@tangle-network/agent-runtime 0.105.0 → 0.106.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (218) hide show
  1. package/README.md +2 -2
  2. package/dist/activation-CRjVfOiy.js +203 -0
  3. package/dist/activation-CRjVfOiy.js.map +1 -0
  4. package/dist/activation-DyWB0K6E.d.ts +98 -0
  5. package/dist/agent.d.ts +216 -234
  6. package/dist/agent.js +615 -547
  7. package/dist/agent.js.map +1 -1
  8. package/dist/analyst-loop-DGdtVENg.js +517 -0
  9. package/dist/analyst-loop-DGdtVENg.js.map +1 -0
  10. package/dist/analyst-loop.d.ts +10 -43
  11. package/dist/analyst-loop.js +2 -11
  12. package/dist/backends-Dy8NujB4.js +713 -0
  13. package/dist/backends-Dy8NujB4.js.map +1 -0
  14. package/dist/candidate-execution/index.d.ts +4 -192
  15. package/dist/candidate-execution/index.js +5 -66
  16. package/dist/candidate-execution-DqMbv1rU.js +488 -0
  17. package/dist/candidate-execution-DqMbv1rU.js.map +1 -0
  18. package/dist/coder-yhVWbdWc.js +40 -0
  19. package/dist/coder-yhVWbdWc.js.map +1 -0
  20. package/dist/conversation-XucG_K3l.js +1265 -0
  21. package/dist/conversation-XucG_K3l.js.map +1 -0
  22. package/dist/conversation.d.ts +2 -804
  23. package/dist/conversation.js +2 -55
  24. package/dist/environment-provider-CfLUmKmf.js +1168 -0
  25. package/dist/environment-provider-CfLUmKmf.js.map +1 -0
  26. package/dist/environment-provider-IUGU3epE.d.ts +964 -0
  27. package/dist/environment-provider.d.ts +2 -78
  28. package/dist/environment-provider.js +2 -18
  29. package/dist/errors-DEAvWQPy.js +111 -0
  30. package/dist/errors-DEAvWQPy.js.map +1 -0
  31. package/dist/improvement-adapter-Cr5TBSji.d.ts +166 -0
  32. package/dist/improvement-cycle-C095kS_g.js +2624 -0
  33. package/dist/improvement-cycle-C095kS_g.js.map +1 -0
  34. package/dist/index-BGdxAwfO.d.ts +838 -0
  35. package/dist/index-C-FYUuFG.d.ts +193 -0
  36. package/dist/index-CbNM40xG.d.ts +140 -0
  37. package/dist/index-DruSKs3Z.d.ts +7224 -0
  38. package/dist/index-Efjb3nrQ.d.ts +671 -0
  39. package/dist/index.d.ts +14 -1058
  40. package/dist/index.js +2106 -2082
  41. package/dist/index.js.map +1 -1
  42. package/dist/intelligence.d.ts +768 -889
  43. package/dist/intelligence.js +1466 -1367
  44. package/dist/intelligence.js.map +1 -1
  45. package/dist/{kb-gate-CwHO0vz6.d.ts → kb-gate-C8z2juK8.d.ts} +33 -31
  46. package/dist/kb-gate-DpaSwXVx.js +80 -0
  47. package/dist/kb-gate-DpaSwXVx.js.map +1 -0
  48. package/dist/knowledge-Ku-mPP4f.js +433 -0
  49. package/dist/knowledge-Ku-mPP4f.js.map +1 -0
  50. package/dist/knowledge.d.ts +2 -150
  51. package/dist/knowledge.js +2 -40
  52. package/dist/local-harness-Dh8PJ0ot.d.ts +455 -0
  53. package/dist/loop-runner-bin-BXsFEUcf.d.ts +162 -0
  54. package/dist/loop-runner-bin-BwjQ1NfH.js +268 -0
  55. package/dist/loop-runner-bin-BwjQ1NfH.js.map +1 -0
  56. package/dist/loop-runner-bin.d.ts +2 -15
  57. package/dist/loop-runner-bin.js +2 -24
  58. package/dist/loops.d.ts +5 -4164
  59. package/dist/loops.js +8 -419
  60. package/dist/mcp/bin.d.ts +1 -1
  61. package/dist/mcp/bin.js +116 -146
  62. package/dist/mcp/bin.js.map +1 -1
  63. package/dist/mcp/index.d.ts +295 -436
  64. package/dist/mcp/index.js +1242 -1177
  65. package/dist/mcp/index.js.map +1 -1
  66. package/dist/mcp/memory-bin.d.ts +1 -1
  67. package/dist/mcp/memory-bin.js +37 -16
  68. package/dist/mcp/memory-bin.js.map +1 -1
  69. package/dist/memory-server-DL6cE2Ag.js +234 -0
  70. package/dist/memory-server-DL6cE2Ag.js.map +1 -0
  71. package/dist/model-resolution-Btd9iIKV.js +98 -0
  72. package/dist/model-resolution-Btd9iIKV.js.map +1 -0
  73. package/dist/openai-tools-B68JaOCx.d.ts +24 -0
  74. package/dist/openai-tools-CKLy1C7M.js +44 -0
  75. package/dist/openai-tools-CKLy1C7M.js.map +1 -0
  76. package/dist/otel-export-D34IQW1C.js +843 -0
  77. package/dist/otel-export-D34IQW1C.js.map +1 -0
  78. package/dist/platform.d.ts +188 -185
  79. package/dist/platform.js +185 -218
  80. package/dist/platform.js.map +1 -1
  81. package/dist/prepare-B53sPV7M.js +4292 -0
  82. package/dist/prepare-B53sPV7M.js.map +1 -0
  83. package/dist/primeintellect/index.d.ts +149 -145
  84. package/dist/primeintellect/index.js +626 -1003
  85. package/dist/primeintellect/index.js.map +1 -1
  86. package/dist/profiles.d.ts +296 -490
  87. package/dist/profiles.js +722 -749
  88. package/dist/profiles.js.map +1 -1
  89. package/dist/protected-model-port-BLPH5VVv.js +722 -0
  90. package/dist/protected-model-port-BLPH5VVv.js.map +1 -0
  91. package/dist/protected-model-port-BP6Z4eau.d.ts +283 -0
  92. package/dist/protected-redaction--F3v1oo8.js +200 -0
  93. package/dist/protected-redaction--F3v1oo8.js.map +1 -0
  94. package/dist/redact-D-u-rrcn.js +82 -0
  95. package/dist/redact-D-u-rrcn.js.map +1 -0
  96. package/dist/redact-kbIOjfjM.d.ts +452 -0
  97. package/dist/researcher-CoVqNhfI.js +433 -0
  98. package/dist/researcher-CoVqNhfI.js.map +1 -0
  99. package/dist/runtime-BG0BziaP.js +5408 -0
  100. package/dist/runtime-BG0BziaP.js.map +1 -0
  101. package/dist/sandbox-events-DeI5xX8P.js +316 -0
  102. package/dist/sandbox-events-DeI5xX8P.js.map +1 -0
  103. package/dist/spawn-journal-CwPvKUTa.js +456 -0
  104. package/dist/spawn-journal-CwPvKUTa.js.map +1 -0
  105. package/dist/structural-rollout-IXUEplky.js +1445 -0
  106. package/dist/structural-rollout-IXUEplky.js.map +1 -0
  107. package/dist/substrate-B0TYNrXn.js +27 -0
  108. package/dist/substrate-B0TYNrXn.js.map +1 -0
  109. package/dist/{substrate-Dyiqmd3o.d.ts → substrate-BcnuSHXm.d.ts} +57 -54
  110. package/dist/supervise-BiRutHS9.js +7321 -0
  111. package/dist/supervise-BiRutHS9.js.map +1 -0
  112. package/dist/supervisor-DTKhF-RV.js +3826 -0
  113. package/dist/supervisor-DTKhF-RV.js.map +1 -0
  114. package/dist/testing.d.ts +4 -3
  115. package/dist/testing.js +2478 -2637
  116. package/dist/testing.js.map +1 -1
  117. package/dist/tool-server-RcWgLIsL.js +141 -0
  118. package/dist/tool-server-RcWgLIsL.js.map +1 -0
  119. package/dist/trace-propagation-CJJC7SVB.js +57 -0
  120. package/dist/trace-propagation-CJJC7SVB.js.map +1 -0
  121. package/dist/types-BaIpb0XT.d.ts +492 -0
  122. package/dist/types-BevOjfTY.d.ts +765 -0
  123. package/dist/types-C9j4qg6l.d.ts +500 -0
  124. package/dist/types-Cja5OTGh.d.ts +190 -0
  125. package/dist/util-MVgdwuIS.js +145 -0
  126. package/dist/util-MVgdwuIS.js.map +1 -0
  127. package/dist/workspace-archive-BosQiotA.js +575 -0
  128. package/dist/workspace-archive-BosQiotA.js.map +1 -0
  129. package/package.json +43 -56
  130. package/skills/build-with-agent-runtime/SKILL.md +1 -0
  131. package/dist/activation-Cm-9MMX1.d.ts +0 -69
  132. package/dist/analyst-loop.js.map +0 -1
  133. package/dist/backends-BO4Jqoe2.d.ts +0 -134
  134. package/dist/candidate-execution/index.js.map +0 -1
  135. package/dist/chunk-2KGAN2HM.js +0 -1256
  136. package/dist/chunk-2KGAN2HM.js.map +0 -1
  137. package/dist/chunk-2VZYRB55.js +0 -834
  138. package/dist/chunk-2VZYRB55.js.map +0 -1
  139. package/dist/chunk-3MDZX7YU.js +0 -125
  140. package/dist/chunk-3MDZX7YU.js.map +0 -1
  141. package/dist/chunk-5IBQAPVB.js +0 -29
  142. package/dist/chunk-5IBQAPVB.js.map +0 -1
  143. package/dist/chunk-5QOB7H74.js +0 -387
  144. package/dist/chunk-5QOB7H74.js.map +0 -1
  145. package/dist/chunk-C5RTIBNZ.js +0 -48
  146. package/dist/chunk-C5RTIBNZ.js.map +0 -1
  147. package/dist/chunk-CTRA64LY.js +0 -680
  148. package/dist/chunk-CTRA64LY.js.map +0 -1
  149. package/dist/chunk-DPEUKJRO.js +0 -27
  150. package/dist/chunk-DPEUKJRO.js.map +0 -1
  151. package/dist/chunk-FD2MBMOH.js +0 -494
  152. package/dist/chunk-FD2MBMOH.js.map +0 -1
  153. package/dist/chunk-H5QPIZNX.js +0 -907
  154. package/dist/chunk-H5QPIZNX.js.map +0 -1
  155. package/dist/chunk-HLKC4UYB.js +0 -3514
  156. package/dist/chunk-HLKC4UYB.js.map +0 -1
  157. package/dist/chunk-HNP72PNU.js +0 -192
  158. package/dist/chunk-HNP72PNU.js.map +0 -1
  159. package/dist/chunk-ISPWRSEC.js +0 -408
  160. package/dist/chunk-ISPWRSEC.js.map +0 -1
  161. package/dist/chunk-KRBFHMV6.js +0 -1228
  162. package/dist/chunk-KRBFHMV6.js.map +0 -1
  163. package/dist/chunk-M22PL4GH.js +0 -171
  164. package/dist/chunk-M22PL4GH.js.map +0 -1
  165. package/dist/chunk-NURIQBJQ.js +0 -72
  166. package/dist/chunk-NURIQBJQ.js.map +0 -1
  167. package/dist/chunk-OPVWXJ2H.js +0 -75
  168. package/dist/chunk-OPVWXJ2H.js.map +0 -1
  169. package/dist/chunk-PZZKQVQV.js +0 -52
  170. package/dist/chunk-PZZKQVQV.js.map +0 -1
  171. package/dist/chunk-Q2JSAVQ3.js +0 -671
  172. package/dist/chunk-Q2JSAVQ3.js.map +0 -1
  173. package/dist/chunk-RDOAVVHY.js +0 -699
  174. package/dist/chunk-RDOAVVHY.js.map +0 -1
  175. package/dist/chunk-SGKPNBXE.js +0 -60
  176. package/dist/chunk-SGKPNBXE.js.map +0 -1
  177. package/dist/chunk-SMQXZGLZ.js +0 -351
  178. package/dist/chunk-SMQXZGLZ.js.map +0 -1
  179. package/dist/chunk-UHEZW5BU.js +0 -2918
  180. package/dist/chunk-UHEZW5BU.js.map +0 -1
  181. package/dist/chunk-UPWGXKXB.js +0 -254
  182. package/dist/chunk-UPWGXKXB.js.map +0 -1
  183. package/dist/chunk-VISA6CI3.js +0 -200
  184. package/dist/chunk-VISA6CI3.js.map +0 -1
  185. package/dist/chunk-WMTCUOQL.js +0 -474
  186. package/dist/chunk-WMTCUOQL.js.map +0 -1
  187. package/dist/chunk-WRTOVNN4.js +0 -4728
  188. package/dist/chunk-WRTOVNN4.js.map +0 -1
  189. package/dist/chunk-WSTRQZYQ.js +0 -5576
  190. package/dist/chunk-WSTRQZYQ.js.map +0 -1
  191. package/dist/chunk-XBG2W2VW.js +0 -6434
  192. package/dist/chunk-XBG2W2VW.js.map +0 -1
  193. package/dist/chunk-YEJR7IXO.js +0 -68
  194. package/dist/chunk-YEJR7IXO.js.map +0 -1
  195. package/dist/chunk-YJZA2BIK.js +0 -1338
  196. package/dist/chunk-YJZA2BIK.js.map +0 -1
  197. package/dist/completion-gate-BMy5LGoP.d.ts +0 -256
  198. package/dist/conversation.js.map +0 -1
  199. package/dist/coordination-BZZSVYpZ.d.ts +0 -2570
  200. package/dist/environment-provider.js.map +0 -1
  201. package/dist/improvement-adapter-HAZz-7vK.d.ts +0 -203
  202. package/dist/knowledge.js.map +0 -1
  203. package/dist/local-harness-BDNcl6jI.d.ts +0 -175
  204. package/dist/loop-runner-bin-68aoM9-7.d.ts +0 -199
  205. package/dist/loop-runner-bin.js.map +0 -1
  206. package/dist/loops.js.map +0 -1
  207. package/dist/openai-tools-fnj6SRVg.d.ts +0 -46
  208. package/dist/protected-model-port-B4t-OQjL.d.ts +0 -271
  209. package/dist/redact-BuFjxBUJ.d.ts +0 -481
  210. package/dist/sanitize-DEbPNtyI.d.ts +0 -328
  211. package/dist/structural-rollout-DEh6CEsa.d.ts +0 -446
  212. package/dist/supervise-4_48wwvy.d.ts +0 -379
  213. package/dist/types-BCoemcXU.d.ts +0 -802
  214. package/dist/types-BwoZWq-i.d.ts +0 -510
  215. package/dist/types-Cbp72qyu.d.ts +0 -491
  216. package/dist/types-CmYCMbFT.d.ts +0 -203
  217. package/dist/types-CvNwMzQt.d.ts +0 -957
  218. package/dist/worktree-fanout-DxhAWr5Z.d.ts +0 -999
@@ -1 +0,0 @@
1
- {"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}
@@ -1,203 +0,0 @@
1
- import { FindingSubject, AnalystFinding } from '@tangle-network/agent-eval';
2
- import { I as ImprovementProposalSource } from './types-CmYCMbFT.js';
3
-
4
- /**
5
- * `AgentSurfaces` — declarative map of the mutable file/directory paths
6
- * the self-improvement loop can edit on behalf of an agent.
7
- *
8
- * The substrate uses this map to resolve every parsed `FindingSubject`
9
- * (from agent-eval) to a real on-disk path. No per-vertical glue;
10
- * no fabricated paths; no silent `existsSync(...)` skips that hide
11
- * misconfiguration from the operator.
12
- *
13
- * Surfaces are validated at `defineAgent` time — missing paths fail
14
- * loud with a list of every offender. A surface that's not needed
15
- * (e.g. an agent with no RAG corpora) is simply omitted; the loop
16
- * refuses to route those subjects rather than fabricating a target.
17
- */
18
-
19
- /**
20
- * Surface declarations. Every path is repo-relative (or absolute) at
21
- * `defineAgent` time. At resolution time, paths are joined against the
22
- * agent's `repoRoot`.
23
- *
24
- * `systemPrompt`, `tools`, `personas` are DIRECTORIES; the loop appends
25
- * `<section>.md`, `<tool>/README.md`, `<persona-id>.yaml` etc.
26
- * `rubric`, `outputSchema` are SINGLE FILES; the loop edits them in
27
- * place.
28
- *
29
- * `knowledge` is the agent-knowledge root (typically `.agent-knowledge`);
30
- * `applyKnowledgeWriteBlocks` writes pages relative to it.
31
- *
32
- * Optional surfaces (`scaffolding`, `memory`, `rag`, `outputSchema`)
33
- * can be omitted — the loop will reject findings targeting them with a
34
- * clear log message instead of fabricating a path.
35
- */
36
- interface AgentSurfaces {
37
- /** Directory containing one markdown file per system-prompt section. */
38
- systemPrompt: string;
39
- /** Directory containing one subdir per tool (`<tool>/README.md`). */
40
- tools: string;
41
- /** Single file (TypeScript module) defining the rubric weights + dimensions. */
42
- rubric: string;
43
- /** Knowledge-base root; typically `.agent-knowledge`. */
44
- knowledge: string;
45
- /** Directory containing one YAML/JSON file per persona. */
46
- personas: string;
47
- /** Optional: directory containing scaffolding rules (precondition checks, retry policies). */
48
- scaffolding?: string;
49
- /** Optional: memory store path (JSONL / SQLite / DB). */
50
- memory?: string;
51
- /** Optional: directory containing RAG corpora (`<corpus>/<doc-id>.md`). */
52
- rag?: string;
53
- /** Optional: single file defining the output schema (Zod / JSON Schema). */
54
- outputSchema?: string;
55
- /** Optional: directory containing Agent Skill packages. */
56
- skills?: string;
57
- /** Optional: directory containing MCP server/tool configuration. */
58
- mcp?: string;
59
- /** Optional: directory containing hook definitions. */
60
- hooks?: string;
61
- /** Optional: directory containing subagent definitions. */
62
- subagents?: string;
63
- /** Optional: directory containing orchestration/workflow policies. */
64
- workflows?: string;
65
- /** Optional: single file containing rollout-policy settings. */
66
- rolloutPolicy?: string;
67
- /** Optional: single canonical AgentProfile file. */
68
- agentProfile?: string;
69
- /** Optional: source root for code findings. */
70
- code?: string;
71
- }
72
- interface ResolvedSurface {
73
- /** Absolute filesystem path the operator can `cat` / `vim`. */
74
- absolutePath: string;
75
- /** Repo-relative path for PR descriptions, diffs, audit logs. */
76
- repoRelativePath: string;
77
- /** Whether the path currently exists on disk. */
78
- exists: boolean;
79
- /** The substrate's intent: edit an existing file or create a new one. */
80
- intent: 'edit-existing' | 'create-new';
81
- }
82
- /**
83
- * Resolve a parsed `FindingSubject` to the file path the substrate
84
- * should edit (or create) on disk.
85
- *
86
- * Returns `null` when:
87
- * - the subject targets a surface the agent didn't declare
88
- * (e.g. `rag:*` when `surfaces.rag` is undefined), OR
89
- * - the subject is a `cluster` (failure-mode emits these as evidence,
90
- * not actionable mutations — they don't route to a file).
91
- *
92
- * Returns a `ResolvedSurface` with `intent: 'create-new'` when the
93
- * subject names a path that doesn't yet exist (e.g. a new wiki page).
94
- * The caller chooses whether to honour the create — for tightly-managed
95
- * surfaces like `systemPrompt` it's usually a contract violation
96
- * (the analyst named a section that doesn't exist); for `knowledge`
97
- * it's the whole point.
98
- */
99
- declare function resolveSubjectPath(subject: FindingSubject, surfaces: AgentSurfaces, repoRoot: string): ResolvedSurface | null;
100
- /**
101
- * Validate that every declared surface exists on disk under `repoRoot`.
102
- *
103
- * Returns an array of `SurfaceValidationIssue` — empty when all required
104
- * surfaces resolve. `defineAgent` throws with the issues rendered, so
105
- * a misconfigured manifest fails at startup (not at the first finding
106
- * the loop produces 20 minutes later).
107
- */
108
- interface SurfaceValidationIssue {
109
- surface: keyof AgentSurfaces;
110
- path: string;
111
- reason: 'missing' | 'not-directory' | 'not-file';
112
- }
113
- /** Validate an `AgentSurfaces` map on disk — missing paths fail loud at `defineAgent` time instead of silently skipping self-improvement edits. */
114
- declare function validateSurfaces(surfaces: AgentSurfaces, repoRoot: string): ReadonlyArray<SurfaceValidationIssue>;
115
- /** Format a list of surface validation issues into a human-readable error string. */
116
- declare function renderSurfaceIssues(issues: ReadonlyArray<SurfaceValidationIssue>, repoRoot: string): string;
117
-
118
- /**
119
- * Surface improvement proposer — resolves analyst findings into LLM-drafted
120
- * candidate patches without changing the caller's repository.
121
- *
122
- * The proposer parses each finding's `subject` via
123
- * `parseFindingSubject` (agent-eval), resolves it to a real file path
124
- * via the agent's `AgentSurfaces`, reads the current content, and asks
125
- * an LLM to draft a unified-diff patch given the finding + current
126
- * content + per-kind editing-discipline rules.
127
- *
128
- * Fail-loud rules:
129
- * - Findings whose subject doesn't parse → counted in `errors`.
130
- * - Findings whose subject targets an undeclared surface → counted in
131
- * `errors` with the offending kind in the message.
132
- * - Findings whose target path doesn't exist AND the kind isn't a
133
- * create-new variant (`new-tool`, `knowledge.wiki`) → counted in
134
- * `errors` with the resolved path in the message.
135
- * - LLM drafts that fail JSON-schema validation → counted in
136
- * `errors` with the schema issue.
137
- *
138
- * No silent skips. Every dropped finding has a recorded reason the
139
- * loop's report surfaces.
140
- */
141
-
142
- interface SurfaceImprovementEdit {
143
- /** Stable id derived from the source finding so re-proposals are idempotent. */
144
- id: string;
145
- /** The finding that produced this edit — for revert + audit trail. */
146
- sourceFindingId: string;
147
- /** Parsed subject; included so the apply step doesn't re-parse. */
148
- subject: FindingSubject;
149
- /** Resolved on-disk target. */
150
- target: ResolvedSurface;
151
- /** SHA-256 of the current file content the patch was drafted against. */
152
- baseSha256: string;
153
- /** Unified-diff patch the LLM drafted (relative to `target.absolutePath`). */
154
- patch: string;
155
- /** One-line summary the operator sees in the report / PR title. */
156
- summary: string;
157
- /** Multi-line rationale for the PR body — finding context + LLM reasoning. */
158
- rationale: string;
159
- /** Carry-forward from the finding so the apply gate can check the threshold. */
160
- confidence: number;
161
- /** Carry-forward severity for prioritization. */
162
- severity: AnalystFinding['severity'];
163
- }
164
- interface CreateSurfaceImprovementProposerOptions {
165
- surfaces: AgentSurfaces;
166
- repoRoot: string;
167
- /**
168
- * LLM-draft callback. Given a finding + current file content + the
169
- * resolved target, returns a unified-diff patch + summary + rationale.
170
- *
171
- * Required — the substrate doesn't ship a hardcoded prompt; the agent
172
- * author picks the model (Haiku for cheap routine drafts, Sonnet for
173
- * substantive prompt rewrites, etc.) via this callback.
174
- */
175
- draftPatch: (input: DraftPatchInput) => Promise<DraftPatchOutput>;
176
- /**
177
- * When the resolved target doesn't exist, allow the substrate to
178
- * CREATE the file (for `knowledge.wiki`, `new-tool` subjects). Default
179
- * true for those kinds, false for `system-prompt` / `rubric` / etc.
180
- * (named sections that don't exist are a contract violation, not a
181
- * scaffolding opportunity).
182
- */
183
- allowCreateForKinds?: ReadonlyArray<FindingSubject['kind']>;
184
- }
185
- interface DraftPatchInput {
186
- finding: AnalystFinding;
187
- subject: FindingSubject;
188
- target: ResolvedSurface;
189
- /** Current file content (empty string when `intent === 'create-new'`). */
190
- currentContent: string;
191
- }
192
- interface DraftPatchOutput {
193
- /** Unified diff against the current file content. Empty string skips this finding. */
194
- patch: string;
195
- /** One-line summary for the operator. */
196
- summary: string;
197
- /** Multi-line rationale for the PR body. */
198
- rationale: string;
199
- }
200
- /** Resolve each finding to a real surface and draft a detached patch candidate. */
201
- declare function createSurfaceImprovementProposer(opts: CreateSurfaceImprovementProposerOptions): ImprovementProposalSource<SurfaceImprovementEdit>;
202
-
203
- export { type AgentSurfaces as A, type CreateSurfaceImprovementProposerOptions as C, type DraftPatchInput as D, type ResolvedSurface as R, type SurfaceImprovementEdit as S, type DraftPatchOutput as a, type SurfaceValidationIssue as b, createSurfaceImprovementProposer as c, resolveSubjectPath as d, renderSurfaceIssues as r, validateSurfaces as v };
@@ -1 +0,0 @@
1
- {"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}
@@ -1,175 +0,0 @@
1
- import { ChildProcess } from 'node:child_process';
2
-
3
- /**
4
- *
5
- * Subprocess wrappers for the local coding-harness CLIs installed in the
6
- * sandbox image (claude-code, codex, opencode). Used by the in-process
7
- * delegation executor (`createInProcessExecutor`) so a delegated coding task
8
- * spawns a real harness on a real git worktree instead of provisioning a
9
- * sibling sandbox.
10
- *
11
- * All harness invocations:
12
- * - run with `cwd` set to the worktree
13
- * - inherit env from the parent (the MCP server inside the sandbox has
14
- * the harness's auth already)
15
- * - capture stdout/stderr
16
- * - support cancellation via AbortSignal
17
- * - enforce a wall-clock timeout
18
- *
19
- * @experimental
20
- */
21
-
22
- /** Local coding harness available inside the sandbox. */
23
- type LocalHarness = 'claude' | 'codex' | 'opencode';
24
- /** @experimental */
25
- interface RunLocalHarnessOptions {
26
- harness: LocalHarness;
27
- /** Working directory for the subprocess (typically a worktree path). */
28
- cwd: string;
29
- /** Prompt forwarded as the harness CLI's task argument. */
30
- taskPrompt: string;
31
- /**
32
- * Pre-built command + args (e.g. from `harnessInvocation` so the full authored
33
- * `AgentProfile` — systemPrompt + model — reaches the harness). When set it OVERRIDES the
34
- * default prompt-only `buildArgs(taskPrompt)` path; `command` defaults to the harness's
35
- * default binary when only `args` is supplied. When absent the legacy prompt-only shape
36
- * is used unchanged.
37
- */
38
- invocation?: {
39
- command?: string;
40
- args: ReadonlyArray<string>;
41
- };
42
- /** Allow autonomous Claude edits without an interactive permission prompt.
43
- * Use only when `cwd` is an isolated candidate worktree. */
44
- dangerouslySkipPermissions?: boolean;
45
- /** Isolate Codex from ambient configuration/instructions and require JSONL token usage.
46
- * The invocation should come from `harnessInvocation(..., { codexReproducible: true })`. */
47
- codexReproducible?: boolean;
48
- /** Absolute host paths that reproducible Codex must not read. The normalized set is compiled
49
- * into the controlled permission profile and its digest is returned in execution evidence. */
50
- codexReadDeniedPaths?: ReadonlyArray<string>;
51
- /** Wall-clock kill deadline (ms). Default 5 min. Subprocess SIGTERMed on expiry. */
52
- timeoutMs?: number;
53
- /** Newest stdout/stderr bytes retained per stream. Default 64 MiB. */
54
- maxOutputBytes?: number;
55
- /** Caller cancellation. SIGTERM is sent on abort. */
56
- signal?: AbortSignal;
57
- /** Override env (defaults to inheriting from the parent). */
58
- env?: NodeJS.ProcessEnv;
59
- /**
60
- * Test seam — inject a custom spawner so unit tests can mock the
61
- * subprocess without touching the OS. Defaults to node's `child_process.spawn`.
62
- */
63
- spawn?: (command: string, args: ReadonlyArray<string>, opts: {
64
- cwd: string;
65
- env: NodeJS.ProcessEnv;
66
- stdio: 'pipe';
67
- detached: boolean;
68
- }) => ChildProcess;
69
- /** Test seam for locating the native Codex executable before it is staged in the worktree. */
70
- resolveCodexExecutable?: (command: string, env: NodeJS.ProcessEnv) => Promise<string>;
71
- }
72
- /** Exact aggregate usage emitted by Codex's terminal `turn.completed` JSONL event. */
73
- interface CodexTokenUsage {
74
- inputTokens: number;
75
- cachedInputTokens: number;
76
- outputTokens: number;
77
- reasoningOutputTokens: number;
78
- }
79
- /** Isolation settings asserted before a reproducible Codex run is allowed to start. */
80
- interface CodexExecutionPolicy {
81
- sessionPersistence: 'ephemeral';
82
- userConfig: false;
83
- rules: false;
84
- projectInstructions: false;
85
- skillInstructions: false;
86
- appInstructions: false;
87
- toolSuggestions: false;
88
- multiAgentInstructions: false;
89
- sandbox: 'workspace-write';
90
- permissionProfile: 'agent_runtime_reproducible';
91
- approvalPolicy: 'never';
92
- shellNetwork: false;
93
- webSearch: false;
94
- serviceTier: 'default';
95
- shellEnvironment: 'core-filtered';
96
- loginShell: false;
97
- credentialsReadable: false;
98
- hostHomeReadable: false;
99
- procEnvironment: 'private-sanitized';
100
- sensitiveEnvironmentNamesVisible: false;
101
- parentRepoRead: false;
102
- gitMetadata: false;
103
- temporaryDirectory: 'workspace-private';
104
- stagedExecutable: 'static-elf-read-only';
105
- callerReadDeniedPaths: 'enforced';
106
- containerSockets: false;
107
- }
108
- /** Zero-model-call evidence for the exact Codex process about to run. */
109
- interface CodexExecutionEvidence {
110
- cliVersion: string;
111
- executableSha256: string;
112
- /** SHA-256 of the exact composed prompt argument proved present in the rendered prompt. */
113
- requestedPromptSha256: string;
114
- effectivePromptSha256: string;
115
- nonPromptArgsSha256: string;
116
- controlledConfigSha256: string;
117
- /** Sorted normalized paths compiled into the permission profile. */
118
- readDeniedPaths: string[];
119
- readDeniedPathsSha256: string;
120
- readDeniedPathCount: number;
121
- policy: CodexExecutionPolicy;
122
- }
123
- /** @experimental */
124
- interface LocalHarnessResult {
125
- /** OS exit code. `null` when killed before exit. */
126
- exitCode: number | null;
127
- /** Concatenated stdout. */
128
- stdout: string;
129
- /** Concatenated stderr. */
130
- stderr: string;
131
- /** Set when the process exited via signal (timeout / abort). */
132
- killedBySignal: NodeJS.Signals | null;
133
- /** Wall-clock duration ms (spawn → exit). */
134
- durationMs: number;
135
- /** Set when timeoutMs elapsed before exit. */
136
- timedOut: boolean;
137
- /**
138
- * Set when the caller's AbortSignal fired before this result settled.
139
- * Optional so injected runners and stored results from older releases remain valid.
140
- */
141
- aborted?: boolean;
142
- /** Present for a reproducible Codex run; parsed from the real terminal JSONL event. */
143
- usage?: CodexTokenUsage;
144
- /** Present for reproducible Codex runs; generated and checked before model execution. */
145
- evidence?: CodexExecutionEvidence;
146
- }
147
- /**
148
- * Spawn a local coding harness CLI as a subprocess + collect its output.
149
- *
150
- * NOT responsible for parsing the harness's output or extracting a diff —
151
- * the in-process executor's `streamPrompt` orchestrates `git diff` against
152
- * the worktree after this resolves. This function is intentionally narrow:
153
- * spawn, wait, capture, return.
154
- *
155
- * Fails loud — throws when:
156
- * - `cwd` doesn't exist (subprocess emits ENOENT; surfaced as Error)
157
- * - the harness binary is not on PATH (ENOENT)
158
- * - the caller signal was already aborted before process launch
159
- *
160
- * Does NOT throw when:
161
- * - the subprocess exits non-zero (`result.exitCode` carries the code)
162
- * - a non-reproducible subprocess is aborted / timed out (`result.aborted` /
163
- * `result.timedOut` carries the reason even when a TERM-aware child exits zero)
164
- *
165
- * Reproducible Codex additionally requires a terminal usage event. If cancellation
166
- * prevents that event, this rejects with `CodexExecutionDiagnosticError` instead of
167
- * returning an incomplete reproducibility receipt.
168
- *
169
- * @experimental
170
- */
171
- declare function runLocalHarness(options: RunLocalHarnessOptions): Promise<LocalHarnessResult>;
172
- /** Parse and validate the one terminal usage event emitted by `codex exec --json`. */
173
- declare function parseCodexTokenUsage(stdout: string): CodexTokenUsage;
174
-
175
- export { type CodexTokenUsage as C, type LocalHarness as L, type RunLocalHarnessOptions as R, type CodexExecutionPolicy as a, type CodexExecutionEvidence as b, type LocalHarnessResult as c, parseCodexTokenUsage as p, runLocalHarness as r };
@@ -1,199 +0,0 @@
1
- import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-CmYCMbFT.js';
2
- import { F as FactCandidate, C as CreateKbGateOptions } from './kb-gate-CwHO0vz6.js';
3
- import { B as Budget } from './types-CvNwMzQt.js';
4
- import { A as AuthoredHarness, W as WinnerStrategy, a as WorktreeFanoutOptions, b as WorktreePatchArtifact } from './worktree-fanout-DxhAWr5Z.js';
5
-
6
- /**
7
- *
8
- * `runDelegatedLoop` — the configured delegated loop-runner.
9
- *
10
- * One typed entrypoint a worker agent (or a scheduled routine) calls to run a
11
- * disciplined loop in a chosen MODE, over agent-runtime's hardened engines:
12
- *
13
- * code → build-in-a-loop on the GENERIC recursive path (worktreeLoopRunner: author one
14
- * `AgentProfile` per harness → worktree-CLI leaves → `patchDelivered` gate)
15
- * review → caller-registered runner — a `code` runner with an approval gate over candidates
16
- * research → research-in-a-loop with valid-only KB growth (createKbGate)
17
- * audit → analyze trace/run data → findings (runAnalystLoop, caller-wired)
18
- * self-improve → caller-registered `improve(profile, options)` run
19
- *
20
- * It is intentionally a thin façade: the value is that EVERY product reuses the
21
- * one hardened engine instead of forking delegation logic. The dispatcher owns
22
- * mode routing, timing, fail-loud on an unregistered mode, and a uniform result
23
- * shape; each mode's engine is a pre-configured runner in the registry (build it
24
- * with the factories below, or inject your own / a stub).
25
- *
26
- * @experimental
27
- */
28
-
29
- /** All valid delegated-loop mode names — used for validation and CLI surfaces. @experimental */
30
- declare const DELEGATED_LOOP_MODES: readonly ["code", "review", "research", "audit", "self-improve"];
31
- /** @experimental */
32
- type DelegatedLoopMode = (typeof DELEGATED_LOOP_MODES)[number];
33
- /** Type guard — returns true when `value` is a valid `DelegatedLoopMode` string. @experimental */
34
- declare function isDelegatedLoopMode(value: unknown): value is DelegatedLoopMode;
35
- /** @experimental A pre-configured loop for one mode. Returns the mode's raw
36
- * output; the dispatcher wraps it in a {@link DelegatedLoopResult}. */
37
- type DelegatedLoopRunner<T = unknown> = (signal: AbortSignal) => Promise<T>;
38
- /** @experimental Mode → configured runner. Partial: only register the modes a
39
- * given product/routine actually uses. */
40
- type DelegatedLoopRegistry = Partial<Record<DelegatedLoopMode, DelegatedLoopRunner>>;
41
- /** @experimental Uniform result — never throws from a registered runner; a
42
- * thrown engine becomes `{ ok: false, error }` so a routine can record + move on. */
43
- interface DelegatedLoopResult<T = unknown> {
44
- mode: DelegatedLoopMode;
45
- ok: boolean;
46
- output?: T;
47
- error?: string;
48
- durationMs: number;
49
- }
50
- /** @experimental */
51
- interface RunDelegatedLoopOptions {
52
- signal?: AbortSignal;
53
- /** Clock override for deterministic tests. */
54
- now?: () => number;
55
- }
56
- /**
57
- *
58
- * Dispatch a configured loop by mode. Fails loud (throws `ConfigError`) when no
59
- * runner is registered for the mode — a routine pointed at an unwired mode is a
60
- * config bug, not a silent no-op. A runner that throws is captured as
61
- * `{ ok: false }` so unattended runs record the failure rather than crash.
62
- *
63
- * @experimental
64
- */
65
- declare function runDelegatedLoop<T = unknown>(mode: DelegatedLoopMode, registry: DelegatedLoopRegistry, options?: RunDelegatedLoopOptions): Promise<DelegatedLoopResult<T>>;
66
- /** @experimental Options for the local-repo `code` runner over the GENERIC recursive path. */
67
- interface WorktreeLoopRunnerOptions {
68
- /** Absolute path to the local git checkout each worktree is cut from. */
69
- repoRoot: string;
70
- /** The instruction handed to every authored harness (composed under each profile's systemPrompt). */
71
- taskPrompt: string;
72
- /** The supervisor-authored harness profiles — one fanout item (one worktree-CLI leaf) each. */
73
- harnesses: ReadonlyArray<AuthoredHarness>;
74
- /** Conserved budget pool bounding the fanout (equal-k holds by construction). */
75
- budget: Budget;
76
- /** Shell command run in each worktree to derive the tests-PASS signal. */
77
- testCmd?: string;
78
- /** Shell command run in each worktree to derive the typecheck-PASS signal. */
79
- typecheckCmd?: string;
80
- /** Which verification signals the deliverable REQUIRES present-and-passing (default none). */
81
- require?: ReadonlyArray<'tests' | 'typecheck'>;
82
- /** Diff-size cap (lines). */
83
- maxDiffLines?: number;
84
- /** Literal path prefixes the patch must not touch (the secret-floor is always on regardless). */
85
- forbiddenPaths?: string[];
86
- /** Winner-selection strategy among gated candidates. Default `highest-score`. */
87
- winnerStrategy?: WinnerStrategy;
88
- /** Test seams forwarded to the worktree-CLI leaves so the runner drives offline. */
89
- runGit?: WorktreeFanoutOptions['runGit'];
90
- runHarness?: WorktreeFanoutOptions['runHarness'];
91
- runCommand?: WorktreeFanoutOptions['runCommand'];
92
- }
93
- /**
94
- *
95
- * `code` mode on the GENERIC recursive path: author one `AgentProfile` per harness, run them as a
96
- * `worktreeFanout` (N `createWorktreeCliExecutor` leaves, each `gateOnDeliverable`) through
97
- * `runPersonified` on the keystone Supervisor. The sandbox-session counterpart that drives the in-box
98
- * harness over a `SandboxClient` is `detachedSessionDelegate` (`./mcp/delegates`); here there is no
99
- * `runAgentRounds` driver, no role-coupled delegate — the harness list is the fanout, the gate is
100
- * `patchDelivered`,
101
- * the winner is the shared valid-only selector (NOT `defaultSelectWinner`, whose non-valid fallback
102
- * would surface an ungated patch). Equal-k holds by the conserved budget pool. Returns the winning
103
- * patch artifact, or throws when no candidate is delivered (fail loud, never a vacuous done).
104
- *
105
- * @experimental
106
- */
107
- declare function worktreeLoopRunner(options: WorktreeLoopRunnerOptions): DelegatedLoopRunner<WorktreePatchArtifact>;
108
- /** @experimental A fact rejected at the KB gate — surfaced, never dropped. */
109
- interface VetoedFact {
110
- candidate: FactCandidate;
111
- vetoedBy?: string;
112
- reason?: string;
113
- }
114
- /** @experimental */
115
- interface ResearchLoopResult {
116
- /** Facts that passed the fail-closed gate — safe to write to the KB. */
117
- accepted: FactCandidate[];
118
- /** Facts the gate vetoed in the final round — escalate, do not silently drop. */
119
- vetoed: VetoedFact[];
120
- /** Research rounds actually run. */
121
- rounds: number;
122
- }
123
- /** @experimental Options for the default `research` runner. */
124
- interface ResearchLoopRunnerOptions {
125
- /**
126
- * The research engine (the consumer's web/doc searcher + extractor). Called
127
- * each round with the prior round's vetoes so it can re-research the gaps.
128
- * Returns fact candidates carrying their grounding (`verbatimPassage` +
129
- * `sourceText`).
130
- */
131
- research: (round: number, vetoed: VetoedFact[]) => Promise<FactCandidate[]>;
132
- /** Gate config (extra judges, self-artifact kinds, …). The floor is always on. */
133
- gate?: CreateKbGateOptions;
134
- /** Max research rounds (correct-on-veto remediation). Default 1. */
135
- maxRounds?: number;
136
- }
137
- /**
138
- * `research` mode — research-in-a-loop with valid-only KB growth.
139
- *
140
- * Each round: research → gate every candidate (fail-closed; passage MUST be in
141
- * the source) → accept the clean ones → re-research the vetoed ones next round,
142
- * up to `maxRounds`. Vetoed facts in the final round are RETURNED (escalate,
143
- * never silently dropped) so the caller audits vs retries.
144
- *
145
- * @experimental
146
- */
147
- declare function researchLoopRunner(o: ResearchLoopRunnerOptions): DelegatedLoopRunner<ResearchLoopResult>;
148
- /**
149
- * `audit` mode — analyst loop over captured trace/run data.
150
- *
151
- * @experimental
152
- */
153
- declare function auditLoopRunner<TProposal = unknown, TEdit = unknown>(options: RunAnalystLoopOpts): DelegatedLoopRunner<RunAnalystLoopResult<TProposal, TEdit>>;
154
-
155
- /**
156
- *
157
- * `agent-runtime-loop` — the schedulable entrypoint for the configured
158
- * delegated loop-runner. A cron job / routine / Makefile target invokes:
159
- *
160
- * agent-runtime-loop --mode research --config ./loops.config.js
161
- *
162
- * The config module wires the registry (with full access to env / creds —
163
- * which is why the deps live there, not in this generic bin). It must default-
164
- * export a `DelegatedLoopRegistry`, or a `() => DelegatedLoopRegistry | Promise<…>`.
165
- * The bin runs the selected mode, prints the `DelegatedLoopResult` as JSON, and
166
- * exits 0 on `ok`, 1 on a recorded failure, 2 on a usage/config error.
167
- *
168
- * @experimental
169
- */
170
-
171
- /** @experimental Parsed CLI invocation. */
172
- interface LoopRunnerCliArgs {
173
- mode: string;
174
- /** Loads the registry — the bin wires this from `--config`; tests inject a stub. */
175
- loadRegistry: () => Promise<DelegatedLoopRegistry> | DelegatedLoopRegistry;
176
- now?: () => number;
177
- }
178
- /** @experimental */
179
- interface LoopRunnerCliResult {
180
- exitCode: number;
181
- result?: DelegatedLoopResult;
182
- error?: string;
183
- }
184
- /**
185
- *
186
- * Pure CLI core (no process / argv / IO) so it's unit-testable: validate the
187
- * mode, load the registry, dispatch, map to an exit code (0 ok / 1 failed /
188
- * 2 usage). Exported for embedding in custom runners + tests.
189
- *
190
- * @experimental
191
- */
192
- declare function runLoopRunnerCli(args: LoopRunnerCliArgs): Promise<LoopRunnerCliResult>;
193
- /** Parse `--mode X --config Y` from an argv tail (`process.argv.slice(2)`). */
194
- declare function parseLoopRunnerArgv(argv: string[]): {
195
- mode?: string;
196
- config?: string;
197
- };
198
-
199
- export { DELEGATED_LOOP_MODES as D, type LoopRunnerCliArgs as L, type ResearchLoopResult as R, type VetoedFact as V, type WorktreeLoopRunnerOptions as W, type DelegatedLoopMode as a, type DelegatedLoopRegistry as b, type DelegatedLoopResult as c, type DelegatedLoopRunner as d, type LoopRunnerCliResult as e, type ResearchLoopRunnerOptions as f, type RunDelegatedLoopOptions as g, auditLoopRunner as h, isDelegatedLoopMode as i, runDelegatedLoop as j, runLoopRunnerCli as k, parseLoopRunnerArgv as p, researchLoopRunner as r, worktreeLoopRunner as w };
@@ -1 +0,0 @@
1
- {"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}
package/dist/loops.js.map DELETED
@@ -1 +0,0 @@
1
- {"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}
@@ -1,46 +0,0 @@
1
- import { O as OpenAIChatTool } from './types-BwoZWq-i.js';
2
-
3
- /**
4
- *
5
- * OpenAI Chat Completions `tools[]` projection of the queue-bound agent-runtime
6
- * MCP delegation tools.
7
- *
8
- * Use when configuring `createOpenAICompatibleBackend({ tools: ... })` so the
9
- * model can call `delegate_feedback`, `delegation_status`, and
10
- * `delegation_history` through the OpenAI-compat transport (tcloud, OpenRouter,
11
- * OpenAI direct, cli-bridge). The runtime surfaces tool calls as `tool_call`
12
- * stream events — execution is the caller's responsibility (typically the
13
- * parent sandbox runtime's MCP mount).
14
- *
15
- * Sandbox-SDK callers do NOT need this helper: the sandbox runtime mounts
16
- * MCP servers natively and the in-sandbox harness discovers tools via the
17
- * runtime, not via an OpenAI tools array.
18
- *
19
- * Tool name + description + JSON-schema are pulled from the canonical
20
- * `DELEGATE_*` constants exported by `./tools/*` so the projection cannot
21
- * drift from the server's own validators.
22
- *
23
- * @experimental
24
- */
25
-
26
- /**
27
- *
28
- * Returns the queue-bound delegation tools projected into OpenAI Chat
29
- * Completions `tools[]` shape. The order is stable: `delegate_feedback`,
30
- * `delegation_status`, `delegation_history`.
31
- *
32
- * @experimental
33
- */
34
- declare function mcpToolsForRuntimeMcp(): OpenAIChatTool[];
35
- /**
36
- *
37
- * Subset filter — return only the projected tools whose `function.name`
38
- * appears in `names`. Useful for curated mounts (e.g. only the queue-bound
39
- * delegation tools, omitting `delegate_feedback`). Unknown names are
40
- * silently ignored; pass an empty array to get an empty result.
41
- *
42
- * @experimental
43
- */
44
- declare function mcpToolsForRuntimeMcpSubset(names: ReadonlyArray<string>): OpenAIChatTool[];
45
-
46
- export { mcpToolsForRuntimeMcpSubset as a, mcpToolsForRuntimeMcp as m };