minicode-ai 0.14.2 → 0.14.4

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 (76) 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 +167 -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 +216 -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.ts +73 -0
  48. package/src/policy/context.ts +55 -3
  49. package/src/policy/ratelimit.ts +47 -18
  50. package/src/policy/usage.ts +15 -8
  51. package/src/policy/work-mode.ts +1 -1
  52. package/src/providers/guards.ts +61 -4
  53. package/src/providers/provision.ts +12 -3
  54. package/src/providers/router.ts +56 -23
  55. package/src/session/context-projection.ts +14 -0
  56. package/src/session/context-selector.ts +58 -20
  57. package/src/session/persistence.ts +1 -1
  58. package/src/task/model.ts +21 -70
  59. package/src/tools/task.ts +7 -0
  60. package/src/ui/footer.ts +131 -66
  61. package/src/ui/i18n/en.ts +0 -11
  62. package/src/ui/i18n/id.ts +0 -11
  63. package/src/ui/input/input.ts +2 -3
  64. package/src/ui/input/prompt-engine.ts +2 -5
  65. package/src/ui/screens/model-manager.ts +3 -10
  66. package/src/ui/screens/picker.ts +23 -40
  67. package/src/ui/screens/provider-manager.ts +12 -33
  68. package/src/ui/screens/wizard.ts +0 -1
  69. package/src/ui/tui/app.ts +10 -6
  70. package/src/ui/tui/transcript.ts +48 -8
  71. package/vendor/minicore/VENDOR.md +9 -5
  72. package/vendor/minicore/src/core/loop.ts +47 -12
  73. package/vendor/minicore/src/core/provider.ts +4 -0
  74. package/vendor/minicore/src/core/session.ts +14 -1
  75. package/docs/OUTPUT_ARCHITECTURE_AUDIT.md +0 -489
  76. package/docs/OUTPUT_IMPLEMENTATION_PLAN.md +0 -391
package/docs/tools.md CHANGED
@@ -1,4 +1,4 @@
1
- # Tools (37)
1
+ # Tools
2
2
 
3
3
  Referensi lengkap tool Minicode: filesystem, search, exec, git, web, memory, plan, agents, MCP, LSP.
4
4
 
@@ -90,7 +90,7 @@ Batas `BASH_BACKGROUND_MAX_JOBS`; semua job mati saat CLI keluar.
90
90
 
91
91
  | Tool | Fungsi | Catatan |
92
92
  |---|---|---|
93
- | `delegate_task` | Sub-agen isolasi `explore`/`plan` | Pool 3, isolasi context/memory/signal/budget, abort-aware. Di-gate (TTY prompt, non-TTY tolak). Anak SELALU mode `auto` tanpa ask: tanpa MCP/commit/memory-tulis/todo-tulis/job/nesting; plan-parent dipaksa explore. Satu approval = delegasi ini saja; turn gagal setelah anak committed → warning `[recovery]` anti re-delegasi buta |
93
+ | `delegate_task` | Sub-agen isolasi `explore`/`plan` | Pool 3, context/memory/signal/budget terpisah, abort-aware. Di ASK, TTY meminta persetujuan dan non-TTY menolak; AUTO berjalan tanpa prompt bila authorized. Parent PermissionMode `ASK`/`plan`/`readonly` memaksa `explore`; WorkMode `DISCUSS`/`PLAN` membatasi tool anak ke subset eksplorasi. Anak tidak mendapat MCP, commit, memory/todo writes, nesting, atau job background. Satu approval mencakup satu delegasi; turn gagal setelah anak committed → warning `[recovery]` |
94
94
 
95
95
  ## MCP
96
96
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "minicode-ai",
3
- "version": "0.14.2",
3
+ "version": "0.14.4",
4
4
  "description": "Coding agent CLI built on the frozen MiniCore runtime — tools, MCP, LSP, sandbox, shadow-git checkpoints",
5
5
  "type": "module",
6
6
  "private": false,
@@ -48,20 +48,22 @@
48
48
  "vendor/minicore/src",
49
49
  "vendor/minicore/package.json",
50
50
  "vendor/minicore/VENDOR.md",
51
+ "scripts/vendor-overlays/minicode-context.patch",
51
52
  "README.md",
52
53
  "LICENSE",
53
- "docs/OUTPUT_ARCHITECTURE_AUDIT.md",
54
54
  "docs/OUTPUT_EVENT_MODEL.md",
55
- "docs/OUTPUT_IMPLEMENTATION_PLAN.md",
56
55
  "docs/OUTPUT_PROTOCOL_SPEC.md",
57
56
  "docs/OUTPUT_RENDERING_SPEC.md",
58
57
  "docs/OUTPUT_UX_RULES.md",
58
+ "docs/SUMMARY.md",
59
59
  "docs/TERMINAL_CONTRACT.md",
60
+ "docs/UI_RENDER_PIPELINE.md",
60
61
  "docs/agents.md",
61
62
  "docs/architecture.md",
62
63
  "docs/choosing-mode.md",
63
64
  "docs/cli.md",
64
65
  "docs/concepts.md",
66
+ "docs/CONTROL-PLANE-MAP.md",
65
67
  "docs/config-providers.md",
66
68
  "docs/config.md",
67
69
  "docs/contributing.md",
@@ -0,0 +1,109 @@
1
+ diff --git a/vendor/minicore/src/core/loop.ts b/vendor/minicore/src/core/loop.ts
2
+ index 11ea5eb..56e4f54 100644
3
+ --- a/vendor/minicore/src/core/loop.ts
4
+ +++ b/vendor/minicore/src/core/loop.ts
5
+ @@ -48,3 +48,6 @@ export async function executeTurn(
6
+ - // maxProviderRetries); the policy flag is set when a policy compaction
7
+ - // is attempted (once per turn, guarding against per-step O(n) repeats).
8
+ - let compactedForBudget = false;
9
+ + // maxProviderRetries); policy compaction has a small per-turn hard cap,
10
+ + // guarding against repeated no-progress work after tool output grows.
11
+ + // Permit one follow-up policy compaction after tool results grow the current
12
+ + // turn. The hard cap prevents a no-progress compactor from looping forever.
13
+ + let budgetCompactions = 0;
14
+ + const maxBudgetCompactions = 2;
15
+ @@ -52,0 +56,13 @@ export async function executeTurn(
16
+ + // Resolve once for the turn: model switching is a between-turn operation,
17
+ + // so budget admission and output accounting must use the same model window.
18
+ + let contextWindowTokens = s.contextWindowTokens;
19
+ + try {
20
+ + const resolved = s.contextWindowTokensForModel?.(s.state.model);
21
+ + if (typeof resolved === "number" && Number.isSafeInteger(resolved) && resolved > 0) {
22
+ + contextWindowTokens = resolved;
23
+ + }
24
+ + } catch {
25
+ + // Keep the validated numeric fallback; a metadata lookup must not bypass
26
+ + // the kernel's context guard.
27
+ + }
28
+ +
29
+ @@ -62 +78 @@ export async function executeTurn(
30
+ - limitTokens: s.contextWindowTokens,
31
+ + limitTokens: contextWindowTokens,
32
+ @@ -64 +80,2 @@ export async function executeTurn(
33
+ - if (s.budget.shouldCompact(pressure) && !compactedForBudget) {
34
+ + let reduced = false;
35
+ + if (s.budget.shouldCompact(pressure) && budgetCompactions < maxBudgetCompactions) {
36
+ @@ -67 +84 @@ export async function executeTurn(
37
+ - compactedForBudget = true;
38
+ + budgetCompactions++;
39
+ @@ -72 +89 @@ export async function executeTurn(
40
+ - const reduced = s.store.messages.length < before;
41
+ + reduced = s.store.messages.length < before;
42
+ @@ -79 +96 @@ export async function executeTurn(
43
+ - limitTokens: s.contextWindowTokens,
44
+ + limitTokens: contextWindowTokens,
45
+ @@ -82,2 +99,9 @@ export async function executeTurn(
46
+ - if (pressure === "critical" && compactedForBudget) {
47
+ - throw new AgentError("budget_exceeded", "context window exceeded after compaction");
48
+ + if (s.budget.shouldCompact(pressure) && reduced && budgetCompactions < maxBudgetCompactions) {
49
+ + continue;
50
+ + }
51
+ + if (pressure === "critical") {
52
+ + const used = contextTokens(s);
53
+ + throw new AgentError(
54
+ + "budget_exceeded",
55
+ + `context window exceeded after compaction (${used} > ${contextWindowTokens}); reduce the prompt, tool output, or fixed instructions, or select a model with a larger context window`,
56
+ + );
57
+ @@ -179 +203,5 @@ export async function executeTurn(
58
+ - throw new AgentError("budget_exceeded", errorMessage ?? "context too large", { cause: providerError ?? action });
59
+ + throw new AgentError(
60
+ + "budget_exceeded",
61
+ + `context window exceeded after recovery compaction: ${errorMessage ?? "context too large"}`,
62
+ + { cause: providerError ?? action },
63
+ + );
64
+ @@ -303,0 +332,3 @@ function toProviderError(error: unknown, signal: AbortSignal): ProviderError {
65
+ + // Host-side admission guards (context budget, shutdown, etc.) preserve their
66
+ + // deterministic AgentError identity instead of entering provider retry.
67
+ + if (error instanceof AgentError) throw error;
68
+ @@ -336 +367,5 @@ function contextTokens(s: SessionInternal): number {
69
+ - return estimateSessionContext(s.store, s.system, s.registry.list(), s.estimator);
70
+ + let system = s.system;
71
+ + try {
72
+ + system = s.systemForContext?.() ?? system;
73
+ + } catch {}
74
+ + return estimateSessionContext(s.store, system, s.registry.list(), s.estimator);
75
+ diff --git a/vendor/minicore/src/core/provider.ts b/vendor/minicore/src/core/provider.ts
76
+ index 489e244..63afbbc 100644
77
+ --- a/vendor/minicore/src/core/provider.ts
78
+ +++ b/vendor/minicore/src/core/provider.ts
79
+ @@ -37,0 +38,4 @@ export interface ModelProvider {
80
+ + /** Optional host metadata for per-model admission budgets. */
81
+ + contextWindowTokensForModel?(model?: string): number;
82
+ + /** Canonical provider/model key for per-model estimator state. */
83
+ + contextModelIdentity?(model?: string): string | undefined;
84
+ diff --git a/vendor/minicore/src/core/session.ts b/vendor/minicore/src/core/session.ts
85
+ index 6e183d7..df80273 100644
86
+ --- a/vendor/minicore/src/core/session.ts
87
+ +++ b/vendor/minicore/src/core/session.ts
88
+ @@ -37,0 +38,2 @@ export interface SessionConfig {
89
+ + /** Dynamic system prompt for host-controlled per-turn mode hints. */
90
+ + systemForContext?: () => string;
91
+ @@ -61,0 +64,3 @@ export interface SessionConfig {
92
+ + /** Resolve a model-specific context limit for each run. Additive app seam;
93
+ + * the numeric option remains the fallback for existing library callers. */
94
+ + contextWindowTokensForModel?: (model?: string) => number;
95
+ @@ -140,0 +146 @@ export interface SessionInternal {
96
+ + readonly systemForContext?: () => string;
97
+ @@ -149,0 +156 @@ export interface SessionInternal {
98
+ + readonly contextWindowTokensForModel?: (model?: string) => number;
99
+ @@ -180,0 +188 @@ export function createSession(config: SessionConfig): Session {
100
+ + systemForContext: config.systemForContext,
101
+ @@ -189,0 +198 @@ export function createSession(config: SessionConfig): Session {
102
+ + contextWindowTokensForModel: config.contextWindowTokensForModel,
103
+ @@ -213 +222,5 @@ export function createSession(config: SessionConfig): Session {
104
+ - return estimateSessionContext(store, config.system, impl.registry.list(), impl.estimator);
105
+ + let system = config.system
106
+ + try {
107
+ + system = config.systemForContext?.() ?? system
108
+ + } catch {}
109
+ + return estimateSessionContext(store, system, impl.registry.list(), impl.estimator);
@@ -1,5 +1,6 @@
1
1
  import { createOpenAICompatProvider } from "#minicore/providers/openai-compat.ts"
2
2
  import { loadConfig, type MinicodeConfig } from "../config.ts"
3
+ import { resolveContextWindow } from "../policy/context-window.ts"
3
4
  import type { RateLimiter } from "../policy/ratelimit.ts"
4
5
  import { createAnthropicProvider } from "../providers/anthropic.ts"
5
6
  import { buildProviderListAsync } from "../providers/build.ts"
@@ -7,6 +8,7 @@ import { withStreamGuards } from "../providers/guards.ts"
7
8
  import { createRouterProvider } from "../providers/router.ts"
8
9
 
9
10
  let currentRouter: ReturnType<typeof createRouterProvider> | null = null
11
+ let currentConfig: MinicodeConfig | null = null
10
12
 
11
13
  /** Setup tanpa provider apa pun. Dilempar (bukan exit di sini) agar
12
14
  * composition root bisa memilih presentasi: REPL/one-shot mencetak pesan
@@ -24,9 +26,17 @@ export async function reloadProviders(
24
26
  ): Promise<void> {
25
27
  if (!currentRouter) return
26
28
  const cfg = await loadConfig(cwd, { allowLocal: opts.allowLocal })
29
+ if (currentConfig)
30
+ currentConfig.providers.splice(0, currentConfig.providers.length, ...cfg.providers)
31
+ else currentConfig = cfg
27
32
  const providers = await buildProviderListAsync(cfg)
28
- const r = currentRouter as unknown as { updateProviders: (list: unknown[]) => void }
33
+ const r = currentRouter as unknown as {
34
+ updateProviders: (list: unknown[]) => void
35
+ }
29
36
  if (typeof r.updateProviders === "function") r.updateProviders(providers)
37
+ // Keep the router's live context resolver. The composition root may have
38
+ // wrapped it with a session override and warning policy; replacing it here
39
+ // would make /sync silently reset child sessions to config-only limits.
30
40
  }
31
41
 
32
42
  export async function createProviderLayer(opts: {
@@ -43,6 +53,7 @@ export async function createProviderLayer(opts: {
43
53
  setupWhenEmpty?: () => Promise<boolean>
44
54
  }): Promise<{ cfg: MinicodeConfig; router: ReturnType<typeof createRouterProvider> }> {
45
55
  const cfg = await loadConfig(opts.cwd, { allowLocal: opts.allowLocalConfig })
56
+ currentConfig = cfg
46
57
  type Provider = ReturnType<typeof createOpenAICompatProvider>
47
58
  let providers = await buildProviderListAsync(cfg)
48
59
 
@@ -119,6 +130,7 @@ export async function createProviderLayer(opts: {
119
130
  }
120
131
  const router = createRouterProvider({
121
132
  providers,
133
+ contextWindowTokensForModel: (model) => resolveContextWindow(cfg, model).tokens,
122
134
  ...(opts.rateLimiter ? { limiter: opts.rateLimiter } : {}),
123
135
  })
124
136
  currentRouter = router
@@ -1,23 +1,31 @@
1
1
  import { resolve } from "node:path"
2
2
  import type { RecoveryAction } from "#minicore/core/errors.ts"
3
- import { ProviderError } from "#minicore/core/errors.ts"
3
+ import { AgentError, ProviderError } from "#minicore/core/errors.ts"
4
4
  import type {
5
5
  ModelProvider,
6
6
  PermissionHandler,
7
+ ProviderEvent,
7
8
  Session,
8
9
  SessionConfig,
9
10
  } from "#minicore/core/index.ts"
10
11
  import { createSession as createCoreSession } from "#minicore/core/index.ts"
11
12
  import { defaultRecoveryPolicy } from "#minicore/core/recovery.ts"
12
- import { estimateMessages, estimateSystem, estimateTools } from "#minicore/core/tokens.ts"
13
+ import {
14
+ estimateMessages,
15
+ estimateSystem,
16
+ estimateTools,
17
+ safeStringify,
18
+ } from "#minicore/core/tokens.ts"
13
19
  import { LIMITS } from "../constants.ts"
14
20
  // [Area B · PROMPT 06] Binding persetujuan ↔ operasi yang dieksekusi.
15
21
  import { ApprovalBindingStore, verifyApprovalBinding } from "../policy/approval-binding.ts"
16
22
  import {
17
23
  buildSystemPrompt,
18
24
  createCalibratedEstimator,
25
+ createIncrementalTokenCounter,
19
26
  minicodeEstimator,
20
27
  } from "../policy/context.ts"
28
+ import { DEFAULT_FALLBACK_CONTEXT_WINDOW_TOKENS } from "../policy/context-window.ts"
21
29
  import { parallelExecutor } from "../policy/executor.ts"
22
30
  import {
23
31
  canonicalArgsHash,
@@ -118,6 +126,8 @@ export async function createMinicodeSession(
118
126
  > & {
119
127
  systemExtra?: string
120
128
  cwd?: string
129
+ /** Live model identity used for per-model token calibration. */
130
+ getContextModel?: () => string | undefined
121
131
  permissionMode?: "auto" | "readonly" | "plan" | "allow-all" | "ask" | "allowlist"
122
132
  /** S2.1: WorkMode adalah lapis intent TERPISAH dari PermissionMode.
123
133
  * Default EXECUTE = perilaku lama (base handler apa adanya). Nilai selain
@@ -230,6 +240,7 @@ export async function createMinicodeSession(
230
240
  workMode: _workMode,
231
241
  allowLocalConfig,
232
242
  systemExtra: _extra,
243
+ getContextModel: _getContextModel,
233
244
  provider,
234
245
  onPermissions,
235
246
  onWorkMode,
@@ -261,6 +272,8 @@ export async function createMinicodeSession(
261
272
  const t = (rest as Record<string, unknown>).timeoutMs
262
273
  if (typeof t === "number" && (!Number.isFinite(t) || t < 0))
263
274
  delete (rest as Record<string, unknown>).timeoutMs
275
+ else if (typeof t === "number" && t > LIMITS.MAX_TIMER_DELAY_MS)
276
+ (rest as Record<string, unknown>).timeoutMs = LIMITS.MAX_TIMER_DELAY_MS
264
277
  const m = (rest as Record<string, unknown>).maxSteps
265
278
  if (typeof m !== "number" || !Number.isFinite(m) || m < 1)
266
279
  delete (rest as Record<string, unknown>).maxSteps
@@ -444,17 +457,172 @@ export async function createMinicodeSession(
444
457
  const planHint = pm === "plan" ? PLAN_PM_HINT : ""
445
458
  return baseSystem + planHint + workModeHint(wm)
446
459
  }
460
+ // [CTX] Kalibrasi dipisahkan per identitas model agar berpindah ke model
461
+ // lain tidak membawa rasio tokenizer/provider dari sesi sebelumnya.
462
+ const calibrations = new Map<string, ReturnType<typeof createCalibratedEstimator>>()
463
+ const contextModelKey = (model?: string): string => model?.trim() || "__default__"
464
+ const calibrationFor = (model?: string) => {
465
+ const key = contextModelKey(model)
466
+ let calibration = calibrations.get(key)
467
+ if (!calibration) {
468
+ calibration = createCalibratedEstimator()
469
+ calibrations.set(key, calibration)
470
+ }
471
+ return calibration
472
+ }
473
+ let activeContextModel = opts.model
474
+ const currentContextModel = () => {
475
+ const selected = _getContextModel?.() ?? activeContextModel
476
+ try {
477
+ return provider.contextModelIdentity?.(selected) ?? selected
478
+ } catch {
479
+ return selected
480
+ }
481
+ }
482
+ const calibratedEstimator = (text: string) =>
483
+ calibrationFor(currentContextModel()).estimator(text)
484
+ let lastRequestLocalEstimate: { model: string; tokens: number } | undefined
485
+ const contextWindowForModel = (model?: string): number => {
486
+ const explicit = rest.contextWindowTokens
487
+ if (
488
+ !rest.contextWindowTokensForModel &&
489
+ typeof explicit === "number" &&
490
+ Number.isSafeInteger(explicit) &&
491
+ explicit > 0
492
+ ) {
493
+ return explicit
494
+ }
495
+ try {
496
+ const resolved =
497
+ rest.contextWindowTokensForModel?.(model) ?? provider.contextWindowTokensForModel?.(model)
498
+ if (typeof resolved === "number" && Number.isSafeInteger(resolved) && resolved > 0)
499
+ return resolved
500
+ } catch {}
501
+ return typeof explicit === "number" && Number.isSafeInteger(explicit) && explicit > 0
502
+ ? explicit
503
+ : DEFAULT_FALLBACK_CONTEXT_WINDOW_TOKENS
504
+ }
447
505
  const dynamicProvider: ModelProvider = {
448
506
  id: provider.id,
449
507
  get models() {
450
508
  return provider.models
451
509
  },
452
- stream(req, signal) {
453
- return provider.stream({ ...req, system: computeLiveSystem() }, signal)
510
+ async *stream(req, signal) {
511
+ const system = computeLiveSystem()
512
+ const model = req.model ?? currentContextModel()
513
+ let limit = contextWindowForModel(model)
514
+ const estimatorModel = model?.includes("::") ? model : (currentContextModel() ?? model)
515
+ let calibration = calibrationFor(estimatorModel)
516
+ let estimator = calibration.estimator
517
+ let admittedFactor = calibration.snapshot().factor
518
+ const tools = req.tools ?? []
519
+ const localInputTokens =
520
+ estimateMessages(req.messages, minicodeEstimator) +
521
+ estimateSystem(system, minicodeEstimator) +
522
+ estimateTools(tools, minicodeEstimator)
523
+ let inputTokens =
524
+ estimateMessages(req.messages, estimator) +
525
+ estimateSystem(system, estimator) +
526
+ estimateTools(tools, estimator)
527
+ lastRequestLocalEstimate = {
528
+ model: contextModelKey(estimatorModel),
529
+ tokens: localInputTokens,
530
+ }
531
+ // Sisakan ruang framing dan respons minimum. Tanpa reserve, request yang
532
+ // tepat menyentuh batas provider tetap ditolak setelah biaya dibakar.
533
+ let reserve = Math.min(256, Math.floor(limit / 10))
534
+ let outputAllowance = Math.max(0, limit - inputTokens - reserve)
535
+ let outputTokens = 0
536
+ const streamedText = createIncrementalTokenCounter()
537
+ const streamedReasoning = createIncrementalTokenCounter()
538
+ const contextLimitEvent = (input: number, output: number): ProviderEvent => ({
539
+ type: "extension",
540
+ kind: "context-window-limit",
541
+ data: { inputTokens: input, outputTokens: output, windowTokens: limit },
542
+ })
543
+ const overflow = (actualInput = inputTokens, actualOutput = outputTokens) =>
544
+ new AgentError(
545
+ "budget_exceeded",
546
+ `context window budget reached for the selected model: input ${actualInput} + output ${actualOutput} exceeds ${limit} tokens; reduce the prompt/tools or select a larger context window`,
547
+ )
548
+ const requestController = new AbortController()
549
+ const forwardAbort = () => requestController.abort(signal.reason)
550
+ if (signal.aborted) forwardAbort()
551
+ else signal.addEventListener("abort", forwardAbort, { once: true })
552
+ try {
553
+ if (inputTokens + reserve >= limit) {
554
+ const error = overflow(inputTokens, 0)
555
+ requestController.abort(error)
556
+ yield contextLimitEvent(inputTokens, 0)
557
+ throw error
558
+ }
559
+
560
+ for await (const event of provider.stream({ ...req, system }, requestController.signal)) {
561
+ if (event.type === "extension" && event.kind === "effective-model") {
562
+ const effective = event.data as { effective?: unknown; provider?: unknown } | undefined
563
+ if (
564
+ typeof effective?.effective === "string" &&
565
+ typeof effective.provider === "string"
566
+ ) {
567
+ // Router fallback/substitution is a different model boundary.
568
+ // Keep the smaller of the requested and effective capacities so
569
+ // a failed primary provider cannot silently widen the budget.
570
+ const effectiveModel = `${effective.provider}::${effective.effective}`
571
+ const effectiveLimit = contextWindowForModel(effectiveModel)
572
+ limit = Math.min(limit, effectiveLimit)
573
+ calibration = calibrationFor(effectiveModel)
574
+ estimator = calibration.estimator
575
+ admittedFactor = calibration.snapshot().factor
576
+ inputTokens = Math.max(
577
+ inputTokens,
578
+ estimateMessages(req.messages, estimator) +
579
+ estimateSystem(system, estimator) +
580
+ estimateTools(tools, estimator),
581
+ )
582
+ reserve = Math.min(256, Math.floor(limit / 10))
583
+ outputAllowance = Math.max(0, limit - inputTokens - reserve)
584
+ }
585
+ }
586
+ const currentFactor = calibration.snapshot().factor
587
+ if (currentFactor > admittedFactor) {
588
+ inputTokens = Math.max(
589
+ inputTokens,
590
+ estimateMessages(req.messages, estimator) +
591
+ estimateSystem(system, estimator) +
592
+ estimateTools(tools, estimator),
593
+ )
594
+ reserve = Math.min(256, Math.floor(limit / 10))
595
+ outputAllowance = Math.max(0, limit - inputTokens - reserve)
596
+ }
597
+ admittedFactor = currentFactor
598
+ let added = 0
599
+ if (event.type === "text") {
600
+ added = streamedText.add(event.text, calibration.snapshot().factor)
601
+ } else if (event.type === "tool_call") {
602
+ added = estimator(event.name) + estimator(safeStringify(event.args))
603
+ } else if (event.type === "extension" && event.kind === "reasoning") {
604
+ const reasoning = (event.data as { text?: unknown } | undefined)?.text
605
+ if (typeof reasoning === "string") {
606
+ added = streamedReasoning.add(reasoning, calibration.snapshot().factor)
607
+ }
608
+ }
609
+ if (added > 0 && outputTokens + added > outputAllowance) {
610
+ const error = overflow(inputTokens, outputTokens + added)
611
+ requestController.abort(error)
612
+ yield contextLimitEvent(inputTokens, outputTokens + added)
613
+ throw error
614
+ }
615
+ outputTokens += added
616
+ yield event as ProviderEvent
617
+ }
618
+ } finally {
619
+ signal.removeEventListener("abort", forwardAbort)
620
+ if (!requestController.signal.aborted) {
621
+ requestController.abort(new AgentError("aborted", "provider stream closed"))
622
+ }
623
+ }
454
624
  },
455
625
  }
456
- // [Fase 3 / CTX-03] Estimator terkalibrasi terhadap usage aktual provider.
457
- const calibration = createCalibratedEstimator()
458
626
  const liveToolSchemas = () =>
459
627
  registeredTools.map((t) => ({
460
628
  name: t.name,
@@ -463,11 +631,14 @@ export async function createMinicodeSession(
463
631
  }))
464
632
  const core = createCoreSession({
465
633
  ...rest,
634
+ contextWindowTokens: contextWindowForModel(currentContextModel()),
635
+ contextWindowTokensForModel: contextWindowForModel,
466
636
  provider: dynamicProvider,
467
637
  system: computeLiveSystem(),
638
+ systemForContext: computeLiveSystem,
468
639
  permissions: withEffectBarrier,
469
640
  permissionMode: livePermissionMode,
470
- estimator: calibration.estimator,
641
+ estimator: calibratedEstimator,
471
642
  recovery: cappedRecovery,
472
643
  executor: parallelExecutor({
473
644
  concurrency: safeConcurrency ?? LIMITS.EXECUTOR_CONCURRENCY,
@@ -479,14 +650,29 @@ export async function createMinicodeSession(
479
650
  })
480
651
  core.events.on("provider:extension", (e) => {
481
652
  if (e.kind !== "usage") return
482
- const d = e.data as { inputTokens?: number; totalTokens?: number } | undefined
483
- if (!d) return
484
- const baseLocal =
653
+ const d = e.data as
654
+ | {
655
+ inputTokens?: number
656
+ totalTokens?: number
657
+ model?: string
658
+ provider?: string
659
+ source?: string
660
+ }
661
+ | undefined
662
+ if (!d || d.source === "compaction") return
663
+ const reportedModel =
664
+ typeof d.model === "string" && d.model.length > 0
665
+ ? typeof d.provider === "string" && d.provider.length > 0
666
+ ? `${d.provider}::${d.model}`
667
+ : d.model
668
+ : currentContextModel()
669
+ const fallbackLocal =
485
670
  estimateSystem(computeLiveSystem(), minicodeEstimator) +
486
671
  estimateTools(liveToolSchemas(), minicodeEstimator) +
487
672
  estimateMessages(core.state.history, minicodeEstimator)
488
- calibration.observeProviderUsage({
489
- localEstimatedTokens: baseLocal,
673
+ const localEstimatedTokens = lastRequestLocalEstimate?.tokens ?? fallbackLocal
674
+ calibrationFor(reportedModel).observeProviderUsage({
675
+ localEstimatedTokens,
490
676
  providerInputTokens: d.inputTokens,
491
677
  providerTotalTokens: d.totalTokens,
492
678
  })
@@ -530,5 +716,22 @@ export async function createMinicodeSession(
530
716
  return origRun(input, runOpts)
531
717
  }
532
718
  }
533
- return core
719
+ return {
720
+ get state() {
721
+ return core.state
722
+ },
723
+ get events() {
724
+ return core.events
725
+ },
726
+ get contextTokens() {
727
+ return core.contextTokens
728
+ },
729
+ run(input: string, runOpts?: { signal?: AbortSignal; model?: string }) {
730
+ activeContextModel = runOpts?.model ?? opts.model
731
+ return core.run(input, runOpts)
732
+ },
733
+ abort() {
734
+ core.abort()
735
+ },
736
+ }
534
737
  }
package/src/config.ts CHANGED
@@ -37,6 +37,10 @@ export interface ProviderEntry {
37
37
  /** API key. Kosong bila provider memakai OAuth (`auth: "oauth"`). */
38
38
  apiKey: string
39
39
  models: string[]
40
+ /** Context window tokens by exact model id; `*` is a provider-wide fallback.
41
+ * Values are operator-supplied because gateways often expose aliases whose
42
+ * actual model capacity cannot be discovered reliably at runtime. */
43
+ contextWindows?: Record<string, number>
40
44
  providerHint?: string
41
45
  /** Knob generik reasoning effort — dipetakan per-wire (openai reasoning_effort, anthropic thinking, gemini). */
42
46
  reasoningEffort?: "low" | "medium" | "high"
@@ -104,9 +108,32 @@ export function localConfigNotice(cwd: string, allowLocal: boolean): string | un
104
108
  export function normalizeConfig(raw: unknown): MinicodeConfig {
105
109
  const cfg = raw as Record<string, unknown>
106
110
  const providers = Array.isArray(cfg?.providers)
107
- ? (cfg.providers as ProviderEntry[]).filter(
108
- (p) => p && typeof p.id === "string" && typeof p.baseUrl === "string",
109
- )
111
+ ? (cfg.providers as ProviderEntry[])
112
+ .filter((p) => p && typeof p.id === "string" && typeof p.baseUrl === "string")
113
+ .map((p) => {
114
+ const source = (p as { contextWindows?: unknown }).contextWindows
115
+ const validContextWindows: Record<string, number> = {}
116
+ if (source && typeof source === "object" && !Array.isArray(source)) {
117
+ for (const [model, tokens] of Object.entries(source as Record<string, unknown>)) {
118
+ if (
119
+ (model === "*" || model.length > 0) &&
120
+ typeof tokens === "number" &&
121
+ Number.isSafeInteger(tokens) &&
122
+ tokens > 0 &&
123
+ tokens <= 20_000_000
124
+ ) {
125
+ validContextWindows[model] = tokens
126
+ }
127
+ }
128
+ }
129
+ const contextWindows =
130
+ Object.keys(validContextWindows).length > 0 ? validContextWindows : undefined
131
+ return {
132
+ ...p,
133
+ contextWindows:
134
+ contextWindows && Object.keys(contextWindows).length > 0 ? contextWindows : undefined,
135
+ }
136
+ })
110
137
  : []
111
138
  // Server MCP sah bila punya `command` (stdio) ATAU `url` (HTTP). Entri tanpa
112
139
  // keduanya dibuang di sini supaya kegagalan tampak saat config dibaca, bukan
package/src/constants.ts CHANGED
@@ -1,5 +1,7 @@
1
1
  /** Centralized magic numbers — full English naming. Single source for limits. */
2
2
  export const LIMITS = {
3
+ /** Maksimum timer JS; nilai lebih besar diklem agar tidak berubah jadi 1ms. */
4
+ MAX_TIMER_DELAY_MS: 2_147_483_647,
3
5
  /** File size limits */
4
6
  READ_FILE_MAX_BYTES: 2_000_000,
5
7
  WRITE_FILE_MAX_CHARS: 5_000_000,
@@ -22,7 +22,7 @@ export interface HostedSessionInfo {
22
22
  runStatus: string | null
23
23
  presentationHead: number
24
24
  /** P0.2 (Q5): WorkMode efektif sesi yang di-host (bila diketahui host). */
25
- workMode?: "DISCUSS" | "PLAN" | "EXECUTE"
25
+ workMode?: "DISCUSS" | "PLAN" | "BUILD"
26
26
  /**
27
27
  * P2.3: hasil recovery run terakhir yang durable (mis. UNKNOWN setelah
28
28
  * restart). Operator dapat membedakan LIVE vs reconstructed/interrupted
package/src/mcp/server.ts CHANGED
@@ -240,9 +240,8 @@ async function invokeTool(
240
240
  * the whole process. As a durable task namespace that is unsafe: it is a
241
241
  * constant, so every MCP context that has ever run against the same `root`
242
242
  * shares one identity space, and it can also collide with anything else that
243
- * ever uses that string. `TASK_ARCHITECTURE_AUDIT.md` already flagged the
244
- * underlying problem - `mcp serve` is single-tenant, and concurrent requests
245
- * overwrite each other's files last-writer-wins.
243
+ * ever uses that string. Because `mcp serve` is single-tenant, concurrent
244
+ * requests must not share this durable namespace and overwrite files last-writer-wins.
246
245
  *
247
246
  * WHAT THIS IDENTIFIES. One server instance = one logical MCP context. That
248
247
  * matches this implementation's actual lifecycle, which was verified rather