| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "os" |
| 9 | "path/filepath" |
| 10 | "sort" |
| 11 | "strings" |
| 12 | "time" |
| 13 | "unicode/utf8" |
| 14 | |
| 15 | "reasonix/internal/ablation" |
| 16 | "reasonix/internal/event" |
| 17 | "reasonix/internal/provider" |
| 18 | ) |
| 19 | |
| 20 | // Compaction is a low-frequency cache-reset point: the prompt grows append-only |
| 21 | // (high cache hits) until a turn nears compactRatio of the window, then it is |
| 22 | // compacted down to a tail budget. The budget is a fixed token count, not a |
| 23 | // fraction of the window, so a huge window still compacts rarely while a small |
| 24 | // one still lands below the trigger (which is what stops the re-compaction loop). |
| 25 | const ( |
| 26 | defaultSoftCompactRatio = 0.5 // report growing context here, but keep the cache-stable prefix intact |
| 27 | defaultToolResultSnipRatio = 0.6 // rewrite stale tool results cheaply before summary compaction |
| 28 | defaultCompactRatio = 0.8 // trigger: prompt at this fraction of the window compacts |
| 29 | defaultCompactForceRatio = 0.9 // force compaction at this high-water mark even for low-value folds |
| 30 | defaultCompactTarget = 0.5 // safety cap: the kept tail never exceeds this fraction of the window |
| 31 | defaultTailTokens = 16384 // verbatim recent-tail budget, in tokens |
| 32 | minRecentKeep = 2 // never keep fewer recent messages than this |
| 33 | minCompactMessages = 2 // skip compaction below this many compactable messages |
| 34 | fallbackTokPerChar = 0.25 // ~4 chars/token, used before any usage is available to calibrate |
| 35 | maxPinnedFirstUserTokens = 1500 // ceiling on pinning the first user turn verbatim; larger first turns (pasted content) stay foldable |
| 36 | pinnedFirstUserWindowFrac = 0.15 // and never pin a first turn worth more than this fraction of the window |
| 37 | ) |
| 38 | |
| 39 | // summaryTag wraps the compaction summary so the model can distinguish it from |
| 40 | // live user input and later strip or skip it when reasoning about the current turn. |
| 41 | const ( |
| 42 | summaryTagOpen = "<compaction-summary>" |
| 43 | summaryTagClose = "</compaction-summary>" |
| 44 | ) |
| 45 | |
| 46 | // summaryTimeout bounds one summarizer call so a stalled stream surfaces a clear |
| 47 | // failure (then a mechanical fold) instead of hanging compaction indefinitely. |
| 48 | const summaryTimeout = 90 * time.Second |
| 49 | |
| 50 | // summarySystemPrompt steers the executor to distill older history into a |
| 51 | // structured briefing it can keep relying on after the originals are dropped. |
| 52 | // The section layout mirrors what a coding agent actually needs to resume work |
| 53 | // mid-task: the goal verbatim, the concrete state of the code, and an explicit |
| 54 | // next step — so the post-compaction turn doesn't lose the thread or re-derive |
| 55 | // decisions already made. |
| 56 | const summarySystemPrompt = `You are compacting the earlier part of a coding agent's conversation to save context. |
| 57 | The agent keeps your summary alongside the user's own turns (kept verbatim) and the recent tail; your job is to fold the assistant/tool work into a briefing it can resume from. |
| 58 | Write under these exact headings, omitting a heading only if it has no content: |
| 59 | |
| 60 | ## Standing facts & constraints |
| 61 | 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. |
| 62 | |
| 63 | ## Goal |
| 64 | The user's request and intent. |
| 65 | |
| 66 | ## Decisions & rationale |
| 67 | Key choices made so far and why — so they are not re-litigated or reversed. |
| 68 | |
| 69 | ## Files & code |
| 70 | 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. |
| 71 | |
| 72 | ## Commands & outcomes |
| 73 | Commands run (builds, tests, git) and their relevant results — what passed, what failed, and the error text that matters. |
| 74 | |
| 75 | ## Errors & fixes |
| 76 | Problems hit and how they were resolved (or not), so the same dead ends are not repeated. |
| 77 | |
| 78 | ## Pending & next step |
| 79 | What is still in progress or unstarted, and the single most concrete next action to take. |
| 80 | |
| 81 | Rules: be terse — bullet points and fragments, not prose. Preserve identifiers, paths, and numbers exactly. Do NOT invent anything not present in the messages; if something is unknown, leave it out rather than guessing.` |
| 82 | |
| 83 | // compactThresholds returns the prompt-token boundaries maybeCompact switches |
| 84 | // on. The compaction ablation arm collapses the snip and fold triggers onto |
| 85 | // soft, so the cache-preserving deferral branch is unreachable and the session |
| 86 | // folds as soon as it grows — what a harness with no prompt-cache strategy does. |
| 87 | func (a *Agent) compactThresholds() (soft, snip, high int) { |
| 88 | high = int(float64(a.contextWindow) * a.compactRatio) |
| 89 | snip = int(float64(a.contextWindow) * a.toolResultSnipRatio) |
| 90 | soft = int(float64(a.contextWindow) * a.softCompactRatio) |
| 91 | if a.ablation.Off(ablation.Compaction) { |
| 92 | high, snip = soft, soft |
| 93 | } |
| 94 | return soft, snip, high |
| 95 | } |
| 96 | |
| 97 | // maybeCompact compacts the session when the last turn's prompt has grown to the |
| 98 | // configured fraction of the context window. It is a no-op when compaction is |
| 99 | // disabled (no window) or usage is unavailable. |
| 100 | func (a *Agent) maybeCompact(ctx context.Context, u *provider.Usage) { |
| 101 | if a.contextWindow <= 0 || u == nil || u.PromptTokens == 0 { |
| 102 | return |
| 103 | } |
| 104 | soft, snip, high := a.compactThresholds() |
| 105 | // A turn that sits under the trigger is the breathing room a healthy |
| 106 | // compaction buys; it clears the stuck latch and the run counter. This has to |
| 107 | // happen before the soft/snip branches return, because a compaction that |
| 108 | // settles the prompt anywhere in [snip, high) is working exactly as intended: |
| 109 | // leaving a stale run count behind there would latch the *next* compaction as |
| 110 | // "the window is too small" and silently disable auto-compaction for the rest |
| 111 | // of the session. |
| 112 | if u.PromptTokens < high { |
| 113 | a.consecutiveCompacts = 0 |
| 114 | a.compactStuck = false |
| 115 | } |
| 116 | // Between the soft ratio and the trigger, report growing context once without |
| 117 | // rewriting the prefix — a compaction here would needlessly crater the cache. |
| 118 | if u.PromptTokens >= soft && u.PromptTokens < snip && !a.softCompactNoticed { |
| 119 | a.softCompactNoticed = true |
| 120 | detail := fmt.Sprintf("context reached %.0f%% of window; keeping cache-first prefix until compact threshold %.0f%%", a.softCompactRatio*100, a.compactRatio*100) |
| 121 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: "Context is getting large; preserving cache until cleanup is needed.", Detail: detail}) |
| 122 | return |
| 123 | } |
| 124 | if u.PromptTokens >= snip && u.PromptTokens < high { |
| 125 | ratio := a.tokPerChar() |
| 126 | if st, err := a.SnipStaleToolResults(); err == nil && st.Results > 0 { |
| 127 | saved := int(float64(st.SavedChars) * ratio) |
| 128 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: fmt.Sprintf( |
| 129 | "snipped %d stale tool results (~%d tokens est.) before compaction", st.Results, saved)}) |
| 130 | } |
| 131 | return |
| 132 | } |
| 133 | if u.PromptTokens < high { |
| 134 | return // the latch was already cleared above |
| 135 | } |
| 136 | if a.compactStuck { |
| 137 | return |
| 138 | } |
| 139 | force := u.PromptTokens >= int(float64(a.contextWindow)*a.compactForceRatio) |
| 140 | // Prune before folding: when eliding stale tool results alone clears the |
| 141 | // trigger, this turn's (paid) summarize call is skipped entirely. |
| 142 | ratio := a.tokPerChar() |
| 143 | if st, err := a.PruneStaleToolResults(); err == nil && st.Results > 0 { |
| 144 | saved := int(float64(st.SavedChars) * ratio) |
| 145 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: fmt.Sprintf( |
| 146 | "pruned %d stale tool results (~%d tokens est.) before compaction", st.Results, saved)}) |
| 147 | if !force && u.PromptTokens-saved < high { |
| 148 | return |
| 149 | } |
| 150 | } |
| 151 | if err := a.compact(ctx, "auto", "", force); err != nil { |
| 152 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: "Context cleanup skipped for now.", Detail: fmt.Sprintf("compaction skipped: %v", err)}) |
| 153 | return |
| 154 | } |
| 155 | // A healthy compaction drops the prompt under the trigger, so the next turn |
| 156 | // won't compact. Compacting on consecutive turns means the kept tail alone |
| 157 | // exceeds the trigger — the system prompt plus one verbatim turn is bigger than |
| 158 | // the window allows. Re-firing every turn is the loop users hit, so pause |
| 159 | // auto-compaction and say why, once. |
| 160 | a.consecutiveCompacts++ |
| 161 | if a.consecutiveCompacts >= 2 { |
| 162 | a.compactStuck = true |
| 163 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: "Automatic context cleanup paused because the context window is too small.", Detail: fmt.Sprintf( |
| 164 | "context_window=%d is too small for compaction to help (the system prompt plus one turn already exceeds %.0f%% of it); raise context_window or shrink tool output. Auto-compaction paused until the prompt drops.", |
| 165 | a.contextWindow, a.compactRatio*100)}) |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | // foldEconomics estimates whether compacting the given region saves enough |
| 170 | // tokens to justify the summarization API call. It returns false when the |
| 171 | // region is too small for the savings to outweigh the extra round-trip cost |
| 172 | // and latency of calling the summarizer. |
| 173 | func foldEconomics(region []provider.Message) bool { |
| 174 | const minFoldTokens = 400 |
| 175 | return estimateMessagesTokens(region) >= minFoldTokens |
| 176 | } |
| 177 | |
| 178 | func estimateMessagesTokens(msgs []provider.Message) int { |
| 179 | total := 0 |
| 180 | for _, m := range msgs { |
| 181 | if m.LocalOnly { |
| 182 | continue |
| 183 | } |
| 184 | total += 4 // chat-message framing overhead |
| 185 | total += estimateTextTokens(m.Content) |
| 186 | total += estimateTextTokens(m.ReasoningContent) |
| 187 | total += estimateTextTokens(m.Name) |
| 188 | total += estimateTextTokens(m.ToolCallID) |
| 189 | for _, tc := range m.ToolCalls { |
| 190 | total += 8 |
| 191 | total += estimateTextTokens(tc.ID) |
| 192 | total += estimateTextTokens(tc.Name) |
| 193 | total += estimateTextTokens(tc.Arguments) |
| 194 | } |
| 195 | for _, item := range m.ResponsesItems { |
| 196 | total += estimateTextTokens(string(item)) |
| 197 | } |
| 198 | } |
| 199 | return total |
| 200 | } |
| 201 | |
| 202 | func estimateTextTokens(s string) int { |
| 203 | if s == "" { |
| 204 | return 0 |
| 205 | } |
| 206 | // A conservative cross-language approximation: English-ish text trends near |
| 207 | // four bytes per token, while CJK-heavy text is closer to one rune per token. |
| 208 | bytes := len(s) |
| 209 | runes := utf8.RuneCountInString(s) |
| 210 | byBytes := (bytes + 3) / 4 |
| 211 | if runes > byBytes { |
| 212 | return runes |
| 213 | } |
| 214 | return byBytes |
| 215 | } |
| 216 | |
| 217 | // compact summarizes the older middle of the session and replaces it in place: |
| 218 | // the session becomes system + summary + recent tail. The dropped originals are |
| 219 | // archived first, so the full history stays traceable. trigger is "auto" (the |
| 220 | // window threshold) or "manual" (/compact); it rides the Compaction events so a |
| 221 | // frontend can label the card. instructions is optional extra summary guidance |
| 222 | // (the user's `/compact <focus>` text); a PreCompact hook can contribute more. |
| 223 | // force bypasses the fold-economics skip (manual /compact and the force-ratio |
| 224 | // high-water mark always compact). A Started event is emitted before the (network) |
| 225 | // summarize so the UI can show a "compacting…" placeholder, and a Done event |
| 226 | // (carrying the summary) replaces it. |
| 227 | func (a *Agent) compact(ctx context.Context, trigger, instructions string, force bool) error { |
| 228 | msgs := a.session.Messages |
| 229 | head, start, ok := a.planCompaction(msgs, minCompactMessages) |
| 230 | if !ok { |
| 231 | // A single huge message can still be worth folding. Keep the normal |
| 232 | // message-count guard for small histories, but let content size decide |
| 233 | // whether a one-message region has real compaction value. |
| 234 | head, start, ok = a.planCompaction(msgs, 1) |
| 235 | } |
| 236 | if !ok { |
| 237 | return nil // recent tail already covers everything worth keeping |
| 238 | } |
| 239 | // A controller in-flight marker records the pre-turn message count, but a |
| 240 | // compaction rewrites message indexes. Keep the entire active turn outside |
| 241 | // the fold so completed tool call/result pairs remain available for a later |
| 242 | // cancellation or crash recovery instead of surviving only as prose in a |
| 243 | // summary. |
| 244 | if active := a.activeTurnStart(msgs); active >= head && active < start { |
| 245 | start = active |
| 246 | if start <= head { |
| 247 | return nil |
| 248 | } |
| 249 | } |
| 250 | region := msgs[head:start] |
| 251 | |
| 252 | // Base layer: every small user turn in the region is kept verbatim (the |
| 253 | // deterministic floor — a fact the user stated is never summarized away, |
| 254 | // wherever in the session they said it); only the rest folds into the digest. |
| 255 | kept, fold := a.partitionFold(region) |
| 256 | if len(fold) == 0 { |
| 257 | return nil // nothing but kept user turns — a fold would save nothing |
| 258 | } |
| 259 | |
| 260 | // Economic check on the foldable part (kept user turns don't count toward the |
| 261 | // savings): skip if too small to justify the call, unless force demands it. |
| 262 | if !force && !foldEconomics(fold) { |
| 263 | return nil |
| 264 | } |
| 265 | |
| 266 | a.sink.Emit(event.Event{Kind: event.CompactionStarted, Compaction: event.Compaction{Trigger: trigger}}) |
| 267 | |
| 268 | // A PreCompact hook can steer what the summary keeps; its stdout joins any |
| 269 | // explicit /compact <focus> text. |
| 270 | if a.hooks != nil { |
| 271 | if hookInstr := a.hooks.PreCompact(ctx, trigger); hookInstr != "" { |
| 272 | if instructions != "" { |
| 273 | instructions += "\n" |
| 274 | } |
| 275 | instructions += hookInstr |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | // compaction.prepare: extensions rule on the fold and the accumulated |
| 280 | // guidance (hook + /compact focus) for THIS pass only. A block skips the |
| 281 | // pass; the caller surfaces the reason through its usual notice path. |
| 282 | var err error |
| 283 | fold, instructions, err = a.interceptCompactionPrepare(ctx, fold, instructions) |
| 284 | if err != nil { |
| 285 | a.emitCompactionAborted(trigger) |
| 286 | return err |
| 287 | } |
| 288 | if len(fold) == 0 { |
| 289 | a.emitCompactionAborted(trigger) |
| 290 | return nil // the extension replaced the fold with nothing to fold |
| 291 | } |
| 292 | |
| 293 | archived := "" |
| 294 | if a.archiveDir != "" { |
| 295 | path, err := archiveMessages(a.archiveDir, fold) |
| 296 | if err != nil { |
| 297 | a.emitCompactionAborted(trigger) |
| 298 | return fmt.Errorf("archive: %w", err) |
| 299 | } |
| 300 | archived = path |
| 301 | } |
| 302 | |
| 303 | // The digest covers only the foldable work; kept user turns and prior digests |
| 304 | // are spliced back verbatim, so a fact that reached a digest once is never |
| 305 | // re-summarized away and the user's own words are never touched. Digests |
| 306 | // accumulate (small) rather than collapsing into one lossy rolling summary. |
| 307 | summary, err := a.summarizeWithRetry(ctx, fold, instructions) |
| 308 | if err != nil { |
| 309 | // Mechanical fold: the foldable region is already archived, so stand in a |
| 310 | // deterministic marker rather than aborting. /compact then always frees |
| 311 | // context (and auto-compaction can't loop on a still-full window); the |
| 312 | // verbatim user turns kept above are untouched. |
| 313 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: "Context was compacted without a generated summary.", Detail: "compaction summary unavailable (" + err.Error() + "); folded mechanically"}) |
| 314 | summary = mechanicalFoldDigest(len(fold), archived) |
| 315 | } |
| 316 | |
| 317 | // compaction.complete: extensions rule on the produced summary before it |
| 318 | // is written into the session; a replacement is persisted as the summary. |
| 319 | summary, err = a.interceptCompactionComplete(ctx, summary) |
| 320 | if err != nil { |
| 321 | a.emitCompactionAborted(trigger) |
| 322 | return err |
| 323 | } |
| 324 | |
| 325 | compacted := make([]provider.Message, 0, head+len(kept)+1+len(msgs)-start) |
| 326 | compacted = append(compacted, msgs[:head]...) |
| 327 | compacted = append(compacted, kept...) |
| 328 | compacted = append(compacted, provider.Message{ |
| 329 | Role: provider.RoleUser, |
| 330 | Content: summaryTagOpen + "\n" + |
| 331 | "Summary of earlier conversation (older messages were compacted to save context):\n" + |
| 332 | summary + "\n" + |
| 333 | summaryTagClose, |
| 334 | }) |
| 335 | compacted = append(compacted, msgs[start:]...) |
| 336 | a.session.Rewrite(compacted, "compact_"+trigger) |
| 337 | |
| 338 | a.sink.Emit(event.Event{Kind: event.CompactionDone, Compaction: event.Compaction{ |
| 339 | Trigger: trigger, Messages: len(fold), Summary: summary, Archive: archived, |
| 340 | }}) |
| 341 | return nil |
| 342 | } |
| 343 | |
| 344 | // emitCompactionAborted resolves a "compacting…" placeholder when a pass fails |
| 345 | // after the Started event: a Done with no summary tells a frontend to drop the |
| 346 | // placeholder. The caller still surfaces the reason (a Notice), so this carries |
| 347 | // no text of its own. |
| 348 | func (a *Agent) emitCompactionAborted(trigger string) { |
| 349 | a.sink.Emit(event.Event{Kind: event.CompactionDone, Compaction: event.Compaction{Trigger: trigger}}) |
| 350 | } |
| 351 | |
| 352 | // SummarizeFrom replaces the messages from fromIdx onward with a single summary, |
| 353 | // keeping everything before it verbatim ("summarize from here"). fromIdx is a turn |
| 354 | // boundary (a user message), so the split never severs a tool_call/result pair — |
| 355 | // those live within one turn. A no-op when the region is empty. |
| 356 | func (a *Agent) SummarizeFrom(ctx context.Context, fromIdx int) error { |
| 357 | msgs := a.session.Messages |
| 358 | if fromIdx < 0 || fromIdx >= len(msgs) { |
| 359 | return nil |
| 360 | } |
| 361 | region, localOnly := splitLocalOnlyMessages(msgs[fromIdx:]) |
| 362 | if len(region) == 0 { |
| 363 | return nil |
| 364 | } |
| 365 | if a.archiveDir != "" { |
| 366 | _, _ = archiveMessages(a.archiveDir, region) // best-effort traceability |
| 367 | } |
| 368 | summary, err := a.summarize(ctx, region, "") |
| 369 | if err != nil { |
| 370 | return err |
| 371 | } |
| 372 | next := make([]provider.Message, 0, fromIdx+1+len(localOnly)) |
| 373 | next = append(next, msgs[:fromIdx]...) |
| 374 | next = append(next, provider.Message{ |
| 375 | Role: provider.RoleUser, |
| 376 | Content: "Summary of the later conversation (compacted from here on):\n" + summary, |
| 377 | }) |
| 378 | next = append(next, localOnly...) |
| 379 | a.session.Rewrite(next, "summarize_from") |
| 380 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, |
| 381 | Text: fmt.Sprintf("summarized %d later messages → summary", len(region))}) |
| 382 | return nil |
| 383 | } |
| 384 | |
| 385 | // SummarizeUpTo replaces the messages before toIdx (after the system prompt) with |
| 386 | // a single summary, keeping toIdx onward verbatim ("summarize up to here"). toIdx |
| 387 | // is a turn boundary, so no tool pair is split. A no-op when the region is empty. |
| 388 | func (a *Agent) SummarizeUpTo(ctx context.Context, toIdx int) error { |
| 389 | msgs := a.session.Messages |
| 390 | head := 0 |
| 391 | if len(msgs) > 0 && msgs[0].Role == provider.RoleSystem { |
| 392 | head = 1 |
| 393 | } |
| 394 | if toIdx <= head || toIdx > len(msgs) { |
| 395 | return nil |
| 396 | } |
| 397 | region, localOnly := splitLocalOnlyMessages(msgs[head:toIdx]) |
| 398 | if len(region) == 0 { |
| 399 | return nil |
| 400 | } |
| 401 | if a.archiveDir != "" { |
| 402 | _, _ = archiveMessages(a.archiveDir, region) |
| 403 | } |
| 404 | summary, err := a.summarize(ctx, region, "") |
| 405 | if err != nil { |
| 406 | return err |
| 407 | } |
| 408 | next := make([]provider.Message, 0, head+1+len(localOnly)+len(msgs)-toIdx) |
| 409 | next = append(next, msgs[:head]...) |
| 410 | next = append(next, provider.Message{ |
| 411 | Role: provider.RoleUser, |
| 412 | Content: "Summary of earlier conversation (compacted up to here):\n" + summary, |
| 413 | }) |
| 414 | next = append(next, localOnly...) |
| 415 | next = append(next, msgs[toIdx:]...) |
| 416 | a.session.Rewrite(next, "summarize_up_to") |
| 417 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, |
| 418 | Text: fmt.Sprintf("summarized %d earlier messages → summary", len(region))}) |
| 419 | return nil |
| 420 | } |
| 421 | |
| 422 | // IsCompactionSummary reports whether m is a rolling digest inserted by a |
| 423 | // prior compaction fold. Exported for session owners outside this package |
| 424 | // (e.g. the guardian) whose turn rollback must not treat a digest as a |
| 425 | // disposable user message. |
| 426 | func IsCompactionSummary(m provider.Message) bool { return isCompactionSummary(m) } |
| 427 | |
| 428 | func (a *Agent) activeTurnStart(msgs []provider.Message) int { |
| 429 | createdAt := a.activeTurnCreatedAt.Load() |
| 430 | if createdAt == 0 { |
| 431 | return -1 |
| 432 | } |
| 433 | for i, m := range msgs { |
| 434 | if m.Role == provider.RoleUser && m.CreatedAt == createdAt { |
| 435 | return i |
| 436 | } |
| 437 | } |
| 438 | return -1 |
| 439 | } |
| 440 | |
| 441 | // splitLocalOnlyMessages removes display-only interrupted output from the |
| 442 | // summarizer/archive input while returning it in transcript order for durable |
| 443 | // reattachment. Explicit range summaries are user-requested rewrites, but they |
| 444 | // must not erase visible output or expose private partial reasoning to a model. |
| 445 | func splitLocalOnlyMessages(msgs []provider.Message) (model, localOnly []provider.Message) { |
| 446 | for _, m := range msgs { |
| 447 | if m.LocalOnly { |
| 448 | localOnly = append(localOnly, m) |
| 449 | continue |
| 450 | } |
| 451 | model = append(model, m) |
| 452 | } |
| 453 | return model, localOnly |
| 454 | } |
| 455 | |
| 456 | // isCompactionSummary reports whether m is a rolling summary from a prior fold. |
| 457 | func isCompactionSummary(m provider.Message) bool { |
| 458 | return m.Role == provider.RoleUser && |
| 459 | strings.HasPrefix(strings.TrimLeft(m.Content, "\n "), summaryTagOpen) |
| 460 | } |
| 461 | |
| 462 | // pinnedPrefixLen counts the leading messages a fold keeps verbatim: the system |
| 463 | // prompt, the first user turn (its task + stated facts/constraints) when it is |
| 464 | // small enough to be a brief, and any prior summaries — so a fold never |
| 465 | // summarizes the user's facts away, and a later fold never re-summarizes an |
| 466 | // earlier summary into nothing (the drift that silently dropped user-stated facts |
| 467 | // after the second compaction). A large first turn (pasted content) stays |
| 468 | // foldable so pinning never starves the window. |
| 469 | func (a *Agent) pinnedPrefixLen(msgs []provider.Message) int { |
| 470 | i := 0 |
| 471 | if i < len(msgs) && msgs[i].Role == provider.RoleSystem { |
| 472 | i++ |
| 473 | } |
| 474 | if i < len(msgs) && msgs[i].Role == provider.RoleUser && !isCompactionSummary(msgs[i]) && a.pinnableUserTurn(msgs[i]) { |
| 475 | i++ |
| 476 | } |
| 477 | for i < len(msgs) && isCompactionSummary(msgs[i]) { |
| 478 | i++ |
| 479 | } |
| 480 | return i |
| 481 | } |
| 482 | |
| 483 | // pinnableUserTurn reports whether a user turn is small enough to keep verbatim. A |
| 484 | // turn larger than a brief (pasted content) folds like any other message so the |
| 485 | // kept-verbatim floor never starves the window. |
| 486 | func (a *Agent) pinnableUserTurn(m provider.Message) bool { |
| 487 | budget := maxPinnedFirstUserTokens |
| 488 | if a.contextWindow > 0 { |
| 489 | if f := int(float64(a.contextWindow) * pinnedFirstUserWindowFrac); f < budget { |
| 490 | budget = f |
| 491 | } |
| 492 | } |
| 493 | return int(float64(msgChars(m))*a.tokPerChar()) <= budget |
| 494 | } |
| 495 | |
| 496 | // partitionFold splits a compaction region into what is kept verbatim — small user |
| 497 | // turns (a fact the user stated is never summarized away) and prior digests (so a |
| 498 | // later fold never re-summarizes an earlier digest and drops the facts it already |
| 499 | // captured) — and the rest, which folds. Order within each group is preserved. |
| 500 | func (a *Agent) partitionFold(region []provider.Message) (kept, fold []provider.Message) { |
| 501 | policyKeep := keepIndexes(region, a.keepPolicy) |
| 502 | for i, m := range region { |
| 503 | if m.LocalOnly || policyKeep[i] || isCompactionSummary(m) || (m.Role == provider.RoleUser && a.pinnableUserTurn(m)) { |
| 504 | kept = append(kept, m) |
| 505 | } else { |
| 506 | fold = append(fold, m) |
| 507 | } |
| 508 | } |
| 509 | return kept, fold |
| 510 | } |
| 511 | |
| 512 | func keepIndexes(region []provider.Message, policy KeepPolicy) []bool { |
| 513 | keep := make([]bool, len(region)) |
| 514 | policyStart := 0 |
| 515 | for i, m := range region { |
| 516 | if isCompactionSummary(m) { |
| 517 | policyStart = i + 1 |
| 518 | } |
| 519 | } |
| 520 | // Retention applies only to messages since the latest digest; older kept |
| 521 | // messages are allowed to fold on the next pass so they cannot grow forever. |
| 522 | for i, m := range region { |
| 523 | if i >= policyStart && shouldKeepMessage(m, policy) { |
| 524 | keep[i] = true |
| 525 | } |
| 526 | } |
| 527 | for i, m := range region { |
| 528 | if !keep[i] { |
| 529 | continue |
| 530 | } |
| 531 | switch m.Role { |
| 532 | case provider.RoleTool: |
| 533 | if j := findToolCaller(region, i, m.ToolCallID); j >= 0 { |
| 534 | keepToolCallGroup(region, keep, j) |
| 535 | } |
| 536 | case provider.RoleAssistant: |
| 537 | keepToolCallGroup(region, keep, i) |
| 538 | } |
| 539 | } |
| 540 | return keep |
| 541 | } |
| 542 | |
| 543 | func keepToolCallGroup(region []provider.Message, keep []bool, assistantIndex int) { |
| 544 | if assistantIndex < 0 || assistantIndex >= len(region) { |
| 545 | return |
| 546 | } |
| 547 | m := region[assistantIndex] |
| 548 | if m.Role != provider.RoleAssistant || len(m.ToolCalls) == 0 { |
| 549 | return |
| 550 | } |
| 551 | keep[assistantIndex] = true |
| 552 | ids := toolCallIDs(m) |
| 553 | for j := assistantIndex + 1; j < len(region) && region[j].Role == provider.RoleTool; j++ { |
| 554 | if ids[region[j].ToolCallID] { |
| 555 | keep[j] = true |
| 556 | } |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | func shouldKeepMessage(m provider.Message, policy KeepPolicy) bool { |
| 561 | if policy&KeepErrors != 0 && isErrorMessage(m) { |
| 562 | return true |
| 563 | } |
| 564 | if policy&KeepUserMarked != 0 && isUserMarked(m) { |
| 565 | return true |
| 566 | } |
| 567 | return false |
| 568 | } |
| 569 | |
| 570 | func isErrorMessage(m provider.Message) bool { |
| 571 | if m.Role != provider.RoleTool { |
| 572 | return false |
| 573 | } |
| 574 | s := strings.TrimSpace(strings.ToLower(m.Content)) |
| 575 | return strings.HasPrefix(s, "error:") || strings.HasPrefix(s, "blocked:") |
| 576 | } |
| 577 | |
| 578 | func isUserMarked(m provider.Message) bool { |
| 579 | if m.Role != provider.RoleUser { |
| 580 | return false |
| 581 | } |
| 582 | content := strings.TrimSpace(strings.ToLower(m.Content)) |
| 583 | return strings.HasPrefix(content, "[[keep]]") || |
| 584 | strings.HasPrefix(content, "[keep]") || |
| 585 | strings.HasPrefix(content, "<keep>") || |
| 586 | strings.HasPrefix(content, "<!-- keep -->") |
| 587 | } |
| 588 | |
| 589 | func findToolCaller(region []provider.Message, toolIndex int, id string) int { |
| 590 | for i := toolIndex - 1; i >= 0; i-- { |
| 591 | if region[i].Role != provider.RoleAssistant { |
| 592 | continue |
| 593 | } |
| 594 | for _, tc := range region[i].ToolCalls { |
| 595 | if tc.ID == id { |
| 596 | return i |
| 597 | } |
| 598 | } |
| 599 | } |
| 600 | return -1 |
| 601 | } |
| 602 | |
| 603 | func toolCallIDs(m provider.Message) map[string]bool { |
| 604 | ids := make(map[string]bool, len(m.ToolCalls)) |
| 605 | for _, tc := range m.ToolCalls { |
| 606 | ids[tc.ID] = true |
| 607 | } |
| 608 | return ids |
| 609 | } |
| 610 | |
| 611 | // planCompaction locates the region to summarize. head is the count of leading |
| 612 | // messages preserved verbatim (see pinnedPrefixLen); start is where the preserved |
| 613 | // recent tail begins, so msgs[head:start] is compacted. The tail is bounded by a |
| 614 | // token budget (not a message count), so a few large tool outputs can't keep it |
| 615 | // above the trigger and re-fire compaction every turn. ok is false when there is |
| 616 | // too little to compact. |
| 617 | func (a *Agent) planCompaction(msgs []provider.Message, min int) (head, start int, ok bool) { |
| 618 | head = a.pinnedPrefixLen(msgs) |
| 619 | if a.contextWindow > 0 { |
| 620 | budget := defaultTailTokens |
| 621 | if maxByWin := int(float64(a.contextWindow) * defaultCompactTarget); maxByWin < budget { |
| 622 | budget = maxByWin |
| 623 | } |
| 624 | start = tailStart(msgs, head, budget, a.tokPerChar(), a.tailFloor()) |
| 625 | } else { |
| 626 | // No window to budget against (manual /compact on an unconfigured |
| 627 | // provider): keep a fixed count of recent messages, aligned off any tool. |
| 628 | start = len(msgs) - a.tailFloor() |
| 629 | for start > head && msgs[start].Role == provider.RoleTool { |
| 630 | start-- |
| 631 | } |
| 632 | } |
| 633 | if start < head { |
| 634 | start = head |
| 635 | } |
| 636 | if start-head < min { |
| 637 | return head, start, false |
| 638 | } |
| 639 | return head, start, true |
| 640 | } |
| 641 | |
| 642 | func (a *Agent) tailFloor() int { |
| 643 | if a.recentKeep > minRecentKeep { |
| 644 | return a.recentKeep |
| 645 | } |
| 646 | return minRecentKeep |
| 647 | } |
| 648 | |
| 649 | // tailStart walks newest→oldest, growing the verbatim tail until the next |
| 650 | // message would push its token estimate past budgetTokens (but never below |
| 651 | // minKeep messages), then aligns the boundary back off any tool result so the |
| 652 | // tail never begins with an orphan whose assistant tool_calls were summarized |
| 653 | // away. |
| 654 | func tailStart(msgs []provider.Message, head, budgetTokens int, tokPerChar float64, minKeep int) int { |
| 655 | start := len(msgs) |
| 656 | acc := 0 |
| 657 | for i := len(msgs) - 1; i > head; i-- { |
| 658 | c := int(float64(msgChars(msgs[i])) * tokPerChar) |
| 659 | if len(msgs)-i > minKeep && acc+c > budgetTokens { |
| 660 | break |
| 661 | } |
| 662 | acc += c |
| 663 | start = i |
| 664 | } |
| 665 | // start == len(msgs) when nothing fit the tail (a session too small to have a |
| 666 | // message after head); there is no msgs[start] to align off, and the caller's |
| 667 | // minCompactMessages check then no-ops the pass. |
| 668 | for start > head && start < len(msgs) && msgs[start].Role == provider.RoleTool { |
| 669 | start-- |
| 670 | } |
| 671 | return start |
| 672 | } |
| 673 | |
| 674 | // tokPerChar derives a tokens-per-character ratio from the last turn's real |
| 675 | // usage so per-message estimates track the provider's tokenizer without a local |
| 676 | // one. Reasoning content is excluded from the char count to match the prompt |
| 677 | // actually sent (the provider strips it). Falls back to ~4 chars/token before |
| 678 | // any usage is known, and ignores absurd ratios. |
| 679 | func (a *Agent) tokPerChar() float64 { |
| 680 | if u := a.lastUsage.Load(); u != nil && u.PromptTokens > 0 { |
| 681 | if c := charsOfMessages(a.session.Messages); c > 0 { |
| 682 | if r := float64(u.PromptTokens) / float64(c); r > 0.05 && r < 2 { |
| 683 | return r |
| 684 | } |
| 685 | } |
| 686 | } |
| 687 | return fallbackTokPerChar |
| 688 | } |
| 689 | |
| 690 | // msgChars counts the characters that ride to the provider for one message — |
| 691 | // content plus tool-call names and arguments, but not reasoning (stripped on |
| 692 | // send). |
| 693 | func msgChars(m provider.Message) int { |
| 694 | if m.LocalOnly { |
| 695 | return 0 |
| 696 | } |
| 697 | n := len(m.Content) |
| 698 | for _, tc := range m.ToolCalls { |
| 699 | n += len(tc.Name) + len(tc.Arguments) |
| 700 | } |
| 701 | return n |
| 702 | } |
| 703 | |
| 704 | func charsOfMessages(msgs []provider.Message) int { |
| 705 | n := 0 |
| 706 | for _, m := range msgs { |
| 707 | n += msgChars(m) |
| 708 | } |
| 709 | return n |
| 710 | } |
| 711 | |
| 712 | // summarize asks the executor's own provider (no tools) to distill the region |
| 713 | // into a briefing, returning the collected text. instructions, when non-empty, |
| 714 | // is appended to the system prompt as extra focus guidance (from /compact <focus> |
| 715 | // and/or a PreCompact hook). |
| 716 | func (a *Agent) summarize(ctx context.Context, region []provider.Message, instructions string) (string, error) { |
| 717 | ctx, cancel := context.WithTimeout(ctx, summaryTimeout) |
| 718 | defer cancel() |
| 719 | ctx = provider.WithRequestAttemptCounter(ctx) |
| 720 | sys := summarySystemPrompt |
| 721 | if strings.TrimSpace(instructions) != "" { |
| 722 | sys += "\n\nAdditional focus for this compaction (prioritize keeping this):\n" + strings.TrimSpace(instructions) |
| 723 | } |
| 724 | var usage *provider.Usage |
| 725 | defer func() { |
| 726 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 727 | if usage != nil && (usage.TotalTokens > 0 || usage.RequestCount > 0) { |
| 728 | a.sink.Emit(event.Event{Kind: event.Usage, ModelRef: a.modelRef, Usage: usage, Pricing: a.pricing, UsageSource: event.UsageSourceCompaction}) |
| 729 | } |
| 730 | }() |
| 731 | ch, err := a.prov.Stream(ctx, provider.Request{ |
| 732 | Messages: []provider.Message{ |
| 733 | {Role: provider.RoleSystem, Content: sys}, |
| 734 | {Role: provider.RoleUser, Content: renderTranscript(region)}, |
| 735 | }, |
| 736 | Temperature: provider.OptionalTemperature(a.temperature), |
| 737 | }) |
| 738 | if err != nil { |
| 739 | return "", err |
| 740 | } |
| 741 | |
| 742 | // select on ctx.Done so a stalled stream (open but never delivering or closing) |
| 743 | // unblocks on timeout instead of pinning the "compacting…" placeholder forever. |
| 744 | var b strings.Builder |
| 745 | for { |
| 746 | select { |
| 747 | case <-ctx.Done(): |
| 748 | return "", ctx.Err() |
| 749 | case chunk, ok := <-ch: |
| 750 | if !ok { |
| 751 | s := strings.TrimSpace(b.String()) |
| 752 | if s == "" { |
| 753 | return "", fmt.Errorf("summarizer returned empty output") |
| 754 | } |
| 755 | return s, nil |
| 756 | } |
| 757 | switch chunk.Type { |
| 758 | case provider.ChunkText: |
| 759 | b.WriteString(chunk.Text) |
| 760 | case provider.ChunkUsage: |
| 761 | usage = chunk.Usage |
| 762 | case provider.ChunkError: |
| 763 | return "", chunk.Err |
| 764 | } |
| 765 | } |
| 766 | } |
| 767 | } |
| 768 | |
| 769 | // summarizeWithRetry retries one non-timeout failure (a transient stream drop or |
| 770 | // rate blip); a timeout or a second failure returns so the caller folds |
| 771 | // mechanically rather than waiting again. |
| 772 | func (a *Agent) summarizeWithRetry(ctx context.Context, fold []provider.Message, instructions string) (string, error) { |
| 773 | summary, err := a.summarize(ctx, fold, instructions) |
| 774 | if err == nil || errors.Is(err, context.DeadlineExceeded) || errors.Is(err, context.Canceled) { |
| 775 | return summary, err |
| 776 | } |
| 777 | return a.summarize(ctx, fold, instructions) |
| 778 | } |
| 779 | |
| 780 | // mechanicalFoldDigest is the deterministic stand-in used when the summarizer is |
| 781 | // unreachable: the foldable region is already archived, so the digest just notes |
| 782 | // the gap and points the model at the user for anything it needs from before it. |
| 783 | func mechanicalFoldDigest(n int, archive string) string { |
| 784 | where := "." |
| 785 | if archive != "" { |
| 786 | where = " (archived to " + archive + ")." |
| 787 | } |
| 788 | return fmt.Sprintf("%d earlier message(s) were folded here to free context, but the automatic summary was unavailable%s Ask the user if you need details from before this point.", n, where) |
| 789 | } |
| 790 | |
| 791 | // renderTranscript flattens messages into a readable transcript for summarization. |
| 792 | func renderTranscript(msgs []provider.Message) string { |
| 793 | var b strings.Builder |
| 794 | for _, m := range msgs { |
| 795 | if m.LocalOnly { |
| 796 | continue |
| 797 | } |
| 798 | switch m.Role { |
| 799 | case provider.RoleUser: |
| 800 | fmt.Fprintf(&b, "[user]\n%s\n\n", m.Content) |
| 801 | case provider.RoleAssistant: |
| 802 | if m.Content != "" { |
| 803 | fmt.Fprintf(&b, "[assistant]\n%s\n", m.Content) |
| 804 | } |
| 805 | for _, tc := range m.ToolCalls { |
| 806 | fmt.Fprintf(&b, "[assistant calls %s] %s\n", tc.Name, summarizeToolArgs(tc.Arguments)) |
| 807 | } |
| 808 | b.WriteString("\n") |
| 809 | case provider.RoleTool: |
| 810 | fmt.Fprintf(&b, "[tool %s result]\n%s\n\n", m.Name, m.Content) |
| 811 | case provider.RoleSystem: |
| 812 | fmt.Fprintf(&b, "[system]\n%s\n\n", m.Content) |
| 813 | } |
| 814 | } |
| 815 | return b.String() |
| 816 | } |
| 817 | |
| 818 | // summarizeToolArgs returns a short summary of tool-call arguments instead of |
| 819 | // the full JSON. This prevents the summarizer from reproducing long argument |
| 820 | // text (like sub-agent task prompts) in the compaction summary, which would |
| 821 | // leak into the session as a user message (#4317). |
| 822 | func summarizeToolArgs(args string) string { |
| 823 | if args == "" { |
| 824 | return "(no arguments)" |
| 825 | } |
| 826 | var parsed map[string]any |
| 827 | if err := json.Unmarshal([]byte(args), &parsed); err != nil { |
| 828 | // Not valid JSON — return a length hint instead of raw text. |
| 829 | return fmt.Sprintf("(%d bytes)", len(args)) |
| 830 | } |
| 831 | keys := make([]string, 0, len(parsed)) |
| 832 | for k := range parsed { |
| 833 | keys = append(keys, k) |
| 834 | } |
| 835 | sort.Strings(keys) |
| 836 | return fmt.Sprintf("{%s} (%d keys)", strings.Join(keys, ", "), len(parsed)) |
| 837 | } |
| 838 | |
| 839 | // archiveMessages writes the dropped originals to a timestamped .jsonl (one |
| 840 | // message per line) under dir, returning the file path. |
| 841 | func archiveMessages(dir string, msgs []provider.Message) (string, error) { |
| 842 | if err := os.MkdirAll(dir, 0o755); err != nil { |
| 843 | return "", err |
| 844 | } |
| 845 | path := filepath.Join(dir, time.Now().Format("20060102-150405.000")+".jsonl") |
| 846 | f, err := os.Create(path) |
| 847 | if err != nil { |
| 848 | return "", err |
| 849 | } |
| 850 | defer f.Close() |
| 851 | |
| 852 | enc := json.NewEncoder(f) |
| 853 | for _, m := range msgs { |
| 854 | if err := enc.Encode(m); err != nil { |
| 855 | return "", err |
| 856 | } |
| 857 | } |
| 858 | return path, nil |
| 859 | } |
| 860 |