minicode-ai 0.27.0 → 0.30.0

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 (145) hide show
  1. package/README.md +1 -1
  2. package/cli/args.ts +184 -57
  3. package/cli/bootstrap/presentation-bridge.ts +383 -0
  4. package/cli/commands/acp.ts +45 -29
  5. package/cli/commands/auth.ts +17 -17
  6. package/cli/commands/config.ts +65 -66
  7. package/cli/commands/daemon.ts +44 -16
  8. package/cli/commands/doctor.ts +18 -15
  9. package/cli/commands/exec.ts +41 -89
  10. package/cli/commands/mcp.ts +25 -8
  11. package/cli/commands/memory.ts +16 -24
  12. package/cli/commands/openwork.ts +6 -2
  13. package/cli/commands/pricing.ts +14 -12
  14. package/cli/commands/providers.ts +36 -57
  15. package/cli/commands/sessions.ts +24 -18
  16. package/cli/commands/skills.ts +33 -18
  17. package/cli/commands/stats.ts +15 -30
  18. package/cli/commands.ts +73 -44
  19. package/cli/context-command-run.ts +30 -18
  20. package/cli/index.ts +51 -100
  21. package/cli/model-manager.ts +7 -2
  22. package/cli/provider-manager.ts +33 -18
  23. package/cli/router.ts +41 -12
  24. package/cli/setup.ts +241 -485
  25. package/cli/tui.ts +303 -100
  26. package/docs/TERMINAL_CONTRACT.md +258 -73
  27. package/docs/UI_RENDER_PIPELINE.md +9 -1
  28. package/docs/agents.md +2 -2
  29. package/docs/architecture.md +17 -32
  30. package/docs/choosing-mode.md +7 -9
  31. package/docs/cli.md +3 -5
  32. package/docs/environment.md +1 -2
  33. package/docs/policy-sandbox.md +3 -3
  34. package/docs/repl.md +2 -2
  35. package/docs/tools.md +1 -1
  36. package/package.json +1 -1
  37. package/src/app/mentions.ts +60 -13
  38. package/src/app/rag-layer.ts +17 -20
  39. package/src/app/session.ts +144 -68
  40. package/src/app/tool-layer.ts +38 -16
  41. package/src/config.ts +64 -104
  42. package/src/constants.ts +4 -6
  43. package/src/daemon/capability.ts +8 -1
  44. package/src/daemon/control.ts +3 -2
  45. package/src/daemon/server.ts +57 -0
  46. package/src/hooks/run.ts +73 -41
  47. package/src/lib/atomic-write.ts +31 -59
  48. package/src/lib/db-path.ts +47 -0
  49. package/src/lib/process-tree.ts +62 -0
  50. package/src/lib/public-http.ts +88 -23
  51. package/src/lib/trusted-exec.ts +48 -1
  52. package/src/lsp/client.ts +62 -17
  53. package/src/mcp/client.ts +85 -38
  54. package/src/mcp/http-transport.ts +28 -16
  55. package/src/mcp/transport.ts +42 -6
  56. package/src/memory/files.ts +164 -134
  57. package/src/memory/vector.ts +371 -420
  58. package/src/policy/approval-binding.ts +122 -45
  59. package/src/policy/bash-guard.ts +18 -7
  60. package/src/policy/compaction.ts +51 -18
  61. package/src/policy/context-window.ts +31 -8
  62. package/src/policy/context.ts +62 -36
  63. package/src/policy/executor.ts +17 -7
  64. package/src/policy/jail.ts +6 -8
  65. package/src/policy/microcompact.ts +4 -3
  66. package/src/policy/no-progress.ts +103 -22
  67. package/src/policy/permission.ts +36 -12
  68. package/src/policy/tool-contract.ts +74 -416
  69. package/src/policy/usage.ts +24 -2
  70. package/src/policy/verifier.ts +50 -2
  71. package/src/policy/work-mode.ts +28 -54
  72. package/src/presentation/adapter.ts +24 -16
  73. package/src/providers/anthropic.ts +16 -5
  74. package/src/providers/provision.ts +6 -56
  75. package/src/providers/responses.ts +38 -7
  76. package/src/providers/sse-bounds.ts +6 -4
  77. package/src/runtime/execution-journal.ts +4 -29
  78. package/src/sandbox/docker.ts +58 -34
  79. package/src/sandbox/mask.ts +157 -0
  80. package/src/sandbox/os.ts +107 -73
  81. package/src/session/authority.ts +29 -10
  82. package/src/session/context-projection.ts +6 -5
  83. package/src/session/db/runs.ts +397 -0
  84. package/src/session/db/schema.ts +358 -0
  85. package/src/session/open-work.ts +26 -3
  86. package/src/session/persistence.ts +74 -998
  87. package/src/session/verification.ts +170 -20
  88. package/src/skills/loader.ts +31 -16
  89. package/src/task/db/schema.ts +325 -0
  90. package/src/task/db/types.ts +269 -0
  91. package/src/task/production-scheduler.ts +13 -1
  92. package/src/task/store.ts +92 -1021
  93. package/src/telemetry/trace.ts +27 -0
  94. package/src/tools/ask_user.ts +1 -13
  95. package/src/tools/bash.ts +27 -92
  96. package/src/tools/code_run.ts +1 -1
  97. package/src/tools/context.ts +133 -0
  98. package/src/tools/delete_file.ts +3 -2
  99. package/src/tools/edit.ts +37 -72
  100. package/src/tools/git.ts +86 -9
  101. package/src/tools/glob.ts +26 -51
  102. package/src/tools/grep.ts +67 -57
  103. package/src/tools/hashline.ts +4 -0
  104. package/src/tools/lsp.ts +11 -8
  105. package/src/tools/mcp_call.ts +9 -6
  106. package/src/tools/memory.ts +56 -33
  107. package/src/tools/move_file.ts +36 -10
  108. package/src/tools/patch.ts +13 -21
  109. package/src/tools/read_file.ts +18 -25
  110. package/src/tools/read_image.ts +35 -33
  111. package/src/tools/submit_result.ts +29 -8
  112. package/src/tools/task.ts +74 -17
  113. package/src/tools/todo.ts +181 -74
  114. package/src/tools/web_fetch.ts +42 -2
  115. package/src/tools/web_search.ts +47 -8
  116. package/src/ui/approval/prompt.ts +206 -47
  117. package/src/ui/assistant/turn-status.ts +14 -5
  118. package/src/ui/footer.ts +4 -4
  119. package/src/ui/i18n/en.ts +47 -23
  120. package/src/ui/i18n/id.ts +76 -51
  121. package/src/ui/input/input.ts +4 -3
  122. package/src/ui/input/prompt-engine.ts +35 -5
  123. package/src/ui/render/activity-panel.ts +4 -2
  124. package/src/ui/render/hint.ts +7 -4
  125. package/src/ui/render/list.ts +32 -0
  126. package/src/ui/render/markdown.ts +6 -3
  127. package/src/ui/render/notice.ts +19 -0
  128. package/src/ui/render/table-grid.ts +7 -8
  129. package/src/ui/render/theme.ts +113 -39
  130. package/src/ui/runtime/screen.ts +74 -10
  131. package/src/ui/runtime/spinner.ts +19 -6
  132. package/src/ui/screens/dialog.ts +19 -1
  133. package/src/ui/screens/form.ts +80 -35
  134. package/src/ui/screens/model-manager.ts +82 -61
  135. package/src/ui/screens/output-window.ts +263 -0
  136. package/src/ui/screens/picker.ts +94 -81
  137. package/src/ui/screens/provider-manager.ts +100 -69
  138. package/src/ui/screens/wizard.ts +117 -28
  139. package/src/ui/tui/app.ts +196 -33
  140. package/src/ui/tui/transcript.ts +158 -25
  141. package/vendor/minicore/src/core/executor.ts +2 -0
  142. package/vendor/minicore/src/core/loop.ts +17 -6
  143. package/vendor/minicore/src/core/session.ts +3 -0
  144. package/vendor/minicore/src/core/tool.ts +1 -0
  145. package/vendor/minicore/src/providers/openai-compat.ts +35 -19
@@ -13,13 +13,16 @@
13
13
  import type { PermissionHandler, ToolCall } from "#minicore"
14
14
  import type { PermissionMode } from "./permission.ts"
15
15
  import {
16
- LOCAL_READONLY_TOOL_NAMES,
17
16
  NETWORK_READ_TOOL_NAMES,
18
17
  PLAN_EXTRA_TOOL_NAMES,
19
18
  READONLY_TOOL_NAMES,
20
19
  } from "./tool-contract.ts"
21
20
 
22
- export type WorkMode = "DISCUSS" | "PLAN" | "EXECUTE"
21
+ export type WorkMode = "DISCUSS" | "EXECUTE"
22
+
23
+ export function normalizeWorkMode(mode?: WorkMode | string | null): WorkMode {
24
+ return mode === "EXECUTE" ? "EXECUTE" : "DISCUSS"
25
+ }
23
26
 
24
27
  /**
25
28
  * Error kebijakan startup saat `--verify` digabung dengan WorkMode non-EXECUTE
@@ -28,7 +31,7 @@ export type WorkMode = "DISCUSS" | "PLAN" | "EXECUTE"
28
31
  */
29
32
  export class VerifyWorkModeConflictError extends Error {
30
33
  readonly workMode: Exclude<WorkMode, "EXECUTE">
31
- constructor(workMode: Exclude<WorkMode, "EXECUTE">) {
34
+ constructor(workMode: Exclude<WorkMode, "EXECUTE"> = "DISCUSS") {
32
35
  super(
33
36
  `[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
37
  )
@@ -63,7 +66,7 @@ export type WorkModePermissionHandler = PermissionHandler &
63
66
  function requireModeControl(base: PermissionHandler): ModeControl {
64
67
  const ctrl = base as Partial<ModeControl>
65
68
  if (typeof ctrl.__getMode !== "function" || typeof ctrl.__setMode !== "function") {
66
- // Fail-closed: tanpa kontrol mode, ceiling PLAN tak bisa mengikuti mode PM
69
+ // Fail-closed: tanpa kontrol mode, ceiling DISCUSS tak bisa mengikuti mode PM
67
70
  // saat runtime — lebih baik menolak konstruksi daripada memakai snapshot
68
71
  // basi yang bisa mencabut ceiling setelah `__setMode`.
69
72
  throw new Error(
@@ -74,27 +77,21 @@ function requireModeControl(base: PermissionHandler): ModeControl {
74
77
  }
75
78
 
76
79
  // Seleksi ceiling per mode PM SAAT INI (bukan snapshot konstruksi):
77
- // DISCUSS selalu read-only; PLAN memakai ceiling read-only saat PM sudah lebih
78
- // ketat (`allowlist`/`readonly`), selain itu ceiling plan.
79
- function ceilingFor(workMode: RestrictiveWorkMode, baseMode: PermissionMode): "readonly" | "plan" {
80
- if (workMode === "DISCUSS") return "readonly"
80
+ // DISCUSS (gabungan diskusi + perencanaan) memakai ceiling read-only saat PM sudah
81
+ // lebih ketat (`allowlist`/`readonly`), selain itu ceiling plan (18 readonly + 3 plan extra).
82
+ function ceilingFor(_workMode: RestrictiveWorkMode, baseMode: PermissionMode): "readonly" | "plan" {
81
83
  return baseMode === "allowlist" || baseMode === "readonly" ? "readonly" : "plan"
82
84
  }
83
85
 
84
86
  // Tool Access Policy — ceiling capability WorkMode dihitung LANGSUNG dari
85
- // registry kontrak tool (satu sumber), bukan lewat handler PermissionMode
86
- // `readonly`/`plan` (yang dulunya diduplikasi sebagai primitif ceiling).
87
- //
88
- // - DISCUSS → tepat 16 `LOCAL_READONLY_TOOL_NAMES` (network read ditolak
89
- // terpisah di check() dengan alasan khusus).
90
- // - PLAN → 18 `READONLY_TOOL_NAMES` (+ `PLAN_EXTRA_TOOL_NAMES`), atau
87
+ // registry kontrak tool (satu sumber):
88
+ // - DISCUSS → 18 `READONLY_TOOL_NAMES` (+ `PLAN_EXTRA_TOOL_NAMES`), atau
91
89
  // read-only saja bila PM sudah `readonly`/`allowlist` (lebih ketat).
92
90
  function ceilingAllows(
93
91
  name: string,
94
- workMode: RestrictiveWorkMode,
92
+ _workMode: RestrictiveWorkMode,
95
93
  baseMode: PermissionMode,
96
94
  ): boolean {
97
- if (workMode === "DISCUSS") return LOCAL_READONLY_TOOL_NAMES.includes(name)
98
95
  if (baseMode === "readonly" || baseMode === "allowlist") return READONLY_TOOL_NAMES.has(name)
99
96
  return READONLY_TOOL_NAMES.has(name) || PLAN_EXTRA_TOOL_NAMES.has(name)
100
97
  }
@@ -105,13 +102,7 @@ export function createWorkModePermissionHandler(opts: {
105
102
  root?: string
106
103
  }): WorkModePermissionHandler {
107
104
  const ctrl = requireModeControl(opts.base)
108
- // P12 S2.2 Slice 5 — WorkMode MUTABLE (bukan snapshot konstruksi). Kernel
109
- // mengkonsultasi handler per tool call, jadi __setWorkMode runtime menggeser
110
- // ceiling tanpa mengganti handler. EXECUTE = delegasi base (perilaku identik).
111
- let wm: WorkMode = opts.workMode
112
- // Handler deny ceiling — hanya untuk seam describeDenial (alasan yang sama
113
- // dengan perilaku lama: "read-only mode" / "plan mode"). Tidak ada lagi
114
- // handler PermissionMode `readonly`/`plan` sebagai primitif ceiling.
105
+ let wm: WorkMode = normalizeWorkMode(opts.workMode)
115
106
  const ceilingDenyHandlers: Record<"readonly" | "plan", PermissionHandler> = {
116
107
  readonly: {
117
108
  async check() {
@@ -176,41 +167,27 @@ export function createWorkModePermissionHandler(opts: {
176
167
 
177
168
  const handler = {
178
169
  async check(call: ToolCall, _deps?: unknown): Promise<"allow" | "deny"> {
179
- // EXECUTE = identitas: delegasi penuh ke base tanpa ceiling (perilaku
180
- // lama). Memungkinkan upgrade runtime EXECUTE→DISCUSS/PLAN walau komposisi
181
- // awal EXECUTE (compose mengembalikan base) TIDAK menyediakan __setWorkMode —
182
- // jalur ini dipakai createMinicodeSession yang SELALU membungkus.
183
170
  if (wm === "EXECUTE") return opts.base.check(call, _deps as never)
184
- // P0.2 (Q2): DISCUSS melarang network read (`web_fetch`, `web_search`) di
185
- // SEMUA PermissionMode tanpa pernah menyentuh base (0 prompt).
186
- if (wm === "DISCUSS" && NETWORK_READ_TOOL_NAMES.has(call.name)) {
187
- rememberDeny(call, discussNetworkDeny)
188
- return "deny"
189
- }
190
- // Baca mode PM sekarang: `__setMode` runtime menggeser seleksi ceiling
191
- // (PLAN × allowlist/readonly = read-only), bukan membekukan yang lama.
192
- // Baca WorkMode sekarang (`wm` mutable) — __setWorkMode runtime menggeser
193
- // ceiling DISCUSS↔PLAN.
194
171
  const baseMode = ctrl.__getMode()
195
172
  const ceilingKind = ceilingFor(wm, baseMode)
196
- // Ceiling DULU (Tool Access Policy): operasi di luar ceiling tak boleh
197
- // memicu prompt base (requirement S2.1 #1) — ceiling tak pernah prompt.
198
173
  if (!ceilingAllows(call.name, wm, baseMode)) {
199
174
  rememberDeny(call, ceilingDenyHandlers[ceilingKind])
200
175
  return "deny"
201
176
  }
202
- // Di dalam ceiling: keputusan mutlak handler PM asli (termasuk prompt
203
- // gated `delegate_task` untuk PLAN × auto/ask — owner decision S2.1).
204
- const decision = await opts.base.check(call, _deps as never)
177
+ // Di dalam ceiling DISCUSS: paksa enforceJail=true agar `DISCUSS × unrestricted`
178
+ // tidak melompati jail workspace maupun proteksi berkas sensitif `.env` (INV-WM1).
179
+ const baseDeps =
180
+ _deps && typeof _deps === "object"
181
+ ? { ...(_deps as Record<string, unknown>), enforceJail: true }
182
+ : { enforceJail: true }
183
+ const decision = await opts.base.check(call, baseDeps as never)
205
184
  if (decision !== "allow") {
206
185
  rememberDeny(call, opts.base)
207
186
  return decision
208
187
  }
209
- // P0.2 (Q2): di PLAN, network read (`web_fetch`, `web_search`) hanya boleh
210
- // berjalan lewat approval prompt eksplisit (headless/tanpa __promptGated =
211
- // fail-closed deny). Dipasang SETELAH `opts.base.check` agar Invariant #1
212
- // terjaga (WorkMode tak pernah melonggarkan penolakan base).
213
- if (wm === "PLAN" && NETWORK_READ_TOOL_NAMES.has(call.name)) {
188
+ // Di DISCUSS, network read (`web_fetch`, `web_search`) wajib melewati
189
+ // gerbang approval prompt eksplisit (headless/tanpa __promptGated = fail-closed deny).
190
+ if (NETWORK_READ_TOOL_NAMES.has(call.name)) {
214
191
  if (typeof ctrl.__promptGated !== "function") {
215
192
  rememberDeny(call, discussNetworkDeny)
216
193
  return "deny"
@@ -225,19 +202,15 @@ export function createWorkModePermissionHandler(opts: {
225
202
  ctrl.__setMode(m)
226
203
  },
227
204
  __getMode(): PermissionMode {
228
- // Metadata yang dilaporkan tetap mode PM asli; WorkMode tidak bocor.
229
205
  return ctrl.__getMode()
230
206
  },
231
207
  __setWorkMode(m: WorkMode): void {
232
- wm = m
208
+ wm = normalizeWorkMode(m)
233
209
  },
234
210
  __getWorkMode(): WorkMode {
235
211
  return wm
236
212
  },
237
213
  describeDenial(call: ToolCall): Promise<string | undefined> | string | undefined {
238
- // Cek owner map lebih dulu: bila call ditolak saat mode non-EXECUTE lalu
239
- // WorkMode diubah ke EXECUTE sebelum describeDenial dipanggil, alasan
240
- // deny tetap diambil dari pemilik yang mencatatnya.
241
214
  const owner = takeDenyOwner(call)
242
215
  if (owner) {
243
216
  const seam = owner.describeDenial
@@ -259,9 +232,10 @@ export function composeWorkModePermissionHandler(opts: {
259
232
  base: PermissionHandler
260
233
  root?: string
261
234
  }): PermissionHandler {
262
- if (opts.workMode === "EXECUTE") return opts.base
235
+ const wm = normalizeWorkMode(opts.workMode)
236
+ if (wm === "EXECUTE") return opts.base
263
237
  return createWorkModePermissionHandler({
264
- workMode: opts.workMode,
238
+ workMode: wm,
265
239
  base: opts.base,
266
240
  root: opts.root,
267
241
  })
@@ -14,7 +14,7 @@
14
14
 
15
15
  import { type CanonicalAssignmentTable, decodeCanonicalAssignments } from "../task/assignment.ts"
16
16
  import { classifyToolResult, denyReasonOf, summarizeArgs } from "../telemetry/trace.ts"
17
- import { normalizeTodos } from "../tools/todo.ts"
17
+ import { extractCommittedTodoStatuses, normalizeTodos } from "../tools/todo.ts"
18
18
  import type {
19
19
  ApprovalEventHook,
20
20
  ApprovalHookEvent,
@@ -483,6 +483,7 @@ export function createPresentationAdapter(
483
483
  * back to the legacy positional shape rather than claim `v2`.
484
484
  */
485
485
  assignment?: CanonicalAssignmentTable,
486
+ resultContent?: unknown,
486
487
  ):
487
488
  | {
488
489
  steps: PlanStep[]
@@ -501,21 +502,27 @@ export function createPresentationAdapter(
501
502
  return undefined
502
503
  }
503
504
  if (list.length === 0) return undefined
504
- const steps: PlanStep[] = list.map((t, i) => ({
505
- stepId: String(i + 1),
506
- // PHASE 4A.4 - the declared id is passed to the resolver as a HINT only. It
507
- // is kept on the step solely when TaskStore confirms that row exists for
508
- // this session; otherwise it is stripped below, so an unverified id can
509
- // never appear on a published plan.
510
- ...(t.taskId ? { taskId: t.taskId } : {}),
511
- title: t.content,
512
- status:
513
- t.status === "completed" || t.status === "cancelled" || t.status === "blocked"
514
- ? t.status
515
- : t.status === "in_progress"
516
- ? "active"
517
- : "pending",
518
- }))
505
+ const committed = extractCommittedTodoStatuses(resultContent, list.length)
506
+ const steps: PlanStep[] = list.map((t, i) => {
507
+ const effectiveStatus = committed?.[i] ?? t.status
508
+ return {
509
+ stepId: String(i + 1),
510
+ // PHASE 4A.4 - the declared id is passed to the resolver as a HINT only. It
511
+ // is kept on the step solely when TaskStore confirms that row exists for
512
+ // this session; otherwise it is stripped below, so an unverified id can
513
+ // never appear on a published plan.
514
+ ...(t.taskId ? { taskId: t.taskId } : {}),
515
+ title: t.content,
516
+ status:
517
+ effectiveStatus === "completed" ||
518
+ effectiveStatus === "cancelled" ||
519
+ effectiveStatus === "blocked"
520
+ ? effectiveStatus
521
+ : effectiveStatus === "in_progress"
522
+ ? "active"
523
+ : "pending",
524
+ }
525
+ })
519
526
  const status = steps.every((step) => step.status === "completed")
520
527
  ? "completed"
521
528
  : steps.every((step) => step.status === "cancelled")
@@ -946,6 +953,7 @@ export function createPresentationAdapter(
946
953
  call.args,
947
954
  planSessionId,
948
955
  decodeCanonicalAssignments(result.content, { sessionId: planSessionId }),
956
+ result.content,
949
957
  )
950
958
  if (plan) {
951
959
  publish({
@@ -104,13 +104,24 @@ export function createAnthropicProvider(config: AnthropicConfig): ModelProvider
104
104
  const effectiveMaxTokens = legacyBudget
105
105
  ? Math.max(config.maxTokens ?? 8192, legacyBudget + 1024)
106
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).
107
+ // Fase 7: sematkan cache_control pada blok konten stabil terakhir.
110
108
  const messagesPayload = toAnthropicMessages(request.messages)
111
109
  if (enableCache && messagesPayload.length >= 2) {
112
- const lastStable = messagesPayload[messagesPayload.length - 2] as Record<string, unknown>
113
- lastStable.cache_control = { type: "ephemeral" }
110
+ const lastStable = messagesPayload[messagesPayload.length - 2] as {
111
+ role: string
112
+ content: unknown
113
+ }
114
+ if (Array.isArray(lastStable.content) && lastStable.content.length > 0) {
115
+ const lastBlock = lastStable.content[lastStable.content.length - 1] as Record<
116
+ string,
117
+ unknown
118
+ >
119
+ lastBlock.cache_control = { type: "ephemeral" }
120
+ } else if (typeof lastStable.content === "string") {
121
+ lastStable.content = [
122
+ { type: "text", text: lastStable.content, cache_control: { type: "ephemeral" } },
123
+ ]
124
+ }
114
125
  }
115
126
  const body = JSON.stringify({
116
127
  model: request.model ?? config.defaultModel ?? config.models[0],
@@ -12,10 +12,10 @@ import {
12
12
  type MinicodeConfig,
13
13
  normalizeConfig,
14
14
  type ProviderEntry,
15
+ updateConfigAtomic,
15
16
  withConfigLock,
16
17
  writeConfigAtomic,
17
18
  } from "../config.ts"
18
- import { atomicWriteText } from "../lib/atomic-write.ts"
19
19
  import { getSecret } from "../lib/keystore.ts"
20
20
  import { clearDetectCache, type DetectResult, detectModels } from "./detect.ts"
21
21
  import { GATEWAY_PRESETS } from "./presets.ts"
@@ -39,44 +39,11 @@ export async function saveProvider(
39
39
  const needsKey = entry.auth !== "oauth"
40
40
  if (!entry.id || !entry.baseUrl || (needsKey && !entry.apiKey))
41
41
  throw new Error("provider id/baseUrl/apiKey required")
42
- // A provider may temporarily have no models: `/model` can remove the last
43
- // entry before `/sync` discovers a new one.
44
- const path =
45
- (opts.global ?? true) ? globalConfigPath() : resolve(opts.cwd ?? process.cwd(), LOCAL)
46
- // Kunci per-path SAMA dengan saveMcpServer/saveLspServer (audit #07):
47
- // tanpa ini baca-modifikasi-tulis paralel last-wins dan menelan entri.
48
- return withConfigLock(path, async () => {
49
- let cfg: MinicodeConfig = { providers: [] }
50
- try {
51
- cfg = normalizeConfig(JSON.parse(await readFile(path, "utf8")))
52
- } catch (e) {
53
- // Audit #08 P1: reset diam-diam saat berkas korup MENGHAPUS semua
54
- // provider + MCP/LSP/verify/allowlist lain saat tulis berikutnya.
55
- // ENOENT (belum ada file) = mulai kosong; korup = backup + gagal keras
56
- // seperti saveMcpServer; error lain (EACCES/…) = teruskan, jangan timpa.
57
- if (e instanceof SyntaxError) {
58
- const raw = await readFile(path, "utf8").catch(() => "")
59
- const backup = `${path}.corrupt.${Date.now()}`
60
- await atomicWriteText(backup, raw).catch(() => {})
61
- throw new Error(`config corrupt: ${path} — backup to ${backup}: ${e.message}`)
62
- }
63
- if ((e as NodeJS.ErrnoException).code !== "ENOENT") throw e
64
- }
42
+ // Entri baru menang mutlak agar perintah `auto` bisa menghapus kunci jendela sebelumnya.
43
+ return updateConfigAtomic(opts, (cfg) => {
65
44
  const idx = cfg.providers.findIndex((p) => p.id === entry.id)
66
- if (idx >= 0) {
67
- cfg.providers[idx] = {
68
- ...entry,
69
- // Entri baru menang MUTLAK atas nilai lama: `saveProvider` menulis apa
70
- // yang diberikan pemanggil. Sengaja tanpa "selamatkan nilai lama bila
71
- // entry tidak punya field ini" — versi lama melakukan itu dan membuat
72
- // `auto` (hapus jendela) mustahil: `assignContextWindow` sengaja MENGHAPUS
73
- // kunci, dan penyelamatan di sini membangkitkannya lagi sehingga perintah
74
- // "kembali ke otomatis" sukses di layar tapi tak berefek di config.
75
- // Pemanggil yang memang ingin mempertahankan nilai lama (mis. `sync`)
76
- // sudah membacanya sendiri lewat `loadConfig` sebelum menyusun entry.
77
- }
78
- } else cfg.providers.push(entry)
79
- await writeConfigAtomic(path, cfg)
45
+ if (idx >= 0) cfg.providers[idx] = { ...entry }
46
+ else cfg.providers.push(entry)
80
47
  })
81
48
  }
82
49
 
@@ -255,25 +222,8 @@ export async function detectAndSave(
255
222
  }
256
223
 
257
224
  export async function removeProvider(id: string, opts: { global?: boolean; cwd?: string } = {}) {
258
- const path =
259
- (opts.global ?? true) ? globalConfigPath() : resolve(opts.cwd ?? process.cwd(), LOCAL)
260
- return withConfigLock(path, async () => {
261
- let cfg: MinicodeConfig = { providers: [] }
262
- try {
263
- cfg = normalizeConfig(JSON.parse(await readFile(path, "utf8")))
264
- } catch (e) {
265
- // Sama seperti saveProvider (audit #08 P1): korup = backup + gagal
266
- // keras, bukan reset diam-diam yang menghapus seluruh config.
267
- if (e instanceof SyntaxError) {
268
- const raw = await readFile(path, "utf8").catch(() => "")
269
- const backup = `${path}.corrupt.${Date.now()}`
270
- await atomicWriteText(backup, raw).catch(() => {})
271
- throw new Error(`config corrupt: ${path} — backup to ${backup}: ${e.message}`)
272
- }
273
- if ((e as NodeJS.ErrnoException).code !== "ENOENT") throw e
274
- }
225
+ return updateConfigAtomic(opts, (cfg) => {
275
226
  cfg.providers = cfg.providers.filter((p) => p.id !== id)
276
- await writeConfigAtomic(path, cfg)
277
227
  })
278
228
  }
279
229
 
@@ -254,6 +254,28 @@ export function createResponsesProvider(config: ResponsesConfig): ModelProvider
254
254
  // response.completed → simpan id untuk chaining turn berikut.
255
255
  // Format nyata: {type:"response.completed", response:{id:"resp_…"}}.
256
256
  const dtype = data.type as string | undefined
257
+ if (dtype === "error" || dtype === "response.failed" || data.error) {
258
+ if (prev) lastResponseByModel.delete(modelKey)
259
+ const errRec =
260
+ (data.error as Record<string, unknown> | undefined) ??
261
+ (data.response as { error?: Record<string, unknown> } | undefined)?.error ??
262
+ data
263
+ const msg = String(errRec.message ?? errRec.msg ?? JSON.stringify(errRec))
264
+ const code = String(errRec.code ?? errRec.type ?? "")
265
+ const combined = `${msg} ${code}`
266
+ const cat = /auth|401|403|unauthorized/i.test(combined)
267
+ ? "auth"
268
+ : /quota_exceeded|insufficient_quota|balance/i.test(combined)
269
+ ? "unknown"
270
+ : /rate|429/i.test(combined)
271
+ ? "rate_limit"
272
+ : /context_length|maximum context|prompt is too long|context window/i.test(
273
+ msg,
274
+ )
275
+ ? "context_length_exceeded"
276
+ : "server"
277
+ throw new ProviderError(cat, cat === "unknown" ? `402: ${msg}` : `${cat}: ${msg}`)
278
+ }
257
279
  if (dtype === "response.completed" || dtype === "completed") {
258
280
  const resp = data.response as { id?: unknown; usage?: unknown } | undefined
259
281
  const rid = resp?.id ?? data.id
@@ -287,6 +309,7 @@ export function createResponsesProvider(config: ResponsesConfig): ModelProvider
287
309
  outputTokens:
288
310
  typeof u.output_tokens === "number" ? u.output_tokens : undefined,
289
311
  totalTokens: typeof u.total_tokens === "number" ? u.total_tokens : undefined,
312
+ cacheIncluded: true,
290
313
  ...(typeof cached === "number" ? { cacheReadTokens: cached } : {}),
291
314
  },
292
315
  }
@@ -390,14 +413,22 @@ export function createResponsesProvider(config: ResponsesConfig): ModelProvider
390
413
  typeof data.arguments === "string" ? data.arguments : "",
391
414
  )
392
415
  }
393
- const isToolEvent =
394
- dtype === "response.output_item.added" ||
395
- dtype === "response.function_call_arguments.delta" ||
396
- dtype === "response.function_call_arguments.done" ||
397
- dtype === "response.output_item.done" ||
398
- dtype === "function_call"
416
+ // Abaikan event lifecycle/summary `.done`/`.added` agar teks utuh di
417
+ // `response.output_text.done` tidak digandakan setelah `.delta`.
418
+ const isNonDeltaEvent =
419
+ typeof dtype === "string" &&
420
+ (dtype.endsWith(".done") ||
421
+ dtype.endsWith(".added") ||
422
+ dtype === "response.created" ||
423
+ dtype === "response.in_progress" ||
424
+ dtype === "response.completed" ||
425
+ dtype === "completed" ||
426
+ dtype === "response.incomplete" ||
427
+ dtype.startsWith("response.function_call") ||
428
+ dtype.startsWith("response.reasoning") ||
429
+ dtype === "function_call")
399
430
  const text =
400
- reasoningText || isToolEvent
431
+ reasoningText || isNonDeltaEvent
401
432
  ? undefined
402
433
  : ((typeof rawDelta === "string" ? rawDelta : undefined) ??
403
434
  drec.text ??
@@ -51,10 +51,7 @@ export function boundedSseBody(
51
51
  if (done) {
52
52
  reachedEof = true
53
53
  if (lineBytes !== 0 || frameBytes !== 0) {
54
- throw new ProviderError(
55
- "invalid_request",
56
- `${provider} returned an incomplete SSE frame`,
57
- )
54
+ throw new ProviderError("network", `${provider} returned an incomplete SSE frame`)
58
55
  }
59
56
  controller.close()
60
57
  release()
@@ -68,6 +65,11 @@ export function boundedSseBody(
68
65
  )
69
66
  }
70
67
  for (const byte of value) {
68
+ if (byte === 0x0d) {
69
+ // CRLF: 0x0D ikut dihitung frameBytes, bukan lineBytes
70
+ frameBytes++
71
+ continue
72
+ }
71
73
  if (byte === 0x0a) {
72
74
  frameBytes++
73
75
  if (frameBytes > maxFrameBytes) {
@@ -37,6 +37,7 @@ import { createHash } from "node:crypto"
37
37
  import { mkdirSync } from "node:fs"
38
38
  import { dirname } from "node:path"
39
39
  import { LIMITS } from "../constants.ts"
40
+ import { isSqliteContention, withSqliteBusyRetrySync } from "../lib/db-path.ts"
40
41
  import type { ExecutionEvent } from "./execution-events.ts"
41
42
 
42
43
  /** Versi skema jurnal (bentuk/encoding) — BUKAN executionVersion. */
@@ -336,36 +337,10 @@ export async function withJournalBusyRetry<T>(
336
337
  throw last instanceof Error ? last : new Error(`SQLITE_BUSY after ${attempts} attempts`)
337
338
  }
338
339
 
339
- // [Phase 12.3] Predikat kontensi SQLite transien lintas-proses — SELARAS dengan
340
- // persistence.ts / store.ts / vector.ts (Fase 12.2). Bun menaruh kode simbolik di
341
- // `error.code` (mis. "SQLITE_IOERR_TRUNCATE") dan teks manusiawi di `error.message`
342
- // ("disk I/O error"), jadi keduanya diperiksa. SENGAJA TIDAK di-retry:
343
- // SQLITE_CORRUPT / SQLITE_NOTADB / constraint error.
344
- const SQLITE_CONTENTION_RE =
345
- /SQLITE_BUSY|database is locked|database table is locked|SQLITE_LOCKED|SQLITE_PROTOCOL|SQLITE_IOERR_TRUNCATE|SQLITE_IOERR_SHMOPEN|SQLITE_IOERR_LOCK|disk I\/O error/i
346
-
347
- function isSqliteContention(e: unknown): boolean {
348
- const code = (e as { code?: unknown } | null)?.code
349
- if (typeof code === "string" && SQLITE_CONTENTION_RE.test(code)) return true
350
- const msg = e instanceof Error ? String(e.message ?? e) : String(e)
351
- return SQLITE_CONTENTION_RE.test(msg)
352
- }
353
-
354
- const DB_INIT_ATTEMPTS = 8
355
- const DB_INIT_BACKOFF_MS = 15
356
-
357
340
  /** Retry sinkron untuk setup PRAGMA/WAL saat open (jalur sinkron openExecutionJournal). */
358
- function withBusyRetrySync<T>(fn: () => T, attempts = DB_INIT_ATTEMPTS): T | null {
359
- for (let i = 0; i < attempts; i++) {
360
- try {
361
- return fn()
362
- } catch (e) {
363
- if (!isSqliteContention(e)) throw e
364
- Bun.sleepSync(DB_INIT_BACKOFF_MS * 2 ** i)
365
- }
366
- }
367
- return null
368
- }
341
+ const DB_INIT_ATTEMPTS = 8
342
+ const withBusyRetrySync = <T>(fn: () => T, attempts = DB_INIT_ATTEMPTS): T | null =>
343
+ withSqliteBusyRetrySync(fn, { attempts })
369
344
 
370
345
  function isCorruptionError(e: unknown): boolean {
371
346
  const msg = (e instanceof Error ? e.message : String(e)).toLowerCase()
@@ -4,6 +4,11 @@ import { resolve } from "node:path"
4
4
  import { LIMITS } from "../constants.ts"
5
5
  import { resolveTrustedExecutable } from "../lib/trusted-exec.ts"
6
6
  import { sanitizeSpawnEnv } from "../policy/scrub.ts"
7
+ import {
8
+ allocateEmptyMaskFile,
9
+ createCappedOutputCollector,
10
+ discoverSandboxMaskTargets,
11
+ } from "./mask.ts"
7
12
 
8
13
  let dockerOk: boolean | null = null
9
14
 
@@ -86,24 +91,26 @@ function removeContainer(name: string, env: NodeJS.ProcessEnv): Promise<boolean>
86
91
  })
87
92
  }
88
93
 
89
- // Run command di dalam container ephemeral, terisolasi (--network none,
90
- // memori/CPU cap), mount workspace read-write. Return { code, output }.
91
- export function runInDocker(
94
+ /** Susun argumen `docker run` termasuk masking `.minicode/` & `.env*` sensitif. */
95
+ export function buildDockerRunArgs(
92
96
  command: string,
93
97
  cwd: string,
94
- opts: DockerRunOptions = {},
95
- ): Promise<{ code: number | null; output: string }> {
96
- if (opts.signal?.aborted) {
97
- return Promise.resolve({ code: null, output: "[docker] aborted before container start" })
98
- }
98
+ opts: DockerRunOptions,
99
+ mask: {
100
+ containerName: string
101
+ hasMinicode: boolean
102
+ envFiles: readonly string[]
103
+ emptyMaskHostPath: string | null
104
+ },
105
+ ): string[] {
99
106
  const image = opts.image ?? process.env.MINICODE_SANDBOX_IMAGE ?? "node:22-alpine"
100
107
  const mount = workspaceMount(cwd)
101
- const containerName = `minicode-${randomUUID()}`
108
+ const net = opts.network === "bridge" ? "bridge" : "none"
102
109
  const args = [
103
110
  "run",
104
111
  "--rm",
105
112
  "--name",
106
- containerName,
113
+ mask.containerName,
107
114
  "-i",
108
115
  "--read-only",
109
116
  "--cap-drop",
@@ -116,10 +123,22 @@ export function runInDocker(
116
123
  "/tmp",
117
124
  "-v",
118
125
  `${mount}:${mount}`,
126
+ ]
127
+ // Sembunyikan state minicode (.minicode/) dan rahasia workspace (.env*) di dalam container.
128
+ if (mask.hasMinicode) {
129
+ args.push("--tmpfs", `${mount}/.minicode:ro,size=64k`)
130
+ }
131
+ if (mask.emptyMaskHostPath) {
132
+ const emptyMount = workspaceMount(mask.emptyMaskHostPath)
133
+ for (const rel of mask.envFiles) {
134
+ args.push("-v", `${emptyMount}:${mount}/${rel}:ro`)
135
+ }
136
+ }
137
+ args.push(
119
138
  "-w",
120
139
  mount,
121
140
  "--network",
122
- "none",
141
+ net,
123
142
  "--memory",
124
143
  opts.memory ?? process.env.MINICODE_SANDBOX_MEMORY ?? "512m",
125
144
  "--cpus",
@@ -130,7 +149,29 @@ export function runInDocker(
130
149
  "sh",
131
150
  "-c",
132
151
  command,
133
- ]
152
+ )
153
+ return args
154
+ }
155
+
156
+ // Run command di dalam container ephemeral, terisolasi (--network none default,
157
+ // memori/CPU cap, masking .env/.minicode), mount workspace read-write. Return { code, output }.
158
+ export function runInDocker(
159
+ command: string,
160
+ cwd: string,
161
+ opts: DockerRunOptions = {},
162
+ ): Promise<{ code: number | null; output: string }> {
163
+ if (opts.signal?.aborted) {
164
+ return Promise.resolve({ code: null, output: "[docker] aborted before container start" })
165
+ }
166
+ const containerName = `minicode-${randomUUID()}`
167
+ const targets = discoverSandboxMaskTargets(cwd)
168
+ const emptyMask = allocateEmptyMaskFile(targets.envFiles.length > 0)
169
+ const args = buildDockerRunArgs(command, cwd, opts, {
170
+ containerName,
171
+ hasMinicode: targets.hasMinicode,
172
+ envFiles: targets.envFiles,
173
+ emptyMaskHostPath: emptyMask.path,
174
+ })
134
175
 
135
176
  return new Promise((resolveResult) => {
136
177
  // env selalu disanitasi dari hasil merge final — secret (API_KEY/TOKEN/...)
@@ -141,25 +182,9 @@ export function runInDocker(
141
182
  env,
142
183
  })
143
184
  // Cap buffer saat streaming — output container raksasa tidak menumpuk di RAM.
144
- const OUT_CAP = LIMITS.DOCKER_OUTPUT_MAX_CHARS
145
- let out = "",
146
- err = ""
147
- let truncated = false
148
- const appendCapped = (target: "out" | "err", d: Buffer | string): void => {
149
- const s = typeof d === "string" ? d : d.toString()
150
- const cur = target === "out" ? out : err
151
- if (cur.length >= OUT_CAP) {
152
- truncated = true
153
- return
154
- }
155
- const room = OUT_CAP - cur.length
156
- if (s.length > room) truncated = true
157
- const piece = s.slice(0, Math.max(0, room))
158
- if (target === "out") out += piece
159
- else err += piece
160
- }
161
- p.stdout.on("data", (d) => appendCapped("out", d))
162
- p.stderr.on("data", (d) => appendCapped("err", d))
185
+ const collector = createCappedOutputCollector(LIMITS.DOCKER_OUTPUT_MAX_CHARS)
186
+ p.stdout.on("data", (d) => collector.append("out", d))
187
+ p.stderr.on("data", (d) => collector.append("err", d))
163
188
  const timeout =
164
189
  typeof opts.timeoutMs === "number" && Number.isFinite(opts.timeoutMs) && opts.timeoutMs > 0
165
190
  ? opts.timeoutMs
@@ -175,9 +200,8 @@ export function runInDocker(
175
200
  settled = true
176
201
  clearTimeout(t)
177
202
  signal?.removeEventListener("abort", onAbort)
178
- const marker = truncated ? "\n… [output truncated]" : ""
179
- const output = (out + (err ? `\n[stderr]\n${err}` : "")).trim() + marker + suffix
180
- resolveResult({ code, output })
203
+ emptyMask.cleanup()
204
+ resolveResult({ code, output: collector.format().output + suffix })
181
205
  }
182
206
  const requestStop = (reason: "timeout" | "abort") => {
183
207
  if (stopReason || settled) return