harn-stdlib 0.10.125

Embedded Harn standard library source catalog
Documentation
/**
 * `harn provider option-probe` empirically binds one portable generation
 * option claim to one concrete provider endpoint.
 *
 * The normal runtime admits and shapes options from the capability catalog.
 * That is correct for product calls and circular evidence for a probe. The
 * Rust dispatch shim therefore selects a typed probe contract for exactly the
 * chosen option. Option extraction carries that contract through spawned
 * transport work. Every unrelated catalog guard remains active.
 */
import { parse, parser } from "std/cli/argparse"
pub import {
  ProviderOptionProbeAttempt,
  ProviderOptionProbeReport,
  ProviderOptionProbeVerdict,
  ProviderPortableProbeOption,
  provider_option_probe_catalog_claim,
  provider_option_probe_route,
} from "std/cli/providers/contracts"
import { provider_portable_option_catalog_field } from "std/cli/providers/contracts"

const __OPTIONS = [
  "temperature",
  "top_p",
  "top_k",
  "seed",
  "frequency_penalty",
  "presence_penalty",
  "stop",
]

fn __probe_verdict(value: string) -> ProviderOptionProbeVerdict {
  if value == "accepted" {
    return "accepted"
  }
  if value == "rejected" {
    return "rejected"
  }
  if value == "inconclusive" {
    return "inconclusive"
  }
  if value == "gated_locally" {
    return "gated_locally"
  }
  throw "invalid provider option probe verdict: ${value}"
}

fn __attempt(
  llm: HarnessLlm,
  provider: string,
  model: string,
  option: ProviderPortableProbeOption,
  max_tokens: int,
) -> unknown {
  const prompt = "Reply with the single word: ok"
  if option == "temperature" {
    return try {
      llm.call(
        prompt,
        nil,
        {provider: provider, model: model, max_tokens: max_tokens, stream: false, temperature: 0.2},
      )
    }
  }
  if option == "top_p" {
    return try {
      llm.call(
        prompt,
        nil,
        {provider: provider, model: model, max_tokens: max_tokens, stream: false, top_p: 0.9},
      )
    }
  }
  if option == "top_k" {
    return try {
      llm.call(
        prompt,
        nil,
        {provider: provider, model: model, max_tokens: max_tokens, stream: false, top_k: 1},
      )
    }
  }
  if option == "seed" {
    return try {
      llm.call(
        prompt,
        nil,
        {provider: provider, model: model, max_tokens: max_tokens, stream: false, seed: 1},
      )
    }
  }
  if option == "frequency_penalty" {
    return try {
      llm.call(
        prompt,
        nil,
        {
          provider: provider,
          model: model,
          max_tokens: max_tokens,
          stream: false,
          frequency_penalty: 0.1,
        },
      )
    }
  }
  if option == "presence_penalty" {
    return try {
      llm.call(
        prompt,
        nil,
        {
          provider: provider,
          model: model,
          max_tokens: max_tokens,
          stream: false,
          presence_penalty: 0.1,
        },
      )
    }
  }
  return try {
    llm.call(
      prompt,
      nil,
      {
        provider: provider,
        model: model,
        max_tokens: max_tokens,
        stream: false,
        stop: ["harn-option-probe-stop"],
      },
    )
  }
}

fn __failure_class(error: unknown) -> string {
  if type_of(error) != "dict" || to_string(error?.origin ?? "") == "local" {
    return "local"
  }
  const kind = to_string(error?.kind ?? "")
  const reason = to_string(error?.reason ?? "")
  if reason == "auth_failure" {
    return "unauthorized"
  }
  if kind == "transient" {
    return "transient"
  }
  if reason == "invalid_request" {
    return "rejected"
  }
  if contains(["model_unavailable", "context_overflow", "invalid_response"], reason) {
    return "unrelated"
  }
  return "unrelated"
}

/**
 * Classifies whether a request reached the provider and whether it accepted the selected option.
 *
 * @effects: []
 * @errors: []
 */
pub fn provider_option_probe_classify(outcome: unknown) -> ProviderOptionProbeAttempt {
  if !is_err(outcome) {
    const usage = unwrap(outcome)?.usage
    return {
      verdict: __probe_verdict("accepted"),
      measured: true,
      request_count: 1,
      input_tokens: to_int(usage?.input_tokens),
      output_tokens: to_int(usage?.output_tokens),
      cost_usd: to_float(usage?.cost_usd),
    }
  }
  const raw = unwrap_err(outcome)
  if type_of(raw) == "dict" && to_string(raw?.reason ?? "") == "empty_generation" {
    return {
      verdict: __probe_verdict("accepted"),
      measured: true,
      request_count: 1,
      reason: "served_empty",
    }
  }
  const message = if type_of(raw) == "dict" {
    to_string(raw?.message ?? raw)
  } else {
    to_string(raw)
  }
  const class = __failure_class(raw)
  if class == "local" {
    return {
      verdict: __probe_verdict("gated_locally"),
      measured: false,
      request_count: 0,
      failure_class: class,
      error: message,
    }
  }
  if class == "rejected" {
    return {
      verdict: __probe_verdict("rejected"),
      measured: true,
      request_count: 1,
      reason: to_string(raw?.reason ?? "invalid_request"),
      error: message,
      cost_usd: 0.0,
    }
  }
  return {
    verdict: __probe_verdict("inconclusive"),
    measured: false,
    request_count: 1,
    failure_class: class,
    reason: if type_of(raw) == "dict" {
      to_string(raw?.reason ?? "")
    } else {
      ""
    },
    error: message,
  }
}

/**
 * Builds the endpoint-keyed claim, observation, and drift receipt for one measured request.
 *
 * @effects: []
 * @errors: []
 */
pub fn provider_option_probe_report(
  provider: string,
  model: string,
  route: string,
  option: ProviderPortableProbeOption,
  claimed: bool,
  ungated: bool,
  attempt: ProviderOptionProbeAttempt,
) -> ProviderOptionProbeReport {
  const observed = if attempt.verdict == "accepted" {
    true
  } else if attempt.verdict == "rejected" {
    false
  } else {
    nil
  }
  const status: "match" | "drift" | "unmeasured" = if observed == nil {
    "unmeasured"
  } else if observed == claimed {
    "match"
  } else {
    "drift"
  }
  return {
    schema_version: "harn.provider_option_probe.v1",
    endpoint: {provider: provider, model: model, route: route},
    option: option,
    catalog: {field: provider_portable_option_catalog_field(option), claimed_supported: claimed},
    probe: {
      ungated: ungated,
      request_count: attempt.request_count,
      measured_count: if attempt.measured {
        1
      } else {
        0
      },
      verdict: attempt.verdict,
      attempt: attempt,
    },
    diff: {status: status, claimed_supported: claimed, observed_supported: observed},
  }
}

fn __render(report: ProviderOptionProbeReport) -> string {
  return report.endpoint.provider + ":" + report.endpoint.model + " " + report.option + " claim="
    + to_string(report.diff.claimed_supported)
    + " observed="
    + to_string(report.diff.observed_supported)
    + " "
    + report.diff.status
}

fn main(harness: Harness) {
  const parsed = parse(
    parser(
      {
        name: "provider_option_probe",
        args: [
          {kind: "flag", name: "provider", long: "--provider", required: true},
          {kind: "flag", name: "model", long: "--model", required: true},
          {kind: "flag", name: "option", long: "--option", required: true},
          {kind: "flag", name: "max_tokens", long: "--max-tokens", parse: "int", default: 8},
          {kind: "switch", name: "plan", long: "--plan"},
          {kind: "switch", name: "ungated", long: "--ungated"},
          {kind: "switch", name: "fail_on_drift", long: "--fail-on-drift"},
        ],
      },
    ),
    argv,
  )
  if is_err(parsed) {
    throw unwrap_err(parsed).message
  }
  const args = unwrap(parsed).options
  const option_text = to_string(args.option)
  if !contains(__OPTIONS, option_text) {
    throw "option-probe: --option must be one of: " + join(__OPTIONS, ", ")
  }
  const option: ProviderPortableProbeOption = option_text
  const provider = to_string(args.provider)
  const model = to_string(args.model)
  const caps = harness.llm.provider_capabilities(provider, model)
  const route = provider_option_probe_route(caps)
  const claimed = provider_option_probe_catalog_claim(caps, option)
  const ungated = args.ungated ?? false
  if args.plan {
    const plan = {
      schema_version: "harn.provider_option_probe_plan.v1",
      endpoint: {provider: provider, model: model, route: route},
      option: option,
      catalog: {field: provider_portable_option_catalog_field(option), claimed_supported: claimed},
      request_count: 1,
      ungated: ungated,
    }
    harness.stdio.println(json_stringify_pretty(plan))
    return
  }
  const outcome = __attempt(harness.llm, provider, model, option, args.max_tokens)
  const report = provider_option_probe_report(
    provider,
    model,
    route,
    option,
    claimed,
    ungated,
    provider_option_probe_classify(outcome),
  )
  if harness.env.get_or("HARN_OUTPUT_JSON", "0") == "1" {
    harness.stdio.println(json_stringify_pretty(report))
  } else {
    harness.stdio.println(__render(report))
  }
  if args.fail_on_drift {
    if report.diff.status == "drift" {
      harness.runtime.exit(1)
    }
    if report.diff.status == "unmeasured" {
      harness.runtime.exit(2)
    }
  }
}