wsr

Java E2E implementation worker

Status: paused. The owner stop recorded in the current checkpoint supersedes the historical authorization below. This document preserves the worker contract; it does not authorize resumption. Before product work, obtain explicit resumption and reconcile the bounded behavior, architecture and external worker prompt. Documentation cleanup does not change or restart any schedule.

flowchart LR
    F[Pinned reference] --> N[Native pass and failure]
    N --> T[One WSR behavioral slice]
    T --> E[Scoped validation and handoff]

Architecture alignment after the documentation refactor

Follow planning, provisioning, cache, engine and evidence. Native reference setup may use independently qualified disposable environments; this does not select Docker or a VM product backend. Preserve the tutorial’s actual versions/commands when claiming tutorial equivalence; any different native fixture/toolchain must be labelled as an adaptation. User stop/hold instructions in the owner session take precedence over automated implementation. The saved automation prompt is external to this document and has not been edited by the refactor.

Historical authorization and scope

On November 1, 2026 the owner authorized replacing the scheduled test-preparation task with a Java E2E TDD worker. This is a bounded exception to the previous design-only implementation gate. The worker may prepare fixtures, add meaningful tests, implement product behavior with settled contracts, run validation, and maintain evidence. It must not silently select unresolved runtime, policy, isolation, canonical-IR, or plugin interfaces. Routine implementation details within an accepted contract can be resolved without repeated permission requests.

Work directly in /Users/carlosferreyra/Development/ectorial/wsr, in local mode. Preserve the existing DX/design changes and prototype stash. Do not reset, clean, overwrite, blindly switch to main, or apply the stash. Resume task-owned work and preserve unrelated changes. No Git worktrees, new repositories, publication, releases, deployments, automatic merges, or product-code pushes/PRs are authorized by this worker. Keep product changes local and reviewable. Existing authorization for test/documentation PRs is unchanged, but do not disguise product changes as test-only work.

Target and reference hierarchy

Use Java DX and its YAML as the intended experience, not execution proof. Follow PLAN.md, the GHA contract, and the TDD contract. Upstream action/tool semantics and observed native behavior establish reference outcomes; synthetic tutorial logs do not. Do not implement an output imitation or special-case fixture names in product code.

Start with one small Maven/Java project using GitHub Actions. Select and record an appropriately licensed, pinned source commit and workflow with checkout, Java setup and wrapper-based verification. Prefer public dependencies and a small deterministic test suite; exclude publication/deployment jobs. Inspect source, build scripts, actions and prerequisites before execution. Use a minimal owned reproduction when reducing a bug, while retaining the original case’s identity and provenance. Do not choose a new project every run. No Gradle/other-language expansion until this milestone passes.

Pin action revisions and record JDK, wrapper, provider, runtime, image, architecture and dependency inputs. Explain any workflow adaptation; a modified example is not unchanged upstream compatibility. Keep canonical example projects/workflows in existing repository conventions. Link DX copies or verify exact agreement rather than maintaining divergent versions.

Execution and environment

Fetching/inspecting the selected public source and dependencies is permitted for explicit worker preparation. Ordinary Cargo tests must not fetch or execute upstream workloads implicitly. Use ignored target/wsr-tests/java/ for disposable sources and raw logs, never personal directories. Use an independently isolated Linux reference environment with narrow mounts, no ambient personal credentials, resource limits and cleanup. WSR’s unimplemented sandbox is not that reference boundary. If no approved environment exists, prepare/review inputs and report the specific prerequisite; do not run unreviewed upstream build scripts directly on the Mac. Use the repository’s existing managed tooling configuration. Do not dispatch hosted workflows or provision paid services automatically. A local Maven command is a native tool reference, not a complete native GHA workflow execution.

Per-run TDD loop

  1. Read progress, live branch/diff, coverage index and any pending task. Acquire exclusive checkout ownership using an atomic run lock under target/wsr-tests/; record owner, process/run identity and start time. If another live writer owns it, do not mutate. Recover a stale lock only after verifying its owner is no longer active. Release ownership on completion/failure.
  2. Resume an unfinished bounded behavior before selecting another. Describe behavior, accepted interface, fixture, exact command, intended failure, non-goals and completion criteria before code.
  3. Establish/refresh native reference success and controlled failure when relevant inputs change. Missing tools, downloads or compilation failures are setup failures, not product red evidence.
  4. Observe a meaningful failing test through the actual owning API/CLI. Implement the minimum general behavior, observe green, then refactor within scope. Do not ignore tests, weaken assertions, accept generic placeholder output as sufficient red, or write a speculative disabled suite.
  5. Run focused tests, relevant regressions and git diff --check. Before claiming a completed slice, run cargo xtask ci; report environmental failures and unexecuted lanes honestly.
  6. Update progress and link behavior/tests/results in docs/testing/coverage.md; update implementation claims only when proven. Preserve the exact next command and branch for interrupted work.

One bounded behavior per invocation is the default, not a requirement to manufacture a commit. GPT-6 Astra (gpt-6-astra) orchestrates the worker. For actionable work, spawn bounded subagents with explicit gpt-6.1-sol or gpt-6.1-terra models and fork_turns=none; supply self-contained contracts, files, commands and acceptance criteria. Sol handles implementation/tests; Sol or Terra can review or handle scoped fixture/documentation work. Astra selects the slice, coordinates TDD, reviews evidence and integrates. Keep one writer per file/slice and parallelize only independent work. Preserve default reasoning effort. Record actual agent/model assignments; report unavailable requested models rather than silently substituting. No second agent may write conflicting files. Preserve deliberate red work locally on the identified development branch, never by breaking main or pretending it passed.

Milestones and acceptance

For each stage, settle only the prerequisite contract needed next. If a design decision blocks it, record two concrete options and a recommendation, request the smallest decision, and continue only independent authorized work. Do not repeat unchanged blockers or invent architecture to get green.

Completion means the selected workflow’s real WSR execution, controlled failure, state, cache and cleanup criteria are backed by runnable tests and recorded identities/results. Report completion and perform no further implementation until another milestone is authorized. A task report does not pause/delete its schedule; scheduling changes remain explicit. No live publishing is in this goal.