v0.41.17 field assessment: false-zero disclosures, import precision and daemon approval concurrency need reproduction

buildagent commented

2026-10-24 17:03:38 +02:00

Member

Two-part local-ai-server/W100 field assessment received after .17, with a hand-reviewed 20-task comparison. Treat observed query results separately from source-only hypotheses. Confirmed event fixes are recorded below; raw tool arguments/responses/source spans are still needed for the remaining cases.

Highest-priority observed cases:

XAML root event KeyDown="CustomWindow_KeyDown" on tl:CustomWindow in wndPosErfassungSchalter.xaml:10 returns no references, while child Click handlers resolve.
x:Name rbPosSchalter has four code-behind uses (26/34/54/81), but returns zero; search_symbols claims measured fallback zero while file_outline reports no_use_reference_channel for the same id. Dataset→C# references similarly appear as measured zero.
C# method groups new RoutedEventHandler(rbPosSchalter_Loaded) are missing. Four ToDoubleNN calls in busMdeGenerator.cs:635–636 reportedly absent despite no extraction diagnostics. Need exact source before classifying.
Python get_dependencies(direction=in) for llm/models.py reportedly returns924 unrelated *.models importers (1correct in first200). Explicit imported _attachment_block at agent.py:1300 remains unresolved despite import at56.
Workspace index_state from another root is attached to primary-only answers; limit clamps and negated-glob/path-only search semantics need raw payload verification.

Source-only hypotheses (UNVERIFIED): daemon load-modify-save approval paths in offer.rs/bundled.rs may bypass the CLI project_mutation_lock; background updater timeout/cooldown/log edges. Reproduce a daemon+CLI schedule before claiming approval loss. Never weaken publication/authority fencing.

Operational findings already tracked: coarse host fingerprint / expensive rebuilds #410; stale last-check package_updates record acknowledged on #442; C# imported static calls #445, generic method arguments #446 (candidate fix), JS→TS imports #447, extension calls #313, member proof #69, Svelte #433/#434.

The sample is not an overall completeness claim: Python187TP/1FP/7misses over194ground-truth items; C#/XAML4correct+2partial tasks, with stronger bridge precision but several uncertain zeros. Keep precision and recall separate. Next steps: obtain raw witnesses, create independent minimal reproductions, then split confirmed fixes into focused issues/PRs. Root XAML events and inconsistent zero disclosures rank above cosmetic output changes.

Raw partner reports retained at /tmp/cosi-04117-local-ai-assessment-raw.jsonl; actual item IDs01M5Q6NRAA232Q67BN4XSXJ8F9 and01M5Q6P9HJBTBNS08Y36V4E73Y. No partner machine or project state was changed.

Revalidation on published 0.41.46 (2026-11-09)

Fresh minimal reproduction using signed C# 0.1.11 and XAML 0.3.3: namespaced tl:CustomWindow root KeyDown resolves to the paired C# method, child Click resolves, all four code-behind uses of x:Name="rbPosSchalter" resolve and public find_references returns four rows, and new RoutedEventHandler(rbPosSchalter_Loaded) emits and resolves a C# method-group reference. Zero parse errors. These particular missing-reference claims no longer reproduce in this minimal case; this does not establish the customer project's cause and does not close the remaining Python, ToDoubleNN, concurrency or workspace cases.

Two remaining concrete findings: (1) the field's public response still labels no_use_reference_channel although it returns four resolved member_access uses; the disclosure needs separate review, (2) XAML Loaded="rbPosSchalter_Loaded" emits no call reference (the same-named C# method-group use is present). The XAML extractor's closed handler list currently covers Click and four keyboard events, not Loaded. Neither finding is included as a fixed behavior in the pending 0.41.47 core batch.

Local reproduction and raw public responses: /tmp/cosi-047-ticket-448-audit/verification.json; runner /tmp/cosi-047-audit-448.py. No customer project was modified.

Confirmed event fixes published: XAML 0.3.4 (PR #511)

PR #511 is merged after all four required workflows passed on b6aeeb7f0a65276a43af5fe2ecd6cac00a068ce3. XAML 0.3.4 is published, and all four package-release jobs passed (run 2510). Digest: sha256:0fb1f6e88701c31a270964f1b230349e4801a49231fb674172952744b96d123b.

The shared C#/XAML regression reproduces and fixes four missing lifecycle handlers (Loaded/Unloaded, root and attached child) and a false bind from NotClick. It checks both endpoint grants, each revoked endpoint, ambiguous paired targets, foreign-file decoys and scoped/full equivalence. Shipped conformance covers eleven positive references and seven negative controls. The complete mandatory release tier passed.

A fresh download test verified the signed public v2 catalog using the previously trusted 0.41.46 CLI, downloaded signed C# 0.1.11 and XAML 0.3.4, and indexed a namespaced root-window example. Root KeyDown, child Click and the XAML Loaded handler resolve to paired C# methods; all four x:Name uses resolve; the C# method-group reference resolves too. Zero parse errors. Public find_references responses are retained at /tmp/cosi-047-xaml034-published/verification.json. No customer machine was changed.

The broader issue remains open: the inconsistent no_use_reference_channel disclosure, Python importer cases, ToDoubleNN, concurrency and workspace evidence still require separate fixes or reproduction. The core 0.41.47 batch separately closes #433, #447 and #305; it does not close this report.

Two-part local-ai-server/W100 field assessment received after .17, with a hand-reviewed 20-task comparison. Treat observed query results separately from source-only hypotheses. Confirmed event fixes are recorded below; raw tool arguments/responses/source spans are still needed for the remaining cases. Highest-priority observed cases: - XAML root event KeyDown="CustomWindow_KeyDown" on tl:CustomWindow in wndPosErfassungSchalter.xaml:10 returns no references, while child Click handlers resolve. - x:Name rbPosSchalter has four code-behind uses (26/34/54/81), but returns zero; search_symbols claims measured fallback zero while file_outline reports no_use_reference_channel for the same id. Dataset→C# references similarly appear as measured zero. - C# method groups `new RoutedEventHandler(rbPosSchalter_Loaded)` are missing. Four ToDoubleNN calls in busMdeGenerator.cs:635–636 reportedly absent despite no extraction diagnostics. Need exact source before classifying. - Python get_dependencies(direction=in) for llm/models.py reportedly returns924 unrelated *.models importers (1correct in first200). Explicit imported _attachment_block at agent.py:1300 remains unresolved despite import at56. - Workspace index_state from another root is attached to primary-only answers; limit clamps and negated-glob/path-only search semantics need raw payload verification. Source-only hypotheses (UNVERIFIED): daemon load-modify-save approval paths in offer.rs/bundled.rs may bypass the CLI project_mutation_lock; background updater timeout/cooldown/log edges. Reproduce a daemon+CLI schedule before claiming approval loss. Never weaken publication/authority fencing. Operational findings already tracked: coarse host fingerprint / expensive rebuilds #410; stale last-check package_updates record acknowledged on #442; C# imported static calls #445, generic method arguments #446 (candidate fix), JS→TS imports #447, extension calls #313, member proof #69, Svelte #433/#434. The sample is not an overall completeness claim: Python187TP/1FP/7misses over194ground-truth items; C#/XAML4correct+2partial tasks, with stronger bridge precision but several uncertain zeros. Keep precision and recall separate. Next steps: obtain raw witnesses, create independent minimal reproductions, then split confirmed fixes into focused issues/PRs. Root XAML events and inconsistent zero disclosures rank above cosmetic output changes. Raw partner reports retained at /tmp/cosi-04117-local-ai-assessment-raw.jsonl; actual item IDs01M5Q6NRAA232Q67BN4XSXJ8F9 and01M5Q6P9HJBTBNS08Y36V4E73Y. No partner machine or project state was changed. ## Revalidation on published 0.41.46 (2026-11-09) Fresh minimal reproduction using signed C# 0.1.11 and XAML 0.3.3: namespaced `tl:CustomWindow` root KeyDown resolves to the paired C# method, child Click resolves, all four code-behind uses of `x:Name="rbPosSchalter"` resolve and public find_references returns four rows, and `new RoutedEventHandler(rbPosSchalter_Loaded)` emits and resolves a C# method-group reference. Zero parse errors. These particular missing-reference claims no longer reproduce in this minimal case; this does not establish the customer project's cause and does not close the remaining Python, ToDoubleNN, concurrency or workspace cases. Two remaining concrete findings: (1) the field's public response still labels `no_use_reference_channel` although it returns four resolved member_access uses; the disclosure needs separate review, (2) XAML `Loaded="rbPosSchalter_Loaded"` emits no call reference (the same-named C# method-group use is present). The XAML extractor's closed handler list currently covers Click and four keyboard events, not Loaded. Neither finding is included as a fixed behavior in the pending 0.41.47 core batch. Local reproduction and raw public responses: `/tmp/cosi-047-ticket-448-audit/verification.json`; runner `/tmp/cosi-047-audit-448.py`. No customer project was modified. ### Confirmed event fixes published: XAML 0.3.4 (PR #511) [PR #511](https://git.h-dv.de/h-dv/code-index/pulls/511) is merged after all four required workflows passed on `b6aeeb7f0a65276a43af5fe2ecd6cac00a068ce3`. [XAML 0.3.4](https://git.h-dv.de/h-dv/code-index/releases/tag/pkg/de.h-dv.xaml/v0.3.4) is published, and all four package-release jobs passed ([run 2510](https://git.h-dv.de/h-dv/code-index/actions/runs/2510)). Digest: `sha256:0fb1f6e88701c31a270964f1b230349e4801a49231fb674172952744b96d123b`. The shared C#/XAML regression reproduces and fixes four missing lifecycle handlers (`Loaded`/`Unloaded`, root and attached child) and a false bind from `NotClick`. It checks both endpoint grants, each revoked endpoint, ambiguous paired targets, foreign-file decoys and scoped/full equivalence. Shipped conformance covers eleven positive references and seven negative controls. The complete mandatory release tier passed. A fresh download test verified the signed public v2 catalog using the previously trusted 0.41.46 CLI, downloaded signed C# 0.1.11 and XAML 0.3.4, and indexed a namespaced root-window example. Root KeyDown, child Click and the XAML Loaded handler resolve to paired C# methods; all four x:Name uses resolve; the C# method-group reference resolves too. Zero parse errors. Public find_references responses are retained at `/tmp/cosi-047-xaml034-published/verification.json`. No customer machine was changed. The broader issue remains open: the inconsistent `no_use_reference_channel` disclosure, Python importer cases, ToDoubleNN, concurrency and workspace evidence still require separate fixes or reproduction. The core 0.41.47 batch separately closes #433, #447 and #305; it does not close this report.

buildagent commented

2026-10-24 17:06:07 +02:00

Author

Member

Reporter reran four witnesses on matched .17 and corrected one claim (A1X01M5Q71QYTRV0EHTFKJ6A32XWT):

Root XAML KeyDown confirmed: private code-behind handler id1059725 at44, framework same-name handler id1317827 at116; find_references(symbol_id1059725,concise) returns total=total_resolved=0, while child Click bridges work.
rbPosSchalter measured zero confirmed, but the claimed search_symbols/file_outline disagreement is RETRACTED: both return field id1059715, ref_count0,name_fallback_count0. Four code-behind uses at26/34/54/81 and XAML ElementName binding at11 remain invisible.
busMdeGenerator.cs635–636: r.Down.ToDoubleNN() three times and r.bab_menge.ToDoubleNN() once; current/diagnostics-empty coverage, text search sees both lines, fully paginated references for all three overload ids omit this file. Whether extraction or fallback shape filtering is responsible remains unverified.
get_dependencies({target:'src/ai_runtime_service/llm/models.py',direction:'in',limit:5}) returns924, beginning with unrelated core.applications.packages.models.Distribution and core.config.models.*. Reporter ground truth61 files importing ai_runtime_service.llm.models.

Independent Linux minimal reproductions are the next step. Keep the corrected tool-agreement claim out of subsequent issue descriptions.

Reporter reran four witnesses on matched .17 and corrected one claim (A1X01M5Q71QYTRV0EHTFKJ6A32XWT): - Root XAML KeyDown confirmed: private code-behind handler id1059725 at44, framework same-name handler id1317827 at116; find_references(symbol_id1059725,concise) returns total=total_resolved=0, while child Click bridges work. - rbPosSchalter measured zero confirmed, but the claimed search_symbols/file_outline disagreement is RETRACTED: both return field id1059715, ref_count0,name_fallback_count0. Four code-behind uses at26/34/54/81 and XAML ElementName binding at11 remain invisible. - busMdeGenerator.cs635–636: r.Down.ToDoubleNN() three times and r.bab_menge.ToDoubleNN() once; current/diagnostics-empty coverage, text search sees both lines, fully paginated references for all three overload ids omit this file. Whether extraction or fallback shape filtering is responsible remains unverified. - get_dependencies({target:'src/ai_runtime_service/llm/models.py',direction:'in',limit:5}) returns924, beginning with unrelated core.applications.packages.models.Distribution and core.config.models.*. Reporter ground truth61 files importing ai_runtime_service.llm.models. Independent Linux minimal reproductions are the next step. Keep the corrected tool-agreement claim out of subsequent issue descriptions.

Rows
Columns

v0.41.17 field assessment: false-zero disclosures, import precision and daemon approval concurrency need reproduction #448

Revalidation on published 0.41.46 (2026-11-09)

Confirmed event fixes published: XAML 0.3.4 (PR #511)