@pugi/cli 0.1.0-beta.4 → 0.1.0-beta.41

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 (250) hide show
  1. package/THIRD_PARTY_NOTICES.md +40 -0
  2. package/assets/pugi-mascot.ansi +15 -25
  3. package/bin/run.js +33 -1
  4. package/dist/commands/jobs-watch.js +201 -0
  5. package/dist/commands/jobs.js +15 -0
  6. package/dist/commands/smoke.js +133 -0
  7. package/dist/core/agent-progress/cleanup.js +134 -0
  8. package/dist/core/agent-progress/schema.js +144 -0
  9. package/dist/core/agent-progress/writer.js +101 -0
  10. package/dist/core/artifact-chain/dispatcher.js +148 -0
  11. package/dist/core/artifact-chain/exporter.js +164 -0
  12. package/dist/core/artifact-chain/state.js +243 -0
  13. package/dist/core/artifact-chain/steps.js +169 -0
  14. package/dist/core/auth/ensure-authenticated.js +129 -0
  15. package/dist/core/auth/env-provider.js +238 -0
  16. package/dist/core/auto-update/channels.js +122 -0
  17. package/dist/core/auto-update/checker.js +241 -0
  18. package/dist/core/auto-update/state.js +235 -0
  19. package/dist/core/bare-mode/index.js +107 -0
  20. package/dist/core/bash-classifier.js +108 -1
  21. package/dist/core/checkpoint/resumer.js +149 -0
  22. package/dist/core/checkpoint/rewinder.js +291 -0
  23. package/dist/core/codegraph/decision-store.js +248 -0
  24. package/dist/core/codegraph/detect-repo.js +459 -0
  25. package/dist/core/codegraph/install.js +134 -0
  26. package/dist/core/codegraph/offer-hook.js +220 -0
  27. package/dist/core/compact/auto-trigger.js +96 -0
  28. package/dist/core/compact/buffer-rewriter.js +115 -0
  29. package/dist/core/compact/summarizer.js +208 -0
  30. package/dist/core/compact/token-counter.js +108 -0
  31. package/dist/core/consensus/diff-capture.js +73 -0
  32. package/dist/core/context/index.js +7 -0
  33. package/dist/core/context/markdown-traverse.js +255 -0
  34. package/dist/core/cost/rate-card.js +129 -0
  35. package/dist/core/cost/tracker.js +221 -0
  36. package/dist/core/denial-tracking/index.js +8 -0
  37. package/dist/core/denial-tracking/state.js +264 -0
  38. package/dist/core/diagnostics/probe-runner.js +93 -0
  39. package/dist/core/diagnostics/probes/api.js +46 -0
  40. package/dist/core/diagnostics/probes/auth.js +86 -0
  41. package/dist/core/diagnostics/probes/bare-mode.js +42 -0
  42. package/dist/core/diagnostics/probes/cli-version.js +127 -0
  43. package/dist/core/diagnostics/probes/config.js +72 -0
  44. package/dist/core/diagnostics/probes/denial-tracking.js +57 -0
  45. package/dist/core/diagnostics/probes/disk.js +81 -0
  46. package/dist/core/diagnostics/probes/git.js +65 -0
  47. package/dist/core/diagnostics/probes/mcp.js +75 -0
  48. package/dist/core/diagnostics/probes/node.js +59 -0
  49. package/dist/core/diagnostics/probes/pnpm.js +36 -0
  50. package/dist/core/diagnostics/probes/pugi-md.js +89 -0
  51. package/dist/core/diagnostics/probes/session.js +74 -0
  52. package/dist/core/diagnostics/probes/status-snapshot.js +488 -0
  53. package/dist/core/diagnostics/probes/workspace.js +63 -0
  54. package/dist/core/diagnostics/types.js +70 -0
  55. package/dist/core/dispatch/cache-cleanup.js +197 -0
  56. package/dist/core/dispatch/cache-handoff.js +295 -0
  57. package/dist/core/edits/dispatch.js +218 -2
  58. package/dist/core/edits/journal.js +199 -0
  59. package/dist/core/edits/layer-d-ast.js +557 -14
  60. package/dist/core/edits/verify-hook.js +273 -0
  61. package/dist/core/edits/worktree.js +322 -0
  62. package/dist/core/engine/anvil-client.js +115 -5
  63. package/dist/core/engine/budgets.js +98 -0
  64. package/dist/core/engine/context-prefix.js +155 -0
  65. package/dist/core/engine/intent.js +260 -0
  66. package/dist/core/engine/native-pugi.js +860 -211
  67. package/dist/core/engine/prompts.js +88 -2
  68. package/dist/core/engine/strip-internal-fields.js +124 -0
  69. package/dist/core/engine/tool-bridge.js +1045 -36
  70. package/dist/core/feedback/queue.js +177 -0
  71. package/dist/core/feedback/submitter.js +145 -0
  72. package/dist/core/file-cache.js +113 -1
  73. package/dist/core/hooks/events.js +44 -0
  74. package/dist/core/hooks/index.js +15 -0
  75. package/dist/core/hooks/registry.js +213 -0
  76. package/dist/core/hooks/runner.js +236 -0
  77. package/dist/core/hooks/v2/event-emitter.js +115 -0
  78. package/dist/core/hooks/v2/executor.js +282 -0
  79. package/dist/core/hooks/v2/index.js +25 -0
  80. package/dist/core/hooks/v2/lifecycle.js +104 -0
  81. package/dist/core/hooks/v2/loader.js +216 -0
  82. package/dist/core/hooks/v2/matcher.js +125 -0
  83. package/dist/core/hooks/v2/trust.js +143 -0
  84. package/dist/core/hooks/v2/types.js +86 -0
  85. package/dist/core/lsp/cache.js +105 -0
  86. package/dist/core/lsp/client.js +776 -0
  87. package/dist/core/lsp/language-detect.js +66 -0
  88. package/dist/core/lsp/post-edit-diagnostics.js +171 -0
  89. package/dist/core/mcp/client.js +75 -6
  90. package/dist/core/mcp/http-server.js +553 -0
  91. package/dist/core/mcp/orchestrator-tools.js +662 -0
  92. package/dist/core/mcp/permission.js +190 -0
  93. package/dist/core/mcp/registry.js +24 -2
  94. package/dist/core/mcp/server-tools.js +219 -0
  95. package/dist/core/mcp/server.js +397 -0
  96. package/dist/core/memory/dual-write.js +416 -0
  97. package/dist/core/memory/phase1-kinds.js +20 -0
  98. package/dist/core/memory-sync/queue.js +158 -0
  99. package/dist/core/onboarding/ensure-initialized.js +133 -0
  100. package/dist/core/onboarding/marker.js +111 -0
  101. package/dist/core/onboarding/telemetry-state.js +108 -0
  102. package/dist/core/output-style/presets.js +176 -0
  103. package/dist/core/output-style/state.js +185 -0
  104. package/dist/core/permissions/auto-classifier.js +124 -0
  105. package/dist/core/permissions/circuit-breaker.js +83 -0
  106. package/dist/core/permissions/gate.js +278 -0
  107. package/dist/core/permissions/index.js +20 -0
  108. package/dist/core/permissions/mode.js +174 -0
  109. package/dist/core/permissions/state.js +241 -0
  110. package/dist/core/permissions/tool-class.js +93 -0
  111. package/dist/core/prd-check/parser.js +215 -0
  112. package/dist/core/prd-check/reporter.js +127 -0
  113. package/dist/core/prd-check/session-review.js +557 -0
  114. package/dist/core/prd-check/verifiers.js +223 -0
  115. package/dist/core/pugi-md/context-injector.js +76 -0
  116. package/dist/core/pugi-md/walk-up.js +207 -0
  117. package/dist/core/release-notes/parser.js +241 -0
  118. package/dist/core/release-notes/state.js +116 -0
  119. package/dist/core/repl/history.js +11 -1
  120. package/dist/core/repl/model-pricing.js +135 -0
  121. package/dist/core/repl/session.js +1899 -38
  122. package/dist/core/repl/slash-commands.js +406 -21
  123. package/dist/core/repl/store/session-store.js +31 -2
  124. package/dist/core/repl/workspace-context.js +22 -0
  125. package/dist/core/repo-map/build.js +125 -0
  126. package/dist/core/repo-map/cache.js +185 -0
  127. package/dist/core/repo-map/extractor.js +254 -0
  128. package/dist/core/repo-map/formatter.js +145 -0
  129. package/dist/core/repo-map/scanner.js +211 -0
  130. package/dist/core/retry-budget/budget.js +284 -0
  131. package/dist/core/retry-budget/index.js +5 -0
  132. package/dist/core/session.js +92 -0
  133. package/dist/core/settings.js +80 -0
  134. package/dist/core/share/formatter.js +271 -0
  135. package/dist/core/share/redactor.js +221 -0
  136. package/dist/core/share/uploader.js +267 -0
  137. package/dist/core/skills/defaults.js +457 -0
  138. package/dist/core/smoke/headless-driver.js +174 -0
  139. package/dist/core/smoke/orchestrator.js +194 -0
  140. package/dist/core/smoke/runner.js +238 -0
  141. package/dist/core/smoke/scenario-parser.js +316 -0
  142. package/dist/core/subagents/dispatcher-real.js +600 -0
  143. package/dist/core/subagents/dispatcher.js +113 -24
  144. package/dist/core/subagents/index.js +18 -5
  145. package/dist/core/subagents/isolation-matrix.js +213 -0
  146. package/dist/core/subagents/spawn.js +19 -4
  147. package/dist/core/telemetry/emitter.js +229 -0
  148. package/dist/core/telemetry/queue.js +251 -0
  149. package/dist/core/theme/context.js +91 -0
  150. package/dist/core/theme/presets.js +228 -0
  151. package/dist/core/theme/state.js +181 -0
  152. package/dist/core/todos/invariant.js +10 -0
  153. package/dist/core/todos/state.js +177 -0
  154. package/dist/core/transport/version-interceptor.js +166 -0
  155. package/dist/core/vim/keymap.js +288 -0
  156. package/dist/core/vim/state.js +92 -0
  157. package/dist/index.js +28 -0
  158. package/dist/runtime/bootstrap.js +190 -0
  159. package/dist/runtime/cli.js +3073 -321
  160. package/dist/runtime/commands/cancel.js +231 -0
  161. package/dist/runtime/commands/chain.js +489 -0
  162. package/dist/runtime/commands/codegraph-status.js +227 -0
  163. package/dist/runtime/commands/compact.js +297 -0
  164. package/dist/runtime/commands/cost.js +199 -0
  165. package/dist/runtime/commands/delegate.js +242 -11
  166. package/dist/runtime/commands/dispatch.js +126 -0
  167. package/dist/runtime/commands/doctor.js +390 -0
  168. package/dist/runtime/commands/feedback.js +184 -0
  169. package/dist/runtime/commands/hooks.js +184 -0
  170. package/dist/runtime/commands/lsp.js +368 -0
  171. package/dist/runtime/commands/mcp.js +879 -0
  172. package/dist/runtime/commands/memory.js +508 -0
  173. package/dist/runtime/commands/model.js +237 -0
  174. package/dist/runtime/commands/onboarding.js +275 -0
  175. package/dist/runtime/commands/patch.js +128 -0
  176. package/dist/runtime/commands/permissions.js +112 -0
  177. package/dist/runtime/commands/plan.js +143 -0
  178. package/dist/runtime/commands/prd-check.js +285 -0
  179. package/dist/runtime/commands/redo-blob-store.js +92 -0
  180. package/dist/runtime/commands/redo.js +361 -0
  181. package/dist/runtime/commands/release-notes.js +229 -0
  182. package/dist/runtime/commands/repo-map.js +95 -0
  183. package/dist/runtime/commands/report.js +299 -0
  184. package/dist/runtime/commands/resume.js +118 -0
  185. package/dist/runtime/commands/review-consensus.js +17 -2
  186. package/dist/runtime/commands/rewind.js +333 -0
  187. package/dist/runtime/commands/sessions.js +163 -0
  188. package/dist/runtime/commands/share.js +316 -0
  189. package/dist/runtime/commands/status.js +186 -0
  190. package/dist/runtime/commands/stickers.js +82 -0
  191. package/dist/runtime/commands/style.js +194 -0
  192. package/dist/runtime/commands/theme.js +196 -0
  193. package/dist/runtime/commands/undo.js +32 -0
  194. package/dist/runtime/commands/update.js +289 -0
  195. package/dist/runtime/commands/vim.js +140 -0
  196. package/dist/runtime/commands/worktree.js +177 -0
  197. package/dist/runtime/headless-repl.js +195 -0
  198. package/dist/runtime/headless.js +543 -0
  199. package/dist/runtime/load-hooks-or-exit.js +71 -0
  200. package/dist/runtime/plan-decompose.js +531 -0
  201. package/dist/runtime/version.js +65 -0
  202. package/dist/tools/agent-tool.js +229 -0
  203. package/dist/tools/apply-patch.js +556 -0
  204. package/dist/tools/ask-user-question.js +213 -0
  205. package/dist/tools/ask-user.js +115 -0
  206. package/dist/tools/file-tools.js +85 -14
  207. package/dist/tools/lsp-tools.js +189 -0
  208. package/dist/tools/mcp-tool.js +260 -0
  209. package/dist/tools/multi-edit.js +361 -0
  210. package/dist/tools/powershell.js +156 -0
  211. package/dist/tools/registry.js +51 -0
  212. package/dist/tools/skill-tool.js +96 -0
  213. package/dist/tools/tasks.js +208 -0
  214. package/dist/tools/todo-write.js +184 -0
  215. package/dist/tools/web-fetch.js +147 -2
  216. package/dist/tools/web-search.js +458 -0
  217. package/dist/tui/agent-progress-card.js +111 -0
  218. package/dist/tui/agent-tree.js +10 -0
  219. package/dist/tui/ask-modal.js +2 -2
  220. package/dist/tui/ask-user-question-prompt.js +192 -0
  221. package/dist/tui/compact-banner.js +81 -0
  222. package/dist/tui/conversation-pane.js +82 -8
  223. package/dist/tui/cost-table.js +111 -0
  224. package/dist/tui/doctor-table.js +46 -0
  225. package/dist/tui/feedback-prompt.js +156 -0
  226. package/dist/tui/input-box.js +69 -2
  227. package/dist/tui/markdown-render.js +4 -4
  228. package/dist/tui/onboarding-wizard.js +240 -0
  229. package/dist/tui/permissions-picker.js +86 -0
  230. package/dist/tui/render.js +35 -0
  231. package/dist/tui/repl-render.js +303 -13
  232. package/dist/tui/repl-splash.js +2 -2
  233. package/dist/tui/repl.js +72 -14
  234. package/dist/tui/splash.js +1 -1
  235. package/dist/tui/status-bar.js +94 -16
  236. package/dist/tui/status-table.js +7 -0
  237. package/dist/tui/stickers-art.js +136 -0
  238. package/dist/tui/style-table.js +28 -0
  239. package/dist/tui/theme-table.js +29 -0
  240. package/dist/tui/tool-stream-pane.js +52 -3
  241. package/dist/tui/update-banner.js +20 -2
  242. package/dist/tui/vim-input.js +267 -0
  243. package/docs/examples/codegraph.mcp.json +10 -0
  244. package/package.json +12 -6
  245. package/test/scenarios/codegen-create-file.scenario.txt +13 -0
  246. package/test/scenarios/compact-force.scenario.txt +11 -0
  247. package/test/scenarios/identity.scenario.txt +11 -0
  248. package/test/scenarios/persona-handoff.scenario.txt +11 -0
  249. package/test/scenarios/walkback.scenario.txt +12 -0
  250. package/dist/core/engine/compaction-hook.js +0 -154
@@ -0,0 +1,260 @@
1
+ import { callTool } from '../core/mcp/client.js';
2
+ import { getMcpPermission, setMcpPermission, } from '../core/mcp/permission.js';
3
+ /**
4
+ * Tool dispatcher for MCP-invoked tools (β4 M1 + M3 + M5).
5
+ *
6
+ * Tool names use the `mcp__<server>__<tool>` namespace (double-underscore
7
+ * separator, mirroring Claude Code's MCP envelope). The triple-underscore
8
+ * forms (`mcp__server__tool__sub`) collapse into the third segment when
9
+ * the upstream server itself uses underscores in its tool names — `split`
10
+ * on the first two `__` only, so any further `__` in the tool name part
11
+ * survive intact (e.g. `mcp__github__create_issue` -> server=`github`,
12
+ * tool=`create_issue`).
13
+ *
14
+ * Why double-underscore: native Pugi tools use single-token names
15
+ * (`read`, `grep`, `edit`, `bash`). The double-underscore prefix
16
+ * unambiguously segregates the MCP namespace from native names without
17
+ * needing per-name regex matching at every dispatch site.
18
+ *
19
+ * Permission flow:
20
+ * 1. Server trust gate (handled at registry-load time). If a server is
21
+ * not `trusted`, its tools never reach the engine loop.
22
+ * 2. Per-(server, tool) permission cache (`./mcp/permission.ts`).
23
+ * Unset on first dispatch -> caller must prompt. Cached `allow_always`
24
+ * auto-passes; cached `deny` auto-refuses.
25
+ *
26
+ * This module is the bridge — it parses the namespaced name, finds the
27
+ * live connection in the registry, consults the cache, and (when
28
+ * approved) routes through `client.callTool`. Prompting is the executor's
29
+ * responsibility; this module exposes the cache lookup + dispatch
30
+ * primitives so the executor stays small.
31
+ */
32
+ /**
33
+ * Prefix every MCP tool name carries on the engine-loop wire.
34
+ */
35
+ export const MCP_TOOL_PREFIX = 'mcp__';
36
+ /**
37
+ * Parse `mcp__<server>__<tool>` into `{ serverName, toolName }`. Returns
38
+ * null when the input does not match the namespace — callers use this as
39
+ * the "is this an MCP tool?" predicate.
40
+ *
41
+ * Server names cannot contain `__` by registry validation (they are JSON
42
+ * object keys); tool names CAN (e.g. `create_issue` has a single `_` but
43
+ * `read_directory` has none, so the only ambiguity is when an upstream
44
+ * tool uses double-underscore in its slug — extremely rare, but if it
45
+ * happens the second `__` boundary still parses correctly because we
46
+ * split on the FIRST occurrence after the prefix).
47
+ */
48
+ export function parseMcpToolName(name) {
49
+ if (!name.startsWith(MCP_TOOL_PREFIX))
50
+ return null;
51
+ const tail = name.slice(MCP_TOOL_PREFIX.length);
52
+ const sep = tail.indexOf('__');
53
+ if (sep === -1)
54
+ return null;
55
+ const serverName = tail.slice(0, sep);
56
+ const toolName = tail.slice(sep + 2);
57
+ if (serverName.length === 0 || toolName.length === 0)
58
+ return null;
59
+ return { serverName, toolName };
60
+ }
61
+ /**
62
+ * Build the namespaced tool name from a server + tool pair. Inverse of
63
+ * `parseMcpToolName`. Used by `buildMcpToolDefs` to emit the schema.
64
+ */
65
+ export function buildMcpToolName(serverName, toolName) {
66
+ return `${MCP_TOOL_PREFIX}${serverName}__${toolName}`;
67
+ }
68
+ /**
69
+ * Build engine-loop tool definitions from every trusted server's
70
+ * surfaced tools. Empty array when no MCP servers are trusted — the
71
+ * schema builder can call this unconditionally without checking first.
72
+ */
73
+ export function buildMcpToolDefs(registry) {
74
+ if (!registry)
75
+ return [];
76
+ const defs = [];
77
+ for (const state of registry.servers.values()) {
78
+ if (state.trust !== 'trusted')
79
+ continue;
80
+ for (const tool of state.surfacedTools) {
81
+ defs.push({
82
+ name: buildMcpToolName(state.name, tool.name),
83
+ description: descriptionFor(state.name, tool),
84
+ // The upstream server returns its own JSON Schema in `inputSchema`.
85
+ // We surface it verbatim — the loop client passes it through to
86
+ // the model, and the model emits arguments matching the upstream
87
+ // shape. Default to `{ type: 'object' }` when missing so the
88
+ // OpenAI-shaped tool envelope still validates.
89
+ parameters: tool.inputSchema ?? { type: 'object' },
90
+ });
91
+ }
92
+ }
93
+ // Sort stable so the schema bundle hash (used for caching/audit) is
94
+ // deterministic regardless of Map iteration order.
95
+ return defs.sort((a, b) => a.name.localeCompare(b.name));
96
+ }
97
+ function descriptionFor(serverName, tool) {
98
+ const base = tool.description?.trim() ?? `MCP tool ${tool.name} on server ${serverName}.`;
99
+ return `[MCP:${serverName}] ${base}`;
100
+ }
101
+ /**
102
+ * Look up the live connection + tool metadata for a parsed MCP tool name.
103
+ * Returns null when the server is not trusted, not connected, or does
104
+ * not expose the named tool. Callers MUST handle null — never throw,
105
+ * because the model may emit stale tool names after a server restart.
106
+ */
107
+ export function resolveMcpTool(registry, parsed) {
108
+ if (!registry)
109
+ return null;
110
+ const state = registry.servers.get(parsed.serverName);
111
+ if (!state || state.trust !== 'trusted' || !state.connection)
112
+ return null;
113
+ const tool = state.surfacedTools.find((t) => t.name === parsed.toolName);
114
+ if (!tool)
115
+ return null;
116
+ return { state, connection: state.connection, tool };
117
+ }
118
+ /**
119
+ * The default prompt — used when no interactive bridge is wired (CI,
120
+ * non-TTY pipes). Returns `deny` so an unattended run never silently
121
+ * fires an MCP call the operator never approved. The deny is NOT
122
+ * persisted, so the next run with a wired prompt still has a chance to
123
+ * approve.
124
+ */
125
+ export const defaultNonInteractiveMcpPrompt = async () => 'unset';
126
+ /**
127
+ * Dispatch one MCP tool call. The flow:
128
+ *
129
+ * 1. Parse the namespaced tool name. Return error string when
130
+ * malformed — the model sees the error and can self-correct.
131
+ * 2. Resolve the live connection. Return error when the server is not
132
+ * trusted/connected or the tool is unknown.
133
+ * 3. Consult the permission cache. `deny` short-circuits. `allow_always`
134
+ * proceeds. `unset` invokes the prompt; the operator's verdict is
135
+ * persisted (allow_always/deny) or used one-shot (allow_once).
136
+ * 4. Parse the arguments string. Bad JSON -> error string.
137
+ * 5. Call `client.callTool` and stringify the content for the model.
138
+ *
139
+ * Throws ONLY on unrecoverable transport failures (e.g. the connection
140
+ * died mid-call). Tool-level errors from the upstream server are
141
+ * surfaced as `[MCP error] <message>` strings so the model can recover.
142
+ */
143
+ export async function dispatchMcpTool(input) {
144
+ const parsed = parseMcpToolName(input.name);
145
+ if (!parsed) {
146
+ return `[MCP dispatch error] tool name "${input.name}" does not match the ${MCP_TOOL_PREFIX}<server>__<tool> namespace`;
147
+ }
148
+ const resolved = resolveMcpTool(input.registry, parsed);
149
+ if (!resolved) {
150
+ return `[MCP dispatch error] no trusted+connected server "${parsed.serverName}" exposes a tool named "${parsed.toolName}"`;
151
+ }
152
+ let args;
153
+ try {
154
+ args = parseArgumentsRaw(input.argumentsRaw);
155
+ }
156
+ catch (error) {
157
+ return `[MCP dispatch error] invalid JSON in arguments for ${input.name}: ${error instanceof Error ? error.message : String(error)}`;
158
+ }
159
+ // Permission gate.
160
+ const cached = getMcpPermission(parsed.serverName, parsed.toolName);
161
+ let effective = cached;
162
+ if (cached === 'unset') {
163
+ const verdict = await input.prompt({
164
+ serverName: parsed.serverName,
165
+ toolName: parsed.toolName,
166
+ toolDescription: resolved.tool.description ?? '',
167
+ callArguments: args,
168
+ });
169
+ effective = verdict;
170
+ if (verdict === 'allow_always' || verdict === 'deny') {
171
+ setMcpPermission(parsed.serverName, parsed.toolName, verdict, resolveDecidedBy(input.decidedBy));
172
+ }
173
+ }
174
+ if (effective === 'deny') {
175
+ return `[MCP refused] operator denied ${parsed.serverName}:${parsed.toolName}`;
176
+ }
177
+ if (effective !== 'allow_once' && effective !== 'allow_always') {
178
+ // Includes `unset` returned by the non-interactive default prompt.
179
+ return `[MCP refused] no operator approval for ${parsed.serverName}:${parsed.toolName} (run from a TTY to approve)`;
180
+ }
181
+ // Dispatch.
182
+ let result;
183
+ try {
184
+ result = await callTool(resolved.connection, parsed.toolName, args, {
185
+ ...(input.timeoutMs !== undefined ? { timeoutMs: input.timeoutMs } : {}),
186
+ });
187
+ }
188
+ catch (error) {
189
+ // Transport-level failure (timeout, child died mid-call). Surface
190
+ // as a recoverable string so the model can degrade gracefully.
191
+ return `[MCP transport error] ${parsed.serverName}:${parsed.toolName}: ${error instanceof Error ? error.message : String(error)}`;
192
+ }
193
+ return renderMcpToolResult(result.content, result.isError, parsed);
194
+ }
195
+ function parseArgumentsRaw(raw) {
196
+ if (!raw || raw.trim() === '')
197
+ return {};
198
+ const parsed = JSON.parse(raw);
199
+ if (parsed === null || typeof parsed !== 'object' || Array.isArray(parsed)) {
200
+ throw new Error('arguments must be a JSON object');
201
+ }
202
+ return parsed;
203
+ }
204
+ function resolveDecidedBy(override) {
205
+ return (override?.trim() ||
206
+ process.env.PUGI_TRUSTED_BY?.trim() ||
207
+ process.env.USER?.trim() ||
208
+ process.env.USERNAME?.trim() ||
209
+ 'cli');
210
+ }
211
+ /**
212
+ * Project the MCP `content` payload into a single text string the model
213
+ * can ingest. MCP servers reply with `content: [{ type: 'text', text }]`
214
+ * by convention; we concatenate every `type: text` chunk and surface a
215
+ * `[MCP non-text content]` marker for other content kinds (images,
216
+ * resource references) which are not yet wired into Pugi's loop.
217
+ *
218
+ * `isError: true` from the upstream maps to a `[MCP error] ...` prefix
219
+ * so the model knows the call failed at the server, not at the
220
+ * transport.
221
+ */
222
+ export function renderMcpToolResult(content, isError, parsed) {
223
+ const text = projectTextContent(content);
224
+ const prefix = isError ? `[MCP error ${parsed.serverName}:${parsed.toolName}] ` : '';
225
+ if (text === null) {
226
+ // Fallback to a JSON dump so the model sees SOMETHING — better than
227
+ // an opaque empty string when the upstream uses image / resource
228
+ // content kinds.
229
+ try {
230
+ return `${prefix}${JSON.stringify(content)}`;
231
+ }
232
+ catch {
233
+ return `${prefix}[MCP non-serialisable content]`;
234
+ }
235
+ }
236
+ return `${prefix}${text}`;
237
+ }
238
+ function projectTextContent(content) {
239
+ if (content === null || content === undefined)
240
+ return '';
241
+ if (typeof content === 'string')
242
+ return content;
243
+ if (!Array.isArray(content))
244
+ return null;
245
+ const parts = [];
246
+ for (const entry of content) {
247
+ if (entry && typeof entry === 'object' && !Array.isArray(entry)) {
248
+ const obj = entry;
249
+ if (obj.type === 'text' && typeof obj.text === 'string') {
250
+ parts.push(obj.text);
251
+ continue;
252
+ }
253
+ }
254
+ // Non-text chunk — record a marker so the model knows something was
255
+ // dropped from the response.
256
+ parts.push('[MCP non-text content chunk]');
257
+ }
258
+ return parts.join('\n');
259
+ }
260
+ //# sourceMappingURL=mcp-tool.js.map
@@ -0,0 +1,361 @@
1
+ /**
2
+ * multi_edit tool — β7 (2026-05-26).
3
+ *
4
+ * Dispatches an ordered batch of file edits as a single transaction. Each
5
+ * edit is one Layer A (oldString -> newString) operation against one
6
+ * workspace file. Either every edit lands, or none do — failures roll
7
+ * the workspace back to the pre-dispatch state using the same journal +
8
+ * snapshot machinery the β1b Pl8 transactional layer uses for the
9
+ * marker-driven dispatcher.
10
+ *
11
+ * Why multi_edit when `edit` already exists:
12
+ *
13
+ * The single-shot `edit` tool is the right primitive for one mutation;
14
+ * the model uses it dozens of times in a typical session. A coordinated
15
+ * refactor (rename across 8 files, add an import to 12 modules, peel a
16
+ * helper into 5 callers) is currently 8/12/5 separate `edit` calls.
17
+ * Each call is its own audit + permission check + atomic write, which
18
+ * is the right shape for the audit story but means the model can leave
19
+ * the workspace half-mutated when one of the calls fails partway. The
20
+ * model also pays the round-trip latency once per call.
21
+ *
22
+ * `multi_edit` collapses the 8/12/5 calls into one tool dispatch with
23
+ * transactional semantics: snapshot every target file, attempt every
24
+ * edit against an in-memory buffer, then commit the writes only after
25
+ * all in-memory edits succeed. A failure rolls back via journal +
26
+ * in-memory snapshot — same code path as the dispatcher.
27
+ *
28
+ * Security: every target file routes through the same `applySecurityGate`
29
+ * chokepoint Layer A/B/C inherit. A path that escapes the workspace,
30
+ * points at a protected basename (`.env`, `*.pem`, ...), or symlinks
31
+ * outside the tree is refused BEFORE any read.
32
+ *
33
+ * Concurrency: marked `concurrencySafe: false` in the tool registry. The
34
+ * model MUST NOT issue another `multi_edit` (or any write tool) in
35
+ * parallel with one in flight; the journal serialises one dispatch per
36
+ * session.
37
+ *
38
+ * Output cap: a 50-edit batch is the soft ceiling. Beyond that the tool
39
+ * refuses with `too_many_edits` — the operator can split the refactor.
40
+ * Empirically a coordinated refactor that needs 50+ atomic edits should
41
+ * be a per-file Layer C rewrite instead.
42
+ *
43
+ * Brand voice: ASCII only, no emoji, no banned words.
44
+ */
45
+ import { existsSync, readFileSync, renameSync, unlinkSync, writeFileSync } from 'node:fs';
46
+ import { applySecurityGate } from '../core/edits/security-gate.js';
47
+ import { appendEntry, snapshotForDispatch, } from '../core/edits/journal.js';
48
+ import { rollbackDispatch } from '../core/edits/dispatch.js';
49
+ import { gateOnCancellation, OperatorAbortedError } from './file-tools.js';
50
+ import { recordFileMutation, recordToolCall, recordToolResult } from '../core/session.js';
51
+ /** Soft cap on per-dispatch edit count. See module docstring. */
52
+ const MULTI_EDIT_MAX = 50;
53
+ /**
54
+ * Apply a batch of file edits transactionally. Returns a structured
55
+ * result; never throws on operator-attributable failure (security,
56
+ * missing file, no_match) — only on infrastructure error (filesystem
57
+ * permission denied mid-write after the snapshot, etc.).
58
+ */
59
+ export function multiEdit(ctx, edits, opts = {}) {
60
+ const toolCallId = recordToolCall(ctx.session, 'multi_edit', `${edits.length} edits across ${new Set(edits.map((e) => e.file)).size} files`);
61
+ try {
62
+ gateOnCancellation(ctx, 'multi_edit');
63
+ }
64
+ catch (error) {
65
+ if (error instanceof OperatorAbortedError) {
66
+ recordToolResult(ctx.session, toolCallId, 'cancelled', error.message);
67
+ throw error;
68
+ }
69
+ throw error;
70
+ }
71
+ if (edits.length === 0) {
72
+ const result = {
73
+ ok: false,
74
+ filesChanged: [],
75
+ editsApplied: 0,
76
+ reason: 'empty_batch',
77
+ detail: 'multi_edit received zero edits',
78
+ perEdit: [],
79
+ };
80
+ recordToolResult(ctx.session, toolCallId, 'error', 'empty_batch');
81
+ return result;
82
+ }
83
+ if (edits.length > MULTI_EDIT_MAX) {
84
+ const result = {
85
+ ok: false,
86
+ filesChanged: [],
87
+ editsApplied: 0,
88
+ reason: 'too_many_edits',
89
+ detail: `multi_edit batch of ${edits.length} exceeds cap ${MULTI_EDIT_MAX}; split the refactor`,
90
+ perEdit: [],
91
+ };
92
+ recordToolResult(ctx.session, toolCallId, 'error', 'too_many_edits');
93
+ return result;
94
+ }
95
+ // SECURITY GATE pass over every distinct file BEFORE any read.
96
+ // A single rejected file aborts the whole batch — the transactional
97
+ // contract requires we never partial-mutate.
98
+ const uniqueFiles = Array.from(new Set(edits.map((e) => e.file)));
99
+ const resolvedByFile = new Map();
100
+ for (const f of uniqueFiles) {
101
+ const gate = applySecurityGate(f, { cwd: ctx.root, toolName: 'layer-c' });
102
+ if (!gate.ok) {
103
+ const result = {
104
+ ok: false,
105
+ filesChanged: [],
106
+ editsApplied: 0,
107
+ reason: gate.reason,
108
+ detail: `${f}: ${gate.detail}`,
109
+ perEdit: edits.map((e, i) => ({
110
+ index: i,
111
+ file: e.file,
112
+ ok: false,
113
+ reason: gate.reason,
114
+ detail: e.file === f ? gate.detail : 'batch aborted by sibling security failure',
115
+ })),
116
+ };
117
+ recordToolResult(ctx.session, toolCallId, 'error', `${gate.reason}: ${f}`);
118
+ return result;
119
+ }
120
+ resolvedByFile.set(f, gate.absPath);
121
+ }
122
+ // Snapshot existing files BEFORE any in-memory edit so a partial-write
123
+ // rollback is deterministic. The snapshot also captures sha256 of each
124
+ // pre-existing file so post-failure restore can verify the in-memory
125
+ // buffer still matches.
126
+ const snapshot = snapshotForDispatch(ctx.root, uniqueFiles);
127
+ const preContent = new Map();
128
+ for (const entry of snapshot) {
129
+ if (!entry.existed)
130
+ continue;
131
+ const abs = resolvedByFile.get(entry.path);
132
+ if (!abs)
133
+ continue;
134
+ try {
135
+ preContent.set(entry.path, readFileSync(abs));
136
+ }
137
+ catch {
138
+ // Best-effort. A read failure here will surface again when the
139
+ // per-edit phase tries to read the same file — let that path
140
+ // produce the operator-facing error.
141
+ }
142
+ }
143
+ // In-memory edit phase. For each edit we work on the latest version
144
+ // of the file (so two edits against the same file stack). Failure
145
+ // here is the common case — `no_match`, `ambiguous_match`, missing
146
+ // file — and aborts the whole batch.
147
+ const bodyByFile = new Map();
148
+ const perEdit = [];
149
+ for (let i = 0; i < edits.length; i += 1) {
150
+ const edit = edits[i];
151
+ const abs = resolvedByFile.get(edit.file);
152
+ if (!abs) {
153
+ // Should be unreachable — every distinct file went through the
154
+ // gate above. Belt + braces.
155
+ perEdit.push({ index: i, file: edit.file, ok: false, reason: 'write_error', detail: 'no resolved path' });
156
+ const result = {
157
+ ok: false,
158
+ filesChanged: [],
159
+ editsApplied: 0,
160
+ reason: 'write_error',
161
+ detail: `${edit.file}: no resolved path`,
162
+ perEdit,
163
+ };
164
+ recordToolResult(ctx.session, toolCallId, 'error', 'write_error');
165
+ return result;
166
+ }
167
+ let body = bodyByFile.get(edit.file);
168
+ if (body === undefined) {
169
+ if (!existsSync(abs)) {
170
+ const detail = `file does not exist: ${edit.file}`;
171
+ perEdit.push({ index: i, file: edit.file, ok: false, reason: 'file_missing', detail });
172
+ const result = {
173
+ ok: false,
174
+ filesChanged: [],
175
+ editsApplied: 0,
176
+ reason: 'file_missing',
177
+ detail,
178
+ perEdit,
179
+ };
180
+ recordToolResult(ctx.session, toolCallId, 'error', 'file_missing');
181
+ return result;
182
+ }
183
+ try {
184
+ body = readFileSync(abs, 'utf8');
185
+ }
186
+ catch (error) {
187
+ const detail = error instanceof Error ? error.message : String(error);
188
+ perEdit.push({ index: i, file: edit.file, ok: false, reason: 'write_error', detail });
189
+ const result = {
190
+ ok: false,
191
+ filesChanged: [],
192
+ editsApplied: 0,
193
+ reason: 'write_error',
194
+ detail: `${edit.file}: ${detail}`,
195
+ perEdit,
196
+ };
197
+ recordToolResult(ctx.session, toolCallId, 'error', 'write_error');
198
+ return result;
199
+ }
200
+ }
201
+ if (edit.oldString === edit.newString) {
202
+ perEdit.push({
203
+ index: i,
204
+ file: edit.file,
205
+ ok: false,
206
+ reason: 'identical_replacement',
207
+ detail: 'oldString and newString are identical',
208
+ });
209
+ const result = {
210
+ ok: false,
211
+ filesChanged: [],
212
+ editsApplied: 0,
213
+ reason: 'identical_replacement',
214
+ detail: `edit ${i} (${edit.file}): oldString and newString are identical`,
215
+ perEdit,
216
+ };
217
+ recordToolResult(ctx.session, toolCallId, 'error', 'identical_replacement');
218
+ return result;
219
+ }
220
+ const matches = countOccurrences(body, edit.oldString);
221
+ if (matches === 0) {
222
+ const detail = `edit ${i} (${edit.file}): oldString not found`;
223
+ perEdit.push({ index: i, file: edit.file, ok: false, reason: 'no_match', detail });
224
+ const result = {
225
+ ok: false,
226
+ filesChanged: [],
227
+ editsApplied: 0,
228
+ reason: 'no_match',
229
+ detail,
230
+ perEdit,
231
+ };
232
+ recordToolResult(ctx.session, toolCallId, 'error', 'no_match');
233
+ return result;
234
+ }
235
+ if (matches > 1) {
236
+ const detail = `edit ${i} (${edit.file}): oldString matches ${matches} times — expand context to make it unique`;
237
+ perEdit.push({ index: i, file: edit.file, ok: false, reason: 'ambiguous_match', detail });
238
+ const result = {
239
+ ok: false,
240
+ filesChanged: [],
241
+ editsApplied: 0,
242
+ reason: 'ambiguous_match',
243
+ detail,
244
+ perEdit,
245
+ };
246
+ recordToolResult(ctx.session, toolCallId, 'error', 'ambiguous_match');
247
+ return result;
248
+ }
249
+ body = body.replace(edit.oldString, edit.newString);
250
+ bodyByFile.set(edit.file, body);
251
+ perEdit.push({ index: i, file: edit.file, ok: true });
252
+ }
253
+ if (opts.dryRun) {
254
+ const result = {
255
+ ok: true,
256
+ filesChanged: Array.from(bodyByFile.keys()),
257
+ editsApplied: edits.length,
258
+ perEdit,
259
+ };
260
+ recordToolResult(ctx.session, toolCallId, 'success', `dry-run ${edits.length} edits ok`);
261
+ return result;
262
+ }
263
+ // Persist the snapshot to the journal BEFORE the first write. A crash
264
+ // mid-write then has a recoverable trail in `.pugi/sessions/<id>/journal.jsonl`.
265
+ // Best-effort; a journal write failure does not block the edits (the
266
+ // in-memory rollback path still covers same-process failures).
267
+ if (ctx.session.enabled) {
268
+ appendEntry(ctx.root, ctx.session.id, {
269
+ ts: Date.now(),
270
+ taskId: `multi_edit-${toolCallId}`,
271
+ files: snapshot,
272
+ });
273
+ }
274
+ // Commit phase. Atomic writes one file at a time. A failure rolls
275
+ // back via the same dispatcher rollback used by the marker layer.
276
+ const written = [];
277
+ for (const [file, body] of bodyByFile) {
278
+ const abs = resolvedByFile.get(file);
279
+ try {
280
+ atomicWrite(abs, body);
281
+ written.push(file);
282
+ }
283
+ catch (error) {
284
+ const detail = error instanceof Error ? error.message : String(error);
285
+ // Roll back every file we already touched plus restore the
286
+ // not-yet-touched ones that existed before (defensive — the
287
+ // rollback function is idempotent on untouched paths).
288
+ const rollback = rollbackDispatch(ctx.root, snapshot, preContent);
289
+ if (!rollback.ok) {
290
+ const result = {
291
+ ok: false,
292
+ filesChanged: [],
293
+ editsApplied: 0,
294
+ reason: 'rollback_failed',
295
+ detail: `${file}: ${detail}; rollback also failed: ${rollback.detail}`,
296
+ perEdit,
297
+ };
298
+ recordToolResult(ctx.session, toolCallId, 'error', 'rollback_failed');
299
+ return result;
300
+ }
301
+ const result = {
302
+ ok: false,
303
+ filesChanged: [],
304
+ editsApplied: 0,
305
+ reason: 'write_error',
306
+ detail: `${file}: ${detail}`,
307
+ perEdit,
308
+ };
309
+ recordToolResult(ctx.session, toolCallId, 'error', `write_error: ${detail}`);
310
+ return result;
311
+ }
312
+ }
313
+ for (const file of written) {
314
+ recordFileMutation(ctx.session, {
315
+ toolCallId,
316
+ path: file,
317
+ operation: 'update',
318
+ });
319
+ }
320
+ recordToolResult(ctx.session, toolCallId, 'success', `applied ${edits.length} edits across ${written.length} files`);
321
+ return {
322
+ ok: true,
323
+ filesChanged: written,
324
+ editsApplied: edits.length,
325
+ perEdit,
326
+ };
327
+ }
328
+ function countOccurrences(haystack, needle) {
329
+ if (needle.length === 0)
330
+ return 0;
331
+ let count = 0;
332
+ let from = 0;
333
+ while (true) {
334
+ const idx = haystack.indexOf(needle, from);
335
+ if (idx === -1)
336
+ return count;
337
+ count += 1;
338
+ from = idx + needle.length;
339
+ }
340
+ }
341
+ /** Atomic write helper — mirrors Layer A / Layer D. */
342
+ function atomicWrite(absPath, contents) {
343
+ const suffix = `${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
344
+ const tmp = `${absPath}.pugi-tmp-${suffix}`;
345
+ try {
346
+ writeFileSync(tmp, contents, { encoding: 'utf8', mode: 0o600 });
347
+ renameSync(tmp, absPath);
348
+ }
349
+ catch (error) {
350
+ try {
351
+ unlinkSync(tmp);
352
+ }
353
+ catch {
354
+ // tmp file may not exist if writeFileSync itself failed.
355
+ }
356
+ throw error;
357
+ }
358
+ }
359
+ /** Test-only surface. */
360
+ export const __test__ = { MULTI_EDIT_MAX };
361
+ //# sourceMappingURL=multi-edit.js.map