Feature-complete roadmap: user-story coverage and gaps

[JarbasAl]

🤖 Auto-generated by Claude Fable 5 (claude-fable-5) via Claude Code — NOT human-reviewed. Verify before acting. Claims marked ✓ were re-verified against source by the model; the rest come from a model-written inventory of the feat/media-backend-v2-port branch (42e7e21).

Feature-complete roadmap

Tracking issue for taking ovos-media from "MediaBackend v2 train ready" to feature complete: every user story implemented, tested (unit + e2e), documented, live-validated, and audited. This body is the living state — it gets edited in place as items land.

Blocked on the release train

Everything in the "fix" sections below lands after OPM #442 merges and #194 follows (the train's commit sequence). Until then only harness, spec, docs, and audit work moves.

Coverage gaps (from the full user-story inventory)

Spec conformance (OCP-1):

• ✓ Skill can_seek declaration (§4.3.1) unimplemented: catalog/catalog.py never reads can_seek from ovos.common_play.announce, and SkillPlayerAdapter.seek() unconditionally warns-and-drops. Skill playback can never become seekable. Implement announce-side ingestion + conditional seek delegation + MPRIS CanSeek passthrough.
• ✓ Relative seek: RESOLVED as a spec amendment. An ecosystem sweep found no emitter uses the spec's position field — the deployed wire is seekValue (absolute ms, GUI seekbar) and seconds (signed relative, OCPInterface.seek_forward/seek_backward — the skill-facing API, a pre-spec surface with mandatory backcompat). Player-side resolution of seconds is atomic against live position, which is stronger against concurrent seekers than requester-side resolution from a stale state read; OCP-1 §4.2.2 is being rewritten to document both real shapes with the corrected rationale. ovos-pydantic-models' seek model (position) needs aligning to the wire.
• §5 per-session isolation: RESOLVED — OCP-1 now states the registry MAY be satisfied by deployment topology (per-satellite embedded daemons; hub serves its local session). Originally: implementation is a binary default/non-default gate on a single player, relying on the one-daemon-per-satellite topology (documented in service.py). Spec text should state explicitly that the per-session registry MAY be satisfied by per-satellite daemon instances, or the gap is real.
• ovos.common_play.register_keyword: RESOLVED — §4.1 carve-out for non-mutating advisory emissions (control requests in, registration hints) landed; state authorship stays the player's alone. Originally: §4.1's reservation wording doesn't carve out advisory non-mutating messages. Spec clarification.

Wrong-session behavior:

• ✓ Wire state reports (media.state/track.state from media_backends/base.py) are bare Message(topic, data) with no context.session — OCP-1 §4/§4.4 require the originating session so hub consumers can demultiplex (peer spec-audit finding, verified file:line). Fix together with the dialog-session item below: stash the requester's session at play time, derive all state reports and notifications from it.
• ✓ All spoken dialogs (track.failed, queue.finished, no.playback.backend, …) land on the default session even for satellite-triggered playback — acknowledged in skill.py's own docstring. Fix: stash the triggering message's session at play time and carry it through END_OF_MEDIA / INVALID_MEDIA / retry-timer paths.

Coverage holes:

• PlaybackType.SKILL full-cycle e2e added on the v2 branch (5 cells: delegation verbs + wire states).
• ✓ Config live-reload: hot knobs (autoplay, validate_source with explicit-override precedence honored end-to-end incl. the voice skill, preferred_*_services) read live configuration; cold knobs pinned by an armed test (feat: hot-reload cheap media config knobs #207, converged).
• Shuffle/repeat/relative-seek/like-unlike e2e cells added on the v2 branch (+11 mutation-proofed, incl. a found-and-fixed unclamped negative-seek defect); MPRIS headless graceful-degrade verified live; MPRIS on by default with a Playlists stub and lifecycle logging, validated on a real bus (feat: MPRIS on by default with a Playlists stub and lifecycle logging #210, converged).
• ovoscope v2 harness built and converged (ovoscope#168: MediaBackendHarness reporter-capture layer + OCPPlayerHarness v2 repair + v2 mock; three adversarial rounds + an executed two-venv gate). Held draft until the v2 train releases, then floors bump and it merges — un-skipping media-plugin e2e cells fleet-wide.
• Remote backends (chromecast, mass, spotify, mpris) have review-verified v2 ports but no real-device live playback validation; remote verb confirmation (silently ignored pause on a dead device) needs a timeout/notify design.

Product polish:

• Voice intents for like/unlike, queue and track repeat (spec'd repeat.set mode), relative seek, shuffle state, and a bounded what's-next preview (spec'd §4.4.1 next_track/next_track_hint) — feat: voice intents for like, repeat, seek, shuffle state and queue peek #209, converged.
• OCP-1 §4.4.3 named-collection read query serving the derived history playlists — feat: named collection read query per OCP-1 §4.4.3 #208, converged.
• Play counts only accumulate for already-liked tracks; ranking can never surface a heavily-played unliked track. Decide global play-count vs documented limitation.
• Album/artist metadata intents mostly answer "no info" (MediaEntry has no album field). Decide enrich vs trim intents.
• Docs: technical-manual one-story page + plugin migration guide (queued with the docs campaign, keyed to the v2 wire-changes table).

Validation plan

• Several independent expert audits (concurrency on the dispatcher/lock model, UX on voice flows, adversarial on the merged train) and user-persona runs (fresh-install quickstart verbatim, satellite user, desktop MPRIS user), each producing findings triaged into this list.
• Final gate: voice-driven live round trip on released packages only, incl. restart/network-kill/bad-stream recovery on all three namespaces.