Repository navigation
Description
Highest-priority delivery objective
Tracking issue — closure rule: Keep this issue OPEN until all body/comment checkboxes are checked and each item has verified supporting evidence. A merged PR, completed subtask, or partial delivery is not grounds to close this issue.
HIGHEST priority: ALL body/comment checkboxes in #4839 and #3731, their child acceptance, and ALL demo-labelled issues. Continue implementation, review, validation, PRs and merges without repeated “continue” prompts. Keep one owner per lane and update both bodies and comments at verified milestones. The overnight/end-of-week target does not waive gates.
Deliver complete parent outcomes, not merely subtask receipts. Preserve unfinished criteria and the HVF-first, library-first, tracing, custody and backend/security requirements below. Source review or scoped tests do not prove end-to-end acceptance. Escalate only a precise unresolved decision with its next executable step. Existing specific approvals govern; this priority adds no signing, publication, paid/live, old-state cleanup or credential authority.
Current demo inventory: 12 open issues, verified 2026-10-30. All twelve owning scopes remain in scope. Unrelated work must not displace these deliveries.
Current implementation and verification batch
Active delivery queue
These are active implementation/delivery assignments, not completion claims. Broad acceptance below remains unchanged.
Live delivery ledger — 2026-10-30
Completion includes body AND comment checklists; keep verified states synchronized. One delivery owner per lane; no new authority.
| Lane / sole owner | Verified progress and next action |
|---|---|
| Client parity / facade owner | #4877 97615b4f: full local/hosted gates passed. Manifest removal/ABI1.12 generation underway; ratchet still 65 files/76 reaches, root exec outside. No native/Studio parity. |
| CLI teardown / PS-21 owner | #4881 d53cc219: scoped/static/gated/hooks +hosted CI pass. Local HOLD: original attestation assertion unresolved; instrumented full run failed earlier on unchanged span wall-clock ceiling. Same-binary failure showed off-CPU time, not its cause. Finite participant build-drain planned; no waiver. |
| Volume safeguard / volume owner | Step1 #4880 landed 511e354e (receipt). Step2 original-FD custody/startup adoption in implementation; Step3 all-path enforcement open. No rollout/cleanup. |
| Capture exclusion / capture owner | V4 landed: #4994 c3fc5831, full gates green. Provider-limited owner handoff pending; no native/all-error claim. |
| Builder ownership / runtime owner | #4879 EOF foundation landed 39afac26. Fresh-owner/terminal/durability +bounded extraction composition active; factory unqualified. First route FlakeImage only; unsupported modes refuse before birth. Reviewed Debug foundation in full validation. |
| Toolchains / product-specific owners | Studio32 c37c2a26, packs86 85ca6897, images73 32bb05ba landed; FC witness skipped. MVM4876 Nix/compiler/two-arch/repro/boot open. ValidationJob design assigned for same-owner implementation, not execution proof. |
| Application / integration + producer owners | 4509 originals1/2/3/4/6 checked; genuine Sigstore positive open. 4510 all7 hermetic criteria checked, producer dependency open. 4511 OCI-fallback fix: RED +58 scoped passes/source clearance; full gates/protected delivery pending. |
Published ABI-1 ARM builder assets and the frozen input packet exist. Fresh M4 isolated build and reviewed/tested builder-only test identity are approved; implementation safety gates still apply. dm-verity finalization is not signing. Official signing/publication, live-model acceptance and old-VM cleanup are separate. PS-06, PS-21 and both umbrellas remain open.
-
Land configured local-policy/Dev defaults #4853,
5bb761c5…; verified source/full-validation/queue receipt. Stable Rust API preserved; not public-pack or native demo acceptance. -
Land signed-budget fold correction #4859,
7a9061aec8fb9d0e71c613a401bd9b062c85d2a0; protected queue40974115784 passed. Budget ordering/transport only; schema, required-egress and no-policy defaults unchanged. No live demo acceptance.
Detailed receipts preserved verbatim for the compact completed milestones below.
Landed bounded milestones (#4843/#4844/#4846/#4849/#4850 and Studio#31) are recorded separately; they do not complete PS-06, PS-21, the <1M-line goal, performance targets, or native acceptance.
-
Repair launch-plan documentation coverage test(bdd): witness the documented manifest launch-plan command #4843: landed; exact post-merge BDD 345 scenarios/1,434 steps passed.
-
Land BDD selection/filter repair fix(bdd): reject empty selections without bypassing safety gates #4846: 355 scenarios/1,472 steps passed; original capability/shard gates retained.
-
Land preparatory FC model docs: define preparatory Firecracker owner contract and model #4849: 14 model tests; not native role/containment approval.
-
Land existing signer-proxy response correlation fix: reject uncorrelated signer responses #4850: 11 tests; no new signing purpose or custody claim.
-
Land isolated verification/cache fixture corrections test(build): stabilize cache verification fixtures #4855 (
e4dc6da7…); protected queue and 20,778 Linux tests passed. Same-size tamper/count assertions and production verification/settling preserved. Preserved detailed receipt. -
Land read-only cgroup diagnostics feat: add read-only cgroup diagnostics #4852 (
1710a20a…); exact-head/merge-group CI, eight Linux tests and ABI invariants passed. Non-atomic observations are not placement/death authority. Preserved detailed receipt. -
Land canonical registry read-error refusal: #4851 merged as
033e6dc093eec805dc7fbe8823c7821ff16adc88; revised-head CI40941022825 passed. Actual readNotFoundalone initializes empty; other failures propagate. No PS-21 closure implied. -
Land the bounded HVF contract/untrusted-parser prerequisite: #4856,
81c1eae8647c216fef7cfcf90dca73b364c42c20; protected merge-group passed. Canonical receipt. No owner runtime/authentication or native acceptance enabled. -
Land prepared-parent lifecycle-history snapshot correction fix(hvf): allow prepared snapshots with retained host history #4857: exact merge/main CI passed; native capture acceptance separate.
-
Reconcile all five MVM-side PS-21: Replace checkout identity and manifest emission with in-process APIs #4412 acceptance criteria against MVM
e8b3c0f…/ imagesf5ed151…: itemized source/platform/artifact evidence. Includes strict consumer/V1 refusal, real mutation and supported raw-name witnesses. Paired-owner final reconciliation and images#67 remain open; not whole PS-21 or release acceptance. -
Land cleanup #4854,
4fdeb6ac…; reviewed source, local workspace/gated checks and protected CI passed. Detailed receipts retained in PR. Broad footprint/performance/parity criteria remain open. -
Land bounded root-duplicate validation #4861,
ee9523883d689210f5fa6aa9e212d3afe17e7c7b: verified receipt, protected queue40978060930 passed. All schema attributes preserved; no new grant or native authority. -
Land pause-during-throttle correction #4862,
8b65f63b87b9c8d79783ec55e60609f8fb33b3e8; protected queue40986056737 passed. Deterministic pause-before-guest-reentry regression retained; not native/capture-race acceptance. -
Land existing-purpose SignPlan validation #4863,
803360d4ba7a914eaf539b4de669fb50316a2cd0; verified queue/source/test receipt. Guard precedes signing, not key loading; no principal/policy/admission authority. -
Land shared checkpoint removal/prune guards #4864,
9850dddf0b5ba904f791a09bfcb4d19ea10d4b81; verified review/test/queue receipt. Client/CLI/hostlib/SDK parity and incomplete-record preservation delivered; no atomicity or blanket lifecycle guarantee. Other PS-21: Finish client execution ownership and hostlib SDK parity #4415 work stays open. -
Land progress-driven poll regression #4865,
0833123bf3c004c13186d949e03820c9e36376c3; merge-group40990793768 passed. Count assertion/watchdog and landed pause fix retained. Stop/application default-suite reruns remain required. -
Land diagnostic export preparation #4866,
59203238c2086974975d3efd94223124f091b468; evidence. Review, seven hermetic tests and protected CI passed. Test adapter/runbook only—not an archive, signing/native witness or runtime prerequisite. -
Land failed-stop state preservation #4867,
58c77967a38ddc4e8c2315b8673fedd8361cadce; protected queue40998829012 passed. Reviewed four-file patch, all 11 default-source validation commands and exact-head CI passed. No native custody/containment or complete macOS HVF: entrypoint boots (machine run --entrypoint, pack build health boot) are refused until native input custody lands #4655 claim. -
Land FC bounds refusal #4868,
708dd28dbf257d3a677be132353b1cd3e154285e; source/test/queue receipt. Guard/fallback preservation verified; not native migration, finite-profile or owner-death qualification. -
Land digest-source reporting fix(core): report sealed sidecar digest provenance accurately #4869 (not latency/native proof). Verified receipt.
-
Land prefilled consumer stdin fix(demo): prepare prompt pipe before consumer spawn #4870 (not live demo acceptance). Verified receipt.
-
Land bounded snapshot refusal reader fix(vmm): bound snapshot refusal record reads #4872. Verified receipt.
-
Land factual SecretStore/KeyProvider ADR correction docs: distinguish credential storage from tenant key retrieval #4871. Verified receipt.
-
Land runtime mock-state isolation/diagnostics test(runtime): isolate mock lifecycle state and expose I/O failure kinds #4874 (
f4dc3c48…). Verified source/test/queue receipt. SDK integration remains separate. -
Land typed Python/TypeScript telemetry reads feat(sdk): expose typed host-local telemetry reads #4873 (
d1fc49b5…). Verified source/SDK/workspace/queue receipt. Existing host ABI reused; Studio live UI/native parity remain open. -
Land PATH-independent owned-child stop fixture test(backends): remove PATH dependency from stopped child fixture #4878 (
4b3db884…); protected queue41089095983 passed. Actual child spawn/wait and assertions retained; not production stop or native proof. -
Land persisted-definition read parity and GC refusal fix(client): observe persisted definitions through local facade #4877 (
97615b4f…). Source landing; all final gates passed. Not full PS-21/native parity.
Cleanup PR#4854 is landed; validation records below retain their original source bindings. Studio local-first retirement landed in PR#31 at 94a470d36f3120c16dc204c7d196c6e6da19f436 after remote CI passed. Broader acceptance below remains open.
- Validate consolidated MCP adapter: 38 adapter tests plus 11 tool-gate tests; shared third-party closure retained.
- Validate removal of the private CaptureRequestBuilder: the 32
commands::deps::tests passed, with direct named fields replacing unreachable missing-field states and existing tracing/audit preserved. - Run pre-Make pair-image tests hermetic and bound macOS validation time #4840 workspace fmt/check/Clippy and Linux capture-test compilation; compilation is not Linux execution.
- Validate cleanup
98efd394…: full workspace tests/doctests, fmt/check/Clippy and both gated arms passed. #4854 retains exact commands/source and historical failures. Not performance/umbrella acceptance. - Review/validate the first Make pair-image tests hermetic and bound macOS validation time #4840 synthetic-checkout slice: five native macOS tests and workspace Clippy passed; production identity/proof logic retained. Remaining Make pair-image tests hermetic and bound macOS validation time #4840 acceptance stays open. Preserved detailed receipt.
- Land reviewed Studio retirement docs(plan-34): fold 12 cross-cutting considerations into the L7 egress plan #31 at
94a470d: stop/remove and delayed-output regressions passed; current/native parity separate. - Correct the independently reviewed HVF snapshot capture concurrency gap before claiming racing requests are nondestructive. Scope includes pause ownership, aliases/handoff, stale-output cleanup and consumption. V3 late-writer reuse was reproduced; V4 landed as #4994 (merge
c3fc58315787593305dfc85b8ed0e4f073584460, all checks green): snapshot-vs-resume serialization with pause ownership, aliases/handoff, stale-output cleanup and request-keyed consumption, with regression tests. fix(hvf): a refused snapshot capture no longer kills the guest, and says why #4836 retries, the 30-second deadline, authenticated observations and diagnostics preserved. Post-land scoped reruns 2026-11-02:cargo test -p mvm-runtime --lib instance_snapshot45/45,cargo test -p mvm-vmm --lib snapshot31/31. No native/cross-generation safety claim from scoped tests.
Next #4840 implementation is no longer discovery-only: the test-only unbounded child wait has been replaced with the existing event observer, execution deadline, owned-child cleanup and file-backed diagnostic capture. All five helper tests and one existing cache/proof regression pass; Clippy and independent review remain in progress. Commands/results. No production VM lifecycle or dependency changes.
Reproduced footprint at public 511e354e: 2,804 Rust files / 1,235,719 inclusive physical lines / 47,170,593 bytes; strict <1M gap 235,720 lines. Exact method, SHA-256 inventory and independent reproduction: vendored 76,259 lines; header-marked generated 378; dedicated tests/benches 93,254; remainder 1,065,828 includes inline tests/unmarked generation, not production SLOC. Historical dirty 2,807/1,227,238 is not a controlled comparison/savings claim. No category exclusion or formatting compression.
Goal and authority
Outcome over checklist count. The maintainer clarified:
We can expand the checklists where necessary... the goal here is to simplify all the AI crap, the project, make it easier to maintain, faster to develop, easier to use, etc. etc. as documented in the issue so far
Expand acceptance criteria when investigation reveals necessary work; do not freeze scope merely to make the list easier to finish. Each addition should name a concrete problem, an owner or existing issue, and observable completion evidence. Avoid process-only ceremony and duplicate implementations. Small additions can be justified when they remove hidden states, unsafe coupling, or maintenance cost, but measure their cost honestly and keep the overall footprint/dependency goals intact. Preserve protected tracing/security/backend behavior and HVF-first demo priority. Checked items must correspond to achieved outcomes, not optimistic wording.
Substantially clean up and refactor MVM—not merely perform cosmetic edits. Reduce unnecessary code, dependencies, crates, features, and maintenance burden; remove AI-generated-looking narration and architectural ceremony; make shipped binaries small and fast and development/release builds predictable.
The maintainer explicitly authorizes removing crates established to be extraneous. A justified crate deletion does not need another permission request just because it removes a crate. Prove that it is extraneous through production consumers, supported targets/features, released contracts, and active issue ownership; small size or lack of default integration alone is not proof.
This umbrella owns scope, acceptance criteria, dependencies, and progress. Link focused child issues and PRs here as workstreams are established. Do not create another shared cleanup dashboard in specs/.
Priority: working demo and cleanup
Maintainer instruction:
Top priority: deliver the working demo across those repositories alongside this cleanup. Order cleanup so it unblocks or safely simplifies the demo critical path; do not disrupt active demo owners, signed artifact compatibility, or security enforcement for unrelated count reductions. Continue non-overlapping bounded cleanup in parallel.
Required demo backend order — latest maintainer instruction:
HVF is the first backend for the demo that must work
- HVF is the first required end-to-end demo backend, on the qualified native macOS execution boundary. A Firecracker, libkrun, WASM, Linux-only, mock, or compile-only result is not a substitute for that witness.
- Prioritize the existing macOS HVF release, input-custody, lifecycle/readiness, and latency owners and their blocking issues (Release lane is red: e2e control-plane failures and archive-smoke bootstrap gap block every release witness #4764, macOS HVF: entrypoint boots (machine run --entrypoint, pack build health boot) are refused until native input custody lands #4655, HVF warm claims take ~3.2 s on the macOS CI runner, so every warm launch is refused #4391) ahead of alternate-backend demo work. Preserve the original cold/warm budgets and do not bypass fail-closed custody/admission gates to obtain a demonstration.
- Keep the other supported backends intact, but their demo validation follows the first successful HVF path rather than replacing it. Trace signed artifact acquisition/admission, host-held credential binding, exact request budget, output/audit correlation, and cleanup through the actual HVF consumer path.
- Complete and record the first final-candidate HVF demo witness before claiming the demo works. Use compatible published/signed CLI-runtime-application inputs and the reviewed/landed budget policy; observe the separately required operator authorization for VM/model execution and publication.
Demo critical-path delivery checklist — all 12 tracked issues
Completion requires: full scope addressed, necessary implementation, reviewed PRs, passing validation, merged commits and required release/live evidence. Existing landed work counts; no redundant PRs. Plans or partial delivery do not. Separate operational/publication approvals still apply.
- mvm#4764 — Release.
- mvm#4655 — HVF custody/recovery.
- mvm#4511 — Signed launch.
- mvm#4510 — Atomic installation. Closed
COMPLETEDon 2026-11-09 with all 7 acceptance criteria checked; credential-free signed application acquisition and installation witness in #4510 (2026-11-09, released macOS aarch64mvmctl 0.27.0, empty privateMVM_HOME, pinned descriptor resolved to the independently trusted author key). Limit recorded there: the v0.27.0 release is a prerelease, so this is not final-demo client compatibility evidence; the demo boxes stay open. - mvm#4509 — Descriptor verification. Closed
COMPLETEDon 2026-11-09 with all 6 acceptance criteria checked; credential-free real signed application positive witness in #4509 (2026-11-09, releasedmvmctl 0.27.0: tar, checksum manifest, Sigstore bundle and provenance attestation verified;pull agent/claude@2.0.1exited 0 and reported the descriptor and bundle digests). Same limit: prerelease, not final-demo client compatibility evidence. - mvm#4508 — Execution baseline.
- mvm#4507 — Claude demo.
- mvm#4391 — HVF latency.
- mvm-packs#84 — Public HVF witness.
- mvm-packs#77 — Allow/deny rehearsal.
- mvm-packs#76 — Signed archive publication.
- mvm-packs#31 — Runnable environments.
Full scope and dependency details. All 12 boxes are umbrella closure requirements; the earlier audit is not delivery acceptance.
- Reconcile each listed issue against current implementation, PR/release state, active ownership, and missing acceptance evidence; record the critical-path dependency order. Read-only reconciliation and limits. This completes the audit task only; implementation/live acceptance remains unchecked.
- Complete the current demo contract across producer, registry, paired CLI/runtime consumer, policy enforcement, and installed-artifact execution. Preserve the exact request-budget constraint documented below; historical descriptors and old multi-request wrappers are not current release evidence.
- Witness the actual supported demo workflow with correlated allow/deny audit evidence and the required platform/artifact/client checks. Unit tests alone do not establish demo readiness. Explicitly record unavailable checks and operator-authorized publication/live/model gates.
- Close/reconcile demo issues only after their own acceptance criteria are verified; link the evidence here without substituting this umbrella for the owning issue.
The original empty image-repository demo-label inventory is not image readiness evidence. Existing owners retain their scopes.
Required working method and skills
Maintainer instructions:
Use subagents where possible to work on the checkboxes in the cleanup issue
Also update the issue to utilize ponytail and rust-skills.
Remember to refer to a full cleanup prompt at https://gist.githubusercontent.com/auser/d62dde280d75372dff19d8be7c549d2d/raw/4b5bea12e01e4a579a44fb3c642c0f88899bdc2d/cleanup.md
- Read and apply the complete, version-pinned MVM cleanup runbook. This issue's checklist is a tracking surface, not a replacement for the full charter, coverage matrix, staged prompts, review requirements, or completion report.
- Use
ponytailfor cleanup, implementation, refactoring, and review. Prefer deletion, existing helpers, direct control flow, and the smallest coherent change. Remove unnecessary concepts and maintained surface—not meaningful security boundaries or behavior tests. - Use
rust-skillsfor all Rust implementation, refactoring, and review. Apply the relevant ownership, error-handling, async/concurrency, unsafe/FFI, API, testing, and dependency guidance; do not merely mention the skill in a report. This is the maintainer's explicit skill requirement alongside the runbook's Rust-review guidance. - Delegate independent checkbox work to subagents where useful. Give each agent a precise worktree, file/surface ownership, read/write scope, execution boundaries, validation requirements, and missing-tool fallback. Do not duplicate active assignments or run competing expensive builds against shared mutable state. Use fresh review context where independence matters.
- Follow the runbook's staged workflow; do not load every stage into one undifferentiated prompt. Current repository instructions and subsequent explicit maintainer decisions take precedence over generic skill/runbook suggestions.
- Each agent reports concrete changes/findings, exact evidence, commands and results, limitations, and remaining acceptance gaps. Integrate and verify changes before checking completion; an agent report or accepted assignment alone is not a completed checkbox.
Source charter and recovered intent
Use the complete version-pinned cleanup runbook, subject to current repository instructions and subsequent explicit maintainer decisions.
This issue consolidates that runbook, surviving coordination context, and the maintainer's subsequent clarifications. It is not a verbatim recovery of the original long prompt.
Our project is based on security. The more dependencies and supply chain deps that we depend on introduce yet another layer of possible attack surfaces and vulnerabilities. I really want to reduce the number of dependencies to deal with supply chain attacks and reduce the amount of possible security vulnerabilities.
In addition, the fewer dependencies we rely on the fewer external dependencies. I want you to explore the dependencies that we have and how we might be able to reduce them, especially for deps that we're only using a tiny amount of their API surface area.
When I build for release I see we build 1017 dependencies and then the build seems to rebuild? Something is wrong. This binary has to be small and fast. It's required for good UX/DX.
Subsequent explicit clarification:
We have to clean up and refactor this project... it's a deep prompt with a strong refactor and cleanup. Reduce dependencies where possible.
Remove the AI slop (it's awful). You have authority to remove crates where they are extraneous
Subsequent maintainer instructions — execution contract
These instructions supplement the recovered charter and govern this cleanup. Keep this section and the owning acceptance criteria current as the maintainer clarifies scope; use issue comments for dated evidence and implementation updates rather than a separate repository progress ledger.
Rust size and completion
We definitely do not want more than 1M lines of rust code at all
I want you to get started on that work
Your goal is address and check all those checkboxes in the umbrella issue
- Reduce the Rust footprint below 1,000,000 lines using a declared, reproducible measurement. Report physical lines and executable SLOC distinctly and break out first-party, inline/dedicated tests, generated, and vendored code. Do not meet the target by silently excluding categories, compressing formatting, moving code to another repository, or weakening tests/security.
- Complete every acceptance item in this umbrella with evidence. A proposal, audit, passing text scan, or one completed cleanup slice is not overall completion; blocked or unverified work remains unchecked.
MvmClient and mvm-studio: library functions, never CLI backfill
We need to make sure that the MvmClient exposes all functionality and never drops into a
Command::new()or the same to execute AND that it supports themvm-studio(../mvm-studio) project's goals through the library functions, NOT throughCommand
Track implementation with #4415 and the native-lifecycle boundary in #4414. Missing library functionality must be implemented in the library, not replaced by a hidden CLI callback, runtime re-export, or a differently named spawn wrapper. CLI, MCP, Studio, hostlib, and language SDKs must consume the same authoritative services.
- Build the Studio→MVM capability matrix: 13 requirement rows with pinned sources/tests, distinguishing verified hermetic behavior, unwitnessed runtime parity, unsupported/planned capabilities and missing integration. Current-pin Studio/live parity remains unchecked.
- Expose the required operations through typed MvmClient/client-library APIs and appropriate hostlib/SDK bindings, including failure, cancellation, cleanup, policy, and audit semantics. Do not claim parity merely because a method exists or returns unsupported.
- Remove Studio's subprocess-backed operation backfill, CLI discovery, and related legacy adapters after library replacements are functional. Studio's local operations must work without finding or executing
mvmctl. Studio #30 records the local-first retirement landed in PR #31 (94a470d36f): gateway/sidecar/CLI-discovery and shell-plugin surfaces removed; existing local machine operations useMvmClient. Native GUI/runtime, missing typed operations, transitive spawn audit and SDK parity remain separate unchecked rows below. - Audit the full reachable production call graph, not just literal
Command::newsites:std::process,tokio::process, aliases, helper factories, shell/exec/spawn APIs, native process calls, deferred callbacks, and transitive runtime paths. Separate test-only processes from production behavior and record every real spawn's purpose and authority. - Resolve the exact backend/process boundary against PS-21: Complete typed VMM lifecycle without shell or utility processes #4414 and the maintainer's stricter instruction. Existing narrowly approved VMM/role-process decisions do not authorize CLI/utility fallbacks. Do not silently preserve an exception that conflicts with this instruction, or collapse signer/parser isolation merely to eliminate a process.
- Extend regression enforcement to the actual client and consumer surfaces. Combine structural checks with real public-API success/failure tests proving supported operations work with CLI execution unavailable and forbidden spawn paths fail the tests.
- Validate Studio integration against the resulting library contract, plus matching hostlib/SDK parity. No Studio integration pass is claimed until its build/tests and relevant runtime witnesses actually run.
Updated evidence after attaching Studio: the earlier sibling checkout was stale. The attached Studio worktree has already retired StudioBackend, SubprocessBackend, and CLI discovery; src-tauri/src/machines.rs::build_client calls connect(Target::Local) directly. Accepted Studio ADR 0005 requires library-only machine operations. Do not duplicate that completed removal. Stale source comments and historical sprint entries still describe the retired backfill and need correction. The remaining explicit Studio process launch found in this bounded scan is the gateway sidecar; ADR 0005 distinguishes it from local machine verbs, but the full transitive process audit remains open.
MVM's current xtask check-no-cli-shellout scans SDK and hostlib sources but explicitly excludes mvm-client; mvm-client::secret::source::SourceResolver can invoke password-manager CLIs through helper_command_with(...).spawn(). A clean direct Command::new grep therefore would not establish the requested guarantee. All four ecosystem worktrees are now attached. No library/Studio parity or live-runtime acceptance is claimed from this source inspection.
Host secret storage and future AWS KMS
The maintainer requires a trait-based host secret-storage boundary: local host storage now, with a future AWS KMS-backed key-management/encryption option. AWS KMS is explicitly confirmed, not Secrets Manager or a guest secret service. This is part of the highest-priority joint #3731/#4839 work, coordinated with PS-21/#3730/#4415 and existing secret ownership.
-
Inspect existing SecretStore, SecretService, SecretResolver and KeyProvider boundaries. Reuse them; no duplicate trait/AWS stub. Confined-resolver/provider custody and KMS envelope acceptance remain open. Preserved detailed receipt.
-
Establish the smallest typed, library-owned local-host implementation and provider seam needed by CLI/Studio. Retain authorization, audit, error/refusal semantics, secret-safe types/logging, and host-only material custody. No hidden provider CLI subprocess, guest plaintext, or parallel source of authority.
-
Define the future KMS integration contract explicitly: what is encrypted/stored where, key identity and access policy, ciphertext/context binding, failure/unavailable behavior, and provider configuration. Do not add an AWS SDK, dummy remote implementation, or Secrets Manager dependency merely to claim extensibility. Remote implementation follows reviewed scope and credential/identity decisions.
-
Add meaningful local success/failure and custody tests plus provider-contract tests where justified. Record dependency impact and client/consumer coverage; do not claim live AWS operation or completed KMS integration from an interface alone.
-
Land direct-keyring false-success/index fix #4858:
11517ede496ba2d262861cc0e9aa3bc614fe8545. Canonical receipt: reviewed file hash retained; PR CI and protected merge-group40965475636 passed. Auto fallback unchanged; native keychain, KMS and broader custody acceptance remain open.
PS-21: shared work with product umbrella #3731
Maintainer direction: work both #3731 and #4839. This cleanup umbrella supplements—not replaces—the agent-sandbox product umbrella. Its remaining PS-21: Keep the CLI thin over mvm-client is owned by #3730, with implementation/closure gates in #4412, #4413, #4414, #4415, #4416 and paired image-contract work in tinylabscom/mvm-images#67. Existing PS-21 acceptance remains authoritative; do not open a duplicate implementation track.
- Complete PS-21 (PS-21: Keep the CLI thin over mvm-client #3730) and its child acceptance in coordination with Agent-sandbox product surface: close every gap (tracking) #3731: zero inventoried production CLI reaches, shared client operations, hostlib/SDK parity, security regressions, required validation, and merge-queue landing. The older 65-file inventory is superseded by later merged extractions; a local
check-thin-cliat public base48c2023957reports 23 allowlisted files, not zero or completed parity. Final unit inventory remains to be measured. - Resolve macOS HVF: entrypoint boots (machine run --entrypoint, pack build health boot) are refused until native input custody lands #4655's native HVF verified input custody and recovery lifetime under that same client/native-boundary ownership. Snapshot observation, completion/audit observations, and foreground mock execution are separate evidence and do not close this requirement.
- Reconcile verified child evidence into both umbrellas before checking PS-21. Uncommitted source, partial handoffs and focused tests are progress, not product closure; keep the parent row unchecked until every original closure gate is satisfied.
CLI recovery lifetime — partner handoff and #4655
The maintainer supplied the partner handoff and asked that it inform CLI work. Live #4655 confirms the blocker: native HVF entrypoint/pack health boots refuse until verified input custody and a held recovery lifetime exist. This is separate from snapshot-exit observability (#4096), from #84 completion/audit observations, and from foreground mock-client diagnostics.
- Resolve macOS HVF: entrypoint boots (machine run --entrypoint, pack build health boot) are refused until native input custody lands #4655 ownership: PS-21 coordinator explicitly released the historical reservation and confirmed no active implementation/PR. One dedicated owner, “Own HVF entrypoint custody recovery,” now owns the bounded current-source lifecycle/design and subsequent implementation. Reservation and boundaries. Recovery design, implementation and live acceptance remain unchecked.
- Specify and verify the approved recovery receiver/lifetime contract. The maintainer approved a minimal unprivileged per-VM birth-in-owner role conditional on safety: original handles/resources, authenticated recovery, signing-key isolation, bounded failure/cleanup and independent review. Approval resolves the role-direction decision; exact implementation, authority identity and post-CLI-exit behavior still require verification.
- Expose that lifecycle through the typed client/library boundary used by CLI and Studio. A process host, if required, needs an approved identity/authority/lifetime contract; no generic command replay, hidden CLI operation backfill, duplicate supervisor, or reuse of another security role without approval.
- Add failure-path evidence for CLI exit/disconnect during uncertain execution, failed recovery handoff, repeated recovery, stale generation/identity, and eventual authenticated completion/cleanup. Refuse unsupported transitions rather than reporting success or losing the owner.
- Only after the product custody/lifetime implementation lands, remove both Firecracker-only gates in macOS HVF: entrypoint boots (machine run --entrypoint, pack build health boot) are refused until native input custody lands #4655 and record the actual macOS HVF entrypoint exit-code and pack-health/signing scenarios. Keep the current refusal and counted gates until then. A mock, compile check, snapshot observer, or Firecracker pass cannot close this item.
This handoff guides implementation; it does not itself choose a new supervisor architecture, assign the exclusive writer, authorize live execution, or supply a passing witness.
No legacy support requirement
For any future prompts in this session, make sure the umbrella issue reflects that and contains the content and intent.
Also we don't need to support any legacy paths
- Retire obsolete compatibility paths, shims, old protocol modes, fallback execution paths, stale flags/configuration, and their duplicate implementations end to end. Do not retain them “just in case” or solely for backward compatibility. Done when (maintainer-approved wording, comment): every live "legacy" item in non-test
crates/*/srchas a row inspecs/contracts/cleanup-inventory.mdstating whether it is a security or format boundary or compatibility, and a retirement condition; expired items are deleted. Done via refactor(cli): retire legacy launch-plan inputs #5104 (legacy launch-plan inputs), chore(cli): remove undocumented hidden subcommand aliases (#4839) #4966 and chore: drop the vcpus manifest alias and the lima_status node alias (#4839) #5210 (hidden aliases and thevcpusalias), chore(image-set): remove the unread [legacy] pin from images.lock (#4839) #4972 (the unread[legacy]pin inimages.lock), fix(runtime): refuse an unrecognised --hypervisor and drop the krun/hypervisor aliases (#4839) #4973, chore: drop the MVM_BACKEND environment alias (#4839) #5209 (MVM_BACKEND), chore: remove dead configuration surface and gate documented env vars (#4839) #5169 (dead configuration), refactor(hostd): drop the broker::control forwarding module (#4839) #5040 and refactor(runtime): drop the driver::spec forwarding module (#4839) #5046 (forwarding modules), docs(contracts): claims witness-adequacy matrix and legacy compatibility classification (#4839) #5432 (classification) and refactor(packs): remove the expired legacy pack-signing identity (#4839) #5459 (the expired legacy pack-signing identity and cutoff,registry_pack/historical.rs, the provenance field and refusal variant, deleted;git grep LEGACY_PACK_SIGNINGis empty; a negative test shows a pack signed only under the former identity is refused with a real Sigstore bundle, with a positive control under an operator policy; all 11 publishedmvm-packsmanifests were verified to carry the current identity before deleting).specs/contracts/cleanup-inventory.mdsection 3.6 now gives every live legacy item a row with its class and a retirement condition. Not retired, by recorded condition and not by neglect: builder boot ABI 0 and 1, theRetiredGuestRuntimeimage-set role, base-image contract 0 andbuilder_boot_abi_or_legacy(blocked on mvm-images publication, ADR-054),legacy_portable_runtime.rs(a live signed-bundle verifier, retirement belongs to feat: one signed file, runnable anywhere — unify .mvm into .mvmpkg and boot it withrun <artifact>#3928) andlegacy_bundle_runtime.rs(no in-tree producer found; blocked on a measurement of persisted metadata and an ABI decision). Narrowing from refactor(packs): remove the expired legacy pack-signing identity (#4839) #5459: a host with a clock before 2026-11-06 no longer accepts the nine historical manifests under the built-in policy. - Identify each intentionally removed interface and update active docs, callers, fixtures, tests, packaging, and CI together. Explicitly distinguish legacy compatibility from an actively required platform or security boundary; naming something “legacy” is not evidence that its security responsibility is obsolete. Done with the box above: section 3.6 of
specs/contracts/cleanup-inventory.mdhas an "intentionally removed" row for each retired interface (aliases, the[legacy]pin, hidden subcommand aliases, the legacy pack signing identity and cutoff) naming what changed and why, and distinguishes legacy compatibility from boundaries that stay (the security boundarieslegacy_portable_runtime.rsandlegacy_bundle_runtime.rs, and the current non-legacy boundariesruntime_evidence.rsandenv verify-runtime). Each landed removal updated its callers, tests, fixtures and docs in the same PR (for example refactor(packs): remove the expired legacy pack-signing identity (#4839) #5459 updatedpack-authoring.md,policy-and-profiles.md,cli-commands.md, an ADR-055 annotation and the fixture notes). - Update the issue with the content and intent of subsequent maintainer prompts, including changed scope, priorities, constraints, and authorizations; keep affected acceptance criteria synchronized. Do not publish secrets or sensitive payloads if a future prompt contains them. Done, as of 2026-11-14: the maintainer decisions, authorizations and process rules given during the cleanup campaign are recorded in this comment and in the edited boxes (the approved done-when wording, option (a) for the sprawling-crates box, bare-string env retired, the expired legacy identity deleted, the rc.3 authorization, the latency stance, and the Firecracker deferral below), with no secrets or sensitive payloads recorded. This is a point-in-time record: each later maintainer decision is recorded as a new comment on this issue and the affected acceptance criteria are edited in step with it.
The no-legacy instruction removes backward compatibility as a reason to retain an obsolete path; it does not waive security requirements or authorize silently losing required Studio/library functionality. Where it supersedes earlier compatibility-preservation wording in this issue or the runbook, this instruction takes precedence.
Protected core goals — cleanup must preserve these
Latest maintainer instruction:
Make sure we retain all "fire-and-forget" tracing log messages as that's a core goal
Other core goals:
- Best UX/DX
- fast startup time -- <300ms cold start, <50ms warm start
- multiple backends (HVF, Firecracker, libkrun, wasm, WHP (future))
- tracing everywhere
- vsock everywhere (microvm to/from host)
- no secrets leak into a microvm
These are protected product requirements, not candidates for deletion to meet line/dependency/CI budgets:
- Retain all fire-and-forget tracing messages and observability coverage. Removing AI-like narration means removing unnecessary prose/structure, not useful tracing. Preserve non-blocking tracing behavior and the existing context/correlation, redaction, delivery/loss reporting, and shutdown/flush contracts. Do not invent an unbounded queue or discard events silently to make an operation appear fast.
- Tracing everywhere: preserve instrumentation of important lifecycle, policy, backend, failure, cleanup, and cross-boundary paths. Payloads must be safe metadata, never secret values. A trace/export change must be reviewed as an observability change, not incidental cleanup.
- Security audit is not disposable diagnostics. Keep signed-audit integrity, required persistence/flush, and fail-closed behavior. The fire-and-forget requirement does not waive mandatory audit evidence or authorize detaching security work that must finish before success.
- Best UX/DX: simplify installation, build, library use, diagnostics, and recovery as well as code. No hidden CLI fallback, repeated compilation, misleading success, or new configuration ceremony. Preserve Studio/library parity and typed actionable failures.
- Startup targets: cold start <300 ms; warm start <50 ms. Use the original invocation-to-ready measurement and declare the exact candidate, backend, host, workload, cache state, and cold/warm definition. Do not exclude authentication, admission, required audit/durability, or other user-visible work to pass. Preserve any stricter existing percentile/per-sample budgets; do not reinterpret these targets as averages or relax the thresholds.
- Backend coverage: preserve HVF, Firecracker, libkrun, and WASM as supported product paths. WHP remains a future goal, not a current-support claim or a reason to add speculative scaffolding. Retiring optional native Cranelift development compilation is not permission to remove WASM execution. “No legacy paths” does not mean deleting a required backend or its real isolation/platform boundary.
- MicroVM ↔ host transport: preserve the vsock architecture, authentication, and no-guest-NIC invariant. Do not introduce a guest TCP/NIC/TAP/bridge alternative or fallback to simplify a backend. Keep backend-specific implementations behind the canonical transport contract.
- No secret values in a microVM: keep secret material and substitution/custody on the host under the approved subsystem. Guest placeholders/references must not become plaintext credentials in environment variables, argv, files, logs, tracing, snapshots, or fixtures. Preserve boundary validation and negative leakage tests; a working demo must satisfy this constraint too.
Acceptance evidence:
- Review every tracing/telemetry-affecting deletion or refactor for retained fire-and-forget behavior, messages, correlation, redaction, bounded resource use, and failure/shutdown semantics; preserve mandatory signed-audit guarantees separately. Done via docs(contracts): telemetry and audit review of every merged #4839 PR (#4839) #5408: a per-PR table in
specs/contracts/diagnostic-data-policy.mdcovers all 86 merged PRs whose title carries#4839(the 33 reviewed earlier in docs(contracts): telemetry-change review and #4840 status (#4839) #4977 re-checked from their diffs, plus every later one), with the telemetry or audit effect, a verdict and a witness test whose existence was checked. No PR removed a bounded queue, redaction step, correlation id, shutdown path or signed-audit emission. Recorded exceptions: build: drop tracing-log forwarding; rank dependency reduction candidates (#4839) #4935 (third-partylog::records no longer forwarded; accepted when merged), test: fix the root-euid sudo-arm test and drop tests that restate code (#4839) #4964 (the literal CLIDEFAULT_FILTERvalue is no longer pinned by a test), and fix(hvf): boot a canonical-kernel bundle that carries no caller registration (#4839) #4968, which relaxes an HVF boot admission check and is tracked separately in Review the HVF kernel-descriptor-only boot path narrowed by #4968 against the admission claims #5411. Not covered: PRs that mention Cleanup umbrella: simplify MVM, reduce dependencies, and improve build performance #4839 only in the body. - Record reproducible cold/warm startup evidence against the <300-ms / <50-ms targets on the applicable supported backends without weakening existing budget definitions or security gates.
- Validate affected supported backend/feature combinations, including WASM; explicitly mark unavailable platform witnesses and WHP as future rather than claiming coverage.
- Prove affected microVM/host paths retain authenticated vsock and host-held secret custody, including meaningful negative tests for forbidden transport and secret leakage.
Security and product boundaries
- Preserve the design governed by Agent-sandbox product surface: close every gap (tracking) #3731, security guarantees, traceability, and mailbox designs. Prioritize demo-tagged work while respecting active ownership.
- Preserve no-guest-NIC architecture, authenticated vsock, host-only secret-value proxying, signed ExecutionPlan enforcement, audit integrity, and meaningful backend/telemetry boundaries.
- Preserve fail-closed behavior, isolation, authorization, network restrictions, secret handling, update verification, supported platforms, and released contracts.
- Prefer deletion,
std, existing dependencies/helpers, and direct Rust. Do not replace mature cryptography, signature verification, hostile-input parsers, serialization, protocols, or memory-safety-critical code merely to improve a count. Follow required owner decisions, threat analysis, and adversarial testing. - Keep real platform, trust, privilege, protocol, lifecycle, unsafe-code, and public-API boundaries. Remove ceremony rather than security structure.
- Respect repository ownership: MVM runtime/interfaces;
mvm-packsapplication/agent packs and pack policy;mvm-imagesLinux image layer under the current migration contract;mvm-assurancescanning/assurance;mvm-opsoutreach/opportunity workflows.
Additional canonical goals found during documentation audit
The maintainer requested coverage against README, ADRs, issues, and PRs—not just the short core-goals list. The following explicit obligations supplement the existing checklist. Audits distinguish documented intent, landed source, and witnessed behavior; this section does not declare them implemented or validated.
Security posture and installation
- Map the current shipped claims in MVM README/ADR-001 to retained enforcement and meaningful witnesses: host-filesystem isolation except explicit shares, privilege separation, dm-verity, sealed-production DevOnly/shell/PTY exclusions, bounded/fuzzed hostile-input protocols, signed audited execution plans, content-addressed artifacts, default-deny egress, bound host services, host-held secret substitution, provenance, CVE/SBOM evidence, and release/update signature verification. Keep preview claims visibly separate from shipped guarantees. Done via docs(contracts): claims witness-adequacy matrix and legacy compatibility classification (#4839) #5432 (
390c5f6c):specs/contracts/claims-witness-adequacy.mdmaps ADR-001 claims 1-15, 19, 20 and 21 to enforcement locations and each cited witness, classifying every witness as a behavioural outcome test, a source-text scan, a fixture or mock, or a CI lane, with CI freshness and the documented limits carried through unchanged (no claim was upgraded). Preview rows 16-18 are in a separate section and not promoted. All 154 distinctfn:witnesses resolve andcheck-claim-catalogpasses (21 claims, 174 witnesses). Recorded, not fixed: text-scan and weak witnesses (claim 20fresh_install_*, claim 2 constant assertions, claim 19 argument parsing, claim 10ci:single-network-pathanchored to a step title) have replacement proposals in section 5 that are not implemented; claim 3 still has no KVM tampered-boot lane. - Preserve relocatable installation and the documented no-host-Nix/no-Docker prerequisites, with source-build prerequisites distinguished from installed-release UX. Preserve prepared/cache-only launch semantics rather than hiding download/build work inside measured startup.
Image and application trust are distinct contracts
Canonical sources: mvm-images README, SECURITY.md and REVOCATIONS.md; mvm-packs README/README.packs.md; MVM ADR-054/055 and the linked migration issues.
- Preserve the Linux-only image producer boundary and generic default/rootless profiles across both architectures. Keep base-contract bytes, fixed identities/boot ABI, full producer-member coverage, and direct VMM/rootless witnesses; do not add application-specific packages, MVM runtime binaries, or MVM-source inputs to mvm-images.
- Verify the staged image migration and release order through mvm-images#49/feat(mvm-supervisor): plan 37 E1 — false-positive circuit breakers #64 and mvm#4100/Release lane is red: e2e control-plane failures and archive-smoke bootstrap gap block every release witness #4764: a compatible same-version signed CLI/runtime and installed-product boot evidence precede Linux-only publication and consumer image-lock advancement. Source definitions without an MVM input are not completed-release evidence.
- Keep image-set identity (root/member table, profile, architecture, base contract, release provenance) distinct from application identity (exact descriptor bytes, archive SHA/size, guest architecture, independently enrolled bundle key, runtime compatibility). Consolidate helpers only where semantics truly match; never create competing identity authorities. — Evidence (2026-11-06): test(packs): guard pack catalog, descriptor and revocation trust boundaries #5142 (merged) adds
an_image_set_signer_cannot_sign_an_application_descriptoranda_flipped_archive_byte_of_the_same_size_is_a_digest_refusalbeside the existing identity refusals; each guarded behaviour was broken locally and its test failed (digest check disabled, publisher trust widened with image-set identities). - Preserve one generated pack catalog with exact-byte manifest hashes, independent manifest/signature verification, publisher policy, and separate bundle trust/admission. Catalog discovery or a descriptor naming a key must not enroll or authorize that key. — Evidence (2026-11-06): test(packs): guard pack catalog, descriptor and revocation trust boundaries #5142 (merged) adds
catalog_hash_binds_the_exact_manifest_bytesandcatalog_discovery_and_a_matching_hash_enroll_and_admit_nothing(fetching the index, searching, and a matching-hash pull that fails on signature leave MVM_HOME with no trust key, policy or pin); mutation-checked (catalog digest check disabled together with a state write). - Preserve separate image and pack revocation authority, signer identities, freshness bounds, and consumer enforcement. Immutable historical signed artifacts remain verification evidence; no-legacy execution cleanup does not authorize rewriting or deleting them. — Evidence (2026-11-09): merged #5142 rejects cross-authority feed signatures and image-set/CLI signer substitution; #5216 confines built-in bundle-author keys to authenticated installed applications with revocation checked before lookup; #5217 refreshes the separately signed official registry-pack feed, checks stale/rollback and additive operator feeds at install and launch/admission. On base
3b827982ff, focused image/registry authority, rollback/wrong-signer, official/operator feed, revoked-bundle and distinct-store tests passed (including the published historical image-set fixture underhost-state,manifest-verify);xtask check-bundle-trust-sitespassed. These PR diffs do not rewrite historical signed fixture artifacts. This is preservation/enforcement evidence only, not a current published/installed product witness or completion of the separate image/pack release boxes. - Preserve atomic immutable publication across the required architectures/profiles, reproducibility and SLSA/SBOM subject coverage. Component checks, observer reports, unit tests, version probes, and local reproduction are bounded evidence, not substitutes for final published/installed artifact admission and boot.
Studio and MVMD scope and contract obligations
- Reconcile Studio#19's explicit single-host/local product scope with accepted legacy fleet ADRs and Studio#30. Record the retirement decision before deleting fleet/auth/observability behavior; do not recreate the absent MVMD client/gateway or claim that sibling symlinks solve a removed API.
- Coordinate—not duplicate—Studio PR#29's local typed MvmClient telemetry adapter and its proposed ADR0008 acceptance gate. Preserve bounded cursor paging, cancellation, selected-machine attribution, stale-cursor refusal/reset, unavailable/not-provisioned states, and explicit fixture mode. Successful PR CI is not a merged/live product witness.
- Preserve MVMD V1's independent contract boundaries: no runtime/framework/transport types in authority contracts; hard placement constraints never become score penalties; billing stays integer microunits; connection establishment is not authorization; remote mutations remain audience/generation/time/nonce-bound with a destination-host refusal path; mailbox plaintext remains unavailable to controller/router/store APIs; queues/frames/retries stay bounded.
Conflicts to resolve rather than silently reinterpret: the README's current 300-ms warm refusal ceiling is not evidence of meeting the maintainer's <50-ms warm goal; Studio's old gateway startup model is not its desired local-library architecture; historical migration release numbers are not the current candidate; future WHP and draft roadmap capabilities are not current support claims.
Accepted contracts and roadmap dispositions from the 58-ADR review
The MVM audit inspected all 58 ADR status headings (001–057 plus 110), README, the live issue/PR inventory, and selected owning contracts/issues. Studio/MVMD and images/packs audits supplied their canonical-source matrices separately. This establishes a goal/disposition inventory—not exhaustive implementation verification or a live witness. Historical/proposed architecture must not silently displace shipped behavior.
- Tracing/stream completeness: reconcile closed Core tracing: every microVM through the host with nonblocking encrypted telemetry #3419 and merged PR#4379 with still-open W3 — Guest capture with bounded non-waiting emission #3422–3426 and Define diagnostic confidentiality and retention policies #4184. Preserve producer coverage, bounded queues/export storage, detached lifetime, cancellation, stopped-reader/full-queue behavior, explicit loss/gaps, and synthetic-secret non-disclosure. Fire-and-forget does not imply lossless delivery during indefinite outages; mandatory signed audit remains separate.
- Current security witnesses: retain no-SSH guest posture and check scheduled witness freshness (Claim-bearing CI lane has stopped running #4788), not just the existence of named tests. Keep privilege/custody follow-ups OCI workload process runs with NoNewPrivs=0 and a full capability bounding set #2101/Firecracker canonical custody: close the review follow-ups #4612/Bind the claim-time layout check to the descriptors Firecracker holds open #4819 visible. Do not promote proposed ADR026's stronger runtime-never-root claim from the narrower shipped anti-escalation guarantee.
- Protocol and readiness (ADR019): preserve explicit version/capability negotiation before dispatch, usable readiness distinct from process-running or disabled state, bounded payload/control capacity, typed backpressure, cancellation, and recovery. Coordinate lifecycle/client owners PS-21: Complete typed VMM lifecycle without shell or utility processes #4414–4416 and input-journal follow-up machine exec: a nonzero guest exit (or a killed client) leaves the input journal unfinished and refuses every later exec #4809. Audit 2026-11-15: readiness, bounded capacity and the closed capability enum are witnessed; the hello-before-dispatch clause is not true as shipped (the authenticated handshake is the prelude, the hello is optional, and the host checks its own capability list, not the guest's). Enforcement is owned by Enforce guest-advertised protocol version and capability negotiation before dispatch (ADR-019) #5582; backpressure, cancellation and recovery limits stay with PS-21: Complete typed VMM lifecycle without shell or utility processes #4414, PS-21: Validate and land complete client-boundary parity #4416 and machine exec: a nonzero guest exit (or a killed client) leaves the input journal unfinished and refuses every later exec #4809. Not ticked until Enforce guest-advertised protocol version and capability negotiation before dispatch (ADR-019) #5582 lands.
- Mailbox operational contract (ADR057): preserve synchronous admission/refusal of cast, no silent loss after acceptance under the declared durability class, timeout ambiguity, deduplication, fencing, host-blindness tiers, and separate topology/acceptance performance measurements. Coordinate Program: signed pass-by-value mailboxes for workload, host, and user messaging #4622, Implement canonical signed and encrypted mailbox envelopes #4628–4631, Implement Iroh mailbox transport with explicit relay policy #4633 and Benchmark mailbox latency, overload, and typed-service isolation #4635; accepted architecture is not automatic security-claim promotion.
- Evidence completeness (ADR110): preserve exactly-once accounting of in-scope entries, the distinction between attested and derivable completeness, and leaf-inclusion proofs—not merely a valid signature on a receipt. Coordinate audit/receipt parity in PS-21: Finish client execution ownership and hostlib SDK parity #4415/PS-21: Validate and land complete client-boundary parity #4416. Done via test(hostd): witness archive completeness accounting (#4839) #5434 (
mvm-hostd): production already enforced the accounting incheck_completeness(crates/mvm-hostd/src/audit/receipt_archive_verify.rs: rejects a leaf listed twice, rejects a derivable archive whose leaves are not exactly0..tree_size, keeps attested a distinct arm), but nothing exercised it throughverify_archive. Three outcome tests now do, each with the manifest mutated and re-signed so integrity and inclusion stay green:a_leaf_listed_twice_fails_completeness_though_signed_and_provable,a_dropped_leaf_fails_a_derivable_archive_completeness,a_signed_attested_archive_missing_a_leaf_is_reported_attested_never_derived. Leaf-inclusion proofs were already witnessed (every_leaf_proof_verifies_and_yields_the_root_for_sizes_1_to_8,a_proof_moved_to_another_leaf_fails_inclusion_with_integrity_intact). The ADR-110 limit stands and the third test pins it: under plan-scope an omitted leaf is undetectable. Parity coordination with PS-21: Validate and land complete client-boundary parity #4416 (PS-21: Finish client execution ownership and hostlib SDK parity #4415 is closed) is recorded in this comment; PS-21: Validate and land complete client-boundary parity #4416 itself was not changed. - Owner-mediated application access (ADR056/Implement owner-mediated application access (ADR-056 accepted; typed channel unimplemented) #4460): reconcile the accepted-but-unimplemented typed channel with stale deferred/Proposed tracker wording. Preserve sealed-console refusal, bounded authenticated application authority, and explicit development-only demo labeling. Do not start a duplicate post-demo protocol implementation or count a dev terminal as sealed-production support.
- GPU remoting (ADR053/EPIC: GPU compute in microVMs by API remoting over vsock #3560): retain compute-over-vsock rather than passthrough and honest unsupported-capability refusal. Merged PR#4727 supplies a bounded harness, not a native-hardware/fork/device-memory witness; retain those evidence gaps and conditional MLX scope. Evidence/limits.
- Capture/authoring/function entrypoints (ADR050/005): retain read-only capture boundaries and actual compile/launch contracts; successful IR generation is not proof of reproduction. Coordinate refactor(compiler): extract mvm-compiler and make mvm-sdk a thin facade #4230 and follow-ups to closed feat(run): consume launch plans with explicit manifest image selection #5 / merged PR#4838 (merge commit
cd150e044fd1ea54e0dd1ca0a883e80b9aaf8c24). The deliberately retained singular-entrypoint/string-environment compatibility needs a coherent follow-up under the newer no-legacy direction; that removal was not part of feat(run): consume launch plans with explicit manifest image selection #5's landed scope. Done, verified 2026-11-13 onmain: read-only capture is witnessed bycapture_honours_ignore_files_and_leaves_the_project_unchanged(crates/mvm-capture/tests/capture_fixture.rs, which also covers ignore-file handling); the retained singular-entrypoint compatibility was retired in refactor(sdk): retire the singular-entrypoint compatibility path #5403 (test_app_rejects_removed_singular_entrypoint_kwarg) and the string-environment compatibility in feat(sdk): retire bare-string env values in the SDKs and decorator parser (#4839) #5463 (bare-string env values refused in both SDKs and in the Rust-side decorator parsers, witha_bare_string_env_value_is_refused_without_echoing_itincrates/mvm-compiler/src/decorator/python.rsandtypescript.rs; the error namesmvm.literal(...)andmvm.secret(...); the Workload IR already rejected a bare string, so no schema change); the compile facade was removed in refactor: remove remaining forwarding files and glob re-exports (#4839) #5448; coordination with refactor(compiler): extract mvm-compiler and make mvm-sdk a thin facade #4230 is recorded in this comment, and refactor(compiler): extract mvm-compiler and make mvm-sdk a thin facade #4230 remains open with its own acceptance. Limit: the clause "successful IR generation is not proof of reproduction" is a caution and is not claimed; byte-reproducibility is witnessed only at the archive level (archive_is_byte_reproducible_across_runs). - Browser/WASM capability tiers (ADR024/049): retain direct-WASI separately from accepted-preview WebLinux/local-builder goals (Deliver Claude through portable pack authoring and protected execution #4222/feat(packs): lower locked package environments through the existing SDK compiler #4224/feat(packs): ship Python and TypeScript build APIs with an archive-only vertical witness #4226/feat(packs): add Rust and Kotlin JVM authoring parity #4228). WASM is not hardware isolation; preview/browser functionality is not a shipped native-tier claim. Record explicit deferral where appropriate instead of speculative implementation. Evidence/limits.
- Future Windows direction: reconcile ADR009 with feat: Windows host support via the libkrun backend (supersedes native-WHP sketch) #3553 and the owner's future-WHP goal. Current Linux/macOS tiers stay unchanged; planned Windows uses existing libkrun/WHP, not a new in-house WHP supervisor. Require qualified boot, private-state ACL, signing/integrity, doctor/refusal, and capability evidence before promotion. A documentation correction is in the cleanup worktree, not a support claim.
- Staged-state contradictions: reconcile README installation assertions with ADR028/Ship the libkrun backend tier in the single signed release archive (macOS 13-25) #4577; ADR054's old-boundary wording with ADR030 and source-landed migration versus publication/boot evidence; Zero-step first run: prepare on install and update, never ask the user to run bootstrap #4794's inline-preparation proposal with measured prepared-launch semantics; and proposed ADR046's intended future replacement with still-required ADR035 streams. Do not erase current behavior or evidence based on a proposed successor. Done 2026-11-15 (docs: reconcile staged-state contradictions (#4839) #5580): (1) README install vs ADR-028 and Ship the libkrun backend tier in the single signed release archive (macOS 13-25) #4577 was a real contradiction and is fixed: the README now says the macOS 13–25 release-install claim rests on the packaged libkrun tier and that a guest boot from the signed archive on that hardware has no recorded witness yet. (2) ADR-054 vs ADR-030 and source-landed vs publication was already consistent (ADR-030's Status carries the dated amendment; ADR-054's separates landed from unpublished
image-set/v0.2.4). (3) Zero-step first run: prepare on install and update, never ask the user to run bootstrap #4794's inline preparation vs measured semantics was already consistent (ADR-054 citeslaunch_refuses_an_unprepared_published_runtime_without_acquiring_it,required_overlay_cache_miss_names_explicit_bootstrap,a_download_mode_host_refuses_an_unprepared_sidecar; moving preparation into the launch layer is proposed and unshipped). (4) Proposed ADR-046 vs required ADR-035 was already consistent: both say ADR-035 stays the shipped design until ADR-046 is accepted and implemented. Remaining: the Zero-step first run: prepare on install and update, never ask the user to run bootstrap #4794 issue body still says install runs bootstrap by default, whileinstall.shdoes so only withMVM_BOOTSTRAP_ON_INSTALL=1; that body is that issue's to correct.
Disposition rule: preserve current required capabilities; explicitly identify preview/proposed/future work and its owner; retire obsolete execution paths only with the current owner decision and complete caller/test/docs cleanup. This goal inventory does not turn every future roadmap feature into a prerequisite to finishing a bounded cleanup slice.
Workstreams and acceptance checklist
Each category requires evidence-backed findings or an explicit, supported conclusion that no change is warranted. Check items only when the outcome is verified, not merely proposed or locally reviewed.
1. Establish baseline and ownership
- Verify the runbook's required release baseline and record its exact tag/commit, or obtain an explicit change to that gate before new implementation.
- Inventory current cleanup diffs, branches, PRs, and active owners; link them here without overwriting unrelated work. Inventory evidence (verified 2026-11-03 at
3d17ff9fdd; 2 open PRs, 75 unmerged cleanup branches reconciled to active lanes). - Record supported binaries, guest artifacts, targets/features, released contracts, and security invariants. Use the repository graph and verify its coverage. Done via docs(contracts): record fresh graph coverage and name workstream owners in the cleanup inventory (#4839) #5383 (
00cdc7f9): the inventory inspecs/contracts/cleanup-inventory.md§1 covers the packages, binaries, guest artifacts, features, released contracts and security invariants, and the graph was reindexed in full mode (2026-11-12,mainataaf74fcb11, 106,437 nodes) and its coverage re-checked: the binary entry points report fresh and themvm-hostd,mvm-agentdandmvm-builder-guestsrc/binscopes are no longer excluded. Remaining gaps are recorded there (two CLI command directories excluded by name, fuzz corpora and virtual environments, five partially parsed files). Coverage is a best-effort signal, not proof of completeness. - Split the inventory into atomic child workstreams with owners, dependencies, acceptance criteria, and validation commands. Done via docs(contracts): record fresh graph coverage and name workstream owners in the cleanup inventory (#4839) #5383 (
00cdc7f9): the workstream table (W1-W17) carries owner, boxes, dependencies, acceptance and validation commands. Owners are accountable roles (maintainer,runtime owner,producer and consumer owners,PS-21 coordinator), not named individuals, because the issue has no assignees and the repository has no CODEOWNERS; a named person replaces the role.
2. Remove AI slop, duplication, and unnecessary structure
- Remove obvious-code narration, process/history commentary, unsupported claims, stale scaffolding/placeholders, and redundant explanations; retain concise invariant and failure-mode reasoning. Done when (maintainer-approved wording, comment): comments carry no process or history references (
Phase N,Plan N,PS-NN,W<n>, "was previously"); the comment gates have empty baselines. Obvious-code narration is a review item, not counted. Done:check-no-history-in-comments(empty baseline,xtask/history-comments-baseline.txtis 0 bytes) andcheck-no-spec-refs-in-comments(extended in chore: gate process labels in comments; drop stale package counts (#4839) #5410 toPS-NN,W<n>andPhase <n>; 73 hit lines in 29 files rewritten; now also scanstests/, where it cleared olderPlan N,Sprint NandADR-Ncitations too) are both incheck-all(xtask/src/check_all.rs) and passed in the merge queue. Earlier slices: chore(xtask): gate history narration in code comments (#4839) #5164, docs(cli): drop plan and phase references from comments (#4839) #4950-docs(vmm): drop plan reference and history narration from comments (#4839) #4956. - Remove or consolidate duplicated helpers, validation, error mapping, and competing sources of authority without erasing meaningful behavioral differences. Done when (maintainer-approved wording, comment): no helper is defined in three or more crates without a recorded reason in
specs/contracts/cleanup-inventory.md. Done via refactor: delete duplicate sha256 and hex helpers onto existing ones (#4839) #4937, refactor: one shell-quote helper and one unix-seconds helper (#4839) #5205, refactor: one sha256_file and one deps read_manifest; trim stale comments (#4839) #5359, refactor: one sha256_hex authority, remaining duplicate helpers recorded (#4839) #5409, refactor: fold duplicated pid-liveness and file-hash helpers; record the rest (#4839) #5447 and docs(contracts): record why exchange is defined in three crates (#4839) #5450:specs/contracts/cleanup-inventory.mdsection 2.1a holds a reproducible scan (non-test, column-0 free functions only, so methods and trait-impl methods are excluded;main,new,defaultandrunskipped) and a recorded reason for every name it finds. The scan finds 93 names defined in three or more crates; re-run on currentmainafter docs(contracts): record why exchange is defined in three crates (#4839) #5450, none is unrecorded. True duplicates were folded: onesha256_hexauthority (mvm_contract::hash::sha256_hex),sha256_fileandhash_fileontomvm_core::crypto::image_verify::sha256_file, and thepid_aliveandread_pidwrappers ontomvm_vmm::host::process_liveness. The rest stay by recorded reason (different frame, error or symlink/mode contracts, or no dependency edge by design, for examplemvm-cli/build.rsandmvm-builder-guest). Limit: the scan is a documented command, not a gate, so a new three-crate name appears only when it is re-run; the one-time re-run after refactor: fold duplicated pid-liveness and file-hash helpers; record the rest (#4839) #5447 foundexchange(three unrelated mailbox and HVF exchanges, recorded in docs(contracts): record why exchange is defined in three crates (#4839) #5450).pid_alivealso remains inmvm-buildandmvm-runtimefor stated reasons. - Remove needless wrappers, forwarding modules, reexports, boilerplate traits/builders/macros, and speculative configuration; narrow unnecessary public APIs. Done when (maintainer-approved wording, comment): no pure-forwarding module and no glob
pub useoutside an allowlist limited to themvm-clientembedder surface and the root facade. Done via refactor: delete pure-forwarding modules, callers use the owning crate (#4839) #5433 and refactor: remove remaining forwarding files and glob re-exports (#4839) #5448: both deleted every pure-forwarding file outside the allowlist (12 files, then the last two:mvm-sdk/src/compile/mod.rswith its facade-only test andmvm-runtime/src/driver/traits.rs) and repointed every caller; every publicpub use …::*outside the allowlist was replaced with explicit names;xtask check-core-module-ownership(incheck-all) now rejects a globpub useor a use-only file anywhere exceptcrates/mvm-client/andsrc/lib.rs(FORWARDING_ALLOWED), with a negative test; the siblingmvmd,mvm-studio,mvm-assuranceandmvmforgecheckouts do not use the removed paths (some checkouts may be stale). A grep overcratesandsrconmainfinds nopub use …::*outside the allowlist. Limits: the allowlist files inmvm-client(artifacts.rs,audit/chain.rs,cache_admin.rs,firecracker_setup.rs,linux_shell.rs,session_store.rs,template.rs,ui.rs,vsock_transport.rs) remain by design as the embedder surface; onepub(crate) use …::*remains inmvm-cli/build_embed_cache.rs(not a public re-export);third_party/is not scanned. - Regroup
crates/mvm-hostd/src/supervisor/(65 flat files at the top level, 116 in total) into responsibility submodules, and splitcrates/mvm-hostd/src/plan_admission.rs(7,173 lines) andcrates/mvm-runtime/src/workload_runner/runner.rs(7,434 lines) by responsibility, with no behavior change, every moved test kept with the code it exercises, andcheck-rust-footprint,check-file-sizeand the full suites of both crates passing. Replaces the open-ended box "Simplify sprawling or fragmented crates/modules based on responsibility and dependency direction—not line-count quotas." by maintainer decision (option (a) in this comment); the one measured slice from the refactor(hostd): fold six single-consumer micro-modules into their owners #5113 follow-up is kept as its own item, and no size threshold is adopted. Done 2026-11-15: all three pieces merged as pure moves — supervisor regrouped into responsibility groups (refactor(hostd): regroup the supervisor modules by responsibility (#4839) #5531; top level 65 flat files to 6 plusmod.rs, 2,523 hostd tests before and after), the workload runner split (refactor(runtime): split the workload runner by responsibility (#4839) #5565; 1,387mvm-runtimetests before and after, rootrunner.rsnow 282 lines), andplan_admissionsplit (refactor(hostd): split plan_admission by responsibility (#4839) #5566; root now 389 lines, test names unchanged apart from the two tests refactor: stage local workload launches through host admission #5564 added,mvm-hostd2,532 passed).check-file-sizepasses with the grandfatheredplan_admissionentry removed.check-rust-footprintexits 0 only because the ratchet was suspended by maintainer decision (chore(xtask): suspend the Rust footprint ratchet while the #4839 cleanup lands (#4839) #5573) while the cleanup lands; restoring it is tracked in Restore the Rust footprint ratchet (suspended during the #4839 cleanup) #5567, so its pass here is not evidence of a footprint reduction. - Shrink
mvm-hostd(340 source files, ~164k lines): consolidate fragmented modules by responsibility, remove duplication and ceremony, and reduce the file/line footprint measurably; record before/after file and line counts. Done via refactor(hostd): fold six single-consumer micro-modules into their owners #5113 (mvm-hostd: 377 → 371 tracked files, 163,335 → 163,304.rslines; six single-consumer micro-modules folded into their owners, no behavior change; fmt, clippy, nextest 2560/2560, claim gates clean). The modest line count is deliberate: it removes module headers and re-export ceremony. The 68-file flatsupervisor/namespace and thenetwork_endpoint_proxysplit are left for follow-up. - Consolidate
scripts/(124 entries, 145 files): merge overlapping helpers, remove obsolete/unused scripts after checking just recipes, CI workflows, and docs consumers, and record before/after counts. Done via chore(scripts): remove the retired linux-gate VMM scripts #5036 (c6b3b701) and chore(scripts): prune dead files and wire orphan script tests into CI #5038 (3fa31d2d): before 124 entries/145 files, after on main 117/137; full consumer map (basename + Python module stem) across just recipes, 35 workflows, scripts, docs, specs, xtask; 8 passing-but-unwired helper test suites wired intoci.ymlinstead of deleted; examined merge candidates each proved distinct with live consumers — evidence.
3. Reduce dependencies and extraneous crates/features
- Map actual API usage and dependency closure for each affected shipped binary and guest artifact, including direct/transitive packages, duplicate versions, features, build scripts, procedural macros, and native code.
- Rank candidates by supply-chain exposure, trust boundary, API usage, maintenance/advisory/license evidence, transitive closure, build cost, and binary contribution.
- Remove unused direct dependencies and unnecessary default/optional features; unify compatible duplicate versions where behavior permits. All actionable ranked candidates verified already landed on main (
cargo macheteclean,cargo tree -dmatches the recorded post-state); dated applied/not-applied record merged in docs(contracts): record the #4839 section 3 applied state #5044 (aa7cec30); crypto/signing retentions (serde_jcs,ring/aws-lc-rs,serde_json,tokio) deliberately kept — evidence. - Remove proven-extraneous first-party crates and obsolete capabilities end to end: manifests, callers, configuration, exports, fixtures, docs, packaging, and CI references. Done, verified 2026-11-12 against current
main:mvm-fc-lifecycledeleted with its crate (specs/contracts/build-dependency-decisions.md, lockfile 729 → 728),mvm-observabilityfolded intomvm-cli(Consolidate small workspace crates into logical crates where the split has no reason #4542), the fourmvm-gpu-*crates folded intomvm-gpu(refactor(gpu): fold the shim cdylibs into mvm-gpu as examples #5021),tracing-logdropped (build: drop tracing-log forwarding; rank dependency reduction candidates (#4839) #4935),mvm-egress-proxyretired;ls crates/shows nomvm-fc-lifecycle, nomvm-observability, and onemvm-gpu. The dependency usage map, evidence matrix and unused-dependency re-scan (docs(deps): dependency usage map and unused-dependency scan (#4839) #4884, docs(contracts): dependency evidence matrix for the #4839 section 3 ranking #4939, docs(contracts): record the unused-dependency and duplicate-version re-scan (#4839) #4959) found no further proven-extraneous package; the packages with no reverse dependants (mvm-conformance,mvm-gpu,mvm-hostlib,runmvm-client) are each a shipped or test surface. - Consolidate the
mvm-gpu-*shim crates (mvm-gpu-cuda-shim,mvm-gpu-cudart-shim,mvm-gpu-nvml-shim,mvm-gpu-shim-core) into the single umbrella cratemvm-gpu; no standalonemvm-gpu-*crates remain. Done in PR refactor(gpu): fold the shim cdylibs into mvm-gpu as examples #5021 (merge commit 90e445b, CI green): shim cdylibs folded intomvm-gpuascrate-type=["cdylib"]examples;git ls-tree origin/main crates/shows onlycrates/mvm-gpu; exhaustive grep for the old crate names is clean;mvm-build/mvm-gpu/mvm-conformancetests and clippy green. - Measure accepted reductions before/after using identical commands and environments. Explain dependencies and boundaries deliberately retained.
4. Explain and fix excessive/repeated release compilation
- Reproduce the maintainer's actual release-build workflow and identify what the approximately 1,017 entries count; do not assume they are unique third-party packages.
- Separate unique packages from host/guest targets, profiles, feature-resolved units, procedural macros/build scripts, toolchains, and genuinely duplicated work.
- Diagnose cache/target-directory mismatches, changing environment inputs, over-broad build-script reruns, repeated hashing/code generation, and redundant packaging/build invocations. Done via docs(deps): record release-build unit counts and repeated compilation (#4839) #4888 and docs(contracts): release-build timings, binary contribution and build diagnosis (#4839) #4988 (
specs/contracts/build-dependency-decisions.md, "Diagnosis"): a five-row table covering cache/target-directory mismatch, environment inputs, over-broad build-script rerun, repeated hashing/generation and redundant invocations, backed by timed runs. The over-broadrerun-if-changedoncratesandsrcinmvm-cli's build script is diagnosed and recorded as found, not fixed; fixing it and demonstrating the result in the release workflow is the separate open box below. - Record comparable clean release-build time, no-change rebuild time, repeated compilation, and final binary/guest-artifact size, with architecture, target, profile, stripping, compression, and environment.
- Demonstrate the accepted fixes in that same real release workflow and document remaining necessary repeated work. Done via docs(deps): record release-build unit counts and repeated compilation (#4839) #4888, docs(contracts): release-build timings, binary contribution and build diagnosis (#4839) #4988 and docs(contracts): record the #4839 section 3 applied state #5044 (
specs/contracts/build-dependency-decisions.md, "The accepted fix, shown in the same workflow" and "Repeated work that remains necessary"): the accepted fix, the embed content store at~/.cache/mvm/embed, was shown in the same release-build command, clean target directory and host with only the store differing (r3againstr4: 512.9 s against 372.2 s, 27 percent less; themvm-clibuild script 150.9 s against 0.9 s; identicalmvmctlSHA-256). The remaining necessary repeated work is documented: any edit undercrates/mvm-clicosts about 160 s (one codegen unit and fat LTO), any file change undercratesorsrcreruns themvm-clibuild script, 343 packages build up to three times (profile and unwind constraints, each recorded), and the first build after the store is lost spends 150 to 180 s cross-compiling five guest binaries. Limits: a warm-store zigbuild run was not taken; the dependency reductions are within the run-to-run spread, so no time saving is claimed for them; the over-broadrerun-if-changedoncratesandsrcwas diagnosed and deliberately not changed, because those directories feed the source-dirty stamp, and is listed as remaining work, not as a fix.
5. Simplify toolchain, development, and CI paths
-
Complete bounded macOS fixture/source-identity and owned-child test-performance follow-up Make pair-image tests hermetic and bound macOS validation time #4840, coordinated with Hard gate: restore fast local feedback and bounded CI/release latency #4657 and PS-21: Replace checkout identity and manifest emission with in-process APIs #4412. Require an immutable source baseline, no production identity/custody bypass, identical-feature timing, and a completed native macOS witness; do not treat the interrupted feat(run): finish canonical launch-plan command support #4838 run as a diagnosed regression or a pass.
-
Verify and complete the ordinary-build/CI Rust 1.97.1 pin and Cargo-wrapper/caller consolidation described in surviving work context.
-
Retire the optional native Cranelift development path while preserving WASM execution and required WASM support.
-
Ensure ordinary Cargo and recipes share the intended cache within a worktree without cross-worktree contamination. Evidence/limits.
-
Preserve supported-platform, WASM, fuzz, eBPF, supply-chain, and release coverage; local validation must match the owning CI gates. Done when (maintainer-approved wording, comment): every CI coverage class has a named owning job in a manifest verified by a gate; removing a lane breaks the gate. Done via test(ci): gate CI coverage classes; classify workflow-reading tests (#4839) #5437:
xtask check-ci-coverage(incheck-all) maps each coverage class (supply-chain, fuzz, eBPF, WASM, release, and the Windows, aarch64 and macOS platforms) to its owning workflow job (10 jobs present), checks the job is a live key with the claims ledger'sci_anchorsparser so a commented-out job does not count, and fails on anycheck-*gate a workflow invokes thatcheck-alldoes not run. Three negative tests show a removed job, a commented-out job and an unrun gate are each reported. Limit: it checks that the job exists, not that the right step is still inside it. -
Measure CI against the ≤30-minute goal. Do not remove a slow gate without identifying its invariant and providing equivalent evidence.
-
Coordinate embedded-toolchain/Nix work with its existing owner rather than duplicating it.
6. Simplify capability surface, Rust, tests, and documentation
- Inventory capabilities, flags, environment/configuration options, modes, backends, compatibility layers, and supported combinations; retire only those shown obsolete.
- Review ownership, allocation, errors, state transitions, cancellation, lock/task ownership, cleanup/rollback, unsafe/FFI boundaries, and panic paths affected by changes. Done when (maintainer-approved wording, comment): split. Checkable part:
clippy::undocumented_unsafe_blocksisdenyworkspace-wide and clippy is clean, with a reviewed-to-date table recorded. The ongoing part moves to the PR checklist. Done via chore(lints): deny undocumented_unsafe_blocks workspace-wide (#4839) #5438 and docs: lifecycle review ledger and PR checklist item (#4839) #5444:clippy::undocumented_unsafe_blocksisdenyworkspace-wide and clippy is clean (about 200SAFETYcomments written after reading each site; three packages that silently skipped the lint table now inherit it; Linux cross-check included); the reviewed-to-date table isspecs/contracts/lifecycle-review-ledger.md(reviewed areas with the PRs that fixed defects, the files read, a "Files NOT read" section, and the five known-unfixed lifecycle gaps with owner and next step); the ongoing part is item 7 of the PR checklist inAGENTS.md(state the state transitions and resource owners and add a failing-before test). Limit: the review is not exhaustive; the ledger says plainly which files were not read (parts ofhvf_restore.rs,kernel_boot.rs, flowmux and libkrun), and Linux-only code behind features outside CI's set, aarch64 and musl arms and the excluded fuzz and eBPF crates were not linted. - Replace misleading existence/text/internal-wiring tests with meaningful outcome tests where practical; retain security, compatibility, recovery, platform, and user-visible BDD coverage. Done when (maintainer-approved wording, comment): tests that read workflow, justfile or source text are limited to an allowlist where each entry states the gate it owns. Done via test(ci): gate CI coverage classes; classify workflow-reading tests (#4839) #5437 and test(ci): allowlist tests that read gate text and ratchet new ones (#4839) #5446:
xtask check-text-reading-tests(incheck-all) fails on a new test file that reads workflow, recipe, script, nix orinstall.shtext, or a new test in a listed file, unlessxtask/data/text-reading-tests.txthas an entry with a maximum test count and the gate it owns (19 files seeded, each with its owned gate; five negative tests);release_assets.rsandgithub_actions_extended_e2e.rsnow carry per-test owner lines. Two pin tests were deleted and one tightened (test(ci): gate CI coverage classes; classify workflow-reading tests (#4839) #5437). Limit: no test undertests/orcrates/*/testsincludes Rust source text throughinclude_str!today, and inline unit tests that read source are not scanned; claim 20fresh_install_*text scans stay and their replacement is a proposal in the claims matrix. - Trace affected security claims through reachable enforcement and regression tests; fix gaps or narrow unsupported claims.
- Synchronize active docs, CLI reference, recipes, CI, release instructions, and actual behavior; remove obsolete scripts/references only after checking consumers. Done when (maintainer-approved wording, comment): the docs gates pass and name every checked surface; prose package counts either match
cargo metadataor are removed. Done via docs: sync manifest sizing key, release instructions and recipes with behavior (#4839) #5351, docs(repo): complete root layout inventory (#4839) #5240, docs(repo): inventory third_party, data dirs and nix/packages owners (#4839) #5347 and chore: gate process labels in comments; drop stale package counts (#4839) #5410:check-cli-help-matches-docs,check-doc-links,check-doc-claims,check-documented-env,check-spec-hygieneandcheck-root-layoutare all incheck-alland passed in the merge queue; the CLAUDE.md prose package count was removed (chore: gate process labels in comments; drop stale package counts (#4839) #5410); the inventory counts were set to the measured 28 packages and 52 bins and the stalemvm-egress-proxyrow removed; the removedvcpuskey andjust release::imagereferences were fixed (docs: sync manifest sizing key, release instructions and recipes with behavior (#4839) #5351). Limit: the inventory counts are a dated measurement and no gate ties them tocargo metadata. - Keep specs limited to active ADRs, stable contracts/security claims, and linked exceptional design attachments; retire obsolete progress ledgers after checking workflow consumers. Done in PR chore(specs): retire the cleanup progress ledgers #5022 (merge commit e4ea8ef, CI green):
cleanup-active-work.mdandcleanup-preservation-ledger.mdretired (history under thespecs-archive-2026-11-02tag), both added tocheck-spec-hygiene's retired-dashboard list; consumer grep found no workflow/tooling references;check-spec-hygieneandcheck-root-layoutclean.
Demo-policy coordination gate
A subsequent #84-owner report adds this constraint; the repository-qualified issue link and authoritative schema/commit still need confirmation before handoff:
-
The schema field
EnvironmentPolicy.request_budgetmay be optional, but the demo requires exactly{"max_requests":1,"max_spend_microusd":500000,"model":"claude-haiku-4-5-20251001"}. Omission or widening must refuse under the strict Anthropic wire policy. -
The prior descriptor abbreviated
663e...is historical, not export-ready; this abbreviation is not a usable artifact identifier. -
Final paired consumer CLI/runtime source must include the contract and enforcement once landed. Coordinate contract/dependency cleanup with that owner.
-
The old three-turn wrapper does not prove client compatibility with beta-header refusal, token caps, or the one-request limit.
-
Link the authoritative demo issue, schema, and landed commit; verify paired CLI/runtime enforcement and required client compatibility before artifact handoff. No live/client/model validation is claimed here.
Current evidence and candidate triage
These are starting points from surviving reports, not completed umbrella acceptance criteria:
- A scoped review of the compiler/wrapper/cache cleanup reported no actionable regressions and passing targeted Python/shell/guard checks. Full Cargo validation, live VM/Nix execution, release measurements, and CI timing are not established by that review.
mvm-fc-lifecycle: deletion candidate reported as an isolated privileged prototype with no default integration. Verify active ownership and deliberate prototype requirements before declaring it extraneous. Its standalone dependency closure is not automatically removable from other shipped artifacts.mvm-mcp: candidate to fold into the CLI while preserving MCP functionality, protocol tests, bounded transport, authorization, and signed-plan/audit behavior. Expect a first-party package-boundary reduction, not removal of its entire shared third-party closure.mvm-net: retain its backend-neutral/default-deny enforcement boundary unless a separately justified architecture change preserves it. Small size alone does not justify removal.network-perfanddev-watch: reported to have live consumers; do not treat them as unused without further evidence.
PS-06 coordination: distinct from the HVF-first demo
The existing PS-06 owner retains #3716 and tinylabscom/mvm-packs#12, including all five stages below. No duplicate source writer, remote validation host, release job, or native lifecycle owner is assigned here.
- Finish current-source Linux workspace/Clippy/all-target/BDD/format validation. — Evidence (2026-11-06, exact source 40ce000, the v0.27.0-rc.3 release commit):
ci-mainrun 42218673590 succeeded for Test (main), Test Linux and conformance, Test workspace (aarch64), Test workspace (doctests + feature suites), BDD conformance, BDD live PS-11 frozen volume (Linux, Firecracker), Test eBPF telemetry load/attach, Test release profile (overflow + no debug assertions) and the lint feature-coverage matrix; PR release: v0.27.0-rc.3 #5162 CI succeeded for Lint core (fmt + clippy), Lint feature combinations (clippy), Invariant (xtask check-all) and Workspace shard 1. A Release dry-run on the previous main (06a6d7a, run 42186498320) also passed end to end. - Establish original-owner persistent Firecracker lifecycle through per-active-VM supervision, detach/stop, and freshly admitted start. Deferred past the demo by maintainer decision (2026-11-14); not complete. Tracked, with the preserved contract and model and what must happen to resume, in Persistent Firecracker lifecycle owner: deferred past the demo (preserve the contract and model) #5522. This is the explicit, linked scope disposition the completion box allows; the box stays unchecked.
- Preserve original signed metadata, failure cleanup/resource census, and one-way owner-only protected completion projection. Deferred past the demo by maintainer decision (2026-11-14); not complete. Tracked, with the preserved contract and model and what must happen to resume, in Persistent Firecracker lifecycle owner: deferred past the demo (preserve the contract and model) #5522. This is the explicit, linked scope disposition the completion box allows; the box stays unchecked.
- Witness genuine released Linux producer A plus independent B, authenticated extracted-ELF H, original signed rootfs/verity/metadata/raw hash, accepted base lock, independent current revocations, byte rebuild, signing and attestation. Deferred past the demo by maintainer decision (2026-11-14); not complete. Tracked, with the preserved contract and model and what must happen to resume, in Persistent Firecracker lifecycle owner: deferred past the demo (preserve the contract and model) #5522. This is the explicit, linked scope disposition the completion box allows; the box stays unchecked.
- Witness transient and persistent consumer lifecycle/refusals/cleanup with exact evidence for the owning checkboxes. Deferred past the demo by maintainer decision (2026-11-14); not complete. Tracked, with the preserved contract and model and what must happen to resume, in Persistent Firecracker lifecycle owner: deferred past the demo (preserve the contract and model) #5522. This is the explicit, linked scope disposition the completion box allows; the box stays unchecked.
- Resolve the process-host design dependency before supervisor wiring: a typed library-owned lifecycle, exact process-host executable/source/signing identity, original Child custody (no PID-file resurrection), bounded startup data/FD handling, invoking UID/privilege bounds, authenticated generation/refusals, detach/stop/fresh admission, and release effects. The prior hidden-CLI bootstrap codec is groundwork, not approval for a hidden CLI supervisor or operation backfill. PS-21: Complete typed VMM lifecycle without shell or utility processes #4414 does not automatically authorize a new host role; do not repurpose broker/host-agent trust boundaries. The owner is preparing a focused review packet. Deferred past the demo by maintainer decision (2026-11-14); not complete. Tracked, with the preserved contract and model and what must happen to resume, in Persistent Firecracker lifecycle owner: deferred past the demo (preserve the contract and model) #5522. This is the explicit, linked scope disposition the completion box allows; the box stays unchecked.
PS-06 source tests do not close live/artifact acceptance. Preserve its existing isolated validation-host allocation; no competing builds/VMs. Firecracker evidence remains necessary for PS-06 but does not substitute for the first required HVF demo.
Low-priority terminal checklist view
- Deliver
just maint::watch-issuesin refactor: simplify builds and retire unused workspace boundaries #4854: plain tasks, 30-second TTY refresh, once/non-TTY modes and explicit child refs; nine tests and exact live once invocation passed. No dependency/crawler; convenience only, not issue implementation.
Delivery and definition of done
For each child workstream: identify callers/contracts, classify compatibility and security effects, make the smallest coherent change, update behavior tests/docs, conduct independent reviews, resolve findings, and record exact validation and measurements. Keep skipped/unavailable/failing checks explicit.
- All child workstreams have verified completion or an explicit, linked scope disposition; no unfinished checklist item is silently checked or deleted to close this umbrella.
- Required validation is recorded:
cargo fmt --all --check,cargo test --workspace,cargo clippy --workspace -- -D warnings,cargo check --workspace, andjust check::gatedwhere shared type shapes change, plus affected BDD/platform/security/integration/image/release gates. - Final before/after dependency/build/binary/CI evidence demonstrates the outcome; remaining risks and retained complexity are documented.
- Active instructions/docs match the final behavior and repository boundaries.
Respect builder-VM and long-operation rules, worktree isolation, and unrelated changes. This issue authorizes tracking the requested cleanup; it does not independently authorize publishing branches, landing changes, billable infrastructure, destructive external operations, or weakening security contracts.
Coordinated runtime → pack → M4 delivery
The single governing checklist records accountable runtime/artifact/consumer owners, landed source milestones, the actual lifecycle-transcript snapshot refusal, retained-state/runner prerequisites, and the separate exact-tag, official-pack-publication and M4 approvals still required. Source merges and the deferred durable helper do not complete the parent release/native checkboxes.
Activity
auser commentedon Oct 28, 2026
Execution started: Rust-size ceiling and first consolidation
The maintainer has set an explicit goal of fewer than 1,000,000 Rust lines and asked that every umbrella acceptance item be addressed and verified. This is the full delivery goal, not permission to close the issue after one slice.
Initial worktree baseline: 1,227,697 physical Rust lines in 2,808
.rsfiles, measured fromrg --files -g "*.rs"(excluding build/state directories) and Pythonbytes.splitlines(). Includes comments and blanks; this is not executable SLOC. Path categories: 81,999 underthird_party/orcrates/deps/, 98,362 in test/bench/example paths, and 1,047,336 elsewhere (including inline tests and potentially generated source). These categories are not a complete generated/test classification. No category is being silently excluded to meet the ceiling. A first-party/generated/vendor/test breakdown and a reproducible final count remain outstanding.Latest published release inspected:
v0.22.0, published 2026-10-02, commitb6b77f83730bed54d52f5f8b9cdcd8d5136ca1de. This identifies a reference release; it does not resolve which release the original runbook meant by “next release.”In progress: #4542’s
mvm-mcpconsolidation into a private CLI module. Source and the issue audit establish a single production consumer and no independent published package boundary. Preserve protocol/drive/policy/error tests and the pinned tool contract; remove the package and redundant standalone example. Expect one fewer first-party dependency node, not hundreds fewer third-party dependencies. This alone will not meet the Rust-size goal.Preserving existing compiler/cache edits and the independently owned demo-policy work. The privileged Firecracker prototype is not being deleted while its relationship to #4414 remains unresolved. Graph lookup timed out, so this slice uses targeted source, manifest, and issue evidence. No new checkbox is marked complete yet.
auser commentedon Oct 28, 2026
Workspace update
The cleanup thread now has attached worktrees for
mvm,mvm-studio,mvm-images, andmvm-packs. This removes the Studio attachment blocker recorded above. Revalidate the library-parity findings against the attached Studio worktree, which is newer than the previously inspected sibling checkout, before implementing.Scope remains the full umbrella, including no legacy support, library-first Studio integration, and the Rust-size ceiling. Attachment is not a merge/publication authorization and does not resolve the separately reported Git index-lock blocker in the MVM worktree.
auser commentedon Oct 28, 2026
First Studio library-parity cleanup implemented (validation blocked)
On the attached Studio branch
chore/library-validation-cleanup:mvm_client::validate_vm_name, preserving typedInvalidSpecerrors and the library error reason. This closes concrete drift: Studio previously accepted 64-character names and leading/trailing hyphens that MVM refuses.async-traitdeclaration/feature edge and obsolete comments describing the deleted subprocess backfill. No third-party closure reduction is claimed; other consumers still need the package.Passed: direct rustfmt checks on both changed Rust files and
git diff --check.Blocked: Cargo cannot load
../mvm/crates/mvm-clientfrom the Studio Delta checkout because the attached repositories live under different parent directories. Studio also requires the../mvmddependency.just ciwas attempted exactly as required but this checkout contains no Justfile. No Rust test, compilation, full CI, or live integration pass is claimed. These blockers must be resolved before checking the corresponding umbrella acceptance items.The MVM-side MCP consolidation remains uncommitted and not fully validated; its moved sources are local-only while the separately reported stale Git index lock prevents tracking them for Delta replication. No umbrella completion is claimed.
auser commentedon Oct 28, 2026
Maintainer clarification and dependency correction
Maintainer authorized removing the stale MVM Git index lock, attached MVMD, and asked: “I thought mvm supplied the mvm-client... can we make sure that’s true?” Direction reaffirmed: “Simplify the heck out of this project.”
Verified:
mvm/crates/mvm-client/Cargo.tomldeclares themvm-clientpackage. Studio consumes that machine-operation library. The separatemvmd-clientdeclaration in Studio is for its legacy fleet integration, not the MVM facade.The newly attached MVMD V1 tree has no
crates/mvmd-clientand nomvmd-gatewaypackage. Its root README explicitly describes a replacement implementation; it exposesmvmd-types,mvmd-core,mvmd-api, themvmdbinary, and a detachedmvmd-mvmadapter consuming MVM. Therefore the Studio build blocker is not merely checkout layout: its old fleet client/sidecar contract is obsolete. Do not resurrect an old checkout or invent a compatibility crate just to make it compile. Retirement/migration of the affected fleet/auth/observability surface needs an explicit current contract and acceptance decision; local MvmClient functionality remains required.The authorized stale-lock removal was performed only after checking for an owner and verifying that the lock did not change during inspection. The four moved MCP files are now Git-tracked and recognized by Delta. No commit, push, or merge occurred. MVM validation restarted with durable
/tmplogs rather than turn-scoped scratch.auser commentedon Oct 28, 2026
Maintainer request: “Post an issue in mvm-studio to give these details.” Created tinylabscom/mvm-studio#30 — Retire obsolete MVMD gateway integration and enforce library-only MVM access, labelled
cleanup.The issue records the verified
mvm-clientvsmvmd-clientownership distinction; current MVMD V1’s missing legacy client/gateway; Studio’s already-completed CLI-backfill removal; the remaining gateway/sidecar and unused shell-plugin surfaces; required scope/ADR decisions; the existing uncommitted validation cleanup; and exact validation blockers. It links Studio #11/#17/#19/#18/#25/#12 and MVM #4414/#4415.Use Studio #30 for implementation ownership and evidence, with this umbrella retaining the cross-repository acceptance criteria. No acceptance item is marked complete by creating the issue.
auser commentedon Oct 28, 2026
Parallel cleanup results and validation status
9cc2502293c78f1e39746a9444ca2ad711aa08ce4afb8d9ca8c3004a77cb106d, and manifest metadata. This is source-review evidence, not a test-execution pass.mvm-sdk; the required Rustmvm-clientremains. The old gateway/sidecar retirement is still a separate design-dependent task in tinylabscom/mvm-studio#30.Validation exposed a harness/environment problem: background jobs inherited Delta’s per-terminal TMPDIR, which disappeared after the terminal call. pnpm failed before install; Clang failed to create a temporary file at MCP-test link time. Those are failures, not test passes. Retried with job-owned durable temporary directories and logs: MVM
/tmp/mvm-cleanup-validated-5953tubt, Studio UI/tmp/studio-cleanup-validated-ib_bazb1. The MVM queue includes MCP/tool-gate tests, capture tests, Linux-target capture test compilation, workspace check, and Clippy. UI queue includes frozen install, tests, lint, typecheck, and build.Separately recovered previous workspace validation records: fmt passed; scoped tests, workspace check, Clippy, and workspace tests each timed out at 1200 seconds. No full-suite pass or ≤30-minute CI claim is made.
The demo-readiness scout is still assigned the 12 current demo issues and active-owner/dependency reconciliation. Demo/security contracts remain unchanged; no publication, live VM/model execution, or issue closure is authorized by these reports.
auser commentedon Oct 28, 2026
Demo critical-path reconciliation — 12 tracked demo issues
The original twelve issues below track full delivery and acceptance, not merely code presence. This comment is reconciled with the live umbrella ledger as of 2026-10-30; linked receipts distinguish completed source work from outstanding artifact/runtime witnesses.
Completion rule for every box below: address the complete owning-issue scope; implement required changes; link reviewed PRs, passing validation and merged commits; and verify any required artifact/release/live acceptance. Existing landed work may supply evidence—do not create redundant PRs. A diagnosis, plan, open PR or partial source delivery is not completion. Separate signing, publication and live-execution approvals still apply.
as i64casts in syscall_nr table #83 archive is not final.mvm-imageshas zero currently open demo-labelled issues. This is not proof of image readiness; preserve its Linux-only ownership and do not add MVM runtime/source roles or publication shortcuts.Current critical path — 2026-10-30
Budget/consumer source #4845/#4859/#4870 is landed; the older “wait for an authoritative contract” instruction is superseded. The exact client contract still requires pinned model
claude-haiku-4-5-20251001,max_tokens <= 20000, no beta headers/service_tier/unsupported fields/server-side tools, and refusal on missing/widened budget, wrong model, second request or excess spend. A version probe, historical three-turn wrapper or budget flag is not the witness.All twelve boxes retain full owning-issue acceptance. Source merges do not grant official publication, paid execution, old-state cleanup or broader credential authority. Both umbrella bodies and every comment checklist remain in completion scope.
auser commentedon Oct 28, 2026
Full-goal audit requested and delegated
Maintainer: “Look at the ADRs, the README.md, and the issues and PRs and make sure we’ve covered all of our goals. And then continue working with subagents.”
Three non-overlapping read-only audits are now reconciling (1) MVM README/current ADRs/issues/PRs, (2) Studio + MVMD goals/contracts, and (3) images + packs ownership/trust/publication goals. Outputs must distinguish documented intent, landed source, and witnessed behavior, identify missing umbrella acceptance criteria, and respect current demo-owner reservations. A fourth agent is identifying substantial safe code-deletion candidates only; implementation selection follows the goals reconciliation.
The audit explicitly preserves the latest core goals: all fire-and-forget tracing and correlation/redaction/flush semantics, signed audit, UX/DX, cold<300ms and warm<50ms, all supported backends plus future WHP, authenticated vsock, host-held secrets, mailbox design, and library-first parity. It also checks README/ADR obligations beyond that short list: UID/filesystem isolation, verity, sealed DevOnly bans, artifact signatures/provenance/revocation, CVE/SBOM evidence, relocatable installation, and real consumer/release validation. No goal-coverage completion claim yet.
New executed evidence: MCP tests 37/37 passed; tool-gate tests 11/11 passed with durable TMPDIR. Studio frozen UI install, lint, typecheck and build passed; UI tests report 60 passed and 3 skipped across 9 files. Rust capture/workspace validation remains in progress. Studio Rust still has the obsolete MVMD integration blocker described in #30; UI success is not full application validation.
auser commentedon Oct 28, 2026
Studio architecture recommendation — not yet approved
Maintainer asked: “Is that a good option?” about retiring Studio’s obsolete gateway/fleet/auth integration.
Recommendation: local-first Studio now, not local-only forever. Retire the nonfunctional legacy
mvmd-client/mvmd-gatewayintegration and its obsolete UI/configuration, preserve all current local machine/library functionality, tracing, telemetry, verified audit, and required authorization, and keep fleet orchestration in MVMD V1. Future remote Studio support should use a current typed client contract once its auth/tenancy/capability semantics are defined—not a compatibility bridge over the retired API.Alternatives: repairing the old integration preserves an obsolete architecture; immediately building full MVMD V1 integration adds new public/auth contracts and competes with the HVF-first demo. Local-first is the smallest current path consistent with Studio #19, ADR0005, and cleanup goals. It does not itself fix MVM’s HVF/runtime/release blockers or prove startup budgets.
Important boundary: retiring legacy fleet authentication does not remove local authorization, signed admission/audit, secret custody, or required authentication on any future loopback server. Existing tracing on retained paths must remain. This is a recommendation for the decision tracked in tinylabscom/mvm-studio#30, not owner approval or permission to silently drop required functionality.
auser commentedon Oct 28, 2026
Owner-approved Studio retirement; implementation assigned
The maintainer explicitly approved local-first Studio now, not local-only forever (“Yes, I approve”). Recorded the decision in tinylabscom/mvm-studio#30. An isolated worker now owns the end-to-end legacy gateway/fleet/auth retirement plus ADR/docs/tests, preserving local MvmClient operations, tracing/telemetry/output, verified audit, authorization, host-held secrets, existing user state, and the independent PR#29 telemetry work. No new remote compatibility stack; future MVMD V1 support requires a current typed contract. This does not authorize landing or publication. HVF remains the first required demo backend.
Separately integrated the private CaptureRequestBuilder simplification: 102 fewer Rust lines across the two CLI deps-capture files, identical values at all seven constructors, no production operation/validator or audit/tracing changes. Parent compilation/outcome tests remain pending.
Current validation: MCP37 and tool-gate11 tests passed; capture host11+4 tests passed. Linux-target capture compilation reached an environment blocker: missing
aarch64-linux-gnu-gccfor tree-sitter-python. No Linux execution or cross-compilation pass is claimed, and workspace check/Clippy have not yet run in that chained attempt. Studio UI validation remains passing (60 tests passed,3 skipped; lint/typecheck/build passed).auser commentedon Oct 28, 2026
New bounded cleanup follow-up from #5 / PR #4838: #4840 — Make pair-image tests hermetic and bound macOS validation time. It tracks the real compiled-from-checkout coupling, stable-baseline reproduction, owned-child cancellation/diagnostics, and identical-configuration timing evidence. All acceptance items remain unchecked. PR #4838 has green reported CI (20,731 Linux workspace tests), but its failed/stopped native macOS run is explicitly not being marked green. The separate local-fixture curl removal in #4838 preserves integrity checks and removes repeated subprocess overhead; it does not claim to solve the pair-checkout coupling.
auser commentedon Oct 28, 2026
Disposition coordinated with the cleanup owner: PR #4838 retains #5’s specifically approved singular-entrypoint/string-env compatibility. The newer no-legacy mandate should remove those forms only in an explicit coherent follow-up with callers/tests/docs migrated together; it is not an unrelated edit or silent landing blocker for #4838. No legacy-removal checkbox is being marked complete. macOS fixture/performance follow-up #4840 is unowned and also links the broader #4657 performance and #4412 canonical-identity owners. This scope coordination does not substitute for independent review or landing approval.
auser commentedon Oct 28, 2026
Coordination with PR #4838 / launch-plan owner
The #5 owner requested overlap coordination for PR #4838 at
8c709b7a817e8c59618d33277257d7917267758a. I independently checked its eight-file list; none matches our intended active cleanup edit scopes. We will not edit that PR branch or its files. This is not independent review or landing approval.The owner reports passing required CI (including 20,731 Linux workspace tests) and scoped local checks, but an interrupted >30-minute macOS workspace run with pair-default-image/cache failures and slow children. HEAD moved during that run; a clean-baseline regression is not established. Reported stack: derive_pair_key → open_mvm_checkout → probe_identity → Git attribute normalization of the real compiled-from checkout. The PR’s fixture curl→in-process-copy change reportedly reduced the same 39-test selection from 81.37s to 5.25s, with a feature-configuration caveat. These are owner-reported results, not rerun here.
No current cleanup agent owns that macOS fixture/source-identity performance slice. Searches found the broad #4657 latency umbrella and #4412 identity-contract owner, but no dedicated bounded child; the #5 owner will create/link a focused issue with exact reproduction, stable-source baseline, hermetic in-process fixture requirements and timing evidence. It must not weaken canonical identity, share mutable Cargo targets, remove diagnostics or inflate timeouts.
PR #4838 retains legacy launch-plan forms under #5’s separately approved compatibility scope. The newer no-legacy cleanup direction will be handled in a coherent explicitly tracked follow-up, not by silently rewriting another owner’s active PR. HVF-first demo remains the priority.
auser commentedon Oct 28, 2026
Continued execution
Maintainer requested continued work on the umbrella checkboxes. The current MVM tree passes
cargo fmt --all --checkandcargo test --locked -p mvm-cli --lib commands::deps::(32 tests, 0 failures). The scoped command took 299.72s including compilation; test execution was 0.11s. This is one run, not the controlled repeated latency evidence required by #4657/#4840.Workspace check, Clippy, the supported Zig-based Linux capture-test compilation, and workspace tests remain queued in the durable validation run
/tmp/mvm-cleanup-final-checks-zrn50yxm. The earlier plain GNU cross-check failed for a missing cross-compiler and is not a pass.The approved Studio local-first retirement remains assigned to its separate worker under Studio#30. A different worker is investigating #4840 read-only before proposing fixture changes; no overlap with PR#4838 or canonical identity work is authorized. PR#4838 remains open at this check, not verified landed. No broad acceptance item has been checked merely because a scoped test passed.
380 remaining items
auser commentedon Nov 15, 2026
Final source-slice handoff for this conversation (not two-architecture release closure): mvm #5549 head
caeb2a0d8a0ffc9cce56353d56f8e19430c89dd3merged by another operator as61348d65dc365131b61af93c9ac255dd20a50548at 2026-11-15T07:49:43Z, with all reported PR checks green. Exact-head localcargo fmt --all --check,TMPDIR=/tmp cargo test --workspaceincluding doctests,TMPDIR=/tmp cargo check --workspace, andTMPDIR=/tmp cargo clippy --workspace -- -D warningspassed. The first chainedxtask check-alltimed out in uvx pinned generator resolution after earlier gates; a separate exact-headUV_OFFLINE=1 UV_NO_PROGRESS=1 TMPDIR=/tmp cargo run -p xtask -- check-allcompleted 86 gates clean. This offline-cache environment is a distinct check from the timed-out default invocation; neither result is hidden. The new production policy DTO and tests were extracted fromcrates/mvm-contract/src/plan/bundle.rs(978→839 lines), preserve schema-8 LocalComposition, and still leave schema-9 reader cap at 8. Independent source review found no actionable issue; it did not independently run tests. No protected exporter, recipient, or signer authority was enabled. I did not perform the shared-branch merge.Two-architecture final-package evidence remains OPEN:
The native x86_64 and aarch64 AF_VSOCK base-probe evidence is real but is not a Python archive witness. Draft mvm-images #80 and mvm-packs #167 are still open, with no authorization in this conversation to merge them. The parallel Claude/skill session acknowledged the nonoverlap and separately signed Claude component decision, but does not assume ownership of this locally authored Python release. Next owners: #5417 schema/policy gate, #5420–#5426 exporter/recipient and signed-input chain, then exact x86_64 #5429 witness and separately authorized #5430 release; repeat the complete release gate for aarch64 under this issue. Protected custodian must provision a distinct official Python author public key/enrollment and retain the private half; owner must authorize exact source/signer/descriptor/rights and final digest before official signing/publication. Keep #4839 open and do not label either architecture shipped.
auser commentedon Nov 15, 2026
Unattended continuation request: proceed with independent schema-9 exporter, recipient, policy and candidate source/test work while maintainer is away; do not reinterpret this as exact-source/signer/descriptor/rights authorization for protected official Python signing, publication, main-branch merge, tag or images.lock advance. No official Python author key is available locally and none will be generated or promoted from a test signer. I am inspecting current main and active #5552 policy PR before editing overlapping authority paths. x86_64 remains first required final-file gate; aarch64 remains a separate required release gate. The evidence table in the prior comment remains open until exact signed archives and clean installed boots actually exist.
auser commentedon Nov 15, 2026
Independent #5425 prerequisite source is now in the attached mvm main worktree as a five-file uncommitted, unmerged diff: a private staged-archive readback verifies the same held file descriptor’s manifest signature, declared member hashes and digest under a caller-supplied public key before
PreparedPortableBootAssets::stagecan return; failure aborts private staging and does not publish a destination. It preserves existing selected Linux/runtime input re-verification and does not accept schema 9, change the reader cap, add publisher enrollment, use an official signer or create a Python package. Test-signer positive/wrong-key/tampered-member/abort cases were added. Focused tests and scoped Clippy/check are running in the isolated source thread; independent security review is also underway, so no pass or protected-export claim yet. This is reusable integrity readback, not a health boot, a fresh revocation feed, or recipient admission.auser commentedon Nov 15, 2026
Evidence update (local feature branch only; not a Python release):
feat/portable-stage-readbackat local commitfad0f3eec6756a167943185f3345f086378fef16adds verification of the exact privately staged.mvmpkgthrough its held file descriptor before publication, then rechecks the selected input closure. Wrong-key and changed-member failures abort staging without publishing. The archive readers now reject duplicate paths consistently; a signed, descriptor-pinned schema-9 production archive remains refused at the registry recipient boundary under reader cap 8.Validation: focused staged readback/refusal/duplicate tests passed; scoped Clippy (
mvm-core,mvm-bundler,mvm-client, all targets), formatting, diff check,xtask check-all(86 gates), andTMPDIR=/tmp MVM_EMBED=0 cargo test --workspace -- --test-threads=1passed. The unmodified parallel workspace command failed twice on different unrelated tests (CLI doctest import errors, then a temporary Git-config file read); the CLI doctest and Git-identity test each passed alone. The serial workspace run is a different check, not a claim that parallel tests are green.This does not close the production evidence boxes: no protected official Python signer, schema-9 producer/recipient admission, final x86_64/aarch64 Python archive SHA, selected-backend installed boot, or publication witness exists yet. No PR, push, merge, signing, tag, publication, or images.lock update was performed for this slice.
auser commentedon Nov 15, 2026
Coordination for the
plan_admission.rs/workload_runner/runner.rssplits: #5538 has an uncommitted behavior-only draft onfix/5538-record-then-respond, based on main7eb015747545051fb5e26ae2c6d5fa4c86e4de00. It is not a ready PR and must not be folded into a behavior-preserving split.Its current shared-file footprint is:
plan_admission.rs::populate_audit_substrate: project the actual admitted plan and production/development posture intoAuditContext, plus an assertion in the existing projection test.workload_runner/runner.rs: pass that context to the endpoint; compare it with the admitted plan; gate supported shutdown; register the target after boot/fork.workload_runner/claim.rs: carry the one-shot audit-control registration handle.The post-boot target-acquisition/registration path has a known ownership blocker, recorded on #5538; it is not a safe original-owner API for other work to consume. No shared cold-HVF custody refactor or FC/warm original-owner expansion is being implemented here. #5522's explicit deferral remains in force. The overlapping mailbox work is also paused on these shared paths pending a single owner and agreed scope.
No further edits are planned in the shared admission/runner/claim files while split coordination is pending. Remaining #5538 changes are bounded FlowMux UDP cleanup and documentation, followed by validation and a clearly blocked draft PR. Please share split source pins/new owning module paths before integration; do not merge stale dirty snapshots across the move. No landing or queue permission is conveyed.
auser commentedon Nov 15, 2026
PS-21 status/coordination, no checklist mutation: current read-only call-graph and operation inventory distinguishes zero
check-thin-cliallowlist entries on clean b45ab83 from functional parity, and names unresolved Firecracker shell/reaper and helper verify-to-exec paths under their existing owners. An isolated, uncommitted metadata-onlysecret.gethostlib/SDK slice is being validated against origin/main 7eb0157; no supervisor-regroup files are in its git diff (Delta's large deleted-file notice came from switching the isolated checkout to newer upstream history). Do not tick boxes 220/222/224/225/236/246/248 on this progress: required checks, source review, native #4414, Studio PR #29 current-pin build/tests and full #3730 child acceptance remain. Direct Delta cross-session messages could not be delivered because remote synchronization is disabled; this comment is the coordination handoff.auser commentedon Nov 15, 2026
Follow-up to the staged-readback evidence update: local-only
feat/portable-stage-readbackcommita80ced98ec83822e5531ed1cfb5cd817464a535bmakes pack completion compare the health-checked output to the verified staged manifest and report the archive SHA returned by that same held-file readback. It reads the existing host public key without minting an identity; missing/mismatched keys and changed payloads abort the owned stage before publication. An independent source review found and resolved a needless-borrow lint; no additional actionable security issue was identified in this narrow flow.Exact validation on the combined branch:
cargo fmt --all --check,git diff --check,TMPDIR=/tmp MVM_EMBED=0 cargo test -p mvm-bundler --lib staged_readback,TMPDIR=/tmp MVM_EMBED=0 cargo test -p mvm-cli --lib(2,881 passed, 1 ignored),TMPDIR=/tmp MVM_EMBED=0 cargo clippy -p mvm-cli -p mvm-bundler --all-targets -- -D warnings,TMPDIR=/tmp MVM_EMBED=0 cargo test --workspace -- --test-threads=1, andUV_OFFLINE=1 UV_NO_PROGRESS=1 TMPDIR=/tmp cargo run -p xtask -- check-all(86 gates) all passed. No live pack health boot, official protected signing, schema-9 admission, final Python package bytes or installed two-architecture boot were run. Cache-hit application resolution already rechecks current revocations and archive bytes; existing tests cover these separately, but their injected feed is not a real signed-feed positive fixture. All production evidence boxes remain open. No PR, push, merge, tag, publication or images.lock update.auser commentedon Nov 15, 2026
Handoff correction to the earlier local-only evidence note: the owner explicitly authorized a feature-branch push for review only. The clean two-commit source slice is now durably available at
feat/portable-stage-readback, heada80ced98ec83822e5531ed1cfb5cd817464a535b. No PR was opened, no merge tomain, official signing, publication, tag or image lock advance occurred. The production Python evidence boxes and limits in the prior comment remain unchanged.auser commentedon Nov 15, 2026
Piece 3 of 3 (split of
workload_runner/runner.rsby responsibility, pure move, identical test counts): #5565auser commentedon Nov 15, 2026
Maintainer authorization recorded (2026-11-15): the cleanup session may queue a #4839 refactor pull request without asking again when it is a pure move (item bodies verbatim, visibility not widened, no compatibility shims), the
cargo nextest listtest names and counts are identical before and after, every required check is green, and it changes no security claim, release behavior or published artifact. Anything that changes behavior, a claim, a budget, a verification or an audit path, and anything outside #4839, still needs an explicit go-ahead. First use: #5565 (the workload runner split).auser commentedon Nov 15, 2026
Piece 2 of 3 (split of
plan_admission.rs) is up as #5566. It is a pure move: same item bodies, names and visibility, identical test list (2,344 before and after inmvm-hostd --lib). One note: #5564 editsadmit_and_start_with_verified_release, which now lives inplan_admission/start.rs.refactor(hostd): fold small plan_admission modules and share one test…
auser commentedon Nov 15, 2026
Audit of boxes 348, 350 and 351 against
mainat 54eb7a0 (read-only; the cited witnesses were checked to exist, not re-run). No box is ticked. None of the three is "preserved" as worded, and ticking would overclaim.348 Tracing/stream completeness
shed_evidence_reaches_the_host_as_a_capacity_loss_summary,a_withheld_record_never_reaches_the_records_file_in_any_encoding). Detached lifetime is live-witnessed on Firecracker cold/warm and HVF cold only.xtask/data/telemetry/certification.mditself lists what is not certified, so "closed" overstates the epic's done criteria.350 Protocol and readiness (ADR-019)
ComponentState), bounded payload and control capacity (data_plane_limit_reserves_control_plane_capacity,a_saturated_data_plane_refuses_data_requests_and_still_admits_control), and a closed capability enum.ProtocolHellois optional; the guest accepts requests with no hello,handshake_as_guestechoes the version without a range check, andnegotiate_protocolchecks requested capabilities against the host's own list, not the guest's. A guest older than the host with fewer capabilities is not detected before dispatch. ADR-019 still says every session begins withProtocolHello.OutputConsumerSlowis ever emitted), cancellation (extension calls only), recovery (stream-level only; host reconnect belongs to open PS-21: Complete typed VMM lifecycle without shell or utility processes #4414 and PS-21: Validate and land complete client-boundary parity #4416, and machine exec: a nonzero guest exit (or a killed client) leaves the input journal unfinished and refuses every later exec #4809 stays open for the typed-task follow-up).351 Mailbox operational contract (ADR-057)
specs/contracts/mailbox-and-evidence-clause-witnesses.mdis stale and needs a refresh.DeliveryOutcome::FullandLateare vocabulary only; no end-to-end durable-acceptance probe; no mailbox benchmark). Each maps to an open owner the box already names: Add dedicated bounded mailbox vsock service and call/cast API #4629, Persist encrypted mailbox journals, snapshots, and effect receipts #4630, Move mailboxes across hosts with execution-epoch fencing #4631, Benchmark mailbox latency, overload, and typed-service isolation #4635 (program Program: signed pass-by-value mailboxes for workload, host, and user messaging #4622). Accepted architecture is not claim promotion: ADR-001 has no mailbox claim row, and that is preserved.Evidence is from a read-only audit; nothing was run. Refs #4839.
auser commentedon Nov 15, 2026
PS-21 footprint gate coordination (correction to prior progress note): isolated
secret.getbranch full workspace nextest 21,177/21,177 and required scoped/SDK checks passed, but itscheck-rust-footprintgreen result required increasing SLOC/physical ceilings to 991,493/1,259,820. The binding instruction says never loosen a check or budget to pass a gate, so do not treat that green result as accepted validation or publish the branch with this change. The worker measured base origin/main 7eb0157 at 991,237/1,259,528 before the slice, already above the current origin/main gate ceilings 991,000/1,259,000; the new operation adds approximately 256/292. This needs coordination with the footprint/supervisor-regroup owner for legitimate reduction or an explicit separate design/scope decision, not hidden compression, test deletion, or a cap bump. #3730/#4415/this umbrella remain open; no PR/merge/box change.auser commentedon Nov 16, 2026
Read-only audit of boxes 339 and 340 against
main(cited test functions were checked to exist; nothing was run). Neither box is ticked.339 Studio PR#29 telemetry adapter
tinylabscom/mvm-studioPR#29 is open and draft, with no reviews. Its body says ADR 0008 is Proposed and not to merge until it is accepted. CI on its head passed, which this box says is not a witness. It pins an old MVM commit, so nothing checks it against the current contract. Our 2026-11-15 comment asking who owns the pin update has no reply.MAX_PAGE_RECORDS=1024,pages_walk_the_stream_oldest_first_and_the_last_cursor_polls_for_more,telemetry_records_pages_from_the_cursor_it_is_given), stale-cursor refusal (a_cursor_off_a_line_boundary_or_past_the_end_is_refused; "reset" is only the caller restarting fromTelemetryCursor::START), unavailable/not-provisioned (NotProvisionedand the status fixtures), and fixtures on the producer side (studio_fixture_streams_match_the_committed_vectors,specs/contracts/studio-telemetry-contract.md).MachineId; the pane is Studio code) and cancellation (a synchronous page call; any cancel is Studio-side and has no MVM witness).340 MVMD V1 contract boundaries
authenticated_connection_is_not_mailbox_authority).stale_or_foreign_epochs_are_fenced_on_both_sidesandonly_the_destination_authority_can_sign_acceptance; no destination-host refusal witness here; durable fencing is Move mailboxes across hosts with execution-epoch fencing #4631), mailbox plaintext hidden from router APIs (a_router_sees_routing_fields_and_opaque_ciphertext_only; no attestation witness), bounded retries and transport (retry_tracking_cannot_grow_past_the_peer_bound,direct_delivery_is_pass_by_value_bounded_and_cancelable; no bounded-queue witness while Add dedicated bounded mailbox vsock service and call/cast API #4629 and Persist encrypted mailbox journals, snapshots, and effect receipts #4630 are open).check-core-runtime-freekeepstokioout ofmvm-core's default build andmvm-contractisno_std, but no gate says authority contract types carry no framework or transport types.Refs #4839.