| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | |
| 6 | "reasonix/internal/provider" |
| 7 | ) |
| 8 | |
| 9 | const minSummaryOutputTokens = 512 |
| 10 | |
| 11 | // summaryOutputBudget scales only shared/unknown-window summaries. Providers |
| 12 | // with an independent completion window keep the full digest cap; smaller |
| 13 | // shared windows reserve one quarter for a useful briefing without crowding |
| 14 | // every fold out of the prompt budget. |
| 15 | func (a *Agent) summaryOutputBudget() int { |
| 16 | if contextBudgetPolicyOf(a.svc.prov).WindowMode == provider.ContextWindowIndependent { |
| 17 | return summaryOutputMaxTokens |
| 18 | } |
| 19 | window := a.effectiveContextWindow() |
| 20 | if window <= 0 { |
| 21 | return summaryOutputMaxTokens |
| 22 | } |
| 23 | return min(summaryOutputMaxTokens, max(window/4, minSummaryOutputTokens)) |
| 24 | } |
| 25 | |
| 26 | // foldSummary is what compaction reports about turning a fold into a digest. |
| 27 | // It is populated even when the call fails, so telemetry still records how |
| 28 | // large the attempt was and that exactly one call was used. |
| 29 | type foldSummary struct { |
| 30 | Text string |
| 31 | Mode string |
| 32 | RequestID string |
| 33 | Usage *provider.Usage |
| 34 | FoldTokens int |
| 35 | Spans int |
| 36 | InputMode string |
| 37 | } |
| 38 | |
| 39 | func summaryInputTokens(msgs []provider.Message) int { |
| 40 | return estimateMessagesTokens(msgs) |
| 41 | } |
| 42 | |
| 43 | func (a *Agent) guardedSummaryInputTokens(msgs []provider.Message) int { |
| 44 | return a.estimatedVisibleRequestTokens(msgs) |
| 45 | } |
| 46 | |
| 47 | func (a *Agent) summaryInputBudget(instructions string) int { |
| 48 | window := a.effectiveContextWindow() |
| 49 | if window <= 0 { |
| 50 | window = a.contextWindow |
| 51 | } |
| 52 | if window <= 0 { |
| 53 | return 0 |
| 54 | } |
| 55 | return max(0, window-a.summaryOutputBudget()-estimateTextTokens(compactionInstruction)-estimateTextTokens(instructions)-protocolReserveTokens) |
| 56 | } |
| 57 | |
| 58 | // foldToSummary turns a fold region into one digest with exactly one provider |
| 59 | // request. Pressure-time tool pruning is durable and happens before this call; |
| 60 | // the summary request never performs a private second transformation. |
| 61 | func (a *Agent) foldToSummary(ctx context.Context, fold []provider.Message, instructions string) (foldSummary, error) { |
| 62 | return a.foldToSummaryMode(ctx, fold, instructions, SummaryInputCachePrefix) |
| 63 | } |
| 64 | |
| 65 | func (a *Agent) foldToSummaryMode(ctx context.Context, fold []provider.Message, instructions, inputMode string) (foldSummary, error) { |
| 66 | res := foldSummary{Mode: CompactionModeSummarized, Spans: 1, FoldTokens: summaryInputTokens(fold), InputMode: inputMode} |
| 67 | if inputMode == SummaryInputSlim { |
| 68 | summary, usage, err := a.summarizeTranscript(ctx, fold, instructions) |
| 69 | res.Text, res.Usage = summary, usage |
| 70 | return res, err |
| 71 | } |
| 72 | return a.singleCallSummary(ctx, res, fold, instructions) |
| 73 | } |
| 74 | |
| 75 | func (a *Agent) singleCallSummary(ctx context.Context, res foldSummary, fold []provider.Message, instructions string) (foldSummary, error) { |
| 76 | summary, mode, usage, reqID, err := a.runCompactionSummary(ctx, fold, instructions) |
| 77 | res.Text, res.Mode, res.Usage, res.RequestID = summary, mode, usage, reqID |
| 78 | return res, err |
| 79 | } |
| 80 | |
| 81 | func (a *Agent) foldSummaryWithTelemetry(ctx context.Context, trigger string, fold []provider.Message, instructions string, sourceTokens int, inputMode string) (foldSummary, CompactionTelemetry, error) { |
| 82 | res, err := a.foldToSummaryMode(ctx, fold, instructions, inputMode) |
| 83 | tele := compactionTelemetryFromSummary(trigger, a.CacheState(), sourceTokens, res) |
| 84 | if err != nil { |
| 85 | tele.Error = err.Error() |
| 86 | } |
| 87 | return res, tele, err |
| 88 | } |
| 89 |