| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "fmt" |
| 6 | "time" |
| 7 | |
| 8 | "reasonix/internal/billing" |
| 9 | "reasonix/internal/event" |
| 10 | "reasonix/internal/provider" |
| 11 | ) |
| 12 | |
| 13 | // TaskBudget bounds one task on the axes its failures are reported in, and |
| 14 | // every axis ships off: stopping a task is the user's call. Tokens is the one |
| 15 | // that generalizes — a slow expensive loop accumulates them and so does a fast |
| 16 | // empty one, where wall clock catches only the first and money is not portable |
| 17 | // across models. |
| 18 | type TaskBudget struct { |
| 19 | Cost float64 |
| 20 | Wall time.Duration |
| 21 | Tokens int |
| 22 | } |
| 23 | |
| 24 | // normalizeTaskBudget reads a negative value as unset, so a disabled axis and |
| 25 | // an unconfigured one behave identically. |
| 26 | func normalizeTaskBudget(b TaskBudget) TaskBudget { |
| 27 | if b.Cost < 0 { |
| 28 | b.Cost = 0 |
| 29 | } |
| 30 | if b.Wall < 0 { |
| 31 | b.Wall = 0 |
| 32 | } |
| 33 | if b.Tokens < 0 { |
| 34 | b.Tokens = 0 |
| 35 | } |
| 36 | return b |
| 37 | } |
| 38 | |
| 39 | // runBudget accumulates what a turn has actually spent. Rounds are a poor proxy |
| 40 | // for it: the same hundred of them cost minutes or hours depending on what each |
| 41 | // one read and how long the model thought, and the failures worth stopping are |
| 42 | // reported in hours and tokens, never in rounds. |
| 43 | type runBudget struct { |
| 44 | started time.Time |
| 45 | rounds int |
| 46 | requests int |
| 47 | promptTokens int |
| 48 | outputTokens int |
| 49 | cost float64 |
| 50 | pricedRounds int |
| 51 | unpricedTurns bool |
| 52 | // limit is configuration, not accumulation: it survives the reset that |
| 53 | // starts a new task. |
| 54 | limit TaskBudget |
| 55 | } |
| 56 | |
| 57 | // observe folds one round's provider usage into the turn's running total. |
| 58 | // A round whose usage never arrived still counts as a round, so the axis never |
| 59 | // reads cheaper than the turn actually was. |
| 60 | func (b *runBudget) observe(usage *provider.Usage, pricing *provider.Pricing) { |
| 61 | var quote *billing.CostQuote |
| 62 | if usage != nil && pricing != nil { |
| 63 | quote = event.EnsureCostQuote(event.Event{Kind: event.Usage, Usage: usage, Pricing: pricing}, nil) |
| 64 | } |
| 65 | b.observeQuote(usage, quote) |
| 66 | } |
| 67 | |
| 68 | func (b *runBudget) observeQuote(usage *provider.Usage, quote *billing.CostQuote) { |
| 69 | b.rounds++ |
| 70 | if usage == nil { |
| 71 | return |
| 72 | } |
| 73 | b.requests += usageRequestCount(usage) |
| 74 | b.promptTokens += usage.PromptTokens |
| 75 | b.outputTokens += usage.CompletionTokens |
| 76 | if usage.Unknown { |
| 77 | b.unpricedTurns = true |
| 78 | } |
| 79 | if quote == nil || (!quote.CostComplete && quote.IncompleteReason != "usage_unknown") || quote.Original.Currency == "" { |
| 80 | b.unpricedTurns = true |
| 81 | return |
| 82 | } |
| 83 | b.cost += quote.Original.Float64() |
| 84 | b.pricedRounds++ |
| 85 | } |
| 86 | |
| 87 | func (b *runBudget) elapsed() time.Duration { |
| 88 | if b.started.IsZero() { |
| 89 | return 0 |
| 90 | } |
| 91 | return time.Since(b.started) |
| 92 | } |
| 93 | |
| 94 | // totals is the shadow reading for one scope: counts and money, never content. |
| 95 | func (b *runBudget) totals() event.RunBudgetTotals { |
| 96 | return event.RunBudgetTotals{ |
| 97 | Rounds: b.rounds, |
| 98 | Requests: b.requests, |
| 99 | PromptTokens: b.promptTokens, |
| 100 | OutputTokens: b.outputTokens, |
| 101 | Cost: b.cost, |
| 102 | Priced: !b.unpricedTurns && b.pricedRounds > 0, |
| 103 | ElapsedMs: b.elapsed().Milliseconds(), |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | // exceeded names the first axis the task has spent past, or "" while inside |
| 108 | // the budget. Cost only counts when the turn was actually priced: an unpriced |
| 109 | // model reads as free, and a free reading must never look like a crossing. |
| 110 | func (b *runBudget) exceeded(limit TaskBudget) (axis, detail string) { |
| 111 | if limit.Tokens > 0 { |
| 112 | if used := b.promptTokens + b.outputTokens; used >= limit.Tokens { |
| 113 | return "token", fmt.Sprintf("task used %d tokens, reaching the %d budget", used, limit.Tokens) |
| 114 | } |
| 115 | } |
| 116 | if limit.Cost > 0 && b.pricedRounds > 0 && b.cost >= limit.Cost { |
| 117 | return "cost", fmt.Sprintf("task spend %.4f reached the %.4f budget", b.cost, limit.Cost) |
| 118 | } |
| 119 | if limit.Wall > 0 { |
| 120 | if elapsed := b.elapsed(); elapsed >= limit.Wall { |
| 121 | return "time", fmt.Sprintf("task ran %s, past the %s budget", |
| 122 | elapsed.Round(time.Second), limit.Wall) |
| 123 | } |
| 124 | } |
| 125 | return "", "" |
| 126 | } |
| 127 | |
| 128 | // taskBudgetLimit resolves this turn's bound: a host-injected budget wins over |
| 129 | // the configured one, which is how an unattended loop gets a ceiling while |
| 130 | // ordinary chat keeps none. |
| 131 | func (a *Agent) taskBudgetLimit(ctx context.Context) TaskBudget { |
| 132 | if b, ok := taskBudgetFromContext(ctx); ok { |
| 133 | return b |
| 134 | } |
| 135 | return a.task.budget.limit |
| 136 | } |
| 137 | |
| 138 | // ResetTaskBudget starts a fresh user-approved spend slice without touching |
| 139 | // Delivery evidence or the persisted Goal usage totals. Callers use this only |
| 140 | // after a resumable explicit-budget pause, while no Agent Run is active. |
| 141 | func (a *Agent) ResetTaskBudget() { |
| 142 | a.task.budget = runBudget{limit: a.task.budget.limit} |
| 143 | } |
| 144 | |
| 145 | // observeRunBudget folds a round into both scopes and reports them. |
| 146 | func (a *Agent) observeRunBudget(state *turnRuntime, usage *provider.Usage, quotes ...*billing.CostQuote) { |
| 147 | if state == nil { |
| 148 | return |
| 149 | } |
| 150 | var quote *billing.CostQuote |
| 151 | if len(quotes) > 0 { |
| 152 | quote = quotes[0] |
| 153 | } else if usage != nil && a.svc.pricing != nil { |
| 154 | e := event.Event{Kind: event.Usage, ModelRef: a.modelRef, Usage: usage, Pricing: a.svc.pricing, UsageSource: a.usageSource} |
| 155 | quote = event.EnsureCostQuote(e, a.svc.quoteContext) |
| 156 | } |
| 157 | state.budget.observeQuote(usage, quote) |
| 158 | if a.task.budget.started.IsZero() { |
| 159 | a.task.budget.started = state.budget.started |
| 160 | } |
| 161 | a.task.budget.observeQuote(usage, quote) |
| 162 | currency := "" |
| 163 | if quote != nil { |
| 164 | currency = billing.CurrencySymbol(quote.Original.Currency) |
| 165 | } |
| 166 | event.RecordRunBudget(a.svc.sink, event.RunBudgetSample{ |
| 167 | Turn: state.budget.totals(), |
| 168 | Task: a.task.budget.totals(), |
| 169 | Currency: currency, |
| 170 | }) |
| 171 | } |
| 172 | |
| 173 | type taskBudgetContextKey struct{} |
| 174 | |
| 175 | // WithTaskBudget overrides a run's task budget for one turn. The agent serving |
| 176 | // an unattended loop and the one serving chat are the same instance, so the |
| 177 | // bound is a property of the turn, not of construction. |
| 178 | func WithTaskBudget(ctx context.Context, b TaskBudget) context.Context { |
| 179 | return context.WithValue(ctx, taskBudgetContextKey{}, normalizeTaskBudget(b)) |
| 180 | } |
| 181 | |
| 182 | func taskBudgetFromContext(ctx context.Context) (TaskBudget, bool) { |
| 183 | if ctx == nil { |
| 184 | return TaskBudget{}, false |
| 185 | } |
| 186 | b, ok := ctx.Value(taskBudgetContextKey{}).(TaskBudget) |
| 187 | return b, ok |
| 188 | } |
| 189 |