minicode-ai 0.22.0 → 0.23.0

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 (46) hide show
  1. package/cli/commands.ts +10 -0
  2. package/cli/index.ts +11 -0
  3. package/cli/setup.ts +64 -33
  4. package/cli/tui.ts +3 -3
  5. package/docs/TERMINAL_CONTRACT.md +39 -6
  6. package/docs/environment.md +3 -0
  7. package/package.json +1 -1
  8. package/src/app/session.ts +129 -20
  9. package/src/config.ts +100 -2
  10. package/src/constants.ts +29 -4
  11. package/src/policy/compaction.ts +27 -0
  12. package/src/policy/context-budget.ts +132 -0
  13. package/src/policy/context-escalation.ts +140 -0
  14. package/src/policy/context-window.ts +47 -9
  15. package/src/policy/context.ts +10 -1
  16. package/src/policy/microcompact.ts +162 -0
  17. package/src/providers/anthropic.ts +17 -3
  18. package/src/providers/detect.ts +128 -13
  19. package/src/providers/provision.ts +78 -1
  20. package/src/providers/responses.ts +13 -1
  21. package/src/session/persistence.ts +56 -0
  22. package/src/ui/assistant/simple.ts +1 -1
  23. package/src/ui/i18n/en.ts +19 -4
  24. package/src/ui/i18n/id.ts +21 -4
  25. package/src/ui/input/input.ts +2 -2
  26. package/src/ui/render/activity-panel.ts +21 -3
  27. package/src/ui/render/hint.ts +28 -0
  28. package/src/ui/render/logo.ts +91 -0
  29. package/src/ui/render/session-summary.ts +22 -1
  30. package/src/ui/render/theme.ts +5 -0
  31. package/src/ui/runtime/screen.ts +4 -2
  32. package/src/ui/screens/form.ts +19 -23
  33. package/src/ui/screens/lifecycle.ts +80 -0
  34. package/src/ui/screens/model-manager.ts +77 -81
  35. package/src/ui/screens/picker.ts +28 -22
  36. package/src/ui/screens/provider-manager.ts +47 -52
  37. package/src/ui/screens/watchdog.ts +27 -0
  38. package/src/ui/screens/wizard.ts +3 -20
  39. package/src/ui/tui/app.ts +5 -2
  40. package/src/ui/tui/transcript.ts +51 -16
  41. package/vendor/minicore/src/core/compact.ts +10 -0
  42. package/vendor/minicore/src/core/errors.ts +11 -1
  43. package/vendor/minicore/src/core/executor.ts +6 -1
  44. package/vendor/minicore/src/core/loop.ts +92 -2
  45. package/vendor/minicore/src/core/session.ts +8 -0
  46. package/vendor/minicore/src/providers/openai-compat.ts +15 -2
@@ -0,0 +1,162 @@
1
+ import type { Message } from "#minicore/core/types.ts"
2
+
3
+ /**
4
+ * Microcompact — pangkas ISI hasil tool lama jadi sentinel pendek: tanpa
5
+ * membuang pesan, tanpa LLM, tanpa menyentuh pasangan `assistant(toolCalls)`
6
+ * → `tool` yang masih berlangsung.
7
+ *
8
+ * KENAPA: pada sesi panjang hasil tool adalah penyumbang payload terbesar
9
+ * (>90% karakter konteks di lapangan). Kompaksi penuh mahal + lossy; memangkas
10
+ * isi lama jauh lebih murah dan mempertahankan struktur percakapan. Lapisan
11
+ * pertama sebelum kompaksi penuh, bukan penggantinya.
12
+ *
13
+ * KENAPA keepRecent=5: model masih butuh hasil terakhir untuk menyimpulkan
14
+ * langkah berikutnya; 5 = titik kumpul lintas implementasi harness produksi.
15
+ *
16
+ * KENAPA gating tiga syarat: memangkas hasil kecil tak menghemat apa pun tapi
17
+ * membuang informasi. Sepadan hanya bila (a) hemat berarti, (b) hasil tool
18
+ * mendominasi, (c) bukan di tengah batch tool yang belum selesai.
19
+ */
20
+
21
+ /** Tool yang hasilnya DAPAT dipulihkan (dibaca ulang kapan saja). Tool
22
+ * kontrol-sesi (`todo`, `memory`) SENGAJA tidak masuk: isinya keadaan yang tak
23
+ * bisa direkonstruksi dari filesystem — memangkasnya menghapus kebenaran. */
24
+ export const MICROCOMPACT_RECOVERABLE_TOOLS: ReadonlySet<string> = new Set([
25
+ "read_file",
26
+ "grep",
27
+ "glob",
28
+ "bash",
29
+ "bash_output",
30
+ "code_run",
31
+ "web_fetch",
32
+ "web_search",
33
+ "list_dir",
34
+ "view_file",
35
+ ])
36
+
37
+ /** Awalan sentinel. Stabil supaya bisa dikenali ulang (idempoten) dan supaya
38
+ * model tahu isinya dipangkas, bukan kosong. */
39
+ export const MICROCOMPACT_SENTINEL_PREFIX = "[hasil tool lama dipangkas"
40
+
41
+ export interface MicrocompactOptions {
42
+ /** Hasil tool TERBARU yang dipertahankan utuh. Default 5. */
43
+ keepRecent?: number
44
+ /** Estimator token (disuntik agar modul tetap murni). */
45
+ estimateTokens: (text: string) => number
46
+ /** Hemat minimum agar pemangkasan sepadan. Default 20.000 token. */
47
+ minSavingTokens?: number
48
+ /** Porsi total hasil tool agar pemangkasan dipertimbangkan (default 40%). */
49
+ minToolShare?: number
50
+ /** Budget konteks efektif (jendela - reserve). Wajib untuk syarat (b). */
51
+ budgetTokens?: number
52
+ }
53
+
54
+ export interface MicrocompactResult {
55
+ readonly messages: readonly Message[]
56
+ readonly clearedCount: number
57
+ readonly savedTokens: number
58
+ /** Diisi bila tak memangkas — dibedakan dari "tak ada yang bisa dipangkas"
59
+ * supaya pemanggil tidak melaporkan no-op palsu. */
60
+ readonly skippedReason?: "below-saving" | "below-share" | "no-candidates" | "in-flight-batch"
61
+ }
62
+
63
+ /** Konten pesan → teks untuk diukur. Objek bersarang di-stringify aman. */
64
+ function contentText(content: unknown): string {
65
+ if (typeof content === "string") return content
66
+ if (content === null || content === undefined) return ""
67
+ try {
68
+ return JSON.stringify(content) ?? ""
69
+ } catch {
70
+ return String(content)
71
+ }
72
+ }
73
+
74
+ /**
75
+ * Indeks hasil tool yang berada dalam batch AKTIF (toolCall terakhir dari
76
+ * assistant terakhir yang belum punya hasil lengkap).
77
+ *
78
+ * KENAPA penting: memangkas sebagian hasil dari satu batch multi-tool membuat
79
+ * model melihat observasi setengah jadi — lebih menyesatkan daripada konteks
80
+ * penuh. Batch dianggap aktif selama ia adalah grup `tool` terakhir.
81
+ */
82
+ function activeBatchIds(messages: readonly Message[]): Set<string> {
83
+ const lastAssistantIdx = (() => {
84
+ for (let i = messages.length - 1; i >= 0; i--) if (messages[i]?.role === "assistant") return i
85
+ return -1
86
+ })()
87
+ if (lastAssistantIdx < 0) return new Set()
88
+ // Semua hasil tool SETELAH assistant terakhir (bila ada) adalah batch aktif.
89
+ const ids = new Set<string>()
90
+ for (let i = lastAssistantIdx + 1; i < messages.length; i++) {
91
+ const m = messages[i]
92
+ if (m?.role === "tool") ids.add(m.toolCallId)
93
+ }
94
+ return ids
95
+ }
96
+
97
+ /**
98
+ * Murni (input tak dimutasi). Semua syarat harus lulus, jika tidak pesan
99
+ * dikembalikan apa adanya: (1) kandidat = hasil tool yang dapat dipulihkan di
100
+ * luar batch aktif, (2) melebihi `keepRecent`, (3) hemat ≥ `minSavingTokens`,
101
+ * (4) total hasil tool ≥ `minToolShare` × budget.
102
+ */
103
+ export function applyMicrocompact(
104
+ messages: readonly Message[],
105
+ opts: MicrocompactOptions,
106
+ ): MicrocompactResult {
107
+ const keepRecent = Math.max(1, opts.keepRecent ?? 5)
108
+ const minSaving = Math.max(0, opts.minSavingTokens ?? 20_000)
109
+ const minShare = opts.minToolShare ?? 0.4
110
+ const inFlight = activeBatchIds(messages)
111
+
112
+ // Dari belakang: `keepRecent` terbaru dipertahankan.
113
+ const candidates: number[] = []
114
+ for (let i = messages.length - 1; i >= 0; i--) {
115
+ const m = messages[i]
116
+ if (m?.role !== "tool") continue
117
+ if (!MICROCOMPACT_RECOVERABLE_TOOLS.has(m.name)) continue
118
+ if (inFlight.has(m.toolCallId)) continue
119
+ const text = contentText(m.content)
120
+ if (text.startsWith(MICROCOMPACT_SENTINEL_PREFIX)) continue // sudah dipangkas
121
+ candidates.push(i)
122
+ }
123
+ if (candidates.length === 0)
124
+ return { messages, clearedCount: 0, savedTokens: 0, skippedReason: "no-candidates" }
125
+ // `candidates` terurut dari belakang; `keepRecent` pertama = terbaru.
126
+ const victims = candidates.slice(keepRecent)
127
+ if (victims.length === 0)
128
+ return { messages, clearedCount: 0, savedTokens: 0, skippedReason: "no-candidates" }
129
+
130
+ // (b): hasil tool harus mendominasi budget, kalau tidak ia hanya membuang
131
+ // informasi demi sedikit token.
132
+ const toolTokens = candidates.reduce(
133
+ (sum, i) => sum + opts.estimateTokens(contentText(messages[i]!.content)),
134
+ 0,
135
+ )
136
+ if (opts.budgetTokens !== undefined && opts.budgetTokens > 0) {
137
+ if (toolTokens < opts.budgetTokens * minShare)
138
+ return { messages, clearedCount: 0, savedTokens: 0, skippedReason: "below-share" }
139
+ }
140
+
141
+ let clearedCount = 0
142
+ let savedTokens = 0
143
+ const out = messages.slice()
144
+ for (const idx of victims) {
145
+ const m = messages[idx] as Extract<Message, { role: "tool" }>
146
+ const original = contentText(m.content)
147
+ const wasError = m.isError === true
148
+ const sentinel = `${MICROCOMPACT_SENTINEL_PREFIX}: ${m.name} ${original.length} char${wasError ? ", gagal" : ""}. Jalankan ulang tool ini bila isinya masih dibutuhkan.]`
149
+ // Hemat = token asli - token sentinel (sentinel tetap memakan token).
150
+ const saving = opts.estimateTokens(original) - opts.estimateTokens(sentinel)
151
+ out[idx] = { ...m, content: sentinel }
152
+ clearedCount++
153
+ // Sentinel bisa lebih panjang dari aslinya: jangan catat "hemat" palsu.
154
+ savedTokens += Math.max(0, saving)
155
+ }
156
+
157
+ // (c): hemat harus berarti. Bila tidak, batalkan seluruh operasi.
158
+ if (savedTokens < minSaving)
159
+ return { messages, clearedCount: 0, savedTokens: 0, skippedReason: "below-saving" }
160
+
161
+ return { messages: out, clearedCount, savedTokens }
162
+ }
@@ -68,7 +68,10 @@ export function createAnthropicProvider(config: AnthropicConfig): ModelProvider
68
68
  .replace(/\/+$/, "")
69
69
  .replace(/\/v1$/, "")
70
70
  const endpoint = `${baseUrl}/v1/messages`
71
- const enableCache = config.enablePromptCaching !== false
71
+ // Prompt caching aktif default; kill-switch config `enablePromptCaching:false`
72
+ // (statis) + env `MINICODE_PROMPT_CACHE=0` (runtime, pola microcompact/probe).
73
+ const enableCache =
74
+ config.enablePromptCaching !== false && process.env.MINICODE_PROMPT_CACHE !== "0"
72
75
 
73
76
  // brand `kind` dipakai router agar tahu konten Uint8Array ditangani native
74
77
  const provider: ModelProvider & { kind: "anthropic" } = {
@@ -101,11 +104,19 @@ export function createAnthropicProvider(config: AnthropicConfig): ModelProvider
101
104
  const effectiveMaxTokens = legacyBudget
102
105
  ? Math.max(config.maxTokens ?? 8192, legacyBudget + 1024)
103
106
  : (config.maxTokens ?? 8192)
107
+ // Fase 7: cache prefix percakapan (system+tools+histori) — tandai pesan
108
+ // stabil terakhir (kedua dari akhir; yang terakhir = pesan user saat ini)
109
+ // agar tool result besar ikut ter-cache tiap turn (bukan dikirim ulang).
110
+ const messagesPayload = toAnthropicMessages(request.messages)
111
+ if (enableCache && messagesPayload.length >= 2) {
112
+ const lastStable = messagesPayload[messagesPayload.length - 2] as Record<string, unknown>
113
+ lastStable.cache_control = { type: "ephemeral" }
114
+ }
104
115
  const body = JSON.stringify({
105
116
  model: request.model ?? config.defaultModel ?? config.models[0],
106
117
  max_tokens: effectiveMaxTokens,
107
118
  system: systemPayload,
108
- messages: toAnthropicMessages(request.messages),
119
+ messages: messagesPayload,
109
120
  tools: toolsPayload,
110
121
  ...(legacyBudget ? { thinking: { type: "enabled", budget_tokens: legacyBudget } } : {}),
111
122
  // Adaptive: tanpa budget & tanpa beta thinking (dikelola model);
@@ -387,7 +398,10 @@ function toContent(content: Content): unknown {
387
398
  )
388
399
  }
389
400
 
390
- function toAnthropicTools(tools: readonly ToolSchema[], enableCache: boolean = true): unknown[] {
401
+ export function toAnthropicTools(
402
+ tools: readonly ToolSchema[],
403
+ enableCache: boolean = true,
404
+ ): unknown[] {
391
405
  return tools.map((t, idx) => {
392
406
  const isLast = idx === tools.length - 1
393
407
  return {
@@ -2,10 +2,24 @@ import { LIMITS } from "../constants.ts"
2
2
 
3
3
  export interface DetectedModel {
4
4
  id: string
5
+ /**
6
+ * Jendela konteks yang DILAPORKAN endpoint `/models`, bila ada.
7
+ *
8
+ * KENAPA dibaca: kapasitas endpoint yang benar-benar dipanggil adalah sumber
9
+ * otoritatif — katalog publik bisa basi atau tidak mengenali alias gateway.
10
+ * Nilai ini tidak dikarang: hanya diisi bila endpoint menyertakannya.
11
+ */
12
+ contextLength?: number
13
+ /** Batas output model (`max_output_tokens`) bila dilaporkan. Dipakai sebagai
14
+ * reserve output proporsional, bukan angka tetap. */
15
+ maxOutputTokens?: number
5
16
  }
6
17
 
7
18
  export interface DetectResult {
8
19
  models: string[]
20
+ /** Metadata per model, hanya untuk model yang endpoint-nya melaporkan
21
+ * kapasitas. Model tanpa metadata tidak muncul di sini (tidak dikarang). */
22
+ modelMetadata?: Record<string, DetectedModel>
9
23
  providerHint: "openai" | "anthropic" | "responses" | "unknown"
10
24
  /**
11
25
  * True bila server menjawab 401/403 di salah satu attempt: key salah/
@@ -40,11 +54,111 @@ function hybridHeaders(apiKey: string): Record<string, string>[] {
40
54
  ]
41
55
  }
42
56
 
57
+ /**
58
+ * Nama field yang dipakai berbagai gateway/agregator untuk melaporkan
59
+ * kapasitas konteks. Tidak ada standar tunggal: OpenAI-compat memakai
60
+ * `context_length`, Anthropic-style memakai `max_input_tokens`, agregator
61
+ * memakai `context_window` atau menyarangkan di `top_provider`.
62
+ *
63
+ * KENAPA daftar eksplisit, bukan tebakan berbasis substring: field bernama
64
+ * `max_tokens` pada sebagian endpoint berarti batas OUTPUT, bukan konteks —
65
+ * memetakannya sebagai konteks akan mengecilkan jendela secara diam-diam,
66
+ * persis kelas bug yang ingin dihilangkan.
67
+ */
68
+ const CONTEXT_FIELD_ALIASES = [
69
+ "context_length",
70
+ "context_window",
71
+ "max_context_length",
72
+ "max_context_tokens",
73
+ "max_input_tokens",
74
+ "input_token_limit",
75
+ ] as const
76
+
77
+ const OUTPUT_FIELD_ALIASES = [
78
+ "max_output_tokens",
79
+ "max_completion_tokens",
80
+ "max_output",
81
+ "output_token_limit",
82
+ ] as const
83
+
84
+ function positiveInt(value: unknown): number | undefined {
85
+ if (typeof value !== "number" || !Number.isFinite(value)) return undefined
86
+ const n = Math.floor(value)
87
+ // Batas atas jauh di atas kapasitas model publik mana pun (2M) — menolak
88
+ // nilai rusak/absurd tanpa menolak model nyata.
89
+ return n > 0 && n <= 20_000_000 ? n : undefined
90
+ }
91
+
92
+ function pickField(obj: Record<string, unknown>, aliases: readonly string[]): number | undefined {
93
+ for (const key of aliases) {
94
+ const v = positiveInt(obj[key])
95
+ if (v !== undefined) return v
96
+ }
97
+ return undefined
98
+ }
99
+
100
+ /**
101
+ * Ekstrak metadata kapasitas dari satu entri `/models`.
102
+ *
103
+ * Diekspor untuk test: perilaku yang paling penting diuji di sini — payload
104
+ * tanpa field kapasitas harus menghasilkan `undefined` (bukan angka karangan).
105
+ */
106
+ export function extractModelCapacity(entry: unknown): {
107
+ contextLength?: number
108
+ maxOutputTokens?: number
109
+ } {
110
+ if (!entry || typeof entry !== "object") return {}
111
+ const obj = entry as Record<string, unknown>
112
+ // Agregator menyarangkan kapasitas di `top_provider`, tapi sebagian juga
113
+ // menduplikasinya di level atas — cek keduanya, level atas menang karena
114
+ // itu yang paling spesifik untuk model ini.
115
+ const nested =
116
+ obj.top_provider && typeof obj.top_provider === "object"
117
+ ? (obj.top_provider as Record<string, unknown>)
118
+ : undefined
119
+ const contextLength =
120
+ pickField(obj, CONTEXT_FIELD_ALIASES) ??
121
+ (nested ? pickField(nested, CONTEXT_FIELD_ALIASES) : undefined)
122
+ const maxOutputTokens =
123
+ pickField(obj, OUTPUT_FIELD_ALIASES) ??
124
+ (nested ? pickField(nested, OUTPUT_FIELD_ALIASES) : undefined)
125
+ return {
126
+ ...(contextLength !== undefined ? { contextLength } : {}),
127
+ ...(maxOutputTokens !== undefined ? { maxOutputTokens } : {}),
128
+ }
129
+ }
130
+
131
+ /** Petakan entri mentah `/models` → id + metadata (hanya yang dilaporkan). */
132
+ function mapEntries(raw: unknown[]): {
133
+ models: string[]
134
+ modelMetadata?: Record<string, DetectedModel>
135
+ } {
136
+ const models: string[] = []
137
+ let metadata: Record<string, DetectedModel> | undefined
138
+ for (const entry of raw) {
139
+ if (!entry || typeof entry !== "object") continue
140
+ const id = (entry as { id?: unknown }).id
141
+ if (typeof id !== "string" || !id) continue
142
+ models.push(id)
143
+ const capacity = extractModelCapacity(entry)
144
+ if (capacity.contextLength !== undefined || capacity.maxOutputTokens !== undefined) {
145
+ metadata ??= {}
146
+ metadata[id] = { id, ...capacity }
147
+ }
148
+ }
149
+ return { models, ...(metadata ? { modelMetadata: metadata } : {}) }
150
+ }
151
+
43
152
  async function tryFetchModels(
44
153
  baseUrl: string,
45
154
  headers: Record<string, string>,
46
155
  signal: AbortSignal,
47
- ): Promise<{ models: string[] | null; contacted: boolean; sawAuth: boolean }> {
156
+ ): Promise<{
157
+ models: string[] | null
158
+ modelMetadata?: Record<string, DetectedModel>
159
+ contacted: boolean
160
+ sawAuth: boolean
161
+ }> {
48
162
  const urls = [`${baseUrl.replace(/\/+$/, "")}/models`, `${baseUrl.replace(/\/+$/, "")}/v1/models`]
49
163
  let contacted = false
50
164
  let sawAuth = false
@@ -61,17 +175,17 @@ async function tryFetchModels(
61
175
  // terlihat sukses tersimpan).
62
176
  if (res.status === 401 || res.status === 403) sawAuth = true
63
177
  if (!res.ok) continue
64
- const json = (await res.json()) as { data?: { id: string }[]; models?: { id: string }[] }
178
+ const json = (await res.json()) as { data?: unknown[]; models?: unknown[] }
65
179
  const data = json.data ?? json.models ?? []
66
- if (Array.isArray(data) && data.length)
67
- return { models: data.map((m) => m.id).filter(Boolean), contacted, sawAuth }
68
- // anthropic format: {data: [{id, display_name}]}
69
- if (Array.isArray((json as unknown as { models: unknown }).models)) {
70
- return {
71
- models: (json as unknown as { models: { id: string }[] }).models.map((m) => m.id),
72
- contacted,
73
- sawAuth,
74
- }
180
+ if (Array.isArray(data) && data.length) {
181
+ const mapped = mapEntries(data)
182
+ if (mapped.models.length)
183
+ return {
184
+ models: mapped.models,
185
+ ...(mapped.modelMetadata ? { modelMetadata: mapped.modelMetadata } : {}),
186
+ contacted,
187
+ sawAuth,
188
+ }
75
189
  }
76
190
  } catch {}
77
191
  }
@@ -94,7 +208,7 @@ export async function detectModels(
94
208
  let everAuthFailed = false
95
209
  for (const h of hybridHeaders(apiKey)) {
96
210
  if (sig.aborted) break
97
- const { models, contacted, sawAuth } = await tryFetchModels(baseUrl, h, sig)
211
+ const { models, modelMetadata, contacted, sawAuth } = await tryFetchModels(baseUrl, h, sig)
98
212
  everContacted = everContacted || contacted
99
213
  everAuthFailed = everAuthFailed || sawAuth
100
214
  if (models?.length) {
@@ -110,8 +224,9 @@ export async function detectModels(
110
224
  ? "anthropic"
111
225
  : "openai"
112
226
  const uniqueModels = Array.from(new Set(models))
113
- const result = {
227
+ const result: DetectResult = {
114
228
  models: uniqueModels,
229
+ ...(modelMetadata ? { modelMetadata } : {}),
115
230
  providerHint: hint as DetectResult["providerHint"],
116
231
  authFailed: everAuthFailed,
117
232
  }
@@ -5,6 +5,7 @@
5
5
  import { readFile } from "node:fs/promises"
6
6
  import { resolve } from "node:path"
7
7
  import {
8
+ type ContextWindowSource,
8
9
  globalConfigPath,
9
10
  LOCAL,
10
11
  loadConfig,
@@ -106,6 +107,66 @@ export function deriveProviderId(baseUrl: string, existingIds: string[], id?: st
106
107
  return uniqId
107
108
  }
108
109
 
110
+ /**
111
+ * Gabungkan metadata kapasitas yang DITEMUKAN endpoint ke entri provider.
112
+ *
113
+ * Aturan presedensi (alasannya sama dengan `resolveContextWindow`):
114
+ * - Nilai ber-`source: "user"` milik operator TIDAK PERNAH ditimpa. Config lama
115
+ * tanpa catatan asal diperlakukan `user` supaya perilaku tidak berubah diam-diam.
116
+ * - Nilai `endpoint`/`probe` boleh diperbarui oleh deteksi berikutnya — endpoint
117
+ * yang dipanggil hari ini lebih tahu daripada hasil deteksi kemarin.
118
+ * - Model yang hilang dari `/models` (sudah tak ada) dibuang agar config tidak
119
+ * mengumpulkan nilai basi untuk model yang tak bisa dipakai lagi.
120
+ *
121
+ * Diekspor untuk test: ini satu-satunya tempat yang menentukan "siapa menang".
122
+ */
123
+ export function mergeDiscoveredCapacity(
124
+ previous: ProviderEntry | undefined,
125
+ detected: DetectResult,
126
+ ): Pick<ProviderEntry, "contextWindows" | "contextWindowSources" | "maxOutputTokens"> {
127
+ const metadata = detected.modelMetadata
128
+ if (!metadata || Object.keys(metadata).length === 0) return {}
129
+ const prevWindows = previous?.contextWindows ?? {}
130
+ const prevSources = previous?.contextWindowSources ?? {}
131
+ const prevMaxOut = previous?.maxOutputTokens ?? {}
132
+ const windows: Record<string, number> = {}
133
+ const sources: Record<string, ContextWindowSource> = {}
134
+ const maxOut: Record<string, number> = {}
135
+ for (const [model, meta] of Object.entries(metadata)) {
136
+ // Buang entri untuk model yang tak lagi ditawarkan endpoint (basi).
137
+ // Model di luar metadata tetap dipertahankan dari config lama di bawah.
138
+ const prevSource = prevSources[model]
139
+ const userOwned =
140
+ prevSource === undefined ? prevWindows[model] !== undefined : prevSource === "user"
141
+ if (meta.contextLength !== undefined && !userOwned) {
142
+ windows[model] = meta.contextLength
143
+ sources[model] = "endpoint"
144
+ }
145
+ if (meta.maxOutputTokens !== undefined) maxOut[model] = meta.maxOutputTokens
146
+ }
147
+ // Pertahankan yang tak dilaporkan endpoint kali ini (endpoint bisa parsial:
148
+ // sebagian gateway hanya menyertakan kapasitas untuk model yang baru
149
+ // ditambahkan). `live` hanya dipakai untuk membuang model yang endpoint
150
+ // TIDAK menawarkan lagi — bukan untuk membuang nilai yang belum dilaporkan.
151
+ const live = new Set(detected.models)
152
+ for (const [model, tokens] of Object.entries(prevWindows)) {
153
+ if (live.size > 0 && !live.has(model)) continue
154
+ if (windows[model] === undefined) {
155
+ windows[model] = tokens
156
+ sources[model] = prevSources[model] ?? "user"
157
+ }
158
+ }
159
+ for (const [model, tokens] of Object.entries(prevMaxOut)) {
160
+ if (live.size > 0 && !live.has(model)) continue
161
+ if (maxOut[model] === undefined) maxOut[model] = tokens
162
+ }
163
+ return {
164
+ ...(Object.keys(windows).length > 0 ? { contextWindows: windows } : {}),
165
+ ...(Object.keys(sources).length > 0 ? { contextWindowSources: sources } : {}),
166
+ ...(Object.keys(maxOut).length > 0 ? { maxOutputTokens: maxOut } : {}),
167
+ }
168
+ }
169
+
109
170
  export async function detectAndSave(
110
171
  baseUrl: string,
111
172
  apiKey: string,
@@ -158,6 +219,9 @@ export async function detectAndSave(
158
219
  // ini tiap add/edit me-reset effort ke default diam-diam ("hilang").
159
220
  const previous = prevCfg.providers.find((p) => p.id === uniqId)
160
221
  const prevEffort = previous?.reasoningEffort
222
+ // Kapasitas dari metadata endpoint (bila dilaporkan) diutamakan atas tebakan
223
+ // statis, tapi tak pernah menimpa nilai `user` — lihat mergeDiscoveredCapacity.
224
+ const capacity = mergeDiscoveredCapacity(previous, detected)
161
225
  const entry: ProviderEntry = {
162
226
  id: uniqId,
163
227
  baseUrl,
@@ -165,6 +229,7 @@ export async function detectAndSave(
165
229
  models: detected.models,
166
230
  providerHint: detected.providerHint,
167
231
  ...(previous?.contextWindows ? { contextWindows: previous.contextWindows } : {}),
232
+ ...capacity,
168
233
  ...(prevEffort ? { reasoningEffort: prevEffort } : {}),
169
234
  }
170
235
  await saveProvider(entry, opts)
@@ -250,7 +315,14 @@ export async function refreshProviderModels(
250
315
  try {
251
316
  const detected = await detectModels(p.baseUrl, apiKey)
252
317
  if (detected.models.length) {
253
- updated.set(p.id, { ...p, models: detected.models, providerHint: detected.providerHint })
318
+ // Metadata kapasitas ikut disegarkan: gateway yang menambah/mengubah
319
+ // model akan memperbarui jendela konteksnya tanpa user menyentuh apa pun.
320
+ updated.set(p.id, {
321
+ ...p,
322
+ models: detected.models,
323
+ providerHint: detected.providerHint,
324
+ ...mergeDiscoveredCapacity(p, detected),
325
+ })
254
326
  } else if (detected.authFailed) {
255
327
  // Key ditolak (401/403): bedakan dari provider tanpa endpoint
256
328
  // (Anthropic-style, tetap diam) — operator harus periksa key-nya.
@@ -290,6 +362,11 @@ export async function refreshProviderModels(
290
362
  if (nu) {
291
363
  p.models = nu.models
292
364
  p.providerHint = nu.providerHint
365
+ // Kapasitas hanya ditulis bila hasil merge memang punya isi — tanpa
366
+ // guard ini provider tanpa metadata akan menghapus nilai lama.
367
+ if (nu.contextWindows) p.contextWindows = nu.contextWindows
368
+ if (nu.contextWindowSources) p.contextWindowSources = nu.contextWindowSources
369
+ if (nu.maxOutputTokens) p.maxOutputTokens = nu.maxOutputTokens
293
370
  changed = true
294
371
  }
295
372
  }
@@ -262,12 +262,23 @@ export function createResponsesProvider(config: ResponsesConfig): ModelProvider
262
262
  // budget/cost (buta spend). Bentuk shape sama dengan adapter
263
263
  // openai-compat agar kolektor tak perlu tahu provider.
264
264
  const u = resp?.usage as
265
- | { input_tokens?: unknown; output_tokens?: unknown; total_tokens?: unknown }
265
+ | {
266
+ input_tokens?: unknown
267
+ output_tokens?: unknown
268
+ total_tokens?: unknown
269
+ cached_tokens?: unknown
270
+ input_tokens_details?: { cached_tokens?: unknown }
271
+ }
266
272
  | undefined
267
273
  if (
268
274
  u &&
269
275
  (u.input_tokens != null || u.output_tokens != null || u.total_tokens != null)
270
276
  ) {
277
+ // Fase 7: cache-hit alias (input_tokens_details.cached_tokens / cached_tokens).
278
+ const cached =
279
+ (u.input_tokens_details?.cached_tokens !== undefined
280
+ ? u.input_tokens_details.cached_tokens
281
+ : u.cached_tokens) ?? undefined
271
282
  yield {
272
283
  type: "extension",
273
284
  kind: "usage",
@@ -276,6 +287,7 @@ export function createResponsesProvider(config: ResponsesConfig): ModelProvider
276
287
  outputTokens:
277
288
  typeof u.output_tokens === "number" ? u.output_tokens : undefined,
278
289
  totalTokens: typeof u.total_tokens === "number" ? u.total_tokens : undefined,
290
+ ...(typeof cached === "number" ? { cacheReadTokens: cached } : {}),
279
291
  },
280
292
  }
281
293
  }
@@ -891,6 +891,22 @@ export function loadPresentationEvents(id: string, cwd?: string): DomainEvent[]
891
891
  return loadPresentationEventsWithStats(id, cwd).events
892
892
  }
893
893
 
894
+ /** Hitung kompaksi durable via kolom `type` (jangan muat+decode seluruh
895
+ * presentation_events — resume dulu tarik semua baris hanya untuk hitung revisi). */
896
+ export function countDurableCompactionEvents(sessionId: string, cwd?: string): number {
897
+ const db = open(cwd)
898
+ try {
899
+ const row = db
900
+ .prepare(
901
+ "SELECT COUNT(*) AS n FROM presentation_events WHERE session_id = ? AND type = 'context.compacted'",
902
+ )
903
+ .get(sessionId) as { n: number } | null
904
+ return row?.n ?? 0
905
+ } finally {
906
+ db.close()
907
+ }
908
+ }
909
+
894
910
  // P2.12 — baca rentang untuk replay konsumen. WAJIB ada: `loadPresentationEvents`
895
911
  // memuat SELURUH riwayat ke memori, dan konsumen yang attach dengan offset lama
896
912
  // akan menarik seluruh sesi (ratusan ribu baris) hanya untuk 200 event terakhir
@@ -1557,6 +1573,46 @@ export function loadSession(
1557
1573
  }
1558
1574
  }
1559
1575
 
1576
+ /**
1577
+ * Meta sesi TANPA decode histori pesan. Jalur resume hanya butuh turnCount/
1578
+ * workMode/system/cwd; histori penuh dibaca SEKALI lewat `loadThreadHistoryWithSeq`.
1579
+ * Memisahkan meta dari decode menutup bug perf: dulu resume memanggil `loadSession`
1580
+ * (decode SEMUA messages) + `loadThreadHistoryWithSeq` (decode lagi) — dua kali
1581
+ * JSON.parse seluruh histori.
1582
+ */
1583
+ export function loadSessionMeta(
1584
+ id: string,
1585
+ cwd?: string,
1586
+ ): {
1587
+ system?: string
1588
+ cwd?: string
1589
+ turnCount: number
1590
+ workMode?: "DISCUSS" | "PLAN" | "EXECUTE"
1591
+ } | null {
1592
+ const db = open(cwd)
1593
+ try {
1594
+ const sess = db.prepare("SELECT system, cwd, work_mode FROM sessions WHERE id = ?").get(id) as {
1595
+ system: string
1596
+ cwd: string
1597
+ work_mode: string | null
1598
+ } | null
1599
+ if (!sess) return null
1600
+ const turnRow = db
1601
+ .prepare("SELECT MAX(turn_idx) AS m FROM turns WHERE session_id = ?")
1602
+ .get(id) as { m: number | null } | null
1603
+ const turnCount = turnRow?.m != null ? turnRow.m + 1 : 0
1604
+ const workMode = parseStoredWorkMode(sess.work_mode)
1605
+ return {
1606
+ system: sess.system,
1607
+ cwd: sess.cwd,
1608
+ turnCount,
1609
+ ...(workMode ? { workMode } : {}),
1610
+ }
1611
+ } finally {
1612
+ db.close()
1613
+ }
1614
+ }
1615
+
1560
1616
  // P2.4: baca histori SATU thread (ketat: tanpa NULL transisional).
1561
1617
  // Beda dari loadSession (kompatibel): di sini thread scope eksplisit.
1562
1618
  export function loadThreadHistory(sessionId: string, threadId: string, cwd?: string): unknown[] {
@@ -615,7 +615,7 @@ export function attachSimpleLogger(bus: UiBus, opts: SimpleOptions = {}): () =>
615
615
  const lines = renderActivityPanel(snap.activities, { maxRows: 10 })
616
616
  if (lines.length > 0) {
617
617
  wErr(c.muted(`${t("panel.header")}\n`))
618
- for (const line of lines) wErr(c.muted(`${line}\n`))
618
+ for (const line of lines) wErr(`${line}\n`)
619
619
  }
620
620
  }
621
621
  }