Agent Completions

std/agent/completions provides shared envelopes for inline suggestions, next-edit predictions, and completion telemetry. Hosts still collect editor state and draw UI; Harn owns the reusable shapes so IDEs, TUIs, replayers, and harnesses do not invent incompatible JSON.

Import it with:

import {
  completion_context,
  completion_model_route,
  completion_policy,
  completion_policy_decision,
  completion_proposal,
  completion_shown_event,
  completion_decision_event,
  completion_trust_record,
  completion_usage_record,
} from "std/agent/completions"

Core helpers:

completion_context(input) normalizes file path, language, stack, cursor, prefix/suffix, recent edits, LSP or IntelliSense facts, diagnostics, symbols, local scope, open files, and codebase snippets into harn.completion.context.v1.
completion_policy(input) normalizes workspace or enterprise controls: enabled state, language/provider/model allow and deny lists, remote-model allowance, codebase-context allowance, response-time and prompt-size limits, prompt recording mode, trust-event emission, and retention.
completion_policy_decision(context, route, policy) returns {allowed, reason_code, reason, policy} with stable machine-readable denial reasons.
completion_model_route(opts) asks the model catalog for a coding-oriented route that can be overridden by the host.
completion_proposal(input) normalizes inline insertions, replacements, next edits, and workspace edits into harn.completion.proposal.v1.
completion_shown_event(...) and completion_decision_event(...) produce harn.completion.event.v1 records for analytics and replay.
completion_trust_record(event) projects a completion event into the TrustGraph record shape.
completion_usage_record(event) projects completion usage into the common usage row shape used by host dashboards.

Example:

import {
  completion_context,
  completion_policy,
  completion_policy_decision,
  completion_proposal,
  completion_shown_event,
} from "std/agent/completions"

pipeline default() {
  const context = completion_context({
    surface: "editor",
    file_path: "src/main.rs",
    language: "rust",
    prefix: "const result = client.",
    suffix: "\nprintln!(\"{result:?}\");",
    symbols: [{name: "client", kind: "variable"}],
  })

  const policy = completion_policy({
    disabled_languages: ["markdown"],
    allow_remote_models: false,
    prompt_recording: "metadata",
  })

  const decision = completion_policy_decision(
    context,
    {provider: "ollama", model_id: "qwen2.5-coder"},
    policy,
  )
  if !decision.allowed {
    return decision.reason
  }

  const shown = completion_shown_event(
    completion_proposal({
      kind: "inline_insert",
      provider: "ollama",
      model_id: "qwen2.5-coder",
      insert_text: "send()?;",
    }),
    context,
    {input_tokens: 800, output_tokens: 12, cost_usd: 0.0},
    {policy: policy, prompt: "rendered private prompt"},
  )
  return shown.record_id
}

prompt_recording: "metadata" stores prompt length and hash. "full" stores prompt text for replay. "off" records neither.