Files
yoi/docs/report/T-743-standalone-correction.md

8.6 KiB

T-743 standalone source correction

Accepted-plan correction and source boundary

The user withdrew the instruction to compose/wait for T-748. T-743 owns its public UI/API/Tools/Prompt work and its own independent review/fix loop against the existing develop contracts. Platform composition and target integration are Orchestrator work. T-748 implementation or approval is not a T-743 return gate. This does not reduce any T-743 product requirement or authorize ignoring T-743-specific authorization, concurrency or public-operation defects.

Observed develop is 56795fb92cb9def9bcd701bd51d92e99526f8eff; a normal fetch confirmed no movement from the original T-743 base. Standalone implementation 9b0d42fdbc2920736c3418bb894f26140e962bbc adds the required public contracts and semantic tests on that base. Published merge 5c5727e3f8ad85bddebda2672fd245a24317fc5f had added unrelated T-748 platform and dependent adapters at the withdrawn instruction. Corrective commit c6d73e82 uses a normal mainline revert of that merge, not reset, rebase, force push or a replacement MR.

Before retaining historical report bodies, the corrective tree was compared and matched 9b0d42fd exactly. After retaining those reports, git diff --quiet 9b0d42fd -- crates web tools resources openapi still passed. Thus all source, wire/generated contracts, UI/Tools/Prompt and semantic tests from the initial standalone result survive without changes or weakened assertions.

Removed only the composition's shared DB/Host/cleanup execution owners, broad async-platform and lifecycle caller changes, Workdir signing adapter, and the new assignment/query execution adapters/test accessors and fixtures requiring that removed owner. No incomplete executor facade or platform API was copied. The baseline's synchronous execution placement is distinguished from T-743 product defects; it is not an excuse to overlook a newly introduced defect. The implementation is separable: no T-743 product API, author/source fence or receipt requires a T-748-only type.

Kept T-743's authenticated source attribution, current singleton/live binding fence including replay, whole IMMEDIATE answer source/CAS/body/audit/receipt/ commit transaction, committed-input delivery scan/deduplication, sanitized assignment diagnostics, retained history, typed SQL-before-page filters, canonical/Browser routes and Normal Tools/WIP transport. Kept case/handling/ Worker/MR separation, same-assignment answer versus new-row reissue, stable tags and closure history, effective Prompt/Profile/Flow and their original tests.

Evidence and review ownership

Historical report bodies (T-743-public-surface-validation.md, T-743-T748-callers.md, T-743-T748-composition.md), published commits and MR/ Ticket events remain evidence of what was done then. Their old composition-wait instructions are superseded, not retroactively rewritten. T-748-specific reports and tests remain readable in the published merge's Git history, but are not current T-743 source contracts. This document carries the current correction and new validation evidence.

Fresh trusted independent review on MR 01a12d19-b61b-79a2-baca-89241fd3c3d3 judges the actual T-743 Ticket requirements, authorization, concurrency and public operations against the exact normally published corrected source. The review thread should include this correction as an append-only event, not replace its original summary. The Coder tool set has no MR-comment mutation tool; OpenMergeRequest on this repository would conflict with the existing MR, not append a comment. The correction is therefore supplied for the trusted review event body, without bypassing typed authority. No target integration, Ticket/MR conclusion, live update or retained Worker/ Workdir cleanup is authorized or performed.

Revalidation

These are fresh results on the corrected standalone implementation, not the historical T-748-composed suite. Only report notices/evidence changed afterward; implementation equality with 9b0d42fd was checked again before publication.

  • Root cargo check passed, including normal TUI/Server compile closure; final repeat passed. Logs: target/t743-standalone-{check,final-check}.log.
  • Public assignment/source/replay/filter focused tests: 28 passed.
  • cargo test --no-fail-fast -p ticket -p worker-runtime -p worker -p yoi-workspace-server -p server-api passed all selected targets with normal default parallelism. Library counts: server-api 98, ticket 99, worker 994, worker-runtime 417, Server 946 with 1 pre-existing ignored (176.76s); Server binary 11, other integration/binary/doc targets passed. Log: target/t743-standalone-crates.log.
  • The full Worker suite includes public Tool/WIP forwarding, source/target distinction, exact-retry and effective research-return/confirmation-wait/ separate-close Prompt fixtures. These are scripted semantic tests, not a measurement of real model judgment. Extra manifest --lib builtin_profile 6 passed and worker --lib flow 17 passed. Initial mistyped filters matched zero tests and are not counted as proof.
  • Fresh Web check 0 errors/0 warnings and production build passed; existing chunk-size advisory remains. Web 613, component 318 and Chromium Ticket/principal 7 tests passed. Logs: target/t743-standalone-{web-check,web-build,web-tests,components,browser}.log.
  • All six TypeScript generator --check commands passed: ticket, repository OpenAPI, runtime, workdir, Worker launch and companion.
  • cargo fmt --all -- --check, git diff --check HEAD and the required env -u LD_LIBRARY_PATH deno run --allow-read --allow-run=git tools/check-state-contract-names.ts passed: 558 matching lines, 0 unapproved, 0 stale exceptions.

Recovered visual evidence and parent review

The visual child was stopped by an interrupted turn, not a successful final handoff. The parent inspected its committed session and recovered six completed capture bundles under target/t743-standalone-visual, then personally used ViewImage on actual images. No missing child verdict was invented, and no child was restarted just to wait for a report.

Fresh bundles at source c6d73e8286d743e0329154d49b52d4bee53e8d60:

  • c6d73e82-tickets / tickets-dark: 63 captures each, covering detail, list/create/current closure filtering, Orchestrator, open+done, waiting, running/stopped Worker, repository/MR/no-MR, read-only/eligibility, closed, long, empty, error and unknown states at 1440/768/390, light/dark.
  • c6d73e82-ticket-principals / ticket-principals-dark: 9 each, explicit agent/account/Worker selection at all three widths.
  • c6d73e82-related-light / related-dark: 21 each, explicitly scrolled capture points for retained result/release history, case history, returned handling, stopped/running Worker, action eligibility and MR separation. This closes the inner-scroll gap: browser full-page capture alone does not expose lower detail sections.

Total 186 captures. Metadata records the actual source SHA and dirty:true for the then-uncommitted report notices, not hidden implementation changes. Parent rechecked the implementation equality, source/viewport/route matrix, per-capture errors and runner-owned fixture shutdown records. Parent actual image review included list desktop/mobile and Orchestrator narrow/mobile, principal light/mobile and dark/desktop sheets, waiting and returned handling, retained result/release mobile and case history narrow, stopped Worker, MR separation, mobile error and long list. Labels/actions remain readable; case state, handling, actual Worker and MR sections remain distinct; no new overflow or visual fix was needed. Original before/after visual evidence remains retained for comparison.

Both Ticket scenario bundles deliberately exit 2 and are completed-with-errors: only the error route at the three widths records its fixture query HTTP503, console/request-abort and visible error/Retry evidence. The other four bundles exit 0, and no other route has capture errors. The parent saved the exact error audit in target/t743-standalone-visual/parent-errors-audit.txt. Do not describe the whole scenario as error-free or the intentional 503 as a product regression. The synthetic member/anonymous transport is not proof of live anonymous permission; actual cookie Origin/signed Runtime/source denial is exercised by Server tests.

All reported fixture processes had owner-scoped shutdown before the child was stopped; this is fixture teardown, not live Yoi/retained Worker cleanup. This publication is ready for its own fresh independent T-743 review, not an approval, model-quality attestation, completion of T-744 live paths or target integration.