Skip to content

LA: match admitted model offers with local context limits #154

Closed

Description

@andersalm

Part of #84. This product fix unblocks the admitted model journeys in #141 and #150.
Base and PR target: develop, approved by Anders under the weekly task rules; fetched base de1db928063425856498785476809fe939663738. Milestone: 0.7.2.
The native provider adds local context limits to its public offer. Runtime compares that offer with a saved activation summary that omits the limits, so Marketplace Get can admit the package while Use and cold Init report the service unavailable. The warm/cold result is recorded in #141; the source comparison identifies the cause.
Runtime will accept the local context limits for its admitted startup profile while keeping saved identity, policy and descriptor bytes exact. Older summaries remain compatible.
Done means: focused regressions accept the correct live limits and reject changed limits, policy and extra authority fields; the basic gate passes; an exact installed candidate completes Marketplace Get → Use → Assistant reply and cold Init with stable descriptor bytes; independent review and Irzhy's required review cover the candidate; its PR is merged into develop.
Reuse accepted pinned inputs and installed proof tools. Preserve existing Homes, keys and data. Record isolated artifact receipts and cleanup in this issue; #150 owns the three-platform CI journey and #98 owns resource qualification.

Activity

  1. added this to the 0.7.2 milestone on Oct 1, 2026
  2. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Prepared the bounded Runtime comparison fix on fetched develop@de1db928; source diff SHA-256 8a376cd4dadf509838423ba31d613f5da90b173292ec5f48f1cc4e9294778183.
    The live offer may report the exact 4096-token, one-slot profile and 1024-token output limit. Saved activation identity, policy and descriptor bytes stay exact; older summaries remain compatible.
    The existing readiness regressions now use provider context metadata and reject changed limits, malformed fields and changed policy.
    git diff --check, Home entropy, workspace format and chain-provider format passed. The focused startup-binding Rust suite owns the single local heavy-build slot and is running; independent source/security review is separate.
    Installed Get → Use → reply and cold Init remain open. The new task uses isolated proof inputs and preserves #141's failed/partial evidence.

  3. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Prepared candidate 409dc1d6 on develop@de1db928; source diff SHA-256 8a376cd4dadf509838423ba31d613f5da90b173292ec5f48f1cc4e9294778183.
    Focused elastos-server startup-binding suite: 34 passed, 0 failed; four explicit native-provider tests remain ignored until their installed fixture is supplied. The test build finished in 2m25s and released its heavy-build slot.
    All four basic gates passed. Independent source/security review round 1 passed with zero findings on that exact diff.
    The change accepts the exact admitted context profile while preserving saved descriptor bytes, identity and policy, including cold startup and retirement comparisons.
    Pre-push inventory: one new commit, one file, 38 insertions / 2 deletions; fetched-base divergence 0 behind / 1 ahead. The named task branch will be published to origin as a draft PR against develop.
    Installed Get → Use → Assistant reply, cold Init and required Irzhy review remain open; source tests alone establish the comparison fix.

  4. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Published 409dc1d6 to origin/fix/0.7.2-model-offer-context as draft PR #155, targeting develop. The task branch is clean and synchronized, divergence 0/0.
    Its 34 focused regressions, four basic gates and independent source/security review round 1 passed; four explicit native-fixture tests remain separate installed gates.
    Source qualification confirms current develop supplies the Mac local-engine reply path. #141 child-file confinement stays a separate task, so this candidate carries only the context comparison fix.
    The single heavy-build slot now belongs to this exact candidate's release Runtime artifact. Qualified pinned providers and existing proof tools will supply an isolated installed warm/cold journey.
    CI may run in parallel. Installed reply, cold Init and required human review remain open.

  5. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Exact release Runtime 409dc1d6d07667b46ec4c89b391d498e95f8de3c / tree 2ac757e79dfa5ea7fe80d3a980f3936679d41b07 built successfully in 1m56s; SHA-256 34c74f7c6a5e0055b7483640d9ed771c895572af1d559f5f17f42bf440822d4d.
    The fresh stable #154 fixture has copied and verified those exact bytes. Its isolated real-model Home proof is being prepared; #141's data and evidence stay intact.
    A separate run of the four existing explicit native-provider tests used the source-qualified package provider 8a76e990…: generated Init and startup retirement passed; refresh under cache pressure and deterministic long-output/cancel/restart failed (2 passed / 2 failed).
    These two failures remain open and separate from the 34 passing startup-binding regressions and actual model journey. Refresh retained an admitted state instead of reclaimed; the deterministic output run settled failed instead of completed. Each has run once; their fixture/product cause needs source inspection before a justified retry.
    The release build and test processes have ended. PR #155 stays draft while installed warm/cold acceptance, failure qualification and required human review remain open.

  6. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Exact installed proof is prepared for draft PR #155, candidate 409dc1d6d07667b46ec4c89b391d498e95f8de3c, tree 2ac757e79dfa5ea7fe80d3a980f3936679d41b07.
    The release Runtime build passed and its stable isolated installation matches SHA-256 34c74f7c6a5e0055b7483640d9ed771c895572af1d559f5f17f42bf440822d4d; the private build receipt preserves exact source and build inputs.
    The verified model-provider is 8a76e990a241f0fbd1a627266f6e0241c698e3b77f53bf1c0e0e8fc85ce86db7, originally built from 26de9c52 in run 36783916734. Native build inputs match this candidate; its original provenance stays intact.
    Candidate-owned served capsule files are staged and hash-bound. SmolLM2, licence, b10516 engine and archive-qualified Kubo pass their retained pins; disposable fixture trust and a fresh passkey keep existing Homes separate.
    The fresh fixture has a budget below 1 GiB and stays above the 10% free-disk reserve. The next bounded run uses actual Marketplace Get and Open in Assistant, one composer reply, owned cleanup, and one cold Init descriptor/readiness check.
    Existing headless CI capture is the changed display input; any readable automated media stays distinct from human UX acceptance.

  7. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    The #154 headless Home capture is readable on exact installed candidate 409dc1d6; automated visual evidence is now available for this changed display input.
    Two browser preflight attempts stopped before Get: the Welcome sheet covered the launcher, then the CI helper expected Marketplace in the launcher. Their private failures are retained. The proof now uses the existing sheet-close and taskbar Marketplace controls.
    Actual Marketplace Models → Details → Get is preparing the verified pinned package in the fresh isolated Home. Source and the retained #141 evidence stay unchanged.
    The bounded run continues to Assistant reply and cold Init; its current size stays below 1 GiB and free disk remains above 10%.

  8. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Source inspection identifies both explicit native fixture gaps, separate from the product comparison fix in PR #155.
    Explicit Content Use now records the caller's Keep intent. The pressure fixture retained model A and then expected capacity pressure to reclaim it; Runtime correctly protected that intent. The fixture needs an explicit release before it tests eviction.
    The deterministic engine accepts only /v1/chat/completions. Current local generation first calls /v1/chat/completions/input_tokens; the fixture needs that existing count contract before output/cancel assertions can run.
    The next change updates only those two accepted fixture inputs, then reruns their four explicit native tests once. Saved authority and production behavior stay exact. A final candidate artifact will bind the installed cold proof after the test correction.
    The warm/cold actual SmolLM2 Home journey continues separately, with readable headless media and owned cleanup.

  9. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    The corrected explicit native fixture run passed generated Init, startup retirement and same-process refresh under cache pressure (3 passed / 1 failed).
    The remaining long-output fixture passed generation, cancellation and restart/replay, then stopped at its second pressure case because that case also kept model A. The next test input explicitly releases Keep so it can reach the intended unknown-run protection gate.
    Only fixture inputs changed; Runtime's production comparison remains byte-identical to reviewed 409dc1d6. One narrow rerun covers the unresolved branch; accepted passing cases are retained.
    The real Home test has admitted its pinned package. Its missing owned temporary directory is being repaired before the first visible composer reply. This fixture result stays separate from installed product acceptance.
    PR #155 remains draft. Final artifact custody, cold descriptor proof and required human review remain open.

  10. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Prepared fixture-only tail bd990bcafc741992e45c518183b5aabc33ecabec / tree acd3ea82389b351b51549ccb6ee424e0c2df563f for PR #155; Runtime production source remains unchanged from 409dc1d6.
    All four explicit native cases now have passing evidence: generated Init, startup retirement and same-process pressure refresh passed together; the remaining long-output/cancel/restart case passed its one narrow rerun after its Keep input was corrected.
    The initial 2 failures and subsequent 1 failure remain recorded as obsolete fixture inputs, separate from the current passes. The 34 accepted regular regressions and unchanged basic gates are retained; diff and workspace format pass after the tail.
    Independent source/security review round 2 passed with zero findings on exact test diff SHA-256 ae6c63e3d81dbedefa460ebbaf1a90544ba532f1aa0b74b48284f21eb7d9d9dc; the two source review rounds are complete.
    Pre-push inventory: one fixture commit, one file, 10 insertions; task-remote divergence 0 behind / 1 ahead. Final candidate artifact and installed cold proof follow this forward publication; required human review stays open.

  11. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    The #154 headless Home capture is readable on exact installed candidate 409dc1d6; automated visual evidence is now available for this changed display input.
    Two browser preflight attempts stopped before Get: the Welcome sheet covered the launcher, then the CI helper expected Marketplace in the launcher. Their private failures are retained. The proof now uses the existing sheet-close and taskbar Marketplace controls.
    Actual Marketplace Models → Details → Get is preparing the verified pinned package in the fresh isolated Home. Source and the retained #141 evidence stay unchanged.
    The bounded run continues to Assistant reply and cold Init; its current size stays below 1 GiB and free disk remains above 10%.

  12. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Warm installed product proof passed for PR #155, original candidate 409dc1d6, Runtime SHA-256 34c74f7c6a5e0055b7483640d9ed771c895572af1d559f5f17f42bf440822d4d.
    Actual Marketplace Models → Details → Get admitted SmolLM2 and reported it ready. Open in Assistant selected that exact offer; one composer Send produced a visible “Hello.” reply, with one completed native-provider run bound to Assistant.
    Readable headless Home, Marketplace-ready and Assistant-reply media are retained with SHA-256 receipts. This automated evidence stays separate from human UX review.
    The owned controller expected a missing temporary directory; preparing that concrete fixture path restored native-provider startup. The original setup, launcher and redundant-picker preflight failures remain in private evidence.
    The saved activation descriptor remained exact through that recovery; canonical SHA-256 7e0b32d86b9372ef88917f87cdb623c113c22638f693f88b7d23a580a7fb57b6, activation pending false.
    All owned warm gateway/browser/provider/engine children stopped, and the gateway port is reusable.
    The final candidate is bd990bcafc741992e45c518183b5aabc33ecabec; its rebuilt Runtime is byte-identical. The next bounded phase binds its final receipt and proves cold Init with unchanged descriptor and current readiness.

  13. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    The isolated Home passed Marketplace Get, Open in Assistant and visible composer Send; SmolLM2 replied "Hello." Readable automated Home, Marketplace and Assistant media remain in the private receipt.
    Final candidate bd990bcafc741992e45c518183b5aabc33ecabec, tree acd3ea82389b351b51549ccb6ee424e0c2df563f, passed cold Init with the admitted model ready and activation pending cleared.
    The saved descriptor equals the warm baseline (canonical SHA256 7e0b32d86b9372ef88917f87cdb623c113c22638f693f88b7d23a580a7fb57b6); its file metadata proves the saved bytes were unchanged through the cold start.
    Built and installed Runtime SHA256 34c74f7c6a5e0055b7483640d9ed771c895572af1d559f5f17f42bf440822d4d match. The warm 409dc1d6 artifact is byte-identical to the final build; its original provenance remains separate.
    Qualified model-provider SHA256 8a76e990a241f0fbd1a627266f6e0241c698e3b77f53bf1c0e0e8fc85ce86db7, manifest, served capsules, pinned model and engine retain their exact source and hash bindings.
    Both successful phases stopped all owned children; the gateway port is reusable. The retained fixture is below 1 GiB, with 12.87% free disk. Existing Homes and #141 evidence remain preserved.
    The task branch is clean and origin divergence is 0/0. PR #155 remains draft for final CI and required human review/UX acceptance; automated media are separate evidence.

  14. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Exact bd990bcafc741992e45c518183b5aabc33ecabec retains the accepted warm Get/Assistant reply and cold readiness with unchanged descriptor bytes; the independent review and final installed receipts remain accepted.
    Seven current checks pass in run 36799566173, including both Linux architectures. The macOS job was still running at the bounded result read; its result remains open.
    PR #155 stays draft until that check completes, then moves to Irzhy's review. #141 and #150 include this exact reviewed dependency and wait for its reviewed integration into develop.
    The owned fixture remains stopped, with bounded private evidence. Next: complete the current CI gate and request the required human review; accepted tests and review rounds stay qualified.

  15. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Candidate: 00de6734b35cf6743086ee16da39a759abac1f89 in PR #155, pushed to origin/fix/0.7.2-model-offer-context; divergence 0/0.
    CI: 36855892880 started fresh checks after the stalled macOS job, as Anders requested.
    Result: the new commit is empty and has the same source tree as bd990bca; accepted source gates, independent review and installed warm/cold proof remain qualified.
    Next: when the exact candidate is green, mark #155 ready and request Irzhy. #141 and #150 wait for its reviewed integration into develop.

  16. andersalm commented on Oct 1, 2026

    @andersalm
    ContributorAuthor

    Candidate: 00de6734b35cf6743086ee16da39a759abac1f89 in PR #155; task branch matches origin.
    CI: 36855892880 passes all eight checks, including Linux x86-64, Linux ARM64 and macOS.
    Result: #155 is ready and irzhywau is requested. Accepted source tests, independent review and installed warm/cold proof remain qualified.
    Next: Irzhy’s review and the remaining human acceptance gate, then reviewed integration into develop; #141 and #150 retain that dependency.

  17. andersalm commented on Oct 2, 2026

    @andersalm
    ContributorAuthor

    PR #155 merged into develop as 0cba4e67edc79e7f25ba95392620c8f3c86cd220, from 00de6734b35cf6743086ee16da39a759abac1f89.
    CI 36855892880 passed all eight checks; 34 startup-binding regressions and installed Get → Use → Assistant reply plus cold Init passed on the same source tree.
    The operator’s Claude adversarial review is PASS; current-base compatibility is clean. This meets the task Done means under Anders’s 2 October merge rule.
    The hard-coded profile-limit follow-up is retained in open #98; branch and installed evidence stay preserved.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions