pbakaus_impeccable/site/data/live-harnesses.ts

export const harnessPaths = [
  {
    name: 'Foreground long-poll',
    harness: 'Codex desktop · observed',
    pickup: 'Immediate when the turn is blocked',
    tradeoff: 'Reliable, but Live occupies the main agent and inherits its model speed.',
    status: 'Baseline',
    statusTone: 'ready',
  },
  {
    name: 'Background terminal + watcher',
    harness: 'Codex app-server · available primitives',
    pickup: 'Not yet measured end-to-end',
    tradeoff: 'The app-server exposes output notifications and fs.watch, but a watcher still has to wake model work quickly.',
    status: 'Benchmark next',
    statusTone: 'measure',
    source: 'https://learn.chatgpt.com/docs/app-server#api-overview',
  },
  {
    name: 'Warm variant producer',
    harness: 'Codex custom agent',
    pickup: 'Spawn/resume cost unknown',
    tradeoff: 'Can pin a faster model and lower reasoning, but fresh context and handoff can erase the gain unless the agent stays warm.',
    status: 'Prototype',
    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: 'Dispatch first, capture second — shipped',
    evidence: 'Plain click-handler → generate fetch is now 2.2 ms median; end-to-end model-free latency fell from 916 ms to 414 ms.',
    move: 'Unannotated picks wait for the helper to accept the event, then capture the shader off-path. Annotated picks still capture and upload before dispatch.',
    expected: 'Delivered a 54.8% median reduction on the same fixture and deterministic agent.',
    confidence: 'Measured + shipped',
  },
  {
    rank: 2,
    title: 'Reveal variants progressively',
    evidence: 'First and all variants arrive together today; the measured delivery gap is only the observer settling time.',
    move: 'Give each variant its own ready event or preview file. Reveal variant 1 as soon as it validates while the remaining variants continue.',
    expected: 'Large perceived win whenever generation is the dominant stage.',
    confidence: 'Protocol change',
  },
  {
    rank: 3,
    title: 'Keep a harness-native producer warm',
    evidence: 'Codex and Claude can assign a different model to a custom subagent; both support continuing agent state.',
    move: 'Resume one narrow Live producer with a compact contract, low reasoning, and source-write tools only. Quality-gate its output before preview.',
    expected: 'Reduces model latency without forcing the user to change the main conversation model.',
    confidence: 'Needs spawn/resume benchmark',
  },
  {
    rank: 4,
    title: 'Move context work into the selection dwell',
    evidence: 'The existing prefetch event was disabled because quick Go clicks paid an extra harness round trip.',
    move: 'Debounce selection, then resolve the source file, style mode, token summary, and identity lock locally. Cancel or reuse the result on Go without queueing a model turn.',
    expected: 'Shortens scaffold and prompt preparation while keeping quick clicks cheap.',
    confidence: 'Low-risk prototype',
  },
  {
    rank: 5,
    title: 'Generate knobs after pixels',
    evidence: 'Parameter manifests are authored in the same response as every variant even though the first job is visual comparison.',
    move: 'Render validated variants first. Infer coarse CSS knobs locally or ask the producer for parameters only for the visible/selected variant.',
    expected: 'Cuts output tokens and lets the first useful preview land sooner.',
    confidence: 'Quality experiment',
  },
  {
    rank: 6,
    title: 'Parallelize variant ideas, centralize writes',
    evidence: 'Modern harnesses can run several specialized agents concurrently, but shared-file edits create coordination risk.',
    move: 'Have workers return one structured variant each; one coordinator validates and performs the only source write. Stream the first valid result.',
    expected: 'Lower time to first option and stronger diversity at higher token cost.',
    confidence: 'Expensive experiment',
  },
  {
    rank: 7,
    title: 'Trim and cache the producer contract',
    evidence: 'The full Live reference is large and stable; only a small slice is needed for one generate event.',
    move: 'Compile a provider-specific generation contract, keep the stable prefix cacheable, and attach only the action reference plus picked-element context.',
    expected: 'Lower prefill latency and cost, especially for fresh subagents.',
    confidence: 'Harness-specific',
  },
  {
    rank: 8,
    title: 'Wake on the journal, not terminal stdout',
    evidence: 'Codex app-server exposes fs.watch and process output notifications, while this desktop probe still required an explicit terminal read.',
    move: 'Benchmark a plugin/app-server bridge that watches the durable Live journal and starts or steers a dedicated turn directly.',
    expected: 'Frees the main turn if watcher-to-turn startup beats foreground polling.',
    confidence: 'Architectural bet',
  },
];

export const currentHarnessProbe = {
  testedAt: '2026-07-11',
  surface: 'Codex desktop unified exec',
  delayedOutputMs: 3000,
  processExitMs: 4000,
  surfacedAutomatically: false,
  result: 'The sentinel appeared only after an explicit session read; no app terminal was attached.',
};