roadmap
This page holds the public 12-month roadmap (dated priorities, release status, adoption evidence and contribution opportunities) and the program structure, gates and traceability that follow it.
Status: Active. Tier 1 (nested developer preview) is implemented; tier 2 (daily-driver candidate) is open. Work items live as GitHub issues on the Tuna Desktop roadmap project board; this document keeps the program structure, gates, and traceability.
Architecture: Program architecture
Parity target: GNOME 51 on the Marlin image family. Decided 2026-10-01. The published Marlin GNOME image measured on 2026-10-04 contains GNOME Shell 50.5, so GNOME 51 comparison evidence must identify its separately upgraded reference payload.
Target platform: v1 targets every TunaOS variant, each qualified from its own image digest or explicitly unsupported with a recorded reason (ADR 0008). Marlin remains the GNOME 51 parity/performance comparator; reference and candidate use the same VM image family.
Variant evidence: Source-backed 14-variant inventory and digest qualification matrix; package availability and stale unsupported declarations do not satisfy runtime gates.
Planning system: Spektacular spec → plan → implement. Every delivery unit has a spek, a reviewed implementation plan, evidence, and an explicit gate.
Public roadmap
This roadmap covers October 2026–October 2027. Updated 2026-10-19. Tuna Desktop was previously named Roost; the code and packages were renamed in #536, while dated evidence keeps the old name (#504). This is the public contribution roadmap; program gates and requirement traceability describe the acceptance process. Work and evidence live on the roadmap board. Dates are planning windows, and release qualification depends on evidence.
Current status
Tuna Desktop has a nested developer preview, an actual Marlin VM test lane, and a signed experimental amd64 Marlin Tuna Desktop image. The daily-driver candidate remains open. TunaOS merged the flavor, then named Tuna Desktop, in PR #2991. Its qualified image passed an actual AWS KubeVirt bootc switch, shipped greeter selection and PAM login (retained evidence). The published marlin:tuna image built on 2026-11-14 (sha256:f7859108a9a3a4538929a8b9a34913bf37ba5d6396e51bb8538a7e59af2581db, tunaOS 48ca80c7) carries tuna-desktop 1.0.0beta.3-1, built from the v1.0.0-beta.3 tag; fixes merged after that tag are not in it. The beta.2 image (sha256:d0e65bce…, observed 2026-10-31) is described in the installed-session evidence. GNOME 51 remains the comparison baseline. The published Marlin GNOME image measured on 2026-10-04 still contains GNOME Shell 50.5; GNOME 51 comparisons require a separately recorded, fully upgraded reference payload. A build or merged feature alone does not establish parity or release readiness.
The latest installable prerelease is v1.0.0-beta.3, published 2026-11-14 (UTC), with an amd64 Debian-format package targeting Ubuntu 26.04 (SHA-256 e595f5c02b20c3aab3a88e0c68d777e151e9b209841405f225efed91b111ba7e). The release notes list its changes, known limitations and supported configurations; installation guidance describes supported routes. Debian 13 needs a separately qualified backport and artifact. This beta claims no v1 checklist item. It supersedes v1.0.0-beta.2, published 2026-10-30. The older v0.0.0-ci media bundle remains documentation media, not an installable release.
On 2026-10-19, the GitHub release API reports 3 downloads of the beta.1 package and 7 downloads across the older media bundle's assets. These are separate asset-download counts, not installs or active users. Installed-user metrics remain unavailable; Tuna Desktop does not collect usage telemetry. The adoption tracker #652 owns supported configurations, tester feedback and maintainer continuity.
Direction: GNOME's functionality, better performance
Tuna Desktop aims to match GNOME 51 capabilities while beating its responsiveness and matching CPU and memory on the same machine. ADR 0008 leaves GPU and total-power comparisons as follow-up work; it does not waive CPU, memory or latency. The v1 checklist remains open.
Performance has improved since the October 4 baseline, but the gap is not closed. The recorded #696 paired Marlin measurement reports whole-trace CPU of 28.3% for Tuna versus 22.8% for GNOME. #719's bounded overview chrome is now merged; its local overview result is not whole-trace or GNOME parity. #503 remains the performance outcome, and #547 is the landed regression-gate implementation for #315. A passing regression budget above ratio 1.0 must never be presented as meeting v1.
Next three months: through 2027-01-04
- Finish and validate the core shell journeys: workspace insertion, app-grid hover/page behavior, notifications, shortcut consent, IME caret placement and live settings interoperability. Keep concrete GNOME differences in the parity ledger and pixel comparisons in GNOME parity captures, with actual graphical acceptance evidence.
- Preserve the qualified packaged portal consent/lock/revoke core from closed #61 (ledger P-SY-05). Complete the remaining provider/frontend and shipped-image acceptance in #776, and modern RemoteDesktop capabilities in #358 / #577. Require verified ownership, revocation and input delivery for each advertised capability.
- Refresh and qualify the admitted signed amd64 Marlin Tuna Desktop image: current-source package, clean installation, graphical login, portals, PAM, boot and update/rollback evidence. Historical flavor admission and boot/login evidence (#69) does not qualify a later package. In parallel, record each other TunaOS variant's install, real-greeter login, session, portals and lock/unlock evidence against its own image digest, or its explicit unsupported reason (ADR 0008). This does not invent new ISO or LUKS acceptance cells.
- Run the actual 24-hour endurance and comparable performance work (#203, #73). Record binary/image provenance, resource growth, frame pacing and failures. Investigate measured regressions before making performance parity claims.
- Record session lifecycle qualification (#62, #68): real VM suspend/resume, VT switching, logout and fail-closed lock behavior, plus the separate physical GPU/driver and mixed-display support matrix. AWS KubeVirt VM evidence covers its own profile only.
- Port every GNOME 51 animation (epic #492, spec). An audit on 2026-10-08 found 6 of GNOME 51's 54 animations ported and 7 partial. Deliver, in order: GNOME's motion policy (reduced motion keeps fades, slow-down factor) and a deterministic test clock; window open and close; minimize; maximize and tiling; workspace slide and swipe; modal dimming; overview corrections; app grid; menus and transient surfaces; lock screen and desktop transitions. Each family needs a CI gate and a paired GNOME 51 capture, and must cost less CPU than GNOME.
- Beat GNOME 51 on CPU, frame pacing and overview response on the same Marlin VM (#503). Profile first, then set regression gates (#315).
- Qualify the remaining transition work after the rename: #505 closed with compatibility removal in #775, but its existing public upgrade review remains unresolved. #506 owns current media provenance, and #577 links the remaining capability and release qualifications. The rename itself already landed in #536; do not reimplement it or infer upgrade qualification from issue closure.
Six to twelve months: 2027-04-04–2027-10-04
The next delivery stage is a daily-driver candidate after the security, usability, hardware, recovery and performance gates pass. Architectural work should reduce unnecessary repainting, complete protocol/application compatibility, improve fault recovery and preserve the bounded compositor critical path. A faster benchmark must retain the same workload and measurement scope.
Ecosystem work centers on GTK, Qt and XWayland application journeys; accessible keyboard and screen-reader use; live GNOME settings compatibility; and reliable existing system services. Broader hardware and distribution support should follow a named, reproducible support matrix rather than an untested compatibility claim.
Adoption work should make the qualified amd64 image and its install/recovery instructions usable by external testers, collect reproducible issue reports, and publish the supported configurations and known gaps. Report package download and contributor counts when the publication tools supply them. No numerical active-user target is set without a way to measure it; collect feedback voluntarily and preserve the no-telemetry default.
A 1.0 release requires the global decision gates. This roadmap does not assign a release date before those gates pass.
Known limitations and scope
Screen-reader speech, physical GPU/VT/suspend behavior and completed 24-hour endurance are not established by nested graphical tests. ADR 0008 permits named emulated-device evidence for its listed hardware-only rows while retaining their implementation requirements; physical qualification remains explicitly separate. GPU and total-power comparisons are follow-up, not current v1 gates. Settings support is key-specific; see the compatibility map. Modern RemoteDesktop features must remain unadvertised until implemented and tested. Exact visual parity requires paired captures from the declared reference version.
Public extension APIs remain optional later work, outside the initial core release unless scope changes through the recorded decision process. In-process shell extensions, an unbounded plugin authority model and unsupported platform claims are outside this roadmap. No shipped feature deprecation is scheduled; protocol version changes must preserve the documented compatibility contract or explicitly reject incompatible peers.
Where contributors can help
Useful contributions include reproducible GNOME 51 comparisons, accessibility and language journeys, hardware support reports with exact kernel/GPU/driver/topology, protocol and service failure regressions, comparable performance traces, and packaging/recovery verification. Start from an open roadmap issue and identify the acceptance evidence your change supplies.
Use the existing test strategy and report real runtime failures. Keep unsupported capability advertising, new image/ISO matrix expansion and public extension work out of unrelated fixes. Source tests, screenshots and a successful boot each establish different parts of the release evidence; retain their limits when updating an issue.
Delivery tiers
The project board's 1 Nested preview, 2 Daily-driver candidate, 3 Later, and Cross-cutting values map to the architecture tiers. A merged implementation is not a passed release gate: physical hardware, security, visual comparison and endurance evidence must still be recorded. ADR 0007 records the baseline and target choices.
Delivery sequence
flowchart LR
S0[000 Baseline and spikes] --> S1[001 Nested compositor and recovery]
S1 --> S2[002 Shell workflow parity]
S1 --> S3[003 Hardware and app compatibility]
S2 --> S4[004 Secure session, accessibility, integration]
S3 --> S4
S0 --> S4
S0 --> S6[006 Performance, packaging, release]
S2 --> S7[007 GNOME 51 animation parity]
S3 --> S7
S7 --> S6
S2 --> S6
S3 --> S6
S4 --> S6
S0 --> S5[005 Extension broker and API]
S1 --> S5
S2 --> S5
S4 --> S5
Specs 002 and 003 may proceed in parallel after the 001 compositor/control contract is stable. Spec 004 may prototype service/accessibility work earlier, but its production acceptance depends on shell and hardware behavior. Spec 006 is the integration and release dossier. Spec 005 is optional post-core work and is not a prerequisite for 1.0 unless release scope explicitly changes.
Near-term delivery queue
Snapshot: 2026-10-19, reconciled through main fb4abac3, after the roadmap refreshes in #770 and #773. The linked PRs distinguish landed implementation, remaining acceptance, and candidates still awaiting qualification. Check the current head and its evidence before acting: a green ancestor, closed PR, merged stack branch or bot label is not qualification of current main.
Landed foundations
-
#758/#763: checkout/token handling; #759/#767/#769: RustSec gate and its API permissions. #766 subsequently landed the tokenless, fail-closed replacement at
ecd46d18, removing redundant Checks/Issues write grants. -
#765: all six previously omitted Python gate suites are wired into CI; #737 is closed. #733's checkout and registry-token work is also closed.
-
#706: bounded native brightness contract. #355 remains open for the complete power-policy and shipped-image outcome.
-
#719: bounded transparent overview chrome. #503's paired performance target remains open.
-
#764: pinned nightly fuzz toolchain update, landed at
da4dee0aafter exact-head CI and decoder fuzz passed. -
#708: accelerator-only reconstruction landed by ancestry-preserving merge at
02e5a678, on protocol 0.38 after all 23 checks passed on reviewed head1baef8ac. This includes genuine GNOME portal press/release and lock cancellation evidence. Shipped-image, real native/Flatpak frontend, rebind/persisted-permission and remaining ownership/race acceptance is tracked by open #776. That focused follow-up preserves the unfinished criteria from #395, which automation prematurely reclosed; already-qualified backend work is not reopened or counted twice. -
#772: wellbeing fade-off proof repair landed at
bd153775after check, installed journey, full GTK shell and aggregate gates passed on reviewed headb7094ae0. The proof now waits for a later neutral-frame submission at the same frozen animation clock; pixel tolerances and the 400 ms resize-animation bound are unchanged, with a separately bounded submission wait. Submission is not physical presentation evidence, and #747 remains open. -
#768:
freedesktop-desktop-entry0.8.4 and the single-decoding Exec adaptation landed atbc7cc21fafter all 21 checks passed on reviewed head99cb66a2. Escaped quotes, backslashes, dollars, spaces and newlines have regression coverage. The final diff contains only the dependency manifest, reviewed lockfile and app-discovery adaptation. All 19 post-merge CI jobs and measured coverage also passed onbc7cc21f. Its DRM result used the older6.17.0-1022-azurekernel; the newer-kernel package-discovery correction landed separately in #774. -
#774: shared CI repairs landed at
7bda47daafter all 22 exact-head checks passed. These cover old-head audit compatibility without privileged candidate execution, exact-kernel VKMS package discovery, isolated Night Light watch timing, fresh portal parent identity and bounded native audio failure receipts. Assertions and product deadlines remain intact; retained transient failures are not erased. -
#771: fingerprint relay saturation causality, owned per-test GSettings backends and bounded brightness failure diagnostics landed at
a249267fafter all 22 exact-head checks passed. The merge tree exactly matches the qualified candidate; the old coverage crash and intermittent native failures remain retained evidence, with no unproved root cause asserted. -
#453: native Orca keyboard interoperability and running-app search activation landed at
ea17be1eafter all 23 exact-head checks passed. Both distribution Orca and separately packaged upstream Orca 51 passed genuine keyboard commands, spoken Search/Disks activation, Quick Settings navigation, lock/PAM, VT/S3, fresh authentication, repaint and owned-reader logout. The landed tree exactly matches the qualified candidate. Mixed-sink PCM receipts do not establish exclusive audio origin; full spoken GTK/shell/lock/login and PointerLocator acceptance remain open in #352. -
#458: bounded native SVG backgrounds and the separately packaged
tuna-wallpaper-svghelper landed at0cd23760after all 39 exact-head checks passed. Actual GNOME Settings selection, independent native reference pixels, generated Fontconfig cache provenance, helper/resource negatives, light/dark/rewrite/restart and locked rendering passed with unchanged pixel tolerance; native VM, GTK, installed-package, visual and stress lanes passed too. The combined tree with #453 was reviewed. Physical multi-output, final-image and broader font/loader/timeline acceptance remain open in #452 and #360. -
#775: compatibility removal landed separately at
1d6e0c16. The existing public upgrade review led to the bounded transition repair in #780: exactly two temporary package identities and non-destructive old-only settings adoption. See upgrade and recovery evidence. Actual GDM reselection, old-compositor unlock, pacman replacement and Marlin activation/rollback remain separately unqualified; the compatibility-removal merge does not establish those outcomes. -
#547: GNOME-relative regression budgets and per-phase/process attribution landed at
2bf79f3bafter all 24 exact-head checks passed. The calibrated stock profile, thresholds, floors, ceilings and repeat counts remain unchanged. ADR 0009 makes measurement-code PRs report-only and scheduled/enforced runs fail on exceeded or missing enforced metrics. A green report-only run does not establish a budget pass, GNOME parity or release qualification; #503 remains open. -
#777: the one-file peer-auth retirement fixture repair landed at
e901747a. All ten native peer-auth tests and measured coverage passed; all 21 check results were acceptable (11 passed and 10 path-selected skips), including the selected GTK and native DRM proofs. It replaces a fixed poll-count assumption with observed retirement under a bounded test-only deadline, retaining exact one-session admission. Production authentication and deadlines are unchanged, and the retained intermittent incident is not assigned an unproven cause. -
#778: the
sha20.11 dependency update and exact receipt-encoding adaptation landed at47fccae9after all 22 checks passed, including the full 20-job native matrix and measured coverage. The original compile failure is covered by known digest vectors, leading-zero file receipts and 98 comparisons with the older encoder. Dependency sources/checksums/licenses were reviewed; original authorship and DCO are preserved.
Unmerged work requiring a decision
| PR | Scope | Landing requirement |
|---|---|---|
| #561 | WM shortcuts and focus policy / #347 | Head b6f3c803 passed all 24 checks, including genuine focus-mouse and native lifecycle proofs. The PR was closed unmerged on October 19; the feature is absent from main. Its disposition needs maintainer reconciliation before any reopening. Preserve #347's unimplemented/unqualified remainder and protocol 0.39's additive ordering. |
Resolved stack dependency, deferred scope: #708's reconstruction excluded #707 (EDID identity), #710 (brightness journal) and #713 (native display lifecycle); the original stack is preserved on backlight/accelerator-stack-preserved-2026-10-19. #707 and #713 closed without merging into main; #710 merged only into the former stack branch. They remain unshipped scope, not prerequisites for the landed accelerator implementation. Track full display acceptance in #340 and power acceptance in #355. If #561 is revived, preserve its additive protocol positions after the landed 0.38 baseline.
Post-merge qualification: main bc7cc21f passed all 19 CI jobs in run 39305891502, and the subsequent documentation tree c5196347 passed all 19 jobs in run 39307728389. Older 02e5a678 / bd153775 runs retained the newer-runner VKMS package failure; #774 now contains the reviewed discovery repair with exact-head native DRM evidence. Main 7bda47da also passed all 19 jobs in run 39320146000. GitHub superseded queued a249267f, ea17be1e and 0cd23760 CI runs as later merges arrived; those cancellations are not test results. The separate a249267f coverage run 39322667046 failed a peer-auth reconnect assertion (two sessions observed instead of one); subsequent ea17be1e coverage passed. The fixture retirement sequence is repaired in landed #777, with native tests and measured coverage passed but no claimed incident root cause. Main 1d6e0c16, including the external compatibility removal, passed all 20 jobs in run 39326409262 and measured coverage. Those fresh-session proofs do not qualify legacy-state upgrades. Main 2bf79f3b CI run 39330737912 subsequently passed all 20 jobs. Queued e901747a and 47fccae9 CI runs were superseded by later merges; their measured coverage passed. The combined main tree fb4abac3, including #777, #778 and the #779 roadmap refresh, passed all 20 jobs in run 39336110655, completed October 19 at 23:27 UTC. This satisfies the stated default-branch incident condition in #735, now closed. The distinct intermittent investigations #747 and #748 remain open. A passing later run does not prove a proposed cause for a retained intermittent failure.
All landing candidates need review of the actual final tree, DCO sign-off, and successful relevant CI on that exact head. Do not waive a failed runtime/ownership/lock assertion or mark a draft ready on the strength of unrelated checks. After each merge, observe main CI before attributing later failures to another change.
Issue priorities and release evidence
- Reliable qualification: #735 is resolved by the verified green combined-main run above. #747 remains open after the bounded #772 proof repair, and #748 tracks lifecycle investigation. Diagnose their retained failure evidence and preserve fail-closed gates; a resolved baseline incident does not close those specific investigations.
- Capability acceptance: #577 is the canonical capability/release epic; #367 tracks Settings coverage, #397 actual Core apps, and #492 animation families. Keep linked issues open where implementation or installed/native acceptance remains incomplete; no mass closure from merged code or nested-only ledger rows.
- Performance: #503 owns the GNOME-relative CPU/latency outcome. #547 landed the calibrated regression gate; #315 remains open for the broader measurement/release commitments, including memory/growth budgets and variance guidance. Run comparable fresh measurements after performance changes, then tighten only with evidence.
- Image and recovery: qualify current packages on Marlin and every other TunaOS variant (or record unsupported reasons), using the beta/release runbook and v1 checklist. Old flavor admission, package build success and a nested tour are distinct evidence.
- Adoption and remaining decisions: #652 owns the support matrix and tester feedback; #506 owns media provenance. #505 closed with the compatibility removal in #775. The bounded repair and remaining native evidence are tracked by #780 and the upgrade guide; release qualification must not infer upgrade safety from #505's closure. RustSec maintenance/soundness warnings still need triage under the existing advisory policy.
Milestone mapping and dates
- GNOME 51 capabilities: #577, landed accelerators, Orca and SVG support, open #776 frontend/shipped-image acceptance, and the WM candidate with its explicit remaining acceptance work.
- Performance versus GNOME: #503 and #315; #719 is an implemented optimization and #547 is the landed regression gate, while the performance outcome remains open.
- Provenance and images: current package/image qualification for every variant or an unsupported disposition; Marlin remains the comparison platform.
- Release checklist: publish additional beta tags only from qualified main under ADR 0008. GA has no promised calendar date before v1 and the global gates pass; this refresh does not invent a Q4 commitment.
- Login screen: Tuna greeter on GDM's daemon, replacing the GNOME Shell greeter on every variant (ADR 0008 item 6, epic #1164, ledger P-LK-03).
Spek and plan registry
| ID | Spek | Purpose | Depends on | Release relationship |
|---|---|---|---|---|
| 000 | Program baseline and technical spikes | Baseline GNOME, compare toolkit/accessibility, Smithay/protocol spike, threat model | None | Required evidence before committing major implementation decisions; nested prototype may proceed with provisional assumptions |
| 001 | Nested compositor and shell recovery | First buildable vertical slice; app mapping and supervised shell recovery | 000 inputs | Developer preview foundation |
| 002 | Shell workflow parity | Overview, launcher/search, panel, quick settings, notifications, workspace journeys | 000, 001 | Required for daily-driver candidate |
| 003 | Hardware and app compatibility | DRM/KMS, outputs, XWayland, scaling, IME, data devices, devices and rendering fallbacks | 000, 001 | Required for daily-driver candidate |
| 004 | Secure session, accessibility, and system integration | Lock/auth, portals, AT-SPI journeys, settings and existing services | 000, 001, 002, 003 | Mandatory 1.0 security/usability gate |
| 005 | Extension broker and public API | Capability-limited, out-of-process extensions | 000, 001, 002, 004 | Optional later capability; excluded from initial 1.0 unless explicitly promoted |
| 006 | Performance, packaging, and release readiness | Benchmarks, CI/conformance, package/rollback and release evidence | 000, 001, 002, 003, 004 | Required for daily-driver candidate |
| 007 | GNOME 51 animation parity | Every GNOME 51 shell and window animation with GNOME timing and motion preferences, cheaper than GNOME on the same VM (epic #492) | 001, 002, 003 | Required for daily-driver candidate |
The plan store contains plan.md, context.md, research.md, and test-plan.md for each unit. Plans are drafts until reviewed against current source, confirmed dependencies, and explicit acceptance evidence. Do not begin implementation from a draft plan if a blocking decision or security boundary is unresolved.
Requirement traceability
Parent requirements R1–R16 are defined in the
parent requirement register. A bare R<n> in this
document is a parent requirement; a spec's own requirements are written
with their spek number (001-R5), per the register's naming rule.
| Parent requirement area | Owning spek(s) | Evidence at completion |
|---|---|---|
| R1 Activities/overview; R2 launch/search; R3 panel/system controls | 002; 004 for service backend | Baseline journeys, recordings, responsiveness/failure cases |
| R4 windows/workspaces/input gestures | 001 foundation; 002 parity; 003 output/input compatibility | lifecycle and focus tests; hotplug and gesture journeys |
| R5 notifications | 002 UI/behavior; 004 service and locked privacy | freedesktop notification matrix, history/actions/DND and locked behavior |
| R6 session/lock | 004 | lock fail-closed, PAM handoff, VT/resume/hotplug fault evidence |
| R7 applications/peripherals | 003; 004 for portal capture | GTK/Qt/XWayland, clipboard, DnD, IME, screen-share consent/revoke |
| R8 accessibility/language | 000 toolkit spike; 002/004 implementation | keyboard/AT-SPI, large text, high contrast, motion, touch, RTL journeys |
| R9 settings interoperability | 004 | published settings compatibility map and migration journey |
| R10 shell failure isolation | 001 | shell kill/restart with app surfaces alive and state resync |
| R11 extension boundary | 005 | quotas, revocation, crash/resource containment/security cases |
| R12 bounded critical path | 001, 002, 005, 006 | tracing and fault tests show no blocking on shell/extensions/search/I/O |
| R13 display correctness | 003, 006 | scale/hotplug/mixed-refresh matrix with measured behavior and fallbacks |
| R14 capture/privileged surfaces | 000 threat model; 003 integration; 004 enforcement; 005 extension denials | spoofing, consent/revoke, unauthorized layer/capture tests |
| R15 recoverable configuration/upgrade | 006 | broken package/config recovery without user-data loss |
| R16 comparable performance | 000 baseline; 006 release comparison; 007 animation cost (#503) | raw comparable traces, variance and soak results |
Global decision gates
- Before 001 implementation: provisional nested backend, Smithay/protocol versions, event-loop approach, control API envelope, recovery affordance scope, and nested-session safety are recorded. Production same-UID identity remains a release blocker, not a prototype assumption.
- Before 002 toolkit lock-in: toolkit spike demonstrates panel/overview, layer surface, AT-SPI screen reader, IME, RTL, large text, reduced motion, and measured resource/frame behavior.
- Before hardware claims: support matrix names distro, kernel, GPU/driver, output topology, protocol revisions, and tested fallback.
- Before secure implementation/release: threat model assigns trusted identities and lock/auth/portal/capture authority end to end; specialist review has no unresolved critical issue.
- Before 1.0: acceptance of 002, 003, 004, and 006; no critical lock/capture/input/clipboard/IME/accessibility regression; known compositor crash boundary documented; upgrade recovery and benchmark artifacts available.
- Before public extensions: initial use cases map to a small capability set, API compatibility policy and permission UI are reviewed, and all broker containment criteria pass.
Research intake protocol
GNOME issue reports are leads, not current-state facts. For every issue used to prioritize work, record exact project and issue ID, state, labels, last activity, affected release, reproduction, linked merge request/fix, verification against the pinned baseline, and the Tuna Desktop spek/test that covers it. Do not call an issue unresolved from search snippets. The intake checklist lives in research intake.
Change control
- A scope move between developer preview, 1.0, and later requires an ADR with user impact, dependencies, schedule/security impact, and updated traceability.
- A parent requirement change updates owning spek(s), acceptance criteria, roadmap, and plan dependencies together.
- Performance thresholds can change only with repeated comparable evidence and a recorded ADR.
- User-facing parity deviations need severity, owner, rationale, and disposition; undocumented gaps are not accepted as parity.