pbakaus_impeccable/site/data/live-harnesses.ts

export const harnessPaths = [
  {
    name: 'Dedicated app-server worker',
    harness: 'Codex · isolated persistent thread',
    pickup: '1.47 ms turn wake',
    tradeoff: 'A separate Live-owned Sol thread generates while the control supervisor keeps polling. Full context protects design quality; cold readiness measured 657 ms and happens during Live initialization.',
    status: 'Codex default · gated',
    statusTone: 'ready',
    source: 'https://learn.chatgpt.com/docs/app-server',
  },
  {
    name: 'Foreground terminal supervisor',
    harness: 'Codex desktop · unified exec session',
    pickup: 'Concurrent shell work proven',
    tradeoff: 'A yielded foreground process keeps running while other commands execute, but new output still needs an explicit session read; it does not wake a finished model turn by itself.',
    status: 'Useful within active turn',
    statusTone: 'measure',
    source: 'https://learn.chatgpt.com/docs/app-server#api-overview',
  },
  {
    name: 'Native generation subagent',
    harness: 'Codex desktop · portable fallback',
    pickup: 'Main turn must dispatch it',
    tradeoff: 'Low-effort compact handoff is safe and fenced, but it cannot start after the desktop task has gone idle.',
    status: 'Current fallback',
    statusTone: 'prototype',
    source: 'https://learn.chatgpt.com/docs/agent-configuration/subagents',
  },
  {
    name: 'Background variant producer',
    harness: 'Claude Code subagent',
    pickup: 'Supported; measure per version',
    tradeoff: 'Model, effort, background execution, and resume are configurable. Preserve the current main-thread path as fallback.',
    status: 'Cross-harness control',
    statusTone: 'control',
    source: 'https://code.claude.com/docs/en/sub-agents',
  },
];

export const liveExperiments = [
  {
    rank: 1,
    title: 'Run a dedicated Codex worker — shipped',
    evidence: 'Matched full tasks reached usable output in 34.9 s warm app-server versus 64.6 s direct, with both paths passing 4/4. Live now returns from prewarm in 87 ms while the worker finishes initializing alongside page startup.',
    move: 'Keep one Live-owned app-server connection and isolated Sol/medium thread. Attach the skill, sub-command, project context, source neighborhood, and annotation image.',
    expected: 'Removes main-turn availability from generation without trading away brand fidelity.',
    confidence: 'Production browser + full-task gate',
  },
  {
    rank: 2,
    title: 'Accept any arrived variant — shipped',
    evidence: 'A verified Vite/React + Sol run committed variant 2 to source, returned the page in 238 ms, and dispatched the next Go in 831 ms. The next first variant arrived in 16.9 s versus the 43.0 s control, a 60.6% reduction.',
    move: 'Durably fence generation on Accept/Discard, rotate the canceled generation thread, prioritize control events, and keep a separate poll supervisor active.',
    expected: 'Removes the full-set wait and keeps the page interactive while canceled work unwinds.',
    confidence: 'Production browser',
  },
  {
    rank: 3,
    title: 'Send every source variant as a fenced delta — shipped',
    evidence: 'Across five Sol/medium samples, first review was 28.1 s median versus the 34.2 s full-source control. Variant 2 followed 24.6 s later versus 43.9 s, and the guarded all-delta run completed in 102.2 s with 2/3 rendered variants passing and 7.7/10 average brand fidelity.',
    move: 'Keep the full skill, plan, project context, and source evidence, but return only each new variant’s markup, scoped CSS, and the final deferred parameter manifest. The supervisor validates and merges each delta without exposing prior output to rewrites.',
    expected: 'Shrinks every model response while preserving full design context and immutable review checkpoints.',
    confidence: 'Production browser + rendered gate',
  },
  {
    rank: 4,
    title: 'Dispatch first, capture second — shipped',
    evidence: 'Plain click-handler → generate fetch is 2.2 ms median; end-to-end model-free latency fell from 916 ms to 414 ms.',
    move: 'Unannotated picks dispatch before off-path shader capture. Annotated picks still upload their visual evidence before dispatch.',
    expected: 'Removes screenshot work from the common critical path.',
    confidence: 'Ten-run browser benchmark',
  },
  {
    rank: 5,
    title: 'Preflight source locally — shipped',
    evidence: 'The production worker resolved and wrapped the selected React source in 52 ms before the model turn.',
    move: 'Attach durable scaffold metadata to the queued event. Every harness reuses it; discovery failures preserve the existing agent-driven fallback.',
    expected: 'Removes one deterministic source-discovery tool round trip from every successful generation.',
    confidence: 'Production browser',
  },
  {
    rank: 6,
    title: 'Generate knobs after pixels — shipped',
    evidence: 'The first progressive write intentionally carries no parameter manifest; tune controls arrive with the complete variant set.',
    move: 'Plan parameter axes with the trio, but defer their manifests and CSS branches until the final delivery edit.',
    expected: 'Keeps parameter output off the first-reviewable critical path without shrinking the final tuning surface.',
    confidence: 'Protocol shipped',
  },
  {
    rank: 7,
    title: 'Acknowledge first, clean up off-path — shipped',
    evidence: 'The production page returned to Pick in 197 ms. Cleanup is durably enqueued before acknowledgement, and replayed Accept events are idempotent after a worker or network failure.',
    move: 'Persist the Accept receipt, enqueue carbonize cleanup before replying, then let the poll supervisor lease new work while cleanup stays source-locked.',
    expected: 'Preserves immediate interaction without leaving temporary Live source behind.',
    confidence: 'Production browser + build control',
  },
  {
    rank: 8,
    title: 'Reject speculative model warmup — decided',
    evidence: 'A prewarmed spare thread made first review slower (35.0 s vs 33.3 s) and the post-Accept next result slower (50.9 s vs 43.0 s). The app-server workload contended instead of hiding latency.',
    move: 'Do not spend model work before a real generation request. Keep cold thread rotation for cancellation safety and optimize the requested output itself.',
    expected: 'Avoids extra tokens, contention, and initialization work that did not improve the user-visible path.',
    confidence: 'Matched production A/B',
  },
  {
    rank: 9,
    title: 'Keep atomic delivery as an opt-in experiment',
    evidence: 'One production Sol/medium run completed all three variants in 40.2 s versus 56.1 s progressive median, but delayed first review by about 18 s and only 1/3 rendered variants passed.',
    move: 'Expose atomic in the benchmark and configuration surface, but do not make it the Codex default until repeated rendered runs clear the same quality gate.',
    expected: 'Preserves a promising full-set path without regressing the default first-review experience.',
    confidence: 'One production run · needs replication',
  },
];

export const currentHarnessProbe = {
  testedAt: '2026-07-13',
  surface: 'Codex desktop unified exec',
  delayedOutputMs: 250,
  processExitMs: null,
  parallelCommandMs: 0.01,
  parallelWorkSucceeded: true,
  backgroundChildSurvivedShellExit: false,
  requiresExplicitSessionRead: true,
  surfacedAutomatically: false,
  result: 'A yielded foreground poller kept producing output while a separate command completed immediately. A traditional shell-backgrounded child did not survive shell exit. Output still required an explicit session read and did not proactively wake a finished turn.',
};