thumbgate 1.27.18 → 1.27.19

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 (96) hide show
  1. package/.claude-plugin/marketplace.json +6 -6
  2. package/.claude-plugin/plugin.json +4 -3
  3. package/.well-known/agentic-verify.txt +1 -0
  4. package/.well-known/llms.txt +33 -12
  5. package/.well-known/mcp/server-card.json +8 -8
  6. package/README.md +249 -30
  7. package/adapters/chatgpt/openapi.yaml +12 -0
  8. package/adapters/claude/.mcp.json +2 -2
  9. package/adapters/codex/config.toml +2 -2
  10. package/adapters/gemini/function-declarations.json +1 -0
  11. package/adapters/mcp/server-stdio.js +263 -11
  12. package/adapters/opencode/opencode.json +1 -1
  13. package/bench/thumbgate-bench.json +2 -2
  14. package/bin/cli.js +1429 -121
  15. package/bin/postinstall.js +1 -8
  16. package/config/gate-classifier-routing.json +98 -0
  17. package/config/gate-templates.json +216 -0
  18. package/config/gates/claim-verification.json +12 -0
  19. package/config/gates/default.json +31 -2
  20. package/config/github-about.json +2 -2
  21. package/config/mcp-allowlists.json +23 -13
  22. package/config/merge-quality-checks.json +0 -1
  23. package/config/model-candidates.json +121 -6
  24. package/config/post-deploy-marketing-pages.json +80 -0
  25. package/config/tessl-tiles.json +1 -3
  26. package/openapi/openapi.yaml +12 -0
  27. package/package.json +1 -1
  28. package/public/blog.html +4 -4
  29. package/public/codex-plugin.html +72 -20
  30. package/public/compare.html +31 -8
  31. package/public/dashboard.html +930 -166
  32. package/public/federal.html +2 -2
  33. package/public/guide.html +33 -13
  34. package/public/index.html +469 -111
  35. package/public/learn.html +183 -18
  36. package/public/lessons.html +168 -10
  37. package/public/numbers.html +7 -7
  38. package/public/pro.html +34 -11
  39. package/scripts/agent-memory-lifecycle.js +211 -0
  40. package/scripts/agent-readiness.js +20 -3
  41. package/scripts/agent-reward-model.js +53 -1
  42. package/scripts/auto-promote-gates.js +82 -10
  43. package/scripts/auto-wire-hooks.js +14 -0
  44. package/scripts/billing.js +93 -1
  45. package/scripts/bot-detection.js +61 -3
  46. package/scripts/build-metadata.js +50 -10
  47. package/scripts/cli-feedback.js +4 -2
  48. package/scripts/cli-schema.js +97 -0
  49. package/scripts/cli-telemetry.js +6 -1
  50. package/scripts/commercial-offer.js +82 -2
  51. package/scripts/context-manager.js +74 -6
  52. package/scripts/dashboard.js +68 -2
  53. package/scripts/export-databricks-bundle.js +5 -1
  54. package/scripts/export-dpo-pairs.js +7 -2
  55. package/scripts/feedback-loop.js +123 -1
  56. package/scripts/feedback-quality.js +87 -0
  57. package/scripts/filesystem-search.js +35 -10
  58. package/scripts/gate-stats.js +89 -0
  59. package/scripts/gates-engine.js +1176 -85
  60. package/scripts/gemini-embedding-policy.js +2 -1
  61. package/scripts/hook-runtime.js +20 -14
  62. package/scripts/hook-thumbgate-cache-updater.js +18 -2
  63. package/scripts/hybrid-feedback-context.js +142 -7
  64. package/scripts/lesson-inference.js +8 -3
  65. package/scripts/lesson-search.js +17 -1
  66. package/scripts/license.js +10 -10
  67. package/scripts/llm-client.js +169 -4
  68. package/scripts/local-model-profile.js +15 -8
  69. package/scripts/mcp-config.js +7 -1
  70. package/scripts/memory-scope-readiness.js +159 -0
  71. package/scripts/meta-agent-loop.js +36 -0
  72. package/scripts/operational-integrity.js +39 -5
  73. package/scripts/oss-pr-opportunity-scout.js +35 -5
  74. package/scripts/plausible-server-events.js +9 -6
  75. package/scripts/pro-local-dashboard.js +4 -4
  76. package/scripts/proxy-pointer-rag-guardrails.js +42 -1
  77. package/scripts/published-cli.js +0 -8
  78. package/scripts/rate-limiter.js +64 -13
  79. package/scripts/secret-scanner.js +44 -5
  80. package/scripts/security-scanner.js +260 -10
  81. package/scripts/self-distill-agent.js +3 -1
  82. package/scripts/seo-gsd.js +916 -7
  83. package/scripts/statusline-cache-path.js +17 -2
  84. package/scripts/statusline-local-stats.js +9 -1
  85. package/scripts/statusline-meta.js +28 -2
  86. package/scripts/statusline.sh +20 -4
  87. package/scripts/telemetry-analytics.js +357 -0
  88. package/scripts/thompson-sampling.js +31 -10
  89. package/scripts/thumbgate-bench.js +16 -1
  90. package/scripts/thumbgate-search.js +85 -19
  91. package/scripts/tool-registry.js +169 -1
  92. package/scripts/vector-store.js +45 -0
  93. package/scripts/workflow-sentinel.js +286 -53
  94. package/scripts/workspace-evolver.js +62 -2
  95. package/src/api/server.js +2683 -319
  96. package/scripts/bot-detector.js +0 -50
package/public/pro.html CHANGED
@@ -11,11 +11,12 @@ __GOOGLE_SITE_VERIFICATION_META__
11
11
  <meta property="og:type" content="website">
12
12
  <meta property="og:url" content="__APP_ORIGIN__/pro">
13
13
  <link rel="canonical" href="__APP_ORIGIN__/pro">
14
+ <link rel="alternate" type="text/markdown" title="ThumbGate LLM context" href="__APP_ORIGIN__/llm-context.md">
14
15
  <link rel="icon" type="image/png" href="/thumbgate-icon.png">
15
16
  <link rel="apple-touch-icon" href="/assets/brand/thumbgate-mark.svg">
16
17
  <meta property="og:image" content="/og.png">
17
18
 
18
- <script defer data-domain="thumbgate-production.up.railway.app" src="https://plausible.io/js/script.js"></script>
19
+ <script defer data-domain="thumbgate.ai" src="https://plausible.io/js/script.js"></script>
19
20
  __GA_BOOTSTRAP__
20
21
 
21
22
  <script>
@@ -36,7 +37,7 @@ __GA_BOOTSTRAP__
36
37
  <script type="application/ld+json">
37
38
  {
38
39
  "@context": "https://schema.org",
39
- "@type": "FAQPage", "mainEntity": [{ "@type": "Question", "name": "How is Pro different from the free install?", "acceptedAnswer": { "@type": "Answer", "text": "Free keeps local recall, checks, and MCP. Pro adds the personal dashboard, DPO export, auto-connect, and founder support." } }, { "@type": "Question", "name": "Does Pro require a cloud account?", "acceptedAnswer": { "@type": "Answer", "text": "No. Pro stays local-first; Team is the hosted rollout lane for shared lessons, org visibility, and reviews." } }, { "@type": "Question", "name": "What happens after checkout?", "acceptedAnswer": { "@type": "Answer", "text": "You activate Pro, connect the local dashboard, and inspect blocked actions, lessons, and exports." } }, { "@type": "Question", "name": "When should I choose Team instead of Pro?", "acceptedAnswer": { "@type": "Answer", "text": "Choose Team when one correction needs to protect multiple developers or agents across shared repositories." } }]
40
+ "@type": "FAQPage", "mainEntity": [{ "@type": "Question", "name": "How is Pro different from the free install?", "acceptedAnswer": { "@type": "Answer", "text": "Free keeps local recall, checks, and MCP. Pro adds the personal dashboard, DPO export, auto-connect, and founder support." } }, { "@type": "Question", "name": "Does Pro require a cloud account?", "acceptedAnswer": { "@type": "Answer", "text": "No. Pro stays local-first; Enterprise is the hosted rollout lane for shared lessons, org visibility, and reviews." } }, { "@type": "Question", "name": "What happens after checkout?", "acceptedAnswer": { "@type": "Answer", "text": "You activate Pro, connect the local dashboard, and inspect blocked actions, lessons, and exports." } }, { "@type": "Question", "name": "When should I choose Enterprise instead of Pro?", "acceptedAnswer": { "@type": "Answer", "text": "Choose Enterprise when one correction needs to protect multiple developers or agents across shared repositories." } }]
40
41
  }
41
42
  </script>
42
43
 
@@ -711,7 +712,7 @@ __GA_BOOTSTRAP__
711
712
  <body>
712
713
  <nav>
713
714
  <div class="container">
714
- <a href="/" class="nav-logo"><img src="/assets/brand/thumbgate-mark-inline.svg" alt="ThumbGate" class="logo-mark" width="28" height="28"><span class="logo-text">ThumbGate Pro</span></a>
715
+ <a href="/" class="nav-logo"><img src="/assets/brand/thumbgate-mark-inline-v3.svg" alt="ThumbGate" class="logo-mark" width="28" height="28"><span class="logo-text">ThumbGate Pro</span></a>
715
716
  <div class="nav-links">
716
717
  <a href="#why-pay">Why Pro</a>
717
718
  <a href="#proof">Proof</a>
@@ -730,7 +731,7 @@ __GA_BOOTSTRAP__
730
731
  <h1>Buy the operator loop that proves your AI agent stopped repeating the mistake.</h1>
731
732
  <p style="font-size:13px;opacity:0.8;margin-bottom:0.5rem;">Updated: <time datetime="2026-04-20">2026-04-20</time> · by <a href="https://github.com/IgorGanapolsky" style="color:inherit;">Igor Ganapolsky</a></p>
732
733
  <p>ThumbGate Pro is for one operator who already hit a repeated AI-agent failure and now needs proof: what was blocked, why it was blocked, and what changed before the next risky run.</p>
733
- <p>Start Pro when you want the local dashboard, DPO export, and a single proof lane for the repeated mistake you need to stop. Team diagnostics and custom services are handled through intake, not this buyer path.</p>
734
+ <p>Start Pro when you want the local dashboard, DPO export, and a single proof lane for the repeated mistake you need to stop. Enterprise diagnostics and custom services are handled through intake, not this buyer path.</p>
734
735
  <div class="hero-proof">
735
736
  <div class="proof-pill">Personal local dashboard</div>
736
737
  <div class="proof-pill">DPO export from real corrections</div>
@@ -815,6 +816,28 @@ __GA_BOOTSTRAP__
815
816
  </div>
816
817
  </section>
817
818
 
819
+ <section class="section" id="deterministic-loop">
820
+ <div class="container">
821
+ <div class="section-label">Why Pro now</div>
822
+ <h2 class="section-title">Black-box thumbs do not prove prevention. Pro gives the operator an audit loop.</h2>
823
+ <p class="section-intro">Native rating buttons can tell a vendor that an answer felt wrong. ThumbGate Pro gives you the operational record: the correction, the lesson, the rule, the blocked tool call, and the export path.</p>
824
+ <div class="grid-3">
825
+ <div class="feature-card">
826
+ <h3>Inspectable memory</h3>
827
+ <p>Search the exact lesson that came from a thumbs-down and see whether it is still active, warning-only, or blocking.</p>
828
+ </div>
829
+ <div class="feature-card">
830
+ <h3>Deterministic checks</h3>
831
+ <p>The enforcement layer evaluates tool name, arguments, working directory, command shape, confidence, and required evidence before the action runs.</p>
832
+ </div>
833
+ <div class="feature-card">
834
+ <h3>Exportable proof</h3>
835
+ <p>Take the same correction history into JSONL, DPO export, review packets, and team rollout conversations instead of trusting hidden memory.</p>
836
+ </div>
837
+ </div>
838
+ </div>
839
+ </section>
840
+
818
841
  <section class="section" id="why-pay">
819
842
  <div class="container">
820
843
  <div class="section-label">Why operators pay</div>
@@ -901,11 +924,11 @@ __GA_BOOTSTRAP__
901
924
 
902
925
  <div class="pricing-sidebar">
903
926
  <div class="team-card">
904
- <div class="section-label" style="text-align:left;margin-bottom:8px;">When Team is better</div>
927
+ <div class="section-label" style="text-align:left;margin-bottom:8px;">When Enterprise is better</div>
905
928
  <h3>Need shared enforcement?</h3>
906
- <p>Choose Team when one correction must protect multiple developers or agents across shared repositories, CI, approval policies, and audit trails. Team is $49/seat/mo with a 3-seat minimum after qualification.</p>
929
+ <p>Choose Enterprise when one correction must protect multiple developers or agents across shared repositories, CI, approval policies, and audit trails. Enterprise is custom pricing, scoped after intake.</p>
907
930
  <div class="hero-actions" style="margin-top:18px;">
908
- <a class="btn-secondary" href="/#workflow-sprint-intake">Book a Team Pilot Call</a>
931
+ <a class="btn-secondary" href="/#workflow-sprint-intake">Book an Enterprise Pilot Call</a>
909
932
  </div>
910
933
  </div>
911
934
  <div class="team-card">
@@ -928,15 +951,15 @@ __GA_BOOTSTRAP__
928
951
  </div>
929
952
  <div class="faq-item">
930
953
  <button class="faq-q" type="button" onclick="toggleFaq(this)" onkeydown="handleFaqKeydown(event)" aria-expanded="false">Does Pro require a cloud account?</button>
931
- <div class="faq-a">No. Pro is still local-first for the individual operator lane. Team is the hosted rollout lane for shared lessons, org visibility, and hosted review views.</div>
954
+ <div class="faq-a">No. Pro is still local-first for the individual operator lane. Enterprise is the hosted rollout lane for shared lessons, org visibility, and hosted review views.</div>
932
955
  </div>
933
956
  <div class="faq-item">
934
957
  <button class="faq-q" type="button" onclick="toggleFaq(this)" onkeydown="handleFaqKeydown(event)" aria-expanded="false">What happens after checkout?</button>
935
958
  <div class="faq-a">You activate Pro, connect the personal local dashboard, and your running agents can appear automatically so you can inspect blocked actions, active lessons, and exportable DPO pairs without adding a separate cloud dashboard dependency.</div>
936
959
  </div>
937
960
  <div class="faq-item">
938
- <button class="faq-q" type="button" onclick="toggleFaq(this)" onkeydown="handleFaqKeydown(event)" aria-expanded="false">When should I choose Team instead of Pro?</button>
939
- <div class="faq-a">Choose Team when one thumbs-down should protect multiple people or agents across shared repositories, or when you need shared hosted lessons, org dashboard visibility, and a workflow hardening pilot with rollout review views.</div>
961
+ <button class="faq-q" type="button" onclick="toggleFaq(this)" onkeydown="handleFaqKeydown(event)" aria-expanded="false">When should I choose Enterprise instead of Pro?</button>
962
+ <div class="faq-a">Choose Enterprise when one thumbs-down should protect multiple people or agents across shared repositories, or when you need shared hosted lessons, org dashboard visibility, and a workflow hardening pilot with rollout review views.</div>
940
963
  </div>
941
964
  </div>
942
965
  </div>
@@ -964,7 +987,7 @@ __GA_BOOTSTRAP__
964
987
  <a href="__VERIFICATION_URL__" target="_blank" rel="noopener">Verification Evidence</a>
965
988
  <a href="https://github.com/IgorGanapolsky/ThumbGate" target="_blank" rel="noopener">GitHub</a>
966
989
  </div>
967
- <div class="footer-copy">ThumbGate Pro for individual operators. Team stays intake-first.</div>
990
+ <div class="footer-copy">ThumbGate Pro for individual operators. Enterprise stays intake-first.</div>
968
991
  </div>
969
992
  </footer>
970
993
 
package/scripts/agent-memory-lifecycle.js CHANGED
@@ -2,6 +2,40 @@
2
2
  'use strict';
3
3
 
4
4
  const MEMORY_TYPES = new Set(['episodic', 'semantic', 'procedural', 'preference', 'working']);
5
+ const MEMORY_SCOPES = new Set(['task', 'session', 'user', 'project', 'org']);
6
+ const HIGH_RISK_TERMS = new Set([
7
+ 'billing',
8
+ 'checkout',
9
+ 'compliance',
10
+ 'credential',
11
+ 'data-loss',
12
+ 'deploy',
13
+ 'deployment',
14
+ 'git',
15
+ 'payment',
16
+ 'production',
17
+ 'release',
18
+ 'secret',
19
+ 'security',
20
+ 'stripe',
21
+ 'verification',
22
+ ]);
23
+ const KNOWN_ENTITY_PATTERNS = [
24
+ ['Claude Code', /\bclaude\s+code\b/i, 'agent'],
25
+ ['Codex', /\bcodex\b/i, 'agent'],
26
+ ['Cursor', /\bcursor\b/i, 'agent'],
27
+ ['Gemini CLI', /\bgemini\s+cli\b/i, 'agent'],
28
+ ['MCP', /\bmcp\b/i, 'protocol'],
29
+ ['Stripe', /\bstripe\b/i, 'service'],
30
+ ['GitHub', /\bgithub\b|\bgh\s+/i, 'service'],
31
+ ['Railway', /\brailway\b/i, 'service'],
32
+ ['Plausible', /\bplausible\b/i, 'service'],
33
+ ['PostHog', /\bposthog\b/i, 'service'],
34
+ ['SQLite', /\bsqlite\b|\bfts5\b/i, 'storage'],
35
+ ['LanceDB', /\blancedb\b/i, 'storage'],
36
+ ['Docker', /\bdocker\b/i, 'runtime'],
37
+ ['npm', /\bnpm\b|\bnpx\b/i, 'runtime'],
38
+ ];
5
39
 
6
40
  function normalizeText(value) {
7
41
  if (value === undefined || value === null) return '';
@@ -13,6 +47,178 @@ function normalizeMemoryType(value) {
13
47
  return MEMORY_TYPES.has(normalized) ? normalized : 'episodic';
14
48
  }
15
49
 
50
+ function tokenize(value) {
51
+ return normalizeText(value)
52
+ .toLowerCase()
53
+ .split(/[^a-z0-9_.:/-]+/)
54
+ .filter(Boolean);
55
+ }
56
+
57
+ function uniqueByName(entities) {
58
+ const seen = new Set();
59
+ return entities.filter((entity) => {
60
+ const key = normalizeText(entity.name).toLowerCase();
61
+ if (!key || seen.has(key)) return false;
62
+ seen.add(key);
63
+ return true;
64
+ });
65
+ }
66
+
67
+ function collectMemoryText(memory = {}) {
68
+ return [
69
+ memory.title,
70
+ memory.content,
71
+ memory.context,
72
+ memory.whatWentWrong,
73
+ memory.whatToChange,
74
+ memory.whatWorked,
75
+ memory.domain,
76
+ memory.skill,
77
+ Array.isArray(memory.tags) ? memory.tags.join(' ') : memory.tags,
78
+ ].filter(Boolean).join(' ');
79
+ }
80
+
81
+ function extractMemoryEntities(memory = {}) {
82
+ const text = collectMemoryText(memory);
83
+ const entities = [];
84
+
85
+ for (const [name, pattern, type] of KNOWN_ENTITY_PATTERNS) {
86
+ if (pattern.test(text)) entities.push({ name, type });
87
+ }
88
+
89
+ const commandMatches = text.match(/`([^`]+)`/g) || [];
90
+ for (const match of commandMatches) {
91
+ const command = match.slice(1, -1).trim();
92
+ if (/^(git|npm|npx|node|gh|curl|docker|python|pytest|stripe)\b/i.test(command)) {
93
+ entities.push({ name: command, type: 'command' });
94
+ } else if (/[./-]/.test(command)) {
95
+ entities.push({ name: command, type: 'path' });
96
+ }
97
+ }
98
+
99
+ const pathMatches = text.match(/\b(?:[a-z0-9_-]+\/)+[a-z0-9_.-]+\b/gi) || [];
100
+ for (const filePath of pathMatches.slice(0, 8)) {
101
+ entities.push({ name: filePath, type: 'path' });
102
+ }
103
+
104
+ return uniqueByName(entities).slice(0, 16);
105
+ }
106
+
107
+ function inferMemoryScope(memory = {}) {
108
+ const explicit = normalizeText(memory.scope || memory.memoryScope).toLowerCase();
109
+ if (MEMORY_SCOPES.has(explicit)) return explicit;
110
+
111
+ const text = collectMemoryText(memory).toLowerCase();
112
+ const tags = new Set(Array.isArray(memory.tags) ? memory.tags.map((tag) => normalizeText(tag).toLowerCase()) : []);
113
+
114
+ if (tags.has('preference') || /\b(prefer|style|tone|my preference|user preference)\b/.test(text)) return 'user';
115
+ if (tags.has('org') || tags.has('team') || /\b(enterprise|seat|team|shared|org|compliance|policy|approval)\b/.test(text)) return 'org';
116
+ if (tags.has('repo') || tags.has('project') || tags.has('release') || tags.has('deployment')
117
+ || /\b(repo|repository|branch|ci|pull request|github|deploy|production|release|publish)\b/.test(text)) return 'project';
118
+ if (tags.has('session') || /\b(this session|current session|today|right now)\b/.test(text)) return 'session';
119
+ return 'task';
120
+ }
121
+
122
+ function scoreMemoryDecay(memory = {}, options = {}) {
123
+ const nowMs = options.now ? new Date(options.now).getTime() : Date.now();
124
+ const timestampMs = memory.timestamp ? new Date(memory.timestamp).getTime() : NaN;
125
+ const ageDays = Number.isFinite(timestampMs)
126
+ ? Math.max(0, (nowMs - timestampMs) / (1000 * 60 * 60 * 24))
127
+ : null;
128
+ const textTokens = new Set(tokenize(collectMemoryText(memory)));
129
+ const tags = Array.isArray(memory.tags) ? memory.tags.map((tag) => normalizeText(tag).toLowerCase()) : [];
130
+ const highRisk = tags.some((tag) => HIGH_RISK_TERMS.has(tag))
131
+ || [...textTokens].some((token) => HIGH_RISK_TERMS.has(token))
132
+ || ['critical', 'high'].includes(normalizeText(memory.importance).toLowerCase());
133
+
134
+ if (highRisk) {
135
+ return {
136
+ state: 'sticky',
137
+ ageDays,
138
+ score: 1,
139
+ reason: 'high-risk memories stay retrievable until explicitly retired',
140
+ };
141
+ }
142
+ if (ageDays === null) {
143
+ return {
144
+ state: 'review',
145
+ ageDays,
146
+ score: 0.6,
147
+ reason: 'memory has no timestamp, so it needs review before durable promotion',
148
+ };
149
+ }
150
+ if (ageDays > 180) {
151
+ return {
152
+ state: 'archive_candidate',
153
+ ageDays,
154
+ score: 0.2,
155
+ reason: 'old low-risk memory should be consolidated or archived',
156
+ };
157
+ }
158
+ if (ageDays > 60) {
159
+ return {
160
+ state: 'review',
161
+ ageDays,
162
+ score: 0.55,
163
+ reason: 'older low-risk memory should be refreshed before it dominates recall',
164
+ };
165
+ }
166
+ return {
167
+ state: 'active',
168
+ ageDays,
169
+ score: 0.85,
170
+ reason: 'recent memory remains eligible for recall',
171
+ };
172
+ }
173
+
174
+ function scoreHybridMemoryMatch(query, memory = {}, options = {}) {
175
+ const queryTokens = new Set(tokenize(query));
176
+ const memoryTokens = new Set(tokenize(collectMemoryText(memory)));
177
+ const queryText = normalizeText(query).toLowerCase();
178
+ const memoryText = collectMemoryText(memory).toLowerCase();
179
+ const memoryEntities = extractMemoryEntities(memory);
180
+ const queryEntityNames = extractMemoryEntities({ content: query }).map((entity) => entity.name.toLowerCase());
181
+
182
+ let lexicalMatches = 0;
183
+ for (const token of queryTokens) {
184
+ if (memoryTokens.has(token)) lexicalMatches++;
185
+ }
186
+ const lexicalScore = queryTokens.size > 0 ? lexicalMatches / queryTokens.size : 0;
187
+ const phraseScore = queryText && memoryText.includes(queryText) ? 0.35 : 0;
188
+ const entityMatches = memoryEntities.filter((entity) => queryEntityNames.includes(entity.name.toLowerCase()));
189
+ const entityScore = queryEntityNames.length > 0 ? entityMatches.length / queryEntityNames.length : 0;
190
+ const decay = scoreMemoryDecay(memory, options);
191
+ const lifecycleScore = decay.state === 'archive_candidate' ? -0.15 : decay.state === 'sticky' ? 0.12 : 0;
192
+ const score = lexicalScore + phraseScore + (entityScore * 0.45) + lifecycleScore;
193
+
194
+ return {
195
+ score: Number(Math.max(0, score).toFixed(4)),
196
+ lexicalScore: Number(lexicalScore.toFixed(4)),
197
+ entityScore: Number(entityScore.toFixed(4)),
198
+ matchedEntities: entityMatches,
199
+ decayState: decay.state,
200
+ };
201
+ }
202
+
203
+ function buildMemoryLifecycleView(memory = {}, options = {}) {
204
+ const scope = inferMemoryScope(memory);
205
+ const entities = extractMemoryEntities(memory);
206
+ const decay = scoreMemoryDecay(memory, options);
207
+ const retrieval = scoreHybridMemoryMatch(options.query || '', memory, options);
208
+
209
+ return {
210
+ scope,
211
+ entities,
212
+ decay,
213
+ retrievalHints: {
214
+ hybridScore: retrieval.score,
215
+ lexicalScore: retrieval.lexicalScore,
216
+ entityScore: retrieval.entityScore,
217
+ matchedEntities: retrieval.matchedEntities,
218
+ },
219
+ };
220
+ }
221
+
16
222
  function buildMemoryLifecyclePolicy(input = {}) {
17
223
  return {
18
224
  generatedAt: normalizeText(input.generatedAt) || new Date().toISOString(),
@@ -91,6 +297,11 @@ function evaluateMemoryPromotion(memory = {}, policy = buildMemoryLifecyclePolic
91
297
 
92
298
  module.exports = {
93
299
  buildMemoryLifecyclePolicy,
300
+ buildMemoryLifecycleView,
94
301
  evaluateMemoryPromotion,
302
+ extractMemoryEntities,
303
+ inferMemoryScope,
95
304
  normalizeMemoryType,
305
+ scoreHybridMemoryMatch,
306
+ scoreMemoryDecay,
96
307
  };
package/scripts/agent-readiness.js CHANGED
@@ -94,9 +94,26 @@ function detectRuntimeIsolation() {
94
94
  };
95
95
  }
96
96
 
97
- function collectBootstrapFiles(projectRoot = PROJECT_ROOT) {
97
+ function findProjectRoot(startDir = process.cwd()) {
98
+ try {
99
+ let curr = path.resolve(startDir);
100
+ while (true) {
101
+ const indicators = ['AGENTS.md', 'CLAUDE.md', 'GEMINI.md', '.mcp.json', '.git'];
102
+ if (indicators.some((f) => fs.existsSync(path.join(curr, f)))) {
103
+ return curr;
104
+ }
105
+ const parent = path.dirname(curr);
106
+ if (parent === curr) break;
107
+ curr = parent;
108
+ }
109
+ } catch (_) { /* fallback to startDir */ }
110
+ return startDir;
111
+ }
112
+
113
+ function collectBootstrapFiles(projectRoot) {
114
+ const effectiveRoot = projectRoot || findProjectRoot();
98
115
  const files = BOOTSTRAP_FILES.map((file) => {
99
- const absolutePath = path.join(projectRoot, file.path);
116
+ const absolutePath = path.join(effectiveRoot, file.path);
100
117
  return {
101
118
  id: file.id,
102
119
  path: file.path,
@@ -118,7 +135,7 @@ function collectBootstrapFiles(projectRoot = PROJECT_ROOT) {
118
135
  missingRequired,
119
136
  recommendation: missingRequired.length === 0
120
137
  ? 'Bootstrap context is present.'
121
- : `Add missing bootstrap files: ${missingRequired.join(', ')}`,
138
+ : `Add missing bootstrap files to project root (${effectiveRoot}): ${missingRequired.join(', ')}`,
122
139
  };
123
140
  }
124
141
 
package/scripts/agent-reward-model.js CHANGED
@@ -331,6 +331,55 @@ function buildRewardReport(episodes = [], options = {}) {
331
331
  deep: 'destructive, public, or production-adjacent work',
332
332
  xhigh: 'payments, secrets, deploy-prod, data-loss, force-push-main',
333
333
  },
334
+ continualAdapterPlan: buildContinualAdapterTrainingPlan(episodes, options),
335
+ };
336
+ }
337
+
338
+ function buildContinualAdapterTrainingPlan(episodes = [], options = {}) {
339
+ const maxAdapters = Math.max(1, Number(options.maxAdapters || 4));
340
+ const minExamples = Math.max(1, Number(options.minExamples || 2));
341
+ const pairs = buildPreferencePairs(episodes, { ...options, maxPairs: Number(options.maxPairs || 24) });
342
+ const candidates = rankGateCandidatesByReward(episodes, {
343
+ ...options,
344
+ minOccurrences: minExamples,
345
+ maxGateCandidates: maxAdapters,
346
+ }).slice(0, maxAdapters);
347
+
348
+ const adapters = candidates.map((candidate, index) => ({
349
+ id: `lora_${candidate.gateId || `adapter_${index + 1}`}`.slice(0, 96),
350
+ source: candidate.key,
351
+ trainingMode: 'concurrent-lora',
352
+ examples: candidate.occurrences,
353
+ priorityScore: candidate.priorityScore,
354
+ targetBehavior: candidate.recommendation,
355
+ retentionChecks: [
356
+ 'baseline gate pass-rate must not regress',
357
+ 'previous adapter reward must stay within tolerance',
358
+ 'new adapter must improve blocked-repeat or verification reward on held-out episodes',
359
+ ],
360
+ }));
361
+
362
+ const enoughData = pairs.length >= minExamples && adapters.length > 0;
363
+ return {
364
+ status: enoughData ? 'ready' : 'needs_more_feedback',
365
+ trainingStack: 'multi-lora-continual-learning',
366
+ baseModelPolicy: 'freeze base model; train small adapters from reward-ranked DPO pairs',
367
+ scheduling: adapters.length > 1 ? 'batch adapters concurrently when they share the same frozen base' : 'train serially until at least two adapter candidates exist',
368
+ adapters,
369
+ dpoPairsAvailable: pairs.length,
370
+ retentionGate: 'ship no adapter unless reward, regression, and prior-domain retention checks pass',
371
+ nextActions: enoughData
372
+ ? [
373
+ 'Export reward-ranked DPO pairs.',
374
+ 'Train candidate LoRA adapters concurrently on the same frozen base.',
375
+ 'Evaluate each adapter against held-out negative episodes and prior positive episodes.',
376
+ 'Promote only the adapter that improves reward without retention regression.',
377
+ ]
378
+ : [
379
+ 'Capture more thumbs-up/down outcomes before training.',
380
+ 'Promote deterministic gates first when examples are sparse.',
381
+ 'Use local semantic recall until enough reward-ranked pairs exist for adapter training.',
382
+ ],
334
383
  };
335
384
  }
336
385
 
@@ -419,8 +468,10 @@ if (isCliInvocation()) {
419
468
  console.log(JSON.stringify(buildPreferencePairs(episodes), null, 2));
420
469
  } else if (args.command === 'gates') {
421
470
  console.log(JSON.stringify(rankGateCandidatesByReward(episodes), null, 2));
471
+ } else if (args.command === 'adapters') {
472
+ console.log(JSON.stringify(buildContinualAdapterTrainingPlan(episodes), null, 2));
422
473
  } else {
423
- console.error(`Unknown command: ${args.command}. Use: report, pairs, gates`);
474
+ console.error(`Unknown command: ${args.command}. Use: report, pairs, gates, adapters`);
424
475
  process.exit(1);
425
476
  }
426
477
  }
@@ -428,6 +479,7 @@ if (isCliInvocation()) {
428
479
  module.exports = {
429
480
  HIGH_RISK_TAGS,
430
481
  allocateTestTimeCompute,
482
+ buildContinualAdapterTrainingPlan,
431
483
  buildPreferencePairFromEpisodes,
432
484
  buildPreferencePairs,
433
485
  buildRewardReport,
package/scripts/auto-promote-gates.js CHANGED
@@ -58,6 +58,47 @@ function readJSONL(filePath) {
58
58
  }).filter(Boolean);
59
59
  }
60
60
 
61
+ // --- Self-Harness stage 3: regression-gated promotion -----------------------
62
+ // Inspired by "Self-Harness: Harnesses That Improve Themselves" (arXiv 2606.09498).
63
+ // Stages 1-2 (weakness mining -> rule extraction) already exist via lesson
64
+ // inference + this promoter. Stage 3 — accept a harness change only after
65
+ // regression-testing it does not degrade behavior — was missing: a noisy 3x
66
+ // capture could hard-block an over-broad pattern with no check that it wouldn't
67
+ // have wrongly blocked actions that were previously ALLOWED. This replays a
68
+ // candidate BLOCK rule against the audit trail's prior `allow` decisions; if it
69
+ // would have blocked safe actions, the caller quarantines it to `warn` instead.
70
+ const REGRESSION_FALSE_BLOCK_LIMIT = 0; // any prior safe action it would block => quarantine
71
+
72
+ function getAuditTrailPath() {
73
+ return path.join(path.dirname(getFeedbackLogPath()), 'audit-trail.jsonl');
74
+ }
75
+
76
+ // Returns { falseBlocks, allowSampleSize } or null when there is no history /
77
+ // matcher available — in which case the caller promotes as usual (fail-open to
78
+ // existing behavior, since regression gating is an enhancement, not a hard gate).
79
+ function regressionCheck(gate, options = {}) {
80
+ const auditPath = options.auditTrailPath || getAuditTrailPath();
81
+ const entries = readJSONL(auditPath);
82
+ if (!entries.length) return null;
83
+ // Lazy-require to avoid the gates-engine <-> auto-promote-gates require cycle.
84
+ let matchesGate;
85
+ try { ({ matchesGate } = require('./gates-engine')); } catch { return null; }
86
+ if (typeof matchesGate !== 'function') return null;
87
+ const allowed = entries.filter((e) => e && e.decision === 'allow' && e.toolName);
88
+ if (!allowed.length) return null;
89
+ let falseBlocks = 0;
90
+ for (const e of allowed) {
91
+ try {
92
+ if (matchesGate(gate, e.toolName, e.toolInput || {})) falseBlocks += 1;
93
+ } catch { /* a bad pattern/entry never counts as a false block */ }
94
+ }
95
+ return { falseBlocks, allowSampleSize: allowed.length };
96
+ }
97
+
98
+ function safeRegressionCheck(gate, options) {
99
+ try { return regressionCheck(gate, options); } catch { return null; }
100
+ }
101
+
61
102
  function loadAutoGates() {
62
103
  const autoGatesPath = getAutoGatesPath();
63
104
  if (!fs.existsSync(autoGatesPath)) {
@@ -193,9 +234,9 @@ function patternToGateId(key) {
193
234
  return 'auto-' + key.replace(/[^a-z0-9]+/gi, '-').replace(/^-|-$/g, '').slice(0, 50).toLowerCase();
194
235
  }
195
236
 
196
- function buildGateRule(group) {
197
- const action = group.count === 'MANUAL' ? group.manualAction || 'block' : (group.count >= BLOCK_THRESHOLD ? 'block' : 'warn');
198
- const severity = action === 'block' ? 'critical' : 'medium';
237
+ function buildGateRule(group, actionOverride) {
238
+ const action = actionOverride || (group.count === 'MANUAL' ? group.manualAction || 'block' : (group.count >= BLOCK_THRESHOLD ? 'block' : 'warn'));
239
+ const severity = action === 'block' ? 'critical' : action === 'approve' ? 'high' : 'medium';
199
240
  const context = group.latestContext.slice(0, 120);
200
241
  const kind = group.key.startsWith('diagnosis:')
201
242
  ? 'repeated diagnosis'
@@ -332,7 +373,8 @@ function forcePromote(context, action = 'block') {
332
373
  return { gateId, action, totalGates: data.gates.length };
333
374
  }
334
375
 
335
- function promote(feedbackLogPath) {
376
+ function promote(feedbackLogPath, options) {
377
+ const opts = options || {};
336
378
  const logPath = feedbackLogPath || getFeedbackLogPath();
337
379
  const entries = readJSONL(logPath);
338
380
  const groups = groupNegativeFeedback(entries, WINDOW_DAYS);
@@ -357,17 +399,38 @@ function promote(feedbackLogPath) {
357
399
  const existing = data.gates[existingIdx];
358
400
  const newAction = group.count >= BLOCK_THRESHOLD ? 'block' : 'warn';
359
401
  if (existing.action !== newAction && newAction === 'block') {
360
- // Upgrade from warn to block
361
- data.gates[existingIdx] = { ...existing, action: 'block', severity: 'critical', occurrences: group.count, upgradedAt: new Date().toISOString() };
362
- promotions.push({ type: 'upgrade', gateId, from: existing.action, to: 'block', occurrences: group.count });
402
+ // Self-Harness stage 3: regression-test before upgrading warn -> block.
403
+ const regression = opts.skipRegression ? null : safeRegressionCheck(buildGateRule(group, 'block'), opts);
404
+ if (regression && regression.falseBlocks > REGRESSION_FALSE_BLOCK_LIMIT) {
405
+ // Would block prior safe actions — hold at warn instead of upgrading.
406
+ promotions.push({ type: 'upgrade-quarantined', gateId, from: existing.action, occurrences: group.count, falseBlocks: regression.falseBlocks });
407
+ } else {
408
+ // Upgrade from warn to block
409
+ data.gates[existingIdx] = { ...existing, action: 'block', severity: 'critical', occurrences: group.count, upgradedAt: new Date().toISOString() };
410
+ promotions.push({ type: 'upgrade', gateId, from: existing.action, to: 'block', occurrences: group.count });
411
+ }
363
412
  }
364
413
  // Update occurrence count even if no action change
365
414
  data.gates[existingIdx].occurrences = group.count;
366
415
  continue;
367
416
  }
368
417
 
369
- // New gate
370
- const gate = buildGateRule(group);
418
+ // New gate — respect explicit gateAction override (e.g. 'approve' for human-approval rules)
419
+ const gate = buildGateRule(group, opts.gateAction);
420
+
421
+ // Self-Harness stage 3: before a feedback rule goes live as a hard block,
422
+ // regression-test it against prior allowed actions. If it would have blocked
423
+ // safe actions, quarantine it to `warn` instead of `block`.
424
+ let regression = null;
425
+ if (gate.action === 'block' && !opts.gateAction && !opts.skipRegression) {
426
+ regression = safeRegressionCheck(gate, opts);
427
+ if (regression && regression.falseBlocks > REGRESSION_FALSE_BLOCK_LIMIT) {
428
+ gate.action = 'warn';
429
+ gate.severity = 'medium';
430
+ gate.quarantined = true;
431
+ gate.regression = regression;
432
+ }
433
+ }
371
434
 
372
435
  // Enforce max limit — rotate oldest
373
436
  if (data.gates.length >= MAX_AUTO_GATES) {
@@ -376,7 +439,13 @@ function promote(feedbackLogPath) {
376
439
  }
377
440
 
378
441
  data.gates.push(gate);
379
- promotions.push({ type: 'new', gateId: gate.id, action: gate.action, occurrences: group.count });
442
+ promotions.push({
443
+ type: gate.quarantined ? 'new-quarantined' : 'new',
444
+ gateId: gate.id,
445
+ action: gate.action,
446
+ occurrences: group.count,
447
+ ...(gate.quarantined ? { falseBlocks: regression.falseBlocks, allowSampleSize: regression.allowSampleSize } : {}),
448
+ });
380
449
  }
381
450
 
382
451
  // Log promotions
@@ -437,6 +506,9 @@ module.exports = {
437
506
  groupNegativeFeedback,
438
507
  patternToGateId,
439
508
  buildGateRule,
509
+ regressionCheck,
510
+ getAuditTrailPath,
511
+ REGRESSION_FALSE_BLOCK_LIMIT,
440
512
  extractPatternKey,
441
513
  normalizeCommandSignature,
442
514
  isNegative,
package/scripts/auto-wire-hooks.js CHANGED
@@ -27,6 +27,7 @@ const {
27
27
  statuslineCommand,
28
28
  userPromptHookCommand,
29
29
  } = require('./hook-runtime');
30
+ const { installShim } = require('./install-shim');
30
31
 
31
32
  function getHome() {
32
33
  return process.env.HOME || process.env.USERPROFILE || '';
@@ -338,6 +339,19 @@ function wireClaudeHooks(options) {
338
339
  options.projectSettingsPath || claudeProjectSettingsPath(options.projectDir);
339
340
  const dryRun = options.dryRun || false;
340
341
  const projectDir = options.projectDir || process.cwd();
342
+
343
+ // --- Install stable shim before resolving hook commands ---
344
+ // The shim at ~/.thumbgate/bin/thumbgate-hook always resolves @latest,
345
+ // so hooks never go stale across version bumps (Volta-style pattern).
346
+ // Skip in source-checkout mode — developers use direct node commands.
347
+ if (!dryRun && !require('./mcp-config').isSourceCheckout(path.join(__dirname, '..'))) {
348
+ try {
349
+ installShim();
350
+ } catch {
351
+ // Non-fatal: fall back to version-pinned commands
352
+ }
353
+ }
354
+
341
355
  const desiredStatusLine = statuslineCommand();
342
356
 
343
357
  // --- Step 0: clean up stale hooks from BOTH settings locations ---