| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "sort" |
| 9 | "strings" |
| 10 | "unicode/utf8" |
| 11 | |
| 12 | "reasonix/internal/ablation" |
| 13 | "reasonix/internal/event" |
| 14 | "reasonix/internal/provider" |
| 15 | ) |
| 16 | |
| 17 | // Compaction is a low-frequency cache-reset point: the prompt grows append-only |
| 18 | // until compactRatio of the window is crossed, then pressure-time tool pruning |
| 19 | // and up to two cache-aligned summary checkpoints restore headroom. |
| 20 | const ( |
| 21 | defaultCompactRatio = 0.80 // sole automatic maintenance trigger (new configs) |
| 22 | recentTailBudgetRatio = 0.16 // recent verbatim tail as a fraction of the window |
| 23 | summaryOutputMaxTokens = 8192 // max digest output; further clipped by remaining candidate space |
| 24 | |
| 25 | // summaryReasoningMaxBytes clamps a surfaced reasoning-only summary |
| 26 | // (~8k tokens of bytes), matching the summaryOutputMaxTokens envelope. |
| 27 | summaryReasoningMaxBytes = 32768 |
| 28 | |
| 29 | minRecentKeep = 2 // never keep fewer recent messages than this |
| 30 | minCompactMessages = 2 // skip compaction below this many compactable messages |
| 31 | fallbackTokPerChar = 0.25 // ~4 chars/token, used before any usage is available to calibrate |
| 32 | protocolReserveTokens = 256 // provider framing and control fields not represented by message estimates |
| 33 | ) |
| 34 | |
| 35 | var ( |
| 36 | errSummaryOutputTruncated = errors.New("summarizer output truncated") |
| 37 | errCheckpointRejected = errors.New("checkpoint candidate rejected") |
| 38 | ) |
| 39 | |
| 40 | // summaryTag wraps the compaction summary so the model can distinguish it from |
| 41 | // live user input and later strip or skip it when reasoning about the current turn. |
| 42 | const ( |
| 43 | summaryTagOpen = "<compaction-summary>" |
| 44 | summaryTagClose = "</compaction-summary>" |
| 45 | ) |
| 46 | |
| 47 | // compactionInstruction is appended as the only new message after an otherwise |
| 48 | // byte-stable sampling prefix. This lets providers reuse the ordinary request's |
| 49 | // system, tools and message-prefix KV cache. |
| 50 | const compactionInstruction = `Compact the preceding conversation prefix into a durable resume briefing. |
| 51 | Write under these exact headings, omitting a heading only if it has no content: |
| 52 | |
| 53 | ## Standing facts & constraints |
| 54 | Everything the user stated that still governs the work — names, paths, IDs, versions, tokens, preferences, and hard "never do X" rules — in their own words. Be exhaustive; this is the durable contract, so prefer over- to under-including. |
| 55 | |
| 56 | ## Goal |
| 57 | The user's request and intent. |
| 58 | |
| 59 | ## Decisions & rationale |
| 60 | Key choices made so far and why — so they are not re-litigated or reversed. |
| 61 | |
| 62 | ## Files & code |
| 63 | Files read or modified, with the specific facts that matter: signatures, line locations, data shapes, and exact edits applied. Be concrete; this is what lets the agent act without re-reading everything. |
| 64 | |
| 65 | ## Commands & outcomes |
| 66 | Commands run (builds, tests, git) and their relevant results — what passed, what failed, and the error text that matters. |
| 67 | |
| 68 | ## Errors & fixes |
| 69 | Problems hit and how they were resolved (or not), so the same dead ends are not repeated. |
| 70 | |
| 71 | ## Pending & next step |
| 72 | What is still in progress or unstarted, and the single most concrete next action to take. |
| 73 | |
| 74 | Rules: be terse — bullet points and fragments, not prose. Preserve identifiers, paths, and numbers exactly. Merge valid facts from any existing <compaction-summary> and remove facts superseded by later messages. Do NOT invent anything not present in the messages; if something is unknown, leave it out rather than guessing. Output only the structured Markdown briefing. Do not call tools. Do not output reasoning.` |
| 75 | |
| 76 | // compactTrigger is the sole automatic context-maintenance boundary. Output |
| 77 | // budgets are intentionally absent: they are clipped against the final request |
| 78 | // at send time and must never make compaction happen earlier than the user's |
| 79 | // configured compact_ratio. |
| 80 | func (a *Agent) compact(ctx context.Context, trigger, instructions string, force bool) error { |
| 81 | _, err := a.compactToProjectionWithChunked(ctx, trigger, instructions, foldRequest{ |
| 82 | force: force, allowChunked: trigger == CompactionTriggerManual, |
| 83 | }) |
| 84 | return err |
| 85 | } |
| 86 | |
| 87 | func (a *Agent) compactToProjection(ctx context.Context, trigger, instructions string, force, mustFree bool) (CompactionOutcome, error) { |
| 88 | return a.compactToProjectionWithChunked(ctx, trigger, instructions, foldRequest{force: force, mustFree: mustFree}) |
| 89 | } |
| 90 | |
| 91 | func (a *Agent) compactToProjectionWithChunked(ctx context.Context, trigger, instructions string, req foldRequest) (outcome CompactionOutcome, err error) { |
| 92 | ctx, finish := a.beginCompactionRun(ctx) |
| 93 | defer func() { err = finish(err) }() |
| 94 | if err := a.sess.compactionRunMu.acquire(ctx); err != nil { |
| 95 | return CompactionNoop, err |
| 96 | } |
| 97 | defer a.sess.compactionRunMu.Unlock() |
| 98 | return a.compactToProjectionLocked(ctx, trigger, instructions, req) |
| 99 | } |
| 100 | |
| 101 | func (a *Agent) compactTrigger() int { |
| 102 | window := a.effectiveContextWindow() |
| 103 | if a == nil || window <= 0 { |
| 104 | return 0 |
| 105 | } |
| 106 | ratio := a.compactRatio |
| 107 | if ratio <= 0 { |
| 108 | ratio = defaultCompactRatio |
| 109 | } |
| 110 | if a.ablation.Off(ablation.Compaction) { |
| 111 | ratio = 0.5 |
| 112 | } |
| 113 | return max(1, int(float64(window)*ratio)) |
| 114 | } |
| 115 | |
| 116 | // hardInputCeiling is a physical input-safety boundary, not another user |
| 117 | // compaction threshold. Reply budgets are resolved independently at send time. |
| 118 | func (a *Agent) hardInputCeiling() int { |
| 119 | window := a.effectiveContextWindow() |
| 120 | if a == nil || window <= 0 { |
| 121 | return 0 |
| 122 | } |
| 123 | return max(1, window-protocolReserveTokens) |
| 124 | } |
| 125 | |
| 126 | // recentTailBudget is the content-construction budget for the recent verbatim |
| 127 | // tail. Harness-style compaction always retains 16% of the model window. |
| 128 | func (a *Agent) recentTailBudget() int { |
| 129 | window := a.effectiveContextWindow() |
| 130 | if a == nil || window <= 0 { |
| 131 | return 1 |
| 132 | } |
| 133 | return max(1, int(float64(window)*recentTailBudgetRatio)) |
| 134 | } |
| 135 | |
| 136 | // foldEconomics estimates whether compacting the given region saves enough |
| 137 | // tokens to justify the summarization API call. It returns false when the |
| 138 | // region is too small for the savings to outweigh the extra round-trip cost |
| 139 | // and latency of calling the summarizer. |
| 140 | func foldEconomics(region []provider.Message) bool { |
| 141 | const minFoldTokens = 400 |
| 142 | return estimateMessagesTokens(region) >= minFoldTokens |
| 143 | } |
| 144 | |
| 145 | func estimateMessagesTokens(msgs []provider.Message) int { |
| 146 | total := 0 |
| 147 | for _, m := range msgs { |
| 148 | if m.LocalOnly || IsPinnedContextRevision(m) { |
| 149 | continue |
| 150 | } |
| 151 | total += 4 // chat-message framing overhead |
| 152 | total += estimateTextTokens(m.Content) |
| 153 | total += estimateTextTokens(m.ReasoningContent) |
| 154 | total += estimateTextTokens(m.Name) |
| 155 | total += estimateTextTokens(m.ToolCallID) |
| 156 | for _, tc := range m.ToolCalls { |
| 157 | total += 8 |
| 158 | total += estimateTextTokens(tc.ID) |
| 159 | total += estimateTextTokens(tc.Name) |
| 160 | total += estimateTextTokens(tc.Arguments) |
| 161 | } |
| 162 | for _, item := range m.ResponsesItems { |
| 163 | total += estimateTextTokens(string(item)) |
| 164 | } |
| 165 | for _, search := range m.ServerSearch { |
| 166 | provider.WalkServerSearchEstimate(search, func(s string) { |
| 167 | total += estimateTextTokens(s) |
| 168 | }) |
| 169 | } |
| 170 | } |
| 171 | return total |
| 172 | } |
| 173 | |
| 174 | func estimateTextTokens(s string) int { |
| 175 | if s == "" { |
| 176 | return 0 |
| 177 | } |
| 178 | // A conservative cross-language approximation: English-ish text trends near |
| 179 | // four bytes per token, while CJK-heavy text is closer to one rune per token. |
| 180 | bytes := len(s) |
| 181 | runes := utf8.RuneCountInString(s) |
| 182 | byBytes := (bytes + 3) / 4 |
| 183 | if runes > byBytes { |
| 184 | return runes |
| 185 | } |
| 186 | return byBytes |
| 187 | } |
| 188 | |
| 189 | // SummarizeFrom keeps the compatibility index contract while installing a |
| 190 | // projection that compresses from that user-turn boundary onward. |
| 191 | func (a *Agent) SummarizeFrom(ctx context.Context, fromIdx int) error { |
| 192 | return a.summarizeAtProjectionBoundary(ctx, fromIdx, "after") |
| 193 | } |
| 194 | |
| 195 | // SummarizeUpTo keeps the compatibility index contract while installing a |
| 196 | // projection that compresses everything before that user-turn boundary. |
| 197 | func (a *Agent) SummarizeUpTo(ctx context.Context, toIdx int) error { |
| 198 | return a.summarizeAtProjectionBoundary(ctx, toIdx, "before") |
| 199 | } |
| 200 | |
| 201 | func (a *Agent) summarizeAtProjectionBoundary(ctx context.Context, canonicalIndex int, direction string) (resultErr error) { |
| 202 | ctx, finish := a.beginCompactionRun(ctx) |
| 203 | defer func() { resultErr = finish(resultErr) }() |
| 204 | if err := ctx.Err(); err != nil { |
| 205 | return err |
| 206 | } |
| 207 | snap := a.snapshotExplicitCompression() |
| 208 | if canonicalIndex < 0 || canonicalIndex >= len(snap.canonical) { |
| 209 | return nil |
| 210 | } |
| 211 | anchor := snap.canonical[canonicalIndex] |
| 212 | if !compressAnchorCandidate(anchor) { |
| 213 | return nil |
| 214 | } |
| 215 | visibleIndex := -1 |
| 216 | for i, msg := range snap.visible { |
| 217 | if !compressAnchorCandidate(msg) { |
| 218 | continue |
| 219 | } |
| 220 | if anchor.CreatedAt != 0 && msg.CreatedAt == anchor.CreatedAt { |
| 221 | visibleIndex = i |
| 222 | break |
| 223 | } |
| 224 | if anchor.CreatedAt == 0 && UserMessageText(msg) == UserMessageText(anchor) { |
| 225 | if visibleIndex >= 0 { |
| 226 | return fmt.Errorf("summarize boundary is ambiguous in the current model context") |
| 227 | } |
| 228 | visibleIndex = i |
| 229 | } |
| 230 | } |
| 231 | if visibleIndex < 0 { |
| 232 | return fmt.Errorf("context compression unavailable: selected turn is no longer present in the model context") |
| 233 | } |
| 234 | result, err := a.compressVisibleRange(ctx, snap, CompactionTriggerManual, direction, visibleIndex, anchorPreview(UserMessageText(anchor)), "") |
| 235 | if err != nil { |
| 236 | return err |
| 237 | } |
| 238 | if result.Status != "ok" { |
| 239 | if result.Status == "noop" && noCompressionHistory(result.Reason) && result.SourceTokens < a.hardInputCeiling() { |
| 240 | return ctx.Err() |
| 241 | } |
| 242 | reason := strings.TrimSpace(result.Reason) |
| 243 | if reason == "" { |
| 244 | reason = "selected range did not reduce the model context" |
| 245 | } |
| 246 | return fmt.Errorf("context compression skipped: %s", reason) |
| 247 | } |
| 248 | return nil |
| 249 | } |
| 250 | |
| 251 | // IsCompactionSummary reports whether m is a rolling digest inserted by a |
| 252 | // prior compaction fold. Exported for session owners outside this package |
| 253 | // (e.g. the guardian) whose turn rollback must not treat a digest as a |
| 254 | // disposable user message. |
| 255 | func IsCompactionSummary(m provider.Message) bool { return isCompactionSummary(m) } |
| 256 | |
| 257 | func (a *Agent) activeTurnStart(msgs []provider.Message) int { |
| 258 | createdAt := a.activeTurnCreatedAt.Load() |
| 259 | if createdAt == 0 { |
| 260 | return -1 |
| 261 | } |
| 262 | for i, m := range msgs { |
| 263 | if m.Role == provider.RoleUser && m.CreatedAt == createdAt { |
| 264 | return i |
| 265 | } |
| 266 | } |
| 267 | return -1 |
| 268 | } |
| 269 | |
| 270 | // isCompactionSummary reports whether m is a rolling summary from a prior fold. |
| 271 | func isCompactionSummary(m provider.Message) bool { |
| 272 | return m.Role == provider.RoleUser && |
| 273 | strings.HasPrefix(strings.TrimLeft(m.Content, "\n "), summaryTagOpen) |
| 274 | } |
| 275 | |
| 276 | // pinnedPrefixLen keeps only the system message. All older user turns, |
| 277 | // failures, and [[keep]] markers enter the Harness-style summary prefix. |
| 278 | func (a *Agent) pinnedPrefixLen(msgs []provider.Message) int { |
| 279 | if len(msgs) > 0 && msgs[0].Role == provider.RoleSystem { |
| 280 | return 1 |
| 281 | } |
| 282 | return 0 |
| 283 | } |
| 284 | |
| 285 | // planCompaction returns [head:start] to fold while retaining the newest 16% |
| 286 | // of the model window and keeping tool-call/result groups balanced. |
| 287 | func (a *Agent) planCompaction(msgs []provider.Message, min int, force bool) (head, start int, ok bool) { |
| 288 | head = a.pinnedPrefixLen(msgs) |
| 289 | if a.contextWindow > 0 { |
| 290 | budget := a.recentTailBudget() |
| 291 | if force { |
| 292 | // Fixed instructions are not compressible history. Including them |
| 293 | // here can reserve the entire conversation as the recent tail and |
| 294 | // leave only a non-summarizable context snapshot in the fold. |
| 295 | _, history, _ := a.partitionFoldForProjectionAt(msgs[head:], head, latestSessionContextIndex(msgs)) |
| 296 | if half := estimateMessagesTokens(modelInputMessages(history)) / 2; half > 0 && half < budget { |
| 297 | budget = half |
| 298 | } |
| 299 | } |
| 300 | start = tailStart(msgs, head, budget, a.tokPerChar(), a.tailFloor()) |
| 301 | // Remeasure when force or non-strict roles; strict-alternating otherwise |
| 302 | // keeps a cheap tokPerChar overestimate of the tail under force. |
| 303 | floor := max(head, len(msgs)-a.tailFloor()) |
| 304 | remeasure := force || !a.strictAlternatingRoles |
| 305 | for remeasure && start < floor && estimateMessagesTokens(provider.ModelMessages(msgs[start:])) > budget { |
| 306 | start++ |
| 307 | for start < floor && start < len(msgs) && msgs[start].Role == provider.RoleTool { |
| 308 | start++ |
| 309 | } |
| 310 | } |
| 311 | } else { |
| 312 | // No window: keep a fixed recent count, aligned off tool results. |
| 313 | start = len(msgs) - a.tailFloor() |
| 314 | for start > head && start < len(msgs) && msgs[start].Role == provider.RoleTool { |
| 315 | start-- |
| 316 | } |
| 317 | } |
| 318 | start = max(start, head) |
| 319 | if start-head < min { |
| 320 | return head, start, false |
| 321 | } |
| 322 | return head, start, true |
| 323 | } |
| 324 | |
| 325 | func (a *Agent) tailFloor() int { |
| 326 | return 0 |
| 327 | } |
| 328 | |
| 329 | // tailStart walks newest→oldest, growing the verbatim tail until the next |
| 330 | // message would push its token estimate past budgetTokens (but never below |
| 331 | // minKeep messages), then aligns the boundary back off any tool result so the |
| 332 | // tail never begins with an orphan whose assistant tool_calls were summarized |
| 333 | // away. |
| 334 | func tailStart(msgs []provider.Message, head, budgetTokens int, tokPerChar float64, minKeep int) int { |
| 335 | start := len(msgs) |
| 336 | acc := 0 |
| 337 | for i := len(msgs) - 1; i > head; i-- { |
| 338 | c := int(float64(msgChars(msgs[i])) * tokPerChar) |
| 339 | if len(msgs)-i > minKeep && acc+c > budgetTokens { |
| 340 | break |
| 341 | } |
| 342 | acc += c |
| 343 | start = i |
| 344 | } |
| 345 | // start == len(msgs) when nothing fit the tail (a session too small to have a |
| 346 | // message after head); there is no msgs[start] to align off, and the caller's |
| 347 | // minCompactMessages check then no-ops the pass. |
| 348 | for start > head && start < len(msgs) && msgs[start].Role == provider.RoleTool { |
| 349 | start-- |
| 350 | } |
| 351 | return start |
| 352 | } |
| 353 | |
| 354 | // tokPerChar derives a tokens-per-character ratio from the last turn's real |
| 355 | // usage so per-message estimates track the provider's tokenizer without a local |
| 356 | // one. Reasoning content is excluded from the char count to match the prompt |
| 357 | // actually sent (the provider strips it). Falls back to ~4 chars/token before |
| 358 | // any usage is known, and ignores absurd ratios. |
| 359 | func (a *Agent) tokPerChar() float64 { |
| 360 | if cal := a.sess.output.promptCalibration.Load(); cal != nil && cal.compactChars > 0 { |
| 361 | if r := float64(cal.promptTokens) / float64(cal.compactChars); r > 0.05 && r < 2 { |
| 362 | return r |
| 363 | } |
| 364 | } |
| 365 | return fallbackTokPerChar |
| 366 | } |
| 367 | |
| 368 | // msgChars counts the characters that ride to the provider for one message — |
| 369 | // content plus tool-call names and arguments, but not reasoning (stripped on |
| 370 | // send). |
| 371 | func msgChars(m provider.Message) int { |
| 372 | if m.LocalOnly { |
| 373 | return 0 |
| 374 | } |
| 375 | n := len(m.Content) |
| 376 | for _, tc := range m.ToolCalls { |
| 377 | n += len(tc.Name) + len(tc.Arguments) |
| 378 | } |
| 379 | return n |
| 380 | } |
| 381 | |
| 382 | func charsOfMessages(msgs []provider.Message) int { |
| 383 | n := 0 |
| 384 | for _, m := range msgs { |
| 385 | n += msgChars(m) |
| 386 | } |
| 387 | return n |
| 388 | } |
| 389 | |
| 390 | // summarize asks the executor's own provider to distill a replayed prefix into |
| 391 | // a briefing. instructions is optional /compact focus + PreCompact text. |
| 392 | // Named returns so defer can attach RequestCount and still return usage. |
| 393 | func compactionInstructionWithFocus(instructions string) string { |
| 394 | instruction := compactionInstruction |
| 395 | if strings.TrimSpace(instructions) != "" { |
| 396 | instruction += "\n\nAdditional focus for this compaction (prioritize keeping this):\n" + strings.TrimSpace(instructions) |
| 397 | } |
| 398 | return instruction |
| 399 | } |
| 400 | |
| 401 | // summaryRequest builds the exact cache-aligned request shape used by |
| 402 | // summarize. Keeping planning and execution on this shared builder prevents a |
| 403 | // supposedly safe overflow fold from being rejected only after it is selected. |
| 404 | func (a *Agent) summaryRequest(region []provider.Message, instructions string) provider.Request { |
| 405 | prefix := append([]provider.Message(nil), region...) |
| 406 | for i := range prefix { |
| 407 | if !a.imageInput.native && prefix[i].VisionSummary != nil { |
| 408 | prefix[i].ImageInputs = nil |
| 409 | } |
| 410 | } |
| 411 | if len(prefix) == 0 || prefix[0].Role != provider.RoleSystem { |
| 412 | visible := a.modelVisibleMessages() |
| 413 | if len(visible) > 0 && visible[0].Role == provider.RoleSystem { |
| 414 | prefix = append([]provider.Message{visible[0]}, prefix...) |
| 415 | } |
| 416 | } |
| 417 | messages := a.normalizeModelRequestMessages(prefix) |
| 418 | messages = append(messages, HostGeneratedUserMessage(compactionInstructionWithFocus(instructions))) |
| 419 | var schemas []provider.ToolSchema |
| 420 | if a.svc.tools != nil { |
| 421 | schemas = a.providerToolSchemas() |
| 422 | } |
| 423 | return provider.Request{ |
| 424 | Messages: messages, |
| 425 | Tools: schemas, |
| 426 | MaxTokens: a.summaryOutputBudget(), |
| 427 | Temperature: provider.OptionalTemperature(a.temperature), |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | // summarize asks the executor's own provider to distill a replayed prefix into |
| 432 | // a briefing. instructions is optional /compact focus + PreCompact text. |
| 433 | func (a *Agent) summarize(ctx context.Context, region []provider.Message, instructions string) (string, *provider.Usage, error) { |
| 434 | req := a.summaryRequest(region, instructions) |
| 435 | summary, usage, err := a.runSummaryRequest(ctx, req) |
| 436 | a.observeSummaryOutcome(req, usage, err) |
| 437 | return summary, usage, err |
| 438 | } |
| 439 | |
| 440 | // runSummaryRequest admits, sends, and drains one summary request. |
| 441 | // Named returns so defer can attach RequestCount and still return usage. |
| 442 | func (a *Agent) runSummaryRequest(ctx context.Context, req provider.Request) (summary string, usage *provider.Usage, err error) { |
| 443 | if err := ctx.Err(); err != nil { |
| 444 | return "", nil, err |
| 445 | } |
| 446 | defer func(operationCtx context.Context) { err = summaryRequestError(operationCtx, err) }(ctx) |
| 447 | req.Messages, err = a.resolveRequestImages(ctx, req.Messages) |
| 448 | if err != nil { |
| 449 | return "", nil, err |
| 450 | } |
| 451 | ctx, cancel := context.WithCancel(ctx) |
| 452 | defer cancel() |
| 453 | ctx = provider.WithRequestAttemptCounter(ctx) |
| 454 | defer func() { |
| 455 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 456 | if usage != nil && (usage.TotalTokens > 0 || usage.RequestCount > 0) { |
| 457 | a.svc.sink.Emit(event.Event{Kind: event.Usage, ModelRef: a.modelRef, Usage: usage, Pricing: a.svc.pricing, UsageSource: event.UsageSourceCompaction}) |
| 458 | } |
| 459 | }() |
| 460 | defer trackPublishedHostStream(ctx, cancel)() |
| 461 | if err := a.applySummaryAdmissionToRequest(&req); err != nil { |
| 462 | return "", usage, err |
| 463 | } |
| 464 | if budget := a.summaryOutputBudget(); req.MaxTokens > budget { |
| 465 | req.MaxTokens = budget |
| 466 | } |
| 467 | if req.MaxTokens < 256 { |
| 468 | return "", usage, fmt.Errorf("summary output budget too small (%d tokens)", req.MaxTokens) |
| 469 | } |
| 470 | if a.svc.prov == nil { |
| 471 | return "", usage, fmt.Errorf("summary unavailable") |
| 472 | } |
| 473 | if err := ctx.Err(); err != nil { |
| 474 | return "", usage, err |
| 475 | } |
| 476 | ch, err := provider.StreamAuxiliary(ctx, a.svc.prov, req) |
| 477 | if err != nil { |
| 478 | return "", usage, err |
| 479 | } |
| 480 | defer func() { |
| 481 | cancel() |
| 482 | for range ch { |
| 483 | } |
| 484 | }() |
| 485 | |
| 486 | // Cancel on timeout; join the buffer worker before releasing execution ownership. |
| 487 | var b strings.Builder |
| 488 | var reasoning strings.Builder |
| 489 | toolCalls := 0 |
| 490 | for { |
| 491 | select { |
| 492 | case <-ctx.Done(): |
| 493 | return "", usage, ctx.Err() |
| 494 | case chunk, ok := <-ch: |
| 495 | if !ok { |
| 496 | if usage != nil && usage.FinishReason == "length" { |
| 497 | return "", usage, fmt.Errorf("%w: provider reached the output token limit", errSummaryOutputTruncated) |
| 498 | } |
| 499 | s := strings.TrimSpace(b.String()) |
| 500 | if s == "" { |
| 501 | // Thinking providers may answer with reasoning_content only. Surface |
| 502 | // it as the briefing unless the turn also reached for tools: that |
| 503 | // reasoning is private chain-of-thought, not digest material. |
| 504 | r := strings.TrimSpace(reasoning.String()) |
| 505 | if r == "" || toolCalls > 0 { |
| 506 | return "", usage, errSummaryEmpty |
| 507 | } |
| 508 | return truncateUTF8Bytes(r, summaryReasoningMaxBytes), usage, nil |
| 509 | } |
| 510 | return s, usage, nil |
| 511 | } |
| 512 | switch chunk.Type { |
| 513 | case provider.ChunkText: |
| 514 | b.WriteString(chunk.Text) |
| 515 | case provider.ChunkReasoning: |
| 516 | reasoning.WriteString(chunk.Text) |
| 517 | case provider.ChunkToolCall, provider.ChunkToolCallStart: |
| 518 | toolCalls++ |
| 519 | case provider.ChunkUsage: |
| 520 | usage = chunk.Usage |
| 521 | case provider.ChunkError: |
| 522 | return "", usage, chunk.Err |
| 523 | } |
| 524 | } |
| 525 | } |
| 526 | } |
| 527 | |
| 528 | // summarizeOnce performs exactly one application-layer summary request. |
| 529 | // Timeouts, empty results, stream errors, and output truncation all fail once |
| 530 | // with no second attempt. |
| 531 | func (a *Agent) summarizeOnce(ctx context.Context, fold []provider.Message, instructions string) (string, *provider.Usage, error) { |
| 532 | return a.summarize(ctx, fold, instructions) |
| 533 | } |
| 534 | |
| 535 | // renderTranscript flattens messages into a bounded transcript for the |
| 536 | // transcript-form summary request. Tool bodies are the provider-visible |
| 537 | // Content cut to slimToolResultRunes; RawContent never enters a summary. |
| 538 | func renderTranscript(msgs []provider.Message) string { |
| 539 | var b strings.Builder |
| 540 | for _, m := range msgs { |
| 541 | if m.LocalOnly { |
| 542 | continue |
| 543 | } |
| 544 | switch m.Role { |
| 545 | case provider.RoleUser: |
| 546 | fmt.Fprintf(&b, "[user]\n%s\n\n", m.Content) |
| 547 | case provider.RoleAssistant: |
| 548 | if m.Content != "" { |
| 549 | fmt.Fprintf(&b, "[assistant]\n%s\n", m.Content) |
| 550 | } |
| 551 | for _, tc := range m.ToolCalls { |
| 552 | fmt.Fprintf(&b, "[assistant calls %s] %s\n", tc.Name, summarizeToolArgs(tc.Arguments)) |
| 553 | } |
| 554 | b.WriteString("\n") |
| 555 | case provider.RoleTool: |
| 556 | fmt.Fprintf(&b, "[tool %s result]\n%s\n\n", m.Name, slimToolResult(m.Content)) |
| 557 | case provider.RoleSystem: |
| 558 | fmt.Fprintf(&b, "[system]\n%s\n\n", m.Content) |
| 559 | } |
| 560 | } |
| 561 | return b.String() |
| 562 | } |
| 563 | |
| 564 | // summarizeToolArgs returns a short summary of tool-call arguments instead of |
| 565 | // the full JSON. This prevents the summarizer from reproducing long argument |
| 566 | // text (like sub-agent task prompts) in the compaction summary, which would |
| 567 | // leak into the session as a user message (#4317). |
| 568 | func summarizeToolArgs(args string) string { |
| 569 | if args == "" { |
| 570 | return "(no arguments)" |
| 571 | } |
| 572 | var parsed map[string]any |
| 573 | if err := json.Unmarshal([]byte(args), &parsed); err != nil { |
| 574 | // Not valid JSON — return a length hint instead of raw text. |
| 575 | return fmt.Sprintf("(%d bytes)", len(args)) |
| 576 | } |
| 577 | keys := make([]string, 0, len(parsed)) |
| 578 | for k := range parsed { |
| 579 | keys = append(keys, k) |
| 580 | } |
| 581 | sort.Strings(keys) |
| 582 | return fmt.Sprintf("{%s} (%d keys)", strings.Join(keys, ", "), len(parsed)) |
| 583 | } |
| 584 |