Infrastructure hardening: CI enforcement, supply chain, updater, diagnostics #14

Merged

dtourolle merged 10 commits from chore/infra-hardening into master 2026-08-21 17:33:54 +00:00

Conversation 0 Commits 10 Files Changed 284

+9859 -7300

10 Commits

Author	SHA1	Message	Date
dtourolle	f11f5eddd5	docs(specs): record the Linux native-video compositing spike ... Adds the spike write-up and re-opens finding 2 of playback-backend-unification.md, which concluded that video cannot unify on a native engine because the webview owns the surface. That finding's general form has since been falsified on Android, where native video composites behind a transparent WebView and ships on by default. The evidence behind it was also entirely about foreign-window embedding -- mpv's render API, drawing into a GL context we own inside Tauri's own GTK tree, was never tested. The spike tests that one claim and comes back green on Linux for both X11 and Wayland, bar the Tauri default_vbox() half of G1. Findings 3-6 are deliberately left standing. Finding 3 in particular -- mpv has no adaptive bitrate -- is an independent disqualifier that a green compositing result does not clear, and the spike says so rather than reading as a green light. The next-free-id line moves to UR-079 / IR-033 / DR-219: this branch allocated UR-077 and UR-078 for the updater and diagnostics work, and DR-215 through DR-218 with them, after that line was last written. Authored in a parallel session working in the same checkout; committed here so it travels with the rest of the branch.	2026-08-21 19:06:13 +02:00
dtourolle	f3fa45f742	feat(diagnostics): persistent redacted logging and an exportable bundle ... The app forgot everything it did the moment it exited. The Rust half logged through env_logger to stdout only -- invisible to anyone who launched from a desktop icon, and on Android worse than that: stdout is not logcat, so the backend produced no visible output at all on the platform carrying this project's hardest bugs. The autoplay deadlock, the truncated-stream restart and the background-audio stall were all diagnosed by talking a user through `adb logcat`, because there was no other way to see anything. A panic left nothing behind at all. Logs now go to a size-capped rotating file, to logcat on Android, and to the webview console in dev. A panic is recorded with its backtrace before the process dies. The frontend's messages are forwarded into the same file, so one timeline holds both halves of the app in order -- which is what makes a race between them legible after the fact, and races between them are the expensive bug class here. Redaction runs in the log FORMATTER, not at export time. A credential sitting in a file on the device is already a disclosure; stripping it on the way out would be too late. The exporter redacts a second time to cover files written by builds that predate this. api_key, X-Emby-Token, Authorization, "AccessToken" and Token="..." all reduce to [REDACTED], while host, item ids and filenames are deliberately kept -- a log scrubbed of those is one nobody can debug anything from. Server URLs keep scheme and host and drop any embedded user:pass@. Two things the tests caught that review would not have: - redact_headers recursed on its own output. The replacement keeps the header NAME, so the next call matched the same header forever; the test died with a stack overflow. It is a forward scan now. - The frontend forwarder used `void plugin.error(...)`. `void` discards a promise's value but not its rejection, so in any webview without IPC -- a unit test, SSR, a browser preview -- every log line became an unhandled rejection. 20 of them showed up the first time coverage ran. Each call now attaches a catch. Only info and above cross the IPC boundary: debug is per-tick player state and forwarding it would be thousands of calls a minute for output nobody reads. A failing forwarder never propagates and never prevents the console write. Nothing is transmitted anywhere. The export writes a zip and reports its path; the user attaches it themselves, which is also what keeps this from becoming telemetry. An Android share intent is explicitly out of scope -- it is Kotlin work that belongs with the other native code. The panic hook chains to the previous hook rather than replacing it, because utils/lock.rs installs a silencing hook around tests that provoke poisoned locks on purpose. Spec in docs/specs/diagnostics-and-logging.md; UR-078 / DR-218 / UT-209. Verified: 1079 frontend tests and the coverage gate, 759 Rust tests, clippy -D warnings, svelte-check 0 errors, and cargo check for aarch64-linux-android.	2026-08-21 18:58:57 +02:00
dtourolle	fb72bf3005	docs(ci): drop the runner-health references the amend missed ... The commit that removed .gitea/workflows/runner-health.yml was amended with a git add whose pathspec named the already-deleted file, so the add aborted and only the deletion was staged -- leaving ci-operations.md still describing a scheduled job that no longer exists. The runner section now says what to check by hand and why there is no job: on a single-slot runner a daily job takes the slot and pulls the builder image to run df, and df inside a container does not reliably describe the host disk.	2026-08-21 18:58:39 +02:00
dtourolle	6897b290ed	docs: add the governance and CI-operations files the project never had ... The repo had no SECURITY.md, CONTRIBUTING.md, code of conduct, or issue and PR templates. For a client that handles Jellyfin credentials and ships signed binaries, the missing one that actually matters is SECURITY.md: there was no stated way to report a vulnerability privately, so the only available channel was the public tracker. CONTRIBUTING.md documents the gates as they now stand, including the three ratchets and which direction each is allowed to move, and the two rules that surprise people: bug fixes start with a failing test, and Jellyfin's taxonomy stays in Rust. The bug template asks the three playback questions -- streaming or downloaded, transcoding or direct, music or video -- because those answers decide which of several very different code paths a report is about, and reconstructing them over several round trips is most of the cost of a playback bug report. docs/build/ci-operations.md is the missing operations manual: how to change the builder image and in what order (image pushed before the workflow that names it, or CI breaks), why tags are dated rather than :latest or per-SHA, what each secret is for, and what losing the updater private key would mean -- installed desktop clients only accept payloads signed by the key matching the public key they shipped with, so losing it means everyone reinstalls by hand. Disk exhaustion on the runner is documented as a manual check rather than a scheduled job. A daily job would occupy the only slot on a single-slot runner and pull the whole builder image to run `df` -- and `df` inside a container does not reliably describe the host's disk, so it would spend real build capacity reporting a number that might be wrong. What the doc records instead is the part that is actually hard to rediscover: the symptoms (cargo dying mid-link, docker refusing to pull, actions/cache quietly not saving) and that `docker volume prune` needs `-a` to touch named volumes, which is how it filled up unnoticed. Two things in these docs are stated plainly because they are true and were not written down anywhere: without branch protection every gate in the pipeline is advisory, and the Gitea instance -- canonical remote, signing secrets, registry, runner -- is not backed up by anything in this repository.	2026-08-21 18:45:40 +02:00
dtourolle	3211c96ecf	feat(updater): in-app update on desktop, releases link on Android ... Anyone who installed an AppImage or ran the Windows installer was frozen on that version forever. Nothing in the app ever mentioned a new release existed, and the release notes were the only announcement. Desktop now checks a signed manifest, shows the version and its notes in Settings, and installs and relaunches on request. The signature check is the whole point: it is what stops a substituted download from being installed by the app itself. Windows binaries stay unsigned for SmartScreen purposes -- that is a code-signing certificate, a separate problem -- but the update payload is verified against our own key. Android is deliberately not wired to the updater. An app may not replace its own APK; that is the package installer's job, and the plugin has no Android implementation. It gets a link to the releases page instead of a button that would throw. The plugins are gated with a target-triple cfg rather than cfg(desktop). Cargo only evaluates target cfgs in a [target.'cfg(..)'] table, so cfg(desktop) matches nothing, silently drops the dependency, and fails much later with "Permission updater:default not found" -- which is exactly what the first attempt here did. Where the manifest lives took some finding. This Gitea serves /releases/download/<tag>/<asset> but 404s on /releases/latest/download/<asset> (verified against a real asset), so there is no stable latest-release URL. The gitea-pages branch is force-pushed wholesale by publish-docs.yml, so it cannot host the file either. latest.json therefore gets its own orphan branch, read over the raw-file URL, and is published from a scratch repo in RUNNER_TEMP rather than by switching branches in the checkout -- doing that would have left the following steps standing on a one-commit history, and the next step but one runs release:notes against the real commit range. Also fixed, all of it release-integrity: - "appimage" is in bundle.targets. The release notes have advertised an AppImage for months; tauri.conf.json never built one, the artifact step globbed for *.AppImage, found nothing, and said nothing. The step now fails instead. - The .AppImage.tar.gz/.sig pair and the NSIS .sig are collected. A manifest referencing a signature that was never uploaded fails only on the user's machine, so the manifest step also refuses to write an entry with an empty signature. - Release notes are generated by release:notes from the traceability graph, which is what CLAUDE.md has asked for all along, instead of a fixed heredoc that said "see CHANGELOG.md for detailed changes" and linked "GitHub Issues" on a Gitea-hosted project. - The notes tell users how to verify a download with SHA256SUMS. Requirements UR-077 / DR-217, tests UT-208 (12 cases over the version comparison and the platform decision, including that a pre-release does not offer itself as an upgrade to the matching release). Verified: 1070 frontend tests, cargo check for both the host and aarch64-linux-android (confirming the plugins are absent there), clippy -D warnings, svelte-check 0 errors.	2026-08-21 18:41:50 +02:00
dtourolle	96abc3afef	docs(specs): make "a spec becomes an architecture doc" the written rule ... The sixteen specs folded in last commit were folded because someone noticed they had gone stale, not because anything said they should be. Without the rule written down the directory drifts straight back to a mix of promises and descriptions, and neither can be trusted: you cannot tell from a file whether it describes the build or proposes a change to it. So: docs/specs/ holds only unshipped work, there is no "Implemented" resting state, and the fold-in and the deletion happen in the same commit. The template now asks for the destination architecture doc **up front**, which is a design check rather than bookkeeping — a feature that fits no existing doc usually has an unclear layer assignment, and it is cheaper to find that out at spec time. It also tells the author which half of what they are writing is durable (invariants, rejected alternatives, the defect a decision prevents) and which half dies with the file (phases, migration steps, acceptance criteria). The review checklist gains a Lifecycle section, including the case that gets lost otherwise: out-of-scope work worth doing has to be written where it will still be found after the spec is gone.	2026-08-21 18:29:09 +02:00
dtourolle	f6653e6a8b	ci(security): add a supply-chain gate, checksums and an SBOM ... The project shipped signed Android builds and unsigned desktop binaries with no vulnerability scanning of any kind. Nothing checked the ~500 crate Rust graph or the JS packages against an advisory feed, and nothing checked that what we redistribute inside an MIT bundle permits it. The first cargo-deny run found eight vulnerabilities and one unsoundness -- bytes, four in rustls-webpki, time, two in quick-xml and rand -- every one of them closed by a `cargo update` nobody had a reason to run. That update is in this commit; 740 Rust tests and clippy -D warnings pass on the new lockfile. Two structural fixes matter as much as the gate itself: - deny.toml scopes the graph to the targets we actually ship. Without it the Apple targets pull in plist -> quick-xml and report two DoS advisories against a crate that is in no binary we release. Ignoring those by ID would silence them everywhere, including where they would matter; scoping makes them correctly absent. - libmpv is pinned by rev instead of branch = "master". A branch means the revision is whatever Cargo.lock happens to hold and any `cargo update` silently substitutes new upstream code -- in the one dependency that is not from crates.io and that links a C library into the player. The rev is the commit already locked, so this pins current behaviour rather than changing it. Licence findings are recorded rather than waved through. libmpv and libmpv-sys are LGPL-2.1, satisfied here by dynamic linking against the system library; deny.toml carries the two obligations that follow (keep the linkage dynamic, ship libmpv's licence text with any bundle carrying the .so). MPL-2.0 crates are file-level copyleft and fine unmodified. Releases now publish SHA256SUMS (verified in-job with `sha256sum -c` before upload) and a CycloneDX SBOM for both halves, so "does this release contain <vulnerable crate>?" has an answer that is not "rebuild the tag and re-resolve it". Workflows pin jellytau-builder:2026.08 instead of :latest. While every job said :latest, rebuilding the image changed what every build compiled against, including rebuilds of old release tags. Also folded in, because both were the same class of problem: - publish-docs.yml downloaded mdBook from GitHub releases into /usr/local/bin at job time -- a toolchain install in CI, which CLAUDE.md explicitly forbids, and a hard dependency on GitHub's CDN at publish time. It is in the builder image now. - extract-traces.ts only ever read .ts/.svelte/.rs, so every requirement implemented by *configuration* was invisible to the matrix that measures it. DR-205, DR-206, DR-207 and DR-215 all carry TRACES comments nothing read, and each counted as uncovered while being covered. Coverage was really 90%, not 88%; MIN_THRESHOLD moves to 89 accordingly. CI workflows stay excluded and there is a test saying why: traceability-check.yml quotes "a TRACES: comment" beside deliberately-undefined example IDs, which the extractor would read as real traces and then fail its own dangling-ID check. Supply-chain requirement is DR-216. 🔴 The builder image must be rebuilt and pushed (scripts/build-builder-image.sh 2026.08) before this reaches master -- the workflows now name a tag and tools that do not exist in the registry yet.	2026-08-21 18:25:38 +02:00
dtourolle	32043a2152	docs: fold shipped specs into the architecture docs and delete them ... A spec was a promise; sixteen of them had become descriptions of code that already shipped, sitting beside four that describe work still outstanding, with nothing in the file telling the two apart. Half the statuses were also wrong — audio-equalizer read "Accepted" with the EQ live on both platforms, the native video spec said the flag stays off after the default was flipped on. The shipped designs move into docs/architecture, which is the maintained description of the build, and the spec files go. Git history keeps the originals; what a future change still needs is carried across: - 01-rust-backend: favourites rewritten (the old section named a file that no longer exists and called shipped buttons "planned"), domain vocabulary owned by Rust (SearchScope, exclusions, the bitrate ladder), background workers - 02-svelte-frontend: app shell and chrome, library mosaic, series/episode navigation, downloaded browse, safe-area insets, native-video store, logging - 03-data-flow: locally-indexed search - 05-platform-backends: audio settings on ExoPlayer, the equalizer's band vocabulary, native video compositing, the background-audio handoff - 06-downloads-and-offline: one storage model, offline catalog visibility - 09-security: path confinement and input binding docs/specs/README.md now says what the directory is for and where each shipped design went. Deferred work the specs recorded is kept beside the code it concerns rather than lost: season-bounded autoplay, the two dead search commands, why indexing is a full crawl. requirements.md had fourteen stale statuses — Android audio parity still read "Linux only", DR-150 still said the native-video default was off, DR-190 was Proposed after DR-196 implemented it, and five tooling requirements were Proposed after landing. Three unbuilt specs suggested requirement ids that have since been allocated to other work; each now carries a warning.	2026-08-21 18:15:58 +02:00
dtourolle	8f5c9023d0	ci: make the frontend gates real, and fix the coverage script ... The repo configured four frontend gates and enforced one of them. eslint and prettier ran in no workflow and no hook; `bun run check` ran only in build-release.yml, so a type error could sit on master until somebody cut a tag; and `bun run test:coverage` had been dead for months. CI (build-and-test.yml) now runs format:check, lint, check and coverage alongside the existing boundary and doc-link tripwires. The coverage script failure was a version mismatch, not a config problem: @vitest/coverage-v8 resolved to 4.1.10, whose peer range pins vitest exactly, while package.json asked for ">=1.0.0 <5.0.0" and got 4.0.16 -- every run died on a missing BaseCoverageProvider export. The loose range is what allowed the pair to drift, so it is now ^4.1.10. Two ratchets, same policy as MIN_THRESHOLD in traceability-check.yml: eslint --max-warnings=159 (0 errors; 159 is today's backlog, only ever lower it) vitest thresholds (statements 51 / branches 45 / functions 46 / lines 52, measured at 54.6 / 48.7 / 49.6 / 55.1) no-console is promoted from "off" to "error": the logger-facade migration it was waiting on is finished -- 8 calls remained, 2 of them real stragglers in the settings page, now on the facade the file already imported. The sink itself, tests, and scripts/ are exempted; a CLI whose stdout is the product is not a stray debug statement. The threshold was verified to bite by raising it to 99 and watching the run go red, not by assuming an unfailed gate works. DR-205 moves to Done; the coverage gate is DR-215.	2026-08-21 18:11:26 +02:00
dtourolle	ad48d89dfe	chore(format): run prettier over src/ and scripts/ ... Formatting was configured but never enforced: `bun run format:check` reported 199 unformatted files and ran in no workflow and in no git hook, so .prettierrc (printWidth 100, trailing commas) described an intention rather than the tree. This is the one-time sweep that makes the check gateable. Whitespace and token-reflow only -- no behavioural change: `bun run check` reports 0 errors and all 1053 frontend tests pass before and after. Kept out of every other commit on purpose. A 199-file diff mixed with real changes is unreviewable, and the next commit turns format:check into a hard CI gate so this cannot silently accumulate again.	2026-08-21 17:41:44 +02:00