Documentation
¶
Overview ¶
Package usage turns the audit log's real, per-call judge token usage into a cost report - both judges this project has: the text provenance judge (internal/provenance, PROVENANCE_JUDGE_USAGE entries) and the screenshot-reasoning one (internal/visualprovenance, VISUAL_PROVENANCE_JUDGE_USAGE entries).
This is the first concrete slice of the metered-judge paid layer: real usage accounting from real API responses, not yet real billing. The point is to have a genuine, working answer to "here's how this would actually be billed" before any billing integration exists - the same "prove the mechanism standalone before wiring it into anything" pattern this project has used everywhere else (biscuit-go, the Anthropic SDK, Vault dynamic secrets all got a throwaway validation step before being built on top of).
One pricing rate covers both judges, deliberately not two: Anthropic bills image content as regular input tokens (confirmed against the real API response when internal/visualprovenance was validated - a screenshot simply produces a larger input_tokens count, not a different billing field), so the per-token rate below is the whole story for either judge. What differs is the token VOLUME per call, which the real numbers below already capture without needing a second rate.
Like internal/compliance, this is a single pass over the audit log - the append-only source of truth - not a running counter kept in memory, so a report generated at any point reflects real historical usage rather than only what happened to occur since the process last restarted.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type JudgeUsage ¶
type JudgeUsage struct {
TotalInvocations int `json:"total_invocations"`
TotalErrors int `json:"total_errors"`
InputTokens int64 `json:"input_tokens"`
OutputTokens int64 `json:"output_tokens"`
EstimatedCostUSD float64 `json:"estimated_cost_usd"`
ByTool map[string]*ToolUsage `json:"by_tool"`
ByDay map[string]*DayUsage `json:"by_day"`
}
JudgeUsage is real invocation counts and real token totals from real API responses, for one judge - text or visual - with an estimated dollar cost computed from the named, cited rate above, never presented as more precise than "estimated."
func (JudgeUsage) SortedDays ¶
func (j JudgeUsage) SortedDays() []string
SortedDays returns ByDay's keys in chronological order, for a report that reads front-to-back rather than in random map order.
type Report ¶
type Report struct {
PricingNote string `json:"pricing_note"`
Text JudgeUsage `json:"text_judge"`
Visual JudgeUsage `json:"visual_judge"`
}
func Generate ¶
Generate reads the audit log once and tallies both judges' usage/error entries (see internal/handler's assessProvenance and assessVisualProvenance for exactly what gets written). Malformed lines are skipped, not fatal - the audit log is append-only and long-running; one bad line (a truncated write mid-crash, say) shouldn't take the whole report down.
Source Files