Skip to content

AGENTS.md

Latest commit

 

History

History
791 lines (737 loc) · 59.1 KB

AGENTS.md

File metadata and controls

791 lines (737 loc) · 59.1 KB
Raw

AGENTS.md

This file provides guidance to coding agents working in this repository.

Plan Mode

  • When an agent is in Plan Mode and produces a final <proposed_plan>, it must also save that finalized plan as a Markdown file in the repo-root .plans/ directory.
  • Save only finalized plans. Do not write interim exploration, questions, or draft revisions to .plans/.
  • Use the filename pattern YYYY-MM-DD-short-topic.md such as .plans/2026-03-12-channel-filtering.md.
  • If the intended filename already exists, append a numeric suffix such as -2, -3, and so on.

Agent Bootstrap

  • In a fresh worktree, run pnpm install --frozen-lockfile before relying on Nx project discovery, lint, test, or build commands. Without node_modules, pnpm nx show projects will fail because the local Nx modules are unavailable.
  • Re-run the install whenever the checkout moves — git pull, git reset --hard, a rebase, or a worktree branch being re-pointed. Git rewrites pnpm-lock.yaml but never re-links node_modules, so a tree installed at an older commit keeps serving the old dependency versions and tests fail locally while CI stays green. Check with cmp pnpm-lock.yaml node_modules/.pnpm/lock.yaml; any difference means the tree is stale, and a plain pnpm install --frozen-lockfile in that directory repairs it. Each worktree needs its own install — with no local node_modules, Nx aborts with Could not find ".modules.yaml".
  • After dependencies are installed, verify workspace discovery with pnpm nx show projects.
  • Use scoped path aliases from tsconfig.base.json such as @iptvnator/services, @iptvnator/shared/interfaces, and @iptvnator/ui/components. Do not add new imports from legacy bare aliases such as services, shared-interfaces, components, m3u-state, or database.
  • Every Nx project should keep scope:*, domain:*, and type:* tags in project.json so @nx/enforce-module-boundaries remains useful for humans and agents.
  • See docs/architecture/nx-workspace-boundaries.md for the current Nx tag and alias policy.
  • Keep nx and every official @nx/* package on the same exact version; run pnpm run deps:nx:validate after dependency updates.
  • Vite 7.3.6, resolved through Angular's build tooling, is patched with bounded transform prefilters and the upstream precise matchers in patches/vite@7.3.6.patch. Keep the patch until supported Angular tooling resolves a Vite version containing the fix, and run pnpm run deps:vite:test after related dependency updates.
  • app-builder-lib 26.15.7 (electron-builder's macOS signing) is patched in patches/app-builder-lib@26.15.7.patch with the upstream backport electron-userland/electron-builder#10172: security set-key-partition-list -k must receive the temporary keychain's own password, not the .p12 import password. macOS runner images since macos-26-arm64 20260831 verify that password, and Build on macos arm64 failed with SecKeychainUnlock: The user name or passphrase you entered is not correct. Keep the patch until electron-builder resolves an app-builder-lib containing the fix (26.16.1+), and run pnpm run deps:electron-builder:test after related dependency updates — the test fails when the patched version no longer matches the installed one.
  • A directory holding files consumed by other projects must be an Nx project. Nx builds its graph from TypeScript imports only, so a relative SCSS @use across project roots creates no edge and the imported file lands in no task hash — edits then return a cache hit instead of rebuilding. Shared partials live in libs/ui/styles (project ui-styles), and each consumer declares "implicitDependencies": ["ui-styles"]. Run pnpm run styles:inputs:validate after adding a cross-project stylesheet import.
  • Update Nx with pnpm nx migrate nx@<target> --skipInstall, regenerate the lockfile, run generated migrations when present, and validate before opening a PR. Major updates are always manual. Replace incomplete Dependabot security PRs with a coordinated update instead of editing the bot branch.
  • ESLint enforces max-lines on TypeScript files: production code targets under 300 with a hard maximum of 400, while tests (**/*.spec.ts, **/*.spec-data.ts, **/*.e2e.ts, apps/*-e2e/**) are held to 1200 — a long spec signals coverage, not the design debt the production limit catches. Blank lines and comments are not counted, so a docblock never forces a split. Limits live in tools/eslint/max-lines-config.mjs, imported by both eslint.config.mjs and the generator so the rule and the baseline cannot drift. Files that predate the rule are baselined in tools/eslint/max-lines-baseline.mjs; after splitting a file, regenerate it with node tools/eslint/generate-max-lines-baseline.mjs (it runs ESLint's own rule rather than counting lines itself). Never add new files to the baseline — the list must only shrink. A new file that genuinely cannot be split (for example a function serialized into another process) instead carries its own file-wide /* eslint-disable max-lines -- <why> */; the generator skips those files, so a justified exemption never lands in the baseline. Remove such a directive once ESLint reports it as unused.
  • Project lint targets that shell out to eslint must quote the glob, e.g. eslint "apps/<project>/**/*.ts". An unquoted ** is expanded by the POSIX shell on Linux and macOS (which has no globstar, so it matches only a shallow subset of files) while Windows passes the literal pattern to ESLint, which expands it recursively — the two hosts then lint different file sets. The target still reports success either way, so a broken glob hides missing coverage instead of failing. After changing such a target, compare the linted file count against find <project> -name '*.ts' | wc -l.
  • Repository-specific skills live under .codex/skills/.
  • Frontmatter descriptions are trigger-only and begin with Use when; keep each skill at or below 500 words.
  • Run pnpm run skills:validate after editing a committed skill or a literal path it documents.
  • Keep .codex and .claude copies of release-notes and release-cut byte-identical.

Documentation After Changes

  • After implementing a meaningful change, agents must assess whether canonical repo docs need updates before considering the task complete.
  • Meaningful changes include new or changed user-visible behavior, architecture or data-flow changes, non-obvious maintenance workflows, new setup/debugging steps, and new subsystem contracts or boundaries.
  • Skip doc updates for trivial refactors with unchanged behavior, formatting-only edits, and isolated test-only changes.
  • Prefer updating an existing authoritative doc before creating a new one:
    1. README.md for top-level developer or user workflows
    2. docs/architecture/ for architecture, ownership, and behavior contracts
    3. the nearest module README.md for local usage or behavior
  • Keep the root CLAUDE.md and this file up to date. They are living documents: whenever a change touches something they describe — monorepo structure (new/moved/renamed apps or libs), routes, database schema/tables, stores and their features, key components, commands, environment behavior, or coding conventions — update the affected sections as part of the same task, and keep the process sections mirrored between AGENTS.md and CLAUDE.md in sync.
  • When adding a new feature area, check whether the Architecture or Key Features sections of CLAUDE.md describe the surrounding area; if they do, reflect the addition there instead of leaving the description stale.
  • Do not let CLAUDE.md or AGENTS.md drift: a stale path or route in these files poisons the context of every future agent session. If you notice an outdated claim while working, fix it (or flag it in the final summary) even if it is unrelated to the current task.
  • Repo docs are canonical even when they were originally drafted by an LLM.
  • Final task summaries should state whether docs were updated and which doc changed.

Release Notes For User-Visible Changes

  • Any change a user could notice — new behavior, changed behavior, bug fix, performance win, breaking change — must add one note file under .changes/ in the same PR. Format, field table, and writing rules: .changes/README.md.
  • Name it <area>-<short-slug>.md; area matches the conventional-commit scope. There is no version field — the release version is chosen at release time.
  • Write the body for a user, not a reviewer: "the player now remembers volume between episodes", not "hoist volume state into the session". Max 400 characters; depth belongs in the release blog post.
  • type: internal records invisible maintenance. Internal notes stay collapsed in CHANGELOG.md, are omitted from the blog scaffold, and are removed from the authored public GitHub body by extract-changelog-section.mjs --public; GitHub's generated commit list remains separate, so an internal-only release can have an empty authored body.
  • highlight: <short headline> (max 60 characters, rejected on type: internal) marks a note as one of the release's two or three headline changes. Highlights lead the Telegram/Reddit announcement drafts, become ready-made blog section headings, and are the input the highlight-card generator renders from. A release where everything is a highlight has none.
  • Skip the note for test-only changes, docs, CI/workflow plumbing, and pure refactors with no behavior change. When skipping on a PR that touches apps/** or libs/**, apply the no-release-note label.
  • CI enforces this: the "Release note gate" job in .github/workflows/ci.yml fails PRs that change runtime code without an added .changes/*.md or the label (policy in tools/release/check-release-note-gate.mjs; tests/e2e/website/mock-server/docs paths are auto-exempt).
  • The release-notes skill covers writing notes; the release-cut skill covers the full release sequence. Canonical contract — surfaces, ordering constraints, the required draft asset set: docs/architecture/release-pipeline.md.
  • Validate before finishing: pnpm run release:notes:validate.
  • Announcement drafts and highlight cards are built from the same notes: pnpm --silent run release:notes:telegram and pnpm --silent run release:notes:reddit print paste-ready posts to stdout (Telegram is guaranteed to fit its 4096-character limit; --silent keeps pnpm's lifecycle banner out of a redirected post), and pnpm run release:cards:generate renders branded 1200×630 highlight cards plus a release hero into dist/release-highlight-cards/v<version>/. All three read highlight: metadata that exists only in the note files, so they must run before build-release-notes.mjs --consume; the cards additionally need release:screenshots to have run. Nothing is posted or copied into the website tree automatically.
  • Pushes to master and v* can publish Docker images. A v* tag build creates a draft GitHub release.
  • pnpm run release:verify:draft waits for that tag build (polling until the run is indexed, then gh run watch) and verifies the draft's status, authored body, and complete required asset set. It is read-only and deliberately fails on an already-published release, because it is the gate that runs before publication.
  • Publishing the GitHub release verifies its Snap assets and automatically uploads them to edge; installed-Snap smoke and candidate/stable promotion remain manual.
  • Release-post screenshots come only from the release capture script running against the mock servers. Never add a screenshot taken from a real playlist or account to apps/website/public/blog/** — real streams, logos, and metadata are copyrighted, and credentials must never reach a published image.
  • Final task summaries should state whether a release note was added or why it was skipped.

Regression Prevention And Test Updates

  • Before the final summary for any feature, behavior change, bug fix, data-flow change, Electron IPC/database change, or user-visible UI workflow change, complete a test impact pass. Identify the affected projects and decide whether unit, integration, E2E, build, lint, or manual/CDP verification is required.
  • Bug fixes must normally include regression coverage that fails on the old behavior and passes with the fix. If automated coverage is not practical, document why in the final summary and include the strongest manual validation performed.
  • Feature work and behavior changes must update existing tests when assertions, fixtures, mocks, routes, or E2E flows are now stale, incomplete, or missing. Prefer extending the closest existing spec or E2E file before adding a new suite.
  • Default validation ladder:
    1. Run targeted unit tests for directly affected projects with pnpm nx test <project> or existing scripts such as pnpm run test:frontend, pnpm run test:backend, or pnpm run test:unit:ci when the scope is broader.
    2. Run affected E2E coverage when changing user-visible workflows, routing, persistence, playback, portals, settings, import flows, or Electron-only behavior.
    3. Use pnpm nx show projects --withTarget test and pnpm nx show projects --withTarget e2e when project ownership or available validation targets are unclear.
    4. Prefer specific atomized E2E targets before broad suites when they cover the changed behavior, for example pnpm nx run web-e2e:e2e-ci--src/xtream.e2e.ts or pnpm nx run electron-backend-e2e:e2e-ci--src/search.e2e.ts.
  • Electron-specific changes affecting IPC, SQLite, packaged runtime, external players, native file access, or Electron-only routes require Electron E2E coverage where available, or CDP/manual verification with agent-browser and the tracing flags documented below.
  • Final task summaries must list tests added or updated, validation commands run with results, and any skipped validation with the reason. For docs-only changes, state that unit/E2E validation was not required and verify the changed Markdown instead.

Electron Debugging (CDP)

  • Start the Electron development app with: nx serve electron-backend
  • Package-script equivalent: pnpm run serve:backend
  • Electron is configured to start with: --remote-debugging-port=9222
  • Connect Chrome DevTools Protocol tools to: 127.0.0.1:9222
  • For Electron automation/debugging tasks, use the electron skill
  • Do not auto-open DevTools during normal CDP automation. In development, DevTools is opt-in via ELECTRON_OPEN_DEVTOOLS=1.
  • If DevTools is open, agent-browser --cdp 9222 ... may attach to the DevTools page instead of the IPTVnator window. Symptoms: tab list shows about:blank, snapshots are empty, and screenshots are black.
  • If that happens, inspect targets with curl http://127.0.0.1:9222/json/list and connect directly to the IPTVnator page websocket from the webSocketDebuggerUrl field.
  • The app holds a single-instance lock (acquireSingleInstanceLock in apps/electron-backend/src/app/services/single-instance.ts): a second launch against the same userData quits immediately and focuses the running window. To attach a second CDP-enabled instance to the same profile, set IPTVNATOR_ALLOW_MULTIPLE_INSTANCES=1 — knowing that only one of the two processes will own the renderer's IndexedDB, so settings written by the other are lost. Before focusing, the guard forwards the second launch's argv to onSecondInstance, which is how a playlist path handed to an already-running app reaches the open queue.

Trace / Debug Startup

  • Full startup tracing:
IPTVNATOR_TRACE_STARTUP=1 nx serve electron-backend

  • Narrower trace flags:

    • IPTVNATOR_TRACE_IPC=1 traces renderer window.electron.* bridge calls
    • IPTVNATOR_TRACE_DB=1 traces DB worker requests and request-scoped DB events
    • IPTVNATOR_TRACE_SQL=1 traces SQLite statements in the main process and DB worker
    • IPTVNATOR_TRACE_WINDOW=1 traces BrowserWindow lifecycle and unresponsive events
    • IPTVNATOR_TRACE_PLAYER=1 traces external-player activity and bounded Embedded MPV runtime-probe stderr
    • IPTVNATOR_TRACE_RENDERER_CONSOLE=1 mirrors renderer console output into the Electron terminal
    • IPTVNATOR_PERF_CAPTURE=1 enables development/test-only, redacted M3U and Xtream preload IPC request/completion markers plus count-only M3U acquire/parse/normalize, Xtream main network/JSON-transform/success-response-ready/cancel-dispatch, and renderer store phase capture; renderer wrappers emit only while the benchmark installs its Symbol hook, benchmark tooling sets the flag explicitly, and production launches must leave it unset
    • IPTVNATOR_PERF_WORKER_PROFILING=1 enables development/test-only, request-scoped worker receive/work/response-post timestamps, thread CPU, event-loop utilization/delay, count-only playlist serialization/SQLite write/read/deserialization plus Xtream category/content/cache-clear/delete/in-source-search phase events, profiling-only worker cancel-receipt acknowledgements, valid-sample-counted isolate peak memory, and the database worker's idle-only one-shot post-GC heap probe; overlapping database requests are explicitly invalidated instead of misattributed, the performance benchmark sets the flag automatically, and production launches must leave it unset

  • Settings, portal request/response, and trace payloads must use @iptvnator/shared/logging or the redacting portal logger before reaching console.*; never log raw credentials while debugging.

  • If local Nx state gets weird before a rerun:

pnpm nx reset

agent-browser (global install)

agent-browser --cdp 9222 tab list
agent-browser --cdp 9222 tab 1
agent-browser --cdp 9222 snapshot -i -c -d 4
agent-browser --cdp 9222 screenshot /tmp/iptvnator-cdp.png

Fallback

npx --yes agent-browser --cdp 9222 tab list

DevTools Workaround

ELECTRON_OPEN_DEVTOOLS=1 nx serve electron-backend
curl http://127.0.0.1:9222/json/list
agent-browser connect ws://127.0.0.1:9222/devtools/page/<iptvnator-page-id>
agent-browser screenshot /tmp/iptvnator-cdp.png

Xtream Category Management

The Electron Live TV, Movies, and Series category dialog applies Select/Deselect to search results while a filter is active and to the whole type otherwise. Button states use the matching group; "Total selected" counts the whole catalog. Save persists the complete draft, Close discards it, and refresh restores hidden categories by provider ID and type. See docs/architecture/category-management.md.

Portal Connectivity Preference

  • Desktop Settings > General > Portal connections exposes default-on Settings.portalConnectivityGuard. Only explicit false opts out. Save mirrors the value to Electron PORTAL_CONNECTIVITY_GUARD and applies it without restart; settings bootstrap restores it before the renderer loads. It controls Xtream and Stalker together. Preference transitions clear cooldowns and invalidate old request completions; unchanged saves preserve evidence. The environment switch IPTVNATOR_DISABLE_CONNECTIVITY_GUARD=1 remains authoritative. PWA clients do not control the shared backend's guard.
  • Both account-info dialogs explain guard refusals with localized paused-request copy and Retry now; Stalker preserves cached account data on a failed refresh. Contract: docs/architecture/host-connectivity-guard.md.

Radio / Audio Player

M3U playlists can contain radio channels identified by the radio="true" attribute on #EXTINF lines. When a radio channel is selected:

  • The dedicated AudioPlayerComponent (libs/ui/playback/src/lib/audio-player/) renders instead of a video player
  • The audio player always uses the built-in inline player — external player settings (MPV/VLC) are ignored
  • The EPG panel is hidden (radio streams have no EPG data)
  • The layout uses a cinematic hero pattern: the station logo is blurred as a full-area backdrop with a vignette overlay, and the artwork card + controls float above it
  • Volume is shared with the video player via localStorage key 'volume'
  • Keyboard shortcuts: ArrowUp/ArrowDown (volume +/-5%), M (mute toggle)
  • Radio detection in the video player template: activeChannel.radio === 'true' — this is a string comparison, not boolean

Key files:

  • libs/ui/playback/src/lib/audio-player/audio-player.component.ts — the audio player component
  • libs/ui/playback/src/lib/audio-player/audio-player.component.scss — cinematic hero styling
  • libs/playlist/m3u/feature-player/src/lib/video-player/video-player.component.html — template conditionals for radio vs video
  • libs/shared/interfaces/src/lib/channel.interface.ts — radio: string field on Channel interface

Shared Player Controls

  • libs/ui/playback/src/lib/player-controls/ contains the additive, engine-neutral PlayerController contract, standalone app-player-controls, generic web-video adapter/helper, and component-scoped WEB_PLAYER_SHARED_CONTROLS rollout token.
  • The subtitle menu carries capability-gated advanced subtitle support (#1408): external subtitle file loading, a ±0.5 s timing-offset row, and size/color styling persisted in the shared subtitleStyle localStorage key. HTML5/ArtPlayer implement it through the neutral source bridge (.srt/.vtt via a DOM file picker with encoding detection, native TextTrack rendering, ::cue styling, delay only while the loaded file is the selected track; picks are source-generation-guarded and engine deselection precedes external track activation). The canonical style shape and clamp/normalize rules are shared with the main process via @iptvnator/shared/interfaces (subtitle-style.util.ts). Embedded MPV frame-copy implements it through helper protocol commands (sub-add/sub-delay/sub-scale/sub-color, main-process file dialog, ASS supported, delay for all tracks). Video.js shared mode, vendor-chrome paths, native-view, and the Linux out-of-process path advertise no such capability and render no UI. Contract details: docs/architecture/player-controls-contract.md ("Advanced subtitle support").
  • In fullscreen, app-player-controls shows a pointer-transparent media-title overlay at the top while controls are revealed (mediaTitle input: movie/channel/series name, plus an S01E03 second line for episodes). Series names flow from the Xtream/Stalker detail views through PortalInlinePlayerComponent.seriesTitle and WebPlayerViewComponent.mediaTitle; movie and live hosts fall back to playback.title, skipping raw stream-URL fallbacks. Outside fullscreen the overlay stays hidden.
  • Auto-hide pauses while the pointer is over the controls bar or keyboard focus is inside it, but only keyboard-originated focus pins the bar open. Chromium also focuses a clicked <button>, so ControlsSurface.wasPointerInteraction attributes a focusin to a recent pointerdown inside the focused element; such focus reveals without blocking auto-hide (otherwise the fullscreen button left the controls on screen until a click-to-pause on the viewport). The press record is discarded on the first bar focus event it is asked about or on any keydown, a pointerdown inside the bar releases a keyboard pin, and a keydown bubbling out of a bar control re-pins it, since operating a focused control produces no focus event. A completed pointer click then releases the focus it left on the control (onBarClick → ControlsSurface.releasePointerFocus, attributed by wasPointerClick: non-empty click pointerType, else a recent press inside the clicked element), because a focused control captures the keyboard: Space and Enter re-activated the clicked button and ControlsShortcuts yields to any interactive element in the key's path, so after a click on fullscreen Space left fullscreen instead of pausing. Keyboard activation (empty pointerType) keeps focus, only buttons and range sliders are released, Chromium keeps its sequential-focus starting point at the blurred control so Tab continues from it, and the volume popover ignores the release's focusout (wasPointerFocusRelease). Contract: docs/architecture/player-controls-contract.md (auto-hide paragraph).
  • Persisted Settings.webPlayerSharedControls is default-ON (absent stored values coerce with !== false; only an explicit false opts out to the legacy vendor chrome), and its checkbox appears only when HTML5, Video.js, or ArtPlayer is selected. WebPlayerViewComponent snapshots the preference into WEB_PLAYER_SHARED_CONTROLS for each new player host. The parent /workspace route awaits the initial SettingsStore load, including cold-start direct links, before this snapshot can occur. Saving applies to the next host without an application restart; an existing session never changes controls mode in place.
  • Settings.showCaptions is deliberately outside this rollout gate: it is engine state, not controls UI. HTML5, Video.js, and ArtPlayer apply it in both modes — shared controls through their controls bridge, the preference-off paths through the same helpers without an adapter (WebVideoSourceTracks for HTML5/ArtPlayer, VjsLegacyTracks for Video.js). Both re-apply the preference as the engine adds or switches text tracks. WebPlayerViewComponent reads it from SettingsStore rather than a host input, so the M3U player, the Xtream/Stalker live layouts, and the portal detail inline player all inherit it (#1155).
  • The modes differ in how long the preference is enforced. Shared controls are authoritative for the session; user intent arrives through setSubtitleTrack and wins until the source changes. Vendor chrome is source-default: the preference seeds each new source and is released once the media element reports playing, so the engine's own caption menu keeps working. The mode is selected by the optional playbackStarted probe the legacy owners pass to all three helpers (HLS, native text tracks, Shaka); in that mode the HLS helper deselects the track (subtitleTrack = -1) instead of hiding it, because subtitleDisplay would silently override whatever the vendor menu picks. For DASH the seed happens in ShakaVideoSession.start() after the manifest loads, so the helper only stops re-suppressing afterwards.
  • Shared controls include a per-session quality menu (Auto + "1080p"-style levels via setQualityLevel; AUTO_QUALITY_LEVEL_ID restores ABR). The capability derives from the manifest — advertised only when the source exposes >1 video rendition (multi-variant HLS via hls.js nextLevel/manualLevel, DASH via Shaka variant tracks pinned to the active variant's exact audio stream (audioId, language fallback) with ABR toggled off for manual picks, Video.js via videojs-contrib-quality-levels) — so single-bitrate VOD and raw MPEG-TS never show it, nothing persists to Settings, and Embedded MPV/external players report the capability false.
  • Embedded MPV ignores the web-player preference. Frame-copy always uses shared DOM controls through its component-scoped EmbeddedMpvControlsAdapter, while native-view retains the legacy compositor-safe dock and external MPV/VLC retain their own UI. The host must render exactly one controls system for the reported Embedded MPV engine.
  • Frame-copy shared controls own DOM surface interactions, shortcuts, fullscreen, and recording feedback. showControls=false detaches the shared surface, modal overlays gate playback shortcuts, fullscreen still triggers bounds sync, and a playback/session transition key prevents engine or session handoff from presenting stale recording feedback while timers and pending commands are cancelled. Same-session IPC replies also yield to a broadcast snapshot received while the command was pending, preventing a successful recording acknowledgement from being rolled back by a stale reply.
  • Embedded MPV seek steps (arrow keys, ±10 s buttons, PlayerController.seekBy) go through the relative seekEmbeddedMpvBy IPC: every backend forwards the delta as mpv seek <delta> relative+exact (addon export seekBy, helper stdin command seek-by, Linux JSON IPC) and never advances the snapshot position itself. Do not derive an absolute target from the renderer's positionSeconds: it is floored to whole seconds, polled every 500 ms, and a seek reply does not carry the new position, so rapid presses computed from it collapse onto one target. Only the timeline scrub commits an absolute seek. Contract: docs/architecture/embedded-mpv-native.md ("Resume And Track Handling").
  • DASH (.mpd) sources play through a lazily imported Shaka Player source engine (libs/ui/playback/src/lib/shaka-engine/) inside the HTML5 and ArtPlayer components; ClearKey keys come from KODIPROP-derived Channel.drm, and the shared bridge exposes Shaka audio/text tracks via source kind shaka. The DOM-free Shaka 5.2.4 diagnostic boundary lives in libs/playback/util; it version-locks public severity/category/code evidence, ignores recoverable error events, treats rejected loads as terminal lifecycle outcomes, preserves exact public DASH text-parser category/code evidence with unknown stage/failure, and never retains or renders raw messages or error.data. A failed browser-support preflight stays generic-unknown but carries the exact app-owned PlaybackRuntimeSupport.ShakaBrowserUnsupported marker, preserving managed external fallback only for clear transferable DASH; PWA capability and KODIPROP DRM still suppress it. See the CLAUDE.md "Video Players" feature entry and the "DASH + ClearKey Playback" section of docs/architecture/m3u-playlist-module.md.
  • mpegts.js 1.8.1 errors from HTML5, Video.js, and ArtPlayer cross one version-locked structured evidence boundary in libs/playback/util. Only exact public type/detail pairs, pair-derived stage/failure, terminal disposition, and the validated HTTP 4xx/5xx status slot are retained; raw messages and arbitrary info never reach diagnostics. This is a sibling of PlayerController, not part of the controls contract.
  • Browser playback diagnostics and recovery policy live in libs/playback/util and are exported by @iptvnator/playback/util. Public engine errors cross allowlisted sanitizers into a PlaybackDiagnostic; recommendPlaybackRecovery(context) then ranks at most three actions, and WebPlayerViewComponent executes only the action the user selects. The policy is a sibling of PlayerController; shared controls only gate interaction while the diagnostic panel is visible. WebPlayerViewComponent owns a host-derived content-session key that is stable for the mounted logical selection, attempted target IDs, the temporary player override, and VOD handoff position. Its PlaybackBinding is exactly { generation, target }, while every source/target/reload application uses a fieldless opaque Symbol token. Diagnostic storage uses a separate fieldless intent Symbol, and source applications advance a third fieldless revision Symbol that clears only the VOD handoff position; target-only switches and Retry leave that revision stable. None of these ownership primitives contains URLs, headers, DRM material, or credentials. The application effect synchronizes the content session before tracking intent, so clearing a temporary player override cannot schedule a duplicate application or header handoff. Every application start clears both the diagnostic owner and backing signal before asynchronous header setup; a current false result or rejection leaves them clear, and a stale completion cannot erase a newer owned diagnostic. Each rendered web or Embedded MPV application captures its nullable binding, the application and source-revision tokens, and live/VOD flag; a time update changes resume state only while that exact capture still owns the current application. A recommended built-in player temporarily outranks the host override and saved player for that mounted content session, never mutates Settings.player, and resumes finite VOD position on a best-effort basis; live playback returns to the live edge. Retry and alternative sources preserve attempts, while a different content-session key or component teardown resets them. Recovery recommendations never auto-switch, persist history, learn across sessions, or emit telemetry. The policy projects attempted inline target IDs through the validated canonical source/target capabilities and excludes every attempted engine family, so HTML5 and ArtPlayer are not separate hls.js recoveries. Network and generic unknown evidence fail closed to Retry/alternative source; the exact Shaka browser-unsupported preflight marker is the sole unknown-code exception. PWA capability suppresses managed MPV/VLC, and ClearKey/KODIPROP DRM suppresses external targets because its payload is not transferable. Raw engine messages, arbitrary data, and credentials never enter recommendation evidence or ownership state. MPV/VLC actions remain mounted after an attempt and expose credential-free per-target launching/started/playing/error state; only an exact Electron playing update is labelled Playing. One handshake is allowed at a time. The renderer claims the credential-free content identity before awaiting Electron, so primary Play is disabled and a launching or closable-error alternative remains owned before the controller commits it. Every route action that can start the same external playback, including Restart and the provider-source shortcut, observes that local pre-IPC guard. The Xtream VOD diagnostic-fallback handler records the same route-scoped destination and pending generation before invoking MPV/VLC, so route reuse cannot orphan that process outside the next route's close-before-play path. Its fieldless intent is bound to the exact session returned by the source owner's launch promise, so a late timed-out attempt cannot take over a retry; later global updates must match that ID. A replacement waits for confirmed teardown of the tracked external process, applies the old exact close before launch, and cancels an unlaunched handoff if diagnostic ownership changes. Process teardown has bounded graceful and forced confirmation windows, and reusable MPV bounds the IPC command that precedes them; if any stage cannot reach a confirmed exit, the exact session stays live and the replacement fails closed instead of overlapping it. A process-wide teardown gate starts before any potentially slow teardown preparation, including VLC position flush and a reused player's protocol quit, and rejects every MPV/VLC spawn until that exact child reports exit. If bounded teardown fails while a fresh launch is still pending, that launch IPC rejects and the exact session remains a closable error instead of hanging forever. If a pre-content reuse failure has no still-live displaced session to restore, the replacement error keeps its attached closer so Stop can retry the orphaned child teardown. A terminal error without a closer is never restorable. A failed close is single-flight only while its promise is pending: Stop can retry the same exact child after a bounded confirmation failure. Reuse maps the child to its current content session, so a stale older closer becomes a no-op instead of terminating a newer loadfile/VLC enqueue handoff. A duplicate close for an already closed session returns its terminal snapshot without re-entering the saved closer, and a late process error cannot revive that terminal session. Reused MPV commands are bound to the socket captured for that exact child, so a later process cannot inherit a stale protocol quit. Stop observed before a pending MPV content command or VLC enqueue command prevents that command from dispatching. A source handoff fails closed while a live session has no closer (canClose: false); renderer Dismiss is not teardown confirmation. That denied handoff advances neither the multi-source switch token nor the playback generation, so it cannot cancel the sole launch already in flight. VLC rechecks the gate at each concrete spawn after port allocation or reuse work; if a post-start fallback is blocked there, the opened session becomes an error rather than retaining a false started status. A failed RC-port allocation never claims reuse ownership, so the fallback VLC child retains its exact one-shot closer. Reuse failures before a content command restore the globally displaced renderer session, not the reusable process's prior owner, and only while the exact displaced-session ID is still active; after loadfile/VLC clear is dispatched, the replacement owns the process and remains a closable error instead of restoring stale content metadata. Stop during an in-flight MPV or VLC reuse command, including during failed-command teardown or the subsequent VLC fallback port-allocation wait, settles that exact close without falling through to a fresh spawn; a stopped VLC spawn error that reports only close also settles its original launch IPC with the exact closed session; a fresh fallback retires the old child's exit under its prior session so it cannot close the replacement. Source handoffs recheck ownership after launch and accept only opened/playing; a stale returned session is closed exactly and a Stop-returned closed session is never committed. If that exact stale close fails, its credential-free destination owner is retained for the next close attempt. Retained destination ownership is scoped to the initiating playlist/VOD route key, so route reuse cannot expose Stop for the previous movie's external session. Play/Resume capture that route key before awaiting close and cancel if navigation changes it; a late diagnostic fallback closes its exact returned session instead of adopting it on the new route. They supersede an older source resolution before awaiting the shared close-before-replacement path, and accepting a diagnostic fallback retires the same older resolution before opening MPV/VLC. They publish the route-source badge, caption evidence, and position only after start succeeds. Closable errors still participate in every replacement close and keep Stop as the global dock's only teardown affordance; Dismiss is reserved for terminal errors that have no closer. The shared isLiveExternalPlayerSession predicate keeps M3U and series ownership while such an error can still be stopped; consumers must not treat every error status as terminal. If the local handshake times out after an exact Electron session is known, that ID remains correlated so a later exact update can recover the UI. The global dock mirrors those statuses, keeps closable errors visible until Stop confirms teardown and terminal errors visible until dismissal, and intentionally has no retry because it does not own the original launch headers or credentials.
  • The built-in HTML5/hls.js player is the second guarded consumer. HtmlVideoPlayerComponent provides a component-scoped WebVideoControlsAdapter; its neutral web-video-support bridge is shared with ArtPlayer and owns HLS/Shaka(DASH)/native tracks, MPEG-TS VOD duration correction, caption preference, and source cleanup. HtmlVideoElementSession owns native video-event lifecycle, persisted volume, start-time/time/ended propagation, and legacy post-play caption suppression. WebPlayerViewComponent.resolvedIsLive supplies authoritative live/VOD metadata, while a visible playback diagnostic disables both shared surface interaction and shortcuts and exits the shared controls' resolved fullscreen owner (the host-supplied fullscreenTarget, else the HTML5 shell) so the diagnostic actions remain visible. The preference-off path keeps native controls and legacy series navigation unchanged, while the playback keyboard shortcuts (Space/K, F, arrow seek/volume, M) attach through LegacyPlayerShortcuts with commands acting on the native video element (html-video-legacy-shortcuts.ts); seek requires authoritative VOD metadata plus a finite positive duration, and a visible diagnostic disables the keys.
  • Video.js is the third guarded consumer. VjsPlayerComponent provides a component-scoped WebVideoControlsAdapter; its bridge binds the current Tech video, rebinds after playerreset, exposes source-stable audio/subtitle IDs, preserves caption preference and explicit subtitle-off state, and reads duration from Video.js. Reset-driven raw MPEG-TS changes pause first, coalesce to the latest desired source, preserve actual volume across Video.js's reset, and restart when authoritative live/VOD metadata changes. The shared-controls path disables native controls, Video.js click/double-click/hotkey actions, and spatial navigation; diagnostic gating and owned-fullscreen exit match HTML5. The preference-off path keeps the existing Video.js skin and legacy series navigation unchanged (still without userActions.hotkeys), while the playback keyboard shortcuts attach through LegacyPlayerShortcuts and drive the player API so the vendor control bar stays in sync (vjs-legacy-shortcuts.ts). That chrome also releases the focus a pointer interaction leaves on a control (vjs-pointer-focus-release.ts, sharing pointer-focus-release.ts's blurFocusedControl with ControlsSurface): a focused Video.js component stops every key before the document and turns Space/Enter into a click, so after a click on fullscreen Space left fullscreen instead of pausing. It is driven mainly by focusin, not the click, because choosing a menu item moves focus to the menu button a tick later and that click never bubbles to the shell: an eligible control (button/role=button/slider, never a menu item) is released when its focus is attributable to a recent shell pointerdown not yet ended by a document keydown, so Tab focus is kept. A click runs the same release for a control clicked while already focused (Tab, then a mouse click), which fires no focusin. The release is scoped to .vjs-control-bar, so the caption-settings dialog (a modal sibling of the bar) keeps its focus trap. Menu buttons live in the bar and are not exempt: a popup is navigated through its focused item, so releasing the button never disturbs an open menu, and the button focus a pointer moves through (open, item selection, toggling an open menu shut) is released so Space works again after the menu closes. ArtPlayer (non-focusable divs) and the native HTML5 controls (focus lands on the <video>) need no counterpart.
  • ArtPlayer is the fourth guarded consumer. ArtPlayerComponent provides a component-scoped WebVideoControlsAdapter; ArtPlayerSourceSession owns HLS/DASH(Shaka)/MPEG-TS/native sources, the neutral web-video bridge, exact cleanup, and a destroyed-session guard for delayed customType callbacks, while ArtPlayerVideoSession owns native media/ArtPlayer events. Shared mode uses authoritative live/VOD metadata, HLS/Shaka/native tracks and caption preference, MPEG-TS VOD duration correction, and reapplies app volume directly after ArtPlayer restores its own stored volume. Vendor chrome/hotkeys are disabled, and a transparent capture layer gives shared controls exclusive click and double-click ownership. Diagnostic interaction gating and owned-fullscreen exit match the other web players. The preference-off path keeps the legacy ArtPlayer skin, source behavior, and series navigation unchanged, while the playback keyboard shortcuts attach through LegacyPlayerShortcuts using the vendor setters ArtPlayer's own hotkeys used (art-player-legacy-shortcuts.ts); the legacy chrome passes hotkey: false because ArtPlayer's focus-scoped hotkeys ignore defaultPrevented and would double-handle every key, and the wiring restores its Escape-exits-web- fullscreen behavior.
  • Shared web picture-in-picture stays inside that default-on rollout. PlayerController exposes capability pictureInPicture, state pictureInPictureActive/canPictureInPicture, and command togglePictureInPicture(). HTML5, Video.js, and ArtPlayer use standard element PiP from the adapter's attached video; shared ArtPlayer keeps vendor pip: false, while preference-off native/vendor controls keep their own UI. The capability-gated button sits before fullscreen and uses active enter/exit semantics; entry is disabled until metadata, and the action is disabled while an operation is pending. Embedded MPV reports capability/state false with a no-op command and has no popup/mini-window.
  • WebVideoControlsAdapter supplies its current video and binding generation to WebVideoPictureInPictureController; the controller reads the video's ownerDocument, while browser enter/leave events remain authoritative. Exact-owner exit stays available if request support changes. Request/exit invocation remains synchronous for user activation, one operation is serialized, and binding generation plus exact video identity protects replacement and teardown from stale completion. Video.js Tech reset and ArtPlayer rebuild rebind with exact-owner cleanup; HTML5 source changes on a retained target preserve PiP. Legacy HTML5/ArtPlayer teardown and Video.js Tech replacement also release exact-owned PiP through web-video-picture-in-picture-lifecycle.ts, independent of the controls preference. A one-shot listener on the retired video closes late native/vendor entries without retaining the host or touching another video's PiP. Legacy WebKit presentation-mode PiP also returns the retired video to inline; its presentation-change listener ignores fullscreen/inline events until a late PiP entry consumes it. Standard PiP shows the browser/OS video surface without Angular control chrome, with browser-dependent subtitles. AirPlay, Cast, Document PiP, a PiP keyboard shortcut, and Embedded MPV popup/native support are out of scope.
  • Canonical docs: docs/architecture/player-controls-contract.md and docs/architecture/embedded-mpv-native.md

Display Sleep During Playback

  • PlaybackKeepAwakeService (apps/web/src/app/services/playback-keep-awake.service.ts) watches every <video> via document-level capture listeners (media events don't bubble; release listeners sit on the tracked element because Chromium's removed-from-DOM pause never reaches the document) and, while any video is playing and the document is visible (or the playing video is in picture-in-picture — the PiP surface survives a minimized window), holds a display-sleep lock.
  • Electron: a main-process powerSaveBlocker behind window.electron.setPlaybackKeepAwake (apps/electron-backend/src/app/services/playback-keep-awake.service.ts); the renderer's vote is auto-cleared on renderer reload, crash (render-process-gone), or destruction. PWA: the Screen Wake Lock API, re-requested after browser auto-release; state changes masked by an in-flight request() queue one re-evaluation on rejection.
  • Radio's <audio> deliberately never blocks display sleep. Embedded MPV holds its own blocker in EmbeddedMpvNativeService; external MPV/VLC inhibit the screensaver themselves.

Windows Embedded MPV Pin Maintenance

  • PR, master, and tag builds resolve the Windows runtime only from tools/embedded-mpv/windows-runtime-pin.json; repository variables are not build inputs.
  • Validate the checked-in schema and provenance with pnpm embedded-mpv:windows-runtime-pin:check.
  • Prepare a manual rotation with pnpm embedded-mpv:windows-runtime-pin:refresh -- --force. The weekly refresh-windows-embedded-mpv-runtime.yaml workflow runs the same updater and opens a reviewable PR before upstream retention expires.
  • The PAT-backed refresh job must keep every third-party action pinned to a full commit. Do not mirror the upstream binary without complete corresponding source, build records, license notices, and a validated transitive license closure.

Linux Embedded MPV Packaging

  • Official Linux frame-copy artifacts are x64-only. AppImage, DEB, RPM, Pacman, Snap, and Flatpak are supported; non-x64 Linux packages must remain marker-only and must never inherit x64 native artifacts from environment overrides.
  • Packaging runs three isolated profiles:
    • system: DEB/RPM/Pacman, no private native/lib, with package dependencies DEB=libmpv2,libegl1,libgl1,libgbm1, RPM=mpv-libs,libglvnd-egl,libglvnd-glx,mesa-libgbm, and Pacman=mpv,libglvnd,mesa
    • portable: AppImage/Snap with the pinned LGPL-compatible closure
    • flatpak: Flatpak with the same pinned closure
  • Flatpak is an isolated packaging pass and keeps iptvnator as the real Electron ELF so Electron Builder's electron-wrapper passes it directly to Zypak. Other Linux targets retain the conditional iptvnator wrapper and iptvnator.bin. Mixed Flatpak/non-Flatpak target sets fail before mutation.
  • The DEB system-runtime contract is Ubuntu 24.04+ (libmpv2). Ubuntu 22.04 provides libmpv1, so use the x64 AppImage on Jammy instead of weakening the package dependency or advertising frame-copy without a compatible runtime.
  • Only iptvnator_mpv_helper may link libmpv. The Electron executable, Electron libraries, embedded_mpv.node, and embedded_mpv_frame_reader.node must not load or link it. Preserve this process-isolation contract in build, package, and smoke checks.
  • electron-backend/native{,/**/*} is excluded from app.asar; afterPack exclusively writes the profile-normalized unpacked native tree. Layout and final-artifact checks must reject every archived /electron-backend/native/** entry so system and marker-only packages cannot hide stale x64 artifacts.
  • Packaged addon, frame-reader, and helper discovery is package-owned app.asar.unpacked only. Writable cwd/dist candidates are development-only and must never satisfy packaged native-view support or the frame-copy gate.
  • Pristine afterPack/unpacked layouts scan Electron libraries recursively. Extracted Snap payloads exclude only the package-manager lib/** and usr/lib/** trees that Snap overlays into the same root; every other directory remains recursive, and Electron-library symlinks still fail closed.
  • Linux frame-copy availability is fail-closed. The packaged manifest, artifact modes, declared bundled hashes/closure, and bounded --runtime-probe must all succeed before frame-copy can relax the renderer sandbox. Any failure reports a stable reason and falls back to native-view without crashing; an environment flag never bypasses this gate.
  • Snap is core22/strict and uses an exact private shared-memory plug plus the graphics-core22 content plug at an empty mode-0755 $SNAP/graphics, with mesa-core22 as default provider. It declares only the canonical provider layouts: /usr/share/libdrm binds from $SNAP/graphics/libdrm, and /usr/share/drirc.d symlinks to $SNAP/graphics/drirc.d. The provider is external shared content, not part of IPTVnator's package size, source archive, or notices. Installed-Snap CI must prove controlled unavailable exit after disconnect, then reconnect and prove success. Static artifact verification requires regular desktop-init.sh, desktop-common.sh, and desktop-gnome-specific.sh files at the Snap root, with desktop-init.sh executable. The helper links libGL.so.1 rather than libOpenGL.so.0.
  • The probe and playback helper share one sanitized loader environment: ambient audit, preload, library, graphics-driver, and shell-startup overrides are removed; the validated private closure wins; trusted Snap GL, graphics-core22, the core22 base x64 root, and exact GNOME-platform roots precede generic in-snap roots. The core22 base must precede GNOME so its libedit.so.2 cannot be replaced by the older copy requiring libtinfo.so.5. The extracted-artifact verifier removes the identical unsafe loader/graphics/shell set before direct helper smoke while preserving feature/debug selectors such as LIBGL_ALWAYS_SOFTWARE. Snap fixes the wrapper PATH, removes exported BASH_FUNC_* functions, and launches probe/playback through the regular executable $SNAP/graphics/bin/graphics-core22-provider-wrapper; a missing or disconnected provider returns snap-graphics-provider-unavailable before helper spawn. The packaging-only --embedded-mpv-runtime-probe app switch runs the complete cached manifest/hash/helper gate before BrowserWindow startup and exits with one availability JSON line. A nonzero helper exit keeps top-level reason helper-probe-failed; helperReason is present only for an exact protocol-v1 line carrying a fixed allowlisted reason, and its optional helperDetail must be 1–1024 printable ASCII characters. Invalid detail suppresses both helper fields. Every probe uses an explicit 16 MiB aggregate captured-output ceiling independent of tracing. With IPTVNATOR_TRACE_PLAYER=1, a non-empty helper stderr capture is emitted separately as one JSON-escaped stderr line whose stderr field is limited to 16,384 characters and whose truncated field is always explicit; trace-write failure cannot change the capability result. Installed-Snap CI enables Mesa EGL/GL diagnostics through this bounded channel. Any loader failure remains a stable native-view fallback, never a flag-enabled success.
  • In the exact packaged Flatpak /app context, reconstruct only Freedesktop Platform 24.08's immutable __EGL_EXTERNAL_PLATFORM_CONFIG_DIRS; its GL extension loader path comes from the sandbox cache. Flatpak CI must invoke the application-level --embedded-mpv-runtime-probe, not a direct helper probe that bypasses capability detection.
  • The packaged x64 Playwright smoke runs its fixture-contract target first and passes Chromium --ignore-gpu-blocklist so CI llvmpipe can expose WebGL2. This launch-only flag does not bypass the manifest, hash, loader, or helper capability gate; --no-sandbox remains root-only.
  • Bundled Linux releases must publish the exact source archives/git records, checksums, licenses, flags, patches, build scripts, and the pinned hwdata pnp.ids input. Each bundled package carries embedded-mpv-notices.json, THIRD_PARTY_NOTICES.txt, and the exact licenses/** files. CI may cache immutable source inputs, but regenerates notices and a VCS-metadata-free linux-frame-copy-runtime-sources.tar.xz for the current checkout on every run while retaining the exact pinned six recursive libplacebo submodule records. Each record is canonical full-commit safe/path; clone-depth dependent git describe annotations are discarded and never form part of the provenance identity. Its source index carries the globally sorted libplacebo directory/file/symlink inventory; file hashes, sizes, executable bits, link targets, aggregates, and canonical tree digest must match the trusted pinned checkout. The archive has an exact member/type layout and its metadata/archive-sha256.txt records must match the actual source archives. Concatenated tar/xz streams are inspected past every end marker. The final archive's SHA-256 and repository revision are copied into every bundled x64 package manifest; system and marker-only packages carry no source-archive binding. Automated Snap Store publication is allowed only after a public v* GitHub release contains both the Snap assets and exactly one matching source archive. Before any upload, the workflow hashes and inspects that archive, verifies its exact member/type set and size bounds, clean tag revision, pinned sources including the six recursive submodule records and exact libplacebo tree digest, legal files, and exact released tooling, then performs bounded extraction and static package validation for every Snap. That public-release boundary independently revalidates the exact strict meta/snap.yaml graphics/shared-memory contract and enumerates resources/app.asar, rejecting any archived electron-backend/native/** payload before publication. Its bounded ASAR header reader uses only Node built-ins and released local tooling, so the clean tag checkout does not require node_modules. Exactly one x64 Snap must have matching sourceArchive and sourceRuntime; any non-x64 Snap must remain marker-only. Checkout and the artifact-transfer actions are pinned to full commits; checkout does not persist credentials, and repository credentials are limited to download steps. A secretless verification job copies assets through no-follow descriptors, checks pre/post hashes, writes an exact receipt, repeats the complete source/package verification on a root-owned read-only snapshot, and transfers only that data through the pinned artifact service while its receipt digest travels separately through a job output. The dependent publish job runs on a bounded ubuntu-latest runner with no checkout or release-tag code, verifies that digest plus the exact receipt, asset hashes, and file-only layout, root-seals the data again, and installs Snapcraft directly. Store credentials exist only in its final fixed shell step, which resolves no PATH command, executes no released code, and exposes the credential only to each exact /snap/bin/snapcraft upload --release=edge process. Candidate/stable promotion is manual after installed-Snap frame-copy and missing-runtime fallback smoke; GitHub Actions never promotes automatically. Canonical maintenance docs: docs/architecture/embedded-mpv-native.md and tools/embedded-mpv/README.md.

Repo Skills

  • .codex/skills/iptvnator-nx-architecture/SKILL.md
  • .codex/skills/iptvnator-sqlite-db-worker/SKILL.md
  • .codex/skills/iptvnator-theme-style/SKILL.md
  • .codex/skills/iptvnator-ui-design/SKILL.md
  • .codex/skills/release-cut/SKILL.md
  • .codex/skills/release-notes/SKILL.md
  • .codex/skills/stalker-portal/SKILL.md
  • .codex/skills/xtream-electron/SKILL.md

Descriptions and trigger conditions are canonical in each skill's frontmatter; do not duplicate them here.

General Guidelines for working with Nx

  • For navigating/exploring the workspace, invoke the nx-workspace skill first when it is available - it has patterns for querying projects, targets, and dependencies. If it is unavailable, use pnpm nx show projects, pnpm nx graph, and project project.json files directly.
  • When running tasks (for example build, lint, test, e2e, etc.), always prefer running the task through nx (i.e. nx run, nx run-many, nx affected) instead of using the underlying tooling directly
  • Prefix nx commands with the workspace's package manager (e.g., pnpm nx build, npm exec nx test) - avoids using globally installed CLI
  • You have access to the Nx MCP server and its tools, use them to help the user
  • For Nx plugin best practices, check node_modules/@nx/<plugin>/PLUGIN.md. Not all plugins have this file - proceed without it if unavailable.
  • NEVER guess CLI flags - always check nx_docs or --help first when unsure

Scaffolding & Generators

  • For scaffolding tasks (creating apps, libs, project structure, setup), ALWAYS invoke the nx-generate skill FIRST before exploring or calling MCP tools

When to use nx_docs

  • USE for: advanced config options, unfamiliar flags, migration guides, plugin configuration, edge cases
  • DON'T USE for: basic generator syntax (nx g @nx/react:app), standard commands, things you already know
  • The nx-generate skill handles generator discovery internally - don't call nx_docs just to look up generator syntax