ultimate-pi 0.11.0 → 0.13.0
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.
- package/.agents/skills/ck-search/SKILL.md +11 -87
- package/.agents/skills/cocoindex-search/SKILL.md +35 -0
- package/.agents/skills/harness-debate-plan/SKILL.md +44 -0
- package/.agents/skills/harness-decisions/SKILL.md +1 -1
- package/.agents/skills/harness-orchestration/SKILL.md +54 -28
- package/.agents/skills/harness-plan/SKILL.md +15 -20
- package/.pi/PACKAGING.md +1 -0
- package/.pi/SYSTEM.md +21 -20
- package/.pi/agents/harness/adversary.md +0 -1
- package/.pi/agents/harness/evaluator.md +0 -1
- package/.pi/agents/harness/executor.md +1 -2
- package/.pi/agents/harness/incident-recorder.md +0 -1
- package/.pi/agents/harness/meta-optimizer.md +0 -1
- package/.pi/agents/harness/planning/decompose.md +3 -4
- package/.pi/agents/harness/planning/execution-plan-author.md +30 -0
- package/.pi/agents/harness/planning/hypothesis-validator.md +23 -0
- package/.pi/agents/harness/planning/hypothesis.md +3 -4
- package/.pi/agents/harness/planning/plan-adversary.md +10 -42
- package/.pi/agents/harness/planning/plan-evaluator.md +18 -0
- package/.pi/agents/harness/planning/review-integrator.md +23 -0
- package/.pi/agents/harness/planning/scout-graphify.md +13 -5
- package/.pi/agents/harness/planning/scout-semantic.md +23 -11
- package/.pi/agents/harness/planning/scout-structure.md +12 -6
- package/.pi/agents/harness/planning/sprint-contract-auditor.md +18 -0
- package/.pi/agents/harness/planning/stack-researcher.md +24 -0
- package/.pi/agents/harness/tie-breaker.md +0 -1
- package/.pi/agents/harness/trace-librarian.md +0 -1
- package/.pi/extensions/debate-orchestrator.ts +90 -53
- package/.pi/extensions/harness-plan-approval.ts +2 -2
- package/.pi/extensions/harness-run-context.ts +150 -5
- package/.pi/extensions/harness-subagents.ts +17 -6
- package/.pi/extensions/lib/harness-cocoindex-refresh.ts +49 -0
- package/.pi/extensions/lib/harness-posthog.ts +6 -1
- package/.pi/extensions/lib/harness-spawn-budget.ts +75 -0
- package/.pi/extensions/lib/harness-subagent-auth.ts +123 -0
- package/.pi/extensions/lib/{harness-subagents/harness-subagent-policy.ts → harness-subagent-policy.ts} +8 -7
- package/.pi/extensions/lib/harness-subagent-precheck.ts +95 -0
- package/.pi/extensions/lib/harness-subagents-bridge.ts +122 -0
- package/.pi/extensions/lib/plan-approval/create-plan.ts +4 -7
- package/.pi/extensions/lib/plan-approval/plan-review.ts +1 -1
- package/.pi/extensions/lib/plan-approval/types.ts +7 -1
- package/.pi/extensions/lib/plan-debate-envelope.ts +84 -0
- package/.pi/extensions/lib/{harness-subagents/spawn-policy.ts → spawn-policy.ts} +1 -0
- package/.pi/extensions/policy-gate.ts +1 -1
- package/.pi/extensions/review-integrity.ts +48 -29
- package/.pi/harness/agents.manifest.json +37 -25
- package/.pi/harness/docs/adrs/0032-harness-command-orchestration.md +4 -3
- package/.pi/harness/docs/adrs/0033-parent-orchestrated-planning.md +2 -2
- package/.pi/harness/docs/adrs/0035-plan-phase-review-gate.md +27 -0
- package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/artifacts/review-round-r1.yaml +25 -0
- package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/artifacts/review-round-r4.yaml +26 -0
- package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/artifacts/sprint-audit-r4.yaml +5 -0
- package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/plan-packet.yaml +196 -0
- package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/plan-review.md +14 -0
- package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/research-brief.yaml +32 -0
- package/.pi/harness/evals/smoke/run-context.fixture.json +1 -1
- package/.pi/harness/evals/smoke/smoke-harness-plan.mjs +88 -0
- package/.pi/harness/specs/harness-posthog-event.schema.json +6 -1
- package/.pi/harness/specs/plan-execution-plan-brief.schema.json +13 -0
- package/.pi/harness/specs/plan-execution-plan.schema.json +255 -0
- package/.pi/harness/specs/plan-packet.schema.json +14 -5
- package/.pi/harness/specs/plan-review-round-draft.schema.json +68 -0
- package/.pi/harness/specs/plan-sprint-audit-turn.schema.json +29 -0
- package/.pi/harness/specs/plan-stack-brief.schema.json +65 -0
- package/.pi/harness/specs/plan-validation-turn.schema.json +42 -0
- package/.pi/harness/specs/round-result.schema.json +16 -9
- package/.pi/lib/debate-orchestrator-types.ts +38 -0
- package/.pi/lib/harness-agent-discovery.mjs +81 -0
- package/.pi/lib/harness-run-context.ts +64 -38
- package/.pi/lib/harness-yaml.mjs +73 -0
- package/.pi/lib/harness-yaml.ts +90 -0
- package/.pi/prompts/harness-auto.md +13 -11
- package/.pi/prompts/harness-critic.md +2 -2
- package/.pi/prompts/harness-eval.md +3 -3
- package/.pi/prompts/harness-incident.md +2 -2
- package/.pi/prompts/harness-plan.md +83 -92
- package/.pi/prompts/harness-review.md +2 -2
- package/.pi/prompts/harness-router-tune.md +1 -1
- package/.pi/prompts/harness-run.md +2 -2
- package/.pi/prompts/harness-setup.md +30 -17
- package/.pi/prompts/harness-trace.md +2 -2
- package/.pi/scripts/README.md +1 -0
- package/.pi/scripts/harness-agents-manifest.mjs +1 -1
- package/.pi/scripts/harness-cli-verify.sh +24 -14
- package/.pi/scripts/harness-cocoindex-bootstrap.sh +182 -0
- package/.pi/scripts/harness-verify.mjs +38 -19
- package/.pi/scripts/validate-plan-dag.mjs +258 -0
- package/.pi/scripts/vendor-sync-pi-subagents.sh +19 -0
- package/.pi/skills/ast-grep/SKILL.md +2 -2
- package/.pi/skills/ccc/SKILL.md +142 -0
- package/.pi/skills/ccc/references/management.md +110 -0
- package/CHANGELOG.md +22 -0
- package/THIRD_PARTY_NOTICES.md +15 -0
- package/biome.json +2 -2
- package/package.json +7 -4
- package/vendor/pi-subagents/LICENSE +21 -0
- package/vendor/pi-subagents/UPSTREAM_PIN.md +11 -0
- package/vendor/pi-subagents/src/agents.ts +357 -0
- package/vendor/pi-subagents/src/subagents.ts +1463 -0
- package/.pi/agents/harness/planner.md +0 -13
- package/.pi/agents/harness/planning/hypothesis-eval.md +0 -59
- package/.pi/agents/harness/planning/planner.md +0 -20
- package/.pi/extensions/lib/harness-subagents/agent-loader.ts +0 -126
- package/.pi/extensions/lib/harness-subagents/agent-manifest.ts +0 -119
- package/.pi/extensions/lib/harness-subagents/agent-parser.ts +0 -87
- package/.pi/extensions/lib/harness-subagents/blackboard-tool.ts +0 -118
- package/.pi/extensions/lib/harness-subagents/blackboard.ts +0 -175
- package/.pi/extensions/lib/harness-subagents/parent-ask-user-bridge.ts +0 -10
- package/.pi/extensions/lib/harness-subagents/parent-harness-ui-bridge.ts +0 -137
- package/.pi/extensions/lib/harness-subagents/parent-harness-ui-hooks.ts +0 -77
- package/.pi/extensions/lib/harness-subagents/types-blackboard.ts +0 -27
- package/.pi/extensions/lib/harness-subagents/vendored/agent-manager.ts +0 -558
- package/.pi/extensions/lib/harness-subagents/vendored/agent-runner.ts +0 -666
- package/.pi/extensions/lib/harness-subagents/vendored/agent-types.ts +0 -175
- package/.pi/extensions/lib/harness-subagents/vendored/context.ts +0 -59
- package/.pi/extensions/lib/harness-subagents/vendored/cross-extension-rpc.ts +0 -134
- package/.pi/extensions/lib/harness-subagents/vendored/custom-agents.ts +0 -5
- package/.pi/extensions/lib/harness-subagents/vendored/default-agents.ts +0 -123
- package/.pi/extensions/lib/harness-subagents/vendored/env.ts +0 -43
- package/.pi/extensions/lib/harness-subagents/vendored/group-join.ts +0 -144
- package/.pi/extensions/lib/harness-subagents/vendored/index.ts +0 -2460
- package/.pi/extensions/lib/harness-subagents/vendored/invocation-config.ts +0 -52
- package/.pi/extensions/lib/harness-subagents/vendored/memory.ts +0 -182
- package/.pi/extensions/lib/harness-subagents/vendored/model-resolver.ts +0 -92
- package/.pi/extensions/lib/harness-subagents/vendored/output-file.ts +0 -115
- package/.pi/extensions/lib/harness-subagents/vendored/prompts.ts +0 -103
- package/.pi/extensions/lib/harness-subagents/vendored/schedule-store.ts +0 -177
- package/.pi/extensions/lib/harness-subagents/vendored/schedule.ts +0 -416
- package/.pi/extensions/lib/harness-subagents/vendored/settings.ts +0 -210
- package/.pi/extensions/lib/harness-subagents/vendored/skill-loader.ts +0 -108
- package/.pi/extensions/lib/harness-subagents/vendored/types.ts +0 -187
- package/.pi/extensions/lib/harness-subagents/vendored/ui/agent-widget.ts +0 -639
- package/.pi/extensions/lib/harness-subagents/vendored/ui/conversation-viewer.ts +0 -324
- package/.pi/extensions/lib/harness-subagents/vendored/ui/schedule-menu.ts +0 -110
- package/.pi/extensions/lib/harness-subagents/vendored/usage.ts +0 -71
- package/.pi/extensions/lib/harness-subagents/vendored/worktree.ts +0 -195
- /package/.pi/extensions/{00-ultimate-pi-system-prompt.ts → custom-system-prompt.ts} +0 -0
|
@@ -0,0 +1,122 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* ultimate-pi harness wrapper around vendored pi-subagents.
|
|
3
|
+
*/
|
|
4
|
+
|
|
5
|
+
import type {
|
|
6
|
+
ExtensionAPI,
|
|
7
|
+
ExtensionContext,
|
|
8
|
+
} from "@earendil-works/pi-coding-agent";
|
|
9
|
+
import type { AgentConfig } from "../../../vendor/pi-subagents/src/agents.js";
|
|
10
|
+
import {
|
|
11
|
+
createSubagentsExtension,
|
|
12
|
+
type HarnessSubagentsOptions,
|
|
13
|
+
type SpawnAuthForward,
|
|
14
|
+
} from "../../../vendor/pi-subagents/src/subagents.js";
|
|
15
|
+
import { refreshHarnessCocoindexIndex } from "./harness-cocoindex-refresh.js";
|
|
16
|
+
import { captureHarnessEvent } from "./harness-posthog.js";
|
|
17
|
+
import {
|
|
18
|
+
checkHarnessSpawnBudget,
|
|
19
|
+
countHarnessAgentsInRequest,
|
|
20
|
+
createSpawnBudgetState,
|
|
21
|
+
recordSpawnEnd,
|
|
22
|
+
recordSpawnStart,
|
|
23
|
+
} from "./harness-spawn-budget.js";
|
|
24
|
+
import {
|
|
25
|
+
isUsableApiKey,
|
|
26
|
+
resolveConcreteSubagentModel,
|
|
27
|
+
} from "./harness-subagent-auth.js";
|
|
28
|
+
import {
|
|
29
|
+
inferPhaseForPrecheck,
|
|
30
|
+
precheckHarnessSubagentSpawn,
|
|
31
|
+
} from "./harness-subagent-precheck.js";
|
|
32
|
+
|
|
33
|
+
const spawnBudget = createSpawnBudgetState();
|
|
34
|
+
let lastSessionId = "harness";
|
|
35
|
+
|
|
36
|
+
async function resolveHarnessSpawnAuth(
|
|
37
|
+
ctx: ExtensionContext,
|
|
38
|
+
agent: AgentConfig,
|
|
39
|
+
): Promise<SpawnAuthForward | undefined> {
|
|
40
|
+
const parentModel = ctx.model
|
|
41
|
+
? { provider: ctx.model.provider, id: ctx.model.id }
|
|
42
|
+
: undefined;
|
|
43
|
+
const concrete = resolveConcreteSubagentModel(ctx.cwd, parentModel, agent);
|
|
44
|
+
if (!concrete) {
|
|
45
|
+
return undefined;
|
|
46
|
+
}
|
|
47
|
+
const apiKey = await ctx.modelRegistry.getApiKeyForProvider(
|
|
48
|
+
concrete.provider,
|
|
49
|
+
);
|
|
50
|
+
if (!isUsableApiKey(apiKey)) return undefined;
|
|
51
|
+
return {
|
|
52
|
+
provider: concrete.provider,
|
|
53
|
+
modelRef: concrete.modelRef,
|
|
54
|
+
apiKey,
|
|
55
|
+
};
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
export function createHarnessSubagentsExtension(
|
|
59
|
+
packageRoot: string,
|
|
60
|
+
): (pi: ExtensionAPI) => void {
|
|
61
|
+
const options: HarnessSubagentsOptions = {
|
|
62
|
+
packageRoot,
|
|
63
|
+
defaultAgentScope: "both",
|
|
64
|
+
defaultConfirmProjectAgents: false,
|
|
65
|
+
truncateDetails: true,
|
|
66
|
+
resolveSpawnAuth: resolveHarnessSpawnAuth,
|
|
67
|
+
beforeExecute: async (params, agents, ctx) => {
|
|
68
|
+
lastSessionId = ctx.sessionManager.getSessionId();
|
|
69
|
+
const { harnessCount } = countHarnessAgentsInRequest(
|
|
70
|
+
params as Parameters<typeof countHarnessAgentsInRequest>[0],
|
|
71
|
+
);
|
|
72
|
+
if (harnessCount > 0) {
|
|
73
|
+
const budget = checkHarnessSpawnBudget(spawnBudget, harnessCount);
|
|
74
|
+
if (!budget.ok) {
|
|
75
|
+
return { ok: false, message: budget.message };
|
|
76
|
+
}
|
|
77
|
+
const phase = inferPhaseForPrecheck(ctx.sessionManager.getEntries());
|
|
78
|
+
const pre = precheckHarnessSubagentSpawn(
|
|
79
|
+
params as Parameters<typeof precheckHarnessSubagentSpawn>[0],
|
|
80
|
+
agents,
|
|
81
|
+
phase,
|
|
82
|
+
);
|
|
83
|
+
if (!pre.ok) {
|
|
84
|
+
return { ok: false, message: pre.message };
|
|
85
|
+
}
|
|
86
|
+
if (phase === "plan" || phase === "execute") {
|
|
87
|
+
const refreshMsg = refreshHarnessCocoindexIndex(ctx.cwd);
|
|
88
|
+
if (refreshMsg?.includes("continuing")) {
|
|
89
|
+
// warn-only path; do not block spawn
|
|
90
|
+
} else if (refreshMsg) {
|
|
91
|
+
return { ok: false, message: refreshMsg };
|
|
92
|
+
}
|
|
93
|
+
}
|
|
94
|
+
}
|
|
95
|
+
return { ok: true };
|
|
96
|
+
},
|
|
97
|
+
onSpawnStart: (harnessCount) => {
|
|
98
|
+
if (harnessCount <= 0) return;
|
|
99
|
+
recordSpawnStart(spawnBudget, harnessCount);
|
|
100
|
+
captureHarnessEvent(lastSessionId, "harness_subagent_spawned", {
|
|
101
|
+
active_after: spawnBudget.active,
|
|
102
|
+
spawn_count: harnessCount,
|
|
103
|
+
});
|
|
104
|
+
},
|
|
105
|
+
onSpawnEnd: (harnessCount) => {
|
|
106
|
+
if (harnessCount <= 0) return;
|
|
107
|
+
recordSpawnEnd(spawnBudget, harnessCount);
|
|
108
|
+
},
|
|
109
|
+
onCompleted: ({ agents, mode, durationMs }) => {
|
|
110
|
+
if (agents.length === 0) return;
|
|
111
|
+
captureHarnessEvent(lastSessionId, "harness_subagent_completed", {
|
|
112
|
+
mode,
|
|
113
|
+
duration_ms: durationMs,
|
|
114
|
+
agent_count: agents.length,
|
|
115
|
+
});
|
|
116
|
+
},
|
|
117
|
+
};
|
|
118
|
+
|
|
119
|
+
return (pi: ExtensionAPI) => {
|
|
120
|
+
createSubagentsExtension(pi, options);
|
|
121
|
+
};
|
|
122
|
+
}
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { mkdir
|
|
1
|
+
import { mkdir } from "node:fs/promises";
|
|
2
2
|
import { dirname, resolve } from "node:path";
|
|
3
3
|
import {
|
|
4
4
|
canonicalPlanPath,
|
|
@@ -9,6 +9,7 @@ import {
|
|
|
9
9
|
saveRunContextToDisk,
|
|
10
10
|
validatePlanPacket,
|
|
11
11
|
} from "../../../lib/harness-run-context.js";
|
|
12
|
+
import { writeYamlFile } from "../../../lib/harness-yaml.js";
|
|
12
13
|
import { writePlanReviewMarkdown } from "./plan-review.js";
|
|
13
14
|
|
|
14
15
|
export const CREATE_PLAN_SNIPPET =
|
|
@@ -17,7 +18,7 @@ export const CREATE_PLAN_SNIPPET =
|
|
|
17
18
|
export const CREATE_PLAN_GUIDELINES = [
|
|
18
19
|
"Call create_plan only after the user approves via approve_plan (Approve selection).",
|
|
19
20
|
"Pass the same plan_packet you showed in approve_plan — path is resolved automatically.",
|
|
20
|
-
"Never use write or edit for plan-packet.
|
|
21
|
+
"Never use write or edit for plan-packet.yaml; create_plan is the only allowed plan write.",
|
|
21
22
|
];
|
|
22
23
|
|
|
23
24
|
export interface CreatePlanDeps {
|
|
@@ -89,11 +90,7 @@ export async function executeCreatePlan(
|
|
|
89
90
|
|
|
90
91
|
try {
|
|
91
92
|
await mkdir(dirname(planPath), { recursive: true });
|
|
92
|
-
await
|
|
93
|
-
planPath,
|
|
94
|
-
`${JSON.stringify(planPacket, null, 2)}\n`,
|
|
95
|
-
"utf-8",
|
|
96
|
-
);
|
|
93
|
+
await writeYamlFile(planPath, planPacket);
|
|
97
94
|
} catch (err) {
|
|
98
95
|
const msg = err instanceof Error ? err.message : String(err);
|
|
99
96
|
return { ok: false, error: `create_plan: write failed — ${msg}` };
|
|
@@ -245,7 +245,7 @@ export function formatPlanPacketMarkdown(
|
|
|
245
245
|
);
|
|
246
246
|
} else if (status === "approved") {
|
|
247
247
|
lines.push(
|
|
248
|
-
"Approved in the harness TUI. Waiting for `create_plan` to write `plan-packet.
|
|
248
|
+
"Approved in the harness TUI. Waiting for `create_plan` to write `plan-packet.yaml`, or run `/harness-plan-commit` if that step failed.",
|
|
249
249
|
);
|
|
250
250
|
} else {
|
|
251
251
|
lines.push(
|
|
@@ -7,11 +7,17 @@ export const DEFAULT_PLAN_APPROVAL_OPTIONS = [
|
|
|
7
7
|
"Cancel",
|
|
8
8
|
] as const;
|
|
9
9
|
|
|
10
|
-
/** Optional Darwin research artifacts from /harness-plan (not
|
|
10
|
+
/** Optional Darwin research artifacts from /harness-plan (research-brief.yaml, not in plan-packet). */
|
|
11
11
|
export interface PlanResearchBrief {
|
|
12
12
|
decomposition?: Record<string, unknown> | null;
|
|
13
13
|
hypothesis?: Record<string, unknown> | null;
|
|
14
14
|
eval?: Record<string, unknown> | null;
|
|
15
|
+
stack?: Record<string, unknown> | null;
|
|
16
|
+
debate?: {
|
|
17
|
+
rounds?: Record<string, unknown>[];
|
|
18
|
+
hypothesis_validations?: Record<string, unknown>[];
|
|
19
|
+
} | null;
|
|
20
|
+
dag_validation?: Record<string, unknown> | null;
|
|
15
21
|
}
|
|
16
22
|
|
|
17
23
|
export interface ApprovePlanParams {
|
|
@@ -0,0 +1,84 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Plan Review Gate — convert integrator YAML to debate bus round JSON.
|
|
3
|
+
*/
|
|
4
|
+
|
|
5
|
+
import type { DebateParticipant } from "../../lib/debate-orchestrator-types.js";
|
|
6
|
+
|
|
7
|
+
export interface PlanReviewRoundDraft {
|
|
8
|
+
schema_version: string;
|
|
9
|
+
round_index: number;
|
|
10
|
+
debate_round_focus?: string;
|
|
11
|
+
round_summary?: string;
|
|
12
|
+
validation_summary?: string;
|
|
13
|
+
adversary_summary?: string;
|
|
14
|
+
disputes?: string[];
|
|
15
|
+
recommended_packet_patches?: Array<{ path: string; value: unknown }>;
|
|
16
|
+
review_gate_ready?: boolean;
|
|
17
|
+
participants?: DebateParticipant[];
|
|
18
|
+
claims?: string[];
|
|
19
|
+
rebuttals?: string[];
|
|
20
|
+
evidence_refs?: string[];
|
|
21
|
+
token_usage?: {
|
|
22
|
+
per_agent: Record<string, number>;
|
|
23
|
+
round_total: number;
|
|
24
|
+
};
|
|
25
|
+
consensus_delta?: number;
|
|
26
|
+
severity_scores?: {
|
|
27
|
+
correctness: number;
|
|
28
|
+
security: number;
|
|
29
|
+
architecture: number;
|
|
30
|
+
test_integrity: number;
|
|
31
|
+
};
|
|
32
|
+
}
|
|
33
|
+
|
|
34
|
+
export function buildPlanReviewRoundEnvelope(
|
|
35
|
+
draft: PlanReviewRoundDraft,
|
|
36
|
+
opts: { runId: string; debateId: string },
|
|
37
|
+
): {
|
|
38
|
+
protocol: "pi-debate-bus/v1";
|
|
39
|
+
kind: "round";
|
|
40
|
+
correlation: {
|
|
41
|
+
run_id: string;
|
|
42
|
+
debate_id: string;
|
|
43
|
+
round_index: number;
|
|
44
|
+
sender: DebateParticipant;
|
|
45
|
+
};
|
|
46
|
+
payload: {
|
|
47
|
+
participants: DebateParticipant[];
|
|
48
|
+
claims: string[];
|
|
49
|
+
rebuttals: string[];
|
|
50
|
+
evidence_refs: string[];
|
|
51
|
+
token_usage: { per_agent: Record<string, number>; round_total: number };
|
|
52
|
+
consensus_delta: number;
|
|
53
|
+
severity_scores?: PlanReviewRoundDraft["severity_scores"];
|
|
54
|
+
};
|
|
55
|
+
} {
|
|
56
|
+
const participants = (draft.participants ?? [
|
|
57
|
+
"PlanEvaluatorAgent",
|
|
58
|
+
"PlanAdversaryAgent",
|
|
59
|
+
"ReviewIntegratorAgent",
|
|
60
|
+
]) as DebateParticipant[];
|
|
61
|
+
|
|
62
|
+
return {
|
|
63
|
+
protocol: "pi-debate-bus/v1",
|
|
64
|
+
kind: "round",
|
|
65
|
+
correlation: {
|
|
66
|
+
run_id: opts.runId,
|
|
67
|
+
debate_id: opts.debateId,
|
|
68
|
+
round_index: draft.round_index,
|
|
69
|
+
sender: "ReviewIntegratorAgent",
|
|
70
|
+
},
|
|
71
|
+
payload: {
|
|
72
|
+
participants,
|
|
73
|
+
claims: draft.claims ?? [draft.round_summary ?? "review round"],
|
|
74
|
+
rebuttals: draft.rebuttals ?? draft.disputes ?? [],
|
|
75
|
+
evidence_refs: draft.evidence_refs ?? [],
|
|
76
|
+
token_usage: draft.token_usage ?? {
|
|
77
|
+
per_agent: { ReviewIntegratorAgent: 0 },
|
|
78
|
+
round_total: 0,
|
|
79
|
+
},
|
|
80
|
+
consensus_delta: draft.consensus_delta ?? 0,
|
|
81
|
+
severity_scores: draft.severity_scores,
|
|
82
|
+
},
|
|
83
|
+
};
|
|
84
|
+
}
|
|
@@ -243,7 +243,7 @@ export default function policyGate(pi: ExtensionAPI) {
|
|
|
243
243
|
|
|
244
244
|
const planPhaseHint =
|
|
245
245
|
state.phase === "plan"
|
|
246
|
-
? "\nPlan phase: scouts → decompose → hypothesis
|
|
246
|
+
? "\nPlan phase: scouts → decompose → hypothesis → stack-researcher → execution-plan-author → validate-plan-dag → 4-round plan debate → approve_plan → create_plan (YAML plan-packet.yaml). Post-execute: /harness-critic."
|
|
247
247
|
: "";
|
|
248
248
|
|
|
249
249
|
return {
|
|
@@ -1,9 +1,8 @@
|
|
|
1
1
|
/**
|
|
2
2
|
* review-integrity — enforce evaluator/adversary isolation from executor session.
|
|
3
3
|
*
|
|
4
|
-
* Parent orchestrators spawn review agents in isolated subagent
|
|
5
|
-
* Direct review tools in the executor session are blocked
|
|
6
|
-
* for harness review agents remain allowed.
|
|
4
|
+
* Parent orchestrators spawn review agents in isolated subprocesses via `subagent`.
|
|
5
|
+
* Direct review tools in the executor session are blocked.
|
|
7
6
|
*/
|
|
8
7
|
|
|
9
8
|
import { appendFile, mkdir } from "node:fs/promises";
|
|
@@ -15,12 +14,6 @@ type HarnessPhase = "plan" | "execute" | "evaluate" | "adversary" | "merge";
|
|
|
15
14
|
const INCIDENTS_DIR = join(process.cwd(), ".pi", "harness", "incidents");
|
|
16
15
|
const INCIDENT_FILE = join(INCIDENTS_DIR, "review-integrity.jsonl");
|
|
17
16
|
|
|
18
|
-
const ORCHESTRATION_TOOLS = new Set([
|
|
19
|
-
"Agent",
|
|
20
|
-
"get_subagent_result",
|
|
21
|
-
"steer_subagent",
|
|
22
|
-
]);
|
|
23
|
-
|
|
24
17
|
const REVIEW_SUBAGENT_TYPES = new Set([
|
|
25
18
|
"harness/evaluator",
|
|
26
19
|
"harness/adversary",
|
|
@@ -104,15 +97,45 @@ function restoreState(ctx: {
|
|
|
104
97
|
};
|
|
105
98
|
}
|
|
106
99
|
|
|
107
|
-
function
|
|
100
|
+
function agentsFromSubagentInput(
|
|
108
101
|
input: Record<string, unknown> | undefined,
|
|
109
|
-
): string {
|
|
110
|
-
if (!input) return
|
|
111
|
-
const
|
|
112
|
-
if (typeof
|
|
113
|
-
const
|
|
114
|
-
if (
|
|
115
|
-
|
|
102
|
+
): string[] {
|
|
103
|
+
if (!input) return [];
|
|
104
|
+
const names: string[] = [];
|
|
105
|
+
if (typeof input.agent === "string") names.push(input.agent);
|
|
106
|
+
const tasks = input.tasks;
|
|
107
|
+
if (Array.isArray(tasks)) {
|
|
108
|
+
for (const t of tasks) {
|
|
109
|
+
if (
|
|
110
|
+
t &&
|
|
111
|
+
typeof t === "object" &&
|
|
112
|
+
typeof (t as { agent?: string }).agent === "string"
|
|
113
|
+
) {
|
|
114
|
+
names.push((t as { agent: string }).agent);
|
|
115
|
+
}
|
|
116
|
+
}
|
|
117
|
+
}
|
|
118
|
+
const chain = input.chain;
|
|
119
|
+
if (Array.isArray(chain)) {
|
|
120
|
+
for (const c of chain) {
|
|
121
|
+
if (
|
|
122
|
+
c &&
|
|
123
|
+
typeof c === "object" &&
|
|
124
|
+
typeof (c as { agent?: string }).agent === "string"
|
|
125
|
+
) {
|
|
126
|
+
names.push((c as { agent: string }).agent);
|
|
127
|
+
}
|
|
128
|
+
}
|
|
129
|
+
}
|
|
130
|
+
const agg = input.aggregator;
|
|
131
|
+
if (
|
|
132
|
+
agg &&
|
|
133
|
+
typeof agg === "object" &&
|
|
134
|
+
typeof (agg as { agent?: string }).agent === "string"
|
|
135
|
+
) {
|
|
136
|
+
names.push((agg as { agent: string }).agent);
|
|
137
|
+
}
|
|
138
|
+
return names;
|
|
116
139
|
}
|
|
117
140
|
|
|
118
141
|
async function appendIncident(payload: Record<string, unknown>): Promise<void> {
|
|
@@ -178,26 +201,26 @@ export default function reviewIntegrity(pi: ExtensionAPI) {
|
|
|
178
201
|
customType: "harness-review-integrity-hint",
|
|
179
202
|
display: true,
|
|
180
203
|
content: [
|
|
181
|
-
"Review phase in executor session: spawn harness/evaluator or harness/adversary via
|
|
182
|
-
"Do not run review checks directly in this session
|
|
204
|
+
"Review phase in executor session: spawn harness/evaluator or harness/adversary via subagent (isolated subprocess).",
|
|
205
|
+
"Do not run review checks directly in this session.",
|
|
183
206
|
].join("\n"),
|
|
184
207
|
},
|
|
185
208
|
};
|
|
186
209
|
});
|
|
187
210
|
|
|
188
211
|
pi.on("tool_call", async (event, ctx) => {
|
|
189
|
-
if (event.toolName === "
|
|
190
|
-
const
|
|
212
|
+
if (event.toolName === "subagent") {
|
|
213
|
+
const agents = agentsFromSubagentInput(
|
|
191
214
|
event.input as Record<string, unknown> | undefined,
|
|
192
215
|
);
|
|
193
|
-
if (
|
|
216
|
+
if (agents.includes(EXECUTOR_SUBAGENT_TYPE)) {
|
|
194
217
|
state.executorSessionId = ctx.sessionManager.getSessionId();
|
|
195
218
|
state.violationActive = false;
|
|
196
219
|
state.updatedAt = nowIso();
|
|
197
220
|
persist();
|
|
198
221
|
return undefined;
|
|
199
222
|
}
|
|
200
|
-
if (REVIEW_SUBAGENT_TYPES.has(
|
|
223
|
+
if (agents.some((a) => REVIEW_SUBAGENT_TYPES.has(a))) {
|
|
201
224
|
state.violationActive = false;
|
|
202
225
|
state.updatedAt = nowIso();
|
|
203
226
|
persist();
|
|
@@ -207,10 +230,6 @@ export default function reviewIntegrity(pi: ExtensionAPI) {
|
|
|
207
230
|
|
|
208
231
|
if (!state.violationActive) return undefined;
|
|
209
232
|
|
|
210
|
-
if (ORCHESTRATION_TOOLS.has(event.toolName)) {
|
|
211
|
-
return undefined;
|
|
212
|
-
}
|
|
213
|
-
|
|
214
233
|
await appendIncident({
|
|
215
234
|
type: "review_integrity_violation",
|
|
216
235
|
session_id: ctx.sessionManager.getSessionId(),
|
|
@@ -218,13 +237,13 @@ export default function reviewIntegrity(pi: ExtensionAPI) {
|
|
|
218
237
|
reason:
|
|
219
238
|
"direct tool use in review phase while sharing executor session context",
|
|
220
239
|
mitigation:
|
|
221
|
-
"spawn harness/evaluator or harness/adversary via
|
|
240
|
+
"spawn harness/evaluator or harness/adversary via subagent instead",
|
|
222
241
|
});
|
|
223
242
|
|
|
224
243
|
return {
|
|
225
244
|
block: true,
|
|
226
245
|
reason:
|
|
227
|
-
"review-integrity: tool blocked in review phase — spawn an isolated review subagent via
|
|
246
|
+
"review-integrity: tool blocked in review phase — spawn an isolated review subagent via subagent.",
|
|
228
247
|
};
|
|
229
248
|
});
|
|
230
249
|
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
{
|
|
2
2
|
"schema_version": "1.0.0",
|
|
3
3
|
"package": "ultimate-pi",
|
|
4
|
-
"package_version": "0.
|
|
5
|
-
"generated_at": "2026-05-
|
|
4
|
+
"package_version": "0.11.0",
|
|
5
|
+
"generated_at": "2026-05-18T09:43:44.563Z",
|
|
6
6
|
"agents": {
|
|
7
7
|
"pi-pi/agent-expert": {
|
|
8
8
|
"path": ".pi/agents/pi-pi/agent-expert.md",
|
|
@@ -46,27 +46,23 @@
|
|
|
46
46
|
},
|
|
47
47
|
"harness/adversary": {
|
|
48
48
|
"path": ".pi/agents/harness/adversary.md",
|
|
49
|
-
"sha256": "
|
|
49
|
+
"sha256": "dd2ef87139cb175d795f4d7bde78dca1a181d2e42c3c3bd0d48832cf5069aa29"
|
|
50
50
|
},
|
|
51
51
|
"harness/evaluator": {
|
|
52
52
|
"path": ".pi/agents/harness/evaluator.md",
|
|
53
|
-
"sha256": "
|
|
53
|
+
"sha256": "2b8039fd79f9177fdafd5319a53a96812719d4f1f68e2de70632030142649cfe"
|
|
54
54
|
},
|
|
55
55
|
"harness/executor": {
|
|
56
56
|
"path": ".pi/agents/harness/executor.md",
|
|
57
|
-
"sha256": "
|
|
57
|
+
"sha256": "b549e9fc802ba23857a1bc6b2ff36f3c169e708fe5ec13857b3bcfe841384f1f"
|
|
58
58
|
},
|
|
59
59
|
"harness/incident-recorder": {
|
|
60
60
|
"path": ".pi/agents/harness/incident-recorder.md",
|
|
61
|
-
"sha256": "
|
|
61
|
+
"sha256": "d7577c911a9e6c9607eb64f76337aab85c4eb9a92e7cd917eb8d989ef3cd1de5"
|
|
62
62
|
},
|
|
63
63
|
"harness/meta-optimizer": {
|
|
64
64
|
"path": ".pi/agents/harness/meta-optimizer.md",
|
|
65
|
-
"sha256": "
|
|
66
|
-
},
|
|
67
|
-
"harness/planner": {
|
|
68
|
-
"path": ".pi/agents/harness/planner.md",
|
|
69
|
-
"sha256": "648b9e4c56a6c0f983ae990238952579bb4745af81eb6c79add5325ea4929c91"
|
|
65
|
+
"sha256": "a4eed88084c7cfb5ace3edc72b72d7ead4134b3eae0d444b391decfe2640a632"
|
|
70
66
|
},
|
|
71
67
|
"harness/sentrux-bootstrap": {
|
|
72
68
|
"path": ".pi/agents/harness/sentrux-bootstrap.md",
|
|
@@ -74,43 +70,59 @@
|
|
|
74
70
|
},
|
|
75
71
|
"harness/tie-breaker": {
|
|
76
72
|
"path": ".pi/agents/harness/tie-breaker.md",
|
|
77
|
-
"sha256": "
|
|
73
|
+
"sha256": "68f02b86e95927f06d7f963e1f61f193159bbef1ba4558d90c84d5457d62b3f7"
|
|
78
74
|
},
|
|
79
75
|
"harness/trace-librarian": {
|
|
80
76
|
"path": ".pi/agents/harness/trace-librarian.md",
|
|
81
|
-
"sha256": "
|
|
77
|
+
"sha256": "03b499a948b8467f1cfe2b4e63190feb7b8b9d96461055638e774253b9b6b2d4"
|
|
82
78
|
},
|
|
83
79
|
"harness/planning/decompose": {
|
|
84
80
|
"path": ".pi/agents/harness/planning/decompose.md",
|
|
85
|
-
"sha256": "
|
|
81
|
+
"sha256": "1b3f85d956d2e203ec87045a731c47f8b40f75b63fce8916fda91cefc39244a8"
|
|
82
|
+
},
|
|
83
|
+
"harness/planning/execution-plan-author": {
|
|
84
|
+
"path": ".pi/agents/harness/planning/execution-plan-author.md",
|
|
85
|
+
"sha256": "a69fb2e8bda9336e71ce9536071f9c8a2f4abd9d9d88930c6a8be29bdc9c5f62"
|
|
86
86
|
},
|
|
87
|
-
"harness/planning/hypothesis-
|
|
88
|
-
"path": ".pi/agents/harness/planning/hypothesis-
|
|
89
|
-
"sha256": "
|
|
87
|
+
"harness/planning/hypothesis-validator": {
|
|
88
|
+
"path": ".pi/agents/harness/planning/hypothesis-validator.md",
|
|
89
|
+
"sha256": "f75312439c441ccee72692d41f44b6e733df08e06c89e930740fc256bed3ba02"
|
|
90
90
|
},
|
|
91
91
|
"harness/planning/hypothesis": {
|
|
92
92
|
"path": ".pi/agents/harness/planning/hypothesis.md",
|
|
93
|
-
"sha256": "
|
|
93
|
+
"sha256": "b20c527d15c2243cd5d3a8f16cea6d44bdfd16e01915d42f3b830bf9938e5f8b"
|
|
94
94
|
},
|
|
95
95
|
"harness/planning/plan-adversary": {
|
|
96
96
|
"path": ".pi/agents/harness/planning/plan-adversary.md",
|
|
97
|
-
"sha256": "
|
|
97
|
+
"sha256": "685926c638ae1377361d7cafda5e400be19cb3880510d8f6d389a5876647575f"
|
|
98
|
+
},
|
|
99
|
+
"harness/planning/plan-evaluator": {
|
|
100
|
+
"path": ".pi/agents/harness/planning/plan-evaluator.md",
|
|
101
|
+
"sha256": "44fd52389d7e43dd5093653cba9694900561318ee5f00e3bc05c3ecef5d43621"
|
|
98
102
|
},
|
|
99
|
-
"harness/planning/
|
|
100
|
-
"path": ".pi/agents/harness/planning/
|
|
101
|
-
"sha256": "
|
|
103
|
+
"harness/planning/review-integrator": {
|
|
104
|
+
"path": ".pi/agents/harness/planning/review-integrator.md",
|
|
105
|
+
"sha256": "d0e8214539d0a78b9e5add70e61dd4e4de36def64172cda18d9b70727e7600ca"
|
|
102
106
|
},
|
|
103
107
|
"harness/planning/scout-graphify": {
|
|
104
108
|
"path": ".pi/agents/harness/planning/scout-graphify.md",
|
|
105
|
-
"sha256": "
|
|
109
|
+
"sha256": "b59916a26afccfe105e29c0bd8637ac54275e8afef1c6cc88a58bd05b0325473"
|
|
106
110
|
},
|
|
107
111
|
"harness/planning/scout-semantic": {
|
|
108
112
|
"path": ".pi/agents/harness/planning/scout-semantic.md",
|
|
109
|
-
"sha256": "
|
|
113
|
+
"sha256": "47b7ea3e65b20a65e6d0ff11b6d5daff59b47a9ed618b8a3b6282f2eb0460572"
|
|
110
114
|
},
|
|
111
115
|
"harness/planning/scout-structure": {
|
|
112
116
|
"path": ".pi/agents/harness/planning/scout-structure.md",
|
|
113
|
-
"sha256": "
|
|
117
|
+
"sha256": "e67b7cd75519e5ae36e1bb5f49ca158888c28d365465863aee50a9b2e8e5b7d7"
|
|
118
|
+
},
|
|
119
|
+
"harness/planning/sprint-contract-auditor": {
|
|
120
|
+
"path": ".pi/agents/harness/planning/sprint-contract-auditor.md",
|
|
121
|
+
"sha256": "f613a4fa937d76936fa01155d4e7956a81878f300100f99f6a78915b0af6f7c7"
|
|
122
|
+
},
|
|
123
|
+
"harness/planning/stack-researcher": {
|
|
124
|
+
"path": ".pi/agents/harness/planning/stack-researcher.md",
|
|
125
|
+
"sha256": "90e2ff1348f54bebc8c0392407bf1bb4d794c942fd8d6f342d80b191c945b34e"
|
|
114
126
|
}
|
|
115
127
|
}
|
|
116
128
|
}
|
|
@@ -5,14 +5,14 @@
|
|
|
5
5
|
|
|
6
6
|
## Context
|
|
7
7
|
|
|
8
|
-
Harness slash prompts duplicated logic already defined in `harness/*` agents.
|
|
8
|
+
Harness slash prompts duplicated logic already defined in `harness/*` agents. The in-process `Agent` / `createAgentSession` stack was heavy and unstable. Review docs told users to fork a new Pi session even though subprocess subagents already provide isolation.
|
|
9
9
|
|
|
10
10
|
## Decision
|
|
11
11
|
|
|
12
12
|
1. **Slash commands** (prompt templates) are orchestrators: spawn `harness/*` agents once, perform policy-gated writes, emit handoff blocks. Command identity is captured on Pi **`input`** as `harness-turn` (raw `/harness-*`), not from expanded prompt markdown.
|
|
13
13
|
2. **Agents** perform multi-turn reads and emit structured JSON drafts. **Planning** (`harness/planning/*`) scouts and plan-adversary are read-only; parent orchestrator runs `ask_user`, `approve_plan`, and `create_plan` (see ADR 0033).
|
|
14
14
|
3. **HarnessSpawnContext** is injected in `[HarnessRunContext]`; orchestrator copies it into spawn prompts. Subagents do not receive `[HarnessActivePlan]` injection.
|
|
15
|
-
4. **Review isolation** uses `
|
|
15
|
+
4. **Review isolation** uses native `subagent` (vendored pi-subagents: isolated `pi --mode json` subprocess). `review-integrity` allows `subagent` when `agent` is evaluator/adversary/tie-breaker; bridge blocks plan-phase mutating spawns and nested `subagent` in children.
|
|
16
16
|
5. **Subagent policy** blocks mutating tools for read-only phase agents; `ask_user` bridged for evaluator/adversary/tie-breaker only (not planning scouts).
|
|
17
17
|
6. **Parent** owns plan-phase `ask_user`, `approve_plan`, and `create_plan` per ADR 0033.
|
|
18
18
|
|
|
@@ -32,6 +32,7 @@ Harness slash prompts duplicated logic already defined in `harness/*` agents. Co
|
|
|
32
32
|
|
|
33
33
|
- `.pi/prompts/harness-*.md`
|
|
34
34
|
- `.pi/agents/harness/*.md`
|
|
35
|
-
- `.pi/extensions/lib/harness-subagents
|
|
35
|
+
- `vendor/pi-subagents/src/subagents.ts`, `.pi/extensions/lib/harness-subagents-bridge.ts`
|
|
36
|
+
- `.pi/extensions/lib/harness-subagent-policy.ts`
|
|
36
37
|
- `.pi/extensions/review-integrity.ts`
|
|
37
38
|
- `.pi/lib/harness-agent-output.ts`
|
|
@@ -5,7 +5,7 @@
|
|
|
5
5
|
|
|
6
6
|
## Context
|
|
7
7
|
|
|
8
|
-
`/harness-plan` delegated the full plan lifecycle to a single `harness/planner` subagent. Plans and approval UI were largely invisible in the parent transcript until
|
|
8
|
+
`/harness-plan` previously delegated the full plan lifecycle to a single `harness/planner` subagent. Plans and approval UI were largely invisible in the parent transcript until subprocess completion, and the orchestrator could not call `ask_user` / `approve_plan` / `create_plan` directly.
|
|
9
9
|
|
|
10
10
|
## Decision
|
|
11
11
|
|
|
@@ -19,7 +19,7 @@
|
|
|
19
19
|
### Positive
|
|
20
20
|
|
|
21
21
|
- Full plan visible in parent session; editor `plan-review.md` path surfaced from parent.
|
|
22
|
-
- Specialized scouts (graphify, ast-grep,
|
|
22
|
+
- Specialized scouts (graphify, ast-grep, cocoindex `ccc`) run in parallel with clear JSON contracts.
|
|
23
23
|
- Pre-approval adversary separate from post-run `harness/adversary`.
|
|
24
24
|
|
|
25
25
|
### Negative
|
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
# ADR-0035: Plan-phase Review Gate and YAML artifacts
|
|
2
|
+
|
|
3
|
+
## Status
|
|
4
|
+
|
|
5
|
+
Accepted (2026-05-18)
|
|
6
|
+
|
|
7
|
+
## Context
|
|
8
|
+
|
|
9
|
+
`/harness-plan` produced thin PlanPackets (scope + bullets). Post-execute adversarial review (`/harness-critic`) ran too late. Graphify corpus (Structured Planning, ADR-020, Generator–Evaluator) defines WBS, validation, and review gate before baseline.
|
|
10
|
+
|
|
11
|
+
## Decision
|
|
12
|
+
|
|
13
|
+
1. **PlanPacket 1.1.0** — required `execution_plan` (phases, work_items, sprint_contract, dag_validation).
|
|
14
|
+
2. **YAML on disk** — `plan-packet.yaml`, `research-brief.yaml`, `run-context.yaml`, `artifacts/*.yaml`. JSON Schema unchanged; instances validated after YAML parse.
|
|
15
|
+
3. **Review Gate agents** — `stack-researcher`, `execution-plan-author`, debate: `hypothesis-validator`, `plan-evaluator`, `plan-adversary`, `sprint-contract-auditor`, `review-integrator`.
|
|
16
|
+
4. **Debate bus** — `debate_id=plan-<run_id>`, plan budget profile (4 rounds, 12k cap), plan-phase consensus prerequisites.
|
|
17
|
+
5. **No legacy JSON** plan paths; no pre-debate standalone `hypothesis-eval`.
|
|
18
|
+
|
|
19
|
+
## Consequences
|
|
20
|
+
|
|
21
|
+
- Positive: PM-grade plans, deterministic DAG gate, blind hypothesis eval in debate R1.
|
|
22
|
+
- Negative: Higher spawn/token cost; `harness-verify` and smoke fixtures must use `.yaml`.
|
|
23
|
+
|
|
24
|
+
## References
|
|
25
|
+
|
|
26
|
+
- [ADR-0033](0033-parent-orchestrated-planning.md), [ADR-0034](0034-darwin-plan-research-pipeline.md)
|
|
27
|
+
- `raw/decisions/adr-020.md`, `raw/modules/structured-planning.md`
|
package/.pi/harness/evals/smoke/fixtures/plan-phase/minimal-med/artifacts/review-round-r1.yaml
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
1
|
+
schema_version: "1.0.0"
|
|
2
|
+
round_index: 1
|
|
3
|
+
debate_round_focus: spec
|
|
4
|
+
round_summary: Spec round passed for fixture
|
|
5
|
+
validation_summary: All spec checks pass
|
|
6
|
+
adversary_summary: No blocking adversarial findings
|
|
7
|
+
disputes: []
|
|
8
|
+
recommended_packet_patches: []
|
|
9
|
+
review_gate_ready: true
|
|
10
|
+
participants:
|
|
11
|
+
- PlanEvaluatorAgent
|
|
12
|
+
- PlanAdversaryAgent
|
|
13
|
+
- HypothesisValidatorAgent
|
|
14
|
+
- ReviewIntegratorAgent
|
|
15
|
+
claims:
|
|
16
|
+
- spec validation complete
|
|
17
|
+
rebuttals: []
|
|
18
|
+
evidence_refs: []
|
|
19
|
+
token_usage:
|
|
20
|
+
per_agent:
|
|
21
|
+
PlanEvaluatorAgent: 100
|
|
22
|
+
PlanAdversaryAgent: 100
|
|
23
|
+
ReviewIntegratorAgent: 50
|
|
24
|
+
round_total: 250
|
|
25
|
+
consensus_delta: 0.1
|