Skip to content

Production-readiness partials: review fixes (tickets #255–#264) ​

Ticket #255 review follow-up. This record covers the two-axis review fixes on ticket/255-review-fixes; product behavior is unchanged beyond the harness, test, and policy corrections listed here.

Findings and fixes ​

  1. Runtime test seam. scripts/live-matrix.sh no longer reads ENSEMBLE_LIVE_SCENARIO_ROOT. Scenario selection is exercised through the public fake/all CLI, and the retry/quarantine policy moved to the pure src/v2/live-matrix.ts module tested directly. The stub-scenario tests were removed.
  2. Constant self-assertions. WORK_PANE_MAX_WIDTH === 53 and WCAG_AA_TEXT === 4.5 / WCAG_AA_UI === 3 were removed. The behavioral checks stay: 49/48 card-mode boundaries, the exact right edge, and the rendered contrast pairs.
  3. Retry literal. The shell no longer restates attempt < 2; the retry ceiling comes from FLAKE_BUDGET.maxRetriesPerSuite through decideFailedAttempt.
  4. Wizard sleeps. The memory add/edit/advance block uses live_require_ready_text, live_require_ready_absence, and live_require_quiescent instead of fixed sleeps. The alignment block was already sleep-free.
  5. Compatibility artifacts. .github/workflows/compatibility.yml sets ENSEMBLE_LIVE_ARTIFACTS for the fake smoke and uploads the harness panes/logs with if: failure(). checkCompatibilityWorkflow guards both the directory and the failure upload.
  6. Header scope. The group header assertion now reads only the contiguous surface-title block at the top of the chat pane. A positive wrapped-title fixture and two negative fixtures prove a transcript mention cannot satisfy a missing leader or leader agent.
  7. Quarantine ordering. Classification comes first: only timing may retry or quarantine. Product, unknown, and infrastructure failures always fail, and a quarantine is reported as its own outcome, never a pass.
  8. Latency geometry. The patch-card fixture renders at the real card content width derived from workspacePaneWidths and workCardContentWidth, not a 49-cell card the TUI never draws.

Judgement calls ​

  • Kept one ASCII glyph map (src/v2/tui/ascii.ts) consumed by motion, Work rows, chrome, and init rows.
  • Kept one workflowValue reader (src/v2/ci-workflows.ts) shared with scripts/docs-check.mjs.
  • Kept one Work-header summary/page grammar for the full and compact forms.
  • Removed TESTED_NODE_LINES; docs:check now derives the tested Node claim from node-matrix.yml.
  • Removed the identity METRIC_LABEL map.
  • Replaced the accumulateCompletions(..., undefined, process.env) call with an options object.
  • Removed retryDecision's unused quarantined option.
  • Kept the ci-workflows.ts job-graph checks and the docs-check package/matrix checks: #261 owns the CI cost split and failure-artifact contract, and #262 owns the docs/version contract. The rationale is recorded in the module comments and in the ticket comments for #261/#262.

Verification ​

CommandResult
npm run verifyPass — 2030 tests, 0 failures; tsc --noEmit clean
npm run docs:checkPass — contract v7; OpenCode 2.2.6, 2.2.7; Node >=22.23.4; live Node 26.12.1
npm run test:live:latencyPass — rail p95 11 ms / 500 ms; patch-card p95 36 ms / 5000 ms; first-token report-only
ENSEMBLE_LIVE_REQUIRE=1 npm run test:live:fake:allPass — fake 412 s, fake-empty 10 s; 0 retries, flake budget ok
git merge integration/prod-partialsClean (see status after the merge)

The latency and fake suites ran on the reference host with the isolated OpenCode 2.2.7 host. No conversation content, credentials, file bytes, or private paths are recorded here.

Released under the MIT License.