@juspay/neurolink/dist/localUsage/copilotCliReader.d.ts

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

/**
 * Reads token usage out of Copilot CLI's local SQLite store.
 *
 * Store: `~/.copilot/session-store.db`. Usage lives on the
 * `assistant_usage_events` table — one INSERT-only row per assistant turn,
 * with `model`, `input_tokens`, `output_tokens`, `cache_read_tokens`,
 * `cache_write_tokens`, `reasoning_tokens` and a `created_at` timestamp
 * (ISO-8601 with a `Z` suffix). Confirmed no duplicate `(session_id,
 * turn_index)` pair exists on a reference machine, so — like OpenCode's
 * `message` table — this needs no dedup logic beyond the table's own
 * uniqueness: every row is counted once, matching the `rowid-high-water-mark`
 * strategy already established for that reader.
 *
 * The same database also holds a JSONL-adjacent `session.shutdown` event with
 * its own `modelMetrics`/`tokenDetails` payload (written to
 * `~/.copilot/session-state/*\/events.jsonl`, not this database). It was
 * deliberately NOT used here: cross-checking real sessions confirmed the two
 * sources report identical totals for the sessions they both cover, so
 * combining them would double-count rather than add coverage. The SQLite
 * table is the sole source; the tradeoff is a disclosed gap for any history
 * that predates the table's introduction.
 *
 * `cache_read_tokens` and `cache_write_tokens` are both SUBSETS of
 * `input_tokens`, not disjoint — confirmed arithmetically against a real
 * `session.shutdown` record (session `c2c38d0b`): `tokenDetails.input.tokenCount
 * = 3`, `cache_write = 24047`, and `usage.inputTokens = 24050` — exactly
 * `3 + 24047`, with nothing left over for a separate additive interpretation.
 * So `inputTokens` here is `input_tokens - cache_read_tokens -
 * cache_write_tokens`, the same style of subtraction `codexReader.ts`,
 * `qwenCodeReader.ts` and `geminiCliReader.ts` make for the same reason: this
 * subsystem's `LocalUsageTotals.inputTokens` + `.cacheReadTokens` must sum to
 * the true prompt size without double-counting. `reasoning_tokens` folds into
 * output — Copilot's own SDK types (`ShutdownModelMetricUsage`) group it
 * alongside `outputTokens` as a further breakdown of output, not a separate
 * accounting bucket.
 *
 * Cost is deliberately `unavailable`. The table's own `request_multiplier`
 * and `total_nano_aiu` columns are Copilot's request-quota accounting, not a
 * USD figure — the SDK's own generated types tag the closest concept,
 * `ShutdownModelMetricUsage.cost`, `@experimental`, and every real sample row
 * on the reference machine had `request_multiplier: 0.0`. Modeling a
 * per-token dollar figure from public list prices would also ignore that
 * Copilot CLI is typically used under a flat-rate Copilot subscription, the
 * same reasoning that keeps Codex's confidence `unavailable`. This holds even
 * when the underlying model is a metered vendor's (e.g. `claude-haiku-4-5` on
 * the reference machine) — Copilot's telemetry layer normalizes across
 * backend vendors into its own accounting convention, and this reader has no
 * way to tell whether a given row was billed by the minute, by the token, or
 * not separately billed at all.
 */
import type { LocalUsageReader } from "../types/index.js";
export declare function createCopilotCliReader(): Promise<LocalUsageReader>;