Files
ARCADIA-Studio/harness/TASK_LEDGER.md
T

7.3 KiB

Task Ledger — first release

Updated 2026-09-11 for Codex-ytr83r. The full user goal and current progress remain authoritative. Exact host ledger archives preserve prior facts, dates and reasoning without transferring their acceptance.

Goal and given requirements

Deliver a complete, usable OA/SA release on Windows and Linux, with one explicit reST semantic contract and equivalent graphical, textual and REST access. Include constraints/calculations, quantities/units, maintained libraries, behavior/state, traceability and navigable offline help. Reduce cognitive load and visual clutter; small and large models must be practical using text alone. Later Arcadia layers remain future scope. Compilation or a pleasant happy path is not completion.

Verified facts

  • Local start was cleanff615d0. Root decoded/hash-verified/imported published3978, c6 and4bb bundles; merged c6 as c81e285 and integrated creation4bb at e207b04. There were no production cherry-pick conflicts; docs/attributes require evidence union. The current goal is copied from the user's latest file.
  • Read-only integration comparison finds equivalent workspace/Sensor code in c6; replaying3978 or d114 is unnecessary. Core source is unchanged betweenff andc6. B4 now has a production Scene bridge and independently reviewed source, with compilation/native acceptance pending. Audit.
  • Imported source implements named behavior creation, exact prepared-batch correlation, success-only selection, raw-draft guards, recoverable original layer and modal-priority Escape. These are source facts; combined local runtime acceptance is pending.
  • Earlier-host desktop642 and653 suites, scoped native scenarios/creation, packaging25 notices and source experiments retain exact attribution in STATE.md. They are not new local passes.
  • The original transferred creation oracle incorrectly equates physical disk files with reachable sources during Undo. Later external v2 corrects the evaluator with explicit origin/retained-byte rules; exact v2/evidence2ec was imported at b853a1e after complete path/hash preflight. New host execution is still pending.
  • Isolated Rust1.98.1/LLVM23.1.1 and locked dependencies are available. Packaging 11/11 and synthetic Windows notice checks pass; full core all-targets currently compiles under one Cargo job. Strict reST/HTML generation passes with Docutils0.22.2.
  • B4 independent criticism found and resolved self-caption overlap and two dense minimum-Fit hit-selection counterexamples. Shared nearest marker/message-stroke arbitration preserves exact marker/header/label selection. Ten real-Scene tests and a separate fixed-ID twelve-scenario native oracle await execution.
  • The separate baseline-context experiment passed218 assertions; explicit one-level overrides reduce repeated stable assumptions while complete frozen definition environments remain necessary. Proposed syntax is not source7.

Assumptions and unresolved questions

  • Combining already coherent desktop slices should preserve behavior; fresh build, automated and native gates must test that assumption.
  • The baseline/library designs can fit existing atomic Project/Command/source boundaries; exact references, deletion, transition, imports, freezing and numeric rules still need ratified contracts and falsification.
  • Pinned package provenance must be distinguished from checked type compatibility; unchanged quantity kinds cannot become incompatible merely through cosmetic repackaging without a deliberate policy. Energy and Torque remain distinct.
  • Which supported environment will complete current Linux GUI/service/package execution remains a release question. Old failed or denied experiments do not establish a current successful environment.
  • Large-model human/AI task usability, cross-file refactoring, actual chooser latency/keyboard/minimum geometry and complete native lifecycle behavior remain to be measured. Parser admission or a widget experiment is insufficient.

Candidate designs and selected directions

The imported analysis proposal, library proposal, paired critique and textual authoring design retain alternatives and counterexamples. Metadata-only analysis and bare source includes do not meet evaluated quantities or maintained libraries.

Codex-956atf root selected property-owned baseline values/formulas with explicit ModelBaseline/Case evaluation and one-level case overrides for contract refinement. No hidden default case or UI-dependent evaluation; inherited formulas use the effective requested case, and complete freeze captures definitions/provenance. Exact numeric, library normalization/type compatibility and transition details remain open. This direction is recorded in DECISIONS.md, not silently treated as completed production semantics.

The external staged prerequisite design reserves source/consumer identities, keeps the parent context/rawtext/slot, stages locally and creates source+consumer in one atomic final command. It must not implicitly allocate work, involve capabilities, retarget an Exchange-fixed endpoint or accept a different editor. Qualified Action allocation remains a distinct engineering choice.

Retained implementation decisions

  • Keep one authoritative typed Project, validated atomic Commands, stable identities and explicit references. Semantic ownership/allocation/order and diagram geometry remain distinct; forms are adapters, not a second metamodel.
  • Preserve source ownership/comments/includes and exact model/source history. Writeback succeeds before adoption. Invalid/rejected raw drafts survive and browsing/tool selection does not silently accept another editor. Recovery validates/replans rather than trusting serialized commands.
  • Semantic reconciliation uses finite typed facets, frozen baselines, explicit mappings and one atomic candidate. Names alone never establish identity.
  • OA↔SA transitions are explicit engineering decisions. New persistent semantics need validation/reference/deletion/transition/source/API/UI/history parity.
  • Exact evidence bytes and host attribution survive integration. No focused test total is added to a full-suite count; old binary acceptance does not transfer to newly merged code.

Current strategy and coordination

Root owns local integration, serialized builds, native inputs and final synthesis. The local audit resolves evidence and reviews source; independent server review has its own report. Codex-9e8r6t owns staged sources/first actions and source CLI; Codex-956atf owns chooser and semantic contracts. Read Gitea issue #1/Wiki before resuming, obtain exact coherent transfer heads, and respect their source ownership.

First finish and verify the integrated current source using the imported corrected evaluator and newly prepared independent lifetime oracle. Then close remaining modeling/textual, analysis/library and platform gates with independent criticism and whole-task acceptance. Update facts, assumptions and decisions as evidence changes them.