minicode-ai 0.14.2 → 0.14.5

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (78) hide show
  1. package/README.md +34 -178
  2. package/cli/args.ts +10 -18
  3. package/cli/commands/acp.ts +20 -27
  4. package/cli/commands/daemon.ts +6 -7
  5. package/cli/commands/exec.ts +12 -10
  6. package/cli/commands.ts +6 -9
  7. package/cli/index.ts +30 -15
  8. package/cli/provider-manager.ts +1 -3
  9. package/cli/setup.ts +188 -143
  10. package/cli/tui.ts +15 -23
  11. package/docs/CONTROL-PLANE-MAP.md +145 -0
  12. package/docs/OUTPUT_EVENT_MODEL.md +1 -1
  13. package/docs/OUTPUT_PROTOCOL_SPEC.md +2 -2
  14. package/docs/OUTPUT_RENDERING_SPEC.md +7 -2
  15. package/docs/OUTPUT_UX_RULES.md +1 -1
  16. package/docs/README.md +14 -102
  17. package/docs/SUMMARY.md +53 -0
  18. package/docs/TERMINAL_CONTRACT.md +84 -51
  19. package/docs/UI_RENDER_PIPELINE.md +204 -0
  20. package/docs/agents.md +5 -5
  21. package/docs/architecture.md +23 -12
  22. package/docs/choosing-mode.md +7 -7
  23. package/docs/cli.md +22 -29
  24. package/docs/concepts.md +1 -1
  25. package/docs/config.md +10 -5
  26. package/docs/contributing.md +6 -6
  27. package/docs/environment.md +3 -1
  28. package/docs/exec.md +5 -5
  29. package/docs/glossary.md +8 -6
  30. package/docs/memory-sessions.md +7 -5
  31. package/docs/policy-sandbox.md +16 -14
  32. package/docs/quickstart.md +2 -2
  33. package/docs/repl.md +4 -3
  34. package/docs/security-model.md +14 -15
  35. package/docs/security.md +8 -29
  36. package/docs/skills.md +1 -1
  37. package/docs/tools.md +2 -2
  38. package/package.json +5 -3
  39. package/scripts/vendor-overlays/minicode-context.patch +109 -0
  40. package/src/app/provider-layer.ts +13 -1
  41. package/src/app/session.ts +231 -13
  42. package/src/config.ts +30 -3
  43. package/src/constants.ts +2 -0
  44. package/src/daemon/control.ts +1 -1
  45. package/src/mcp/server.ts +2 -3
  46. package/src/policy/compaction.ts +272 -16
  47. package/src/policy/context-window-catalog.ts +64 -0
  48. package/src/policy/context-window.ts +173 -0
  49. package/src/policy/context.ts +55 -3
  50. package/src/policy/pricing.ts +52 -0
  51. package/src/policy/ratelimit.ts +47 -18
  52. package/src/policy/usage.ts +15 -8
  53. package/src/policy/work-mode.ts +1 -1
  54. package/src/providers/guards.ts +61 -4
  55. package/src/providers/provision.ts +12 -3
  56. package/src/providers/router.ts +56 -23
  57. package/src/session/context-projection.ts +14 -0
  58. package/src/session/context-selector.ts +58 -20
  59. package/src/session/persistence.ts +1 -1
  60. package/src/task/model.ts +21 -70
  61. package/src/tools/task.ts +7 -0
  62. package/src/ui/footer.ts +131 -66
  63. package/src/ui/i18n/en.ts +0 -11
  64. package/src/ui/i18n/id.ts +0 -11
  65. package/src/ui/input/input.ts +2 -3
  66. package/src/ui/input/prompt-engine.ts +2 -5
  67. package/src/ui/screens/model-manager.ts +3 -10
  68. package/src/ui/screens/picker.ts +23 -40
  69. package/src/ui/screens/provider-manager.ts +12 -33
  70. package/src/ui/screens/wizard.ts +0 -1
  71. package/src/ui/tui/app.ts +10 -6
  72. package/src/ui/tui/transcript.ts +48 -8
  73. package/vendor/minicore/VENDOR.md +9 -5
  74. package/vendor/minicore/src/core/loop.ts +47 -12
  75. package/vendor/minicore/src/core/provider.ts +4 -0
  76. package/vendor/minicore/src/core/session.ts +14 -1
  77. package/docs/OUTPUT_ARCHITECTURE_AUDIT.md +0 -489
  78. package/docs/OUTPUT_IMPLEMENTATION_PLAN.md +0 -391
@@ -14,8 +14,60 @@ import { isRealPathOutsideRoot } from "./jail.ts"
14
14
 
15
15
  const execFileAsync = promisify(execFile)
16
16
 
17
- export const minicodeEstimator: TokenEstimator = (text: string) =>
18
- Math.ceil(text.length / DEFAULT_CHARS_PER_TOKEN)
17
+ const denseTokenCharacters =
18
+ /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}\p{Extended_Pictographic}]/gu
19
+ const hasDenseTokenCharacter =
20
+ /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}\p{Extended_Pictographic}]/u
21
+
22
+ export function countDenseTokenCharacters(text: string): number {
23
+ let dense = 0
24
+ for (const match of text.matchAll(denseTokenCharacters)) {
25
+ if (match[0]) dense++
26
+ }
27
+ return dense
28
+ }
29
+
30
+ /** Track a streamed text field without retaining/re-tokenizing its full prefix. */
31
+ export function createIncrementalTokenCounter(): {
32
+ add(chunk: string, factor?: number): number
33
+ } {
34
+ let codeUnits = 0
35
+ let denseCharacters = 0
36
+ let pendingHighSurrogate = ""
37
+ let previousEstimate = 0
38
+ return {
39
+ add(chunk, factor = 1) {
40
+ let complete = pendingHighSurrogate + chunk
41
+ pendingHighSurrogate = ""
42
+ if (complete.length > 0) {
43
+ const last = complete.charCodeAt(complete.length - 1)
44
+ if (last >= 0xd800 && last <= 0xdbff) {
45
+ pendingHighSurrogate = complete.slice(-1)
46
+ complete = complete.slice(0, -1)
47
+ }
48
+ }
49
+ codeUnits += complete.length
50
+ denseCharacters += countDenseTokenCharacters(complete)
51
+ const baseEstimate = Math.ceil(codeUnits / DEFAULT_CHARS_PER_TOKEN + denseCharacters * 0.75)
52
+ const safeFactor = Number.isFinite(factor) && factor >= 1 ? factor : 1
53
+ const estimate = Math.ceil(baseEstimate * safeFactor)
54
+ const added = Math.max(0, estimate - previousEstimate)
55
+ previousEstimate = Math.max(previousEstimate, estimate)
56
+ return added
57
+ },
58
+ }
59
+ }
60
+
61
+ export const minicodeEstimator: TokenEstimator = (text: string) => {
62
+ // Keep the cheap chars/4 baseline for ordinary text, but price CJK and emoji
63
+ // closer to one token per code point. Plain UTF-16 chars/4 can undercount a
64
+ // multilingual prompt by several times before the first provider usage sample.
65
+ if (!hasDenseTokenCharacter.test(text)) {
66
+ return Math.ceil(text.length / DEFAULT_CHARS_PER_TOKEN)
67
+ }
68
+ const dense = countDenseTokenCharacters(text)
69
+ return Math.ceil(text.length / DEFAULT_CHARS_PER_TOKEN + dense * 0.75)
70
+ }
19
71
 
20
72
  // helper for image content: base64 overhead 4/3 — use when estimating tool image results
21
73
  export function estimateImageTokens(bytes: number): number {
@@ -90,7 +142,7 @@ export function createCalibratedEstimator(opts: { maxFactor?: number } = {}): Ca
90
142
  // mengirim `usage` progresif-kumulatif per-chunk SSE. Sampel "actual naik,
91
143
  // local beku" adalah artefak penghitungan intra-request — bukan drift
92
144
  // tokenizer — dan sampel jauh di atas baseline lokal yang beku (actual >
93
- // local × 3) bisa membangkaskan factor secara global (uji live tiprouter:
145
+ // local × 3) bisa membangkaskan factor model ini (uji live tiprouter:
94
146
  // local=5026 beku sementara input merambat 6413 → 28024 = 5.58× → seluruh
95
147
  // estimasi konteks terinflasi → kompaksi palsu → turn dibunuh
96
148
  // `context window exceeded after compaction`). Sampel demikian DIKIP yang
@@ -63,8 +63,20 @@ export function pricingCachePath(): string {
63
63
  return join(homeDir(), ".minicode", "pricing.json")
64
64
  }
65
65
 
66
+ /** Cache terpisah untuk jendela konteks (diisi dari `models.dev` saat sync). */
67
+ export function contextWindowsCachePath(): string {
68
+ return join(homeDir(), ".minicode", "context-windows.json")
69
+ }
70
+
66
71
  const MODELS_DEV_URL = "https://models.dev/api.json"
67
72
 
73
+ interface ContextWindowsCache {
74
+ fetchedAt: number
75
+ source: string
76
+ /** modelId → jendela konteks (token). Kunci di-lowercase. */
77
+ models: Record<string, number>
78
+ }
79
+
68
80
  interface PricingCache {
69
81
  fetchedAt: number
70
82
  source: string
@@ -157,6 +169,7 @@ export interface ModelsDevPayload {
157
169
  {
158
170
  id?: string
159
171
  cost?: { input?: number; output?: number; cache_read?: number; cache_write?: number }
172
+ context_length?: unknown
160
173
  }
161
174
  >
162
175
  }
@@ -201,6 +214,36 @@ export function extractPricing(payload: ModelsDevPayload): Record<string, ModelP
201
214
  return out
202
215
  }
203
216
 
217
+ /** Normalisasi field `context_length` models.dev (bisa number atau objek). */
218
+ function contextLengthNumber(value: unknown): number | undefined {
219
+ if (typeof value === "number" && Number.isFinite(value) && value > 0) return Math.floor(value)
220
+ if (value && typeof value === "object") {
221
+ const v = value as { total?: unknown; input?: unknown; output?: unknown }
222
+ for (const c of [v.total, v.input, v.output]) {
223
+ if (typeof c === "number" && Number.isFinite(c) && c > 0) return Math.floor(c)
224
+ }
225
+ }
226
+ return undefined
227
+ }
228
+
229
+ /** Ekstrak jendela konteks (token total) dari payload models.dev. Diekspor untuk test. */
230
+ export function extractContextWindows(payload: ModelsDevPayload): Record<string, number> {
231
+ const out: Record<string, number> = {}
232
+ for (const provider of Object.values(payload ?? {})) {
233
+ for (const [key, model] of Object.entries(provider?.models ?? {})) {
234
+ const cl = contextLengthNumber(model?.context_length)
235
+ if (cl === undefined) continue
236
+ const id = (model?.id ?? key).toLowerCase()
237
+ if (!id) continue
238
+ // Context length adalah properti model (bukan provider); bila satu id
239
+ // muncul di beberapa provider dengan nilai beda, ambil max (kapasitas
240
+ // asli model). Over-estimasi menurun anggun via recovery compact-and-retry.
241
+ out[id] = out[id] === undefined ? cl : Math.max(out[id]!, cl)
242
+ }
243
+ }
244
+ return out
245
+ }
246
+
204
247
  /** Median harga, mengabaikan entri gratis bila ada yang berbayar. Diekspor untuk test. */
205
248
  export function pickRepresentativePrice(list: ModelPrice[]): ModelPrice | undefined {
206
249
  if (list.length === 0) return undefined
@@ -283,6 +326,15 @@ export async function syncPricing(url: string = MODELS_DEV_URL): Promise<SyncRes
283
326
  await atomicWriteText(pricingCachePath(), serialized)
284
327
  overlay = models
285
328
  overlayMeta = { fetchedAt: cache.fetchedAt, count }
329
+
330
+ // Sekalian simpan jendela konteks (field `context_length`) ke cache terpisah;
331
+ // satu fetch, dua artefak. Gagal menulis cache jendela jangan menggagalkan
332
+ // sync harga — harga tetap lebih penting dan path-nya sudah teruji.
333
+ const windows = extractContextWindows(payload)
334
+ if (Object.keys(windows).length > 0) {
335
+ const wCache: ContextWindowsCache = { fetchedAt: Date.now(), source: url, models: windows }
336
+ await atomicWriteText(contextWindowsCachePath(), JSON.stringify(wCache))
337
+ }
286
338
  return { count, bytes: serialized.length, path: pricingCachePath() }
287
339
  }
288
340
 
@@ -1,20 +1,22 @@
1
1
  // Token bucket rate limiter — mencegah request LLM beruntun kena 429.
2
2
  // Capacity = burst allowance; refill per ms dari rpm.
3
+ import { LIMITS } from "../constants.ts"
3
4
 
4
5
  export class RateLimiter {
6
+ private readonly capacity: number
7
+ private readonly refillPerMs: number
5
8
  private tokens: number
6
9
  private lastRefill: number
7
10
 
8
- constructor(
9
- private capacity: number,
10
- private refillPerMs: number,
11
- ) {
12
- this.tokens = capacity
13
- this.lastRefill = Date.now()
11
+ constructor(capacity: number, refillPerMs: number) {
12
+ this.capacity = Number.isSafeInteger(capacity) && capacity >= 1 ? capacity : 1
13
+ this.refillPerMs = Number.isFinite(refillPerMs) && refillPerMs > 0 ? refillPerMs : 1 / 60_000
14
+ this.tokens = this.capacity
15
+ this.lastRefill = performance.now()
14
16
  }
15
17
 
16
18
  private refill(): void {
17
- const now = Date.now()
19
+ const now = performance.now()
18
20
  const elapsed = now - this.lastRefill
19
21
  if (elapsed > 0) {
20
22
  this.tokens = Math.min(this.capacity, this.tokens + elapsed * this.refillPerMs)
@@ -22,17 +24,21 @@ export class RateLimiter {
22
24
  }
23
25
  }
24
26
 
25
- /** Ambil 1 token; tunggu bila bucket kosong. */
26
- async acquire(): Promise<void> {
27
- this.refill()
28
- if (this.tokens >= 1) {
29
- this.tokens -= 1
30
- return
27
+ /** Ambil 1 token; tunggu dibatalkan bila pemanggilnya timeout/cancel. */
28
+ async acquire(signal?: AbortSignal): Promise<void> {
29
+ while (true) {
30
+ if (signal?.aborted) throw signal.reason ?? new DOMException("Aborted", "AbortError")
31
+ this.refill()
32
+ if (this.tokens >= 1) {
33
+ this.tokens -= 1
34
+ return
35
+ }
36
+ const waitMs = Math.max(1, Math.ceil((1 - this.tokens) / this.refillPerMs))
37
+ // Node/Bun clamp delays above 2^31-1ms to 1ms. Chunk long waits so a
38
+ // very low configured RPM cannot accidentally become a hot loop.
39
+ const boundedWaitMs = Math.min(waitMs, LIMITS.MAX_TIMER_DELAY_MS)
40
+ await waitAbortable(boundedWaitMs, signal)
31
41
  }
32
- const waitMs = Math.ceil((1 - this.tokens) / this.refillPerMs)
33
- await new Promise((r) => setTimeout(r, Math.max(0, waitMs)))
34
- this.refill()
35
- this.tokens = Math.max(0, this.tokens - 1)
36
42
  }
37
43
 
38
44
  /** Info status (untuk /status). */
@@ -48,7 +54,30 @@ export class RateLimiter {
48
54
 
49
55
  // Buat limiter dari rpm (requests per minute). Burst = min(rpm, 10).
50
56
  export function createRateLimiter(rpm: number): RateLimiter {
51
- const safe = Math.max(1, rpm)
57
+ const safe = Number.isFinite(rpm) && rpm > 0 ? rpm : 60
52
58
  const capacity = Math.max(1, Math.min(Math.floor(safe), 10))
53
59
  return new RateLimiter(capacity, safe / 60_000)
54
60
  }
61
+
62
+ export function waitAbortable(ms: number, signal?: AbortSignal): Promise<void> {
63
+ return new Promise((resolve, reject) => {
64
+ if (signal?.aborted) {
65
+ reject(signal.reason ?? new DOMException("Aborted", "AbortError"))
66
+ return
67
+ }
68
+ const timer = setTimeout(() => {
69
+ cleanup()
70
+ resolve()
71
+ }, ms)
72
+ const onAbort = () => {
73
+ clearTimeout(timer)
74
+ cleanup()
75
+ reject(signal?.reason ?? new DOMException("Aborted", "AbortError"))
76
+ }
77
+ signal?.addEventListener("abort", onAbort, { once: true })
78
+ if (signal?.aborted) onAbort()
79
+ function cleanup() {
80
+ signal?.removeEventListener("abort", onAbort)
81
+ }
82
+ })
83
+ }
@@ -236,13 +236,15 @@ export function watchBudgetLimit(opts: {
236
236
  export function watchContextWindowLimit(opts: {
237
237
  bus: EventBus
238
238
  /** Batas window konteks (W). Bila 0/undefined → no-op. */
239
- windowTokens: number
239
+ windowTokens: number | (() => number)
240
240
  /** Snapshot token request provider aktif; window berlaku per request. */
241
241
  getRequestTokens: () => { inputTokens: number; outputTokens: number } | null
242
- onOver: (allowed: number, actual: number) => void
242
+ onOver: (allowed: number, actual: number, inputTokens?: number) => void
243
243
  }): () => void {
244
- const W = opts.windowTokens
245
- if (!Number.isFinite(W) || W <= 0) return () => {}
244
+ const readWindow = () =>
245
+ typeof opts.windowTokens === "function" ? opts.windowTokens() : opts.windowTokens
246
+ const initialWindow = readWindow()
247
+ if (!Number.isFinite(initialWindow) || initialWindow <= 0) return () => {}
246
248
  let fired = false
247
249
  const check = () => {
248
250
  if (fired) return
@@ -250,15 +252,20 @@ export function watchContextWindowLimit(opts: {
250
252
  // Belum ada usage untuk request aktif: jangan memakai snapshot request
251
253
  // sebelumnya atau total token kumulatif turn sebagai pengganti.
252
254
  if (!request) return
253
- const allowedOutput = Math.max(256, Math.floor(W - request.inputTokens))
255
+ const W = readWindow()
256
+ if (!Number.isFinite(W) || W <= 0) return
257
+ const allowedOutput = Math.max(0, Math.floor(W - request.inputTokens))
254
258
  const output = request.outputTokens
255
- if (output > allowedOutput) {
259
+ if (request.inputTokens > W || output > allowedOutput) {
256
260
  fired = true
257
- opts.onOver(allowedOutput, output)
261
+ opts.onOver(allowedOutput, output, request.inputTokens)
258
262
  }
259
263
  }
260
264
  const offExt = opts.bus.on("provider:extension", (e) => {
261
- if (e.kind === "usage") check()
265
+ if (e.kind === "usage") {
266
+ const data = e.data as { source?: unknown } | undefined
267
+ if (data?.source !== "compaction") check()
268
+ }
262
269
  })
263
270
  check()
264
271
  return () => {
@@ -30,7 +30,7 @@ export class VerifyWorkModeConflictError extends Error {
30
30
  readonly workMode: Exclude<WorkMode, "EXECUTE">
31
31
  constructor(workMode: Exclude<WorkMode, "EXECUTE">) {
32
32
  super(
33
- `[work-mode] --verify cannot be combined with WorkMode ${workMode} (verification runs shell commands and self-heal turns; use --work-mode EXECUTE or omit --verify)`,
33
+ `[work-mode] --verify cannot be combined with WorkMode ${workMode} (verification runs shell commands and self-heal turns; use --work-mode BUILD or omit --verify)`,
34
34
  )
35
35
  this.name = "VerifyWorkModeConflictError"
36
36
  this.workMode = workMode
@@ -25,7 +25,13 @@ export function withStreamGuards(
25
25
  maxTotalToolArgBytes?: number
26
26
  } = {},
27
27
  ): ModelProvider {
28
- const timeoutMs = opts.timeoutMs ?? LIMITS.PROVIDER_REQUEST_TIMEOUT_MS
28
+ const configuredTimeout = opts.timeoutMs
29
+ const timeoutMs =
30
+ typeof configuredTimeout === "number" &&
31
+ Number.isFinite(configuredTimeout) &&
32
+ configuredTimeout > 0
33
+ ? Math.min(LIMITS.MAX_TIMER_DELAY_MS, Math.max(1, Math.floor(configuredTimeout)))
34
+ : LIMITS.PROVIDER_REQUEST_TIMEOUT_MS
29
35
  const maxTextChars = opts.maxTextChars ?? LIMITS.PROVIDER_TEXT_MAX_CHARS
30
36
  // Reasoning = bukan bagian jawaban; cap agresif (2MB logis ≈ 500k token —
31
37
  // jauh di atas reasoning sah mana pun, jauh di bawah OOM).
@@ -61,8 +67,6 @@ export function withStreamGuards(
61
67
  timedOut = true
62
68
  ctl.abort(new Error(`provider request timed out after ${timeoutMs}ms`))
63
69
  }, timeoutMs)
64
- // Jangan menahan exit proses bila turn sudah selesai dari sisi lain.
65
- ;(timer as unknown as { unref?: () => void }).unref?.()
66
70
  let textChars = 0
67
71
  let reasonChars = 0
68
72
  let toolArgumentBytes = 0
@@ -71,8 +75,13 @@ export function withStreamGuards(
71
75
  let finishReason: "stop" | "tool_calls" | "length" | "abort" | "error" | undefined
72
76
  let hadVisibleOutput = false
73
77
  let invalidated = false
78
+ let iterator: AsyncIterator<ProviderEvent> | undefined
74
79
  try {
75
- for await (const ev of provider.stream(request, ctl.signal)) {
80
+ iterator = provider.stream(request, ctl.signal)[Symbol.asyncIterator]()
81
+ while (true) {
82
+ const next = await nextOrAbort(iterator, ctl.signal)
83
+ if (next.done) break
84
+ const ev = next.value
76
85
  if (
77
86
  finishReason !== undefined &&
78
87
  (ev.type === "text" || ev.type === "tool_call" || ev.type === "finish")
@@ -186,7 +195,55 @@ export function withStreamGuards(
186
195
  } finally {
187
196
  clearTimeout(timer)
188
197
  signal.removeEventListener("abort", onParentAbort)
198
+ if (!ctl.signal.aborted && finishReason !== "stop" && finishReason !== "tool_calls") {
199
+ ctl.abort(new DOMException("Provider stream closed", "AbortError"))
200
+ }
201
+ if (ctl.signal.aborted && iterator?.return) {
202
+ try {
203
+ void Promise.resolve(iterator.return()).catch(() => {})
204
+ } catch {}
205
+ }
189
206
  }
190
207
  },
191
208
  }
192
209
  }
210
+
211
+ function nextOrAbort<T>(
212
+ iterator: AsyncIterator<T>,
213
+ signal: AbortSignal,
214
+ ): Promise<IteratorResult<T>> {
215
+ if (signal.aborted)
216
+ return Promise.reject(signal.reason ?? new DOMException("Aborted", "AbortError"))
217
+ return new Promise((resolve, reject) => {
218
+ const onAbort = () => {
219
+ cleanup()
220
+ reject(signal.reason ?? new DOMException("Aborted", "AbortError"))
221
+ }
222
+ signal.addEventListener("abort", onAbort, { once: true })
223
+ if (signal.aborted) {
224
+ onAbort()
225
+ return
226
+ }
227
+ let next: Promise<IteratorResult<T>>
228
+ try {
229
+ next = Promise.resolve(iterator.next())
230
+ } catch (error) {
231
+ cleanup()
232
+ reject(error)
233
+ return
234
+ }
235
+ next.then(
236
+ (value) => {
237
+ cleanup()
238
+ resolve(value)
239
+ },
240
+ (error) => {
241
+ cleanup()
242
+ reject(error)
243
+ },
244
+ )
245
+ function cleanup() {
246
+ signal.removeEventListener("abort", onAbort)
247
+ }
248
+ })
249
+ }
@@ -62,8 +62,15 @@ export async function saveProvider(
62
62
  if ((e as NodeJS.ErrnoException).code !== "ENOENT") throw e
63
63
  }
64
64
  const idx = cfg.providers.findIndex((p) => p.id === entry.id)
65
- if (idx >= 0) cfg.providers[idx] = entry
66
- else cfg.providers.push(entry)
65
+ if (idx >= 0) {
66
+ const previous = cfg.providers[idx]!
67
+ cfg.providers[idx] = {
68
+ ...entry,
69
+ ...(entry.contextWindows === undefined && previous.contextWindows
70
+ ? { contextWindows: previous.contextWindows }
71
+ : {}),
72
+ }
73
+ } else cfg.providers.push(entry)
67
74
  await writeConfigAtomic(path, cfg)
68
75
  })
69
76
  }
@@ -149,13 +156,15 @@ export async function detectAndSave(
149
156
  const uniqId = deriveProviderId(baseUrl, existing, id)
150
157
  // Pertahankan knob user (thinking effort) bila provider sudah ada — tanpa
151
158
  // ini tiap add/edit me-reset effort ke default diam-diam ("hilang").
152
- const prevEffort = prevCfg.providers.find((p) => p.id === uniqId)?.reasoningEffort
159
+ const previous = prevCfg.providers.find((p) => p.id === uniqId)
160
+ const prevEffort = previous?.reasoningEffort
153
161
  const entry: ProviderEntry = {
154
162
  id: uniqId,
155
163
  baseUrl,
156
164
  apiKey,
157
165
  models: detected.models,
158
166
  providerHint: detected.providerHint,
167
+ ...(previous?.contextWindows ? { contextWindows: previous.contextWindows } : {}),
159
168
  ...(prevEffort ? { reasoningEffort: prevEffort } : {}),
160
169
  }
161
170
  await saveProvider(entry, opts)
@@ -3,17 +3,25 @@ import { ProviderError } from "#minicore/core/errors.ts"
3
3
  import type { ModelProvider, ProviderEvent, StreamRequest } from "#minicore/core/provider.ts"
4
4
  import type { Message } from "#minicore/core/types.ts"
5
5
  import { LIMITS } from "../constants.ts"
6
- import type { RateLimiter } from "../policy/ratelimit.ts"
6
+ import { DEFAULT_FALLBACK_CONTEXT_WINDOW_TOKENS } from "../policy/context-window.ts"
7
+ import { type RateLimiter, waitAbortable } from "../policy/ratelimit.ts"
7
8
 
8
9
  export interface RouterConfig {
9
10
  providers: ModelProvider[]
10
11
  defaultProviderId?: string
12
+ contextWindowTokensForModel?: (model?: string) => number
11
13
  // P2 cap
12
14
  maxRetryAfterMs?: number // default LIMITS.RETRY_AFTER_MAX_MS
13
15
  // Token bucket rate limiter (opsional) — cegah request beruntun kena 429
14
16
  limiter?: RateLimiter
15
17
  }
16
18
 
19
+ export type RouterProvider = ModelProvider & {
20
+ updateProviders: (list: ModelProvider[]) => void
21
+ updateContextWindowResolver: (resolver: (model?: string) => number) => void
22
+ contextModelIdentity: (model?: string) => string | undefined
23
+ }
24
+
17
25
  // C4 fix: convert Uint8Array tool content to base64 before an openai-compat
18
26
  // provider sees it. Anthropic handles Uint8Array natively (image →
19
27
  // source.base64 + media_type in toAnthropicMessages) — converting up-front
@@ -75,15 +83,21 @@ function requestFor(
75
83
  return { req: fixed, substituted: false }
76
84
  }
77
85
 
78
- export function createRouterProvider(config: RouterConfig): ModelProvider {
79
- const maxRetry = config.maxRetryAfterMs ?? LIMITS.RETRY_AFTER_MAX_MS
86
+ export function createRouterProvider(config: RouterConfig): RouterProvider {
87
+ const configuredMaxRetry = config.maxRetryAfterMs
88
+ const maxRetry =
89
+ typeof configuredMaxRetry === "number" &&
90
+ Number.isSafeInteger(configuredMaxRetry) &&
91
+ configuredMaxRetry >= 0
92
+ ? configuredMaxRetry
93
+ : LIMITS.RETRY_AFTER_MAX_MS
80
94
  // byId dan defaultId dihitung per-stream agar provider yang baru ditambah via /provider
81
95
  // langsung dikenali tanpa restart. Provider list bisa berubah mid-session.
82
96
  const getById = () => new Map(config.providers.map((p) => [p.id, p]))
83
97
  const getDefaultId = () => config.defaultProviderId ?? config.providers[0]?.id ?? "router"
84
98
  const getModels = () => config.providers.flatMap((p) => [...p.models])
85
99
 
86
- const router: ModelProvider & { updateProviders: (list: ModelProvider[]) => void } = {
100
+ const router: RouterProvider = {
87
101
  id: "router",
88
102
  get models() {
89
103
  return getModels()
@@ -91,6 +105,21 @@ export function createRouterProvider(config: RouterConfig): ModelProvider {
91
105
  updateProviders(list: ModelProvider[]) {
92
106
  config.providers.splice(0, config.providers.length, ...list)
93
107
  },
108
+ updateContextWindowResolver(resolver) {
109
+ config.contextWindowTokensForModel = resolver
110
+ },
111
+ contextWindowTokensForModel(model?: string) {
112
+ return config.contextWindowTokensForModel?.(model) ?? DEFAULT_FALLBACK_CONTEXT_WINDOW_TOKENS
113
+ },
114
+ contextModelIdentity(model?: string) {
115
+ if (model?.includes("::")) return model
116
+ let provider = model
117
+ ? config.providers.find((entry) => entry.models.includes(model))
118
+ : (getById().get(getDefaultId()) ?? config.providers[0])
119
+ provider ??= getById().get(getDefaultId()) ?? config.providers[0]
120
+ const effectiveModel = model && provider?.models.includes(model) ? model : provider?.models[0]
121
+ return provider && effectiveModel ? `${provider.id}::${effectiveModel}` : effectiveModel
122
+ },
94
123
  async *stream(request: StreamRequest, signal: AbortSignal): AsyncIterable<ProviderEvent> {
95
124
  // route by model name — first match wins (default/daftar urutan provider)
96
125
  // Format "providerId::modelName" = PIN eksplisit (kontrak): provider itu
@@ -144,7 +173,7 @@ export function createRouterProvider(config: RouterConfig): ModelProvider {
144
173
  let hasYieldedContent = false
145
174
  try {
146
175
  // rate limit: tunggu token bucket sebelum tiap request
147
- if (config.limiter) await config.limiter.acquire()
176
+ if (config.limiter) await config.limiter.acquire(signal)
148
177
  const fixed = needsBinaryFix(current) ? fixRequest(request) : request
149
178
  const withSys = needsSystemMessage(current) ? withSystemMessage(fixed) : fixed
150
179
  const { req, effectiveModel, substituted } = requestFor(current, { ...withSys, model })
@@ -180,7 +209,19 @@ export function createRouterProvider(config: RouterConfig): ModelProvider {
180
209
  ) {
181
210
  hasYieldedContent = true
182
211
  }
183
- yield ev
212
+ if (ev.type === "extension" && ev.kind === "usage") {
213
+ const data = ev.data && typeof ev.data === "object" ? ev.data : {}
214
+ yield {
215
+ ...ev,
216
+ data: {
217
+ ...(data as Record<string, unknown>),
218
+ model: req.model ?? model,
219
+ provider: current.id,
220
+ },
221
+ }
222
+ } else {
223
+ yield ev
224
+ }
184
225
  }
185
226
  return
186
227
  } catch (e) {
@@ -189,8 +230,13 @@ export function createRouterProvider(config: RouterConfig): ModelProvider {
189
230
  if (hasYieldedContent) throw e
190
231
  // cap retryAfter without mutating original
191
232
  let err: ProviderError = e
192
- if (e.retryAfterMs != null && e.retryAfterMs > maxRetry) {
193
- err = new ProviderError(e.category, e.message, maxRetry)
233
+ if (
234
+ e.retryAfterMs != null &&
235
+ (!Number.isFinite(e.retryAfterMs) || e.retryAfterMs < 0)
236
+ ) {
237
+ err = new ProviderError(e.category, e.message, undefined, e.cause)
238
+ } else if (e.retryAfterMs != null && e.retryAfterMs > maxRetry) {
239
+ err = new ProviderError(e.category, e.message, maxRetry, e.cause)
194
240
  }
195
241
  // P11 P1.3 — honori retry-after. Audit #14: fallback DULU bila ada
196
242
  // alternatif — jangan bakar sleep (maks 30 dtk) saat provider
@@ -212,20 +258,7 @@ export function createRouterProvider(config: RouterConfig): ModelProvider {
212
258
  }
213
259
  }
214
260
  const waitMs = Math.min(err.retryAfterMs, maxRetry)
215
- if (signal.aborted) throw new DOMException("Aborted", "AbortError")
216
- let onAbort: (() => void) | undefined
217
- try {
218
- await Promise.race([
219
- Bun.sleep(waitMs),
220
- new Promise<void>((_, rej) => {
221
- onAbort = () => rej(new DOMException("Aborted", "AbortError"))
222
- if (signal.aborted) onAbort()
223
- else signal.addEventListener("abort", onAbort, { once: true })
224
- }),
225
- ])
226
- } finally {
227
- if (onAbort) signal.removeEventListener("abort", onAbort)
228
- }
261
+ await waitAbortable(waitMs, signal)
229
262
  // Pin: tetap di provider yang dipilih (tunggu lalu ulangi di
230
263
  // tempat) — jangan pindah ke provider lain.
231
264
  if (pinned) continue
@@ -256,5 +289,5 @@ export function createRouterProvider(config: RouterConfig): ModelProvider {
256
289
  }
257
290
  },
258
291
  }
259
- return router as ModelProvider
292
+ return router
260
293
  }
@@ -29,6 +29,7 @@
29
29
 
30
30
  import { ContextStore, mechanicalCompaction } from "#minicore"
31
31
  import type { Message } from "#minicore/core/types.ts"
32
+ import { LIMITS } from "../constants.ts"
32
33
  import { renderFullHistoryFold } from "./full-history-fold.ts"
33
34
  import {
34
35
  buildProjection,
@@ -100,6 +101,18 @@ export function deriveSummaryFromCanonical(
100
101
  return { summaryText, baseSeq }
101
102
  }
102
103
 
104
+ function boundProjectionSummary(text: string): string {
105
+ const cap = LIMITS.COMPACTION_SUMMARY_MAX_CHARS
106
+ if (text.length <= cap) return text
107
+ const marker = "\n[projection summary truncated; older details may be missing]"
108
+ let head = text.slice(0, Math.max(0, cap - marker.length))
109
+ if (head.length > 0) {
110
+ const code = head.charCodeAt(head.length - 1)
111
+ if (code >= 0xd800 && code <= 0xdbff) head = head.slice(0, -1)
112
+ }
113
+ return `${head}${marker}`
114
+ }
115
+
103
116
  /**
104
117
  * Bungkus P3.7 `renderFullHistoryFold` untuk seam produsen: ubah baris thread
105
118
  * menjadi input renderer, dan ubah `FoldError` (cakupan tak-terbukti) menjadi
@@ -182,6 +195,7 @@ export function produceSummaryProjection(
182
195
  if (baseSeq === undefined || baseSeq <= 0) {
183
196
  return { produced: false, detail: "no explicit/derivable baseSeq — refusing empty projection" }
184
197
  }
198
+ summaryText = boundProjectionSummary(summaryText)
185
199
  // Tulis lewat primitif P2.7 (epoch-fenced, satu txn, delete+insert, sumber
186
200
  // kanonik UTUH). Satu-satunya efek tulis produsen = history_projections.
187
201
  const row = buildProjection(sessionId, threadId, summaryText, cwd, {