FIFO ... H6 (HIGH) — `resolve_host_path` canonicalised the host path and then re-ran the *lexical* policy over the answer, hidden-component rule included. Because canonicalisation resolves through symlinks, that rule started judging where a path happens to live rather than where the user pointed: uploading out of a dependency under pnpm (`node_modules/pkg` → `node_modules/.pnpm/…`) was refused, and so was every download into, or upload out of, a visible directory that leads to `~/.local/share`, `~/.cache`, `~/.var/app`, `~/.nvm` or `~/.cargo`. None of it was refused before the H4 fix landed. The two questions are now separate functions. `validate_host_path` judges the string the user chose, unchanged. `validate_resolved_host_path` judges the canonical form for the things only it can answer — the system roots (a Mac's `/etc` *is* `/private/etc`), the login-item directories, and a new `HOST_CREDENTIAL_DIRS` list. That last one is what keeps H4's escape closed: `Downloads/pub` → `~/.ssh` with a leaf of `authorized_keys` is refused because of where it lands, not because of how the directory is spelled. macOS handling is untouched — `/private/tmp` stays out of `HOST_SYSTEM_ROOTS` and `/var/folders` stays in the exceptions. H8 (HIGH) — the upload reservation claimed its destination with `sh -c 'set -C; : > "$0"'`, and the comment claiming that is `O_EXCL` was wrong for a destination that is not a regular file. Against a FIFO the shell opens it and blocks in `open(2)` forever; `exec_oneshot_raw` has no timeout, so `upload_file_to_container` never returned and the Files pane sat on "Uploading…" for the session with the rest of the batch abandoned. Verified in a fresh ubuntu:24.04: the old form times out and the blocked `sh` stays in `ps`; the new form answers in 35 ms. The reservation is now a `link(2)` — it claims a name atomically, never opens anything, and `EEXIST` is immediate whatever is in the way. A staging file at an unguessable name in the same directory is linked into place and unlinked, under a `trap … EXIT`. `exec_oneshot_as_within` adds a wall-clock ceiling as the second line of defence, opt-in per call site so migration's `apt-get` is unaffected. The upload contract is unchanged: default-refuse, `overwrite: Option<bool>`, and `FILE_EXISTS: <full container path> already exists`. Also fixed, all in the same surface: * A dangling symlink destination was a permanent dead end — `set -C` refused, the confirming `test -e` followed the link and said no, and raw shell text came back with no Replace on offer. `link(2)` does not follow the new-path link, and the script confirms with `[ -L ]`, so it reports as a collision. * An upload through a symlink renamed the file: the leaf came off the *resolved* path, so `~/Downloads/latest.log` landed as `2026-08-23.log` and the collision prompt named a file the user never chose. The name now comes from the path the user gave; the resolved path is still what gets opened. * `download_container_backup` leaked its partial file when the descriptor check fired. It now tracks `created` the way `stream_container_file_to_host` already did. * The failed-upload cleanup was `rm -f` on a path that, the reservation having succeeded, held whatever was written in the interim — a host file under `/workspace/…`. It now removes only an empty regular file, and the comment says what it is doing. * `resolve_container_dir` parsed a combined stdout+stderr buffer as a path. It uses the split-stream helper, like the listing next to it. * `verify_opened_path` failed open on a readlink error (`if let Ok(actual)`). A check that cannot see is not a check that saw nothing wrong; the macOS compile-time no-op is now spelled out too. * A trailing slash on a write path resolved to the directory itself. Nine new tests, all mutation-checked against the pre-fix behaviour. Two more cases added to the ignored live-Docker test: a FIFO and a dangling symlink, both timed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 15:40:05 -07:00
shadow-testandClaude Opus 5	ed91423666	Hold back the Disk panel and OS drag-out from the ship branch ... This is a scope reduction, not an abandonment. Both subsystems are preserved in full on `hold/disk-and-dragout` and are intended to come back once they have been hardened separately. Nothing here is a judgement that the features are unwanted — three successive audit-and-fix cycles each closed a critical defect in these two areas and each opened a new one, so the rest of the round ships now and these two get their own cycle rather than holding it up. Removed: the Disk settings panel and its whole reclaim / destroy / compaction surface — `DiskSettings`, `DiskProjectTable`, `useDiskUsage`, `docker/disk.rs`, `disk_tests.rs`, the disk commands in `docker_commands.rs`, and their `generate_handler!` entries. Dropping the IPC entries is the point: a UI-only removal would have left five commands callable by a compromised webview, one of them a verified arbitrary-DELETE primitive. `sweep_orphaned_snapshots`'s *command* goes with them (the panel was its only caller); the sweep itself stays. Removed: OS drag-out from the Files tab — `stage_container_file_for_drag` and its host staging lifecycle, the pointer gesture and `dragPreview`, `stageForDrag` / `isStagedHostPath`, the `tauri-plugin-drag` and `@crabnebula/tauri-plugin-drag` dependencies, and the `drag:allow-start-drag` capability grant, which could not be scoped. The capability test's expected list is updated; its `*:default` and `store:*` assertions are untouched. Kept, deliberately: drag-and-drop *into* the app (Files pane and terminal) and "Save to host…", which is now the only route out of a container. The prevention work is untouched — the pre-commit scrub and `SNAPSHOT_SCRUB_PATHS`, capped container logs, the `triple-c.base` / `triple-c.managed` labels, `sweep_orphaned_snapshots` and the startup housekeeping, the migration pin/probe reapers, scheduler log pruning, `formatBytes.ts`, and `project_lock.rs` in full with every acquisition site outside `disk.rs`. Entanglements, resolved rather than deleted blind: * `container.rs`'s `a_compaction_runs_this_module_s_scrub_script_byte_for_byte` pinned the compaction Dockerfile against `snapshot_scrub_script()`. Dropped — it existed only for compaction. `snapshot_scrub_script` and its containment tests are untouched. * `lib.rs`'s startup reap of `:compacting` tags and `triple-c-compact-*` containers is dropped: nothing on this branch creates them. * `project_lock`'s `Compaction` / `CacheClear` variants and `any_held_excluding`, `migration_commands::is_migrating`, and `formatBytes{Delta,Ceiling}` lose their last production caller but are kept and still tested, annotated with why. * `projects_store::corrupt_since` and `migration_store::peek_ownerless_since` were read only by the disk survey and are removed. The corrupt-load marker and `.bak` are still written. Verified: `npm run test` 611 passing, `npx tsc --noEmit` clean, `npm run build` green; `cargo test` 419 passed / 2 ignored, `cargo build` 0 warnings. Every test removed belongs to a removed feature — no kept-behaviour test was weakened or deleted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 15:20:22 -07:00
shadow-testandClaude Opus 5	6a8972980d	Stop an untouched secret field from deleting the stored credential ... Making `null` actually clear a secret — which it had to, since a blanked token was previously never revoked — turned the config editors into a credential shredder. Secrets are `#[serde(skip_serializing)]`, so the inputs always render empty whether or not one is stored, and the blur handlers sent `value || null` unconditionally. Focusing the Git token field and tabbing away deleted it, with nothing shown and no undo. `useSecretField` encodes the rule: only a field the user typed in may speak about a secret. Untouched, `patch()` contributes no key at all, and Rust already distinguishes an absent key from an explicit null. `withoutUntouchedSecrets` covers the structural half. `saveBedrock` spreads `{ ...bedrock, ...patch }`, and when that falls back to DEFAULT_BEDROCK_CONFIG the literal spells every secret out as `null` — so editing the AWS region would have wiped the credentials as a side effect. Also here: `WorkspaceSection` no longer saves a half-filled folder row, which `update_project`'s new validation would refuse on every keystroke between the two inputs; `snapshots_skipped` is declared on the wire type rather than widened locally; and `#[must_use]` on `ProjectGuard` and `ScrubOutcome` — which immediately caught the migration path discarding its scrub outcome, the one scrub whose silence is expensive because the layer it declined to clean is about to be committed. The capability test reads the real file and is mutation-verified: adding `core:default` back makes it fail. That grant pulls in an unscoped `std::fs::read` of any host path and went unnoticed for months, because nothing in the suite read the file at all. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 13:23:00 -07:00
shadow-test	7bbb699e4e	Merge branch 'r2/front' into integration/round-1	2026-08-23 13:15:02 -07:00
shadow-test	5df3e7996d	Merge branch 'r2/files' into integration/round-1	2026-08-23 13:15:02 -07:00
shadow-test	5d4d5d37df	Merge branch 'r2/disk' into integration/round-1	2026-08-23 13:15:02 -07:00
shadow-test	bb1c7696f9	Merge branch 'r2/scrub' into integration/round-1	2026-08-23 13:15:02 -07:00
shadow-testandClaude Opus 5	f2a84c18f9	Judge a host path by where it leads, not by how it is spelled ... `validate_host_path` was a string test. Nothing in the module called `canonicalize`, `read_link` or `O_NOFOLLOW`, so a path whose components are all visible could still land somewhere hidden: with `~/Downloads/pub` a symlink to `~/.ssh`, a `host_path` of `~/Downloads/pub/authorized_keys` has no hidden component, no `..` and no system root — and writes into `~/.ssh`. The container end is not hypothetical: `/proc/self/mountinfo` inside a Triple-C container spells the host's project paths out verbatim, so code in there knows both where to plant the link and what host path to ask for. The same bypass read host files back the other way. So the policy now runs twice: once on the string, and once on what the OS says the string resolves to. A write resolves the parent and keeps the caller's leaf, because the leaf is never followed — the partial file is created with `O_EXCL` and the download finishes with a rename, which replaces a link rather than writing through it. A read resolves the whole path, because the whole path is opened. On Linux the descriptor is then checked against the path that was validated (`/proc/self/fd`), which is what closes the window between resolving and opening; elsewhere that window stays open and the comment says so. Also here: * The upload's overwrite guard is a guard again. `noOverwriteDirNonDir` refuses only dir-over-non-dir and the reverse — file-over-file extraction proceeds, which is exactly the `.credentials.json` case (verified against a live daemon). The probe and the write are now one `set -C` exclusive create, with the path travelling as `$0` rather than as script. The `FILE_EXISTS: <path> already exists` contract with the frontend is unchanged, and now pinned by a test — as is the claim the old comment made. * Windows normalisation stopped being a string swap: `\\?\`, `\\?\UNC\` and administrative shares all reach the same places and are compared as such, and the rules are pure functions over a string, so the Windows entries are exercised on any platform. The old test passed on Linux only because `Path::is_absolute` was false for a Windows path. * Container write roots are resolved inside the container too, and the comment no longer claims more than the check does. * A failed download can no longer delete a pre-existing file that happened to collide with the partial's name. * One-shot exec output is buffered as bytes and decoded once, so a filename split across two Docker frames survives; stdout and stderr are tellable apart, so `find`'s diagnostics stay out of the listing parser; and a directory too big to buffer is described as one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 13:13:35 -07:00
shadow-testandClaude Opus 5	dcd2dfe5a3	Stop a project id from steering a Docker API DELETE, and stop trusting a store that lost its list ... C-2 (critical). `destroy_ownerless_rollback_pin` validated its tag and not its project id, then interpolated both into `triple-c-snapshot-{id}:{tag}` and handed the result to bollard. bollard does not percent-encode: `Uri::parse` joins an absolute path onto the base URL, which replaces the path outright and applies RFC 3986 dot-segment removal. An id of `a/../../v1.47/volumes/<name>?` turns a "remove image tag" into `DELETE /v1.47/volumes/<name>`. That arm is reached *because* `find_project` failed, so the id is unconstrained IPC input, and the typed confirmation is no barrier — it compares the caller's own two strings. Reproduced against the live daemon, and now a test: with the check removed the volume is gone and the test fails; with it, the volume survives and a legitimate ownerless pin still deletes. The reference that reaches `remove_image` is now the daemon's own repo_tag, matched on the parsed pair, so nothing built from IPC input addresses the API at all. The same id check now guards the owned arms of `destroy` and `compact_snapshot`, which build volume names and image references from a `projects.json` field. H-1. The ownerless arm decided ownership from the in-memory list alone and then called `sweep_orphaned_snapshots()`, which deletes the freshly dangling image on the same pass — so a corrupt `projects.json` could reap a pin whose migration is still awaiting confirmation, the one thing `pin_is_reapable` orders its conditions to prevent. It now re-reads the store from disk, runs `project_store_trust`, refuses an id the store knows, takes the project lock before reading anything a decision rests on, and checks `has_record`. H-3. The corrupt-store guard keyed on "empty list + file exists", and `ProjectsStore::new()` swallows a corrupt file without rewriting it — so the first `save()`, as little as starting a project, wrote `[{new}]` over it and the guard passed with every other project's volumes unclaimed. A corrupt load is now recorded in a sticky `projects.json.corrupt` marker beside the file, and the existing `.bak` is no longer clobbered by a second corruption. A missing `projects.json` is refused too: it cannot be told from a moved or partially restored data directory, and the genuinely fresh case has nothing to find. Also: the three migration commands surface the lock's real refusal instead of substituting "a migration is already running"; `note_ownerless_since` re-checks `has_record` after writing a tombstone, closing the window that could plant one behind a valid record and reap the pin with zero grace; corrupt migration-record copies are capped at four; `reconcile_migration` yields to any lock holder, not only a migration; and a 22-space run in a refusal string is gone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 13:11:42 -07:00
shadow-testandClaude Opus 5	e70a40507c	Stop granting an unscoped host-file read, and make a refused credential scrub recoverable ... `core:default` was an alias for nine core plugins' default sets, and one of them — `core:image:default` — carries `allow-from-path`, whose handler is a bare `std::fs::read(path)` with no scope mechanism at all. Nothing imports `@tauri-apps/api/image`, so the plugin is dropped rather than scoped; there is nothing to scope it with. The capability file now enumerates what `app/src` actually invokes, which is `core:event`'s listen/unlisten and nothing else from core — every emit in this app originates in Rust. `core:menu`, `core:tray`, `core:window`, `core:path`, `core:resources` and the three dead `dialog:` grants go with it. `core:webview:allow-internal-toggle-devtools` stays because Tauri's own injected debug script calls it; both it and the command behind it are `cfg(any(debug_assertions, feature = "devtools"))`, so it is absent from a release bundle. Verified empirically: an unknown identifier fails the build, so every identifier kept is real and the regenerated `gen/schemas/capabilities.json` carries the opener scope verbatim rather than silently dropping it. `opener:allow-open-url` cannot be host-narrowed — the terminal opens links Claude printed inside the container — so what it does and does not buy is recorded instead, including the verified fact that each scope entry's `app` defaults to `Application::Default`, which matches only `with == None` and therefore refuses `openUrl(url, "/bin/sh")`. `clear_claude_token` deleted the keychain entry first and swept the snapshot images second. The sweep runs once and skips a project another operation holds, the deleted entry made `has_claude_token` false, and Revoke rendered only while a token was stored — so a project that happened to be starting during a revoke kept a live ~1-year OAuth token in its snapshot's `Config.Env` permanently, with Reset (which destroys both volumes) as the only remaining remedy. The sweep now runs first, so a crash mid-revoke leaves the app still saying "authenticated" with the same button still able to finish; a busy project is reported as `snapshots_skipped` rather than folded in with images that genuinely cannot be rewritten; and the panel keeps a retry visible independent of token status, plus offers the sweep outright when nothing is stored, because a snapshot committed by an older build carries the token either way. The retry is the same command — it is idempotent, and the images are the durable record. Also: `openai-compatible-api-key` was written but never deleted, so it outlived its project. The key list is now the single definition and an unlisted key is refused outright, so the writer cannot get ahead of the deleter again. `store_or_clear_project_secret` lands here unused on purpose: the editors send a blanked field as `null` and `store_secrets_for_project` skips `None`, so clearing a secret through the UI is impossible today. Its one call site is in `commands/project_commands.rs`, which belongs to another change in this round. No `devCsp` was added. `tauri dev` loads the main document straight from Vite, and Tauri only attaches a CSP to documents it serves itself — the dev server is proxied through `tauri://` only when `PROXY_DEV_SERVER`, which is `cfg!(all(dev, mobile))`. A `devCsp` here would be inert config that reads as protection. The reasoning, and the one place that could set one, are recorded. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 13:07:07 -07:00
shadow-testandClaude Opus 5	5926a52ff6	Stop the terminal and Files panes refusing drops onto their own chrome ... The z-order gate added last round asked `el.contains(elementFromPoint(x, y))` — "is the thing painted here mine?" — and was handed `TerminalView`'s inner xterm host while every overlay in that pane is a *sibling* of it. So any point under the pane's own chrome answered "not mine" and the drop was refused, with no message and no log line. The "▼ Following / ▽ Paused" toggle is rendered unconditionally at `absolute top-2 right-4`, and `ToastHost` is `fixed bottom-4 right-4` 24rem wide with error cards that never time out: two corners of the terminal, and one of the Files pane, that could not accept a file for as long as the app was running. It shipped green because jsdom has no `elementFromPoint`, so not one of the 81 drop tests entered that branch. The tests here install one. The question the gate asks is now "is a *blocking overlay* painted here?". Chrome the pane paints over itself is not one; a dialog backdrop is, and `ui/Modal` marks its own backdrop so the element `elementFromPoint` actually returns is the one carrying the marker. `classifyDrop` also separates "aimed at me and swallowed" from "not my drop", so the first gets a toast and a log line and the second stays silent. Three defects around it: - **A dialog now refuses only the points it covers.** `dropIsBlocked` is document-wide and `ui/Modal` portals to `document.body`, so any open dialog refused every drop in the window. The deeper half of that is that a dialog opened in project A really was still on screen after a tab switch — the pane hides itself with a `hidden` class, which a portal does not inherit — so `PaneVisibility` lets `App` tell a `Modal` its pane stepped aside, and a hidden one paints nothing, traps no focus, answers no Escape and blocks no drop while staying mounted with its state intact. - **`devicePixelRatio` is applied on Windows only.** Only wry's WebView2 backend hands over physical pixels; the macOS and GTK ones deliver logical points and `tauri-runtime-wry` does not rescale them. Halving those was survivable while the test was a bare rect and is a refused drop once z-order joins in. Read from the wry/tauri sources, not verified on a HiDPI Mac or GTK box. - **`isFileExistsError` can no longer be forged by a filename.** It matched `fileexists` anywhere in a normalised error, so uploading a host file called `file-exists.txt` turned *any* failure into a collision — and Replace re-invoked the upload with `overwrite: true`. The marker now has to stand alone in the backend's canonical form, or be a whole discriminant value. - **A refused compaction or cache-clear keeps its dialog.** `reclaim` reports refusals inside `Ok`, so "did it throw" read one as success: the dialog closed, the tick list was dropped, and the explanation appeared in the outcome panel several screens above the row that was clicked. The dialog now stays put and renders the backend's own sentence verbatim. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 13:03:57 -07:00
shadow-testandClaude Opus 5	42ef1865cc	Remove a stray .deb committed by an over-broad git add -A ... An audit agent downloaded findutils to check GNU find's operand parsing; a76f2c0 swept the 298 KB artifact into the tree along with the capability changes, unmentioned in its message. Second time this session that `git add -A` has staged something unrelated — ignore the pattern so it cannot happen a third time. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 12:35:55 -07:00
shadow-testandClaude Opus 5	4f6c012071	Make a rollback pin outliving its project visible and deletable ... `survey_rollback_pins` walks images, not projects, and deliberately tolerates an absent project by falling back to the raw id as the display name. Two things then dropped it on the floor: `destroy` called `find_project` before the confirmation check, so it refused such a pin every time, and the per-project table joins destructive items to rows by project_id, where rows come only from projects in the store. The result was a multi-GB `pre-migration-*` image that the scan measured, the panel never rendered, and nothing could remove — in the one screen built to find exactly that. `destroy` takes the same early return `OrphanVolume` already takes, and still validates the tag: `latest` names the project's live snapshot, so the ownerless path must not be a way around that check. The UI grows a bucket for destructive items matching no row, rather than filtering them away. The typed gate already compared against the id via `project_name`; the dialog now says "project id" instead of asking for a project name that no longer exists. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 12:18:04 -07:00
shadow-test	6b8d43414d	Merge branch 'fix/reconcile' into integration/round-1	2026-08-23 12:13:55 -07:00
shadow-testandClaude Opus 5	1768240861	Unstick a project row whose lifecycle command was refused ... Two more pieces of drift from the same merges, both invisible to `tsc`. **A refused Start/Stop/Reset strands the row.** `start`, `stop` and `rebuild` paint an optimistic "starting"/"stopping" so a click moves the row at once. That was safe while the only way these could fail was after the backend had begun changing things. `fix/sec`'s per-project lock ended that: all three now take the lock and are refused *before* any state changes, and `stop` could not fail this way at all before — it took no exclusion. So the optimistic paint has nothing to become, `isTransitioning` disables both Start and Stop, and the only thing that clears it is `reconcileProjectStatuses`, which runs once from `App.tsx` when Docker first appears. Clicking Stop during a compaction left the project unusable until the app was restarted. `withOptimisticStatus` re-reads the authoritative list when the command throws, and falls back to the status that was on screen if even that call fails — two failures in a row must not land on the one state there is no way out of. The error is rethrown unchanged, so the toast is unaffected. Five of the six new tests fail against the previous code; the sixth pins that the optimistic paint still happens on the way in. **Six secrets are typed as if they arrive, and they never do.** `git_token`, the four Bedrock credentials and `OpenAiCompatibleConfig .api_key` are `#[serde(skip_serializing)]` in Rust, so the key is absent from every project the backend returns — reading one gives `undefined`, not the `null` the type promised. Every current reader happens to use `?? ""`, so nothing is broken today; a single `=== null` would have been a branch that silently never ran. They are optional now, which makes that a compile error, and documented as write-only, which is what they are. 669 frontend tests pass, `tsc --noEmit` clean, `npm run build` green. Nothing under `src-tauri/` touched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 12:12:50 -07:00
shadow-testandClaude Opus 5	7e1f8df1ff	Reconcile the frontend with the round-1 backend contracts ... Five backend branches merged and the TypeScript still compiled, because none of this is a type error: a field that arrives `undefined`, a variant nothing emits any more, a prompt whose loop never closes. Six things. **Orphaned volumes are destructive now, not safe.** `ReclaimTarget:: OrphanVolume` is gone from Rust; the object is a `DestructiveTarget:: OrphanVolume { name, project_id }` confirmed against the *volume's* name, there being no project to name. The TS union still listed it under `ReclaimTarget`, and — worse — `DiskProjectTable` keys destructive items off `project_id`, which an orphan's never matches. So the item existed in the plan and appeared nowhere on screen. `DiskSettings` now splits the plan's destructive list and gives orphans their own section with a per-volume `TypedConfirmModal`. The copy says what a `triple-c-claude-config-*` volume actually is — a Claude login credential, every plugin and skill, every transcript that project had — and keeps the sentence explaining that "no matching project" is a lookup against the project list and is never inferred from a project being stopped or having no image, which is the inference that once flagged two live projects. `TypedConfirmModal` grew a `subject` prop: asking a user for "the exact project name" of a volume that has no project is asking for a string that does not exist. **Snapshot and Total reconcile.** `ProjectDiskRow.snapshot_attributed_bytes` is the single figure `snapshot_attribution()` exists to produce. The column rendered `snapshot_above_base_bytes` and fell back to `—` while the Total was `size - shared` regardless — and in that branch `size - shared` is the whole 4.7 GB base image, charged per project and then added again as a base-image row. One field, one rule. The one branch where the figure *is* the whole image says so rather than passing itself off as a share. **The overwrite loop closes.** Traced end to end: a `FILE_EXISTS:` refusal raises the prompt, Replace re-invokes with `overwrite: true`, Skip advances, "…all" answers the rest without asking, and picker and host-drop both reach `uploadFileToContainer` through `uploadPaths`. Two gaps: a second batch's `askOverwrite` overwrote the first's resolver, leaving that batch awaiting an answer no dialog could produce; and the backend's written refusals — a hidden host folder, a path outside the write roots — were passed as a toast `detail`, which `ToastHost` renders as collapsed monospace behind a "Details" button, so the only sentence that explained anything was the part nobody saw. `readableRefusal` promotes it to the headline when a batch failed the same way. **The browser pane's sandbox is pinned.** `allow-same-origin` must stay (the proxy's gate reads `Origin`/`Referer`, and an opaque origin sends `null`); every top-navigation grant and `allow-popups-to-escape-sandbox` must stay absent, and the test names the offending token rather than printing a set diff. **`@tauri-apps/plugin-store` is gone** from `package.json` — its capability grants were removed as a host-file-write primitive and nothing in `app/src` imports it. The lockfile was updated with `--package-lock-only`, deliberately: `node_modules` is a symlink shared with other worktrees and a real install would have pulled it out from under them. Nothing under `src-tauri/` is touched. 663 frontend tests pass (was 635), `tsc --noEmit` clean, `npm run build` green, `cargo test` 446 unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 12:02:34 -07:00
shadow-testandClaude Opus 5	a76f2c0a17	Close the cross-stream gaps the parallel fix round left open ... Three items each of which fell between two agents' file lists. `scrub_secrets_from_snapshots` was the third unsynchronised writer of `triple-c-snapshot-{id}:latest`, after a recreate's commit and a compaction. It has the same read-modify-write shape — create a scratch container from the snapshot, commit back over the same tag — and loses the same race, which here means re-baking the very credential it exists to remove. It now takes the project's claim, under a new `ProjectOp::SecretScrub`, and reports a snapshot it had to skip rather than rewriting it unsafely. Its scratch container also now carries `triple-c.scrub=true`, so the Disk panel's reclaim bucket discriminates by label and by the live claim rather than by a clock. The 15-minute age gate stays as the backstop for the cross-process case the claim cannot see. The store plugin is unregistered and its dependency dropped. Its capability grants were removed as a host-file-write primitive; the registration without a grant was unreachable but dead. Finally, container.rs's fold test was pinning a fold that no longer exists — disk.rs now emits the JSON exec form. It asserts the stronger property instead: the script a compaction runs is byte-for-byte the one snapshot_scrub_script() produces, so the two files cannot drift silently. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 11:53:25 -07:00
shadow-test	17f031a5d7	Merge branch 'fix/disk' into integration/round-1	2026-08-23 11:46:17 -07:00
shadow-test	5fba7d6d35	Merge branch 'fix/front' into integration/round-1	2026-08-23 11:46:17 -07:00
shadow-test	433afa5a49	Merge branch 'fix/sec' into integration/round-1	2026-08-23 11:46:17 -07:00
shadow-test	fcea506dce	Merge branch 'fix/files' into integration/round-1	2026-08-23 11:46:17 -07:00
shadow-testandClaude Opus 5	6abc7f27a4	Fix disk/migration defects and add a real per-project lock ... The compaction panel's headline action had never worked, three reclaim paths could delete data with no confirmation and no grace period, and the app's only mutual-exclusion primitive was one-way. **A per-project lock (`project_lock.rs`).** `ACTIVE_MIGRATIONS` was the app's only exclusion and everything but migration merely *polled* it once at entry. Compaction, start/stop/recreate, Reset and destroy now **acquire** a `ProjectGuard` and hold it for the whole operation; `is_migrating` is a view onto the same registry. Closes the three verified interleavings where a compaction commits `flat(A)` over a `:latest` that a migration, a recreate or a Reset had already moved. In-process only — the two-instance case is documented in the module, not solved, and the daemon-wide reapers gained age gates to bound it. **H1: compaction never ran.** `fold_shell_script` joined the scrub script's lines with a space, so every build died on `syntax error: unexpected "do"`. Replaced with the JSON exec form, which carries any script verbatim; `sh -n` and a real end-to-end build now cover it (159.5 MB / 9 layers -> 33.7 MB / 1 layer, setuid and multi-line env preserved). **H2/H4:** `reclaim_migration_pins` and `survey_rollback_pins` apply `parse_rollback_tag` and `pin_is_reapable` like every other path, and stop double-counting an image with two pin tags. The 14-day grace period is re-anchored from the tag's timestamp (when the migration *started*) to a tombstone recording when the record went missing, with clock skew handled in both directions. **H3:** `migration_store::load` no longer renames a corrupt record aside — that destroyed the `has_record` signal both pin reapers depend on. `save` fsyncs the file and the directory, and corruption backups are timestamped. **H2b/M2:** a crashed compaction's `:compacting` tag and `triple-c-compact-*` container are reaped at startup; the stale-container sweep moved from the end of a compaction to the start, where its doc always claimed it was. **M5/M6:** orphan-volume deletion moved from a `Safety::Safe` tick to the destructive path with a typed volume name; `project_store_trust` reads the real `projects.json` so a second instance's project is not offered as an orphan. Numbers: `images_total_bytes` uses `df()`'s deduplicated `layers_size`; the Total column is derived from the same figure the Snapshot column shows; partial container/staging reclaims report their failure count; `human()` no longer prints "1000.0 KB"; `docker_cli` has a timeout; blocking `fs` calls moved to `spawn_blocking`. Also fixes `ProbeContainerGuard::remove_now`, which disarmed before the await and so did nothing on the cancellation path it exists for. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 11:45:28 -07:00
shadow-testandClaude Opus 5	2b6501d8e5	Give the exit-status poll room now that it fails closed ... `wait_for_exec_exit` returning `None` used to mean "call it 0"; it now fails the call, so a busy daemon that has not settled within ~1s would turn into a spurious "the rename failed". The loop exits on the first poll that reports finished, so a wider window costs nothing when things are normal. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 11:35:41 -07:00
shadow-testandClaude Opus 5	3329e07d3d	Fix the file command surface: argument injection, unbounded reads, unchecked paths ... `list_container_files` could delete the user's files. It builds `["find", path, "-mindepth", …]`, and GNU find ends its starting-point list at the first argument beginning with `-`, so a `path` of `-delete` gave it zero starting points (defaulting to `.`, which for an exec that sets no working_dir is the container's WorkingDir — the bind-mounted project) and an expression starting with `-delete`. Verified against a live container on findutils 4.10.0: files and empty directories went out of the host bind mount, and because `exec_oneshot` discards the exit code the panel then reported an empty folder. The rest of the module had the same shape of hole: * Every path parameter — `path`, `from_path`, `parent_path`, `container_dir`, `container_path`, `host_path` — arrived over IPC unchecked. There is now one validator for container paths (absolute, no `..`, no NUL, length-capped), a second for the ones that *change* something (contained in /workspace, /home/claude or /tmp), and one for host paths, which refuses traversal, system locations and hidden components. The `save()`/`open()` dialog in front of these commands is a UI convention, not a boundary. * `download_container_file` passed `None` for the fetch cap, so the cap was inert: the whole transfer was buffered in host RAM twice, and the directory refusal came *after* the buffer, so `/` meant buffering the container's filesystem before erroring. Downloads now stream through a bounded channel into the tar reader, which refuses a non-regular entry and an oversize one before the host file is created at all. Verified against a real container: a 300 MiB download peaks at 10 MiB RSS, a 9 GiB sparse file is refused in 0.01s with nothing written. * Both download paths (file and backup) used to create — i.e. truncate — the user's destination up front and delete it on a stream error, which is precisely the wrong order for a path that already holds something. They now write beside it and rename on success. * `upload_file_to_container` silently clobbered: no existence check anywhere in the stack. It now refuses by default with a FILE_EXISTS marker the frontend turns into a Replace/Skip prompt, and takes an `overwrite` flag for the retry. * `exec_oneshot_inner` read an undeterminable exit status as 0, so rename and mkdir reported success for an exec nobody could read the outcome of. It fails closed now. * A tab in a filename forged the type/size/permission columns of a listing row, and a newline forged a whole row. `find` now prints the name last with NUL-terminated records. While verifying the size ceiling against a real container, the tar header's size field turned out to be unusable past ustar's 8 GiB octal limit — Docker's Go writer puts the real size in a PAX record — so both readers take it from `entry.size()` instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 11:34:04 -07:00
shadow-testandClaude Opus 5	092972fe92	security: close capability, CSP and auth-bridge holes ... capabilities/default.json - Drop every `store:*` grant. `@tauri-apps/plugin-store` has no caller in `app/src`, and the plugin's `resolve_store_path` is a `PathBuf::push` against AppData — `push` discards the base for an absolute path, so the grant was an arbitrary host read/write from the webview. - Replace `opener:default` with a scoped `opener:allow-open-url` (http/https only). That drops `reveal_item_in_dir`, which the plugin does not scope-check and nothing here calls, and the unused mailto:/tel: scope. - Record the unscopable `drag:allow-start-drag` residual risk in `description`. tauri.conf.json - Add `form-action 'none'`, `base-uri 'none'`, `object-src 'none'`. `form-action` has no `default-src` fallback, so an injected auto-submitting form was unblocked even though `script-src 'self'` blocks XSS. - Remove the dead `asset:` / `https://asset.localhost` img-src and `data:` font-src grants; `blob:` stays (the file viewer uses it). auth_bridge - The reserved-port set covered only this project's mappings and the two browser-view ranges. It now also covers the gateway, STT and web-terminal host ports (configured value and shipped default, read off the settings models) and every other project's published host port. A container binding container-loopback 4000 / 9876 / 7681 while those services were stopped had that port mirrored onto the host, unauthenticated, within one poll. - Gate the host listener on fetch metadata: refuse a request that is a cross-site sub-resource, allow navigations (the OAuth redirect) and anything without `Sec-Fetch-*`. Non-HTTP connections are classified from their first line and forwarded verbatim. Residual risk is spelled out in the module docs. - Bound the forwards: max concurrent connections per port, a first-byte deadline enforced before any `docker exec` is created, and an idle timeout. browser_view/mod.rs - `pick_viewer_port` reads procfs with `/usr/bin/cat`, not a bare `cat` the container can shim via its writable PATH entry. - Treat port choice as check-then-bind: walk to the next free candidate when the viewer does not come up, instead of failing the start. BrowserTab.tsx - Sandbox the viewer iframe. Container-controlled content could `top.location` the app's webview away. `allow-top-navigation*` and `allow-popups-to-escape-sandbox` are deliberately absent. HelpDialog.tsx - Escape the quote characters in the entity pass and escape captured attribute values. `href="$2"` with `$2` = `[^)]+` let remote GitHub markdown close the attribute and open another, in a document rendered with `dangerouslySetInnerHTML`. web_terminal/terminal.html - SRI hashes plus `crossorigin` on the three jsdelivr bundles and the stylesheet, and a CSP for the page — it is served 0.0.0.0 behind a permissive CORS layer and nothing else gives it one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 11:13:15 -07:00
shadow-testandClaude Opus 5	d6f065a2b6	Fix HIGH and MEDIUM frontend defects ... Files pane - F16: a drag-out released back inside the app no longer re-imports its own staged copy over the container original. An in-flight flag (cleared from the drag plugin's `onEvent` channel, with a watchdog) suppresses the drop and the "Drop files into …" hint, and an exact staged-path filter is the second line of defence — the `path|size|modified` cache could otherwise write a minutes-old snapshot over a file an agent had since rewritten. - F17: a slow upload/rename no longer yanks the user back to the directory the operation started in. Every operation captures its target path and re-lists only if the user is still there; failures go to the toast host either way. - The grid keeps keyboard focus. Roving tabindex (one tab stop, not one per row) plus focus restore after navigation, rename commit/cancel and Escape. - Transient failures now surface in `ToastHost` (z-[60], persistent aria-live) instead of a `role="alert"` 300 rows down a scroller or behind a modal overlay. The inline error is kept only for the listing failure. - `navigate` is sequenced by generation; "Save to host…" sets `busy`. - Grid a11y: column headers, a text affordance for folder vs file, a live region that is mounted empty and announces completion, Label-in-Name fixed. - FileViewerModal: the blob URL is released only once its replacement exists; the preview is a focusable, named, scrollable region. Native drop routing - New `lib/dropTarget.ts`: the hit test now refuses a drop while any `[aria-modal="true"]` dialog or `[data-blocks-drop]` overlay is up, and checks z-order where the environment can answer it. Shared by FilesTab and TerminalView; App's shutdown overlay opts in. Disk - A partially failed reclaim says so in words ("… — 2 of 5 failed"), not by hue alone. - The scan/reclaim race is closed: every mutation retires an in-flight scan, so a scan can no longer repaint a pre-reclaim report plus a clickable plan of objects that are gone. Scan is disabled while working; the status is a live region; a failed destructive action keeps its dialog open and reports there. - The "unknown" layer count gets a screen-reader fallback; `--text-disabled` no longer carries live information. Terminal / OAuth - After the toast is dismissed, a truncated heuristic guess can no longer fill the slot that an exact OSC 8 or relay URL occupied — the detector remembers every exact URL and drops any candidate that is a strict prefix of one. - The prompt is reachable by keyboard: Ctrl+Shift+O jumps to the default action, Escape dismisses, focus returns to the terminal, and auto-dismiss holds off while focus is inside. It deliberately does not steal focus. - UrlToast renders through `ui/Button` and `--shadow-overlay`. Elsewhere - AuthBridgeRow: a pushed `auth-bridge-changed` status always outranks an older awaited toggle result. - The last two ad-hoc byte formatters route through `lib/formatBytes`. Contract for the backend agent: `upload_file_to_container` refusing to overwrite must satisfy `isFileExistsError` in `src/lib/uploadErrors.ts` (marker `FILE_EXISTS`) and accept an `overwrite` argument; the frontend turns that into an `ui/Modal` Replace/Skip prompt rather than a raw error string. Tests: 536 -> 627 passing. `npm run build` and `npx tsc --noEmit` green. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 11:11:43 -07:00
shadow-test	0003793abb	Merge branch 'feat/disk-ui' into integration/round-1	2026-08-23 09:51:44 -07:00
shadow-test	2b9bf56f25	Merge branch 'feat/drag-out' into integration/round-1	2026-08-23 09:51:44 -07:00
shadow-testandClaude Opus 5	611f67cca7	Fix what review found in the Disk section ... Safety: - `destroy`'s rollback-pin arm took a tag over IPC and interpolated it straight into an image reference it then removed. `tag: "latest"` named the project's live snapshot, deleted under a dialog saying "rollback pin". It is the one destructive variant carrying a free-form string, so it now goes through `parse_rollback_tag`. - The compaction's scratch container was named `triple-c-scrub-*`, which is what the scrub reclaim bucket hunts and force-removes. A reclaim from a second window would have destroyed the container a running compaction was about to commit. It gets `triple-c-compact-*`, swept at the start of the next compaction rather than from a bucket anything else can fire. - Deleting a home or config volume only refused a *running* container, but a stopped one still pins its volumes — the resting state of every project ever started — so the user typed the project name and met a raw 409. The container is now removed first and `loses` says so. Correctness: - The compaction Dockerfile emitted no `LABEL`, so the flattened intermediate could never match the sweep's `dangling` + `triple-c.managed` filter that three cleanup paths rely on. Verified on Docker 29.7.2 that the label lands on the final stage, the build still yields one layer, and untagging the staging tag after the commit leaves the committed snapshot intact and startable. - `snapshot_commit_layers` silently meant something else when `triple-c.base-image-id` was absent — the normal case for a pre-label project — counting the base's own layers and letting a never-recreated project qualify for compaction. `base_lineage_known` now carries that, the column says "unknown", and the plan does not offer the rewrite. - `destroy` returned a `ReclaimResult` wearing a `ReclaimTarget` that named work it had not done (a home-volume deletion came back as `OrphanVolume`). Split into `target` / `destroyed`, exactly one set. - `formatBytes` ran `toFixed` after the divide loop, so 999,999 rendered as "1000.0 KB" — in the app's only byte formatter, in a panel full of near-boundary sizes. - `is_base_image_reference` split on the first colon, so a registry port ate the repo name. UI: - `snapshot_above_base_bytes: null` — deliberately unmeasurable — rendered as "0 B", the one guessed number in the table. - Layer count was flagged by colour alone; it now says "stacked". - The tick list survived a reclaim, so the same call could be re-fired at objects that no longer existed. The plan is dropped after any action and the panel says the totals predate it. - `setReport` landed before the plan call was awaited, so a plan failure rendered fresh totals above the previous scan's rows. - Both confirmation modals unmounted before awaiting, making the entire busy path dead code during multi-second work. - `buildx du` failures silently showed `docker system df`'s under-reported build-cache figure with no explanation. - Tooltip text reached no assistive tech, so two headers announced as "Help"; hardcoded input id; error-toned glyph in warning-toned panels; `sweepOrphanedSnapshots` and `clearOutcome` had no callers. - Four docstrings claimed things the code did not do, and two tests were named for behaviour they did not assert. Tests: 513 frontend (was 502), 370 Rust (was 365). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 09:51:09 -07:00
shadow-testandClaude Opus 5	77ef2291d7	Add a Disk section: see where the bytes went, and get them back ... Every recreation runs `docker commit`, which stacks a layer and never rewrites one, and 24 conditions in `container_needs_recreation` trigger a recreation. Prevention landed earlier on this branch; this is the half a user can act on. The per-project table leads with the two numbers that explain the mechanism rather than just the total: how many commit layers a snapshot has stacked above its base, and what the container's writable layer will add at the next commit. Backend (`docker/disk.rs`, commands in `docker_commands.rs`): - `get_docker_disk_usage` — one `df()` joined against the project store, behind an explicit Scan button because it walks the whole daemon. - `list_reclaimable` / `reclaim` — classified buckets with measured bytes, planned off the existing report so re-planning costs no second scan. - `destroy_project_disk_object` — one object, typed confirmation. - `sweep_orphaned_snapshots` — exposed, so its report is finally visible. Safety is structural: `reclaim` takes `ReclaimTarget`, which has no variant that can name a live project's data. Destructive work is a separate type reached only through `destroy`. No unfiltered prune is called anywhere, and nothing outside a `triple-c*` name or `triple-c.*` label is touched. Orphan detection subtracts ids from the project store and consults nothing else. From the daemon's side an idle live project and a deleted one are indistinguishable — volumes present, no container, no image — so inferring from container or image absence would offer a live project's credentials and transcripts for deletion. A store that loaded empty from an existing `projects.json` is treated as a failed load, not as "no projects", because `ProjectsStore::new()` recovers from a corrupt file by starting empty. Three things verified against a live Docker 29.7.2 rather than assumed: - Compaction is a two-stage build (`FROM scratch` + `COPY --from`), which keeps every byte inside the daemon; bollard's import buffers a whole image into memory. uid/gid and setuid survive; a 192.6 MB/4-layer synthetic came out 45.7 MB/1 layer. Image config does not survive, so it is replayed via create+commit, which round-trips a multi-line env var that a Dockerfile `ENV` could not. - Flattening breaks base-layer sharing, so the result carries its own copy of the base. Eight of ten real projects had a 0.10–1.32 GB delta over a 4.72 GB shared base — compacting those costs ~4 GB. The bound now subtracts that penalty, such projects are not offered at all, and the run compares unique bytes and abandons a rewrite that would grow. - `docker builder prune` reports `Total:`, not `Total reclaimed space:`, so the first parser scored every prune as freeing nothing. The Windows/WSL2 note is mandatory and its copy lives in Rust beside the tests that pin it: pruning frees space inside `ext4.vhdx`, which never shrinks on its own, so C: does not change until the disk is compacted. Also adds `lib/formatBytes.ts` — the app had four disagreeing copies, and `projects/home/format.ts` and `migrationCopy.ts` now delegate to it with byte-identical output. Base 1000 by default, matching what Docker prints. Tests: 502 frontend (was 453), 365 Rust (was 322). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 09:39:54 -07:00
shadow-testandClaude Opus 5	1c834a0b08	Drag a file out of the Files tab onto the host desktop ... The Files tab could accept a drop but never produce one: getting a file out meant "Save to host…" and a file picker. This adds the other direction. Two constraints shape it. `dragDropEnabled` is on — TerminalView needs it, since the native drag-drop event is the only one carrying dropped file paths — and it blocks HTML5 drag inside the webview, so `draggable` plus `DataTransfer.setData("DownloadURL", …)` was never available. The gesture is therefore pointer events into `tauri-plugin-drag`, the same shape and the same reason as the tab strip's drag. And the file being dragged does not exist on the host at all: it lives in a container, and the OS can only drag a real host path. So a drag-out is a copy first and a drag second. `stage_container_file_for_drag` materialises the file into `<os-temp>/triple-c-drag-out/<session>/<slot>/<name>` through the same `fetch_container_file` the download and the viewer use, keeps the original filename (a dropped `tmp1234` is not a file anyone wants), and caps at the 256 MiB an upload already caps at, naming "Save to host…" in the refusal. The path comes from Tauri's path API rather than `/tmp`, because on Windows it is neither. The staging directory has a lifecycle, because whole files accumulating in the host temp dir would be the disk problem this project just fixed, in a new place: cleared on exit inside the existing teardown (still guarded on the main window), and reaped at startup for whatever a crash left behind. The copy is also an async gap in the middle of a gesture that feels instantaneous, and the OS only adopts a drag while the button is still down. Small files beat the pointer; large ones do not — so the staged path is cached per entry (keyed on size and mtime, so an edited file re-stages) and the pane says the copy is ready and to drag again, which is an instruction rather than an apology because the retry is immediate. A per-file slot keeps `a/notes.txt` and `b/notes.txt` from becoming the same host path. "Save to host…" stays exactly as it was. Drag-out is the enhancement; a platform that refuses `startDrag` says so and points back at it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 09:12:59 -07:00
shadow-testandClaude Opus 5	0a022dfcf0	Let a project turn a globally-enabled Claude Code setting back off ... The six boolean settings were plain `bool`s merged with `if p.x { true } else { g.x }`, so a project could only ever add to the global set. There was no project value that produced `false` — turning a switch off at project level simply fell through to the global value and the control did nothing. Widen them to `Option<bool>`. `None` means "not set at this level": inherit the global on a project, leave Claude Code's own default alone globally. `Some(false)` is a deliberate off and wins outright. The fingerprint now formats with `{:?}` rather than `{}` — `None` and `Some(false)` mean different things, and conflating them would leave the container un-recreated when a project switched from inherit to off. The project editor grows a third "Global" state per switch; the global editor has nothing to inherit from, so it stays a plain toggle and keeps collapsing to null at the default. Its three existing tests passed unchanged and caught a first attempt that rendered unset as off, which would have told every user their session recap was disabled. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 09:05:39 -07:00
shadow-testandClaude Opus 5	bb41275cea	Keep the managed settings payload safe for un-migrated projects ... The null-means-delete convention is only understood by the entrypoint.sh shipped alongside this code. An existing project recreates from its own snapshot image, which carries whatever entrypoint it was built with, and an older one merges with a plain `.[0] * .[1]` — so the literal nulls would land in the user's settings.json rather than clearing the keys. Verified against jq: that produces `"tui": null, "effortLevel": null, "viewMode": null, "awaySummaryEnabled": null`, risking the whole file being rejected and taking the user's own `model` and `statusLine` with it. Split the payload instead. CLAUDE_CODE_SETTINGS_JSON now carries only keys that have a value and is safe under either merge; the new CLAUDE_CODE_SETTINGS_CLEAR carries the key names to delete and is ignored by an entrypoint that predates it. Such a project keeps the old sticky behaviour until it is migrated or Reset, which is the pre-existing state rather than a regression. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 08:40:27 -07:00
shadow-testandClaude Opus 5	2ca86bb5d8	Scrub the writable layer on the migration path too, and de-duplicate CLAUDE_JSON ... Two integration fixes after merging the three feature branches. `scrub_writable_layer` is a `docker exec`, so it only works while the container runs. `migrate_project_to_base` stops its container one line before the pre-swap commit, which meant the single largest snapshot Triple-C ever takes was the one path that committed unscrubbed. Call the scrub explicitly before the stop instead of relying on the call inside `commit_container_snapshot`. Also drop a duplicate `CLAUDE_JSON=` assignment in entrypoint.sh. The Shift+Enter block re-declared it defensively to avoid a merge conflict with the awsAuthRefresh block; the conflict did not materialise. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 08:38:01 -07:00
shadow-test	df6d2f1ca4	Merge branch 'feat/file-manager' into integration/round-1	2026-08-23 08:35:51 -07:00
shadow-test	dacc1157ec	Merge branch 'feat/disk-and-settings' into integration/round-1	2026-08-23 08:35:51 -07:00
shadow-testandClaude Opus 5	dd2894cc60	Stop Docker disk growth, and fix the Claude Code settings that never worked ... 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	2026-08-23 08:35:10 -07:00
shadow-testandClaude Opus 5	22d142c70d	Shift+Enter newline, OAuth URL truncation, and the auth bridge toggle ... Three fixes that all land on the same journey: sign in, paste a prompt, and have the terminal behave the way every other Claude Code host does. Shift+Enter inserts a newline ----------------------------- xterm.js does not consult `shiftKey` for Enter (`Keyboard.ts`, case 13), so Shift+Enter was byte-identical to Enter and submitted the prompt. Both terminals now send `\x1b\r` (ESC+CR) instead, which Claude Code parses as return+meta — the same bytes its own `/terminal-setup` writes into the VS Code, Cursor, Alacritty and Zed keymaps, so this is in-band rather than a guess. Not `\n`: Claude Code accepts it, but a shell would run the line, so the two session types would diverge. Bound in Claude sessions only for that reason. `entrypoint.sh` sets `shiftEnterKeyBindingInstalled` in `~/.claude.json` so the CLI stops printing its "run /terminal-setup" tip. Purely cosmetic — the decoding is unconditional either way. Alt+Enter has always done the same thing (xterm ESC-prefixes on altKey) and was simply never documented. It is now, along with the rest. OAuth login URL truncation -------------------------- Two producers wrote one toast slot, last-writer-wins. The OSC 7777 relay delivers the URL base64-encoded and therefore exact; ~300 ms later the screen-scraper's debounce fired and overwrote it with a truncated guess at the same link — a URL that parses, points at the right host, and authorises nothing. The user is the one who has to notice. Why the scraper truncated: `ANSI_RE` strips OSC sequences wholesale, including the OSC 8 hyperlink whose parameter carries the complete URL. Claude Code slices the *visible* text of that hyperlink to the terminal width while every emission carries the whole URL in its parameter. The backend already knew this (`commands/auth_token_commands.rs`); the frontend did not. - `urlDetector` now reads OSC 8 targets out of the raw buffer before stripping, filtered by a port of `usable_sign_in_link`, and tags every candidate with its provenance. - The prompt slot gained `supersedes`: better provenance always wins, worse never does, and between equals only a candidate that *extends* what is showing may replace it. That last rule is `extendsUrl`, factored out of `pickSignInUrl` rather than copied — same rule, same reason, one implementation. - `flatten` splits on a bare `\r` as well as on `\r?\n`, so a `\r`-repainted TUI frame no longer inflates a line past the width and suppresses a join that should have happened; and the width is now sampled at `feed()` rather than read at `scan()`, so a resize inside the 300 ms debounce cannot reassemble 80-column text against a 120-column rule. Also corrects the comment claiming `acquire_claude_token` enables the auth bridge. It deliberately does not, and the module comment in `auth_token_commands.rs` explains at length why not. The auth bridge toggle ---------------------- `setAuthBridgeEnabled` and `getAuthBridgeStatus` had zero call sites: the Rust was complete, the IPC wrapper shipped, and there was nowhere to click — so the docs told users to "enable the Auth Bridge" for a switch that did not exist. `AuthBridgeRow` is that switch, in Config → Runtime. It deliberately does not go through the tab's stopped-only save: the dedicated command exists so the bridge can be flipped while a login is hanging in a running container, which is the only moment anyone reaches for it. It also subscribes to `auth-bridge-changed`, which the poller has been emitting to nobody — so a host port the bridge could not take was a completely silent failure, indistinguishable from a login that hung. `tunnel.rs` promotes the best-effort `::1` bind failure from debug to a warning recorded on the port. Half-bound is the failure mode that looks like success: the status says bridged, and a client that resolves `localhost` to `::1` without falling back is still refused. Finally, for a recognised Anthropic sign-in URL the toast now leads with "In container" and demotes the host "Open". The callback listener is inside the container, so the container-side browser closes the loop with no host round trip and no auth bridge; the host button stays as the fallback. Ordinary URLs are unchanged. Tests: 402 frontend (was 359), 285 Rust (unchanged). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 08:31:39 -07:00
shadow-testandClaude Opus 5	15e05e2197	Turn the Files tab into a real file manager ... Rename, an in-app viewer for text and images, host-to-container drag and drop, New folder, keyboard operation — plus the pre-existing bugs the new surface would otherwise have been built on top of. New Tauri commands (file_commands.rs, registered in lib.rs): * rename_container_path — `mv -n -- <from> <parent>/<name>` through exec_oneshot_as, so the *exit code* is checked. exec_oneshot discards the status and interleaves stderr into stdout, which would have made a permission failure look like a success. `mv -n` on its own is not enough either: GNU coreutils makes its refusal to clobber silent and exits 0, so an explicit `test -e` on the destination is what turns a name clash into an error the user sees. `mv`'s own words are surfaced, since renames outside /workspace legitimately fail on permissions. The new name is validated in Rust (no `/`, no NUL, not "." / ".." / empty, ≤255 bytes) — it is user text going into argv, and a name with a separator would be a move rather than a rename. * read_container_file — exact bytes via Docker's archive endpoint, returned as base64. Deliberately not exec_oneshot, which runs every chunk through String::from_utf8_lossy and merges stderr, so it would corrupt any non-UTF-8 file and could splice diagnostics into content. Base64 rather than Vec<u8> because Tauri serialises a byte vec as a JSON number array. Capped and truncation-reporting; the caller picks the cap (images get 5 MiB against text's 1 MiB, being the kind that blows a text-sized budget) and Rust clamps it to 8 MiB regardless. * create_container_directory — `mkdir` without -p, so a clash is an error rather than a silent success. Named for its siblings rather than the bare `create_directory` in the brief. The tar-extraction half of download_container_file is now the shared fetch_container_file() both commands use, and it abandons the transfer once a capped read has what it needs. Frontend: * Single click selects, double click opens. Directory navigation moved onto double click too — a single click used to navigate, which made it impossible to select a directory in order to rename it. Rows are now focusable and the table is a real `grid`: Enter opens, F2 renames, arrows walk the rows. No outline suppression; the global :focus-visible ring is what shows focus. * FileViewerModal (built on ui/Modal, the only correct dialog) renders text in a <pre> and images from a revocable blob: URL. tauri.conf.json's img-src had neither `data:` nor `blob:`, so an in-app image was blocked by CSP; `blob:` is added — revocable, and no megabytes of base64 in the DOM. The asset protocol stays disabled. Anything else gets a "Save to host" state instead of a broken preview, decided by extension and then by sniffing the bytes for NUL. * Host drag-and-drop uses Tauri's native onDragDropEvent, mirroring TerminalView: HTML5 ondrop carries no paths and is blocked in the webview on Windows by dragDropEnabled, which the terminal needs. The listener is window-wide, so it routes by hit-testing the payload position (physical pixels, hence the devicePixelRatio divide) against the pane's rect — a hidden pane has a zero-size rect and never matches, which is what keeps this and the terminal's listener apart. enter/over/leave drive a drop highlight. * Per-row Download is now "Save to host…"; directories no longer offer it. Pre-existing bugs fixed: * Uploaded files landed root:root with a 1970 mtime. tar::Header::new_gnu() zeroes uid/gid/mtime and Docker honours the header verbatim, so uploads were not writable by `claude`. All four single-file tar builds now go through build_single_file_tar() with the container user's ids, read from the container because entrypoint.sh remaps them to the host user on Unix and deliberately does not on Windows. * Symlinked directories could not be opened: `find -printf '%y'` reports `l`. The listing now prints `%Y` as well, so is_directory dereferences and a new is_symlink carries what `%y` used to say. The row labels the link. * upload_file_to_container had no size cap and did a synchronous fs::read on an async worker. Now 256 MiB (matching the terminal drop path) with the read and tar build in spawn_blocking, and the host mtime preserved. * A directory passed to upload reached fs::read and produced an opaque "Is a directory". Rejected with an explanation instead — recursive upload is a larger feature than this panel needs. * download_container_file wrote the *first tar entry*, so downloading a directory silently produced a garbage file. Non-regular entries are now an explicit error. Tests: 46 new (33 frontend across FilesTab, useFileManager and filePreview; 12 Rust covering the find-output parser and the rename validator, neither of which had any). 405 frontend / 297 Rust, both green. No drag-out dependency was added — tauri-plugin-drag is not introduced and OS drag-out is not attempted; that stays deferred, with "Save to host…" as the way files leave the container. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc	2026-08-23 08:30:48 -07:00
shadow-testandClaude Opus 5	48d0c3249a	Fix two bugs in last round's fixes, and stop --full hiding the Docker host ... Round 3 found defects in code written an hour earlier. Both reproduced. **The handshake poll accepted empty output as a completed handshake.** `[ "$(… | awk '{print $2}')" != 0 ]` is *true* when `wg show` prints nothing — which it does when the interface has no peer, and when the interface is gone (that message goes to stderr). `until` suspends `set -e` and `pipefail`, so nothing else caught it. The poll added last round to make "success without a tunnel" impossible produced exactly that. Now requires a number greater than zero, and waits 20s rather than 10 so a slow link is not rolled back needlessly. **`down` still sat above the key registration.** Last round moved it below the token and server-list fetches but not below `addKey`, which is the most failure-prone of the three — one gateway, by CN, pinned certificate. So a refused registration still tore down a working tunnel. It now runs after the last fetch; the key is generated before but written after, since `down` deletes it. SKILL.md said "after every network fetch has succeeded", which was false; corrected. **`up --full` made `host.docker.internal` unresolvable — and `status` said DNS was fine.** That name is answered only by the resolver being replaced; it is not in `/etc/hosts`. `gateway.rs` hands it to every container for the LiteLLM gateway, and Ollama and custom endpoints default to it, so an agent running `up --full` silently removed the project's model backend. The route was already excluded; only the name was lost. Now resolved with the old resolver and pinned into `/etc/hosts` before the swap, restored on teardown, and `status` probes it — PIA answers public names happily, which is precisely why probing only `api.anthropic.com` reported "ok". Documented as Trap 4. **The rollback could abort halfway.** The trap's `{ … }` is not exempt from `set -e`, and `down`'s `cat`/`tac`/`rm` had no `|| true` — so one failure left the interface up with all traffic captured, after printing "rolling back". `down` now runs under `set +e`, the trap tolerates its failure, and the interface is deleted *first*, since that removes every route pointing at it. **The account password had a real argv window.** curl does blank `-u`, but only once running: sampling /proc/<pid>/cmdline caught the plaintext in 2 of 400 tries, between exec and the overwrite. Small, but it is the permanent password and the token already had the fix. Moved onto the same stdin config — 0 of 400. Review reported this as a 25-second exposure; that was a wrapper's argv, not curl's. entrypoint: the skill install stages into `$_dest.new` and swaps, so a failed copy leaves the previous copy intact instead of a truncated SKILL.md and no script, root-owned, on a persisted volume. Verified against a size-limited filesystem. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 17:15:52 -07:00
shadow-testandClaude Opus 5	5b96ad4823	Stop a failed up from tearing down a working tunnel ... The previous commit moved `down` to the top of `up` to fix a resolv.conf idempotency bug, and in doing so put it *before* every network fetch that can fail. Re-review caught it and I reproduced it: with a tunnel up, an `up` that fails on bad credentials left `pia0` gone and traffic silently back on the real address, while the error talked only about credentials. A privacy regression introduced by a correctness fix. `down` now runs after the token, server list and key registration have all succeeded — nothing above that line touches the network stack — and still clears the stale backup it was added for. Everything after it is covered by a rollback. Note this is an EXIT trap with a flag, not `trap ... ERR`: my first attempt used ERR and did not fire at all, because ERR is not inherited by shell functions without `set -E`, so a failure inside add_route missed it, and `die` exits explicitly, which is not an error. Verified by forcing a route collision — the tunnel is torn down and DNS is intact, where before the fix it was left half-configured with DNS dead. Also from the review: - **The private key lived on disk for the whole life of the tunnel.** `commit_container_snapshot` blanks env vars, never files, and nothing tears the tunnel down before a recreate or migrate — so the `down`-time cleanup never covered the path that put a key in a snapshot in the first place. It is now deleted the moment `wg set` has read it; the kernel keeps its own copy, verified by checking the interface still works afterwards. - **`up` claimed success without a handshake.** An unreachable peer still routes — into a black hole — so `up --full` could exit 0 having pointed all traffic and resolv.conf at a peer that never answered, with `status` printing "mode: full tunnel". Now polls for a handshake and rolls back if none arrives. - **`status` needed root and did not check.** `wg show` fails unprivileged and was swallowed, so an unprivileged run printed "no tunnel up" and then "mode: full tunnel" in the same breath. An agent reading the first line would re-run `up` — which, before the fix above, destroyed the tunnel it failed to see. - **The killswitch bullet was false.** It said `iptables` is not in the image; it is, so a killswitch is buildable. It stays unbuilt because it would cut Claude Code's own API traffic — an honest reason, unlike the previous one. - `install_feature_skill` rejects path-traversal names, not just blank ones — verified `../skills` would have deleted the whole skills directory including Mission Control's — and reports `mkdir`/`cp` failures instead of printing a success line regardless. - The usage text ended by printing `set -euo pipefail`, off by one line. - HOW-TO-USE claimed "the container says so on start". entrypoint prints to PID 1's stdout, which no terminal or UI surfaces — `docker logs` appears nowhere in the repo. Now points at the migration pre-flight, which does. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 17:11:50 -07:00
shadow-testandClaude Opus 5	dcb13d23ea	Fix what review found in the skill: five real defects ... Adversarial review of #29 found bugs I confirmed by reproducing each one. **Every hand-written error message was unreachable.** `tok=$(curl ...)` is a plain assignment, so `set -e` acts on the command substitution before the following `|| die` can run. A wrong password produced exit 22 and no output at all — the most likely way this gets used wrongly, and the least explained. All four captures now go through a `run` helper that takes a *description* rather than echoing the command, because one of them carries the account password in `-u`. **`up` was not idempotent, and the second run destroyed DNS.** The resolv.conf backup was copied unconditionally, so `up --full` twice overwrote the good backup with PIA's own resolvers; the later `down` then "restored" those and left the container with no working DNS and no way back. `up` now runs `down` first. Verified: two `up --full` runs, then `down`, and the backup still holds the original 192.168.65.7. **An empty gateway produced total connectivity loss, reported as healthy.** `$gw` was never validated and `add_route` swallowed every failure to /dev/null. The two half-routes need no gateway and would succeed, so the tunnel captured everything while the exclusions keeping DNS and the Docker host reachable silently did not exist — and `status` still printed "full tunnel". Routes are now fatal on failure, and a via-less default (`$3` is the literal "eth0") is rejected. **The PIA session token was in the process arguments** — confirmed in `ps` and /proc/*/cmdline, a ~24h bearer credential for the account readable by anything in the container. It now goes to curl on stdin as a config. Verified: 60 polls across a full `up`, zero sightings. **The preflight diagnosed the wrong kernel module.** It checked /dev/net/tun and blamed the tun module, but kernel WireGuard is a netlink interface and does not use it — verified by creating one with NET_ADMIN and no tun device. The check is dropped (the container could not have started without the device anyway) and `ip link add` now reports the real dependency. Also: a full tunnel with no DNS servers from PIA used to warn and carry on, which is a tunnel leaking every lookup while reporting itself healthy — now fatal. `down` validates the backup before restoring it, so a truncated one cannot leave the container with no resolver at all. `wg.priv` is shredded on teardown and created under umask 077, because /run rides `docker commit` into the snapshot image. A mistyped `up --ful` is rejected instead of silently giving a test route. entrypoint: `install_feature_skill` gets `local`, a blank-name guard (the disabled branch would otherwise `rm -rf` the whole skills directory under a persisted volume), `-e`/`-L` so a leftover *file* at the destination is cleaned up, and a chown of the parent so `claude` can still add skills of their own when Mission Control is off. When the base image predates the skill it now says so instead of returning silently — and `/opt/triple-c-skills` joins FEATURE_PROBES so the migration pre-flight reports it. Docs corrected to match: neither half reaches an existing project without a migration. `vpn_env_var` extracted and tested, pinning the property the whole removal path rests on — that the variable is emitted as 0 rather than omitted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 17:11:50 -07:00
shadow-testandClaude Opus 5	7a8bbcbef7	Report which exit is which, instead of one ambiguous "public IP" ... `status` probed https://1.1.1.1/cdn-cgi/trace and printed the answer as "public IP". In test mode 1.1.1.1 is the *only* address routed into the tunnel, so that line reported a PIA exit while every other packet left directly — a test tunnel reading exactly like a full one. Found on a live container: default route still via eth0, one 1.1.1.1/32 route through pia0, and the old status line claiming a PIA public IP. This is a plausible route to concluding the VPN is on when it is not, which is close to the confusion this skill exists to prevent. Status now names the mode and, in test mode, prints both exits with the real address called out. 1.0.0.1 serves the same trace endpoint as 1.1.1.1 and is never routed into the tunnel, so the direct exit can be probed without DNS. Verified against all four states: no tunnel, test mode on a live tunnel that was already up, full tunnel, and after teardown. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 17:11:23 -07:00
shadow-testandClaude Opus 5	3bd3caa101	Ship a pia-vpn skill with the VPN support toggle ... The toggle grants CAP_NET_ADMIN and /dev/net/tun and stops there, which users reasonably read as "turn the VPN on" — the gap between the two is the reported bug that the default network does not route through a VPN. Close it by giving the container an agent-usable way to build the tunnel, rather than leaving each project to rediscover it. container/skills/ is baked to /opt/triple-c-skills and installed into ~/.claude/skills/ by entrypoint.sh from VPN_SUPPORT_ENABLED, mirroring how Mission Control installs its own. Staged under /opt because ~/.claude is a volume mount that would mask an image copy from first start. Three details that are not incidental: - The variable is sent as 0 rather than omitted when off, because ~/.claude persists: entrypoint has to be *told* to remove a skill left by an earlier run with the toggle on, and an absent variable cannot say that. A stale skill is worse than none, since it instructs an agent to use a capability the container no longer has. - It is reserved in RESERVED_ENV_EXACT alongside MISSION_CONTROL_ENABLED, or a custom env var of the same name could claim the skill without the capability behind it. Covered by a test. - The skill is re-copied on every start, rm -rf'd first, so fixes reach existing projects and files dropped from a later version do not linger. The skill itself carries the three things that are easy to get wrong: that a full tunnel captures the Docker resolver and takes DNS down with it, that an IP-literal health check cannot see a dead resolver, and that no tunnel survives a restart while /run state riding the snapshot makes it look as though one did. It also states what it deliberately does not do — no killswitch, no autostart — so an agent proposes those as decisions rather than improvising them. pia-wg.sh preflights CAP_NET_ADMIN by capability bit rather than letting the first `ip` call fail with a bare EPERM that points nowhere near the setting that needs changing. Credentials stay in a file (~/pia-creds, PIA_CREDS to override) rather than the environment, where docker inspect and every process in the container would see them. Tested: install/refresh/remove/no-op paths of install_feature_skill against the real function; preflight with and without the capability; and a full up --full / down round trip, confirming DNS via PIA's resolvers, api.anthropic.com reachable through the exit, and routes and resolv.conf restored on teardown. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 17:11:23 -07:00
shadow-testandClaude Opus 5	5dd1ab5217	Stop resting the iptables case on a kernel config I cannot verify ... Round 3 argued the macOS rationale is stale: that Docker Desktop no longer builds from linuxkit/linuxkit and has enabled nft_fib_ipv4 since 4.35. I could not confirm or refute that from a Linux host — searching turned up no version matrix either way. But the decision does not depend on it, and the comment should not have implied it did. `xt_CONNMARK`, which the iptables path needs, was present in every kernel config examined. `nft_fib_ipv4`, which the nft path needs, was absent from the config read here and may be present in current Docker Desktop. That asymmetry is the actual argument: nftables' viability varies by Docker Desktop version in a way nobody here can pin down, iptables' requirement did not vary anywhere it was checked. If nft_fib_ipv4 is present this costs 1.6 MB and nothing else; if it is absent it is the difference between a working full tunnel and none. Rewritten to say that, and to say plainly what is verified versus assumed — this is the third round in which the previous round's central premise did not survive, and a confidently-worded paragraph is what the next round inherits. Also from review: - CLAUDE.md still said "`iptables` is deliberately absent", the opposite of what this PR now does, contradicting the Dockerfile and both other docs. - The Dockerfile referenced "the pia-vpn skill", which does not exist on this branch — the third forward reference of that kind, now gone. - "full tunnels work on native Linux, Mac and WSL2 6.6" was unconditional and contradicted ten lines later by the `DNS =` concession, which stops them on every platform. Reordered so the DNS hurdle is named as the first one. - The WSL2 gap was written as a permanent platform limitation. It is a stale install: `wsl --update` moves the host to a current kernel that has the symbol. That remedy was missing from the user-facing doc. - The migration probe label carried an internal comma, which `joinFeatures` renders into a comma-joined list. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 17:11:17 -07:00
shadow-testandClaude Opus 5	92d64cf252	Ship iptables, not nftables — nftables forfeits macOS ... Re-review overturned the previous commit's package choice, and verifying it proved the reviewer right. `wg-quick` picks nft *unconditionally* when it is present (`type -p nft`, line 241), so installing nftables makes the iptables path unreachable. Its nft ruleset then needs a third expression family the iptables path does not. Isolating the rules on this host, the two connmark rules install fine and this is what fails: nft add rule ... fib saddr type != local drop Error: Could not process rule: No such file or directory That decides it, because of how the hosts differ. LinuxKit's kernel config — Docker Desktop for Mac, identical on x86_64 and aarch64: CONFIG_NETFILTER_XT_CONNMARK=y <- the iptables path works # CONFIG_NFT_FIB_IPV4 is not set <- the nft path does not So nftables would have broken the platform it was added to fix. With iptables, full tunnels work on native Linux, Docker Desktop for Mac, and WSL2 from 6.6. Costs 7,203 kB rather than 5,614 kB on amd64. That also means the mechanism the previous commit documented was wrong: with nftables installed `xt_CONNMARK` is never consulted, and the real blocker on that path is `nft_fib_ipv4`. Rewritten around what actually fails. A second failure neither round had found: `wireguard-tools` only *Suggests* `openresolv | resolvconf`, so neither is installed, and every provider's stock config has a `DNS =` line. That fails in `set_dns()` — before any routing — so it takes split tunnels down too, contradicting what this PR previously claimed: [#] resolvconf -a sp -m 0 -x /usr/bin/wg-quick: line 32: resolvconf: command not found EXIT=127 Not fixed, deliberately: `openresolv` has no installation candidate on noble, and `resolvconf` resolves only by pulling in systemd-resolved — a resolver daemon and systemd units, into a container with no systemd. Documented instead. Smaller corrections from the same review: - the size caveat blamed ~209 kB of libelf1t64; for this package set the real over-count is libelf1t64 + netbase. Restated, and arm64 now given against the real base rather than left as a bare-ubuntu figure. - the manual-install fallback omitted `iproute2`, so it left the user without `ip` — the command the tunnel needs most. - "Without it" had been orphaned from its antecedent by inserted paragraphs and read as referring to configuring a tunnel. - the migration probe said "VPN support", presenting VPN as a feature gained to users who never enabled it. Now names the tools and the toggle. - "What's Inside the Container" gains a row; the key-material-in-snapshot hazard was in CLAUDE.md only, and is the one genuinely user-facing warning here. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 16:24:07 -07:00
shadow-testandClaude Opus 5	ab2c75d0b2	Ship a firewall backend, and correct three claims review disproved ... Review of #28 found the iptables exclusion was justified by a false premise, and I confirmed it: `wireguard-tools` declares `Recommends: nftables | iptables`, `--no-install-recommends` strips it, and `wg-quick`'s add_default() shells out to a firewall backend with no `type -p` guard. Measured on the image as this PR shipped it: [#] iptables-restore -n /usr/bin/wg-quick: line 32: iptables-restore: command not found wg-quick EXIT=127 That fires for `AllowedIPs = 0.0.0.0/0` — every stock full-tunnel config from every provider — not for a desktop client's killswitch as the comment claimed. Split tunnels are unaffected. Ship `nftables` rather than `iptables`: wg-quick prefers it (`type -p nft`, so with both installed iptables is dead weight), it is first in the package's own Recommends, and it is half the size. The review's proposed fix stopped there; it does not hold. Adding nftables does not make wg-quick work on this host, and neither does iptables: Warning: Extension CONNMARK revision 0 not supported, missing kernel module? `Table=auto` routes by fwmark and needs xt_CONNMARK from the *host* kernel. WSL2 has none and containers have no /lib/modules to load one from. So this fixes native Linux and Docker Desktop for Mac — which other WHP users are on — and cannot fix Docker Desktop for Windows, where the answer is to add routes with `ip route` directly. Documented rather than left to be rediscovered. Also from review: - "`ip` and `wg` are always present" was false. A project keeps the base image it was first built from, so this reaches new projects only. Reworded to match the wording already used for the Playwright libraries, and `/usr/bin/wg` added to FEATURE_PROBES so an existing project is *told* it is missing VPN tooling and prompted to migrate, rather than finding out via `wg: command not found`. - "no client is installed" contradicted shipping `wg` four lines earlier. The true claim is that no tunnel is configured or started. - The size figure measured against bare ubuntu:24.04, which over-counts by the ~209 kB of libelf1t64 the real base already has, and covered one arch. Now measured against the current base on amd64 and stated for arm64 too, per the standard CLAUDE.md sets for the Playwright layer. - `/run` persistence conflated two mechanisms: same-container files on a stop/start, `docker commit` on a recreation. Both stated, plus the corollary that key material written to /run ends up inside a snapshot image — observed, a `wg.priv` was already sitting in one. - The DNS bullet presented a Docker Desktop address as the general case. Now leads with the mechanism, notes 127.0.0.11 on a user-defined network is unaffected, and adds the two things the advice omitted: a resolver the tunnel can reach (or it leaks every query), and pinning the endpoint via the old gateway (or the tunnel routes through itself). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 13:55:44 -07:00
shadow-testandClaude Opus 5	00937745f7	Ship the tools the VPN toggle grants capability for ... `vpn_support_enabled` hands a project CAP_NET_ADMIN and /dev/net/tun, and the image then contains no `ip` and no `wg` — a capability with nothing able to exercise it. Bake `iproute2` and `wireguard-tools` (~4.3 MB with deps). They belong in the image rather than a runtime install for the reason the Dockerfile already gives for the Playwright libraries: the writable layer is lost on base-image migration. A hand-installed `wg` works until an upgrade and then vanishes, which presents as a tunnel that will not come up rather than as a missing package. One project only had `ip` at all because MariaDB pulled in iproute2 as a transitive dependency. `iptables` stays out. Only a desktop client's killswitch wants it, and those clients need a GUI the container cannot provide. Also correct three things the docs left users to discover: - the toggle grants capability and routes nothing, which is being reported as the default network "not routing through the VPN automatically" - no tunnel survives a restart, and `/run` state riding the snapshot makes it look as though one did while traffic goes out the real address - a full tunnel captures the Docker resolver, which sits outside the container's subnet, and takes DNS down with it — Claude Code then reports a connection failure because it cannot resolve api.anthropic.com, and a health check aimed at an IP literal passes throughout Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-17 12:38:37 -07:00
shadow-testandClaude Opus 5	2b35aa8c16	Explain a missing tun device where the failure actually happens ... Review caught that the device guard was wired to the wrong call. The daemon does not resolve `--device` at create: verified against Docker 29.7, `docker create --device /dev/does-not-exist` succeeds and prints an id, and runc only resolves the device — and validates sysctls — when it builds the container. So on a host with no tun module the create returns fine and `start` fails, which means the explanation never ran and the user saw the raw daemon string naming a path they would go looking for on the wrong machine. The unit tests fed the create-side string straight in, so they confirmed a function no real failure could reach. Move the guard onto `start_container`, covering create as well in case a future daemon checks earlier. It no longer takes `vpn_support_enabled` — `start_container` has a container id and no project, and nothing else in Triple-C ever requests a device, so an error naming /dev/net/tun is unambiguous on its own. The test now uses the daemon's verbatim message via bollard's real Display format. Also from review: * Soften the security claim. Docker does not enable user-namespace remapping by default, so this is a real CAP_NET_ADMIN in the initial user namespace with only the network namespace confining it. It cannot touch host interfaces, but "confers no authority outside the container" was too strong: within its namespace it can set promiscuous mode and add addresses, routes and NAT on the shared docker0 segment, which puts sibling containers — the LiteLLM gateway among them — within ARP-spoofing reach, and it can flush netfilter rules sandbox mode may rely on. Said plainly in the code, CLAUDE.md and HOW-TO-USE. * Drop Tailscale from the list of clients needing this. Its --tun=userspace-networking mode needs neither the capability nor the device, and listing it invites granting NET_ADMIN for nothing. * Say in the toggle's own hint that changing it recreates the container, matching how every other recreation-triggering setting is labelled. The tab's generic "stop the container first" chip does not tell the user what is about to happen. * Add RuntimeSection tests: saves on, saves off explicitly rather than dropping the key, reflects state, is disabled while running, and carries the recreation warning. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-14 08:15:05 -07:00
shadow-testandClaude Opus 5	65a3d4eb29	Let a project's container run a VPN client ... A VPN client installed in a container today starts, runs, and then hangs until its connection times out. Nothing reports an error: a default container has no /dev/net/tun to open and no CAP_NET_ADMIN to add an interface or a route with, and clients surface that as a generic timeout rather than a permissions failure. Add an opt-in per-project "VPN support" switch granting the three things a tunnel needs. They are useless individually, which is why vpn_host_config() defines the set in one place and the tests assert all of it: * CAP_NET_ADMIN — Docker's default bounding set has net_raw but not net_admin, so a client can ping but never connect. * /dev/net/tun — passed through from the host so the kernel's tun module backs it, rather than mknod-ed inside. * net.ipv4.conf.all.src_valid_mark — WireGuard's wg-quick sets this and cannot from inside a container, /proc/sys being read-only, so its handshakes are dropped by reverse-path filtering. Off by default and deliberately opt-in: NET_ADMIN lets anything in the container reconfigure that container's network stack. It is namespaced — no authority over the host's interfaces or any other container. Capabilities and devices are fixed when a container is created, so this is container state and takes the label-and-compare treatment. triple-c.vpn-support is written unconditionally, false included, for the usual docker commit reason: a true stamped once would ride the snapshot image into every future container and make the switch impossible to turn back off. A missing label reads as false and off is byte-identical to today, so no existing project is churned. Requesting the device fails at creation when the host kernel has no tun module, which would otherwise surface as a project that simply refuses to start. explain_create_failure() rewrites that one error to name the switch and the Docker-Desktop-VM-versus-your-machine distinction, and leaves every other failure untouched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-14 08:06:39 -07:00
shadow-testandClaude Opus 5	3741e0fef5	Number releases by the highest one already published, not by drift ... The Linux release upload failed with a bare "exitcode '1'" and no output. The cause was not the upload: compute-version handed it a version that had already been released three days earlier. The patch number was `git rev-list --count <highest tag>..HEAD` — how far HEAD has drifted from whichever tag sorts highest, which resets to zero every time a tag is cut. It is not a counter, and the published history is what the old formula returned at each point: v0.4.0 -> 3 commits -> v0.4.3 looked fine v0.4.3 -> 4 commits -> v0.4.4 fine by luck, 4 > 3 v0.4.4 -> 2 commits -> v0.4.2 went backwards v0.4.4 -> 6 commits -> v0.4.6 jumped, skipping .5 v0.4.6 -> 3 commits -> v0.4.3 already taken So the line published 0.4.0, 0.4.3, 0.4.4, 0.4.2, 0.4.6 in that order, never used 0.4.1 or 0.4.5, and then came back round to 0.4.3. The patch is now one past the highest already used. Suffixed tags count towards that: create-tag is skipped whenever a platform job fails, so a run can publish v0.4.7-mac and never create the plain v0.4.7, and reading only unsuffixed tags would hand the same number out twice. A commit that is already tagged reuses its own tag, so re-running a build does not mint a version. Reusing a number was doing real damage, not just failing. macOS and Windows delete-then-upload each asset, so they took the duplicate in their stride and rewrote v0.4.3-mac and v0.4.3-win — public since Aug 11 — with today's binaries. Linux is the only platform that failed, and failing was the correct outcome; its v0.4.3 assets are the only ones still original. Linux also gets the idempotent get-or-create the other two already had, plus `set -euo pipefail` and `-fsS`. Its `curl -s` with no `-f` is why a 409 produced no diagnostic at all: the HTTP error was swallowed, the id grep came back empty, and the step died without ever printing why. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-13 22:57:36 -07:00
shadow-testandClaude Opus 5	84a67fcd0d	Stop an empty base-image label from silencing the migration notice ... A project can be out of date and say nothing about it, in two ways that compound: the lineage lookup treats "unknown" as an answer, and the fallback that exists for unknown lineage disappears when its probe fails. `create_container` always writes triple-c.base-image-id, even when the value is unknown — deliberately, so an inherited image label cannot ride a snapshot forever. That makes Some("") the ordinary reading from a container whose lineage was never established. The lookup filtered for emptiness only on the final result, so that empty string satisfied the container branch and skipped the snapshot entirely: a snapshot that had recorded a real lineage was never consulted, and the project reported "unknown" with the answer one lookup away. Each source is now filtered before it can answer, in pick_recorded_lineage, which is a plain function so the case has a test that fails against the old logic. A genuinely pre-label project stays unknown, and should: its ancestor is not knowable, and inventing one would make it look permanently current. The probe is the intended signal for those — but if the probe failed, get_container_staleness returned early with nothing populated, the banner found no gaps and rendered null, and the probe_error it already knew how to display sat behind a gate that returned before reaching it. Silence there is indistinguishable from "up to date", and it is likeliest for the oldest and largest projects, whose manifests are the ones apt to exceed the inspection limit — one real project measured 6.93 MB against an 8 MB cap. An unknown-lineage container whose probe failed now says the check could not be completed, with the reason, under the tone that means unresolved rather than the one that means something is wrong. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-13 21:24:51 -07:00
shadow-testandClaude Opus 5	f3cc1c4c17	Stop Playwright setup from deleting the package it just installed ... Setting up the browser view failed on every container, and re-running it reproduced the same broken state, because the setup destroyed its own work. `install_packages` ran two `npm install --no-save` commands into /workspace, which has no package.json. With no manifest, npm treats the command line as the whole statement of what the tree should contain and prunes the rest, so installing `playwright` second removed the `@playwright/cli` installed first: "removed 3 packages", leaving an empty node_modules/@playwright/ behind playwright and playwright-core. That empty directory is exactly what the pane then reported as missing. The second install now names both specs; the first one is already present, so it costs nothing and is only there to stop npm pruning it. Two failures were waiting behind that one: Nothing in the tree ever configured the browser, so playwright-cli fell back to channel `chrome` — system Google Chrome — with the Chromium sandbox on. These containers forbid unprivileged user namespaces, so it aborted with "Failed to move to new namespace ... Operation not permitted"; on a base image without Google Chrome the same default failed as "Chromium distribution 'chrome' is not found". entrypoint.sh now seeds ~/.playwright/cli.config.json on every start, which is the only way to reach existing projects: ~/.playwright is inside the home volume, so an image copy would reach new projects only. The launch check passed for a configuration the viewer never uses. It launched bundled chromium with no channel, which resolves to chromium-headless-shell, while the viewer's config pins chrome-for-testing — the full chromium build, a separate download. A container could pass every check and still fail in the pane with 'Browser "chrome-for-testing" is not installed', which is what a stale chromium-1217 against a wanted chromium-1237 did. Chromium is now verified on both channels, the sandbox setting is stated rather than inherited from a default, and a failure names the channel. triple-c-playwright-heal repairs all of it on a container that is already broken, including the missing socat that makes the pane report "127.0.0.1 sent an invalid response" while the container side is perfectly healthy. It verifies by launching a browser rather than trusting the preceding steps — which is how the stale-revision case was found — and lives in /usr/local/bin so a fix to it can still reach an existing project. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-13 21:24:37 -07:00
shadow-testandClaude Opus 5	fa4940dd7d	Say when a scheduled task is running ... A run is detached — cron has no terminal, and the app fires it as a detached exec — so triggering one and watching the log was indistinguishable from triggering one that died. Worse, `claude -p` writes its answer in a single burst at the end, so a healthy run shows nothing but its log header for as long as it is thinking. The honest reading of the old UI was "it stalled". triple-c-task-runner now publishes a state file per run (pid, start time, log path) and removes it from an EXIT trap. flock remains what actually prevents overlapping runs; this is purely observability, so every reader verifies the pid rather than trusting the file — a container stopped mid-run cannot fire a trap, and a task stuck on "running" forever would be a worse lie than no indicator at all. Stale files are cleared on read. On top of that: - `list` grows a status column: "running 4m12s" or "idle". - `status [--id] [--watch]` answers "is it still going?" directly, with elapsed time and the tail of the log when there is any output yet. - `run` streams the log instead of blocking silently, and refuses to start a task that is already running. - The Automation tab marks a running task, disables its Run now button, and polls while anything is in flight — including the second or two between firing a run and the runner registering it, which is the exact window that used to read as dead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-12 06:33:49 -07:00
shadow-testandClaude Opus 5	9027fa9ad4	Stop the scheduler handing Claude root's HOME ... Every scheduled task failed with "Not logged in · Please run /login" while the container's OAuth credential sat there, valid, the whole time. The entrypoint snapshots the environment into ~/.claude/scheduler/.env so cron jobs get more than cron's minimal env. It runs as root, and HOME was in the capture list, so the file recorded HOME=/root. The task runner then sources that file with `set -a`, overwriting the HOME cron gave the job. `claude -p` looks for its credential under $HOME, finds no /root/.claude, and exits 1. Logging still worked — SCHEDULER_DIR is expanded before the sourcing — which is why this presents as a well-formed log of a task that never authenticated. Drop HOME from the captured set and write it explicitly instead; cron does still need one. Then restore HOME across the source in the task runner too: .env lives on the home volume, so every project created before this ships keeps a stale copy of it until its container restarts, and the runner is what has to survive that. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-12 06:13:59 -07:00
shadow-testandClaude Opus 5	5f990dd28b	Sweep the snapshot commits recreation leaves behind ... Every recreation commits the container to triple-c-snapshot-{id}:latest and moves that tag; the image it pointed at keeps its layers and loses its name. Nothing deleted those, so they accumulate — measured on one real host, 7 orphans holding 7.4 GB, three of them from a single day's work. `sweep_orphaned_snapshots` removes them, under two conditions that are the whole safety argument. Untagged: every image the app depends on carries a tag, so a project's live `:latest` and a migration's `pre-migration-*` rollback pin cannot match the filter at all. And labelled `triple-c.managed=true`, which `docker commit` copies from the container onto the image — the user's own dangling images are not ours to delete. Removal is unforced on top of that, so Docker refuses while any container is still built from the image, including the stopped containers of projects that are not running; those are counted and left for the next sweep. It runs after a recreation, which is when the orphan it just made becomes removable, and after a migration is accepted, which is the moment dropping the pin turns the pre-migration snapshot into an orphan. Both detached: this is housekeeping, and a full disk beats a project that will not start. Each sweep clears every orphan it finds, so recreations that predate it are cleaned up too. The label string is now a constant rather than four literals, and a test pins both filter conditions in place. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-11 19:01:20 -07:00
shadow-testandClaude Opus 5	9b2f4fe79f	Give the env var its value box back, and stop labelling the secret ... Two separate faults, both reachable from one screenshot of the Global Environment Variables editor. The value input was collapsed to a sliver, so a variable looked like it had lost its value. `inputClass` carries `w-full`, and the `w-2/5` on the key input did not beat it — class-attribute order is not what resolves that conflict, stylesheet order is. The key therefore asked for the whole row, and the value input, whose `flex-1` gives it a basis of 0 and only the leftover space, got almost nothing. Widths now live on wrapper divs, where nothing competes with them. The fingerprint that detects custom-env changes was a plaintext `KEY=VALUE` join, and it is written as the `triple-c.custom-env-fingerprint` label. Labels are readable by anything on the host via `docker inspect`, `docker commit` copies them onto the project's snapshot image, and the recreation check logs both sides on a mismatch — so an API token set as a custom variable was published to all three. It is hashed now, exactly as `triple-c.git-token-hash` already was. Empty stays empty, so "nothing configured" still reads as an empty label. Changing the fingerprint format means every project's label mismatches once: expect a single container recreation per project on next start. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-11 15:27:19 -07:00
shadow-testandClaude Opus 5	e9ec2f8e26	Bring the README back in step with the code, and give it a spine ... Four feature commits landed after the last doc sweep without reaching the README, and cc5f691 documented only half of what it shipped. Adds the sections for browser view, corporate CA certificates, base-image migration and the LiteLLM model gateway, and refreshes the Key Files table, which was missing about ten modules. Fixes what had gone stale: the AWS directory mounts read-only at /tmp/.host-aws and is copied in by the entrypoint, not mounted at ~/.aws; Project Home has six tabs, not five; the Automation tab creates tasks as well as running them; Ctrl+Shift+left/right moves the active tab, and tabs can be dragged. Structurally, the sections are now grouped under Containers, Models and Authentication, Bridges to the Host, and Inside a Project, with a contents list and a table pointing at the other docs. Every section from before survives, in the same words where nothing changed — the file was 414 lines of internals with no way in and no map of where anything was. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-11 14:28:52 -07:00
shadow-testandClaude Opus 5	4c962ebd9c	Archive the marks the new icon replaces ... Neither is referenced by the app or the build; branding/archive/README.md says what each one was and why it did not survive the sizes an app icon is actually drawn at. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-11 12:03:28 -07:00