vibeostheog 0.22.0 → 0.22.6

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.
package/CHANGELOG.md CHANGED
@@ -1,3 +1,22 @@
1
+ ## 0.22.6
2
+ - feat: wire CostAnomalyDetector into tool-execute hook
3
+ - feat: replace TokenAnomalyDetector with CostAnomalyDetector
4
+ - feat: replace TokenAnomalyDetector with CostAnomalyDetector
5
+ - fix: bin/setup.js now delegates to deploy.mjs for proper plugin install
6
+ - fix: restore anomaly detector class in TS source, add mega regression tests
7
+ - fix: read path prefers global scratchpad over stale session-local copies
8
+ - fix: cross-session cache corruptions and hallucinations
9
+ - fix: VibeUltraX mode, anomaly token guard, regression suite
10
+ - chore: bump to 0.22.5
11
+ - chore: bump to 0.22.3
12
+ - chore: add regression tests for anomaly throttle permanent break fix
13
+ - chore: bump to 0.22.2
14
+ Merge pull request #103 from DrunkkToys/fix/anomaly-class-ts-source
15
+ Merge pull request #101 from DrunkkToys/fix/anomaly-throttle-permanent-break
16
+ Merge pull request #100 from DrunkkToys/fix/cross-session-corruptions
17
+ reorder policy comparison table by quality descending; deduplicate VibeUltraX formatting fix
18
+
19
+
1
20
  ## 0.20.16
2
21
  - fix: skip cache savings for free models + add modelCostPerTurn fallback + regression tests
3
22
  - fix: wire incrementTurnCounter into onToolExecuteAfter so session compaction fires at turn 7+
package/README.md CHANGED
@@ -35,22 +35,25 @@ Benchmarked on the DeepSeek v4 family. Prices based on 700 input + 300 output to
35
35
 
36
36
  ## Optimization Modes
37
37
 
38
- ### Mode Comparison — Sorted by Quality Descending
38
+ ### Policy Comparison — Sorted by Quality Descending
39
39
 
40
- | # | Mode | Pipeline | Quality vs Brain | Cost vs Brain | Cost/Turn | Saves |
41
- |---|------|----------|-----------------|--------------|-----------|-------|
42
- | 1 | **Raw Brain** | v4 Pro (no framework) | baseline | 1.00x | $0.00057 | - |
43
- | 2 | **VibeQMaX** (quality) | v4 Pro + full guardrails | ~baseline | 0.50x | $0.00029 | 50% |
44
- | 3 | **VibeUltraX** | MagicCoder:7b -> v4 Flash -> v4 Pro (debate) | **107%** | 0.58x | $0.00033 | 42% |
45
- | 4 | **VibeMaX** (default) | medium (auto-escalate via trained cascade) | ~75% | 0.18x | $0.00010 | 82% |
46
- | 5 | **Speed** | v4 Flash | ~55% | 0.32x | $0.000182 | 68% |
47
- | 6 | **Budget** | DeepSeek Chat | ~40% | 0.00x | $0.00 | 100% |
40
+ | Policy | Quality vs Brain | Cost vs Brain | Savings | Method |
41
+ |--------|-----------------|--------------|---------|--------|
42
+ | VibeUltraX | 107% | 0.58x | 42% | local -> Flash -> Pro cascade |
43
+ | VibeQMaX | ~100% | 0.50x | 50% | same model, framework optimizations |
44
+ | Raw Brain | 100% | 1.00x | - | baseline |
45
+ | VibeMaX | ~75% | 0.18x | 82% | trained cascade (conservative escalate) |
46
+ | Budget | ~40% | 0.00x | 100% | direct routing |
47
+
48
+ **VibeUltraX** — Cascade pipeline: MagicCoder:7b (local Ollama) proposes, v4 Flash reviews, v4 Pro refines. Benchmarked at **107% of Brain quality** at 58% cost (local inference is free, only Flash/Pro API calls cost).
48
49
 
49
50
  **VibeQMaX (Quality Max)** — Routes strategic turns through v4 Pro with full thinking, strict enforcement, strict flow checks, and quality TDD. Write/edit turns delegated to cheaper tiers per enforcement rules. Effective blended cost ~$0.00029/turn (50% of Raw Brain).
50
51
 
52
+ **Raw Brain** — v4 Pro (no framework). Baseline for all comparisons. 100% quality at 1.00x cost.
53
+
51
54
  **VibeMaX (ML-Optimized, Default)** — Intelligent cost-quality sweet spot. Routes through v4 Flash (medium) and uses a random forest classifier (29 trees, gini-split, trained on telemetry) to decide each turn. Classifies on 11 derived features: message length, code block density, urgency, complexity, repetition, question ratio, and more. Benchmarked at ~75% of Brain quality at 18% of cost.
52
55
 
53
- **VibeUltraX** — Cascade pipeline: MagicCoder:7b (local Ollama) proposes, v4 Flash reviews, v4 Pro refines. Benchmarked at **107% of Brain quality** at 58% cost (local inference is free, only Flash/Pro API calls cost).
56
+ **Budget** — DeepSeek Chat. Direct routing. ~40% quality at 0.00x cost. 100% savings.
54
57
 
55
58
  ### Cost vs Quality Visual
56
59
 
@@ -73,7 +76,7 @@ Quality
73
76
  |------|-------|----------|-------------|------|-----|
74
77
  | Raw Brain | v4 Pro | full | - | - | - |
75
78
  | VibeQMaX | v4 Pro | full | strict | strict | quality |
76
- | VibeUltraX | cascade (local->Flash->Pro) | auto | auto | auto | auto | **107%** |
79
+ | VibeUltraX | cascade (local->Flash->Pro) | auto | auto | auto | auto |
77
80
  | VibeMaX | v4 Flash (auto-escalate) | auto | auto | auto | auto |
78
81
  | Speed | v4 Flash | off | relaxed | audit | lazy |
79
82
  | Budget | DeepSeek Chat | off | relaxed | audit | lazy |
@@ -82,14 +85,6 @@ Quality
82
85
 
83
86
  All tests run with DeepSeek v4 family. Quality scores measured against Raw Brain (v4 Pro, full thinking, no vibeOS overhead). VibeMaX quality benchmark derived from real session telemetry with bootstrap confidence intervals (36 bootstrap samples). Pareto frontier computed from 70 holdout scenarios across 170 training samples via hyperparameter sweep. VibeUltra is the first mode that beats Raw Brain on both accuracy and cost — Pareto-dominant.
84
87
 
85
- | Policy | Quality vs Brain | Cost vs Brain | Savings | Method |
86
- |--------|-----------------|--------------|---------|--------|
87
- | VibeUltraX | **107%** | 0.58x | 42% | local -> Flash -> Pro cascade |
88
- | VibeMaX | ~75% | 0.18x | 82% | trained cascade (conservative escalate) |
89
- | VibeQMaX | ~100% | 0.50x | 50% | same model, framework optimizations |
90
- | Raw Brain | 100% | 1.00x | - | baseline |
91
- | Budget | ~40% | 0.00x | 100% | direct routing |
92
-
93
88
  Benchmarked on 1000 simulated questions across 20 runs, using model accuracies from MMLU-Pro / GPQA Diamond with real error correlation data.
94
89
 
95
90
  ## Features
@@ -156,7 +151,7 @@ npx vibeostheog setup --project # per-project
156
151
  npx vibeostheog setup # global ~/.config/opencode/
157
152
  ```
158
153
 
159
- Adds `vibeostheog` to opencode.json. Restart OpenCode Desktop.
154
+ One-command setup: deploys plugin files and registers in opencode.json. Restart OpenCode Desktop.
160
155
 
161
156
  Local dev checkout:
162
157
 
package/bin/setup.js CHANGED
@@ -1,48 +1,45 @@
1
1
  #!/usr/bin/env node
2
2
 
3
- import { mkdir, readFile, writeFile } from "node:fs/promises";
3
+ import { execSync } from "node:child_process";
4
+ import { readFileSync, writeFileSync, existsSync, mkdirSync } from "node:fs";
4
5
  import { dirname, resolve } from "node:path";
6
+ import { fileURLToPath } from "node:url";
5
7
  import { homedir } from "node:os";
6
8
 
7
- const pkgName = "vibeostheog";
9
+ const __dirname = dirname(fileURLToPath(import.meta.url));
10
+ const root = resolve(__dirname, "..");
8
11
  const args = process.argv.slice(2);
9
- const command = args[0] ?? "setup";
10
12
  const isProject = args.includes("--project");
11
13
 
12
- if (command !== "setup") {
13
- console.error(`Usage: ${pkgName} setup [--project]`);
14
+ // Deploy plugin files to ~/.config/opencode/plugins/ and register globally
15
+ const deployScript = resolve(root, "scripts", "deploy.mjs");
16
+ if (!existsSync(deployScript)) {
17
+ console.error("Fatal: scripts/deploy.mjs not found at", deployScript);
14
18
  process.exit(1);
15
19
  }
16
-
17
- const configPath = isProject
18
- ? resolve(process.cwd(), "opencode.json")
19
- : resolve(homedir(), ".config", "opencode", "opencode.json");
20
-
21
- let config = {};
22
- try {
23
- config = JSON.parse(await readFile(configPath, "utf8"));
24
- } catch {
25
- config = {};
26
- }
27
-
28
- if (!config || typeof config !== "object" || Array.isArray(config)) {
29
- config = {};
30
- }
31
-
32
- if (!config.$schema) {
33
- config.$schema = "https://opencode.ai/config.json";
20
+ execSync(`node "${deployScript}"`, { stdio: "inherit", cwd: root });
21
+
22
+ // For per-project setup, also register in project-level opencode.json
23
+ if (isProject) {
24
+ const configPath = resolve(process.cwd(), "opencode.json");
25
+ let config = {};
26
+ if (existsSync(configPath)) {
27
+ try {
28
+ config = JSON.parse(readFileSync(configPath, "utf8"));
29
+ } catch {
30
+ config = {};
31
+ }
32
+ }
33
+ if (!config || typeof config !== "object" || Array.isArray(config)) config = {};
34
+ if (!config.$schema) config.$schema = "https://opencode.ai/config.json";
35
+ if (!Array.isArray(config.plugin)) config.plugin = [];
36
+ const pluginRef = resolve(homedir(), ".config", "opencode", "plugins", "vibeOS.js");
37
+ if (!config.plugin.includes(pluginRef)) {
38
+ config.plugin.push(pluginRef);
39
+ mkdirSync(dirname(configPath), { recursive: true });
40
+ writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
41
+ }
42
+ console.log(`vibeOS registered in ${configPath}`);
34
43
  }
35
44
 
36
- if (!Array.isArray(config.plugin)) {
37
- config.plugin = [];
38
- }
39
-
40
- if (!config.plugin.includes(pkgName)) {
41
- config.plugin.push(pkgName);
42
- }
43
-
44
- await mkdir(dirname(configPath), { recursive: true });
45
- await writeFile(configPath, `${JSON.stringify(config, null, 2)}\n`);
46
-
47
- console.log(`${pkgName} registered in ${configPath}`);
48
- console.log("Restart OpenCode to activate the plugin.");
45
+ console.log("Done. Restart OpenCode to activate the plugin.");
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "vibeostheog",
3
- "version": "0.22.0",
3
+ "version": "0.22.6",
4
4
  "description": "Cost-aware delegation enforcer for OpenCode. Tracks model usage, routes Task subagents to cheaper tiers, surfaces cumulative savings in chat. Includes research audit, reporting framework, project memory, progressive scratchpad decadence, and trinity CLI for brain/medium/cheap slot switching.",
5
5
  "scripts": {
6
6
  "release": "node scripts/release.mjs",
@@ -14,8 +14,8 @@
14
14
  "checkpoint:validate": "node scripts/checkpoint-validate.mjs",
15
15
  "test:scripts": "node --test scripts/tests/checkpoint-validate.test.mjs tests/release-pack.test.mjs",
16
16
  "ts:audit": "node scripts/ts-audit.mjs",
17
- "test": "VIBEOS_MCP_PORT=0 node --test --test-timeout=240000 tests/deep_integration.test.mjs tests/production_regressions.test.mjs tests/release_hardening_tigerteam.test.mjs tests/test_api_migration.neutral.test.mjs tests/test_const_assignment_regression.test.mjs tests/test_delegation_enforcer.test.mjs tests/test_diagnose_cmd.test.mjs tests/test_install_and_recovery.test.mjs tests/test_internals_stress_patterns_offtopic.test.mjs tests/test_saveos_e2e_cleanup.test.mjs tests/test_tdd_enforcer.test.mjs src/tests/*.test.js src/utils/tests/*.test.mjs \"src/vibeOS-lib/tests/auto-select-mode.test.mjs\" \"src/vibeOS-lib/tests/blackbox-regression.test.mjs\" \"src/vibeOS-lib/tests/blackbox-smoke.test.mjs\" \"src/vibeOS-lib/tests/budget-first-mode.test.mjs\" \"src/vibeOS-lib/tests/flow-enforcer.test.mjs\" \"src/vibeOS-lib/tests/flow-secrets.test.mjs\" \"src/vibeOS-lib/tests/session-metrics.test.mjs\" \"src/vibeOS-lib/tests/test_stress.test.mjs\"",
18
- "test:ci": "VIBEOS_MCP_PORT=0 node --test --test-timeout=30000 tests/production_regressions.test.mjs tests/release_hardening_tigerteam.test.mjs tests/test_const_assignment_regression.test.mjs tests/test_diagnose_cmd.test.mjs tests/test_install_and_recovery.test.mjs tests/test_saveos_e2e_cleanup.test.mjs tests/test_tdd_enforcer.test.mjs src/tests/*.test.js src/utils/tests/*.test.mjs \"src/vibeOS-lib/tests/auto-select-mode.test.mjs\" \"src/vibeOS-lib/tests/blackbox-regression.test.mjs\" \"src/vibeOS-lib/tests/blackbox-smoke.test.mjs\" \"src/vibeOS-lib/tests/budget-first-mode.test.mjs\" \"src/vibeOS-lib/tests/flow-enforcer.test.mjs\" \"src/vibeOS-lib/tests/flow-secrets.test.mjs\" \"src/vibeOS-lib/tests/session-metrics.test.mjs\" \"src/vibeOS-lib/tests/test_stress.test.mjs\"",
17
+ "test": "VIBEOS_MCP_PORT=0 node --test --test-timeout=240000 tests/deep_integration.test.mjs tests/production_regressions.test.mjs tests/release_hardening_tigerteam.test.mjs tests/test_api_migration.neutral.test.mjs tests/test_const_assignment_regression.test.mjs tests/test_delegation_enforcer.test.mjs tests/test_diagnose_cmd.test.mjs tests/test_install_and_recovery.test.mjs tests/test_internals_stress_patterns_offtopic.test.mjs tests/test_saveos_e2e_cleanup.test.mjs tests/test_tdd_enforcer.test.mjs tests/test_10fixes_regression.test.mjs tests/test_cross_session_regression.test.mjs tests/test_mega_all_fixes.test.mjs tests/test_smart_cache_regression.test.mjs src/tests/*.test.js src/utils/tests/*.test.mjs \"src/vibeOS-lib/tests/auto-select-mode.test.mjs\" \"src/vibeOS-lib/tests/blackbox-regression.test.mjs\" \"src/vibeOS-lib/tests/blackbox-smoke.test.mjs\" \"src/vibeOS-lib/tests/budget-first-mode.test.mjs\" \"src/vibeOS-lib/tests/flow-enforcer.test.mjs\" \"src/vibeOS-lib/tests/flow-secrets.test.mjs\" \"src/vibeOS-lib/tests/session-metrics.test.mjs\" \"src/vibeOS-lib/tests/test_stress.test.mjs\"",
18
+ "test:ci": "VIBEOS_MCP_PORT=0 node --test --test-timeout=30000 tests/production_regressions.test.mjs tests/release_hardening_tigerteam.test.mjs tests/test_const_assignment_regression.test.mjs tests/test_diagnose_cmd.test.mjs tests/test_install_and_recovery.test.mjs tests/test_saveos_e2e_cleanup.test.mjs tests/test_tdd_enforcer.test.mjs tests/test_10fixes_regression.test.mjs tests/test_cross_session_regression.test.mjs tests/test_mega_all_fixes.test.mjs tests/test_smart_cache_regression.test.mjs src/tests/*.test.js src/utils/tests/*.test.mjs \"src/vibeOS-lib/tests/auto-select-mode.test.mjs\" \"src/vibeOS-lib/tests/blackbox-regression.test.mjs\" \"src/vibeOS-lib/tests/blackbox-smoke.test.mjs\" \"src/vibeOS-lib/tests/budget-first-mode.test.mjs\" \"src/vibeOS-lib/tests/flow-enforcer.test.mjs\" \"src/vibeOS-lib/tests/flow-secrets.test.mjs\" \"src/vibeOS-lib/tests/session-metrics.test.mjs\" \"src/vibeOS-lib/tests/test_stress.test.mjs\"",
19
19
  "codex:guard": "bash plugins/vibetheog-codex/scripts/run-guard.sh",
20
20
  "codex:guard:full": "VIBETHEOG_GUARD_FULL=1 bash plugins/vibetheog-codex/scripts/run-guard.sh",
21
21
  "codex:hook:precommit": "bash plugins/vibetheog-codex/hooks/pre-commit.sh",
@@ -38,6 +38,66 @@ export class VibeOSNetworkError extends Error {
38
38
  this.name = "VibeOSNetworkError";
39
39
  }
40
40
  }
41
+ const ANOMALY_BURST_WINDOW_MS = 5000;
42
+ const ANOMALY_BURST_THRESHOLD = 10;
43
+ const ANOMALY_FREQ_WINDOW_MS = 600_000;
44
+ const ANOMALY_STDDEV_FACTOR = 3;
45
+ const ANOMALY_WARMUP_MS = 30_000;
46
+ const ANOMALY_COOLDOWN_MS = 120_000;
47
+ class TokenAnomalyDetector {
48
+ burstHistory = [];
49
+ freqHistory = [];
50
+ lastWarnTime = 0;
51
+ anomalyTriggered = false;
52
+ disabled = false;
53
+ startedAt = Date.now();
54
+ get isWarmup() {
55
+ return Date.now() - this.startedAt < ANOMALY_WARMUP_MS;
56
+ }
57
+ record() {
58
+ if (this.disabled || this.isWarmup)
59
+ return;
60
+ const now = Date.now();
61
+ this.burstHistory = this.burstHistory.filter(t => now - t < ANOMALY_BURST_WINDOW_MS);
62
+ this.burstHistory.push(now);
63
+ this.freqHistory.push(now);
64
+ }
65
+ checkBurst() {
66
+ return this.burstHistory.length > ANOMALY_BURST_THRESHOLD;
67
+ }
68
+ checkFrequency() {
69
+ const now = Date.now();
70
+ const window = this.freqHistory.filter(t => now - t < ANOMALY_FREQ_WINDOW_MS);
71
+ if (window.length < 10)
72
+ return false;
73
+ const mean = window.length / (ANOMALY_FREQ_WINDOW_MS / 60_000);
74
+ const recent = this.burstHistory.length / (ANOMALY_BURST_WINDOW_MS / 1000);
75
+ return recent > mean * ANOMALY_STDDEV_FACTOR;
76
+ }
77
+ throttleIfAnomalous() {
78
+ const now = Date.now();
79
+ if (this.disabled || this.isWarmup)
80
+ return false;
81
+ if (this.anomalyTriggered)
82
+ return true;
83
+ if (this.checkBurst() || this.checkFrequency()) {
84
+ this.anomalyTriggered = true;
85
+ this.lastWarnTime = now;
86
+ console.error("[vibeOS] Token anomaly detected — throttling API calls");
87
+ return true;
88
+ }
89
+ if (this.lastWarnTime && now - this.lastWarnTime > ANOMALY_COOLDOWN_MS) {
90
+ this.anomalyTriggered = false;
91
+ }
92
+ return this.anomalyTriggered;
93
+ }
94
+ reset() {
95
+ this.burstHistory = [];
96
+ this.freqHistory = [];
97
+ this.anomalyTriggered = false;
98
+ this.lastWarnTime = 0;
99
+ }
100
+ }
41
101
  function normalizeApiToken(token, fallback = "") {
42
102
  const clean = String(token || "").trim();
43
103
  return API_TOKEN_RE.test(clean) ? clean : fallback;
@@ -433,6 +493,19 @@ export let VIBEOS_API_DISABLED = readApiDisabledFromDisk() || isTruthyFlag(proce
433
493
  export let VIBEOS_API_TOKEN = VIBEOS_API_DISABLED ? "" : (readTokenFromDisk() || normalizeApiToken(process.env.VIBEOS_API_TOKEN, "") || EMBEDDED_API_TOKEN);
434
494
  export let VIBEOS_API_BOOTSTRAP_TOKEN = VIBEOS_API_DISABLED ? "" : (readBootstrapTokenFromDisk() || process.env.VIBEOS_API_BOOTSTRAP_TOKEN || "");
435
495
  export let VIBEOS_API_ENABLED = !VIBEOS_API_DISABLED && process.env.VIBEOS_API_ENABLED !== "false" && (!!VIBEOS_API_TOKEN || !!VIBEOS_API_BOOTSTRAP_TOKEN);
496
+ let _anomalyDetector = null;
497
+ function getAnomalyDetector() {
498
+ if (!_anomalyDetector)
499
+ _anomalyDetector = new TokenAnomalyDetector();
500
+ return _anomalyDetector;
501
+ }
502
+ export function setAnomalyDetection(enabled) {
503
+ const d = getAnomalyDetector();
504
+ d.disabled = !enabled;
505
+ if (enabled)
506
+ d.reset();
507
+ console.error(`[vibeOS] Anomaly detection ${enabled ? "enabled" : "disabled"}`);
508
+ }
436
509
  function persistBootstrapToken(token) {
437
510
  const clean = String(token || "").trim();
438
511
  try {
@@ -460,6 +533,8 @@ export function setApiToken(newToken) {
460
533
  VIBEOS_API_BOOTSTRAP_TOKEN = readBootstrapTokenFromDisk() || VIBEOS_API_BOOTSTRAP_TOKEN;
461
534
  VIBEOS_API_ENABLED = process.env.VIBEOS_API_ENABLED !== "false" && (!!VIBEOS_API_TOKEN || !!VIBEOS_API_BOOTSTRAP_TOKEN);
462
535
  persistPrimaryApiEnvState({ token: VIBEOS_API_TOKEN, disabled: false });
536
+ if (_anomalyDetector)
537
+ _anomalyDetector.reset();
463
538
  console.error("[vibeOS] API token updated via setApiToken");
464
539
  }
465
540
  catch (e) {
@@ -475,6 +550,8 @@ export function invalidateApiToken() {
475
550
  _apiClient = null;
476
551
  _apiFallbackMode = false;
477
552
  _apiFallbackSince = null;
553
+ if (_anomalyDetector)
554
+ _anomalyDetector.reset();
478
555
  persistBootstrapToken("");
479
556
  persistPrimaryApiEnvState({ token: "", disabled: true });
480
557
  resetApiConnection();
@@ -624,6 +701,15 @@ export async function remoteCall(method, args, fallbackFn) {
624
701
  return fallbackFn();
625
702
  return null;
626
703
  }
704
+ const detector = getAnomalyDetector();
705
+ detector.record();
706
+ if (detector.throttleIfAnomalous()) {
707
+ // Don't set _apiFallbackMode — detector's own cooldown resets it.
708
+ // This lets the API retry naturally after the throttle window.
709
+ if (fallbackFn)
710
+ return fallbackFn();
711
+ return null;
712
+ }
627
713
  try {
628
714
  const client = getApiClient();
629
715
  if (!client) {
@@ -0,0 +1,75 @@
1
+ // @ts-nocheck
2
+ /**
3
+ * SPDX-License-Identifier: MIT
4
+ * SPDX-FileCopyrightText: 2026 vibeOS <https://github.com/DrunkkToys/vibeOS>
5
+ *
6
+ * Cost anomaly detector — monitors per-turn model costs and warns
7
+ * when a model cost spikes significantly above the running average.
8
+ */
9
+
10
+ const COST_WINDOW_SIZE = 20
11
+ const COST_ANOMALY_THRESHOLD = 3
12
+ const COST_WARMUP_SAMPLES = 5
13
+
14
+ export class CostAnomalyDetector {
15
+ constructor() {
16
+ this.costHistory = []
17
+ this.disabled = false
18
+ this.currentAnomalyModel = null
19
+ this.currentAnomalyCost = 0
20
+ this.currentAnomalyMean = 0
21
+ }
22
+
23
+ record(cost) {
24
+ if (this.disabled) return
25
+ this.costHistory.push(cost)
26
+ if (this.costHistory.length > COST_WINDOW_SIZE) {
27
+ this.costHistory.shift()
28
+ }
29
+ }
30
+
31
+ get mean() {
32
+ if (this.costHistory.length === 0) return 0
33
+ return this.costHistory.reduce((a, b) => a + b, 0) / this.costHistory.length
34
+ }
35
+
36
+ checkAnomaly(model, cost) {
37
+ if (this.disabled) return false
38
+ if (this.costHistory.length < COST_WARMUP_SAMPLES) return false
39
+ const avg = this.mean
40
+ if (avg <= 0 || cost <= avg) return false
41
+ const ratio = cost / avg
42
+ if (ratio > COST_ANOMALY_THRESHOLD) {
43
+ this.currentAnomalyModel = model
44
+ this.currentAnomalyCost = cost
45
+ this.currentAnomalyMean = avg
46
+ return true
47
+ }
48
+ return false
49
+ }
50
+
51
+ clearAnomaly() {
52
+ this.currentAnomalyModel = null
53
+ this.currentAnomalyCost = 0
54
+ this.currentAnomalyMean = 0
55
+ }
56
+
57
+ reset() {
58
+ this.costHistory = []
59
+ this.clearAnomaly()
60
+ }
61
+ }
62
+
63
+ let _costDetector = null
64
+
65
+ export function getCostAnomalyDetector() {
66
+ if (!_costDetector) _costDetector = new CostAnomalyDetector()
67
+ return _costDetector
68
+ }
69
+
70
+ export function setCostAnomalyDetection(enabled) {
71
+ const d = getCostAnomalyDetector()
72
+ d.disabled = !enabled
73
+ if (enabled) d.reset()
74
+ console.error(`[vibeOS] Cost anomaly detection ${enabled ? "enabled" : "disabled"}`)
75
+ }
@@ -1,8 +1,8 @@
1
1
  // @ts-nocheck
2
- import { readFileSync, writeFileSync, appendFileSync, existsSync, mkdirSync } from "node:fs";
2
+ import { readFileSync, writeFileSync, appendFileSync, existsSync, mkdirSync, rmSync } from "node:fs";
3
3
  import { join, basename } from "node:path";
4
4
  import { createHash } from "node:crypto";
5
- import { currentModel, currentProjectFingerprint, currentProjectName, _blackboxEnabled, loadSelection, writeSelection, safeJsonParse, applyDecadence, getSessionScratchpadDir, ensureSessionScratchpadDirs, indexAppend, briefedProjects, getActiveJobForProject, loadTodos, promotedProjectPatterns, detectTechStack, projectFingerprint, TRINITY_OPENCODE_CONFIG, TIERS_FILE, loadGlobalLearning, setCurrentProjectFingerprint, setCurrentProjectName, stableJson, TOOL_NAME_NORMALIZE, _cacheDb, recordCacheSaving, } from "../state.js";
5
+ import { currentModel, currentProjectFingerprint, currentProjectName, _blackboxEnabled, loadSelection, writeSelection, safeJsonParse, applyDecadence, getSessionScratchpadDir, ensureSessionScratchpadDirs, indexAppend, briefedProjects, getActiveJobForProject, loadTodos, promotedProjectPatterns, detectTechStack, projectFingerprint, SCRATCHPAD_ROOT, TRINITY_OPENCODE_CONFIG, TIERS_FILE, loadGlobalLearning, setCurrentProjectFingerprint, setCurrentProjectName, stableJson, TOOL_NAME_NORMALIZE, _cacheDb, recordCacheSaving, } from "../state.js";
6
6
  import { applySlot, TRINITY_CHEAP, TRINITY_MEDIUM, cacheSavePer1MInputTokens, } from "../pricing.js";
7
7
  import { scoreStress, classifyTurnSimple, loadOptimizationMode, saveOptimizationMode, selectOptimizationModeRemote, computeControlVector, getBlackboxTracker, loadBlackboxState as loadBlackboxStateFromCtx, saveBlackboxState as saveBlackboxStateToCtx, extractLastUserText, isLikelyOffTopic, fetchBlackboxEnrichment, estimateContextBudget, buildControlHistoryEntry, } from "../turn-classify.js";
8
8
  import { applyBudgetFirstMode, peekBudgetFirstMode } from "../mode-policy.js";
@@ -277,25 +277,31 @@ function compressToolOutputs(messages) {
277
277
  continue;
278
278
  const hash = createHash("sha256")
279
279
  .update(`tool_result\n${raw}\n`).digest("hex").slice(0, 16);
280
- const fullPath = join(getSessionScratchpadDir(), `${hash}.txt`);
280
+ const globalDir = join(SCRATCHPAD_ROOT, "by-hash");
281
+ const sessPath = join(getSessionScratchpadDir(), `${hash}.txt`);
282
+ const globalPath = join(globalDir, `${hash}.txt`);
281
283
  try {
284
+ mkdirSync(globalDir, { recursive: true });
282
285
  ensureSessionScratchpadDirs();
283
- if (!existsSync(fullPath)) {
284
- writeFileSync(fullPath, raw);
286
+ if (!existsSync(globalPath)) {
287
+ writeFileSync(globalPath, raw);
285
288
  indexAppend(hash, part.tool, raw.length);
286
- // Create pointer file for input-hash-based lookup
287
- const invPart = parts.slice(0, parts.indexOf(part)).reverse().find((p) => p?.type === "tool" && p?.tool === part.tool && p?.state?.input && p?.state?.status !== "completed");
288
- if (invPart?.state?.input) {
289
- const toolKey = TOOL_NAME_NORMALIZE[part.tool] || part.tool;
290
- const inputHash = createHash("sha256")
291
- .update(`${toolKey}\n${stableJson(invPart.state.input)}\n`)
292
- .digest("hex").slice(0, 16);
293
- const ptrPath = join(getSessionScratchpadDir(), `${inputHash}.ptr`);
294
- try {
295
- writeFileSync(ptrPath, JSON.stringify({ contentHash: hash, tool: part.tool }));
296
- }
297
- catch { }
289
+ // Clean up any existing session-local copy
290
+ if (existsSync(sessPath))
291
+ rmSync(sessPath, { force: true });
292
+ }
293
+ // Create pointer file for input-hash-based lookup
294
+ const invPart = parts.slice(0, parts.indexOf(part)).reverse().find((p) => p?.type === "tool" && p?.tool === part.tool && p?.state?.input && p?.state?.status !== "completed");
295
+ if (invPart?.state?.input) {
296
+ const toolKey = TOOL_NAME_NORMALIZE[part.tool] || part.tool;
297
+ const inputHash = createHash("sha256")
298
+ .update(`${toolKey}\n${stableJson(invPart.state.input)}\n`)
299
+ .digest("hex").slice(0, 16);
300
+ const ptrPath = join(getSessionScratchpadDir(), `${inputHash}.ptr`);
301
+ try {
302
+ writeFileSync(ptrPath, JSON.stringify({ contentHash: hash, tool: part.tool }));
298
303
  }
304
+ catch { }
299
305
  }
300
306
  }
301
307
  catch (err) {
@@ -305,7 +311,7 @@ function compressToolOutputs(messages) {
305
311
  if (!isCold)
306
312
  continue;
307
313
  const summary = raw.slice(0, 200).replace(/\n+/g, " ").trim() + (raw.length > 200 ? "\u2026" : "");
308
- const ref = `${COMPRESS_MARKER} [${raw.length} chars compressed -- cold storage at ${fullPath}] ` +
314
+ const ref = `${COMPRESS_MARKER} [${raw.length} chars compressed -- cold storage at ${globalPath}] ` +
309
315
  `[summary] ${summary}`;
310
316
  state.output = ref;
311
317
  compressedBytes += raw.length - ref.length;
@@ -9,6 +9,7 @@ import { scoreStress, extractFirstWordFromArgs, shouldLogWarn, isUserAskingForTe
9
9
  import { saveReport } from "../reporting.js";
10
10
  import { loadCredit } from "../credit-api.js";
11
11
  import { remoteCall, VIBEOS_API_ENABLED } from "../api-client.js";
12
+ import { getCostAnomalyDetector } from "../cost-anomaly.js";
12
13
  import { checkFlowRules } from "../../vibeOS-lib/flow-enforcer.js";
13
14
  import { computeDifficulty, addRouteEdge, predictBestModel, hashQuery } from "../../vibeOS-lib/ml-router.js";
14
15
  import { addCacheEntry, recordCacheStats, predictCacheHit } from "../../vibeOS-lib/smart-cache.js";
@@ -487,6 +488,25 @@ export const onToolExecuteBefore = async (input, output) => {
487
488
  return;
488
489
  }
489
490
  }
491
+ // Cost anomaly detection: warn if this model's per-turn cost spikes
492
+ // significantly above the session rolling average.
493
+ const costDetector = getCostAnomalyDetector();
494
+ if (!costDetector.disabled && currentModel) {
495
+ const modelCost = modelCostPerTurn(currentModel);
496
+ const fullModelName = currentModel;
497
+ if (costDetector.checkAnomaly(fullModelName, modelCost)) {
498
+ const avg = costDetector.currentAnomalyMean;
499
+ const ratio = avg > 0 ? (modelCost / avg).toFixed(1) : "?";
500
+ const msg = `Cost spike: ${shortModelName(fullModelName)} at $${modelCost.toFixed(4)}/turn — ${ratio}x higher than recent avg of $${avg.toFixed(4)}. Run \`trinity cheap\` or \`trinity medium\` to save.`;
501
+ if (shouldLogWarn(`${t}|cost-anomaly|${fullModelName}|${modelCost.toFixed(4)}`)) {
502
+ console.error(`[vibeOS] [cost-anomaly] ${msg}`);
503
+ }
504
+ pendingUiNote = `🚨 ${msg}`;
505
+ enforcementBlocked = true;
506
+ return;
507
+ }
508
+ costDetector.record(modelCost);
509
+ }
490
510
  // Credit < 40%: non-task tool — record and nudge to step aside.
491
511
  if (_credit < 40 && !compatibilityMode) {
492
512
  const total = recordSaving(t, "credit<40% high-tier", _estOpus, { firstWord: _firstWord });
@@ -0,0 +1,100 @@
1
+ // Mode Router — 10 modes, 4 tiers. Full type-safe hierarchy.
2
+ // Branded modes: user-selected strategy + tier pipeline.
3
+ // Runtime modes: classifier-selected behavior per query.
4
+ export const TIERS = {
5
+ brain: { cost: 0.002, desc: "v4 Pro tier — max quality" },
6
+ medium: { cost: 0.000182, desc: "v4 Flash tier — balanced" },
7
+ cheap: { cost: 0, desc: "Chat tier — free" },
8
+ local: { cost: 0, desc: "Ollama local model" },
9
+ };
10
+ export const BRANDED_MODES = [
11
+ {
12
+ id: "vibeultrax", index: 1, name: "VibeUltraX", icon: "\u{1F3C6}",
13
+ pipeline: ["local", "medium", "brain"],
14
+ thinking: "full", tdd: "quality", enforcement: "strict", flow: "strict",
15
+ qualityVsBrain: 107, costVsBrain: 58,
16
+ desc: "3-model debate: local proposes, medium reviews, brain refines.",
17
+ },
18
+ {
19
+ id: "vibeqmax", index: 2, name: "VibeQMaX", icon: "\u{2B50}",
20
+ pipeline: ["brain"],
21
+ thinking: "full", tdd: "quality", enforcement: "strict", flow: "strict",
22
+ qualityVsBrain: 100, costVsBrain: 50,
23
+ desc: "Brain tier only. Same quality as Raw Brain at half cost.",
24
+ },
25
+ {
26
+ id: "vibemax", index: 3, name: "VibeMaX", icon: "\u{26A1}",
27
+ pipeline: ["medium"],
28
+ thinking: "off", tdd: "lazy", enforcement: "relaxed", flow: "audit",
29
+ qualityVsBrain: 75, costVsBrain: 18, default: true,
30
+ desc: "Default mode. Medium tier auto-escalate. Speed-first.",
31
+ },
32
+ ];
33
+ export const RUNTIME_MODES = [
34
+ {
35
+ id: "balanced", index: 4, name: "Balanced", icon: "\u{2696}\u{FE0F}",
36
+ pipeline: ["medium"],
37
+ thinking: "brief", tdd: "lazy", enforcement: "relaxed", flow: "audit",
38
+ qualityVsBrain: 70, costVsBrain: 30, defaultRuntime: true,
39
+ desc: "Default runtime. Auto-selects behavior per query.",
40
+ },
41
+ {
42
+ id: "speed", index: 5, name: "Speed", icon: "\u{1F680}",
43
+ pipeline: ["medium"],
44
+ thinking: "off", tdd: "off", enforcement: "relaxed", flow: "off",
45
+ qualityVsBrain: 55, costVsBrain: 32,
46
+ desc: "Medium tier. Fast responses, no overhead.",
47
+ },
48
+ {
49
+ id: "budget", index: 6, name: "Budget", icon: "\u{1F4B8}",
50
+ pipeline: ["cheap"],
51
+ thinking: "off", tdd: "off", enforcement: "off", flow: "off",
52
+ qualityVsBrain: 40, costVsBrain: 100,
53
+ desc: "Cheap tier only. Zero cost.",
54
+ },
55
+ {
56
+ id: "quality", index: 7, name: "Quality", icon: "\u{1F4AF}",
57
+ pipeline: ["brain"],
58
+ thinking: "full", tdd: "quality", enforcement: "strict", flow: "strict",
59
+ qualityVsBrain: 100, costVsBrain: 60,
60
+ desc: "Brain tier with full thinking and enforcement.",
61
+ },
62
+ {
63
+ id: "audit", index: 8, name: "Audit", icon: "\u{1F50D}",
64
+ pipeline: ["brain"],
65
+ thinking: "full", tdd: "quality", enforcement: "strict", flow: "strict",
66
+ qualityVsBrain: 100, costVsBrain: 55,
67
+ desc: "Brain tier security audit. OWASP validation.",
68
+ },
69
+ {
70
+ id: "longrun", index: 9, name: "Longrun", icon: "\u{1F3C3}",
71
+ pipeline: ["brain"],
72
+ thinking: "full", tdd: "quality", enforcement: "strict", flow: "strict",
73
+ qualityVsBrain: 100, costVsBrain: 70,
74
+ desc: "Brain tier extended session. Full context.",
75
+ },
76
+ ];
77
+ export const RAW_MODE = {
78
+ id: "raw", index: 10, name: "Raw Brain", icon: "\u{1F9E0}",
79
+ pipeline: ["brain"],
80
+ thinking: "full", tdd: "\u2014", enforcement: "\u2014", flow: "\u2014",
81
+ qualityVsBrain: 100, costVsBrain: 0,
82
+ desc: "Pure v4 Pro baseline. No vibeOS overhead.",
83
+ };
84
+ export const ALL_MODES = [...BRANDED_MODES, ...RUNTIME_MODES, RAW_MODE];
85
+ export function getMode(id) {
86
+ return ALL_MODES.find(m => m.id === id) ?? getDefault();
87
+ }
88
+ export function getDefault() {
89
+ return BRANDED_MODES.find(m => m.default);
90
+ }
91
+ export function getDefaultRuntime() {
92
+ return RUNTIME_MODES.find(m => m.defaultRuntime);
93
+ }
94
+ export function getBrandedModes() { return BRANDED_MODES; }
95
+ export function getRuntimeModes() { return RUNTIME_MODES; }
96
+ export function resolveTierModels(mode, tierMap) {
97
+ const models = mode.pipeline.map(t => tierMap[t] ?? t);
98
+ const costs = mode.pipeline.map(t => TIERS[t]?.cost ?? 0);
99
+ return { models, totalCost: costs.reduce((s, c) => s + c, 0) };
100
+ }
package/src/lib/state.js CHANGED
@@ -907,7 +907,7 @@ function getScratchpadHit(toolLower, args, baseDir = null) {
907
907
  const globalDir = SCRATCHPAD_GLOBAL_DIR;
908
908
  const sessionPath = join(sessionDir, `${hash}.txt`);
909
909
  const globalPath = join(globalDir, `${hash}.txt`);
910
- let fullPath = existsSync(sessionPath) ? sessionPath : (existsSync(globalPath) ? globalPath : null);
910
+ let fullPath = existsSync(globalPath) ? globalPath : (existsSync(sessionPath) ? sessionPath : null);
911
911
  if (!fullPath) {
912
912
  // Try pointer files (created by compressToolOutputs mapping input hash -> content hash)
913
913
  const ptrSessionPath = join(sessionDir, `${hash}.ptr`);
@@ -921,7 +921,7 @@ function getScratchpadHit(toolLower, args, baseDir = null) {
921
921
  resolvedHash = ptrData.contentHash;
922
922
  const rSessionPath = join(sessionDir, `${resolvedHash}.txt`);
923
923
  const rGlobalPath = join(globalDir, `${resolvedHash}.txt`);
924
- fullPath = existsSync(rSessionPath) ? rSessionPath : (existsSync(rGlobalPath) ? rGlobalPath : null);
924
+ fullPath = existsSync(rGlobalPath) ? rGlobalPath : (existsSync(rSessionPath) ? rSessionPath : null);
925
925
  }
926
926
  }
927
927
  catch { }
@@ -1454,9 +1454,9 @@ function recordDelegation(tool, saveEst, meta = {}) {
1454
1454
  s.sessions ??= {};
1455
1455
  const sid = _OC_SID;
1456
1456
  s.sessions[sid] ??= { started: now, session_started_at: now, source: "opencode", tool_counts: {}, warns: [] };
1457
- if (currentProjectFingerprint)
1457
+ if (currentProjectFingerprint && !s.sessions[sid].project_fingerprint)
1458
1458
  s.sessions[sid].project_fingerprint = currentProjectFingerprint;
1459
- if (currentProjectName)
1459
+ if (currentProjectName && !s.sessions[sid].project_name)
1460
1460
  s.sessions[sid].project_name = currentProjectName;
1461
1461
  s.sessions[sid].total_savings_usd = roundUsd(Number(s.sessions[sid].total_savings_usd || 0) + delta);
1462
1462
  _pruneOldSessions(s);
@@ -1478,9 +1478,9 @@ function recordCacheSaving(tool, saveEst, meta = {}) {
1478
1478
  s.sessions ??= {};
1479
1479
  const sid = _OC_SID;
1480
1480
  s.sessions[sid] ??= { started: now, session_started_at: now, source: "opencode", tool_counts: {}, warns: [] };
1481
- if (currentProjectFingerprint)
1481
+ if (currentProjectFingerprint && !s.sessions[sid].project_fingerprint)
1482
1482
  s.sessions[sid].project_fingerprint = currentProjectFingerprint;
1483
- if (currentProjectName)
1483
+ if (currentProjectName && !s.sessions[sid].project_name)
1484
1484
  s.sessions[sid].project_name = currentProjectName;
1485
1485
  s.sessions[sid].session_cache_dir = getSessionScratchpadDir();
1486
1486
  s.sessions[sid].tool_counts[tool] = (s.sessions[sid].tool_counts[tool] || 0) + 1;
@@ -45,18 +45,20 @@ export function bootstrapOptimizationSession() {
45
45
  const state = loadBlackboxState();
46
46
  if (!state.sessions)
47
47
  state.sessions = {};
48
- if (!state.sessions[sid])
49
- state.sessions[sid] = {};
50
- state.sessions[sid].optimization_mode = resolvedMode;
51
- state.sessions[sid].active_slot = resolvedSlot;
52
- state.sessions[sid].sub_regime = state.sessions[sid].sub_regime || "INIT";
53
- state.sessions[sid].regime = state.sessions[sid].regime || "INIT";
54
- state.sessions[sid].resolution = state.sessions[sid].resolution || "unresolved";
55
- state.sessions[sid].momentum = Number(state.sessions[sid].momentum || 0);
56
- state.sessions[sid].loop_count = Number(state.sessions[sid].loop_count || 0);
57
- state.sessions[sid].loop_intervention_level = state.sessions[sid].loop_intervention_level || "none";
58
- state.sessions[sid].loop_start_turn = Number(state.sessions[sid].loop_start_turn || 0);
59
- state.sessions[sid].loop_pattern_count = Number(state.sessions[sid].loop_pattern_count || 0);
48
+ if (sid && sid !== "undefined") {
49
+ if (!state.sessions[sid])
50
+ state.sessions[sid] = {};
51
+ state.sessions[sid].optimization_mode = resolvedMode;
52
+ state.sessions[sid].active_slot = resolvedSlot;
53
+ state.sessions[sid].sub_regime = state.sessions[sid].sub_regime || "INIT";
54
+ state.sessions[sid].regime = state.sessions[sid].regime || "INIT";
55
+ state.sessions[sid].resolution = state.sessions[sid].resolution || "unresolved";
56
+ state.sessions[sid].momentum = Number(state.sessions[sid].momentum || 0);
57
+ state.sessions[sid].loop_count = Number(state.sessions[sid].loop_count || 0);
58
+ state.sessions[sid].loop_intervention_level = state.sessions[sid].loop_intervention_level || "none";
59
+ state.sessions[sid].loop_start_turn = Number(state.sessions[sid].loop_start_turn || 0);
60
+ state.sessions[sid].loop_pattern_count = Number(state.sessions[sid].loop_pattern_count || 0);
61
+ }
60
62
  saveBlackboxState(state);
61
63
  }
62
64
  catch { }
@@ -314,11 +316,11 @@ export function getBlackboxTracker() {
314
316
  if (state.enabled !== undefined)
315
317
  _setGlobalBlackboxEnabled(state.enabled);
316
318
  const sid = _OC_SID;
317
- if (state.sessions?.[sid]?.history) {
319
+ if (sid && sid !== "undefined" && state.sessions?.[sid]?.history) {
318
320
  _blackboxTracker = _BlackboxStub.deserialize(state.sessions[sid]);
319
321
  }
320
- else if (currentProjectFingerprint) {
321
- const projectKeys = Object.keys(state.sessions || {}).filter(k => state.sessions[k].project_fingerprint === currentProjectFingerprint);
322
+ else if (currentProjectFingerprint && sid && sid !== "undefined") {
323
+ const projectKeys = Object.keys(state.sessions || {}).filter(k => state.sessions[k].project_fingerprint === currentProjectFingerprint && k !== "undefined" && k !== null && k.trim() !== "");
322
324
  const latest = projectKeys.sort().slice(-1)[0];
323
325
  if (latest && state.sessions[latest]?.history) {
324
326
  const data = state.sessions[latest];
@@ -589,12 +591,14 @@ export function incrementTurnCounter() {
589
591
  const sid = _OC_SID;
590
592
  if (!state.sessions)
591
593
  state.sessions = {};
592
- if (!state.sessions[sid])
593
- state.sessions[sid] = {};
594
- const next = (state.sessions[sid].turn_counter || 0) + 1;
595
- state.sessions[sid].turn_counter = next;
594
+ if (sid && sid !== "undefined") {
595
+ if (!state.sessions[sid])
596
+ state.sessions[sid] = {};
597
+ const next = (state.sessions[sid].turn_counter || 0) + 1;
598
+ state.sessions[sid].turn_counter = next;
599
+ }
596
600
  saveBlackboxState(state);
597
- return next;
601
+ return 0;
598
602
  }
599
603
  catch {
600
604
  return 0;