konoha-mcp 2.2.23 → 2.2.25

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 (132) hide show
  1. package/NOTICE +14 -8
  2. package/README.md +20 -5
  3. package/apps/web/build/adapter-node.js +1 -1
  4. package/apps/web/build/build-meta.json +2 -2
  5. package/apps/web/build/client/_app/immutable/entry/{app.TYH87iYy.js → app.DtoBIrs4.js} +2 -2
  6. package/apps/web/build/client/_app/immutable/entry/app.DtoBIrs4.js.br +0 -0
  7. package/apps/web/build/client/_app/immutable/entry/app.DtoBIrs4.js.gz +0 -0
  8. package/apps/web/build/client/_app/immutable/nodes/{0.D7dPIEWA.js → 0.B8Iw7Xl1.js} +1 -1
  9. package/apps/web/build/client/_app/immutable/nodes/0.B8Iw7Xl1.js.br +0 -0
  10. package/apps/web/build/client/_app/immutable/nodes/{0.D7dPIEWA.js.gz → 0.B8Iw7Xl1.js.gz} +0 -0
  11. package/apps/web/build/client/_app/immutable/nodes/{13.TOXVaTBr.js → 13.CQB_Ed6W.js} +1 -1
  12. package/apps/web/build/client/_app/immutable/nodes/13.CQB_Ed6W.js.br +0 -0
  13. package/apps/web/build/client/_app/immutable/nodes/{13.TOXVaTBr.js.gz → 13.CQB_Ed6W.js.gz} +0 -0
  14. package/apps/web/build/client/_app/version.json +1 -1
  15. package/apps/web/build/client/_app/version.json.br +0 -0
  16. package/apps/web/build/client/_app/version.json.gz +0 -0
  17. package/apps/web/build/server/chunks/instance.js +3 -3
  18. package/apps/web/build/server/chunks/internal.js +1 -1
  19. package/apps/web/build/server/entries/pages/_layout.svelte.js +1 -1
  20. package/apps/web/build/server/entries/pages/_layout.svelte.js.map +1 -1
  21. package/apps/web/build/server/entries/pages/docs/_page.svelte.js +1 -1
  22. package/apps/web/build/server/entries/pages/docs/_page.svelte.js.map +1 -1
  23. package/apps/web/build/server/manifest-full.js +1 -1
  24. package/apps/web/build/server/manifest.js +1 -1
  25. package/apps/web/build/server/nodes/0.js +1 -1
  26. package/apps/web/build/server/nodes/13.js +1 -1
  27. package/apps/web/build/server/server.js +1 -1
  28. package/apps/web/package.json +1 -1
  29. package/assets/demo-agy.webp +0 -0
  30. package/assets/demo-claude.webp +0 -0
  31. package/assets/demo-codex.webp +0 -0
  32. package/assets/demo-commandcode.webp +0 -0
  33. package/assets/demo-cursor.webp +0 -0
  34. package/assets/demo-opencode.webp +0 -0
  35. package/assets/demo-pi.webp +0 -0
  36. package/assets/demo-skill-embed.webp +0 -0
  37. package/assets/demo-skills.webp +0 -0
  38. package/assets/demo-web.webp +0 -0
  39. package/assets/demo.webp +0 -0
  40. package/assets/konoha-akatsuki-flow.webp +0 -0
  41. package/assets/konoha-anbu-flow.webp +0 -0
  42. package/assets/konoha-flow-12-cli-upgrade-progress-engine.webp +0 -0
  43. package/assets/konoha-jonin-flow.webp +0 -0
  44. package/assets/konoha-kage-gate.webp +0 -0
  45. package/assets/konoha-langgraph-flow.webp +0 -0
  46. package/assets/konoha-orchestration-flow.webp +0 -0
  47. package/assets/konoha-qa-flow.webp +0 -0
  48. package/assets/konoha-telegram-flow.webp +0 -0
  49. package/assets/testing.webp +0 -0
  50. package/bin/commands/bridge.js +62 -4
  51. package/bin/commands/doctor.js +2 -0
  52. package/bin/commands/doctor_services.js +57 -0
  53. package/bin/commands/init.js +29 -8
  54. package/bin/commands/init_clients.js +95 -8
  55. package/bin/commands/init_migration.js +32 -1
  56. package/bin/lib/config_defaults.js +3 -0
  57. package/bin/lib/runtime_install.js +147 -54
  58. package/node_modules/jszip/package.json +2 -2
  59. package/node_modules/string_decoder/package.json +1 -1
  60. package/package.json +10 -3
  61. package/patches/gpt-tokenizer@4.1.0.patch +15 -0
  62. package/patches/jszip@3.10.2.patch +11 -0
  63. package/patches/string_decoder@1.3.0.patch +12 -0
  64. package/src/agent_contract.js +2 -2
  65. package/src/agent_deployer.js +48 -35
  66. package/src/agent_file_safety.js +57 -0
  67. package/src/antigravity_subagent_hook.js +30 -12
  68. package/src/bridge/gateway.js +2 -0
  69. package/src/bridge/gateway_adapters.js +155 -29
  70. package/src/bridge/handlers/anthropic.js +95 -26
  71. package/src/bridge/handlers/openai.js +18 -5
  72. package/src/bridge/sidecar/raw.js +10 -2
  73. package/src/bridge/stream_usage.js +66 -0
  74. package/src/chunin_knowledge_ddl.js +95 -0
  75. package/src/client_commands.js +190 -0
  76. package/src/db_bridges.js +53 -7
  77. package/src/db_schema.js +14 -2
  78. package/src/deploy_installer.js +212 -11
  79. package/src/deploy_utils.js +4 -0
  80. package/src/file_tools/common.js +65 -3
  81. package/src/file_tools_mcp.js +17 -9
  82. package/src/file_tools_router.js +67 -6
  83. package/src/global_skill_sync.js +3 -1
  84. package/src/mcp/chunin_knowledge.js +509 -0
  85. package/src/mcp/langgraph/edges/router.js +3 -0
  86. package/src/mcp/langgraph/index.js +22 -3
  87. package/src/mcp/langgraph/nodes/dispatch.js +3 -0
  88. package/src/mcp/langgraph/nodes/execute.js +87 -13
  89. package/src/mcp/langgraph/nodes/review.js +13 -1
  90. package/src/mcp/langgraph/slice_executor_registry.js +22 -0
  91. package/src/mcp/langgraph/state.js +5 -0
  92. package/src/mcp/lazy_subagent.js +11 -1
  93. package/src/mcp/memory_project_context.js +4 -1
  94. package/src/mcp/memory_subagent_runner.js +2 -1
  95. package/src/mcp/skills_common.js +1 -0
  96. package/src/mcp/subagent_spawner.js +7 -0
  97. package/src/mcp/todo_walkthrough_sync.js +7 -0
  98. package/src/mcp/tool_dispatch.js +18 -1
  99. package/src/mcp/web_search.js +27 -3
  100. package/src/mcp/workflow_completion_helpers.js +6 -1
  101. package/src/mcp/workflow_gates.js +118 -5
  102. package/src/mcp/workflow_sannin.js +9 -3
  103. package/src/mcp_clients/common.js +31 -0
  104. package/src/mcp_tool_manifest.json +36 -1
  105. package/src/pi_fs_helpers.js +31 -0
  106. package/src/pi_manager.js +68 -28
  107. package/src/provider_usage_logger.js +92 -0
  108. package/src/schema_ddl.js +17 -1
  109. package/src/templates/AGENTS.md +1 -1
  110. package/src/templates/GEMINI.md +1 -1
  111. package/src/templates/skills/chunin-skill/SKILL.md +34 -0
  112. package/src/templates/skills/chunin-skill/references/deep-research-analyst.md +27 -10
  113. package/src/templates/skills/kage-skill/SKILL.md +1 -1
  114. package/src/templates/skills/konoha/SKILL.md +23 -165
  115. package/src/templates/skills/konoha/references/contract.md +2 -2
  116. package/src/templates/skills/konoha/references/peragent-usage.md +1 -1
  117. package/src/token_usage.js +57 -0
  118. package/src/tools_savings_logger.js +26 -8
  119. package/src/url_canonical.js +58 -0
  120. package/src/vector_pipeline.js +6 -4
  121. package/src/vector_search.js +4 -2
  122. package/src/web/controllers/core_controller.js +10 -1
  123. package/src/workflow/jonin-builder/live_browser_monitor.js +23 -7
  124. package/apps/web/build/client/_app/immutable/entry/app.TYH87iYy.js.br +0 -0
  125. package/apps/web/build/client/_app/immutable/entry/app.TYH87iYy.js.gz +0 -0
  126. package/apps/web/build/client/_app/immutable/nodes/0.D7dPIEWA.js.br +0 -0
  127. package/apps/web/build/client/_app/immutable/nodes/13.TOXVaTBr.js.br +0 -0
  128. package/src/templates/skills/konoha/references/architecture.md +0 -99
  129. package/src/templates/skills/konoha/references/benchmark.md +0 -123
  130. package/src/templates/skills/konoha/references/docs.md +0 -46
  131. package/src/templates/skills/konoha/references/test.md +0 -71
  132. package/src/templates/skills/konoha/references/workflow.md +0 -104
@@ -10,6 +10,7 @@ const { normalizeAgentName } = require('../../../canonical_agents');
10
10
  const { getActiveClient } = require('../../runtime_state');
11
11
  const { detectActiveClient, resolveFirstSpawnClient } = require('../../client_detection');
12
12
  const { finalizeNodeResult } = require('../budget');
13
+ const { getSliceExecutor } = require('../slice_executor_registry');
13
14
  const { syncTodoAndWalkthrough } = require('../../todo_walkthrough_sync');
14
15
  const worktree = require('../../worktree_isolation');
15
16
 
@@ -21,6 +22,15 @@ function recordCompletedDispatch(status, dispatchId) {
21
22
  }
22
23
  }
23
24
 
25
+ function _branchFailed(branchResult) {
26
+ if (!branchResult) return false;
27
+ if (branchResult.spawn_state) return branchResult.spawn_state === 'failed';
28
+ return branchResult.status === 'failed'
29
+ || branchResult.status === 'error'
30
+ || branchResult.status === 'timeout'
31
+ || Boolean(branchResult.error_code);
32
+ }
33
+
24
34
  async function _executeTaskBranchConcurrently(task, idx, status, state, sliceExecutor) {
25
35
  const startedAt = Date.now();
26
36
  await new Promise((resolve) => setTimeout(resolve, 2));
@@ -28,6 +38,13 @@ async function _executeTaskBranchConcurrently(task, idx, status, state, sliceExe
28
38
  const targetAgent = normalizeAgentName(task.agent || status.assigned_agent || 'anbu');
29
39
  if (typeof sliceExecutor === 'function') {
30
40
  branchResult = await sliceExecutor(task, idx, status, state);
41
+ // A sliceExecutor is the caller's own trusted executor: it ran the work and
42
+ // reports the outcome. Absent an explicit denial, its documented `completed`
43
+ // status is the completion signal — the MCP runner path below is the one
44
+ // that never spawns anything and therefore never earns that status.
45
+ if (branchResult && branchResult.child_execution_verified === undefined) {
46
+ branchResult.child_execution_verified = branchResult.status === 'completed';
47
+ }
31
48
  } else {
32
49
  const projectPath = state.projectPath || status.project_path || process.cwd();
33
50
  const taskDir = state.taskDir || status.task_dir || '';
@@ -47,18 +64,27 @@ async function _executeTaskBranchConcurrently(task, idx, status, state, sliceExe
47
64
  });
48
65
  let parsedMcp = {};
49
66
  try { parsedMcp = JSON.parse(rawMcp); } catch (_) { parsedMcp = {}; }
67
+ // `runMcpAgentAsync` prepares a prompt and a spawn directive; it never starts
68
+ // a child process. Reporting the slice as completed with "exit 0" made an
69
+ // unspawned slice indistinguishable from a finished one, which is how a run
70
+ // could reach Gate 2 having executed nothing. Without a sliceExecutor there
71
+ // is no exit code to cite, so the slice stays pending and carries the
72
+ // directive the client must actually run.
50
73
  branchResult = {
51
74
  task_id: task.id,
52
75
  agent: parsedMcp.agent || targetAgent,
53
76
  model: parsedMcp.model || 'inherit',
54
77
  spawn_directive: parsedMcp.spawn_directive || null,
55
- summary: `Executed specialist @${targetAgent} slice (${task.id}): ${task.task || task.id}`,
56
- status: 'completed',
57
- child_execution_verified: true,
58
- validation: [`verified_specialist_slice:${targetAgent} (exit 0)`]
78
+ spawn_error: parsedMcp.spawn_error || null,
79
+ summary: `Spawn directive prepared for @${targetAgent} slice (${task.id}); no child process was started by the MCP runner`,
80
+ status: 'spawn_pending',
81
+ spawn_state: 'directive_only',
82
+ child_execution_verified: false,
83
+ validation: []
59
84
  };
60
85
  }
61
86
  const completedAt = Date.now();
87
+ const childVerified = branchResult?.child_execution_verified === true;
62
88
  const advisorEval = await evaluateWorkerTurnAsync({
63
89
  agent: task.agent || status.assigned_agent || 'anbu',
64
90
  taskId: task.id,
@@ -68,8 +94,8 @@ async function _executeTaskBranchConcurrently(task, idx, status, state, sliceExe
68
94
  ownedFiles: task.owned_files || [],
69
95
  filesModified: branchResult?.filesModified || [],
70
96
  worktreePath: task.worktree_path || null,
71
- status: 'completed',
72
- validation: branchResult?.validation || ['verified_specialist_slice (exit 0)']
97
+ status: childVerified ? 'completed' : 'spawn_pending',
98
+ validation: branchResult?.validation || []
73
99
  });
74
100
  return {
75
101
  task,
@@ -82,6 +108,7 @@ async function _executeTaskBranchConcurrently(task, idx, status, state, sliceExe
82
108
  completed_at: completedAt,
83
109
  elapsed_ms: Math.max(1, completedAt - startedAt),
84
110
  worktree_path: task.worktree_path || null,
111
+ child_execution_verified: childVerified,
85
112
  advisor_status: advisorEval.status,
86
113
  advisor_judge_tokens: (advisorEval.judge_usage && advisorEval.judge_usage.total_tokens) || 0
87
114
  }
@@ -91,6 +118,10 @@ async function _executeTaskBranchConcurrently(task, idx, status, state, sliceExe
91
118
  async function completePendingTasks(status, state, config) {
92
119
  const tasks = status.tasks || [];
93
120
  const pendingTasks = tasks.filter(t => t.status !== 'completed');
121
+ // The previous turn's directives are stale by definition: a slice that was
122
+ // reported complete is no longer waiting on a child, and leaving the old entry
123
+ // in place made the graph end on a spawn that had already happened.
124
+ status.spawn_pending = [];
94
125
  if (pendingTasks.length === 0) return tasks;
95
126
 
96
127
  const sliceResults = state.sliceResults;
@@ -105,7 +136,9 @@ async function completePendingTasks(status, state, config) {
105
136
  return tasks;
106
137
  }
107
138
 
108
- const sliceExecutor = state.sliceExecutor || config?.configurable?.sliceExecutor || null;
139
+ // `state.sliceExecutor` covers a direct node call; the registry covers a graph run,
140
+ // because LangGraph rebuilds `configurable` and would drop a caller's function.
141
+ const sliceExecutor = state.sliceExecutor || getSliceExecutor(state.threadId || config?.configurable?.thread_id);
109
142
  if ((state.requireExplicitDelegation || status.require_explicit_delegation) && !sliceExecutor) {
110
143
  return tasks;
111
144
  }
@@ -119,14 +152,43 @@ async function completePendingTasks(status, state, config) {
119
152
  : [];
120
153
 
121
154
  for (const outcome of branchOutcomes) {
122
- outcome.task.status = 'completed';
123
- outcome.task.result = outcome.branchResult;
155
+ const verified = outcome.branchResult && outcome.branchResult.child_execution_verified === true;
156
+ if (verified) {
157
+ outcome.task.status = 'completed';
158
+ outcome.task.spawn_state = 'completed';
159
+ outcome.task.result = outcome.branchResult;
160
+ recordCompletedDispatch(status, outcome.task.dispatch_id);
161
+ } else {
162
+ // No child process produced this slice. Keep it visible as pending work and
163
+ // hand the directive back so the client can actually spawn it; the review
164
+ // node's delegation guard then routes to dispatch instead of Gate 2.
165
+ outcome.task.status = 'spawn_pending';
166
+ // `failed` when the attempt actually ran and lost; `directive_only` when
167
+ // nothing was ever started and the client still owes the spawn.
168
+ outcome.task.spawn_state = _branchFailed(outcome.branchResult) ? 'failed' : 'directive_only';
169
+ outcome.task.spawn_directive = (outcome.branchResult && outcome.branchResult.spawn_directive) || null;
170
+ // The result stays attached so a caller can read the directive it must run;
171
+ // `status` on the task is what marks the slice as unfinished.
172
+ outcome.task.result = outcome.branchResult;
173
+ status.spawn_pending = Array.isArray(status.spawn_pending) ? status.spawn_pending : [];
174
+ status.spawn_pending.push({
175
+ task_id: outcome.task.id,
176
+ agent: outcome.task.agent || null,
177
+ spawn_directive: outcome.task.spawn_directive,
178
+ spawn_error: (outcome.branchResult && outcome.branchResult.spawn_error) || null
179
+ });
180
+ }
124
181
  outcome.task.started_at = outcome.telemetry.started_at;
125
- outcome.task.completed_at = outcome.telemetry.completed_at;
126
- recordCompletedDispatch(status, outcome.task.dispatch_id);
182
+ outcome.task.completed_at = verified ? outcome.telemetry.completed_at : null;
127
183
  status.parallel_execution_telemetry.push(outcome.telemetry);
128
184
  }
129
- if (branchOutcomes.length >= 2) {
185
+ // The status phase is what a resumed call and the delegation guard read; a slice
186
+ // waiting on its client belongs to the execute stage, not to whatever phase the
187
+ // previous node left behind.
188
+ if (status.spawn_pending.length > 0) {
189
+ status.phase = 'execute';
190
+ }
191
+ if (branchOutcomes.length >= 2 && branchOutcomes.every(o => o.branchResult && o.branchResult.child_execution_verified === true)) {
130
192
  status.parallel_execution_proof = {
131
193
  concurrent_count: branchOutcomes.length,
132
194
  branches: status.parallel_execution_telemetry.slice(-branchOutcomes.length),
@@ -233,6 +295,7 @@ async function executeNode(state, config) {
233
295
  const breakerStatus = breaker.getStatus();
234
296
  const circuitOpen = breakerStatus.state === 'OPEN';
235
297
 
298
+ const spawnPending = Array.isArray(status.spawn_pending) && status.spawn_pending.length > 0;
236
299
  const outputPayload = {
237
300
  status,
238
301
  phase: status.phase || 'review',
@@ -243,10 +306,20 @@ async function executeNode(state, config) {
243
306
  circuitOpen,
244
307
  circuitState: breakerStatus.state,
245
308
  error: executionError,
309
+ // A slice whose child was never started cannot be reviewed. Returning here
310
+ // hands the directive to the client, which spawns the child and reports back
311
+ // through `report_from_agent`; the next invoke resumes at this node with the
312
+ // tasks completed. Without this the graph bounced execute -> review ->
313
+ // dispatch until the remediation cap opened the circuit breaker.
314
+ spawnPending,
315
+ spawn_pending: spawnPending ? status.spawn_pending : undefined,
246
316
  };
247
317
 
248
- const nextNode = outputPayload.phase === 'document' ? 'node_dispatch' : 'node_review';
318
+ const nextNode = spawnPending
319
+ ? '__end__'
320
+ : (outputPayload.phase === 'document' ? 'node_dispatch' : 'node_review');
249
321
  return finalizeNodeResult(state, outputPayload, nextNode, {
322
+ resumeNode: spawnPending ? 'node_execute' : undefined,
250
323
  source: {
251
324
  status,
252
325
  phase: outputPayload.phase,
@@ -254,6 +327,7 @@ async function executeNode(state, config) {
254
327
  circuitOpen,
255
328
  circuitState: breakerStatus.state,
256
329
  error: executionError,
330
+ spawnPending,
257
331
  }
258
332
  });
259
333
  }
@@ -157,6 +157,14 @@ function _delegationGuard(state, status, taskDir) {
157
157
  status.hitl_gate = null;
158
158
  if (taskDir) saveWorkflowStatus(taskDir, status);
159
159
 
160
+ const pendingSpawns = pendingTasks
161
+ .filter((t) => t.status === 'spawn_pending' && t.spawn_directive)
162
+ .map((t) => ({
163
+ task_id: t.id,
164
+ agent: t.agent || null,
165
+ spawn_state: t.spawn_state || 'directive_only',
166
+ spawn_directive: t.spawn_directive
167
+ }));
160
168
  const payload = {
161
169
  status,
162
170
  phase: 'execute',
@@ -164,11 +172,15 @@ function _delegationGuard(state, status, taskDir) {
164
172
  hitlStage: null,
165
173
  delegationPending: true,
166
174
  pending_task_ids: pendingTasks.map((t) => t.id),
175
+ // A slice the MCP runner only prepared a directive for has to be spawned by
176
+ // the client. Without handing the directive back here the run bounced
177
+ // dispatch -> review with no way to see what was waiting on it.
178
+ spawn_pending: pendingSpawns.length > 0 ? pendingSpawns : undefined,
167
179
  error: null,
168
180
  };
169
181
  return finalizeNodeResult(state, payload, 'node_dispatch', {
170
182
  resumeNode: 'node_dispatch',
171
- source: { delegationPending: true, pending: payload.pending_task_ids },
183
+ source: { delegationPending: true, pending: payload.pending_task_ids, spawnPending: pendingSpawns.length },
172
184
  });
173
185
  }
174
186
 
@@ -0,0 +1,22 @@
1
+ 'use strict';
2
+
3
+ // LangGraph rebuilds `config.configurable` for its own runtime keys, so a caller's
4
+ // executor cannot ride along there. A function also cannot live in the state
5
+ // schema: the checkpointer serialises state. This registry is the one channel that
6
+ // survives both, keyed by thread so concurrent runs stay isolated.
7
+ const executors = new Map();
8
+
9
+ function setSliceExecutor(threadId, fn) {
10
+ if (!threadId || typeof fn !== 'function') return;
11
+ executors.set(threadId, fn);
12
+ }
13
+
14
+ function getSliceExecutor(threadId) {
15
+ return threadId ? executors.get(threadId) || null : null;
16
+ }
17
+
18
+ function clearSliceExecutor(threadId) {
19
+ if (threadId) executors.delete(threadId);
20
+ }
21
+
22
+ module.exports = { setSliceExecutor, getSliceExecutor, clearSliceExecutor };
@@ -100,6 +100,11 @@ const KonohaWorkflowState = Annotation.Root({
100
100
  hitlRefinement: Annotation({ default: () => null }),
101
101
  autoSignoff: Annotation({ default: () => false }),
102
102
 
103
+ // Declared as channels or LangGraph drops them: the execute node ends the turn
104
+ // here when a slice only has a spawn directive, and the router reads the flag.
105
+ spawnPending: Annotation({ default: () => false }),
106
+ spawn_pending: Annotation({ default: () => null }),
107
+
103
108
  result: Annotation({ default: () => null }),
104
109
  error: Annotation({ default: () => null }),
105
110
  });
@@ -2,6 +2,7 @@
2
2
 
3
3
  const { isReadOnlyAgent } = require('../agent_file_safety');
4
4
  const { hasPiSubagentSupport } = require('../pi_manager');
5
+ const { normalizeAgentName } = require('../canonical_agents');
5
6
 
6
7
  const SPAWNABLE_SUBAGENTS = Object.freeze([
7
8
  'genin',
@@ -109,6 +110,9 @@ function buildUnsupportedSpawnPayload(client, reason) {
109
110
  return {
110
111
  ok: false,
111
112
  supported: false,
113
+ // Nothing was started, so there is no child to report on. This is a
114
+ // capability gap, not a failed run.
115
+ spawn_state: 'directive_only',
112
116
  error_code: 'UNSUPPORTED_NATIVE_SPAWN',
113
117
  client: client || 'unknown',
114
118
  error: reason,
@@ -192,8 +196,14 @@ function _cliTaskExcerpt(prompt) {
192
196
  return _shellQuotePosix(truncateSpawnTask(raw).replace(/`/g, ''));
193
197
  }
194
198
 
199
+ // Agent tokens must survive to the deployed profile name verbatim. The previous
200
+ // sanitizer stripped every non `[a-z0-9-]` character, so `tokubetsu_jonin` became
201
+ // `tokubetsujonin` while the profile on disk is `tokubetsu-jonin.md`; the native
202
+ // spawn then named an agent no client could resolve and the CLI fallback failed
203
+ // the same way. normalizeAgentName is the single canonical mapping (`_`/space ->
204
+ // `-`, `@` stripped, lowercase), so every directive and the profile agree.
195
205
  function _safeAgentToken(raw) {
196
- const cleaned = String(raw || 'subagent').replace(/^mcp_/, '').toLowerCase().trim().replace(/[^a-z0-9-]/g, '');
206
+ const cleaned = normalizeAgentName(String(raw || 'subagent').replace(/^mcp_/, ''));
197
207
  return cleaned || 'subagent';
198
208
  }
199
209
 
@@ -316,7 +316,10 @@ function _reportFromAgent(delivery, rest, useModelJudge) {
316
316
  // `cleanAgent`, not `opts.agentName`: the dispatch guard below compares against
317
317
  // the normalized name, so `mcp_kage` and `Anbu` were rejected as mismatches.
318
318
  const { error: wfError, workflowStatus } = _updateWorkflowStatus(safeTaskDir || opts.taskDir, opts.dispatchId, cleanAgent, finalStatus, {
319
- summary: reportSummary, validation: opts.validation, verified, filesCreated: opts.filesCreated, filesModified: opts.filesModified
319
+ summary: reportSummary, validation: opts.validation, verified, filesCreated: opts.filesCreated, filesModified: opts.filesModified,
320
+ // A report may name the slice it belongs to; without it a parallel workflow
321
+ // attributes every child's result to whichever task happens to be open.
322
+ taskId: opts.taskId
320
323
  });
321
324
  if (wfError) return JSON.stringify({ status: 'error', message: wfError });
322
325
 
@@ -694,5 +694,6 @@ async function runMcpAgentAsync(agentName, task = null, context = null, constrai
694
694
  module.exports = {
695
695
  getMainModel, applyFileEdits, autoloadSkillsFromPrompt, buildSubagentMcpBlock,
696
696
  buildResearchFetchRule, truncateAtBoundary, cleanRedundantPersonaText,
697
- _loadBaseProjectContext, _prepareAgentRun, _finalizeAgentRun, runMcpAgent, runMcpAgentAsync
697
+ _loadBaseProjectContext, _prepareAgentRun, _finalizeAgentRun, runMcpAgent, runMcpAgentAsync,
698
+ PROMPT_CEILING_BYTES, SYSTEM_PROMPT_FLOOR_BYTES, MIN_SKILL_PREVIEW_BYTES
698
699
  };
@@ -363,6 +363,7 @@ function resolveFlowHashFromTask(flowHash, taskId) {
363
363
  // stays authoritative: the database is an index of them, never a replacement.
364
364
  const PROJECT_SKILL_SUBDIRS = [
365
365
  ['.agents', 'skills'],
366
+ ['.agents', 'dev-skills'],
366
367
  ['.agent', 'skills'],
367
368
  ['skills'],
368
369
  ['.skills'],
@@ -332,6 +332,8 @@ function _prepareSpawnSpec(options) {
332
332
  aborted: true,
333
333
  agent: normalizedAgent,
334
334
  client: capability.client,
335
+ // Cancelled before launch: no child ever existed.
336
+ spawn_state: 'directive_only',
335
337
  error_code: 'SPAWN_ABORTED',
336
338
  error: 'Subagent spawn cancelled before launch.'
337
339
  }
@@ -454,6 +456,9 @@ function _buildFailureExitResult({
454
456
  return {
455
457
  ok: false,
456
458
  status: kind.status,
459
+ // Fixed vocabulary: directive_only | completed | failed. A timeout or abort
460
+ // is a failure, not a state of its own.
461
+ spawn_state: 'failed',
457
462
  ...(kind.flagKey ? { [kind.flagKey]: true } : {}),
458
463
  agent: normalizedAgent,
459
464
  client,
@@ -566,6 +571,7 @@ function _executeSpawnSpec(spec, options, childEnv, startedAt) {
566
571
  client: capability.client,
567
572
  child_pid: childPid || null,
568
573
  elapsed_ms: Date.now() - startedAt,
574
+ spawn_state: 'failed',
569
575
  error_code: err.code === 'ENOENT' ? 'HARNESS_BINARY_NOT_FOUND' : 'SPAWN_PROCESS_ERROR',
570
576
  error: err.message
571
577
  });
@@ -609,6 +615,7 @@ function _executeSpawnSpec(spec, options, childEnv, startedAt) {
609
615
  ...parsed,
610
616
  ok: true,
611
617
  status: parsed.status || 'completed',
618
+ spawn_state: 'completed',
612
619
  agent: normalizedAgent,
613
620
  client: capability.client,
614
621
  child_pid: parsed.child_pid || childPid,
@@ -47,6 +47,10 @@ function normalizeTodoItems(tasks = [], status = {}) {
47
47
  if (rawStatus === 'completed' || rawStatus === 'done') mappedStatus = 'completed';
48
48
  else if (rawStatus === 'in_progress' || rawStatus === 'running') mappedStatus = 'in_progress';
49
49
  else if (['unverified', 'failed', 'rejected', 'blocked'].includes(rawStatus)) mappedStatus = rawStatus;
50
+ // `spawn_pending` is a real state, not a typo: the slice has a directive the
51
+ // client must run and is only completed by `report_from_agent`. Mapping it to
52
+ // `pending` here erased the marker on every sync and hid the waiting spawn.
53
+ else if (rawStatus === 'spawn_pending') mappedStatus = 'spawn_pending';
50
54
  const text = base.task || base.content || base.description || `Step ${idx + 1}`;
51
55
  return {
52
56
  ...base,
@@ -196,6 +200,9 @@ function applyAgentReportToTodos(workflowStatus, cleanAgent, dispatchId, finalSt
196
200
 
197
201
  if (finalStatus === 'completed' || finalStatus === 'unverified') {
198
202
  task.status = finalStatus;
203
+ // A report without passing validation evidence cannot have produced a
204
+ // finished child: `completed` requires the evidence gate to have passed.
205
+ task.spawn_state = finalStatus === 'completed' ? 'completed' : 'directive_only';
199
206
  task.result = updateData.summary || '';
200
207
  task.validation = updateData.validation || [];
201
208
  task.verified = updateData.verified;
@@ -326,6 +326,22 @@ function _mergeSanninGraphResponse(parsedSannin, graphInitRes, threadId) {
326
326
  };
327
327
  }
328
328
 
329
+ // A slice that only has a spawn directive is not routed work: the client has to
330
+ // run the directive and report back. Answering it with the routing envelope made
331
+ // the caller believe a subagent had been dispatched.
332
+ if (graphInitRes.spawnPending) {
333
+ return {
334
+ ...parsedSannin,
335
+ status: 'spawn_pending',
336
+ phase: graphInitRes.phase,
337
+ spawn_pending: graphInitRes.spawn_pending || (wfStatus && wfStatus.spawn_pending) || [],
338
+ next_action: 'spawn_and_report',
339
+ resume_node: 'node_execute',
340
+ langgraph_checkpoint: checkpoint,
341
+ workflow_state: typeof graphInitRes.status === 'object' ? graphInitRes.status : undefined
342
+ };
343
+ }
344
+
329
345
  if (parsedSannin.status === 'routed') {
330
346
  return {
331
347
  ...graphInitRes,
@@ -569,7 +585,8 @@ async function _executeHeavyToolAsync(toolName, args, agent) {
569
585
  const query = args.query;
570
586
  const numResults = Math.min(Math.max(parseInt(args.num_results || 5, 10), 1), 50);
571
587
  const searchDepth = ['standard', 'deep'].includes(args.search_depth) ? args.search_depth : 'standard';
572
- return await runWebSearch(query, numResults, searchDepth, agent);
588
+ const excludeUrls = Array.isArray(args.exclude_urls) ? args.exclude_urls : undefined;
589
+ return await runWebSearch(query, numResults, searchDepth, agent, { excludeUrls });
573
590
  }
574
591
  const docEngineMethod = {
575
592
  document_create: 'documentCreate',
@@ -12,6 +12,7 @@ const { detectActiveAgent } = require('./client_detection');
12
12
  const { logToolCall } = require('./skills');
13
13
  const { globalCircuitRegistry } = require('../circuit_breaker');
14
14
  const { getConfigValue } = require('../config_access');
15
+ const { canonicalizeUrl } = require('../url_canonical');
15
16
 
16
17
  const searxngDir = path.join(KONOHA_DIR, 'searxng');
17
18
  const instancesCachePath = path.join(searxngDir, 'instances_cache.json');
@@ -437,13 +438,29 @@ async function _queryBlendedSearch(q, numResults, isBlended, requestedEngine) {
437
438
 
438
439
  function _appendUniqueResults(currentResults, seenUrls, allResults) {
439
440
  for (const r of currentResults) {
440
- if (!seenUrls.has(r.url)) {
441
- seenUrls.add(r.url);
441
+ const key = canonicalizeUrl(r.url) || r.url;
442
+ if (!seenUrls.has(key)) {
443
+ seenUrls.add(key);
442
444
  allResults.push(r);
443
445
  }
444
446
  }
445
447
  }
446
448
 
449
+ /**
450
+ * Normalizes the caller's `exclude_urls` into canonical keys. Accepts raw URLs,
451
+ * bare hosts, and URLs already in canonical form; anything unparseable is
452
+ * dropped rather than throwing, because a bad exclusion must not fail a search.
453
+ */
454
+ function _normalizeExclusions(excludeUrls) {
455
+ const set = new Set();
456
+ if (!Array.isArray(excludeUrls)) return set;
457
+ for (const entry of excludeUrls) {
458
+ const canonical = canonicalizeUrl(entry);
459
+ if (canonical) set.add(canonical);
460
+ }
461
+ return set;
462
+ }
463
+
447
464
  async function runWebSearch(query, numResults = 5, searchDepth = 'standard', agentName = null, options = {}) {
448
465
  if (!query || !query.trim()) {
449
466
  return JSON.stringify({ status: 'error', message: 'Query is required.' });
@@ -460,6 +477,7 @@ async function runWebSearch(query, numResults = 5, searchDepth = 'standard', age
460
477
 
461
478
  const allResults = [];
462
479
  const seenUrls = new Set();
480
+ const excluded = _normalizeExclusions(options && options.excludeUrls);
463
481
  const requestedEngine = (options && options.engine) ? String(options.engine).toLowerCase() : 'all';
464
482
  const isBlended = options && options.blended !== undefined ? options.blended : (requestedEngine === 'all');
465
483
 
@@ -469,11 +487,16 @@ async function runWebSearch(query, numResults = 5, searchDepth = 'standard', age
469
487
  ? specific
470
488
  : await _queryBlendedSearch(q, numResults, isBlended, requestedEngine);
471
489
 
490
+ // Fetch a wider candidate set when exclusions are in play, so filtering out
491
+ // already-known URLs still leaves `numResults` net-new results.
472
492
  _appendUniqueResults(currentResults, seenUrls, allResults);
473
493
  }
474
494
 
475
495
  allResults.sort((a, b) => rankScore(b) - rankScore(a));
476
- const formatted = allResults.slice(0, numResults).map((r, i) => ({
496
+ const fresh = excluded.size
497
+ ? allResults.filter((r) => !excluded.has(canonicalizeUrl(r.url) || r.url))
498
+ : allResults;
499
+ const formatted = fresh.slice(0, numResults).map((r, i) => ({
477
500
  citation_id: i + 1,
478
501
  title: r.title,
479
502
  url: r.url,
@@ -485,6 +508,7 @@ async function runWebSearch(query, numResults = 5, searchDepth = 'standard', age
485
508
  status: 'success',
486
509
  query,
487
510
  search_depth: searchDepth,
511
+ excluded_count: excluded.size,
488
512
  results_count: formatted.length,
489
513
  results: formatted
490
514
  });
@@ -150,7 +150,12 @@ function extractModifiedFiles(tasks) {
150
150
  function formatTaskSummary(tasks) {
151
151
  const list = (tasks || []).slice(0, 5);
152
152
  return list.map(t => {
153
- const res = (t.result || 'Completed').replace(/\n+/g, ' ').trim();
153
+ // Executor slices report `summary`; agent reports use `result`. Either may
154
+ // arrive as a non-string (object payload), which used to throw here and
155
+ // took the whole review node down.
156
+ const raw = typeof t.result === 'string' ? t.result
157
+ : (typeof t.summary === 'string' ? t.summary : 'Completed');
158
+ const res = raw.replace(/\n+/g, ' ').trim();
154
159
  const short = res.length > 60 ? res.slice(0, 57) + '...' : res;
155
160
  return `• [${t.agent || 'agent'}] ${t.task}: ${short}`;
156
161
  }).join('\n');