zob-harness 0.9.0 → 0.9.2
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/.pi/capabilities/zob-public-runtime-capabilities.json +16 -17
- package/.pi/extensions/zob-harness/index.ts +1 -1
- package/.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime/dry-run.ts +1107 -0
- package/.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime/report-writers.ts +325 -0
- package/.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime/smoke-run.ts +1286 -0
- package/.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime/types.ts +30 -0
- package/.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime/validation.ts +184 -0
- package/.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime.ts +4 -2912
- package/.pi/extensions/zob-harness/src/domains/compute/compute-profile.ts +2 -1
- package/.pi/extensions/zob-harness/src/domains/coms/coms-v2/registry.ts +24 -3
- package/.pi/extensions/zob-harness/src/domains/coms/coms-v2/types.ts +1 -0
- package/.pi/extensions/zob-harness/src/domains/coms/coms-v2/zpeer.ts +5 -3
- package/.pi/extensions/zob-harness/src/domains/delegation/child-runner.ts +28 -3
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos/constants.ts +19 -0
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos/formatting.ts +148 -0
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos/normalize.ts +476 -0
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos/operations.ts +393 -0
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos/parsing.ts +277 -0
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos/reducer.ts +110 -0
- package/.pi/extensions/zob-harness/src/domains/goal/goal-todos.ts +6 -1429
- package/.pi/extensions/zob-harness/src/domains/governance/governed-requests.ts +3 -1
- package/.pi/extensions/zob-harness/src/domains/governance/merge-queue.ts +3 -1
- package/.pi/extensions/zob-harness/src/domains/governance/sandbox/helpers.ts +124 -0
- package/.pi/extensions/zob-harness/src/domains/governance/sandbox/runners.ts +444 -0
- package/.pi/extensions/zob-harness/src/domains/governance/sandbox/simulation.ts +569 -0
- package/.pi/extensions/zob-harness/src/domains/governance/sandbox/types.ts +127 -0
- package/.pi/extensions/zob-harness/src/domains/governance/sandbox/validation.ts +273 -0
- package/.pi/extensions/zob-harness/src/domains/governance/sandbox.ts +4 -1508
- package/.pi/extensions/zob-harness/src/domains/governance/worker-pool.ts +3 -1
- package/.pi/extensions/zob-harness/src/domains/governance/workspace-claims.ts +3 -1
- package/.pi/extensions/zob-harness/src/domains/orchestration/room.ts +8 -2
- package/.pi/extensions/zob-harness/src/domains/promotion/coms.ts +8 -1
- package/.pi/extensions/zob-harness/src/runtime/commands/autonomy.ts +188 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/compute.ts +165 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/daemon.ts +191 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/delegates.ts +47 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/goal.ts +70 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/intent.ts +383 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/misc.ts +229 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/project-dna.ts +130 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/types.ts +3 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/zcommit.ts +145 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/zlive.ts +1606 -0
- package/.pi/extensions/zob-harness/src/runtime/commands/zmode.ts +42 -0
- package/.pi/extensions/zob-harness/src/runtime/commands.ts +26 -3109
- package/.pi/extensions/zob-harness/src/runtime/events.ts +67 -33
- package/.pi/extensions/zob-harness/src/runtime/goal-runtime/commands.ts +194 -0
- package/.pi/extensions/zob-harness/src/runtime/goal-runtime/events.ts +81 -0
- package/.pi/extensions/zob-harness/src/runtime/goal-runtime/state.ts +662 -0
- package/.pi/extensions/zob-harness/src/runtime/goal-runtime/tools.ts +1005 -0
- package/.pi/extensions/zob-harness/src/runtime/goal-runtime.ts +5 -1949
- package/.pi/extensions/zob-harness/src/runtime/tools-delegation/helpers.ts +786 -0
- package/.pi/extensions/zob-harness/src/runtime/tools-delegation/register.ts +1120 -0
- package/.pi/extensions/zob-harness/src/runtime/tools-delegation/types.ts +77 -0
- package/.pi/extensions/zob-harness/src/runtime/tools-delegation.ts +1 -1904
- package/.pi/extensions/zob-harness/src/runtime/zob-intro.ts +46 -15
- package/.pi/factories/project-dna/batch-manifest.json +1 -1
- package/.pi/factories/project-dna/pilot-manifest.json +1 -1
- package/.pi/factories/project-dna/smoke-manifest.json +1 -1
- package/README.md +29 -8
- package/package.json +14 -5
- package/scripts/git-ops/commit-policy-smoke.mjs +33 -6
- package/scripts/goal-todo/child-goal-ref-smoke.mjs +30 -3
- package/scripts/goal-todo/handoff-static-smoke.mjs +31 -3
- package/scripts/harness-intake/lib/cli-io.mjs +89 -0
- package/scripts/harness-intake/lib/constants.mjs +59 -0
- package/scripts/harness-intake/lib/infer-spec.mjs +127 -0
- package/scripts/harness-intake/lib/profiles.mjs +458 -0
- package/scripts/harness-intake/lib/run-init.mjs +307 -0
- package/scripts/harness-intake/lib/scan.mjs +266 -0
- package/scripts/harness-intake/lib/tmux.mjs +92 -0
- package/scripts/harness-intake/lib/validate.mjs +152 -0
- package/scripts/harness-intake/lib.mjs +8 -1521
- package/scripts/harness-switch/static-smoke.mjs +1 -1
- package/scripts/model-catalog/validate-economy.mjs +3 -1
- package/scripts/model-catalog/validate.mjs +3 -1
- package/scripts/project-dna/scan/scan.mjs +5 -2
- package/scripts/project-dna/scan/validate-scan-artifacts.mjs +3 -1
- package/scripts/project-dna/validation/validate-ontology.mjs +3 -1
- package/scripts/project-dna/validation/validate-scaffold.mjs +2 -2
- package/scripts/zagent-static-smoke.mjs +30 -2
- package/scripts/zpeer-local-e2e-smoke.mjs +18 -0
- package/scripts/zpeer-static-smoke.mjs +40 -5
- package/scripts/zteam-hot-add/smoke.mjs +30 -2
|
@@ -1,2912 +1,4 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
3
|
-
|
|
4
|
-
|
|
5
|
-
import { runFactoryRun } from "../factory/run.js";
|
|
6
|
-
import { selectFactoryForDemands } from "../factory/factory-selector.js";
|
|
7
|
-
import { evaluateStrictBudgetDispatchGate } from "../governance/budget-policy.js";
|
|
8
|
-
import { evaluateBudgetPreflightDryRun } from "../telemetry/chronicle.js";
|
|
9
|
-
import { buildBrainLookupResult, buildContextPack, buildDefaultContextScope, validateContextPack, validateContextScope } from "../context/context-gbrain.js";
|
|
10
|
-
import { validateDaemonPolicyConfig } from "./daemon-policy.js";
|
|
11
|
-
import { MISSION_CONTROL_COMMANDS, buildMissionControlCommandProposal, buildMissionControlSnapshot, buildZobComsTransportReadiness, buildZobCommunicationReadinessAudit } from "../coms/mission-control.js";
|
|
12
|
-
import { evaluateModelRoutingDispatchGate, evaluateModelRoutingDryRun } from "../models/model-routing.js";
|
|
13
|
-
import { loadTeamDefinition, validateTeamDefinition } from "../topology/teams.js";
|
|
14
|
-
import { sha256 } from "../../core/utils/hashing.js";
|
|
15
|
-
import { safeFileStem } from "../../core/utils/paths.js";
|
|
16
|
-
import { isRecord } from "../../core/utils/records.js";
|
|
17
|
-
|
|
18
|
-
export type AutonomousApplyPolicy = "no_apply" | "sandbox_simulation" | "manual_apply_only" | "auto_apply_in_scope";
|
|
19
|
-
export type AutonomousBudgetProfile = "advisory" | "strict_requested";
|
|
20
|
-
export type AutonomousRisk = "low" | "medium" | "high";
|
|
21
|
-
export type AutonomousLevel = "L4" | "L5" | "L6";
|
|
22
|
-
|
|
23
|
-
export interface AutonomousRuntimeDryRunInput {
|
|
24
|
-
userNeed: string;
|
|
25
|
-
refinedSpec?: string;
|
|
26
|
-
runId?: string;
|
|
27
|
-
constraints?: string[];
|
|
28
|
-
acceptanceCriteria?: string[];
|
|
29
|
-
expectedArtifacts?: string[];
|
|
30
|
-
allowedPaths?: string[];
|
|
31
|
-
forbiddenPaths?: string[];
|
|
32
|
-
allowedSources?: string[];
|
|
33
|
-
maxContextTokens?: number;
|
|
34
|
-
applyPolicy?: AutonomousApplyPolicy;
|
|
35
|
-
budgetProfile?: AutonomousBudgetProfile;
|
|
36
|
-
risk?: AutonomousRisk;
|
|
37
|
-
authorizedAutonomyLevel?: AutonomousLevel;
|
|
38
|
-
userLaunchConfirmed?: boolean;
|
|
39
|
-
launchConfirmedAt?: string;
|
|
40
|
-
allowedActions?: string[];
|
|
41
|
-
}
|
|
42
|
-
|
|
43
|
-
export interface AutonomousReadOnlySmokeRunInput extends AutonomousRuntimeDryRunInput {
|
|
44
|
-
factoryRunId?: string;
|
|
45
|
-
}
|
|
46
|
-
|
|
47
|
-
const FORBIDDEN_BODY_KEYS = new Set(["task", "prompt", "output", "body", "content", "patch", "diff", "raw", "messages", "conversationHistory"]);
|
|
48
|
-
const AMBIGUOUS_PATTERNS = [/\bstuff\b/i, /\bsomething\b/i, /\bwhatever\b/i, /\bfix it\b/i, /\bmake it better\b/i, /\betc\.?\b/i];
|
|
49
|
-
const SECRET_PATTERNS = [/\.env\b/i, /api[_-]?key/i, /secret/i, /password/i, /token/i, /private[_-]?key/i, /ssh[_-]?key/i];
|
|
50
|
-
const DEFAULT_AUTONOMOUS_FORBIDDEN_PATHS = [".env", ".env.*", "secrets", "raw-conversation-history", "node_modules", "dist", "build"];
|
|
51
|
-
const UNSAFE_AUTONOMOUS_ALLOWED_SEGMENTS = new Set(["node_modules", "dist", "build", "secrets", "raw-conversation-history"]);
|
|
52
|
-
const AUTONOMOUS_CURRENT_SOURCE_FINGERPRINT_FILES = [
|
|
53
|
-
"package.json",
|
|
54
|
-
"tsconfig.json",
|
|
55
|
-
"scripts/harness-smoke.mjs",
|
|
56
|
-
"docs/AUTONOMY_FACTORY_AGENT_DETAILED_PLAN.md",
|
|
57
|
-
".pi/budget-policy.json",
|
|
58
|
-
".pi/model-routing.json",
|
|
59
|
-
".pi/daemon-policy.json",
|
|
60
|
-
".pi/teams/zob-core.json",
|
|
61
|
-
".pi/extensions/zob-harness/index.ts",
|
|
62
|
-
".pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime.ts",
|
|
63
|
-
".pi/extensions/zob-harness/src/domains/autonomy/autonomy-readiness.ts",
|
|
64
|
-
".pi/extensions/zob-harness/src/runtime/tools-autonomous.ts",
|
|
65
|
-
".pi/extensions/zob-harness/src/runtime/schemas.ts",
|
|
66
|
-
".pi/extensions/zob-harness/src/domains/factory/run.ts",
|
|
67
|
-
".pi/extensions/zob-harness/src/domains/models/model-routing.ts",
|
|
68
|
-
".pi/extensions/zob-harness/src/domains/governance/budget-policy.ts",
|
|
69
|
-
".pi/extensions/zob-harness/src/domains/autonomy/daemon-policy.ts",
|
|
70
|
-
".pi/extensions/zob-harness/src/domains/coms/mission-control.ts",
|
|
71
|
-
".pi/extensions/zob-harness/src/domains/governance/sandbox.ts",
|
|
72
|
-
];
|
|
73
|
-
|
|
74
|
-
function hasForbiddenBodyKeys(value: unknown): boolean {
|
|
75
|
-
if (!value || typeof value !== "object") return false;
|
|
76
|
-
if (Array.isArray(value)) return value.some(hasForbiddenBodyKeys);
|
|
77
|
-
return Object.entries(value as Record<string, unknown>).some(([key, child]) => FORBIDDEN_BODY_KEYS.has(key) || hasForbiddenBodyKeys(child));
|
|
78
|
-
}
|
|
79
|
-
|
|
80
|
-
function stableStrings(values: string[] | undefined): string[] {
|
|
81
|
-
return [...new Set((values ?? []).filter((value) => typeof value === "string" && value.trim().length > 0).map((value) => value.trim()))].sort();
|
|
82
|
-
}
|
|
83
|
-
|
|
84
|
-
function normalizeAutonomousSpecPath(value: string): string {
|
|
85
|
-
return value.trim().replace(/\\+/g, "/").replace(/\/+/g, "/").replace(/^\.\//, "").replace(/\/+$/, "");
|
|
86
|
-
}
|
|
87
|
-
|
|
88
|
-
function isBroadAutonomousAllowedPath(value: string): boolean {
|
|
89
|
-
const normalized = normalizeAutonomousSpecPath(value);
|
|
90
|
-
return normalized === "" || normalized === "." || normalized === "/" || normalized === "*" || normalized === "**" || normalized === "~" || normalized === "~/";
|
|
91
|
-
}
|
|
92
|
-
|
|
93
|
-
function forbiddenPatternBase(value: string): string {
|
|
94
|
-
return normalizeAutonomousSpecPath(value)
|
|
95
|
-
.replace(/\/\*\*$/, "")
|
|
96
|
-
.replace(/\/\*$/, "")
|
|
97
|
-
.replace(/\.\*$/, "");
|
|
98
|
-
}
|
|
99
|
-
|
|
100
|
-
function pathConflict(left: string, right: string): boolean {
|
|
101
|
-
const a = forbiddenPatternBase(left).toLowerCase();
|
|
102
|
-
const b = forbiddenPatternBase(right).toLowerCase();
|
|
103
|
-
if (!a || !b) return false;
|
|
104
|
-
return a === b || a.startsWith(`${b}/`) || b.startsWith(`${a}/`);
|
|
105
|
-
}
|
|
106
|
-
|
|
107
|
-
function unsafeAllowedPathReason(value: string, forbiddenPaths: string[]): string | undefined {
|
|
108
|
-
const normalized = normalizeAutonomousSpecPath(value);
|
|
109
|
-
const lower = normalized.toLowerCase();
|
|
110
|
-
if (value.includes("\0")) return "allowed_path_contains_nul_byte";
|
|
111
|
-
if (isBroadAutonomousAllowedPath(value)) return "allowed_path_too_broad";
|
|
112
|
-
if (lower.startsWith("/") || lower.startsWith("~/") || /^[a-z]:\//i.test(lower)) return "allowed_path_must_be_repo_relative";
|
|
113
|
-
if (lower.split("/").includes("..")) return "allowed_path_must_not_traverse_parent";
|
|
114
|
-
const segments = lower.split("/").filter(Boolean);
|
|
115
|
-
if (segments.some((segment) => segment === ".env" || segment.startsWith(".env."))) return "allowed_path_references_secret_path";
|
|
116
|
-
if (segments.some((segment) => UNSAFE_AUTONOMOUS_ALLOWED_SEGMENTS.has(segment))) return "allowed_path_references_forbidden_segment";
|
|
117
|
-
if (forbiddenPaths.some((forbidden) => pathConflict(normalized, forbidden))) return "allowed_path_conflicts_with_forbidden_path";
|
|
118
|
-
return undefined;
|
|
119
|
-
}
|
|
120
|
-
|
|
121
|
-
function validateAutonomousSpecPathGate(allowedPaths: string[], forbiddenPaths: string[]): string[] {
|
|
122
|
-
if (allowedPaths.length === 0) return ["allowed_paths_required_before_execution"];
|
|
123
|
-
return allowedPaths.flatMap((allowedPath) => {
|
|
124
|
-
const reason = unsafeAllowedPathReason(allowedPath, forbiddenPaths);
|
|
125
|
-
return reason ? [reason] : [];
|
|
126
|
-
});
|
|
127
|
-
}
|
|
128
|
-
|
|
129
|
-
function hashes(values: string[]): string[] {
|
|
130
|
-
return values.map((value) => sha256(value)).sort();
|
|
131
|
-
}
|
|
132
|
-
|
|
133
|
-
function defaultAutonomousAllowedActions(applyPolicy: AutonomousApplyPolicy): string[] {
|
|
134
|
-
const base = ["read_repo", "context_lookup", "select_factory", "run_factory_smoke", "run_factory_pilot", "run_factory_batch", "post_run_validation", "post_run_oracle"];
|
|
135
|
-
return applyPolicy === "auto_apply_in_scope"
|
|
136
|
-
? [...base, "sandbox_edit", "apply_in_scope", "post_apply_validation", "post_apply_oracle"]
|
|
137
|
-
: base;
|
|
138
|
-
}
|
|
139
|
-
|
|
140
|
-
function inferAutonomousLevel(applyPolicy: AutonomousApplyPolicy): AutonomousLevel {
|
|
141
|
-
return applyPolicy === "auto_apply_in_scope" ? "L6" : "L4";
|
|
142
|
-
}
|
|
143
|
-
|
|
144
|
-
function capabilityFactories(capabilityIndex: Record<string, unknown>): Record<string, unknown>[] {
|
|
145
|
-
const capabilities = Array.isArray(capabilityIndex.capabilities) ? capabilityIndex.capabilities.filter(isRecord) : [];
|
|
146
|
-
return capabilities.filter((capability) => capability.kind === "factory");
|
|
147
|
-
}
|
|
148
|
-
|
|
149
|
-
function textSignals(...values: string[]): string[] {
|
|
150
|
-
const text = values.join("\n").toLowerCase();
|
|
151
|
-
const signals: string[] = [];
|
|
152
|
-
const add = (condition: boolean, signal: string): void => {
|
|
153
|
-
if (condition) signals.push(signal);
|
|
154
|
-
};
|
|
155
|
-
add(/\b(code review|review code|review changes|oracle matrix|security review|qa review|correctness|architecture)\b/.test(text), "code_review");
|
|
156
|
-
add(/\b(budget|cost|costs|cap|caps|preflight|strict budget|max runs|max cost|parallel children)\b/.test(text), "budget_preflight");
|
|
157
|
-
add(/\b(roadmap|lot|lots|milestone|unchecked item|execution queue)\b/.test(text), "roadmap_lots");
|
|
158
|
-
add(/\b(opencode|pattern|patterns|canonizer|canonical|taxonomy|workflow rules|quality gates)\b/.test(text), "opencode_patterns");
|
|
159
|
-
add(/\b(projectdna|project dna|project-dna|knowledge graph|code knowledge|context pack|repo scan|reference project)\b/.test(text), "project_dna");
|
|
160
|
-
add(/\b(new factory|create factory|generate factory|factory scaffold|quarantine|factory-forge|forge)\b/.test(text), "factory_forge");
|
|
161
|
-
return [...new Set(signals)].sort();
|
|
162
|
-
}
|
|
163
|
-
|
|
164
|
-
function scoreFactoryCandidate(factory: Record<string, unknown>, signals: string[]): Record<string, unknown> {
|
|
165
|
-
const id = typeof factory.id === "string" ? factory.id : "unknown";
|
|
166
|
-
const reasonCodes: string[] = [];
|
|
167
|
-
let score = 0;
|
|
168
|
-
const add = (condition: boolean, points: number, reason: string): void => {
|
|
169
|
-
if (condition) {
|
|
170
|
-
score += points;
|
|
171
|
-
reasonCodes.push(reason);
|
|
172
|
-
}
|
|
173
|
-
};
|
|
174
|
-
add(id === "code-review-matrix" && signals.includes("code_review"), 8, "signal:code_review");
|
|
175
|
-
add(id === "budget-preflight-dry-run" && signals.includes("budget_preflight"), 8, "signal:budget_preflight");
|
|
176
|
-
add(id === "roadmap-smoke-lots" && signals.includes("roadmap_lots"), 8, "signal:roadmap_lots");
|
|
177
|
-
add(id === "opencode-pattern-canonizer" && signals.includes("opencode_patterns"), 8, "signal:opencode_patterns");
|
|
178
|
-
add(id === "project-dna" && signals.includes("project_dna"), 8, "signal:project_dna");
|
|
179
|
-
add(id === "factory-forge" && signals.includes("factory_forge"), 8, "signal:factory_forge");
|
|
180
|
-
const metadata = isRecord(factory.metadata) ? factory.metadata : {};
|
|
181
|
-
const manifests = Array.isArray(metadata.manifests) ? metadata.manifests.filter((item): item is string => typeof item === "string") : [];
|
|
182
|
-
add(manifests.includes("smoke-manifest.json"), 1, "manifest:smoke");
|
|
183
|
-
add(manifests.includes("pilot-manifest.json"), 1, "manifest:pilot");
|
|
184
|
-
add(manifests.includes("batch-manifest.json"), 1, "manifest:batch");
|
|
185
|
-
const confidence = Math.max(0, Math.min(0.99, score / 12));
|
|
186
|
-
return {
|
|
187
|
-
kind: "factory",
|
|
188
|
-
id,
|
|
189
|
-
sourcePath: factory.sourcePath,
|
|
190
|
-
score,
|
|
191
|
-
confidence,
|
|
192
|
-
reasonCodes: reasonCodes.sort(),
|
|
193
|
-
summaryHash: typeof factory.summary === "string" ? sha256(factory.summary) : undefined,
|
|
194
|
-
};
|
|
195
|
-
}
|
|
196
|
-
|
|
197
|
-
function selectAutonomousFactory(input: { factories: Record<string, unknown>[]; refinedSpec: string; acceptanceCriteria: string[]; expectedArtifacts: string[] }): Record<string, unknown> {
|
|
198
|
-
return selectFactoryForDemands({
|
|
199
|
-
schema: "zob.autonomous-factory-selection-score.v1",
|
|
200
|
-
factories: input.factories,
|
|
201
|
-
refinedSpec: input.refinedSpec,
|
|
202
|
-
acceptanceCriteria: input.acceptanceCriteria,
|
|
203
|
-
expectedArtifacts: input.expectedArtifacts,
|
|
204
|
-
}) as unknown as Record<string, unknown>;
|
|
205
|
-
}
|
|
206
|
-
|
|
207
|
-
function factoryManifestAvailability(repoRoot: string, factoryName: string | undefined): Record<string, unknown> {
|
|
208
|
-
if (!factoryName) return { smoke: false, pilot: false, batch: false };
|
|
209
|
-
return {
|
|
210
|
-
smoke: existsSync(join(repoRoot, ".pi", "factories", factoryName, "smoke-manifest.json")),
|
|
211
|
-
pilot: existsSync(join(repoRoot, ".pi", "factories", factoryName, "pilot-manifest.json")),
|
|
212
|
-
batch: existsSync(join(repoRoot, ".pi", "factories", factoryName, "batch-manifest.json")),
|
|
213
|
-
};
|
|
214
|
-
}
|
|
215
|
-
|
|
216
|
-
function readFactoryRegistryReadiness(repoRoot: string, factoryName: string | undefined): Record<string, unknown> {
|
|
217
|
-
const reportPath = "reports/factory-registry-readiness-audit-smoke.json";
|
|
218
|
-
const absolutePath = join(repoRoot, reportPath);
|
|
219
|
-
const base = {
|
|
220
|
-
schema: "zob.autonomous-factory-readiness-snapshot.v1",
|
|
221
|
-
reportPath,
|
|
222
|
-
reportPresent: false,
|
|
223
|
-
reportHash: undefined,
|
|
224
|
-
selectedFactory: factoryName,
|
|
225
|
-
registeredBatchReady: false,
|
|
226
|
-
arbitraryFactoryNoShip: true,
|
|
227
|
-
currentSourceProofRequired: true,
|
|
228
|
-
proofBeforeExecutionRequired: true,
|
|
229
|
-
readinessFreshness: "missing_registry_snapshot",
|
|
230
|
-
bodyStored: false,
|
|
231
|
-
promptBodiesStored: false,
|
|
232
|
-
outputBodiesStored: false,
|
|
233
|
-
};
|
|
234
|
-
if (!existsSync(absolutePath)) return base;
|
|
235
|
-
try {
|
|
236
|
-
const rawJson = readFileSync(absolutePath, "utf8");
|
|
237
|
-
const parsed = JSON.parse(rawJson) as unknown;
|
|
238
|
-
const registry = isRecord(parsed) ? parsed : {};
|
|
239
|
-
const readyFactories = Array.isArray(registry.registeredAgenticBatchReadyFactories) ? registry.registeredAgenticBatchReadyFactories.filter((item): item is string => typeof item === "string") : [];
|
|
240
|
-
const missingFactories = Array.isArray(registry.factoriesMissingRegisteredBatchProof) ? registry.factoriesMissingRegisteredBatchProof.filter((item): item is string => typeof item === "string") : [];
|
|
241
|
-
const registeredBatchReady = typeof factoryName === "string" && readyFactories.includes(factoryName);
|
|
242
|
-
return {
|
|
243
|
-
...base,
|
|
244
|
-
reportPresent: true,
|
|
245
|
-
reportHash: sha256(rawJson),
|
|
246
|
-
registeredBatchReady,
|
|
247
|
-
arbitraryFactoryNoShip: registry.arbitraryFactoryNoShip !== false,
|
|
248
|
-
currentSourceProofRequired: true,
|
|
249
|
-
proofBeforeExecutionRequired: !registeredBatchReady,
|
|
250
|
-
readinessFreshness: "snapshot_requires_current_source_refresh_before_execution",
|
|
251
|
-
registeredAgenticBatchReadyFactoryCount: readyFactories.length,
|
|
252
|
-
factoriesMissingRegisteredBatchProofCount: missingFactories.length,
|
|
253
|
-
selectedFactoryMissingRegisteredBatchProof: typeof factoryName === "string" ? missingFactories.includes(factoryName) : undefined,
|
|
254
|
-
};
|
|
255
|
-
} catch {
|
|
256
|
-
return { ...base, reportPresent: true, readinessFreshness: "invalid_registry_snapshot", reportHash: undefined };
|
|
257
|
-
}
|
|
258
|
-
}
|
|
259
|
-
|
|
260
|
-
function buildAutonomousContextArtifacts(repoRoot: string, scope: Record<string, unknown>, queryHash: string): Record<string, unknown> {
|
|
261
|
-
const scopeErrors = validateContextScope(repoRoot, scope);
|
|
262
|
-
if (scopeErrors.length > 0) return { lookupResults: [], contextPack: undefined, contextPackValid: false, contextPackErrors: scopeErrors };
|
|
263
|
-
const allowedSources = Array.isArray(scope.allowedSources) ? scope.allowedSources.filter((source): source is string => typeof source === "string") : [];
|
|
264
|
-
try {
|
|
265
|
-
const lookupResults: Record<string, unknown>[] = [];
|
|
266
|
-
if (allowedSources.includes("zob-harness-docs")) {
|
|
267
|
-
const citation = "harness-system:zob-harness-docs:docs/AUTONOMOUS_SUPER_FACTORY_GOAL.md#phase-3";
|
|
268
|
-
lookupResults.push(buildBrainLookupResult(repoRoot, {
|
|
269
|
-
scope,
|
|
270
|
-
brainId: "harness-system",
|
|
271
|
-
sourceId: "zob-harness-docs",
|
|
272
|
-
queryHash,
|
|
273
|
-
facts: [{ factHash: sha256("autonomous loop requires spec context factory oracle final report"), citations: [citation], confidence: "HIGH" }],
|
|
274
|
-
gaps: [{ gapHash: sha256("P0 dry-run does not execute live factories"), citations: [citation], noShipIfTreatedAsPass: true }],
|
|
275
|
-
confidence: "HIGH",
|
|
276
|
-
}));
|
|
277
|
-
}
|
|
278
|
-
if (allowedSources.includes("factory-run-reports")) {
|
|
279
|
-
const citation = "factory-evidence:factory-run-reports:reports/factory-registry-readiness-audit-smoke.json";
|
|
280
|
-
lookupResults.push(buildBrainLookupResult(repoRoot, {
|
|
281
|
-
scope,
|
|
282
|
-
brainId: "factory-evidence",
|
|
283
|
-
sourceId: "factory-run-reports",
|
|
284
|
-
queryHash,
|
|
285
|
-
facts: [{ factHash: sha256("registered factory current-source proof is required before execution"), citations: [citation], confidence: "HIGH", sourcePresent: existsSync(join(repoRoot, "reports", "factory-registry-readiness-audit-smoke.json")) }],
|
|
286
|
-
gaps: [{ gapHash: sha256("selected factory may still require current-source proof refresh"), citations: [citation], noShipIfTreatedAsPass: true }],
|
|
287
|
-
confidence: "HIGH",
|
|
288
|
-
}));
|
|
289
|
-
}
|
|
290
|
-
if (lookupResults.length === 0 && allowedSources.includes("zob-harness-src")) {
|
|
291
|
-
const citation = "harness-system:zob-harness-src:.pi/extensions/zob-harness/src/domains/autonomy/autonomous-runtime.ts";
|
|
292
|
-
lookupResults.push(buildBrainLookupResult(repoRoot, {
|
|
293
|
-
scope,
|
|
294
|
-
brainId: "harness-system",
|
|
295
|
-
sourceId: "zob-harness-src",
|
|
296
|
-
queryHash,
|
|
297
|
-
facts: [{ factHash: sha256("autonomous dry-run implementation is repo-local metadata-only"), citations: [citation], confidence: "MEDIUM" }],
|
|
298
|
-
confidence: "MEDIUM",
|
|
299
|
-
}));
|
|
300
|
-
}
|
|
301
|
-
if (lookupResults.length === 0) return { lookupResults: [], contextPack: undefined, contextPackValid: false, contextPackErrors: ["no_allowed_context_source_for_autonomous_lookup"] };
|
|
302
|
-
const contextPack = buildContextPack(repoRoot, scope, lookupResults);
|
|
303
|
-
const contextPackErrors = validateContextPack(repoRoot, contextPack);
|
|
304
|
-
return { lookupResults, contextPack, contextPackValid: contextPackErrors.length === 0, contextPackErrors };
|
|
305
|
-
} catch (error) {
|
|
306
|
-
const message = error instanceof Error ? error.message : String(error);
|
|
307
|
-
return { lookupResults: [], contextPack: undefined, contextPackValid: false, contextPackErrors: [message] };
|
|
308
|
-
}
|
|
309
|
-
}
|
|
310
|
-
|
|
311
|
-
function buildAutonomousRuntimeGates(input: { runId: string; risk: AutonomousRisk; budgetProfile: AutonomousBudgetProfile; applyPolicy: AutonomousApplyPolicy; maxContextTokens?: number }): Record<string, unknown> {
|
|
312
|
-
const budgetGate = evaluateBudgetPreflightDryRun({
|
|
313
|
-
runs: 0,
|
|
314
|
-
durationMs: 0,
|
|
315
|
-
parallelChildren: 0,
|
|
316
|
-
caps: { maxRuns: 0, maxParallelChildren: 0 },
|
|
317
|
-
strictRequested: input.budgetProfile === "strict_requested",
|
|
318
|
-
});
|
|
319
|
-
const modelRoutingGate = evaluateModelRoutingDryRun({
|
|
320
|
-
mode: "factory",
|
|
321
|
-
taskType: "autonomous-runtime-dry-run",
|
|
322
|
-
risk: input.risk,
|
|
323
|
-
contextTokens: input.maxContextTokens,
|
|
324
|
-
estimatedRuns: 0,
|
|
325
|
-
estimatedParallelChildren: 0,
|
|
326
|
-
caps: { maxRuns: 0, maxParallelChildren: 0 },
|
|
327
|
-
strictRequested: input.budgetProfile === "strict_requested",
|
|
328
|
-
});
|
|
329
|
-
const sandboxGate = {
|
|
330
|
-
schema: "zob.autonomous-sandbox-gate.v1",
|
|
331
|
-
applyPolicy: input.applyPolicy,
|
|
332
|
-
sandboxSimulationPlanned: input.applyPolicy === "sandbox_simulation",
|
|
333
|
-
manualApplyOnly: input.applyPolicy === "manual_apply_only",
|
|
334
|
-
launchAuthorizedApplyPlanned: input.applyPolicy === "auto_apply_in_scope",
|
|
335
|
-
launchAuthorizationRequiredForApply: input.applyPolicy === "auto_apply_in_scope",
|
|
336
|
-
productionWritesPerformed: false,
|
|
337
|
-
autoApply: false,
|
|
338
|
-
rollbackRequiredBeforeRealApply: true,
|
|
339
|
-
oracleDiffReviewRequiredBeforeRealApply: true,
|
|
340
|
-
childDispatchAllowed: false,
|
|
341
|
-
noExecution: true,
|
|
342
|
-
bodyStored: false,
|
|
343
|
-
promptBodiesStored: false,
|
|
344
|
-
outputBodiesStored: false,
|
|
345
|
-
};
|
|
346
|
-
const daemonGate = {
|
|
347
|
-
schema: "zob.autonomous-daemon-gate.v1",
|
|
348
|
-
autoStartDaemon: false,
|
|
349
|
-
continuousLoop: false,
|
|
350
|
-
daemonStarted: false,
|
|
351
|
-
manualOneShotOnly: true,
|
|
352
|
-
killSwitchRequiredBeforeLiveAutonomy: true,
|
|
353
|
-
childDispatchAllowed: false,
|
|
354
|
-
noExecution: true,
|
|
355
|
-
bodyStored: false,
|
|
356
|
-
promptBodiesStored: false,
|
|
357
|
-
outputBodiesStored: false,
|
|
358
|
-
};
|
|
359
|
-
const missionControlGate = {
|
|
360
|
-
schema: "zob.autonomous-mission-control-gate.v1",
|
|
361
|
-
proposalOnly: true,
|
|
362
|
-
directWorkerWrites: false,
|
|
363
|
-
transportDispatch: false,
|
|
364
|
-
networkComsEnabled: false,
|
|
365
|
-
topologyGuardRequired: true,
|
|
366
|
-
hashOnlyLedgerRequired: true,
|
|
367
|
-
childDispatchAllowed: false,
|
|
368
|
-
noExecution: true,
|
|
369
|
-
bodyStored: false,
|
|
370
|
-
promptBodiesStored: false,
|
|
371
|
-
outputBodiesStored: false,
|
|
372
|
-
};
|
|
373
|
-
const autonomousStrictBudgetGate = {
|
|
374
|
-
schema: "zob.autonomous-strict-budget-gate.v1",
|
|
375
|
-
strictBudgetRequiredForAutonomy: true,
|
|
376
|
-
budgetProfile: input.budgetProfile,
|
|
377
|
-
strictRequested: input.budgetProfile === "strict_requested",
|
|
378
|
-
strictEnabled: false,
|
|
379
|
-
globalDefaultEnabled: false,
|
|
380
|
-
budgetEnforced: false,
|
|
381
|
-
childDispatchAllowed: false,
|
|
382
|
-
dispatchBlockedUntilLiveStrictGate: true,
|
|
383
|
-
noExecution: true,
|
|
384
|
-
bodyStored: false,
|
|
385
|
-
promptBodiesStored: false,
|
|
386
|
-
outputBodiesStored: false,
|
|
387
|
-
};
|
|
388
|
-
const checks = [
|
|
389
|
-
{ name: "autonomous_strict_budget_requested_no_dispatch", passed: autonomousStrictBudgetGate.strictRequested === true && autonomousStrictBudgetGate.noExecution === true && autonomousStrictBudgetGate.budgetEnforced === false && autonomousStrictBudgetGate.childDispatchAllowed === false },
|
|
390
|
-
{ name: "budget_preflight_advisory_no_dispatch", passed: budgetGate.noExecution === true && budgetGate.childDispatchAllowed === false && budgetGate.budgetEnforced === false && budgetGate.wouldBlockDispatch === false },
|
|
391
|
-
{ name: "model_routing_dry_run_no_dispatch", passed: modelRoutingGate.noExecution === true && modelRoutingGate.modelRouterUsed === false && modelRoutingGate.routingApplied === false && modelRoutingGate.childDispatchAllowed === false },
|
|
392
|
-
{ name: "sandbox_no_production_apply", passed: sandboxGate.productionWritesPerformed === false && sandboxGate.autoApply === false && sandboxGate.noExecution === true },
|
|
393
|
-
{ name: "daemon_not_started", passed: daemonGate.daemonStarted === false && daemonGate.autoStartDaemon === false && daemonGate.continuousLoop === false },
|
|
394
|
-
{ name: "mission_control_proposals_only", passed: missionControlGate.proposalOnly === true && missionControlGate.directWorkerWrites === false && missionControlGate.transportDispatch === false && missionControlGate.networkComsEnabled === false },
|
|
395
|
-
];
|
|
396
|
-
const failedChecks = checks.filter((check) => !check.passed).map((check) => check.name);
|
|
397
|
-
const gates = {
|
|
398
|
-
schema: "zob.autonomous-runtime-gates.v1",
|
|
399
|
-
runId: input.runId,
|
|
400
|
-
passed: failedChecks.length === 0,
|
|
401
|
-
failedChecks,
|
|
402
|
-
checks,
|
|
403
|
-
budgetGate,
|
|
404
|
-
autonomousStrictBudgetGate,
|
|
405
|
-
modelRoutingGate,
|
|
406
|
-
sandboxGate,
|
|
407
|
-
daemonGate,
|
|
408
|
-
missionControlGate,
|
|
409
|
-
dryRun: true,
|
|
410
|
-
noExecution: true,
|
|
411
|
-
childDispatchAllowed: false,
|
|
412
|
-
globalBudgetEnforced: false,
|
|
413
|
-
globalModelRoutingEnabled: false,
|
|
414
|
-
daemonStarted: false,
|
|
415
|
-
productionWritesPerformed: false,
|
|
416
|
-
autoApply: false,
|
|
417
|
-
networkAccessed: false,
|
|
418
|
-
bodyStored: false,
|
|
419
|
-
promptBodiesStored: false,
|
|
420
|
-
outputBodiesStored: false,
|
|
421
|
-
};
|
|
422
|
-
if (hasForbiddenBodyKeys(gates)) throw new Error("autonomous runtime gates would store forbidden plaintext body keys");
|
|
423
|
-
return gates;
|
|
424
|
-
}
|
|
425
|
-
|
|
426
|
-
function buildAutonomousStrictBudgetProofPlan(input: { runId: string; runtimeGates: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
427
|
-
const scenarioSpecs = [
|
|
428
|
-
{ name: "strict_gate_allows_within_caps", gateInput: { runId: `${input.runId}-strict-budget-allow`, mode: "smoke", execution: "agentic", taskCount: 1, budget: { strictEnabled: true, maxRuns: 2, estimatedRuns: 1, maxParallelChildren: 1, estimatedParallelChildren: 1 } } },
|
|
429
|
-
{ name: "strict_gate_blocks_exceedance_pre_dispatch", gateInput: { runId: `${input.runId}-strict-budget-block`, mode: "smoke", execution: "agentic", taskCount: 3, budget: { strictEnabled: true, maxRuns: 2, estimatedRuns: 3, maxParallelChildren: 1, estimatedParallelChildren: 1 } } },
|
|
430
|
-
{ name: "strict_gate_default_disabled_advisory", gateInput: { runId: `${input.runId}-strict-budget-default`, mode: "smoke", execution: "agentic", taskCount: 3, budget: { strictRequested: true, maxRuns: 2, estimatedRuns: 3, maxParallelChildren: 1, estimatedParallelChildren: 1 } } },
|
|
431
|
-
].map((scenario) => {
|
|
432
|
-
const gate = evaluateStrictBudgetDispatchGate(scenario.gateInput);
|
|
433
|
-
return {
|
|
434
|
-
name: scenario.name,
|
|
435
|
-
inputHash: sha256(JSON.stringify(scenario.gateInput)),
|
|
436
|
-
strictRequested: gate.strictRequested === true,
|
|
437
|
-
strictEnabled: gate.strictEnabled === true,
|
|
438
|
-
budgetEnforced: gate.budgetEnforced === true,
|
|
439
|
-
wouldExceed: gate.wouldExceed === true,
|
|
440
|
-
wouldBlockDispatch: gate.wouldBlockDispatch === true,
|
|
441
|
-
gateChildDispatchAllowed: gate.childDispatchAllowed === true,
|
|
442
|
-
dispatchDecision: gate.dispatchDecision,
|
|
443
|
-
stopCondition: gate.stopCondition,
|
|
444
|
-
noExecution: gate.noExecution === true,
|
|
445
|
-
failures: Array.isArray(gate.failures) ? gate.failures.filter((failure): failure is string => typeof failure === "string") : [],
|
|
446
|
-
bodyStored: false,
|
|
447
|
-
promptBodiesStored: false,
|
|
448
|
-
outputBodiesStored: false,
|
|
449
|
-
};
|
|
450
|
-
});
|
|
451
|
-
const allow = scenarioSpecs.find((scenario) => scenario.name === "strict_gate_allows_within_caps");
|
|
452
|
-
const block = scenarioSpecs.find((scenario) => scenario.name === "strict_gate_blocks_exceedance_pre_dispatch");
|
|
453
|
-
const defaultDisabled = scenarioSpecs.find((scenario) => scenario.name === "strict_gate_default_disabled_advisory");
|
|
454
|
-
const autonomousStrictBudgetGate = isRecord(input.runtimeGates.autonomousStrictBudgetGate) ? input.runtimeGates.autonomousStrictBudgetGate : {};
|
|
455
|
-
const checks = [
|
|
456
|
-
{ name: "strict_gate_allows_within_caps", passed: allow?.strictEnabled === true && allow.budgetEnforced === true && allow.wouldBlockDispatch === false && allow.gateChildDispatchAllowed === true && allow.dispatchDecision === "allow_strict" },
|
|
457
|
-
{ name: "strict_gate_blocks_exceedance_pre_dispatch", passed: block?.strictEnabled === true && block.budgetEnforced === true && block.wouldBlockDispatch === true && block.gateChildDispatchAllowed === false && block.dispatchDecision === "block" && block.stopCondition === "blocked" },
|
|
458
|
-
{ name: "strict_gate_default_disabled_advisory", passed: defaultDisabled?.strictRequested === true && defaultDisabled.strictEnabled === false && defaultDisabled.budgetEnforced === false && defaultDisabled.wouldBlockDispatch === false && defaultDisabled.dispatchDecision === "allow_advisory" },
|
|
459
|
-
{ name: "autonomous_runtime_global_budget_still_disabled", passed: autonomousStrictBudgetGate.strictRequested === true && autonomousStrictBudgetGate.strictEnabled === false && autonomousStrictBudgetGate.budgetEnforced === false && autonomousStrictBudgetGate.childDispatchAllowed === false },
|
|
460
|
-
{ name: "run_scope_no_global_autonomy", passed: input.validation.globalAutonomyReady === false && input.validation.globalAutonomyNoShip === true && input.validation.childDispatchAllowed === false && input.validation.productionWritesPerformed === false && input.validation.autoApply === false },
|
|
461
|
-
];
|
|
462
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
463
|
-
const proof = {
|
|
464
|
-
schema: "zob.autonomous-strict-budget-proof-plan.v1",
|
|
465
|
-
runId: input.runId,
|
|
466
|
-
phase: "5B",
|
|
467
|
-
status: failedChecks.length === 0 ? "strict_budget_dispatch_gate_proof_ready_global_default_blocked" : "strict_budget_dispatch_gate_proof_incomplete",
|
|
468
|
-
strictBudgetProofReady: failedChecks.length === 0,
|
|
469
|
-
strictBudgetDispatchGateAvailable: true,
|
|
470
|
-
strictBudgetAllowProofPassed: allow?.dispatchDecision === "allow_strict" && allow.budgetEnforced === true,
|
|
471
|
-
strictBudgetBlockProofPassed: block?.dispatchDecision === "block" && block.wouldBlockDispatch === true,
|
|
472
|
-
strictBudgetDefaultDisabledProofPassed: defaultDisabled?.dispatchDecision === "allow_advisory" && defaultDisabled.strictEnabled === false,
|
|
473
|
-
finalE2ERequirementCleared: false,
|
|
474
|
-
no_ship: true,
|
|
475
|
-
checks,
|
|
476
|
-
failedChecks,
|
|
477
|
-
scenarios: scenarioSpecs,
|
|
478
|
-
evidenceRefs: [
|
|
479
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/runtime-gates.json`,
|
|
480
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
481
|
-
],
|
|
482
|
-
blockers: [
|
|
483
|
-
"live_autonomous_strict_budget_not_enforced",
|
|
484
|
-
"global_default_strict_budget_disabled",
|
|
485
|
-
"final_e2e_strict_budget_evidence_required",
|
|
486
|
-
],
|
|
487
|
-
autonomousRuntimeStrictBudgetEnforced: false,
|
|
488
|
-
globalStrictBudgetEnabled: false,
|
|
489
|
-
globalBudgetEnforced: false,
|
|
490
|
-
liveAutonomousBudgetEnforced: false,
|
|
491
|
-
budgetEnforced: false,
|
|
492
|
-
strictEnabled: false,
|
|
493
|
-
childDispatchAllowed: false,
|
|
494
|
-
daemonStarted: false,
|
|
495
|
-
productionWritesPerformed: false,
|
|
496
|
-
autoApply: false,
|
|
497
|
-
noExecution: true,
|
|
498
|
-
globalAutonomyReady: false,
|
|
499
|
-
globalAutonomyNoShip: true,
|
|
500
|
-
bodyStored: false,
|
|
501
|
-
promptBodiesStored: false,
|
|
502
|
-
outputBodiesStored: false,
|
|
503
|
-
};
|
|
504
|
-
if (hasForbiddenBodyKeys(proof)) throw new Error("autonomous strict budget proof plan would store forbidden plaintext body keys");
|
|
505
|
-
return proof;
|
|
506
|
-
}
|
|
507
|
-
|
|
508
|
-
function buildAutonomousModelRoutingPlan(input: { runId: string; risk: AutonomousRisk; budgetProfile: AutonomousBudgetProfile; maxContextTokens?: number; specText: string; selectedFactory?: string }): Record<string, unknown> {
|
|
509
|
-
const strictRequested = input.budgetProfile === "strict_requested";
|
|
510
|
-
const contextTokens = typeof input.maxContextTokens === "number" && Number.isFinite(input.maxContextTokens) ? Math.max(0, Math.floor(input.maxContextTokens)) : 0;
|
|
511
|
-
const highContextRequired = contextTokens >= 120_000;
|
|
512
|
-
const securitySensitive = input.risk === "high" || /\b(security|secure|vulnerability|vulnerabilities|auth|permission|permissions|sandbox|apply|write|privilege|injection)\b/i.test(input.specText);
|
|
513
|
-
const routeSpecs: Array<{ stage: string; mode: "explore" | "plan" | "implement" | "oracle" | "factory"; taskType: string; outputContract?: string; risk: AutonomousRisk; contextTokens?: number; expectedModelClass: string; securityCritical?: boolean; oracleCritical?: boolean }> = [
|
|
514
|
-
{ stage: "context_reuse_scout", mode: "explore", taskType: "read-only-inspection", risk: "low", expectedModelClass: "cheap_scout" },
|
|
515
|
-
{ stage: "context_pack", mode: "explore", taskType: "context-pack-read-only", risk: "low", contextTokens: contextTokens || undefined, expectedModelClass: highContextRequired ? "high_context" : "cheap_scout" },
|
|
516
|
-
{ stage: "factory_selection", mode: "factory", taskType: securitySensitive ? "security-critical-plan" : "factory-selection", risk: securitySensitive ? "high" : input.risk, expectedModelClass: securitySensitive ? "strong_reasoning" : "balanced_worker", securityCritical: securitySensitive },
|
|
517
|
-
{ stage: "factory_smoke_plan", mode: "factory", taskType: `factory-smoke:${input.selectedFactory ?? "unselected"}`, risk: input.risk, expectedModelClass: input.risk === "high" ? "strong_reasoning" : "balanced_worker" },
|
|
518
|
-
{ stage: "smoke_oracle", mode: "oracle", taskType: "oracle-review", outputContract: "oracle.v1", risk: "high", expectedModelClass: "strong_oracle", oracleCritical: true },
|
|
519
|
-
{ stage: "pilot_oracle", mode: "oracle", taskType: "oracle-review", outputContract: "oracle.v1", risk: "high", expectedModelClass: "strong_oracle", oracleCritical: true },
|
|
520
|
-
{ stage: "final_report_synthesis", mode: "implement", taskType: "synthesis-final-report", risk: input.risk, expectedModelClass: input.risk === "high" ? "strong_reasoning" : "balanced_worker" },
|
|
521
|
-
];
|
|
522
|
-
if (securitySensitive) {
|
|
523
|
-
routeSpecs.push({ stage: "security_reasoning_gate", mode: "plan", taskType: "security-critical-plan", risk: "high", expectedModelClass: "strong_reasoning", securityCritical: true });
|
|
524
|
-
}
|
|
525
|
-
const routes = routeSpecs.map((spec) => {
|
|
526
|
-
const route = evaluateModelRoutingDryRun({
|
|
527
|
-
mode: spec.mode,
|
|
528
|
-
taskType: spec.taskType,
|
|
529
|
-
outputContract: spec.outputContract,
|
|
530
|
-
risk: spec.risk,
|
|
531
|
-
contextTokens: spec.contextTokens,
|
|
532
|
-
estimatedRuns: 0,
|
|
533
|
-
estimatedParallelChildren: 0,
|
|
534
|
-
caps: { maxRuns: 0, maxParallelChildren: 0 },
|
|
535
|
-
strictRequested,
|
|
536
|
-
});
|
|
537
|
-
return {
|
|
538
|
-
stage: spec.stage,
|
|
539
|
-
mode: spec.mode,
|
|
540
|
-
taskType: spec.taskType,
|
|
541
|
-
outputContract: spec.outputContract,
|
|
542
|
-
risk: spec.risk,
|
|
543
|
-
contextTokens: spec.contextTokens,
|
|
544
|
-
expectedModelClass: spec.expectedModelClass,
|
|
545
|
-
recommendedModelClass: route.recommendedModelClass,
|
|
546
|
-
matchedExpectedClass: route.recommendedModelClass === spec.expectedModelClass,
|
|
547
|
-
reasonCodes: Array.isArray(route.reasonCodes) ? route.reasonCodes : [],
|
|
548
|
-
budgetWouldExceed: route.budgetWouldExceed === true,
|
|
549
|
-
budgetEnforced: route.budgetEnforced === true,
|
|
550
|
-
strictRequested: route.strictRequested === true,
|
|
551
|
-
strictEnabled: route.strictEnabled === true,
|
|
552
|
-
modelRouterUsed: route.modelRouterUsed === true,
|
|
553
|
-
routingApplied: route.routingApplied === true,
|
|
554
|
-
childDispatchAllowed: route.childDispatchAllowed === true,
|
|
555
|
-
noExecution: route.noExecution === true,
|
|
556
|
-
securityCritical: spec.securityCritical === true,
|
|
557
|
-
oracleCritical: spec.oracleCritical === true,
|
|
558
|
-
bodyStored: false,
|
|
559
|
-
promptBodiesStored: false,
|
|
560
|
-
outputBodiesStored: false,
|
|
561
|
-
};
|
|
562
|
-
});
|
|
563
|
-
const oracleRoutes = routes.filter((route) => route.oracleCritical === true);
|
|
564
|
-
const securityRoutes = routes.filter((route) => route.securityCritical === true);
|
|
565
|
-
const checks = [
|
|
566
|
-
{ name: "model_routing_plan_present", passed: routes.length >= 7 },
|
|
567
|
-
{ name: "low_risk_readonly_uses_cheap_scout", passed: routes.some((route) => route.stage === "context_reuse_scout" && route.recommendedModelClass === "cheap_scout") },
|
|
568
|
-
{ name: "oracle_uses_strong_oracle", passed: oracleRoutes.length > 0 && oracleRoutes.every((route) => route.recommendedModelClass === "strong_oracle") },
|
|
569
|
-
{ name: "high_context_uses_high_context", passed: !highContextRequired || routes.some((route) => route.stage === "context_pack" && route.recommendedModelClass === "high_context") },
|
|
570
|
-
{ name: "security_not_downgraded", passed: !securitySensitive || (securityRoutes.length > 0 && securityRoutes.every((route) => route.recommendedModelClass === "strong_reasoning" || route.recommendedModelClass === "strong_oracle" || route.recommendedModelClass === "high_context")) },
|
|
571
|
-
{ name: "budget_aware_strict_requested", passed: strictRequested && routes.every((route) => route.strictRequested === true && route.budgetEnforced === false && route.strictEnabled === false) },
|
|
572
|
-
{ name: "routing_dry_run_no_dispatch", passed: routes.every((route) => route.noExecution === true && route.modelRouterUsed === false && route.routingApplied === false && route.childDispatchAllowed === false) },
|
|
573
|
-
{ name: "global_live_routing_disabled", passed: true },
|
|
574
|
-
];
|
|
575
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
576
|
-
const plan = {
|
|
577
|
-
schema: "zob.autonomous-model-routing-plan.v1",
|
|
578
|
-
runId: input.runId,
|
|
579
|
-
selectedFactory: input.selectedFactory,
|
|
580
|
-
routingRequiredForAutonomy: true,
|
|
581
|
-
routingPlanReady: failedChecks.length === 0,
|
|
582
|
-
failedChecks,
|
|
583
|
-
checks,
|
|
584
|
-
routes,
|
|
585
|
-
highContextRequired,
|
|
586
|
-
securitySensitive,
|
|
587
|
-
budgetAware: strictRequested,
|
|
588
|
-
strictRequested,
|
|
589
|
-
strictEnabled: false,
|
|
590
|
-
budgetEnforced: false,
|
|
591
|
-
liveRoutingEnabled: false,
|
|
592
|
-
globalLiveRoutingEnabled: false,
|
|
593
|
-
modelRouterUsed: false,
|
|
594
|
-
routingApplied: false,
|
|
595
|
-
childDispatchAllowed: false,
|
|
596
|
-
noExecution: true,
|
|
597
|
-
bodyStored: false,
|
|
598
|
-
promptBodiesStored: false,
|
|
599
|
-
outputBodiesStored: false,
|
|
600
|
-
};
|
|
601
|
-
if (hasForbiddenBodyKeys(plan)) throw new Error("autonomous model routing plan would store forbidden plaintext body keys");
|
|
602
|
-
return plan;
|
|
603
|
-
}
|
|
604
|
-
|
|
605
|
-
function buildAutonomousModelRoutingProofPlan(input: { runId: string; modelRoutingPlan: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
606
|
-
const proofModelByClass = { cheap_scout: "scout-model", balanced_worker: "worker-model", strong_reasoning: "reasoning-model", strong_oracle: "oracle-model", high_context: "long-context-model" };
|
|
607
|
-
const scenarioSpecs = [
|
|
608
|
-
{ name: "routing_default_disabled", gateInput: { runId: `${input.runId}-routing-default`, mode: "factory" as const, stage: "map", agent: "explore", outputContract: "explore.v1", modelRouting: { enabled: false, modelByClass: proofModelByClass } } },
|
|
609
|
-
{ name: "routing_oracle_applies_when_enabled", gateInput: { runId: `${input.runId}-routing-oracle`, mode: "factory" as const, stage: "validate", agent: "oracle-merge", outputContract: "oracle-merge.v1", modelRouting: { enabled: true, modelByClass: proofModelByClass } } },
|
|
610
|
-
{ name: "routing_high_context_applies_when_enabled", gateInput: { runId: `${input.runId}-routing-high-context`, mode: "factory" as const, stage: "reduce", agent: "synthesis", outputContract: "synthesis.v1", modelRouting: { enabled: true, contextTokens: 150_000, modelByClass: proofModelByClass } } },
|
|
611
|
-
{ name: "routing_security_not_downgraded_when_enabled", gateInput: { runId: `${input.runId}-routing-security`, mode: "plan" as const, stage: "security_reasoning_gate", agent: "planner", outputContract: "plan.v1", risk: "high" as const, taskType: "security-critical-plan", modelRouting: { enabled: true, modelByClass: proofModelByClass } } },
|
|
612
|
-
].map((scenario) => {
|
|
613
|
-
const gate = evaluateModelRoutingDispatchGate(scenario.gateInput);
|
|
614
|
-
return {
|
|
615
|
-
name: scenario.name,
|
|
616
|
-
inputHash: sha256(JSON.stringify(scenario.gateInput)),
|
|
617
|
-
liveRoutingRequested: gate.liveRoutingRequested === true,
|
|
618
|
-
liveRoutingEnabled: gate.liveRoutingEnabled === true,
|
|
619
|
-
modelRouterUsed: gate.modelRouterUsed === true,
|
|
620
|
-
routingApplied: gate.routingApplied === true,
|
|
621
|
-
selectedModelClass: gate.selectedModelClass,
|
|
622
|
-
recommendedModelClass: gate.recommendedModelClass,
|
|
623
|
-
selectedModelHash: typeof gate.selectedModelHash === "string" ? gate.selectedModelHash : undefined,
|
|
624
|
-
selectedModelStored: gate.selectedModelStored === true,
|
|
625
|
-
modelByClassProvided: gate.modelByClassProvided === true,
|
|
626
|
-
gateChildDispatchAllowed: gate.childDispatchAllowed === true,
|
|
627
|
-
defaultDispatchDecision: gate.defaultDispatchDecision,
|
|
628
|
-
noExecution: gate.noExecution === true,
|
|
629
|
-
budgetEnforced: gate.budgetEnforced === true,
|
|
630
|
-
strictEnabled: gate.strictEnabled === true,
|
|
631
|
-
reasonCodes: Array.isArray(gate.reasonCodes) ? gate.reasonCodes.filter((reason): reason is string => typeof reason === "string") : [],
|
|
632
|
-
bodyStored: false,
|
|
633
|
-
promptBodiesStored: false,
|
|
634
|
-
outputBodiesStored: false,
|
|
635
|
-
};
|
|
636
|
-
});
|
|
637
|
-
const defaultDisabled = scenarioSpecs.find((scenario) => scenario.name === "routing_default_disabled");
|
|
638
|
-
const oracle = scenarioSpecs.find((scenario) => scenario.name === "routing_oracle_applies_when_enabled");
|
|
639
|
-
const highContext = scenarioSpecs.find((scenario) => scenario.name === "routing_high_context_applies_when_enabled");
|
|
640
|
-
const security = scenarioSpecs.find((scenario) => scenario.name === "routing_security_not_downgraded_when_enabled");
|
|
641
|
-
const checks = [
|
|
642
|
-
{ name: "routing_default_disabled", passed: defaultDisabled?.liveRoutingEnabled === false && defaultDisabled.modelRouterUsed === false && defaultDisabled.routingApplied === false && defaultDisabled.gateChildDispatchAllowed === false && defaultDisabled.defaultDispatchDecision === "allow_default_model" },
|
|
643
|
-
{ name: "routing_oracle_uses_strong_oracle_when_enabled", passed: oracle?.liveRoutingEnabled === true && oracle.modelRouterUsed === true && oracle.routingApplied === true && oracle.selectedModelClass === "strong_oracle" && typeof oracle.selectedModelHash === "string" && oracle.selectedModelStored === false },
|
|
644
|
-
{ name: "routing_high_context_uses_high_context_when_enabled", passed: highContext?.liveRoutingEnabled === true && highContext.selectedModelClass === "high_context" && typeof highContext.selectedModelHash === "string" && highContext.selectedModelStored === false },
|
|
645
|
-
{ name: "routing_security_not_downgraded_when_enabled", passed: security?.liveRoutingEnabled === true && security.selectedModelClass === "strong_reasoning" && typeof security.selectedModelHash === "string" && security.selectedModelStored === false },
|
|
646
|
-
{ name: "autonomous_model_routing_plan_still_disabled", passed: input.modelRoutingPlan.liveRoutingEnabled === false && input.modelRoutingPlan.globalLiveRoutingEnabled === false && input.modelRoutingPlan.modelRouterUsed === false && input.modelRoutingPlan.routingApplied === false && input.modelRoutingPlan.childDispatchAllowed === false },
|
|
647
|
-
{ name: "run_scope_no_global_autonomy", passed: input.validation.globalAutonomyReady === false && input.validation.globalAutonomyNoShip === true && input.validation.childDispatchAllowed === false && input.validation.productionWritesPerformed === false && input.validation.autoApply === false },
|
|
648
|
-
];
|
|
649
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
650
|
-
const proof = {
|
|
651
|
-
schema: "zob.autonomous-model-routing-proof-plan.v1",
|
|
652
|
-
runId: input.runId,
|
|
653
|
-
phase: "6B",
|
|
654
|
-
status: failedChecks.length === 0 ? "model_routing_dispatch_gate_proof_ready_global_default_blocked" : "model_routing_dispatch_gate_proof_incomplete",
|
|
655
|
-
modelRoutingProofReady: failedChecks.length === 0,
|
|
656
|
-
liveRoutingDispatchGateAvailable: true,
|
|
657
|
-
routingDefaultDisabledProofPassed: defaultDisabled?.defaultDispatchDecision === "allow_default_model" && defaultDisabled.liveRoutingEnabled === false,
|
|
658
|
-
routingOracleProofPassed: oracle?.selectedModelClass === "strong_oracle" && oracle.routingApplied === true,
|
|
659
|
-
routingHighContextProofPassed: highContext?.selectedModelClass === "high_context" && highContext.routingApplied === true,
|
|
660
|
-
routingSecurityNoDowngradeProofPassed: security?.selectedModelClass === "strong_reasoning" && security.routingApplied === true,
|
|
661
|
-
selectedModelsStored: false,
|
|
662
|
-
selectedModelHashesOnly: scenarioSpecs.every((scenario) => scenario.selectedModelStored === false && (scenario.liveRoutingEnabled === false || typeof scenario.selectedModelHash === "string")),
|
|
663
|
-
finalE2ERequirementCleared: false,
|
|
664
|
-
no_ship: true,
|
|
665
|
-
checks,
|
|
666
|
-
failedChecks,
|
|
667
|
-
scenarios: scenarioSpecs,
|
|
668
|
-
evidenceRefs: [
|
|
669
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/model-routing-plan.json`,
|
|
670
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
671
|
-
],
|
|
672
|
-
blockers: [
|
|
673
|
-
"live_autonomous_model_routing_not_enabled",
|
|
674
|
-
"global_default_model_routing_disabled",
|
|
675
|
-
"final_e2e_model_routing_evidence_required",
|
|
676
|
-
],
|
|
677
|
-
globalLiveRoutingEnabled: false,
|
|
678
|
-
liveAutonomousRoutingApplied: false,
|
|
679
|
-
modelRouterUsed: false,
|
|
680
|
-
routingApplied: false,
|
|
681
|
-
budgetEnforced: false,
|
|
682
|
-
strictEnabled: false,
|
|
683
|
-
childDispatchAllowed: false,
|
|
684
|
-
daemonStarted: false,
|
|
685
|
-
productionWritesPerformed: false,
|
|
686
|
-
autoApply: false,
|
|
687
|
-
noExecution: true,
|
|
688
|
-
globalAutonomyReady: false,
|
|
689
|
-
globalAutonomyNoShip: true,
|
|
690
|
-
bodyStored: false,
|
|
691
|
-
promptBodiesStored: false,
|
|
692
|
-
outputBodiesStored: false,
|
|
693
|
-
};
|
|
694
|
-
if (hasForbiddenBodyKeys(proof)) throw new Error("autonomous model routing proof plan would store forbidden plaintext body keys");
|
|
695
|
-
return proof;
|
|
696
|
-
}
|
|
697
|
-
|
|
698
|
-
function buildAutonomousRunGraph(input: {
|
|
699
|
-
runId: string;
|
|
700
|
-
specLocked: boolean;
|
|
701
|
-
contextScopeValid: boolean;
|
|
702
|
-
contextPackValid: boolean;
|
|
703
|
-
runtimeGatesValid: boolean;
|
|
704
|
-
modelRoutingPlanValid: boolean;
|
|
705
|
-
factorySelected: boolean;
|
|
706
|
-
registeredBatchReady: boolean;
|
|
707
|
-
proofBeforeExecutionRequired: boolean;
|
|
708
|
-
selectedFactory?: string;
|
|
709
|
-
}): Record<string, unknown> {
|
|
710
|
-
const nodes = [
|
|
711
|
-
{ id: "spec_gate", kind: "gate", artifactRef: "spec-gate.json", passed: input.specLocked, stopOnFail: true },
|
|
712
|
-
{ id: "context_scope", kind: "context", artifactRef: "context-scope.json", passed: input.contextScopeValid, stopOnFail: true },
|
|
713
|
-
{ id: "context_lookup", kind: "context", artifactRef: "context-lookup.json", passed: input.contextPackValid, stopOnFail: true },
|
|
714
|
-
{ id: "context_pack", kind: "context", artifactRef: "context-pack.json", passed: input.contextPackValid, stopOnFail: true },
|
|
715
|
-
{ id: "runtime_gates", kind: "gate", artifactRef: "runtime-gates.json", passed: input.runtimeGatesValid, stopOnFail: true },
|
|
716
|
-
{ id: "model_routing_plan", kind: "gate", artifactRef: "model-routing-plan.json", passed: input.modelRoutingPlanValid, stopOnFail: true },
|
|
717
|
-
{ id: "factory_selection", kind: "selection", artifactRef: "factory-selection.json", passed: input.factorySelected, selectedFactory: input.selectedFactory, stopOnFail: true },
|
|
718
|
-
{ id: "registered_factory_current_source_proof", kind: "proof", artifactRef: "proof-plan.json", passed: input.registeredBatchReady, proofBeforeExecutionRequired: input.proofBeforeExecutionRequired, stopOnFail: false },
|
|
719
|
-
{ id: "smoke", kind: "future_execution", artifactRef: "SMOKE_PASSED.sentinel", passed: false, dispatchAllowed: false },
|
|
720
|
-
{ id: "smoke_oracle", kind: "future_oracle", artifactRef: "oracle-review-pass.json", passed: false, dispatchAllowed: false },
|
|
721
|
-
{ id: "pilot", kind: "future_execution", artifactRef: "PILOT_PASSED.sentinel", passed: false, dispatchAllowed: false },
|
|
722
|
-
{ id: "pilot_oracle", kind: "future_oracle", artifactRef: "oracle-review-pass.json", passed: false, dispatchAllowed: false },
|
|
723
|
-
{ id: "batch", kind: "future_execution", artifactRef: "BATCH_PASSED.sentinel", passed: false, dispatchAllowed: false },
|
|
724
|
-
{ id: "final_report", kind: "report", artifactRef: "final-report.md", passed: false, dispatchAllowed: false },
|
|
725
|
-
];
|
|
726
|
-
const edges = [
|
|
727
|
-
["spec_gate", "context_scope"],
|
|
728
|
-
["context_scope", "context_lookup"],
|
|
729
|
-
["context_lookup", "context_pack"],
|
|
730
|
-
["context_pack", "runtime_gates"],
|
|
731
|
-
["runtime_gates", "model_routing_plan"],
|
|
732
|
-
["model_routing_plan", "factory_selection"],
|
|
733
|
-
["factory_selection", "registered_factory_current_source_proof"],
|
|
734
|
-
["registered_factory_current_source_proof", "smoke"],
|
|
735
|
-
["smoke", "smoke_oracle"],
|
|
736
|
-
["smoke_oracle", "pilot"],
|
|
737
|
-
["pilot", "pilot_oracle"],
|
|
738
|
-
["pilot_oracle", "batch"],
|
|
739
|
-
["batch", "final_report"],
|
|
740
|
-
].map(([from, to]) => ({ from, to, parentOwned: true, dispatchAllowed: false }));
|
|
741
|
-
const runGraph = {
|
|
742
|
-
schema: "zob.autonomous-run-graph.v1",
|
|
743
|
-
runId: input.runId,
|
|
744
|
-
status: input.specLocked && input.contextScopeValid && input.contextPackValid && input.runtimeGatesValid && input.modelRoutingPlanValid && input.factorySelected ? "dry_run_graph_ready" : "dry_run_graph_blocked",
|
|
745
|
-
parentOwned: true,
|
|
746
|
-
nodes,
|
|
747
|
-
edges,
|
|
748
|
-
stopConditions: ["clarification_required", "context_scope_invalid", "context_pack_invalid", "runtime_gates_invalid", "model_routing_plan_invalid", "factory_selection_missing", "current_source_proof_missing", "oracle_fail", "budget_exceeded", "secret_reference", "production_apply_requested"],
|
|
749
|
-
futureExecutionNodesDispatchAllowed: false,
|
|
750
|
-
childDispatchAllowed: false,
|
|
751
|
-
noExecution: true,
|
|
752
|
-
daemonStarted: false,
|
|
753
|
-
productionWritesPerformed: false,
|
|
754
|
-
autoApply: false,
|
|
755
|
-
globalAutonomyReady: false,
|
|
756
|
-
globalAutonomyNoShip: true,
|
|
757
|
-
bodyStored: false,
|
|
758
|
-
promptBodiesStored: false,
|
|
759
|
-
outputBodiesStored: false,
|
|
760
|
-
};
|
|
761
|
-
if (hasForbiddenBodyKeys(runGraph)) throw new Error("autonomous run graph would store forbidden plaintext body keys");
|
|
762
|
-
return runGraph;
|
|
763
|
-
}
|
|
764
|
-
|
|
765
|
-
function detectContextNeedTags(specText: string): string[] {
|
|
766
|
-
const lower = specText.toLowerCase();
|
|
767
|
-
const tags = new Set<string>(["harness-system", "factory-evidence"]);
|
|
768
|
-
if (/factory|batch|pilot|smoke|oracle/.test(lower)) tags.add("factory-evidence");
|
|
769
|
-
if (/context|gbrain|brain|citation|source/.test(lower)) tags.add("context-gbrain");
|
|
770
|
-
if (/budget|cost|cap/.test(lower)) tags.add("budget-policy");
|
|
771
|
-
if (/model|routing|oracle/.test(lower)) tags.add("model-routing");
|
|
772
|
-
if (/sandbox|apply|write|patch|edit/.test(lower)) tags.add("sandbox-policy");
|
|
773
|
-
if (/daemon|queue|worker|mission|coms|heartbeat/.test(lower)) tags.add("mission-control");
|
|
774
|
-
return [...tags].sort();
|
|
775
|
-
}
|
|
776
|
-
|
|
777
|
-
function buildClarificationQuestions(input: {
|
|
778
|
-
shortSpec: boolean;
|
|
779
|
-
missingAcceptance: boolean;
|
|
780
|
-
missingArtifacts: boolean;
|
|
781
|
-
missingAllowedPaths: boolean;
|
|
782
|
-
unsafeAllowedPaths: boolean;
|
|
783
|
-
missingApplyPolicy: boolean;
|
|
784
|
-
missingBudgetProfile: boolean;
|
|
785
|
-
riskyApply: boolean;
|
|
786
|
-
sensitiveReference: boolean;
|
|
787
|
-
}): string[] {
|
|
788
|
-
const questions: string[] = [];
|
|
789
|
-
if (input.shortSpec) questions.push("What exact deliverable should the autonomous dry-run plan target?");
|
|
790
|
-
if (input.missingAcceptance) questions.push("What acceptance criteria must be true before the run can proceed past spec lock?");
|
|
791
|
-
if (input.missingArtifacts) questions.push("Which final artifacts or reports should be produced?");
|
|
792
|
-
if (input.missingAllowedPaths) questions.push("Which repo-relative allowed_paths should bound future autonomous work?");
|
|
793
|
-
if (input.unsafeAllowedPaths) questions.push("Please remove unsafe or forbidden entries from allowed_paths and keep them in forbidden_paths only.");
|
|
794
|
-
if (input.missingApplyPolicy) questions.push("Which apply_policy should govern this autonomous run: no_apply, sandbox_simulation, manual_apply_only, or auto_apply_in_scope?");
|
|
795
|
-
if (input.missingBudgetProfile) questions.push("Which budget_profile should govern this autonomous run: advisory or strict_requested?");
|
|
796
|
-
if (input.riskyApply) questions.push("Should this remain no-apply/manual-apply, or is a sandbox simulation explicitly required?");
|
|
797
|
-
if (input.sensitiveReference) questions.push("Please remove or replace any secret-like values with safe placeholders before planning.");
|
|
798
|
-
return questions;
|
|
799
|
-
}
|
|
800
|
-
|
|
801
|
-
export function buildAutonomousRuntimeDryRun(repoRoot: string, input: AutonomousRuntimeDryRunInput): Record<string, unknown> {
|
|
802
|
-
const userNeed = typeof input.userNeed === "string" ? input.userNeed.trim() : "";
|
|
803
|
-
const refinedSpec = typeof input.refinedSpec === "string" && input.refinedSpec.trim().length > 0 ? input.refinedSpec.trim() : userNeed;
|
|
804
|
-
const constraints = stableStrings(input.constraints);
|
|
805
|
-
const acceptanceCriteria = stableStrings(input.acceptanceCriteria);
|
|
806
|
-
const expectedArtifacts = stableStrings(input.expectedArtifacts);
|
|
807
|
-
const allowedPaths = stableStrings(input.allowedPaths);
|
|
808
|
-
const forbiddenPaths = stableStrings(input.forbiddenPaths ?? DEFAULT_AUTONOMOUS_FORBIDDEN_PATHS);
|
|
809
|
-
const applyPolicyProvided = typeof input.applyPolicy === "string";
|
|
810
|
-
const budgetProfileProvided = typeof input.budgetProfile === "string";
|
|
811
|
-
const applyPolicy = input.applyPolicy ?? "no_apply";
|
|
812
|
-
const budgetProfile = input.budgetProfile ?? "advisory";
|
|
813
|
-
const risk = input.risk ?? "medium";
|
|
814
|
-
const authorizedAutonomyLevel = input.authorizedAutonomyLevel ?? inferAutonomousLevel(applyPolicy);
|
|
815
|
-
const userLaunchConfirmed = input.userLaunchConfirmed === true;
|
|
816
|
-
const allowedActions = stableStrings(input.allowedActions ?? defaultAutonomousAllowedActions(applyPolicy));
|
|
817
|
-
const runId = safeFileStem(input.runId ?? `autonomous-dry-run-${sha256(refinedSpec || "missing-spec").slice(0, 12)}`);
|
|
818
|
-
const specText = `${userNeed}\n${refinedSpec}\n${constraints.join("\n")}\n${acceptanceCriteria.join("\n")}\n${expectedArtifacts.join("\n")}`;
|
|
819
|
-
const shortSpec = refinedSpec.length < 24;
|
|
820
|
-
const ambiguous = AMBIGUOUS_PATTERNS.some((pattern) => pattern.test(refinedSpec));
|
|
821
|
-
const sensitiveReference = SECRET_PATTERNS.some((pattern) => pattern.test(specText));
|
|
822
|
-
const missingAcceptance = acceptanceCriteria.length === 0;
|
|
823
|
-
const missingArtifacts = expectedArtifacts.length === 0;
|
|
824
|
-
const pathGateErrors = validateAutonomousSpecPathGate(allowedPaths, forbiddenPaths);
|
|
825
|
-
const missingAllowedPaths = allowedPaths.length === 0;
|
|
826
|
-
const unsafeAllowedPaths = pathGateErrors.some((error) => error !== "allowed_paths_required_before_execution");
|
|
827
|
-
const missingApplyPolicy = !applyPolicyProvided;
|
|
828
|
-
const missingBudgetProfile = !budgetProfileProvided;
|
|
829
|
-
const riskyApply = applyPolicy !== "no_apply" && allowedPaths.length === 0;
|
|
830
|
-
const clarificationRequired = !userNeed || shortSpec || ambiguous || sensitiveReference || missingAcceptance || missingArtifacts || missingAllowedPaths || unsafeAllowedPaths || missingApplyPolicy || missingBudgetProfile || riskyApply;
|
|
831
|
-
const clarificationQuestions = buildClarificationQuestions({ shortSpec: !userNeed || shortSpec || ambiguous, missingAcceptance, missingArtifacts, missingAllowedPaths, unsafeAllowedPaths, missingApplyPolicy, missingBudgetProfile, riskyApply, sensitiveReference });
|
|
832
|
-
|
|
833
|
-
const contextScope = buildDefaultContextScope(repoRoot, {
|
|
834
|
-
runId,
|
|
835
|
-
allowedSources: input.allowedSources,
|
|
836
|
-
forbiddenSources: forbiddenPaths,
|
|
837
|
-
agentProfile: "autonomous-runtime-dry-run-p0",
|
|
838
|
-
maxContextTokens: input.maxContextTokens,
|
|
839
|
-
});
|
|
840
|
-
const contextScopeErrors = validateContextScope(repoRoot, contextScope);
|
|
841
|
-
const contextArtifacts = buildAutonomousContextArtifacts(repoRoot, contextScope, sha256(refinedSpec));
|
|
842
|
-
const contextPackValid = contextArtifacts.contextPackValid === true;
|
|
843
|
-
const contextPackErrors = Array.isArray(contextArtifacts.contextPackErrors) ? contextArtifacts.contextPackErrors : [];
|
|
844
|
-
const runtimeGates = buildAutonomousRuntimeGates({ runId, risk, budgetProfile, applyPolicy, maxContextTokens: input.maxContextTokens });
|
|
845
|
-
const runtimeGatesValid = runtimeGates.passed === true;
|
|
846
|
-
const capabilityIndex = buildCapabilityIndex(repoRoot);
|
|
847
|
-
const reuseScout = buildReuseScoutReport(repoRoot, { query: refinedSpec, run_id: runId, limit: 8 });
|
|
848
|
-
const candidates = Array.isArray(reuseScout.candidates) ? reuseScout.candidates : [];
|
|
849
|
-
const factories = capabilityFactories(capabilityIndex);
|
|
850
|
-
const factoryForgeAvailable = factories.some((factory) => factory.id === "factory-forge");
|
|
851
|
-
const factoryScore = selectAutonomousFactory({ factories, refinedSpec, acceptanceCriteria, expectedArtifacts });
|
|
852
|
-
const selectedFactory = typeof factoryScore.selectedFactory === "string" ? factoryScore.selectedFactory : (factoryForgeAvailable ? "factory-forge" : undefined);
|
|
853
|
-
const selectionStatus = typeof factoryScore.selectionStatus === "string" && factoryScore.selectionStatus !== "no_factory_available"
|
|
854
|
-
? factoryScore.selectionStatus
|
|
855
|
-
: (factoryForgeAvailable ? "factory_forge_quarantine_recommended" : "no_factory_available");
|
|
856
|
-
const manifestAvailability = factoryManifestAvailability(repoRoot, selectedFactory);
|
|
857
|
-
const factoryReadiness = readFactoryRegistryReadiness(repoRoot, selectedFactory);
|
|
858
|
-
const proofBeforeExecutionRequired = factoryReadiness.proofBeforeExecutionRequired === true;
|
|
859
|
-
const modelRoutingPlan = buildAutonomousModelRoutingPlan({ runId, risk, budgetProfile, maxContextTokens: input.maxContextTokens, specText, selectedFactory });
|
|
860
|
-
const modelRoutingPlanValid = modelRoutingPlan.routingPlanReady === true;
|
|
861
|
-
const selectionBlockers = selectedFactory ? [] : ["no_factory_available_for_spec"];
|
|
862
|
-
|
|
863
|
-
const strictBudgetMissing = budgetProfile !== "strict_requested";
|
|
864
|
-
const blockers = [
|
|
865
|
-
...(!userNeed ? ["user_need_required"] : []),
|
|
866
|
-
...(clarificationRequired ? ["clarification_required_before_execution"] : []),
|
|
867
|
-
...pathGateErrors,
|
|
868
|
-
...(missingApplyPolicy ? ["apply_policy_required_before_execution"] : []),
|
|
869
|
-
...(missingBudgetProfile ? ["budget_profile_required_before_execution"] : []),
|
|
870
|
-
...(!missingBudgetProfile && strictBudgetMissing ? ["autonomous_strict_budget_required_before_execution"] : []),
|
|
871
|
-
...(contextScopeErrors.length > 0 ? ["context_scope_invalid"] : []),
|
|
872
|
-
...(!contextPackValid ? ["context_pack_invalid"] : []),
|
|
873
|
-
...(!runtimeGatesValid ? ["runtime_gates_invalid"] : []),
|
|
874
|
-
...(!modelRoutingPlanValid ? ["model_routing_plan_invalid"] : []),
|
|
875
|
-
...selectionBlockers,
|
|
876
|
-
];
|
|
877
|
-
const specLocked = blockers.length === 0;
|
|
878
|
-
const launchAuthorization = {
|
|
879
|
-
schema: "zob.launch-authorization.v1",
|
|
880
|
-
runId,
|
|
881
|
-
originalUserAskHash: sha256(userNeed),
|
|
882
|
-
refinedSpecHash: sha256(refinedSpec),
|
|
883
|
-
specLocked,
|
|
884
|
-
userLaunchConfirmed,
|
|
885
|
-
launchConfirmedAt: userLaunchConfirmed && typeof input.launchConfirmedAt === "string" && input.launchConfirmedAt.trim().length > 0 ? input.launchConfirmedAt.trim() : undefined,
|
|
886
|
-
authorizedAutonomyLevel,
|
|
887
|
-
allowedActions,
|
|
888
|
-
allowedPaths,
|
|
889
|
-
forbiddenPaths,
|
|
890
|
-
applyPolicy: {
|
|
891
|
-
mode: applyPolicy,
|
|
892
|
-
rollbackRequired: applyPolicy !== "no_apply",
|
|
893
|
-
exactDiffHashRequired: applyPolicy !== "no_apply",
|
|
894
|
-
postApplyValidationRequired: applyPolicy !== "no_apply",
|
|
895
|
-
postApplyOracleRequired: applyPolicy !== "no_apply",
|
|
896
|
-
},
|
|
897
|
-
budgetPolicy: {
|
|
898
|
-
mode: budgetProfile,
|
|
899
|
-
strict: budgetProfile === "strict_requested",
|
|
900
|
-
strictBudgetRequired: true,
|
|
901
|
-
strictBudgetSatisfied: budgetProfile === "strict_requested",
|
|
902
|
-
},
|
|
903
|
-
stopConditions: ["scope_drift", "secret_required", "validation_fail_exhausted", "oracle_no_ship", "budget_exceeded", "stale_worker_unrecoverable"],
|
|
904
|
-
launchAuthorizesInScopeActions: specLocked && userLaunchConfirmed,
|
|
905
|
-
actionExecutionBlockedUntilLaunch: !userLaunchConfirmed,
|
|
906
|
-
exceptionApprovalRequiredOnlyForOutOfScope: true,
|
|
907
|
-
bodyStored: false,
|
|
908
|
-
promptBodiesStored: false,
|
|
909
|
-
outputBodiesStored: false,
|
|
910
|
-
};
|
|
911
|
-
const status = specLocked ? "dry_run_plan_ready" : "clarification_or_gate_required";
|
|
912
|
-
const runGraph = buildAutonomousRunGraph({
|
|
913
|
-
runId,
|
|
914
|
-
specLocked,
|
|
915
|
-
contextScopeValid: contextScopeErrors.length === 0,
|
|
916
|
-
contextPackValid,
|
|
917
|
-
runtimeGatesValid,
|
|
918
|
-
modelRoutingPlanValid,
|
|
919
|
-
factorySelected: Boolean(selectedFactory),
|
|
920
|
-
registeredBatchReady: factoryReadiness.registeredBatchReady === true,
|
|
921
|
-
proofBeforeExecutionRequired,
|
|
922
|
-
selectedFactory,
|
|
923
|
-
});
|
|
924
|
-
|
|
925
|
-
const report = {
|
|
926
|
-
schema: "zob.autonomous-runtime-dry-run.v1",
|
|
927
|
-
runId,
|
|
928
|
-
status,
|
|
929
|
-
no_ship: true,
|
|
930
|
-
noShipReason: "P0 dry-run only; global autonomy remains disabled until current-source proof and policy gates pass.",
|
|
931
|
-
globalAutonomyReady: false,
|
|
932
|
-
globalAutonomyNoShip: true,
|
|
933
|
-
dryRun: true,
|
|
934
|
-
noExecution: true,
|
|
935
|
-
childDispatchAllowed: false,
|
|
936
|
-
daemonStarted: false,
|
|
937
|
-
productionWritesPerformed: false,
|
|
938
|
-
autoApply: false,
|
|
939
|
-
networkAccessed: false,
|
|
940
|
-
bodyStored: false,
|
|
941
|
-
promptBodiesStored: false,
|
|
942
|
-
outputBodiesStored: false,
|
|
943
|
-
specGate: {
|
|
944
|
-
schema: "zob.autonomous-spec-gate.v1",
|
|
945
|
-
userNeedHash: sha256(userNeed),
|
|
946
|
-
refinedSpecHash: sha256(refinedSpec),
|
|
947
|
-
constraintHashes: hashes(constraints),
|
|
948
|
-
acceptanceCriteriaHashes: hashes(acceptanceCriteria),
|
|
949
|
-
expectedArtifactHashes: hashes(expectedArtifacts),
|
|
950
|
-
allowedPaths,
|
|
951
|
-
forbiddenPaths,
|
|
952
|
-
allowedPathsRequired: true,
|
|
953
|
-
pathGatePassed: pathGateErrors.length === 0,
|
|
954
|
-
pathGateErrors,
|
|
955
|
-
applyPolicyRequired: true,
|
|
956
|
-
applyPolicyProvided,
|
|
957
|
-
applyPolicy,
|
|
958
|
-
budgetProfileRequired: true,
|
|
959
|
-
budgetProfileProvided,
|
|
960
|
-
budgetProfile,
|
|
961
|
-
autonomousStrictBudgetRequired: true,
|
|
962
|
-
autonomousStrictBudgetSatisfied: budgetProfile === "strict_requested",
|
|
963
|
-
risk,
|
|
964
|
-
specTextStored: false,
|
|
965
|
-
specLocked,
|
|
966
|
-
launchAuthorizationRequired: true,
|
|
967
|
-
userLaunchConfirmed,
|
|
968
|
-
launchAuthorizesInScopeActions: launchAuthorization.launchAuthorizesInScopeActions,
|
|
969
|
-
actionExecutionBlockedUntilLaunch: launchAuthorization.actionExecutionBlockedUntilLaunch,
|
|
970
|
-
clarificationRequired,
|
|
971
|
-
clarificationQuestions,
|
|
972
|
-
blockers,
|
|
973
|
-
},
|
|
974
|
-
launchAuthorization,
|
|
975
|
-
contextPlan: {
|
|
976
|
-
schema: "zob.autonomous-context-plan.v1",
|
|
977
|
-
contextNeedTags: detectContextNeedTags(specText),
|
|
978
|
-
contextScope,
|
|
979
|
-
contextScopeValid: contextScopeErrors.length === 0,
|
|
980
|
-
contextScopeErrors,
|
|
981
|
-
lookupResults: contextArtifacts.lookupResults,
|
|
982
|
-
contextPack: contextArtifacts.contextPack,
|
|
983
|
-
contextPackValid,
|
|
984
|
-
contextPackErrors,
|
|
985
|
-
lookupPlan: {
|
|
986
|
-
queryHash: sha256(refinedSpec),
|
|
987
|
-
allowedSources: contextScope.allowedSources,
|
|
988
|
-
citationRequired: true,
|
|
989
|
-
boundedContextOnly: true,
|
|
990
|
-
gbrainImportEnabled: false,
|
|
991
|
-
gbrainEmbedEnabled: false,
|
|
992
|
-
gbrainSyncEnabled: false,
|
|
993
|
-
gbrainWriteEnabled: false,
|
|
994
|
-
},
|
|
995
|
-
},
|
|
996
|
-
runtimeGates,
|
|
997
|
-
modelRoutingPlan,
|
|
998
|
-
runGraph,
|
|
999
|
-
factorySelection: {
|
|
1000
|
-
schema: "zob.autonomous-factory-selection.v1",
|
|
1001
|
-
queryHash: reuseScout.queryHash,
|
|
1002
|
-
selectionStatus,
|
|
1003
|
-
selectedFactory,
|
|
1004
|
-
selectedFactorySourcePath: selectedFactory ? `.pi/factories/${selectedFactory}/factory.json` : undefined,
|
|
1005
|
-
manifestAvailability,
|
|
1006
|
-
candidateCount: candidates.length,
|
|
1007
|
-
reuseScoutCandidates: candidates.filter(isRecord).map((candidate) => ({
|
|
1008
|
-
kind: candidate.kind,
|
|
1009
|
-
id: candidate.id,
|
|
1010
|
-
sourcePath: candidate.sourcePath,
|
|
1011
|
-
score: candidate.score,
|
|
1012
|
-
reasonCodes: candidate.reasonCodes,
|
|
1013
|
-
summaryHash: candidate.summaryHash,
|
|
1014
|
-
})),
|
|
1015
|
-
deterministicScoring: factoryScore,
|
|
1016
|
-
factoryForgeAvailable,
|
|
1017
|
-
factoryReadiness,
|
|
1018
|
-
currentSourceProofRequired: true,
|
|
1019
|
-
proofBeforeExecutionRequired,
|
|
1020
|
-
noAutoActivation: true,
|
|
1021
|
-
quarantineRequiredForNewFactory: selectionStatus === "factory_forge_quarantine_recommended",
|
|
1022
|
-
},
|
|
1023
|
-
proofPlan: {
|
|
1024
|
-
schema: "zob.autonomous-proof-plan.v1",
|
|
1025
|
-
parentOwned: true,
|
|
1026
|
-
stages: [
|
|
1027
|
-
{ name: "spec_lock", required: true, passed: specLocked },
|
|
1028
|
-
{ name: "launch_authorization", required: true, passed: launchAuthorization.launchAuthorizesInScopeActions, executionBlockedUntilLaunch: launchAuthorization.actionExecutionBlockedUntilLaunch, dispatchAllowed: false, noExecution: true },
|
|
1029
|
-
{ name: "context_scope", required: true, passed: contextScopeErrors.length === 0, citationRequired: true },
|
|
1030
|
-
{ name: "context_lookup_and_pack", required: true, passed: contextPackValid, citationRequired: true, boundedContextOnly: true, dispatchAllowed: false },
|
|
1031
|
-
{ name: "runtime_gates_preflight", required: true, passed: runtimeGatesValid, dispatchAllowed: false, noExecution: true },
|
|
1032
|
-
{ name: "model_routing_plan", required: true, passed: modelRoutingPlanValid, dispatchAllowed: false, liveRoutingEnabled: false },
|
|
1033
|
-
{ name: "factory_selection", required: true, passed: Boolean(selectedFactory) },
|
|
1034
|
-
{ name: "registered_factory_current_source_proof", required: true, passed: factoryReadiness.registeredBatchReady === true, dispatchAllowed: false, proofBeforeExecutionRequired },
|
|
1035
|
-
{ name: "smoke", required: true, dispatchAllowed: false, sentinelRequired: "SMOKE_PASSED.sentinel" },
|
|
1036
|
-
{ name: "smoke_oracle", required: true, dispatchAllowed: false, verdictRequired: "PASS", noShipRequired: false },
|
|
1037
|
-
{ name: "pilot", required: true, dispatchAllowed: false, requiresSmokeOracle: true, sentinelRequired: "PILOT_PASSED.sentinel" },
|
|
1038
|
-
{ name: "pilot_oracle", required: true, dispatchAllowed: false, verdictRequired: "PASS", noShipRequired: false },
|
|
1039
|
-
{ name: "batch", required: true, dispatchAllowed: false, requiresPilotOracle: true, sentinelRequired: "BATCH_PASSED.sentinel", batchConcurrencyCapRequired: true },
|
|
1040
|
-
{ name: "final_report", required: true, dispatchAllowed: false, validationJsonRequired: true, doneSentinelRequired: true },
|
|
1041
|
-
],
|
|
1042
|
-
gates: ["strict_goal_spec", "context_scope", "context_lookup_and_pack", "runtime_gates_preflight", "model_routing_plan", "run_graph_ready", "registered_factory_current_source_proof", "budget_preflight", "model_routing_policy", "sandbox_apply_policy", "daemon_policy", "oracle_review", "mission_control_proposals_only"],
|
|
1043
|
-
noShipConditions: ["clarification_required", "context_scope_invalid", "context_pack_invalid", "runtime_gates_invalid", "model_routing_plan_invalid", "factory_selection_missing", "oracle_fail", "budget_exceeded", "secret_reference", "production_apply_requested", "current_source_proof_missing"],
|
|
1044
|
-
},
|
|
1045
|
-
finalReportPlan: {
|
|
1046
|
-
schema: "zob.autonomous-final-report-plan.v1",
|
|
1047
|
-
artifactPath: `reports/autonomous-runs/${runId}/dry-run-report.json`,
|
|
1048
|
-
includesEvidenceRefs: true,
|
|
1049
|
-
includesBlockers: true,
|
|
1050
|
-
includesNoShipDecision: true,
|
|
1051
|
-
rawSpecStored: false,
|
|
1052
|
-
},
|
|
1053
|
-
validation: {
|
|
1054
|
-
passed: blockers.length === 0,
|
|
1055
|
-
blockers,
|
|
1056
|
-
warnings: ["dry_run_only", "global_autonomy_no_ship", "no_child_dispatch", ...(proofBeforeExecutionRequired ? ["selected_factory_current_source_proof_required_before_execution"] : [])],
|
|
1057
|
-
},
|
|
1058
|
-
generatedAt: new Date().toISOString(),
|
|
1059
|
-
};
|
|
1060
|
-
if (hasForbiddenBodyKeys(report)) throw new Error("autonomous dry-run report would store forbidden plaintext body keys");
|
|
1061
|
-
return report;
|
|
1062
|
-
}
|
|
1063
|
-
|
|
1064
|
-
export function buildAutonomousRuntimeDryRunValidation(report: Record<string, unknown>): Record<string, unknown> {
|
|
1065
|
-
const validation = {
|
|
1066
|
-
schema: "zob.autonomous-runtime-dry-run-validation.v1",
|
|
1067
|
-
runId: report.runId,
|
|
1068
|
-
status: report.status,
|
|
1069
|
-
passed: report.status === "dry_run_plan_ready",
|
|
1070
|
-
no_ship: report.no_ship,
|
|
1071
|
-
dryRun: report.dryRun,
|
|
1072
|
-
noExecution: report.noExecution,
|
|
1073
|
-
childDispatchAllowed: report.childDispatchAllowed,
|
|
1074
|
-
daemonStarted: report.daemonStarted,
|
|
1075
|
-
productionWritesPerformed: report.productionWritesPerformed,
|
|
1076
|
-
autoApply: report.autoApply,
|
|
1077
|
-
networkAccessed: report.networkAccessed,
|
|
1078
|
-
globalAutonomyReady: report.globalAutonomyReady,
|
|
1079
|
-
globalAutonomyNoShip: report.globalAutonomyNoShip,
|
|
1080
|
-
blockers: isRecord(report.validation) && Array.isArray(report.validation.blockers) ? report.validation.blockers : [],
|
|
1081
|
-
requiredArtifacts: ["spec-gate.json", "context-scope.json", "context-lookup.json", "context-pack.json", "runtime-gates.json", "model-routing-plan.json", "run-graph.json", "factory-selection.json", "proof-plan.json", "dry-run-report.json", "validation.json", "final-report.md", "DRY_RUN_READY.sentinel"],
|
|
1082
|
-
sentinel: report.status === "dry_run_plan_ready" ? "DRY_RUN_READY.sentinel" : undefined,
|
|
1083
|
-
bodyStored: false,
|
|
1084
|
-
promptBodiesStored: false,
|
|
1085
|
-
outputBodiesStored: false,
|
|
1086
|
-
generatedAt: new Date().toISOString(),
|
|
1087
|
-
};
|
|
1088
|
-
if (hasForbiddenBodyKeys(validation)) throw new Error("autonomous dry-run validation would store forbidden plaintext body keys");
|
|
1089
|
-
return validation;
|
|
1090
|
-
}
|
|
1091
|
-
|
|
1092
|
-
export function buildAutonomousRuntimeDryRunFinalReport(report: Record<string, unknown>): string {
|
|
1093
|
-
const factorySelection = isRecord(report.factorySelection) ? report.factorySelection : {};
|
|
1094
|
-
const validation = isRecord(report.validation) ? report.validation : {};
|
|
1095
|
-
const blockers = Array.isArray(validation.blockers) ? validation.blockers.map((blocker) => `- ${String(blocker)}`).join("\n") : "";
|
|
1096
|
-
return [
|
|
1097
|
-
"# Autonomous Runtime Dry-Run Report",
|
|
1098
|
-
"",
|
|
1099
|
-
`Run ID: ${String(report.runId ?? "unknown")}`,
|
|
1100
|
-
`Status: ${String(report.status ?? "unknown")}`,
|
|
1101
|
-
`No-ship: ${String(report.no_ship ?? true)}`,
|
|
1102
|
-
"",
|
|
1103
|
-
"## Safety posture",
|
|
1104
|
-
"",
|
|
1105
|
-
`- Dry-run only: ${String(report.dryRun === true)}`,
|
|
1106
|
-
`- No execution: ${String(report.noExecution === true)}`,
|
|
1107
|
-
`- Child dispatch allowed: ${String(report.childDispatchAllowed === true)}`,
|
|
1108
|
-
`- Daemon started: ${String(report.daemonStarted === true)}`,
|
|
1109
|
-
`- Production writes performed: ${String(report.productionWritesPerformed === true)}`,
|
|
1110
|
-
`- Auto-apply: ${String(report.autoApply === true)}`,
|
|
1111
|
-
`- Global autonomy no-ship: ${String(report.globalAutonomyNoShip === true)}`,
|
|
1112
|
-
"",
|
|
1113
|
-
"## Factory selection",
|
|
1114
|
-
"",
|
|
1115
|
-
`- Status: ${String(factorySelection.selectionStatus ?? "unknown")}`,
|
|
1116
|
-
`- Selected factory: ${String(factorySelection.selectedFactory ?? "none")}`,
|
|
1117
|
-
"",
|
|
1118
|
-
"## Blockers",
|
|
1119
|
-
"",
|
|
1120
|
-
blockers || "- None for dry-run plan readiness.",
|
|
1121
|
-
"",
|
|
1122
|
-
"## Evidence refs",
|
|
1123
|
-
"",
|
|
1124
|
-
"- spec-gate.json",
|
|
1125
|
-
"- context-scope.json",
|
|
1126
|
-
"- context-lookup.json",
|
|
1127
|
-
"- context-pack.json",
|
|
1128
|
-
"- runtime-gates.json",
|
|
1129
|
-
"- model-routing-plan.json",
|
|
1130
|
-
"- run-graph.json",
|
|
1131
|
-
"- factory-selection.json",
|
|
1132
|
-
"- proof-plan.json",
|
|
1133
|
-
"- dry-run-report.json",
|
|
1134
|
-
"- validation.json",
|
|
1135
|
-
"- DRY_RUN_READY.sentinel when status=dry_run_plan_ready",
|
|
1136
|
-
"",
|
|
1137
|
-
"Compliance: P0 dry-run artifact only; no global autonomy claim.",
|
|
1138
|
-
"",
|
|
1139
|
-
].join("\n");
|
|
1140
|
-
}
|
|
1141
|
-
|
|
1142
|
-
function readJsonArtifact(path: string): { parsed?: Record<string, unknown> | unknown[]; hash?: string; error?: string } {
|
|
1143
|
-
try {
|
|
1144
|
-
const raw = readFileSync(path, "utf8");
|
|
1145
|
-
const parsed = JSON.parse(raw) as unknown;
|
|
1146
|
-
return { parsed: isRecord(parsed) || Array.isArray(parsed) ? parsed : undefined, hash: sha256(raw), error: isRecord(parsed) || Array.isArray(parsed) ? undefined : "artifact JSON root must be object or array" };
|
|
1147
|
-
} catch (error) {
|
|
1148
|
-
return { error: error instanceof Error ? error.message : String(error) };
|
|
1149
|
-
}
|
|
1150
|
-
}
|
|
1151
|
-
|
|
1152
|
-
function artifactHashIfPresent(path: string): string | undefined {
|
|
1153
|
-
if (!existsSync(path)) return undefined;
|
|
1154
|
-
return sha256(readFileSync(path, "utf8"));
|
|
1155
|
-
}
|
|
1156
|
-
|
|
1157
|
-
function relativeFactoryRunPath(runId: string, artifact?: string): string {
|
|
1158
|
-
return artifact ? `reports/factory-runs/${runId}/${artifact}` : `reports/factory-runs/${runId}`;
|
|
1159
|
-
}
|
|
1160
|
-
|
|
1161
|
-
function buildAutonomousPromotionPlan(input: { runId: string; selectedFactory?: string; factoryRunRef: Record<string, unknown>; oracleReview: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1162
|
-
const smokeGatePassed = input.validation.status === "smoke_autonomy_passed"
|
|
1163
|
-
&& input.validation.no_ship === false
|
|
1164
|
-
&& input.factoryRunRef.status === "done"
|
|
1165
|
-
&& input.factoryRunRef.phaseSentinelPresent === true
|
|
1166
|
-
&& input.factoryRunRef.doneSentinelPresent === true
|
|
1167
|
-
&& input.oracleReview.verdict === "PASS"
|
|
1168
|
-
&& input.oracleReview.no_ship === false;
|
|
1169
|
-
const factoryRunId = typeof input.factoryRunRef.factoryRunId === "string" ? input.factoryRunRef.factoryRunId : "unknown";
|
|
1170
|
-
const plan = {
|
|
1171
|
-
schema: "zob.autonomous-promotion-plan.v1",
|
|
1172
|
-
runId: input.runId,
|
|
1173
|
-
selectedFactory: input.selectedFactory,
|
|
1174
|
-
sourceSmoke: {
|
|
1175
|
-
autonomousRunId: input.runId,
|
|
1176
|
-
factoryRunId,
|
|
1177
|
-
smokeSentinelRef: relativeFactoryRunPath(factoryRunId, "SMOKE_PASSED.sentinel"),
|
|
1178
|
-
doneSentinelRef: relativeFactoryRunPath(factoryRunId, "DONE.sentinel"),
|
|
1179
|
-
oracleReviewRef: `reports/autonomous-runs/${input.runId}/oracle-review.json`,
|
|
1180
|
-
},
|
|
1181
|
-
smokeGate: {
|
|
1182
|
-
passed: smokeGatePassed,
|
|
1183
|
-
sentinelPresent: input.factoryRunRef.phaseSentinelPresent === true,
|
|
1184
|
-
doneSentinelPresent: input.factoryRunRef.doneSentinelPresent === true,
|
|
1185
|
-
oraclePass: input.oracleReview.verdict === "PASS" && input.oracleReview.no_ship === false,
|
|
1186
|
-
structuralOracleOnly: input.oracleReview.oracleType === "deterministic_structural",
|
|
1187
|
-
},
|
|
1188
|
-
pilotGate: {
|
|
1189
|
-
preconditionsMet: smokeGatePassed,
|
|
1190
|
-
executionAllowed: false,
|
|
1191
|
-
dispatchAllowed: false,
|
|
1192
|
-
reason: "phase_7a_promotion_metadata_only_no_pilot_execution",
|
|
1193
|
-
requiredBeforeExecution: ["strict_budget_gate", "live_or_structural_oracle_pass", "current_source_factory_proof", "operator_or_parent_gate", "PILOT_PASSED.sentinel_on_success"],
|
|
1194
|
-
maxItems: 10,
|
|
1195
|
-
prerequisiteSmokeRunId: factoryRunId,
|
|
1196
|
-
oracleReviewPath: `reports/autonomous-runs/${input.runId}/oracle-review.json`,
|
|
1197
|
-
},
|
|
1198
|
-
batchGate: {
|
|
1199
|
-
preconditionsMet: false,
|
|
1200
|
-
executionAllowed: false,
|
|
1201
|
-
dispatchAllowed: false,
|
|
1202
|
-
reason: "pilot_not_executed_in_phase_7a",
|
|
1203
|
-
requiredBeforeExecution: ["PILOT_PASSED.sentinel", "pilot_oracle_PASS_no_ship_false", "batch_concurrency_cap", "strict_budget_gate", "final_oracle_gate"],
|
|
1204
|
-
batchConcurrencyCapRequired: true,
|
|
1205
|
-
},
|
|
1206
|
-
resumePolicy: {
|
|
1207
|
-
resumeOnlyIncompleteItems: true,
|
|
1208
|
-
doNotRerunPassedItems: true,
|
|
1209
|
-
stopOnRepeatedFail: true,
|
|
1210
|
-
maxRetryPolicyRequiredBeforeLiveScale: true,
|
|
1211
|
-
},
|
|
1212
|
-
oraclePolicy: {
|
|
1213
|
-
smokeOracleRequired: true,
|
|
1214
|
-
pilotOracleRequired: true,
|
|
1215
|
-
batchOracleRequired: true,
|
|
1216
|
-
warnOrFailTreatedAsNoShip: true,
|
|
1217
|
-
},
|
|
1218
|
-
strictBudgetRequired: true,
|
|
1219
|
-
liveRoutingEnabled: false,
|
|
1220
|
-
childDispatchAllowed: false,
|
|
1221
|
-
pilotExecuted: false,
|
|
1222
|
-
batchExecuted: false,
|
|
1223
|
-
noExecutionBeyondSmoke: true,
|
|
1224
|
-
globalAutonomyReady: false,
|
|
1225
|
-
globalAutonomyNoShip: true,
|
|
1226
|
-
bodyStored: false,
|
|
1227
|
-
promptBodiesStored: false,
|
|
1228
|
-
outputBodiesStored: false,
|
|
1229
|
-
};
|
|
1230
|
-
if (hasForbiddenBodyKeys(plan)) throw new Error("autonomous promotion plan would store forbidden plaintext body keys");
|
|
1231
|
-
return plan;
|
|
1232
|
-
}
|
|
1233
|
-
|
|
1234
|
-
function buildAutonomousPromotionProofPlan(input: { runId: string; promotionPlan: Record<string, unknown>; factoryRunRef: Record<string, unknown>; oracleReview: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1235
|
-
const smokeGate = isRecord(input.promotionPlan.smokeGate) ? input.promotionPlan.smokeGate : {};
|
|
1236
|
-
const pilotGate = isRecord(input.promotionPlan.pilotGate) ? input.promotionPlan.pilotGate : {};
|
|
1237
|
-
const batchGate = isRecord(input.promotionPlan.batchGate) ? input.promotionPlan.batchGate : {};
|
|
1238
|
-
const resumePolicy = isRecord(input.promotionPlan.resumePolicy) ? input.promotionPlan.resumePolicy : {};
|
|
1239
|
-
const oraclePolicy = isRecord(input.promotionPlan.oraclePolicy) ? input.promotionPlan.oraclePolicy : {};
|
|
1240
|
-
const scenarios = [
|
|
1241
|
-
{
|
|
1242
|
-
name: "smoke_to_pilot_preconditions_met",
|
|
1243
|
-
passed: smokeGate.passed === true && smokeGate.sentinelPresent === true && smokeGate.doneSentinelPresent === true && smokeGate.oraclePass === true,
|
|
1244
|
-
evidenceRefs: [`reports/autonomous-runs/${safeFileStem(input.runId)}/promotion-plan.json`, `reports/autonomous-runs/${safeFileStem(input.runId)}/oracle-review.json`],
|
|
1245
|
-
dispatchAllowed: false,
|
|
1246
|
-
},
|
|
1247
|
-
{
|
|
1248
|
-
name: "pilot_execution_blocked_until_live_gates",
|
|
1249
|
-
passed: pilotGate.preconditionsMet === true && pilotGate.executionAllowed === false && pilotGate.dispatchAllowed === false && input.promotionPlan.pilotExecuted === false,
|
|
1250
|
-
blockers: ["pilot_execution_disabled_until_live_budget_routing_daemon_oracle_gates"],
|
|
1251
|
-
dispatchAllowed: false,
|
|
1252
|
-
},
|
|
1253
|
-
{
|
|
1254
|
-
name: "batch_execution_blocked_without_pilot_oracle_and_concurrency",
|
|
1255
|
-
passed: batchGate.preconditionsMet === false && batchGate.executionAllowed === false && batchGate.dispatchAllowed === false && batchGate.batchConcurrencyCapRequired === true && input.promotionPlan.batchExecuted === false,
|
|
1256
|
-
blockers: ["pilot_not_executed", "pilot_oracle_missing", "batch_concurrency_cap_missing"],
|
|
1257
|
-
dispatchAllowed: false,
|
|
1258
|
-
},
|
|
1259
|
-
{
|
|
1260
|
-
name: "resume_retry_policy_safe_before_scale",
|
|
1261
|
-
passed: resumePolicy.resumeOnlyIncompleteItems === true && resumePolicy.doNotRerunPassedItems === true && resumePolicy.stopOnRepeatedFail === true && resumePolicy.maxRetryPolicyRequiredBeforeLiveScale === true,
|
|
1262
|
-
dispatchAllowed: false,
|
|
1263
|
-
},
|
|
1264
|
-
{
|
|
1265
|
-
name: "oracle_policy_no_warn_fail_promotion",
|
|
1266
|
-
passed: oraclePolicy.smokeOracleRequired === true && oraclePolicy.pilotOracleRequired === true && oraclePolicy.batchOracleRequired === true && oraclePolicy.warnOrFailTreatedAsNoShip === true,
|
|
1267
|
-
dispatchAllowed: false,
|
|
1268
|
-
},
|
|
1269
|
-
];
|
|
1270
|
-
const checks = [
|
|
1271
|
-
{ name: "smoke_gate_passed", passed: scenarios[0].passed === true },
|
|
1272
|
-
{ name: "pilot_blocked_no_execution", passed: scenarios[1].passed === true },
|
|
1273
|
-
{ name: "batch_blocked_no_execution", passed: scenarios[2].passed === true },
|
|
1274
|
-
{ name: "resume_retry_policy_safe", passed: scenarios[3].passed === true },
|
|
1275
|
-
{ name: "oracle_policy_safe", passed: scenarios[4].passed === true },
|
|
1276
|
-
{ name: "run_scope_no_global_autonomy", passed: input.validation.globalAutonomyReady === false && input.validation.globalAutonomyNoShip === true && input.validation.childDispatchAllowed === false && input.validation.productionWritesPerformed === false && input.validation.autoApply === false },
|
|
1277
|
-
];
|
|
1278
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
1279
|
-
const proof = {
|
|
1280
|
-
schema: "zob.autonomous-promotion-proof-plan.v1",
|
|
1281
|
-
runId: input.runId,
|
|
1282
|
-
phase: "7B",
|
|
1283
|
-
status: failedChecks.length === 0 ? "pilot_batch_promotion_proof_ready_execution_blocked" : "pilot_batch_promotion_proof_incomplete",
|
|
1284
|
-
promotionProofReady: failedChecks.length === 0,
|
|
1285
|
-
smokeToPilotPreconditionsProved: scenarios[0].passed === true,
|
|
1286
|
-
pilotExecutionBlockedProofPassed: scenarios[1].passed === true,
|
|
1287
|
-
batchExecutionBlockedProofPassed: scenarios[2].passed === true,
|
|
1288
|
-
resumeRetryPolicyProofPassed: scenarios[3].passed === true,
|
|
1289
|
-
oraclePolicyProofPassed: scenarios[4].passed === true,
|
|
1290
|
-
finalE2ERequirementCleared: false,
|
|
1291
|
-
no_ship: true,
|
|
1292
|
-
checks,
|
|
1293
|
-
failedChecks,
|
|
1294
|
-
scenarios,
|
|
1295
|
-
evidenceRefs: [
|
|
1296
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/promotion-plan.json`,
|
|
1297
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/factory-run-ref.json`,
|
|
1298
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/oracle-review.json`,
|
|
1299
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
1300
|
-
],
|
|
1301
|
-
blockers: [
|
|
1302
|
-
"pilot_execution_disabled",
|
|
1303
|
-
"batch_execution_disabled",
|
|
1304
|
-
"pilot_oracle_missing",
|
|
1305
|
-
"batch_concurrency_cap_missing",
|
|
1306
|
-
"final_e2e_pilot_batch_evidence_required",
|
|
1307
|
-
],
|
|
1308
|
-
pilotExecutionAllowed: false,
|
|
1309
|
-
batchExecutionAllowed: false,
|
|
1310
|
-
pilotExecuted: false,
|
|
1311
|
-
batchExecuted: false,
|
|
1312
|
-
pilotOraclePass: false,
|
|
1313
|
-
batchConcurrencyCapPresent: false,
|
|
1314
|
-
childDispatchAllowed: false,
|
|
1315
|
-
daemonStarted: false,
|
|
1316
|
-
productionWritesPerformed: false,
|
|
1317
|
-
autoApply: false,
|
|
1318
|
-
noExecutionBeyondSmoke: true,
|
|
1319
|
-
globalAutonomyReady: false,
|
|
1320
|
-
globalAutonomyNoShip: true,
|
|
1321
|
-
bodyStored: false,
|
|
1322
|
-
promptBodiesStored: false,
|
|
1323
|
-
outputBodiesStored: false,
|
|
1324
|
-
};
|
|
1325
|
-
if (hasForbiddenBodyKeys(proof)) throw new Error("autonomous promotion proof plan would store forbidden plaintext body keys");
|
|
1326
|
-
return proof;
|
|
1327
|
-
}
|
|
1328
|
-
|
|
1329
|
-
function buildAutonomousSchedulerPlan(repoRoot: string, input: { runId: string; promotionPlan: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1330
|
-
const daemonPolicy = validateDaemonPolicyConfig(repoRoot);
|
|
1331
|
-
const allowedJobTypes = Array.isArray(daemonPolicy.allowedJobTypes) ? daemonPolicy.allowedJobTypes.filter((item): item is string => typeof item === "string") : [];
|
|
1332
|
-
const requiredStopConditions = Array.isArray(daemonPolicy.requiredStopConditions) ? daemonPolicy.requiredStopConditions.filter((item): item is string => typeof item === "string") : [];
|
|
1333
|
-
const promotionReady = isRecord(input.promotionPlan.smokeGate) && input.promotionPlan.smokeGate.passed === true;
|
|
1334
|
-
const checks = [
|
|
1335
|
-
{ name: "daemon_policy_valid", passed: daemonPolicy.present === true && daemonPolicy.valid === true },
|
|
1336
|
-
{ name: "daemon_disabled_by_default", passed: daemonPolicy.daemonEnabled === false && daemonPolicy.alwaysOnDaemonEnabled === false && daemonPolicy.autoStartDaemon === false && daemonPolicy.continuousLoop === false },
|
|
1337
|
-
{ name: "one_worker_bounded", passed: daemonPolicy.maxWorkersDefault === 1 && daemonPolicy.maxWorkersMax === 1 },
|
|
1338
|
-
{ name: "read_only_work_kinds_only", passed: allowedJobTypes.length > 0 && daemonPolicy.writeAdaptersEnabled === false && daemonPolicy.productionWritesPerformed === false && daemonPolicy.autoApply === false },
|
|
1339
|
-
{ name: "terminal_stop_conditions_available", passed: requiredStopConditions.includes("timeout") && requiredStopConditions.includes("blocked") && requiredStopConditions.includes("fail_loop") && requiredStopConditions.includes("no_ship") },
|
|
1340
|
-
{ name: "kill_switch_required", passed: true },
|
|
1341
|
-
{ name: "leases_and_heartbeat_required", passed: true },
|
|
1342
|
-
{ name: "budget_required_before_dispatch", passed: true },
|
|
1343
|
-
{ name: "scheduler_execution_disabled", passed: true },
|
|
1344
|
-
];
|
|
1345
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
1346
|
-
const plan = {
|
|
1347
|
-
schema: "zob.autonomous-scheduler-plan.v1",
|
|
1348
|
-
runId: input.runId,
|
|
1349
|
-
schedulerPlanReady: failedChecks.length === 0,
|
|
1350
|
-
failedChecks,
|
|
1351
|
-
checks,
|
|
1352
|
-
promotionReady,
|
|
1353
|
-
daemonPolicy,
|
|
1354
|
-
queuePolicy: {
|
|
1355
|
-
allowedWorkKinds: allowedJobTypes,
|
|
1356
|
-
initialWorkKinds: ["docs_watch", "repo_audit_readonly", "todo_risk_report", "session_analysis"].filter((kind) => allowedJobTypes.includes(kind)),
|
|
1357
|
-
writeWorkAccepted: false,
|
|
1358
|
-
claimAtMostOnePerTick: daemonPolicy.claimAtMostOneJobPerTick === true,
|
|
1359
|
-
boundedQueueOnly: true,
|
|
1360
|
-
},
|
|
1361
|
-
leasePolicy: {
|
|
1362
|
-
leaseRequired: true,
|
|
1363
|
-
leaseTtlMs: 300_000,
|
|
1364
|
-
maxLeaseRenewals: 0,
|
|
1365
|
-
staleLeaseFailsCleanly: true,
|
|
1366
|
-
leaseOwnerHashOnly: true,
|
|
1367
|
-
},
|
|
1368
|
-
heartbeatPolicy: {
|
|
1369
|
-
heartbeatRequired: true,
|
|
1370
|
-
heartbeatIntervalMs: 30_000,
|
|
1371
|
-
staleHeartbeatTimeoutMs: 90_000,
|
|
1372
|
-
staleHeartbeatStopCondition: "timeout",
|
|
1373
|
-
},
|
|
1374
|
-
timeoutPolicy: {
|
|
1375
|
-
perWorkItemTimeoutMs: 300_000,
|
|
1376
|
-
globalTickTimeoutMs: 600_000,
|
|
1377
|
-
timeoutStopCondition: "timeout",
|
|
1378
|
-
},
|
|
1379
|
-
retryPolicy: {
|
|
1380
|
-
retriesCapped: true,
|
|
1381
|
-
maxRetriesPerWorkItem: 1,
|
|
1382
|
-
failLoopThreshold: 3,
|
|
1383
|
-
repeatedFailureStopCondition: "fail_loop",
|
|
1384
|
-
doNotRetryNoShip: true,
|
|
1385
|
-
},
|
|
1386
|
-
killSwitch: {
|
|
1387
|
-
required: true,
|
|
1388
|
-
defaultState: "stopped_until_manual_start",
|
|
1389
|
-
stopFileRef: ".pi/queue/STOP_DAEMON.sentinel",
|
|
1390
|
-
checkedBeforeEachTick: true,
|
|
1391
|
-
stopsQueueCleanly: true,
|
|
1392
|
-
},
|
|
1393
|
-
workerPool: {
|
|
1394
|
-
defaultWorkers: 1,
|
|
1395
|
-
maxWorkers: 1,
|
|
1396
|
-
workerPoolBounded: true,
|
|
1397
|
-
},
|
|
1398
|
-
budgetPolicy: {
|
|
1399
|
-
strictBudgetRequired: true,
|
|
1400
|
-
perRunCapsRequired: true,
|
|
1401
|
-
perDayCapsRequired: true,
|
|
1402
|
-
budgetEnforced: false,
|
|
1403
|
-
dispatchBlockedUntilStrictBudgetEnforced: true,
|
|
1404
|
-
},
|
|
1405
|
-
stopConditions: requiredStopConditions,
|
|
1406
|
-
approvalPolicy: {
|
|
1407
|
-
manualStartRequiresApproval: true,
|
|
1408
|
-
alwaysOnRequiresApproval: true,
|
|
1409
|
-
schedulerAutostartAllowed: false,
|
|
1410
|
-
},
|
|
1411
|
-
schedulerExecutionAllowed: false,
|
|
1412
|
-
daemonEnabled: false,
|
|
1413
|
-
alwaysOnDaemonEnabled: false,
|
|
1414
|
-
autoStartDaemon: false,
|
|
1415
|
-
continuousLoop: false,
|
|
1416
|
-
daemonStarted: false,
|
|
1417
|
-
cronEnabled: false,
|
|
1418
|
-
childDispatchAllowed: false,
|
|
1419
|
-
liveChildExecution: false,
|
|
1420
|
-
networkAccessed: false,
|
|
1421
|
-
writeAdaptersEnabled: false,
|
|
1422
|
-
productionWritesPerformed: false,
|
|
1423
|
-
autoApply: false,
|
|
1424
|
-
pilotExecuted: false,
|
|
1425
|
-
batchExecuted: false,
|
|
1426
|
-
globalAutonomyReady: false,
|
|
1427
|
-
globalAutonomyNoShip: true,
|
|
1428
|
-
bodyStored: false,
|
|
1429
|
-
promptBodiesStored: false,
|
|
1430
|
-
outputBodiesStored: false,
|
|
1431
|
-
};
|
|
1432
|
-
if (hasForbiddenBodyKeys(plan)) throw new Error("autonomous scheduler plan would store forbidden plaintext body keys");
|
|
1433
|
-
return plan;
|
|
1434
|
-
}
|
|
1435
|
-
|
|
1436
|
-
function buildAutonomousSchedulerProofPlan(input: { runId: string; schedulerPlan: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1437
|
-
const queuePolicy = isRecord(input.schedulerPlan.queuePolicy) ? input.schedulerPlan.queuePolicy : {};
|
|
1438
|
-
const leasePolicy = isRecord(input.schedulerPlan.leasePolicy) ? input.schedulerPlan.leasePolicy : {};
|
|
1439
|
-
const heartbeatPolicy = isRecord(input.schedulerPlan.heartbeatPolicy) ? input.schedulerPlan.heartbeatPolicy : {};
|
|
1440
|
-
const timeoutPolicy = isRecord(input.schedulerPlan.timeoutPolicy) ? input.schedulerPlan.timeoutPolicy : {};
|
|
1441
|
-
const retryPolicy = isRecord(input.schedulerPlan.retryPolicy) ? input.schedulerPlan.retryPolicy : {};
|
|
1442
|
-
const killSwitch = isRecord(input.schedulerPlan.killSwitch) ? input.schedulerPlan.killSwitch : {};
|
|
1443
|
-
const workerPool = isRecord(input.schedulerPlan.workerPool) ? input.schedulerPlan.workerPool : {};
|
|
1444
|
-
const budgetPolicy = isRecord(input.schedulerPlan.budgetPolicy) ? input.schedulerPlan.budgetPolicy : {};
|
|
1445
|
-
const approvalPolicy = isRecord(input.schedulerPlan.approvalPolicy) ? input.schedulerPlan.approvalPolicy : {};
|
|
1446
|
-
const daemonPolicy = isRecord(input.schedulerPlan.daemonPolicy) ? input.schedulerPlan.daemonPolicy : {};
|
|
1447
|
-
const stopConditions = Array.isArray(input.schedulerPlan.stopConditions) ? input.schedulerPlan.stopConditions.filter((condition): condition is string => typeof condition === "string") : [];
|
|
1448
|
-
const scenarios = [
|
|
1449
|
-
{
|
|
1450
|
-
name: "daemon_default_disabled",
|
|
1451
|
-
passed: input.schedulerPlan.schedulerExecutionAllowed === false && input.schedulerPlan.daemonStarted === false && input.schedulerPlan.autoStartDaemon === false && input.schedulerPlan.continuousLoop === false && input.schedulerPlan.cronEnabled === false,
|
|
1452
|
-
dispatchAllowed: false,
|
|
1453
|
-
},
|
|
1454
|
-
{
|
|
1455
|
-
name: "one_worker_bounded_readonly_queue",
|
|
1456
|
-
passed: workerPool.defaultWorkers === 1 && workerPool.maxWorkers === 1 && workerPool.workerPoolBounded === true && queuePolicy.writeWorkAccepted === false && queuePolicy.claimAtMostOnePerTick === true,
|
|
1457
|
-
dispatchAllowed: false,
|
|
1458
|
-
},
|
|
1459
|
-
{
|
|
1460
|
-
name: "lease_heartbeat_timeout_stop_conditions_required",
|
|
1461
|
-
passed: leasePolicy.leaseRequired === true && heartbeatPolicy.heartbeatRequired === true && heartbeatPolicy.staleHeartbeatStopCondition === "timeout" && timeoutPolicy.timeoutStopCondition === "timeout" && stopConditions.includes("timeout") && stopConditions.includes("blocked") && stopConditions.includes("fail_loop") && stopConditions.includes("no_ship"),
|
|
1462
|
-
dispatchAllowed: false,
|
|
1463
|
-
},
|
|
1464
|
-
{
|
|
1465
|
-
name: "kill_switch_and_retry_stop_clean",
|
|
1466
|
-
passed: killSwitch.required === true && killSwitch.checkedBeforeEachTick === true && killSwitch.stopsQueueCleanly === true && retryPolicy.retriesCapped === true && retryPolicy.doNotRetryNoShip === true && retryPolicy.repeatedFailureStopCondition === "fail_loop",
|
|
1467
|
-
dispatchAllowed: false,
|
|
1468
|
-
},
|
|
1469
|
-
{
|
|
1470
|
-
name: "strict_budget_required_before_daemon_dispatch",
|
|
1471
|
-
passed: budgetPolicy.strictBudgetRequired === true && budgetPolicy.perRunCapsRequired === true && budgetPolicy.perDayCapsRequired === true && budgetPolicy.budgetEnforced === false && budgetPolicy.dispatchBlockedUntilStrictBudgetEnforced === true,
|
|
1472
|
-
dispatchAllowed: false,
|
|
1473
|
-
},
|
|
1474
|
-
{
|
|
1475
|
-
name: "always_on_requires_future_approval",
|
|
1476
|
-
passed: approvalPolicy.manualStartRequiresApproval === true && approvalPolicy.alwaysOnRequiresApproval === true && approvalPolicy.schedulerAutostartAllowed === false && daemonPolicy.alwaysOnDaemonEnabled === false,
|
|
1477
|
-
dispatchAllowed: false,
|
|
1478
|
-
},
|
|
1479
|
-
];
|
|
1480
|
-
const checks = [
|
|
1481
|
-
{ name: "scheduler_plan_ready", passed: input.schedulerPlan.schedulerPlanReady === true },
|
|
1482
|
-
{ name: "daemon_default_disabled", passed: scenarios[0].passed === true },
|
|
1483
|
-
{ name: "one_worker_bounded_readonly_queue", passed: scenarios[1].passed === true },
|
|
1484
|
-
{ name: "lease_heartbeat_timeout_stop_conditions_required", passed: scenarios[2].passed === true },
|
|
1485
|
-
{ name: "kill_switch_and_retry_stop_clean", passed: scenarios[3].passed === true },
|
|
1486
|
-
{ name: "strict_budget_required_before_daemon_dispatch", passed: scenarios[4].passed === true },
|
|
1487
|
-
{ name: "always_on_requires_future_approval", passed: scenarios[5].passed === true },
|
|
1488
|
-
{ name: "run_scope_no_global_autonomy", passed: input.validation.globalAutonomyReady === false && input.validation.globalAutonomyNoShip === true && input.validation.childDispatchAllowed === false && input.validation.productionWritesPerformed === false && input.validation.autoApply === false },
|
|
1489
|
-
];
|
|
1490
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
1491
|
-
const proof = {
|
|
1492
|
-
schema: "zob.autonomous-scheduler-proof-plan.v1",
|
|
1493
|
-
runId: input.runId,
|
|
1494
|
-
phase: "8B",
|
|
1495
|
-
status: failedChecks.length === 0 ? "scheduler_daemon_proof_ready_execution_blocked" : "scheduler_daemon_proof_incomplete",
|
|
1496
|
-
schedulerProofReady: failedChecks.length === 0,
|
|
1497
|
-
daemonDefaultDisabledProofPassed: scenarios[0].passed === true,
|
|
1498
|
-
oneWorkerBoundedProofPassed: scenarios[1].passed === true,
|
|
1499
|
-
stopConditionsProofPassed: scenarios[2].passed === true,
|
|
1500
|
-
killSwitchRetryProofPassed: scenarios[3].passed === true,
|
|
1501
|
-
strictBudgetBeforeDispatchProofPassed: scenarios[4].passed === true,
|
|
1502
|
-
alwaysOnApprovalProofPassed: scenarios[5].passed === true,
|
|
1503
|
-
finalE2ERequirementCleared: false,
|
|
1504
|
-
no_ship: true,
|
|
1505
|
-
checks,
|
|
1506
|
-
failedChecks,
|
|
1507
|
-
scenarios,
|
|
1508
|
-
evidenceRefs: [
|
|
1509
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/scheduler-plan.json`,
|
|
1510
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
1511
|
-
".pi/daemon-policy.json",
|
|
1512
|
-
],
|
|
1513
|
-
blockers: [
|
|
1514
|
-
"daemon_scheduler_not_started",
|
|
1515
|
-
"always_on_daemon_not_enabled",
|
|
1516
|
-
"daemon_autostart_disabled",
|
|
1517
|
-
"final_e2e_daemon_scheduler_evidence_required",
|
|
1518
|
-
],
|
|
1519
|
-
schedulerExecutionAllowed: false,
|
|
1520
|
-
daemonStarted: false,
|
|
1521
|
-
autoStartDaemon: false,
|
|
1522
|
-
continuousLoop: false,
|
|
1523
|
-
cronEnabled: false,
|
|
1524
|
-
childDispatchAllowed: false,
|
|
1525
|
-
liveChildExecution: false,
|
|
1526
|
-
writeAdaptersEnabled: false,
|
|
1527
|
-
productionWritesPerformed: false,
|
|
1528
|
-
autoApply: false,
|
|
1529
|
-
noExecutionBeyondSmoke: true,
|
|
1530
|
-
globalAutonomyReady: false,
|
|
1531
|
-
globalAutonomyNoShip: true,
|
|
1532
|
-
bodyStored: false,
|
|
1533
|
-
promptBodiesStored: false,
|
|
1534
|
-
outputBodiesStored: false,
|
|
1535
|
-
};
|
|
1536
|
-
if (hasForbiddenBodyKeys(proof)) throw new Error("autonomous scheduler proof plan would store forbidden plaintext body keys");
|
|
1537
|
-
return proof;
|
|
1538
|
-
}
|
|
1539
|
-
|
|
1540
|
-
function buildAutonomousMissionControlPlan(repoRoot: string, input: { runId: string; schedulerPlan: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1541
|
-
const team = loadTeamDefinition(repoRoot, "zob-core");
|
|
1542
|
-
const teamErrors = [...team.errors, ...validateTeamDefinition(repoRoot, team.definition)];
|
|
1543
|
-
const definition = team.definition;
|
|
1544
|
-
const snapshot = definition ? buildMissionControlSnapshot(repoRoot, definition, { runId: input.runId, limit: 5 }) : undefined;
|
|
1545
|
-
const communicationAudit = definition ? buildZobCommunicationReadinessAudit(repoRoot, definition) : undefined;
|
|
1546
|
-
const transport = buildZobComsTransportReadiness(repoRoot);
|
|
1547
|
-
const orchestratorId = definition?.orchestrator.id ?? "orchestrator";
|
|
1548
|
-
const commandProposals = definition ? MISSION_CONTROL_COMMANDS.map((command) => buildMissionControlCommandProposal(definition, {
|
|
1549
|
-
proposalId: `auto-${sha256(`${input.runId}-${command}`).slice(0, 16)}-${command}`,
|
|
1550
|
-
runId: input.runId,
|
|
1551
|
-
command,
|
|
1552
|
-
targetRole: orchestratorId,
|
|
1553
|
-
rationaleHash: sha256(`autonomous mission control ${command}`),
|
|
1554
|
-
artifactRefs: [`reports/autonomous-runs/${input.runId}/validation.json`],
|
|
1555
|
-
})) : [];
|
|
1556
|
-
const directWorkerErrors = definition && definition.workers[0] ? (() => {
|
|
1557
|
-
try {
|
|
1558
|
-
buildMissionControlCommandProposal(definition, {
|
|
1559
|
-
proposalId: `auto-${sha256(`${input.runId}-blocked-worker`).slice(0, 16)}-blocked-worker`,
|
|
1560
|
-
runId: input.runId,
|
|
1561
|
-
command: "stop",
|
|
1562
|
-
targetRole: definition.workers[0].id,
|
|
1563
|
-
rationaleHash: sha256("direct worker blocked"),
|
|
1564
|
-
});
|
|
1565
|
-
return [] as string[];
|
|
1566
|
-
} catch (error) {
|
|
1567
|
-
return [error instanceof Error ? error.message : String(error)];
|
|
1568
|
-
}
|
|
1569
|
-
})() : ["no worker role available for direct worker guard"];
|
|
1570
|
-
const transportFailedChecks = Array.isArray(transport.failedChecks) ? transport.failedChecks.filter((check): check is string => typeof check === "string") : [];
|
|
1571
|
-
const communicationFailedChecks = Array.isArray(communicationAudit?.failedChecks) ? communicationAudit.failedChecks.filter((check): check is string => typeof check === "string") : [];
|
|
1572
|
-
const communicationChecks = Array.isArray(communicationAudit?.checks) ? communicationAudit.checks.filter(isRecord) : [];
|
|
1573
|
-
const communicationCheckPassed = (name: string): boolean => communicationChecks.some((check) => check.name === name && check.passed === true);
|
|
1574
|
-
const livePeerAbsenceBlocksDelivery = transport.mode === "required_local"
|
|
1575
|
-
&& transport.networkEnabled === false
|
|
1576
|
-
&& transport.dispatchAllowed === true
|
|
1577
|
-
&& transportFailedChecks.length > 0
|
|
1578
|
-
&& transportFailedChecks.every((check) => check === "required_local_live_ready_when_enabled");
|
|
1579
|
-
const proposalOnlyCommunicationSafety = communicationAudit?.verdict === "PASS" && communicationAudit.no_ship === false
|
|
1580
|
-
|| (livePeerAbsenceBlocksDelivery
|
|
1581
|
-
&& communicationFailedChecks.length > 0
|
|
1582
|
-
&& communicationFailedChecks.every((check) => check === "transport_policy_safe" || check === "registry_observe_only_available")
|
|
1583
|
-
&& communicationCheckPassed("topology_guard_active")
|
|
1584
|
-
&& communicationCheckPassed("worker_to_worker_blocked")
|
|
1585
|
-
&& communicationCheckPassed("message_body_storage_blocked")
|
|
1586
|
-
&& communicationCheckPassed("existing_ledgers_body_free")
|
|
1587
|
-
&& communicationCheckPassed("stale_transport_not_completion")
|
|
1588
|
-
&& communicationCheckPassed("dashboard_commands_are_proposals")
|
|
1589
|
-
&& communicationCheckPassed("dashboard_direct_worker_commands_blocked"));
|
|
1590
|
-
const requiredLocalTransportSafeForNoDispatchPlan = transport.mode === "required_local"
|
|
1591
|
-
&& transport.verdict === "PASS"
|
|
1592
|
-
&& transport.noExecution === true
|
|
1593
|
-
&& transport.networkEnabled === false
|
|
1594
|
-
&& transport.dispatchAllowed === true;
|
|
1595
|
-
const transportSafeForAutonomySmoke = transport.networkEnabled === false
|
|
1596
|
-
&& (
|
|
1597
|
-
(transport.verdict === "PASS" && transport.enabled === false && transport.dispatchAllowed === false)
|
|
1598
|
-
|| requiredLocalTransportSafeForNoDispatchPlan
|
|
1599
|
-
|| livePeerAbsenceBlocksDelivery
|
|
1600
|
-
|| (transport.mode === "required_local" && transport.verdict === "PASS")
|
|
1601
|
-
);
|
|
1602
|
-
const checks = [
|
|
1603
|
-
{ name: "team_topology_valid", passed: teamErrors.length === 0 },
|
|
1604
|
-
{ name: "dashboard_snapshot_available", passed: snapshot?.schema === "zob.mission-control-snapshot.v1" },
|
|
1605
|
-
{ name: "communication_readiness_pass", passed: proposalOnlyCommunicationSafety, detail: { readinessVerdict: communicationAudit?.verdict ?? "FAIL", livePeerAbsenceBlocksDelivery, livePeerReadyRequiredForDispatch: true, proposalOnlyNoDispatchPlan: true } },
|
|
1606
|
-
{ name: "transport_disabled", passed: transportSafeForAutonomySmoke, detail: { transportVerdict: transport.verdict, requiredLocalTransportSafeForNoDispatchPlan, livePeerAbsenceBlocksDelivery, dispatchAllowedInPlan: false, networkComsEnabledInPlan: false } },
|
|
1607
|
-
{ name: "commands_proposal_only", passed: commandProposals.length === MISSION_CONTROL_COMMANDS.length && commandProposals.every((proposal) => proposal.proposalOnly === true && proposal.parentOwned === true && proposal.directWorkerWrite === false && proposal.transportDispatch === false && proposal.networkTransport === false) },
|
|
1608
|
-
{ name: "direct_worker_commands_blocked", passed: directWorkerErrors.some((error) => error.includes("direct worker")) },
|
|
1609
|
-
{ name: "no_body_storage", passed: commandProposals.every((proposal) => proposal.bodyStored === false && proposal.promptBodiesStored === false && proposal.outputBodiesStored === false) },
|
|
1610
|
-
{ name: "scheduler_still_disabled", passed: input.schedulerPlan.schedulerExecutionAllowed === false && input.schedulerPlan.daemonStarted === false },
|
|
1611
|
-
];
|
|
1612
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
1613
|
-
const plan = {
|
|
1614
|
-
schema: "zob.autonomous-mission-control-plan.v1",
|
|
1615
|
-
runId: input.runId,
|
|
1616
|
-
missionControlPlanReady: failedChecks.length === 0,
|
|
1617
|
-
failedChecks,
|
|
1618
|
-
checks,
|
|
1619
|
-
dashboard: {
|
|
1620
|
-
snapshotAvailable: snapshot?.schema === "zob.mission-control-snapshot.v1",
|
|
1621
|
-
dashboardReads: ["runs", "factories", "queue", "budget", "model_routing", "coms", "blockers", "autonomy_status"],
|
|
1622
|
-
latestRunLimit: 5,
|
|
1623
|
-
directWorkerWrites: false,
|
|
1624
|
-
bypassesParentGates: false,
|
|
1625
|
-
networkComsEnabled: false,
|
|
1626
|
-
snapshotSchema: snapshot?.schema,
|
|
1627
|
-
},
|
|
1628
|
-
commandPolicy: {
|
|
1629
|
-
proposalOnly: true,
|
|
1630
|
-
parentOwned: true,
|
|
1631
|
-
directWorkerWrites: false,
|
|
1632
|
-
transportDispatch: false,
|
|
1633
|
-
networkTransport: false,
|
|
1634
|
-
allowedCommands: [...MISSION_CONTROL_COMMANDS],
|
|
1635
|
-
proposalCount: commandProposals.length,
|
|
1636
|
-
proposalHashes: commandProposals.map((proposal) => proposal.commandHash).filter((hash): hash is string => typeof hash === "string"),
|
|
1637
|
-
directWorkerCommandBlocked: directWorkerErrors.some((error) => error.includes("direct worker")),
|
|
1638
|
-
},
|
|
1639
|
-
comsPolicy: {
|
|
1640
|
-
readinessVerdict: communicationAudit?.verdict ?? "FAIL",
|
|
1641
|
-
proposalOnlyReadinessAccepted: proposalOnlyCommunicationSafety,
|
|
1642
|
-
livePeerAbsenceBlocksDelivery,
|
|
1643
|
-
livePeerReadyRequiredForDispatch: true,
|
|
1644
|
-
topologyGuardActive: Array.isArray(communicationAudit?.checks) && communicationAudit.checks.some((check) => isRecord(check) && check.name === "topology_guard_active" && check.passed === true),
|
|
1645
|
-
hashOnlyLedgers: true,
|
|
1646
|
-
workerToWorkerFreeChatAllowed: false,
|
|
1647
|
-
rawPromptOutputStored: false,
|
|
1648
|
-
bodyStored: false,
|
|
1649
|
-
promptBodiesStored: false,
|
|
1650
|
-
outputBodiesStored: false,
|
|
1651
|
-
},
|
|
1652
|
-
transportPolicy: {
|
|
1653
|
-
enabled: false,
|
|
1654
|
-
localDispatchEnabled: false,
|
|
1655
|
-
dispatchAllowed: false,
|
|
1656
|
-
networkEnabled: false,
|
|
1657
|
-
stalePeerCountsAsCompletion: false,
|
|
1658
|
-
responseBodiesStored: false,
|
|
1659
|
-
livePeerAbsenceBlocksDelivery,
|
|
1660
|
-
livePeerReadyRequiredForDispatch: true,
|
|
1661
|
-
futureHeartbeatRequired: true,
|
|
1662
|
-
futureStaleDetectionRequired: true,
|
|
1663
|
-
futureActiveWorkerRegistryRequired: true,
|
|
1664
|
-
futureResponseCaptureRequired: true,
|
|
1665
|
-
},
|
|
1666
|
-
autonomyStatus: {
|
|
1667
|
-
globalAutonomyReady: false,
|
|
1668
|
-
globalAutonomyNoShip: true,
|
|
1669
|
-
noGlobalAutonomyClaim: true,
|
|
1670
|
-
},
|
|
1671
|
-
schedulerExecutionAllowed: false,
|
|
1672
|
-
daemonStarted: false,
|
|
1673
|
-
childDispatchAllowed: false,
|
|
1674
|
-
directWorkerWrites: false,
|
|
1675
|
-
transportDispatch: false,
|
|
1676
|
-
networkComsEnabled: false,
|
|
1677
|
-
productionWritesPerformed: false,
|
|
1678
|
-
autoApply: false,
|
|
1679
|
-
globalAutonomyReady: false,
|
|
1680
|
-
globalAutonomyNoShip: true,
|
|
1681
|
-
bodyStored: false,
|
|
1682
|
-
promptBodiesStored: false,
|
|
1683
|
-
outputBodiesStored: false,
|
|
1684
|
-
};
|
|
1685
|
-
if (hasForbiddenBodyKeys(plan)) throw new Error("autonomous mission control plan would store forbidden plaintext body keys");
|
|
1686
|
-
return plan;
|
|
1687
|
-
}
|
|
1688
|
-
|
|
1689
|
-
function buildAutonomousMissionControlProofPlan(input: { runId: string; missionControlPlan: Record<string, unknown>; schedulerProofPlan: Record<string, unknown>; modelRoutingProofPlan: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1690
|
-
const commandPolicy = isRecord(input.missionControlPlan.commandPolicy) ? input.missionControlPlan.commandPolicy : {};
|
|
1691
|
-
const comsPolicy = isRecord(input.missionControlPlan.comsPolicy) ? input.missionControlPlan.comsPolicy : {};
|
|
1692
|
-
const transportPolicy = isRecord(input.missionControlPlan.transportPolicy) ? input.missionControlPlan.transportPolicy : {};
|
|
1693
|
-
const allowedCommands = Array.isArray(commandPolicy.allowedCommands) ? commandPolicy.allowedCommands.filter((command): command is string => typeof command === "string") : [];
|
|
1694
|
-
const proposalHashes = Array.isArray(commandPolicy.proposalHashes) ? commandPolicy.proposalHashes.filter((hash): hash is string => typeof hash === "string") : [];
|
|
1695
|
-
const proposalHashesValid = proposalHashes.length === MISSION_CONTROL_COMMANDS.length && proposalHashes.every((hash) => /^[a-f0-9]{64}$/.test(hash));
|
|
1696
|
-
const allMissionCommandsCovered = MISSION_CONTROL_COMMANDS.every((command) => allowedCommands.includes(command));
|
|
1697
|
-
const approvalCommandAvailable = allowedCommands.includes("approve");
|
|
1698
|
-
const scenarios = [
|
|
1699
|
-
{
|
|
1700
|
-
name: "proposal_only_commands_parent_owned",
|
|
1701
|
-
passed: commandPolicy.proposalOnly === true && commandPolicy.parentOwned === true && commandPolicy.transportDispatch === false && commandPolicy.networkTransport === false && commandPolicy.proposalCount === MISSION_CONTROL_COMMANDS.length && allMissionCommandsCovered && proposalHashesValid,
|
|
1702
|
-
dispatchAllowed: false,
|
|
1703
|
-
},
|
|
1704
|
-
{
|
|
1705
|
-
name: "topology_hash_only_comms",
|
|
1706
|
-
passed: comsPolicy.topologyGuardActive === true && comsPolicy.hashOnlyLedgers === true && comsPolicy.workerToWorkerFreeChatAllowed === false && comsPolicy.rawPromptOutputStored === false && comsPolicy.bodyStored === false && comsPolicy.promptBodiesStored === false && comsPolicy.outputBodiesStored === false,
|
|
1707
|
-
dispatchAllowed: false,
|
|
1708
|
-
},
|
|
1709
|
-
{
|
|
1710
|
-
name: "direct_worker_commands_blocked",
|
|
1711
|
-
passed: commandPolicy.directWorkerWrites === false && commandPolicy.directWorkerCommandBlocked === true && input.missionControlPlan.directWorkerWrites === false,
|
|
1712
|
-
dispatchAllowed: false,
|
|
1713
|
-
},
|
|
1714
|
-
{
|
|
1715
|
-
name: "live_transport_and_network_disabled",
|
|
1716
|
-
passed: transportPolicy.enabled === false && transportPolicy.localDispatchEnabled === false && transportPolicy.dispatchAllowed === false && transportPolicy.networkEnabled === false && input.missionControlPlan.transportDispatch === false && input.missionControlPlan.networkComsEnabled === false,
|
|
1717
|
-
dispatchAllowed: false,
|
|
1718
|
-
},
|
|
1719
|
-
{
|
|
1720
|
-
name: "approval_required_before_live_global_routing",
|
|
1721
|
-
passed: approvalCommandAvailable && commandPolicy.proposalOnly === true && commandPolicy.parentOwned === true && input.modelRoutingProofPlan.globalLiveRoutingEnabled === false && input.modelRoutingProofPlan.liveAutonomousRoutingApplied === false && input.modelRoutingProofPlan.finalE2ERequirementCleared === false && input.modelRoutingProofPlan.no_ship === true,
|
|
1722
|
-
dispatchAllowed: false,
|
|
1723
|
-
},
|
|
1724
|
-
{
|
|
1725
|
-
name: "post_8b_scheduler_still_blocked",
|
|
1726
|
-
passed: input.schedulerProofPlan.schedulerProofReady === true && input.schedulerProofPlan.schedulerExecutionAllowed === false && input.schedulerProofPlan.daemonStarted === false && input.schedulerProofPlan.childDispatchAllowed === false,
|
|
1727
|
-
dispatchAllowed: false,
|
|
1728
|
-
},
|
|
1729
|
-
];
|
|
1730
|
-
const checks = [
|
|
1731
|
-
{ name: "mission_control_plan_ready", passed: input.missionControlPlan.missionControlPlanReady === true },
|
|
1732
|
-
{ name: "proposal_only_commands_parent_owned", passed: scenarios[0].passed === true },
|
|
1733
|
-
{ name: "topology_hash_only_comms", passed: scenarios[1].passed === true },
|
|
1734
|
-
{ name: "direct_worker_commands_blocked", passed: scenarios[2].passed === true },
|
|
1735
|
-
{ name: "live_transport_and_network_disabled", passed: scenarios[3].passed === true },
|
|
1736
|
-
{ name: "approval_required_before_live_global_routing", passed: scenarios[4].passed === true },
|
|
1737
|
-
{ name: "post_8b_scheduler_still_blocked", passed: scenarios[5].passed === true },
|
|
1738
|
-
{ name: "run_scope_no_global_autonomy", passed: input.validation.globalAutonomyReady === false && input.validation.globalAutonomyNoShip === true && input.validation.childDispatchAllowed === false && input.validation.productionWritesPerformed === false && input.validation.autoApply === false && input.missionControlPlan.globalAutonomyReady === false && input.missionControlPlan.globalAutonomyNoShip === true },
|
|
1739
|
-
];
|
|
1740
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
1741
|
-
const proof = {
|
|
1742
|
-
schema: "zob.autonomous-mission-control-proof-plan.v1",
|
|
1743
|
-
runId: input.runId,
|
|
1744
|
-
phase: "9B",
|
|
1745
|
-
status: failedChecks.length === 0 ? "mission_control_comms_proof_ready_execution_blocked" : "mission_control_comms_proof_incomplete",
|
|
1746
|
-
missionControlProofReady: failedChecks.length === 0,
|
|
1747
|
-
proposalOnlyCommandsProofPassed: scenarios[0].passed === true,
|
|
1748
|
-
topologyHashOnlyCommsProofPassed: scenarios[1].passed === true,
|
|
1749
|
-
directWorkerCommandsBlockedProofPassed: scenarios[2].passed === true,
|
|
1750
|
-
liveTransportNetworkDisabledProofPassed: scenarios[3].passed === true,
|
|
1751
|
-
liveGlobalRoutingApprovalProofPassed: scenarios[4].passed === true,
|
|
1752
|
-
post8bSchedulerBlockedProofPassed: scenarios[5].passed === true,
|
|
1753
|
-
finalE2ERequirementCleared: false,
|
|
1754
|
-
no_ship: true,
|
|
1755
|
-
checks,
|
|
1756
|
-
failedChecks,
|
|
1757
|
-
scenarios,
|
|
1758
|
-
commandProposalProof: {
|
|
1759
|
-
proposalOnly: true,
|
|
1760
|
-
parentOwned: true,
|
|
1761
|
-
allowedCommands,
|
|
1762
|
-
proposalCount: commandPolicy.proposalCount,
|
|
1763
|
-
proposalHashesValid,
|
|
1764
|
-
directWorkerCommandBlocked: commandPolicy.directWorkerCommandBlocked === true,
|
|
1765
|
-
transportDispatch: false,
|
|
1766
|
-
networkTransport: false,
|
|
1767
|
-
},
|
|
1768
|
-
comsTopologyProof: {
|
|
1769
|
-
topologyGuardActive: comsPolicy.topologyGuardActive === true,
|
|
1770
|
-
hashOnlyLedgers: comsPolicy.hashOnlyLedgers === true,
|
|
1771
|
-
workerToWorkerFreeChatAllowed: false,
|
|
1772
|
-
rawPromptOutputStored: false,
|
|
1773
|
-
bodyStored: false,
|
|
1774
|
-
promptBodiesStored: false,
|
|
1775
|
-
outputBodiesStored: false,
|
|
1776
|
-
},
|
|
1777
|
-
transportProof: {
|
|
1778
|
-
enabled: false,
|
|
1779
|
-
localDispatchEnabled: false,
|
|
1780
|
-
dispatchAllowed: false,
|
|
1781
|
-
networkEnabled: false,
|
|
1782
|
-
stalePeerCountsAsCompletion: false,
|
|
1783
|
-
responseBodiesStored: false,
|
|
1784
|
-
futureHeartbeatRequired: transportPolicy.futureHeartbeatRequired === true,
|
|
1785
|
-
futureStaleDetectionRequired: transportPolicy.futureStaleDetectionRequired === true,
|
|
1786
|
-
futureActiveWorkerRegistryRequired: transportPolicy.futureActiveWorkerRegistryRequired === true,
|
|
1787
|
-
futureResponseCaptureRequired: transportPolicy.futureResponseCaptureRequired === true,
|
|
1788
|
-
},
|
|
1789
|
-
routingApprovalPolicy: {
|
|
1790
|
-
approvalCommandAvailable,
|
|
1791
|
-
approvalCommandProposalOnly: true,
|
|
1792
|
-
liveRoutingRequiresApproval: true,
|
|
1793
|
-
globalRoutingRequiresApproval: true,
|
|
1794
|
-
approvalDoesNotEnableRoutingInThisRun: true,
|
|
1795
|
-
liveRoutingAllowedAfterThisProof: false,
|
|
1796
|
-
globalLiveRoutingAllowedAfterThisProof: false,
|
|
1797
|
-
},
|
|
1798
|
-
evidenceRefs: [
|
|
1799
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/mission-control-plan.json`,
|
|
1800
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/scheduler-proof-plan.json`,
|
|
1801
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/model-routing-proof-plan.json`,
|
|
1802
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
1803
|
-
".pi/teams/zob-core.json",
|
|
1804
|
-
".pi/mission-control/zob_coms_transport.json",
|
|
1805
|
-
".pi/coms/messages.jsonl",
|
|
1806
|
-
".pi/coms/status.jsonl",
|
|
1807
|
-
],
|
|
1808
|
-
blockers: [
|
|
1809
|
-
"live_transport_not_enabled",
|
|
1810
|
-
"network_coms_disabled",
|
|
1811
|
-
"global_live_routing_approval_not_granted",
|
|
1812
|
-
"final_e2e_mission_control_live_comms_evidence_required",
|
|
1813
|
-
],
|
|
1814
|
-
schedulerExecutionAllowed: false,
|
|
1815
|
-
daemonStarted: false,
|
|
1816
|
-
childDispatchAllowed: false,
|
|
1817
|
-
liveChildExecution: false,
|
|
1818
|
-
directWorkerWrites: false,
|
|
1819
|
-
transportDispatch: false,
|
|
1820
|
-
networkComsEnabled: false,
|
|
1821
|
-
liveRoutingEnabled: false,
|
|
1822
|
-
globalLiveRoutingEnabled: false,
|
|
1823
|
-
productionWritesPerformed: false,
|
|
1824
|
-
autoApply: false,
|
|
1825
|
-
noExecutionBeyondSmoke: true,
|
|
1826
|
-
globalAutonomyReady: false,
|
|
1827
|
-
globalAutonomyNoShip: true,
|
|
1828
|
-
bodyStored: false,
|
|
1829
|
-
promptBodiesStored: false,
|
|
1830
|
-
outputBodiesStored: false,
|
|
1831
|
-
};
|
|
1832
|
-
if (hasForbiddenBodyKeys(proof)) throw new Error("autonomous mission control proof plan would store forbidden plaintext body keys");
|
|
1833
|
-
return proof;
|
|
1834
|
-
}
|
|
1835
|
-
|
|
1836
|
-
function buildAutonomousSandboxApplyPlan(repoRoot: string, input: { runId: string; missionControlPlan: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
1837
|
-
const isolatedPath = "reports/sandbox-isolated-execution-smoke.json";
|
|
1838
|
-
const reviewPath = "reports/sandbox-diff-review-gate-smoke.json";
|
|
1839
|
-
const simulationPath = "reports/sandbox-apply-simulation-smoke.json";
|
|
1840
|
-
const preflightPath = "reports/sandbox-manual-apply-preflight-smoke.json";
|
|
1841
|
-
const isolatedRead = readJsonArtifact(join(repoRoot, isolatedPath));
|
|
1842
|
-
const reviewRead = readJsonArtifact(join(repoRoot, reviewPath));
|
|
1843
|
-
const simulationRead = readJsonArtifact(join(repoRoot, simulationPath));
|
|
1844
|
-
const preflightRead = readJsonArtifact(join(repoRoot, preflightPath));
|
|
1845
|
-
const isolated = isRecord(isolatedRead.parsed) ? isolatedRead.parsed : {};
|
|
1846
|
-
const review = isRecord(reviewRead.parsed) ? reviewRead.parsed : {};
|
|
1847
|
-
const simulation = isRecord(simulationRead.parsed) ? simulationRead.parsed : {};
|
|
1848
|
-
const preflight = isRecord(preflightRead.parsed) ? preflightRead.parsed : {};
|
|
1849
|
-
const isolatedDiffHash = typeof isolated.diffHash === "string" ? isolated.diffHash : undefined;
|
|
1850
|
-
const reviewDiffHash = typeof review.diffHash === "string" ? review.diffHash : undefined;
|
|
1851
|
-
const simulationDiffHash = typeof simulation.diffHash === "string" ? simulation.diffHash : undefined;
|
|
1852
|
-
const diffHashesMatch = Boolean(isolatedDiffHash && isolatedDiffHash === reviewDiffHash && reviewDiffHash === simulationDiffHash);
|
|
1853
|
-
const checks = [
|
|
1854
|
-
{ name: "isolated_temp_workspace_executed", passed: isolated.status === "executed_in_sandbox" && isolated.isolatedExecutionPerformed === true && isolated.productionWritesPerformed === false },
|
|
1855
|
-
{ name: "diff_hash_consistent", passed: diffHashesMatch },
|
|
1856
|
-
{ name: "oracle_diff_review_passed", passed: review.status === "diff_review_passed" && review.reviewPassed === true && review.applyReadyUnlocked === true && review.applyPerformed === false },
|
|
1857
|
-
{ name: "rollback_metadata_required", passed: isolated.rollbackPrepared === true && simulation.rollbackPrepared === true && isolated.rollbackApplied === false && simulation.rollbackApplied === false },
|
|
1858
|
-
{ name: "apply_simulation_temp_workspace_only", passed: simulation.status === "simulated_apply_in_temp_workspace" && simulation.simulatedApplyPerformed === true && simulation.tempTargetWorkspaceScoped === true && simulation.productionWritesPerformed === false },
|
|
1859
|
-
{ name: "manual_apply_preflight_packet_ready", passed: preflight.status === "manual_apply_preflight_passed" && preflight.manualApplyPreflightPassed === true && preflight.executionAllowedByThisTool === false && preflight.realApplyExecuted === false && preflight.productionWritesPerformed === false },
|
|
1860
|
-
{ name: "manual_approval_required", passed: isolated.manualApplyRequired === true && review.manualApplyRequired === true && simulation.manualApplyRequired === true && preflight.manualApplyRequired === true && isolated.humanApprovalRequired === true && review.humanApprovalRequired === true && simulation.humanApprovalRequired === true && preflight.humanApprovalRequired === true },
|
|
1861
|
-
{ name: "auto_apply_disabled", passed: isolated.autoApply === false && review.autoApply === false && simulation.autoApply === false && preflight.autoApply === false },
|
|
1862
|
-
{ name: "mission_control_still_proposal_only", passed: input.missionControlPlan.directWorkerWrites === false && input.missionControlPlan.transportDispatch === false },
|
|
1863
|
-
];
|
|
1864
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
1865
|
-
const plan = {
|
|
1866
|
-
schema: "zob.autonomous-sandbox-apply-plan.v1",
|
|
1867
|
-
runId: input.runId,
|
|
1868
|
-
sandboxApplyPlanReady: failedChecks.length === 0,
|
|
1869
|
-
failedChecks,
|
|
1870
|
-
checks,
|
|
1871
|
-
evidenceRefs: [isolatedPath, reviewPath, simulationPath, preflightPath],
|
|
1872
|
-
evidenceHashes: {
|
|
1873
|
-
isolatedExecution: isolatedRead.hash,
|
|
1874
|
-
diffReviewGate: reviewRead.hash,
|
|
1875
|
-
applySimulation: simulationRead.hash,
|
|
1876
|
-
manualApplyPreflight: preflightRead.hash,
|
|
1877
|
-
},
|
|
1878
|
-
sandboxRunIdHash: typeof isolated.runId === "string" ? sha256(isolated.runId) : undefined,
|
|
1879
|
-
isolatedTempWorkspace: {
|
|
1880
|
-
required: true,
|
|
1881
|
-
executed: isolated.isolatedExecutionPerformed === true,
|
|
1882
|
-
workspacePathStored: false,
|
|
1883
|
-
productionWritesPerformed: false,
|
|
1884
|
-
},
|
|
1885
|
-
applyPlan: {
|
|
1886
|
-
required: true,
|
|
1887
|
-
plaintextContentStored: false,
|
|
1888
|
-
allowedPathsRequired: true,
|
|
1889
|
-
forbiddenPathsRequired: true,
|
|
1890
|
-
writeTargetsMustExcludeSecretsVendorGenerated: true,
|
|
1891
|
-
},
|
|
1892
|
-
diffGate: {
|
|
1893
|
-
diffHashRequired: true,
|
|
1894
|
-
diffHash: isolatedDiffHash,
|
|
1895
|
-
diffHashesMatch,
|
|
1896
|
-
changedPathCount: Array.isArray(isolated.changedPaths) ? isolated.changedPaths.length : undefined,
|
|
1897
|
-
rawDiffStored: false,
|
|
1898
|
-
},
|
|
1899
|
-
oracleDiffReview: {
|
|
1900
|
-
required: true,
|
|
1901
|
-
reviewPassed: review.reviewPassed === true,
|
|
1902
|
-
applyReadyUnlocked: review.applyReadyUnlocked === true,
|
|
1903
|
-
liveOracleDispatchRequiredBeforeRealApply: true,
|
|
1904
|
-
},
|
|
1905
|
-
rollbackPolicy: {
|
|
1906
|
-
rollbackMetadataRequired: true,
|
|
1907
|
-
rollbackPrepared: isolated.rollbackPrepared === true && simulation.rollbackPrepared === true,
|
|
1908
|
-
rollbackApplied: false,
|
|
1909
|
-
rollbackSnapshotRequiredBeforeMainApply: true,
|
|
1910
|
-
},
|
|
1911
|
-
approvalPolicy: {
|
|
1912
|
-
manualApprovalRequired: true,
|
|
1913
|
-
humanApprovalRequired: true,
|
|
1914
|
-
approvalMetadataHashOnly: true,
|
|
1915
|
-
approvedForMainWorkspaceApply: false,
|
|
1916
|
-
},
|
|
1917
|
-
applySimulation: {
|
|
1918
|
-
required: true,
|
|
1919
|
-
simulatedApplyPerformed: simulation.simulatedApplyPerformed === true,
|
|
1920
|
-
tempTargetWorkspaceScoped: simulation.tempTargetWorkspaceScoped === true,
|
|
1921
|
-
productionWritesPerformed: false,
|
|
1922
|
-
autoApply: false,
|
|
1923
|
-
},
|
|
1924
|
-
manualApplyPreflight: {
|
|
1925
|
-
required: true,
|
|
1926
|
-
preflightPassed: preflight.manualApplyPreflightPassed === true,
|
|
1927
|
-
executionAllowedByThisTool: false,
|
|
1928
|
-
confirmationPhraseMatched: isRecord(preflight.gates) && preflight.gates.confirmationPhraseMatched === true,
|
|
1929
|
-
approvalHashOnly: isRecord(preflight.gates) && preflight.gates.approvalHashOnly === true,
|
|
1930
|
-
realApplyExecuted: false,
|
|
1931
|
-
productionWritesPerformed: false,
|
|
1932
|
-
autoApply: false,
|
|
1933
|
-
},
|
|
1934
|
-
mainWorkspaceApply: {
|
|
1935
|
-
realApplyAllowed: false,
|
|
1936
|
-
realApplyExecuted: false,
|
|
1937
|
-
productionWritesPerformed: false,
|
|
1938
|
-
autoApply: false,
|
|
1939
|
-
reason: "phase_10b_manual_apply_preflight_only_no_main_workspace_apply",
|
|
1940
|
-
requiredBeforeRealApply: ["fresh_oracle_diff_review_PASS", "rollback_snapshot", "manual_human_approval", "strict_budget_gate", "allowed_paths_write_policy", "separate_manual_apply_executor", "post_apply_validation_oracle"],
|
|
1941
|
-
},
|
|
1942
|
-
writeSafety: {
|
|
1943
|
-
noDirectAutonomousWritesToMainWorkspace: true,
|
|
1944
|
-
noAutoApplyByDefault: true,
|
|
1945
|
-
noSecretsVendorGeneratedPaths: true,
|
|
1946
|
-
rollbackMetadataRequiredBeforeScalingWrites: true,
|
|
1947
|
-
},
|
|
1948
|
-
childDispatchAllowed: false,
|
|
1949
|
-
daemonStarted: false,
|
|
1950
|
-
directWorkerWrites: false,
|
|
1951
|
-
transportDispatch: false,
|
|
1952
|
-
networkComsEnabled: false,
|
|
1953
|
-
productionWritesPerformed: false,
|
|
1954
|
-
autoApply: false,
|
|
1955
|
-
realApplyExecuted: false,
|
|
1956
|
-
globalAutonomyReady: false,
|
|
1957
|
-
globalAutonomyNoShip: true,
|
|
1958
|
-
bodyStored: false,
|
|
1959
|
-
promptBodiesStored: false,
|
|
1960
|
-
outputBodiesStored: false,
|
|
1961
|
-
};
|
|
1962
|
-
if (hasForbiddenBodyKeys(plan)) throw new Error("autonomous sandbox apply plan would store forbidden plaintext body keys");
|
|
1963
|
-
return plan;
|
|
1964
|
-
}
|
|
1965
|
-
|
|
1966
|
-
function expectedAutonomousCurrentSourceFingerprintFiles(factorySelection: Record<string, unknown>): string[] {
|
|
1967
|
-
const selectedFactory = typeof factorySelection.selectedFactory === "string" ? factorySelection.selectedFactory : undefined;
|
|
1968
|
-
const factoryFiles = selectedFactory ? [`.pi/factories/${selectedFactory}/factory.json`, `.pi/factories/${selectedFactory}/smoke-manifest.json`] : [];
|
|
1969
|
-
return [...AUTONOMOUS_CURRENT_SOURCE_FINGERPRINT_FILES, ...factoryFiles].filter((file, index, items) => items.indexOf(file) === index).sort();
|
|
1970
|
-
}
|
|
1971
|
-
|
|
1972
|
-
function stringRecord(value: unknown): Record<string, string> {
|
|
1973
|
-
if (!isRecord(value)) return {};
|
|
1974
|
-
return Object.fromEntries(Object.entries(value).filter((entry): entry is [string, string] => typeof entry[1] === "string"));
|
|
1975
|
-
}
|
|
1976
|
-
|
|
1977
|
-
function sameStringArray(left: unknown, right: string[]): boolean {
|
|
1978
|
-
return Array.isArray(left) && left.every((item) => typeof item === "string") && JSON.stringify([...left].sort()) === JSON.stringify([...right].sort());
|
|
1979
|
-
}
|
|
1980
|
-
|
|
1981
|
-
function buildAutonomousCurrentSourceFingerprint(repoRoot: string, input: { runId: string; factorySelection: Record<string, unknown> }): Record<string, unknown> {
|
|
1982
|
-
const selectedFactory = typeof input.factorySelection.selectedFactory === "string" ? input.factorySelection.selectedFactory : undefined;
|
|
1983
|
-
const sourceFiles = expectedAutonomousCurrentSourceFingerprintFiles(input.factorySelection);
|
|
1984
|
-
const missingFiles = sourceFiles.filter((file) => !existsSync(join(repoRoot, file)));
|
|
1985
|
-
const fileHashes = Object.fromEntries(sourceFiles.filter((file) => existsSync(join(repoRoot, file))).map((file) => [file, sha256(readFileSync(join(repoRoot, file), "utf8"))]));
|
|
1986
|
-
const fingerprintHash = sha256(JSON.stringify(fileHashes));
|
|
1987
|
-
const fingerprint = {
|
|
1988
|
-
schema: "zob.autonomous-current-source-fingerprint.v1",
|
|
1989
|
-
runId: input.runId,
|
|
1990
|
-
phase: "11C",
|
|
1991
|
-
status: missingFiles.length === 0 ? "current_source_fingerprint_captured" : "blocked_missing_source_files",
|
|
1992
|
-
currentSourceFingerprintReady: missingFiles.length === 0,
|
|
1993
|
-
currentSourceFingerprintCaptured: missingFiles.length === 0,
|
|
1994
|
-
noMockCurrentSourceE2EProved: false,
|
|
1995
|
-
finalE2EProofReady: false,
|
|
1996
|
-
sourceFiles,
|
|
1997
|
-
sourceFileCount: sourceFiles.length,
|
|
1998
|
-
hashedFileCount: Object.keys(fileHashes).length,
|
|
1999
|
-
missingFiles,
|
|
2000
|
-
fileHashes,
|
|
2001
|
-
fingerprintHash,
|
|
2002
|
-
selectedFactory,
|
|
2003
|
-
evidencePolicy: {
|
|
2004
|
-
hashOnly: true,
|
|
2005
|
-
sourceBodiesStored: false,
|
|
2006
|
-
currentSourceOnly: true,
|
|
2007
|
-
rehashRequiredBeforeFinalE2E: true,
|
|
2008
|
-
noMockProofRequiredSeparately: true,
|
|
2009
|
-
},
|
|
2010
|
-
safety: {
|
|
2011
|
-
noExecution: true,
|
|
2012
|
-
childDispatchAllowed: false,
|
|
2013
|
-
daemonStarted: false,
|
|
2014
|
-
productionWritesPerformed: false,
|
|
2015
|
-
autoApply: false,
|
|
2016
|
-
networkAccessed: false,
|
|
2017
|
-
globalAutonomyReady: false,
|
|
2018
|
-
globalAutonomyNoShip: true,
|
|
2019
|
-
},
|
|
2020
|
-
bodyStored: false,
|
|
2021
|
-
promptBodiesStored: false,
|
|
2022
|
-
outputBodiesStored: false,
|
|
2023
|
-
};
|
|
2024
|
-
if (hasForbiddenBodyKeys(fingerprint)) throw new Error("autonomous current-source fingerprint would store forbidden plaintext body keys");
|
|
2025
|
-
return fingerprint;
|
|
2026
|
-
}
|
|
2027
|
-
|
|
2028
|
-
function validateAutonomousCurrentSourceFingerprintFreshness(repoRoot: string, fingerprint: Record<string, unknown>, factorySelection: Record<string, unknown>): Record<string, unknown> {
|
|
2029
|
-
const expectedSourceFiles = expectedAutonomousCurrentSourceFingerprintFiles(factorySelection);
|
|
2030
|
-
const currentMissingFiles = expectedSourceFiles.filter((file) => !existsSync(join(repoRoot, file)));
|
|
2031
|
-
const currentFileHashes = Object.fromEntries(expectedSourceFiles.filter((file) => existsSync(join(repoRoot, file))).map((file) => [file, sha256(readFileSync(join(repoRoot, file), "utf8"))]));
|
|
2032
|
-
const storedFileHashes = stringRecord(fingerprint.fileHashes);
|
|
2033
|
-
const currentFingerprintHash = sha256(JSON.stringify(currentFileHashes));
|
|
2034
|
-
const storedFingerprintHash = typeof fingerprint.fingerprintHash === "string" ? fingerprint.fingerprintHash : undefined;
|
|
2035
|
-
const hashValuesValid = Object.values(storedFileHashes).every((hash) => /^[a-f0-9]{64}$/.test(hash));
|
|
2036
|
-
const expectedHashKeys = Object.keys(currentFileHashes).sort();
|
|
2037
|
-
const storedHashKeys = Object.keys(storedFileHashes).sort();
|
|
2038
|
-
const fileHashesMatch = JSON.stringify(expectedHashKeys) === JSON.stringify(storedHashKeys) && expectedHashKeys.every((file) => storedFileHashes[file] === currentFileHashes[file]);
|
|
2039
|
-
const sourceFilesMatch = sameStringArray(fingerprint.sourceFiles, expectedSourceFiles);
|
|
2040
|
-
const missingFilesMatch = sameStringArray(fingerprint.missingFiles, currentMissingFiles);
|
|
2041
|
-
const fingerprintHashMatches = storedFingerprintHash === currentFingerprintHash;
|
|
2042
|
-
const fresh = fingerprint.schema === "zob.autonomous-current-source-fingerprint.v1"
|
|
2043
|
-
&& fingerprint.currentSourceFingerprintReady === true
|
|
2044
|
-
&& fingerprint.currentSourceFingerprintCaptured === true
|
|
2045
|
-
&& sourceFilesMatch
|
|
2046
|
-
&& missingFilesMatch
|
|
2047
|
-
&& currentMissingFiles.length === 0
|
|
2048
|
-
&& hashValuesValid
|
|
2049
|
-
&& fileHashesMatch
|
|
2050
|
-
&& fingerprintHashMatches
|
|
2051
|
-
&& fingerprint.noMockCurrentSourceE2EProved === false;
|
|
2052
|
-
const result = {
|
|
2053
|
-
schema: "zob.autonomous-current-source-fingerprint-freshness.v1",
|
|
2054
|
-
fresh,
|
|
2055
|
-
sourceFilesMatch,
|
|
2056
|
-
missingFilesMatch,
|
|
2057
|
-
hashValuesValid,
|
|
2058
|
-
fileHashesMatch,
|
|
2059
|
-
fingerprintHashMatches,
|
|
2060
|
-
currentFingerprintHash,
|
|
2061
|
-
storedFingerprintHash,
|
|
2062
|
-
expectedSourceFileCount: expectedSourceFiles.length,
|
|
2063
|
-
storedSourceFileCount: Array.isArray(fingerprint.sourceFiles) ? fingerprint.sourceFiles.length : 0,
|
|
2064
|
-
expectedHashedFileCount: expectedHashKeys.length,
|
|
2065
|
-
storedHashedFileCount: storedHashKeys.length,
|
|
2066
|
-
currentMissingFiles,
|
|
2067
|
-
noMockCurrentSourceE2EProved: false,
|
|
2068
|
-
noExecution: true,
|
|
2069
|
-
childDispatchAllowed: false,
|
|
2070
|
-
globalAutonomyReady: false,
|
|
2071
|
-
globalAutonomyNoShip: true,
|
|
2072
|
-
bodyStored: false,
|
|
2073
|
-
promptBodiesStored: false,
|
|
2074
|
-
outputBodiesStored: false,
|
|
2075
|
-
};
|
|
2076
|
-
if (hasForbiddenBodyKeys(result)) throw new Error("autonomous current-source freshness validation would store forbidden plaintext body keys");
|
|
2077
|
-
return result;
|
|
2078
|
-
}
|
|
2079
|
-
|
|
2080
|
-
function buildAutonomousFinalNoShipOracle(input: { runId: string; finalE2EProofPlan: Record<string, unknown>; currentSourceFingerprint: Record<string, unknown> }): Record<string, unknown> {
|
|
2081
|
-
const blockers = Array.isArray(input.finalE2EProofPlan.blockers) ? input.finalE2EProofPlan.blockers.filter((blocker): blocker is string => typeof blocker === "string") : ["final_e2e_proof_plan_missing_blockers"];
|
|
2082
|
-
const checks = [
|
|
2083
|
-
{ name: "final_e2e_proof_ready", passed: input.finalE2EProofPlan.finalE2EProofReady === true },
|
|
2084
|
-
{ name: "final_e2e_no_ship_cleared", passed: input.finalE2EProofPlan.no_ship === false },
|
|
2085
|
-
{ name: "no_mock_current_source_e2e_proved", passed: isRecord(input.finalE2EProofPlan.currentSourceProof) && input.finalE2EProofPlan.currentSourceProof.noMockCurrentSourceE2EProved === true },
|
|
2086
|
-
{ name: "current_source_fingerprint_captured", passed: input.currentSourceFingerprint.currentSourceFingerprintCaptured === true && typeof input.currentSourceFingerprint.fingerprintHash === "string" },
|
|
2087
|
-
{ name: "pilot_executed", passed: isRecord(input.finalE2EProofPlan.promotionStatus) && input.finalE2EProofPlan.promotionStatus.pilotExecuted === true },
|
|
2088
|
-
{ name: "batch_executed", passed: isRecord(input.finalE2EProofPlan.promotionStatus) && input.finalE2EProofPlan.promotionStatus.batchExecuted === true },
|
|
2089
|
-
{ name: "live_strict_budget_enforced", passed: isRecord(input.finalE2EProofPlan.liveGateStatus) && input.finalE2EProofPlan.liveGateStatus.strictBudgetEnforced === true },
|
|
2090
|
-
{ name: "live_model_routing_enabled", passed: isRecord(input.finalE2EProofPlan.liveGateStatus) && input.finalE2EProofPlan.liveGateStatus.liveModelRoutingEnabled === true },
|
|
2091
|
-
{ name: "daemon_started", passed: isRecord(input.finalE2EProofPlan.liveGateStatus) && input.finalE2EProofPlan.liveGateStatus.daemonStarted === true },
|
|
2092
|
-
{ name: "real_apply_executed", passed: isRecord(input.finalE2EProofPlan.liveGateStatus) && input.finalE2EProofPlan.liveGateStatus.sandboxRealApplyExecuted === true },
|
|
2093
|
-
{ name: "global_ready_sentinel_allowed", passed: isRecord(input.finalE2EProofPlan.sentinelPolicy) && input.finalE2EProofPlan.sentinelPolicy.globalReadySentinelAllowed === true },
|
|
2094
|
-
];
|
|
2095
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
2096
|
-
const oracle = {
|
|
2097
|
-
schema: "zob.autonomous-final-no-ship-oracle.v1",
|
|
2098
|
-
runId: input.runId,
|
|
2099
|
-
phase: "11E",
|
|
2100
|
-
oracleType: "deterministic_structural_no_ship",
|
|
2101
|
-
verdict: "FAIL",
|
|
2102
|
-
no_ship: true,
|
|
2103
|
-
finalOracleReady: false,
|
|
2104
|
-
finalOraclePass: false,
|
|
2105
|
-
finalE2EProofReady: false,
|
|
2106
|
-
evidenceChecked: true,
|
|
2107
|
-
checks,
|
|
2108
|
-
failedChecks,
|
|
2109
|
-
blockers,
|
|
2110
|
-
evidenceRefs: [
|
|
2111
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/final-e2e-proof-plan.json`,
|
|
2112
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/current-source-fingerprint.json`,
|
|
2113
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
2114
|
-
],
|
|
2115
|
-
decision: {
|
|
2116
|
-
globalAutonomyReady: false,
|
|
2117
|
-
globalAutonomyNoShip: true,
|
|
2118
|
-
claim100PercentAutonomyAllowed: false,
|
|
2119
|
-
writeGlobalReadySentinelAllowed: false,
|
|
2120
|
-
writeFinalDoneSentinelAllowed: false,
|
|
2121
|
-
},
|
|
2122
|
-
childDispatchAllowed: false,
|
|
2123
|
-
daemonStarted: false,
|
|
2124
|
-
directWorkerWrites: false,
|
|
2125
|
-
transportDispatch: false,
|
|
2126
|
-
networkComsEnabled: false,
|
|
2127
|
-
productionWritesPerformed: false,
|
|
2128
|
-
autoApply: false,
|
|
2129
|
-
realApplyExecuted: false,
|
|
2130
|
-
liveRoutingEnabled: false,
|
|
2131
|
-
budgetEnforced: false,
|
|
2132
|
-
globalAutonomyReady: false,
|
|
2133
|
-
globalAutonomyNoShip: true,
|
|
2134
|
-
bodyStored: false,
|
|
2135
|
-
promptBodiesStored: false,
|
|
2136
|
-
outputBodiesStored: false,
|
|
2137
|
-
};
|
|
2138
|
-
if (hasForbiddenBodyKeys(oracle)) throw new Error("autonomous final no-ship oracle would store forbidden plaintext body keys");
|
|
2139
|
-
return oracle;
|
|
2140
|
-
}
|
|
2141
|
-
|
|
2142
|
-
function buildAutonomousFinalE2EProofPlan(repoRoot: string, input: { runId: string; runtimeGates: Record<string, unknown>; strictBudgetProofPlan: Record<string, unknown>; modelRoutingProofPlan: Record<string, unknown>; modelRoutingPlan: Record<string, unknown>; factorySelection: Record<string, unknown>; factoryRunRef: Record<string, unknown>; oracleReview: Record<string, unknown>; promotionPlan: Record<string, unknown>; promotionProofPlan: Record<string, unknown>; schedulerPlan: Record<string, unknown>; schedulerProofPlan: Record<string, unknown>; missionControlPlan: Record<string, unknown>; missionControlProofPlan: Record<string, unknown>; sandboxApplyPlan: Record<string, unknown>; currentSourceFingerprint: Record<string, unknown>; validation: Record<string, unknown> }): Record<string, unknown> {
|
|
2143
|
-
const safeRunId = safeFileStem(input.runId);
|
|
2144
|
-
const runRoot = `reports/autonomous-runs/${safeRunId}`;
|
|
2145
|
-
const globalReadySentinelPath = join(repoRoot, runRoot, "GLOBAL_AUTONOMY_READY.sentinel");
|
|
2146
|
-
const smokeGate = isRecord(input.promotionPlan.smokeGate) ? input.promotionPlan.smokeGate : {};
|
|
2147
|
-
const pilotGate = isRecord(input.promotionPlan.pilotGate) ? input.promotionPlan.pilotGate : {};
|
|
2148
|
-
const batchGate = isRecord(input.promotionPlan.batchGate) ? input.promotionPlan.batchGate : {};
|
|
2149
|
-
const strictBudgetGate = isRecord(input.runtimeGates.autonomousStrictBudgetGate) ? input.runtimeGates.autonomousStrictBudgetGate : {};
|
|
2150
|
-
const mainWorkspaceApply = isRecord(input.sandboxApplyPlan.mainWorkspaceApply) ? input.sandboxApplyPlan.mainWorkspaceApply : {};
|
|
2151
|
-
const globalReadySentinelPresent = existsSync(globalReadySentinelPath);
|
|
2152
|
-
const proofSteps = [
|
|
2153
|
-
{ name: "user_need", state: "captured", evidenceRef: `${runRoot}/spec-gate.json`, dispatchAllowed: false },
|
|
2154
|
-
{ name: "spec_lock", state: "proved", evidenceRef: `${runRoot}/spec-gate.json`, dispatchAllowed: false },
|
|
2155
|
-
{ name: "context_scope", state: "proved", evidenceRef: `${runRoot}/context-scope.json`, dispatchAllowed: false },
|
|
2156
|
-
{ name: "context_pack", state: "proved", evidenceRef: `${runRoot}/context-pack.json`, dispatchAllowed: false },
|
|
2157
|
-
{ name: "strict_budget", state: input.strictBudgetProofPlan.strictBudgetProofReady === true ? "proved_run_scope_gate_only" : "blocked", evidenceRef: `${runRoot}/strict-budget-proof-plan.json`, blocker: "live_autonomous_strict_budget_not_enforced", dispatchAllowed: false },
|
|
2158
|
-
{ name: "model_routing", state: input.modelRoutingProofPlan.modelRoutingProofReady === true ? "proved_run_scope_gate_only" : "blocked", evidenceRef: `${runRoot}/model-routing-proof-plan.json`, blocker: "live_autonomous_model_routing_not_enabled", dispatchAllowed: false },
|
|
2159
|
-
{ name: "factory_selection", state: "proved", evidenceRef: `${runRoot}/factory-selection.json`, dispatchAllowed: false },
|
|
2160
|
-
{ name: "smoke", state: input.factoryRunRef.status === "done" ? "proved" : "blocked", evidenceRef: `${runRoot}/factory-run-ref.json`, dispatchAllowed: false },
|
|
2161
|
-
{ name: "smoke_oracle", state: input.oracleReview.verdict === "PASS" && input.oracleReview.no_ship === false ? "proved" : "blocked", evidenceRef: `${runRoot}/oracle-review.json`, dispatchAllowed: false },
|
|
2162
|
-
{ name: "promotion_proof", state: input.promotionProofPlan.promotionProofReady === true ? "proved_run_scope_gate_only" : "blocked", evidenceRef: `${runRoot}/promotion-proof-plan.json`, blocker: "pilot_batch_execution_not_enabled", dispatchAllowed: false },
|
|
2163
|
-
{ name: "pilot", state: "blocked", evidenceRef: `${runRoot}/promotion-plan.json`, blocker: "pilot_execution_disabled_until_live_gates", dispatchAllowed: false },
|
|
2164
|
-
{ name: "pilot_oracle", state: "blocked", evidenceRef: `${runRoot}/promotion-plan.json`, blocker: "pilot_oracle_missing_until_pilot_executes", dispatchAllowed: false },
|
|
2165
|
-
{ name: "batch", state: "blocked", evidenceRef: `${runRoot}/promotion-plan.json`, blocker: "batch_execution_disabled_until_pilot_oracle_and_concurrency_gate", dispatchAllowed: false },
|
|
2166
|
-
{ name: "final_oracle", state: "blocked", evidenceRef: `${runRoot}/final-e2e-proof-plan.json`, blocker: "final_oracle_missing_until_full_current_source_e2e", dispatchAllowed: false },
|
|
2167
|
-
{ name: "scheduler_proof", state: input.schedulerProofPlan.schedulerProofReady === true ? "proved_run_scope_gate_only" : "blocked", evidenceRef: `${runRoot}/scheduler-proof-plan.json`, blocker: "daemon_scheduler_not_started", dispatchAllowed: false },
|
|
2168
|
-
{ name: "mission_control_proof", state: input.missionControlProofPlan.missionControlProofReady === true ? "proved_run_scope_gate_only" : "blocked", evidenceRef: `${runRoot}/mission-control-proof-plan.json`, blocker: "live_mission_control_transport_not_enabled", dispatchAllowed: false },
|
|
2169
|
-
{ name: "stop_clean", state: "planned", evidenceRef: `${runRoot}/scheduler-plan.json`, dispatchAllowed: false },
|
|
2170
|
-
];
|
|
2171
|
-
const requiredArtifacts = [
|
|
2172
|
-
{ name: "spec-gate.json", present: true, evidenceRef: `${runRoot}/spec-gate.json` },
|
|
2173
|
-
{ name: "context-scope.json", present: true, evidenceRef: `${runRoot}/context-scope.json` },
|
|
2174
|
-
{ name: "context-pack.json", present: true, evidenceRef: `${runRoot}/context-pack.json` },
|
|
2175
|
-
{ name: "factory-selection.json", present: true, evidenceRef: `${runRoot}/factory-selection.json` },
|
|
2176
|
-
{ name: "runtime-gates.json", present: true, evidenceRef: `${runRoot}/runtime-gates.json` },
|
|
2177
|
-
{ name: "strict-budget-proof-plan.json", present: input.strictBudgetProofPlan.strictBudgetProofReady === true, evidenceRef: `${runRoot}/strict-budget-proof-plan.json` },
|
|
2178
|
-
{ name: "model-routing-proof-plan.json", present: input.modelRoutingProofPlan.modelRoutingProofReady === true, evidenceRef: `${runRoot}/model-routing-proof-plan.json` },
|
|
2179
|
-
{ name: "run-graph.json", present: true, evidenceRef: `${runRoot}/run-graph.json` },
|
|
2180
|
-
{ name: "current-source-fingerprint.json", present: input.currentSourceFingerprint.currentSourceFingerprintReady === true, evidenceRef: `${runRoot}/current-source-fingerprint.json` },
|
|
2181
|
-
{ name: "factory-run-smoke", present: input.factoryRunRef.status === "done", evidenceRef: `${runRoot}/factory-run-ref.json` },
|
|
2182
|
-
{ name: "factory-run-pilot", present: false, missingReason: "pilot_not_executed" },
|
|
2183
|
-
{ name: "factory-run-batch", present: false, missingReason: "batch_not_executed" },
|
|
2184
|
-
{ name: "oracle-review-smoke.json", present: input.oracleReview.verdict === "PASS", evidenceRef: `${runRoot}/oracle-review.json` },
|
|
2185
|
-
{ name: "promotion-proof-plan.json", present: input.promotionProofPlan.promotionProofReady === true, evidenceRef: `${runRoot}/promotion-proof-plan.json` },
|
|
2186
|
-
{ name: "scheduler-proof-plan.json", present: input.schedulerProofPlan.schedulerProofReady === true, evidenceRef: `${runRoot}/scheduler-proof-plan.json` },
|
|
2187
|
-
{ name: "mission-control-proof-plan.json", present: input.missionControlProofPlan.missionControlProofReady === true, evidenceRef: `${runRoot}/mission-control-proof-plan.json` },
|
|
2188
|
-
{ name: "oracle-review-pilot.json", present: false, missingReason: "pilot_oracle_missing" },
|
|
2189
|
-
{ name: "final-oracle.json", present: false, missingReason: "final_oracle_missing" },
|
|
2190
|
-
{ name: "validation.json", present: input.validation.status === "smoke_autonomy_passed", evidenceRef: `${runRoot}/validation.json` },
|
|
2191
|
-
{ name: "DONE.sentinel", present: false, missingReason: "final_e2e_not_run" },
|
|
2192
|
-
{ name: "GLOBAL_AUTONOMY_READY.sentinel", present: globalReadySentinelPresent, missingReason: globalReadySentinelPresent ? undefined : "global_autonomy_not_ready" },
|
|
2193
|
-
];
|
|
2194
|
-
const blockers = [
|
|
2195
|
-
"final_e2e_no_mock_current_source_not_proven",
|
|
2196
|
-
...(strictBudgetGate.strictEnabled === true && strictBudgetGate.budgetEnforced === true ? [] : ["live_strict_budget_not_enforced"]),
|
|
2197
|
-
...(input.modelRoutingPlan.liveRoutingEnabled === true ? [] : ["live_model_routing_not_enabled"]),
|
|
2198
|
-
...(input.schedulerPlan.schedulerExecutionAllowed === true && input.schedulerPlan.daemonStarted === true ? [] : ["daemon_scheduler_not_started"]),
|
|
2199
|
-
...(input.missionControlProofPlan.finalE2ERequirementCleared === true ? [] : ["live_mission_control_transport_not_enabled"]),
|
|
2200
|
-
...(smokeGate.passed === true ? [] : ["smoke_gate_not_passed"]),
|
|
2201
|
-
...(pilotGate.executionAllowed === true ? [] : ["pilot_execution_disabled"]),
|
|
2202
|
-
...(batchGate.executionAllowed === true ? [] : ["batch_execution_disabled"]),
|
|
2203
|
-
...(mainWorkspaceApply.realApplyExecuted === true ? [] : ["real_apply_not_executed"]),
|
|
2204
|
-
"final_oracle_missing",
|
|
2205
|
-
...(globalReadySentinelPresent ? [] : ["global_autonomy_ready_sentinel_absent"]),
|
|
2206
|
-
];
|
|
2207
|
-
const plan = {
|
|
2208
|
-
schema: "zob.autonomous-final-e2e-proof-plan.v1",
|
|
2209
|
-
runId: input.runId,
|
|
2210
|
-
phase: "11A",
|
|
2211
|
-
status: "blocked_for_final_e2e_proof",
|
|
2212
|
-
finalE2EProofReady: false,
|
|
2213
|
-
no_ship: true,
|
|
2214
|
-
blockers,
|
|
2215
|
-
proofSteps,
|
|
2216
|
-
requiredArtifacts,
|
|
2217
|
-
currentSourceProof: {
|
|
2218
|
-
required: true,
|
|
2219
|
-
noMockRequired: true,
|
|
2220
|
-
currentSourceFingerprintCaptured: input.currentSourceFingerprint.currentSourceFingerprintCaptured === true,
|
|
2221
|
-
currentSourceFingerprintHash: typeof input.currentSourceFingerprint.fingerprintHash === "string" ? input.currentSourceFingerprint.fingerprintHash : undefined,
|
|
2222
|
-
sourceFileCount: typeof input.currentSourceFingerprint.sourceFileCount === "number" ? input.currentSourceFingerprint.sourceFileCount : undefined,
|
|
2223
|
-
missingFiles: Array.isArray(input.currentSourceFingerprint.missingFiles) ? input.currentSourceFingerprint.missingFiles : [],
|
|
2224
|
-
noMockCurrentSourceE2EProved: false,
|
|
2225
|
-
registeredFactoryPathChecked: input.factorySelection.currentSourceProofRequired === true,
|
|
2226
|
-
arbitraryFactoryAutonomyReady: false,
|
|
2227
|
-
},
|
|
2228
|
-
liveGateStatus: {
|
|
2229
|
-
strictBudgetEnforced: strictBudgetGate.budgetEnforced === true,
|
|
2230
|
-
strictBudgetEnabled: strictBudgetGate.strictEnabled === true,
|
|
2231
|
-
strictBudgetProofReady: input.strictBudgetProofPlan.strictBudgetProofReady === true,
|
|
2232
|
-
strictBudgetBlockProofPassed: input.strictBudgetProofPlan.strictBudgetBlockProofPassed === true,
|
|
2233
|
-
strictBudgetFinalE2ERequirementCleared: input.strictBudgetProofPlan.finalE2ERequirementCleared === true,
|
|
2234
|
-
modelRoutingProofReady: input.modelRoutingProofPlan.modelRoutingProofReady === true,
|
|
2235
|
-
modelRoutingOracleProofPassed: input.modelRoutingProofPlan.routingOracleProofPassed === true,
|
|
2236
|
-
modelRoutingFinalE2ERequirementCleared: input.modelRoutingProofPlan.finalE2ERequirementCleared === true,
|
|
2237
|
-
liveModelRoutingEnabled: input.modelRoutingPlan.liveRoutingEnabled === true,
|
|
2238
|
-
schedulerExecutionAllowed: input.schedulerPlan.schedulerExecutionAllowed === true,
|
|
2239
|
-
schedulerProofReady: input.schedulerProofPlan.schedulerProofReady === true,
|
|
2240
|
-
schedulerFinalE2ERequirementCleared: input.schedulerProofPlan.finalE2ERequirementCleared === true,
|
|
2241
|
-
daemonStarted: input.schedulerPlan.daemonStarted === true,
|
|
2242
|
-
missionControlProposalOnly: isRecord(input.missionControlPlan.commandPolicy) && input.missionControlPlan.commandPolicy.proposalOnly === true,
|
|
2243
|
-
missionControlProofReady: input.missionControlProofPlan.missionControlProofReady === true,
|
|
2244
|
-
missionControlFinalE2ERequirementCleared: input.missionControlProofPlan.finalE2ERequirementCleared === true,
|
|
2245
|
-
directWorkerCommandsBlocked: input.missionControlProofPlan.directWorkerCommandsBlockedProofPassed === true,
|
|
2246
|
-
liveTransportNetworkDisabled: input.missionControlProofPlan.liveTransportNetworkDisabledProofPassed === true,
|
|
2247
|
-
liveGlobalRoutingApprovalRequired: input.missionControlProofPlan.liveGlobalRoutingApprovalProofPassed === true,
|
|
2248
|
-
sandboxRealApplyExecuted: mainWorkspaceApply.realApplyExecuted === true,
|
|
2249
|
-
},
|
|
2250
|
-
promotionStatus: {
|
|
2251
|
-
smokeGatePassed: smokeGate.passed === true,
|
|
2252
|
-
pilotPreconditionsMet: pilotGate.preconditionsMet === true,
|
|
2253
|
-
promotionProofReady: input.promotionProofPlan.promotionProofReady === true,
|
|
2254
|
-
promotionFinalE2ERequirementCleared: input.promotionProofPlan.finalE2ERequirementCleared === true,
|
|
2255
|
-
pilotExecutionAllowed: pilotGate.executionAllowed === true,
|
|
2256
|
-
batchPreconditionsMet: batchGate.preconditionsMet === true,
|
|
2257
|
-
batchExecutionAllowed: batchGate.executionAllowed === true,
|
|
2258
|
-
pilotExecuted: input.promotionPlan.pilotExecuted === true,
|
|
2259
|
-
batchExecuted: input.promotionPlan.batchExecuted === true,
|
|
2260
|
-
},
|
|
2261
|
-
sentinelPolicy: {
|
|
2262
|
-
doneSentinelAllowed: false,
|
|
2263
|
-
globalReadySentinelAllowed: false,
|
|
2264
|
-
globalReadySentinelPresent,
|
|
2265
|
-
globalReadySentinelPath: `${runRoot}/GLOBAL_AUTONOMY_READY.sentinel`,
|
|
2266
|
-
},
|
|
2267
|
-
finalReportPolicy: {
|
|
2268
|
-
finalGlobalReportAllowed: false,
|
|
2269
|
-
currentReportIsRunScopedSmokeOnly: true,
|
|
2270
|
-
claim100PercentAutonomyAllowed: false,
|
|
2271
|
-
},
|
|
2272
|
-
childDispatchAllowed: false,
|
|
2273
|
-
daemonStarted: false,
|
|
2274
|
-
directWorkerWrites: false,
|
|
2275
|
-
transportDispatch: false,
|
|
2276
|
-
networkComsEnabled: false,
|
|
2277
|
-
productionWritesPerformed: false,
|
|
2278
|
-
autoApply: false,
|
|
2279
|
-
realApplyExecuted: false,
|
|
2280
|
-
globalAutonomyReady: false,
|
|
2281
|
-
globalAutonomyNoShip: true,
|
|
2282
|
-
bodyStored: false,
|
|
2283
|
-
promptBodiesStored: false,
|
|
2284
|
-
outputBodiesStored: false,
|
|
2285
|
-
};
|
|
2286
|
-
if (hasForbiddenBodyKeys(plan)) throw new Error("autonomous final e2e proof plan would store forbidden plaintext body keys");
|
|
2287
|
-
return plan;
|
|
2288
|
-
}
|
|
2289
|
-
|
|
2290
|
-
function buildAutonomousCompletionGate(input: { runId: string; finalE2EProofPlan: Record<string, unknown>; finalNoShipOracle: Record<string, unknown>; currentSourceFingerprintFreshness: Record<string, unknown> }): Record<string, unknown> {
|
|
2291
|
-
const finalBlockers = Array.isArray(input.finalE2EProofPlan.blockers) ? input.finalE2EProofPlan.blockers.filter((blocker): blocker is string => typeof blocker === "string") : [];
|
|
2292
|
-
const oracleFailedChecks = Array.isArray(input.finalNoShipOracle.failedChecks) ? input.finalNoShipOracle.failedChecks.filter((check): check is string => typeof check === "string") : [];
|
|
2293
|
-
const checks = [
|
|
2294
|
-
{ name: "final_e2e_proof_ready", passed: input.finalE2EProofPlan.finalE2EProofReady === true },
|
|
2295
|
-
{ name: "final_no_ship_oracle_pass", passed: input.finalNoShipOracle.finalOraclePass === true && input.finalNoShipOracle.no_ship === false },
|
|
2296
|
-
{ name: "current_source_fingerprint_fresh", passed: input.currentSourceFingerprintFreshness.fresh === true },
|
|
2297
|
-
{ name: "global_ready_sentinel_allowed", passed: isRecord(input.finalE2EProofPlan.sentinelPolicy) && input.finalE2EProofPlan.sentinelPolicy.globalReadySentinelAllowed === true },
|
|
2298
|
-
{ name: "final_done_sentinel_allowed", passed: isRecord(input.finalNoShipOracle.decision) && input.finalNoShipOracle.decision.writeFinalDoneSentinelAllowed === true },
|
|
2299
|
-
{ name: "claim_100_percent_allowed", passed: isRecord(input.finalNoShipOracle.decision) && input.finalNoShipOracle.decision.claim100PercentAutonomyAllowed === true },
|
|
2300
|
-
];
|
|
2301
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
2302
|
-
const gate = {
|
|
2303
|
-
schema: "zob.autonomous-completion-gate.v1",
|
|
2304
|
-
runId: input.runId,
|
|
2305
|
-
phase: "11F",
|
|
2306
|
-
status: "blocked_for_goal_completion",
|
|
2307
|
-
completionReady: false,
|
|
2308
|
-
allRequirementsVerified: false,
|
|
2309
|
-
no_ship: true,
|
|
2310
|
-
updateGoalAllowed: false,
|
|
2311
|
-
requiredGoalStatus: "in_progress",
|
|
2312
|
-
completionToolAvailableInThisRun: false,
|
|
2313
|
-
checks,
|
|
2314
|
-
failedChecks,
|
|
2315
|
-
blockers: [...new Set([...finalBlockers, ...oracleFailedChecks, ...failedChecks])].sort(),
|
|
2316
|
-
evidenceRefs: [
|
|
2317
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/final-e2e-proof-plan.json`,
|
|
2318
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/final-no-ship-oracle.json`,
|
|
2319
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/current-source-fingerprint.json`,
|
|
2320
|
-
`reports/autonomous-runs/${safeFileStem(input.runId)}/validation.json`,
|
|
2321
|
-
],
|
|
2322
|
-
requiredBeforeCompletion: [
|
|
2323
|
-
"final_e2e_no_mock_current_source_proof",
|
|
2324
|
-
"final_oracle_PASS_no_ship_false",
|
|
2325
|
-
"smoke_pilot_batch_artifacts_current_source",
|
|
2326
|
-
"live_strict_budget_enforced",
|
|
2327
|
-
"live_model_routing_enabled",
|
|
2328
|
-
"daemon_scheduler_proven_with_kill_switch",
|
|
2329
|
-
"sandbox_real_apply_or_explicit_no_apply_final_policy",
|
|
2330
|
-
"GLOBAL_AUTONOMY_READY_sentinel_allowed_by_final_oracle",
|
|
2331
|
-
"update_goal_tool_available_and_called_only_after_all_gates_pass",
|
|
2332
|
-
],
|
|
2333
|
-
decision: {
|
|
2334
|
-
updateGoalStatusCompleteAllowed: false,
|
|
2335
|
-
globalAutonomyReady: false,
|
|
2336
|
-
globalAutonomyNoShip: true,
|
|
2337
|
-
writeGlobalReadySentinelAllowed: false,
|
|
2338
|
-
writeFinalDoneSentinelAllowed: false,
|
|
2339
|
-
claim100PercentAutonomyAllowed: false,
|
|
2340
|
-
},
|
|
2341
|
-
childDispatchAllowed: false,
|
|
2342
|
-
daemonStarted: false,
|
|
2343
|
-
directWorkerWrites: false,
|
|
2344
|
-
transportDispatch: false,
|
|
2345
|
-
networkComsEnabled: false,
|
|
2346
|
-
productionWritesPerformed: false,
|
|
2347
|
-
autoApply: false,
|
|
2348
|
-
realApplyExecuted: false,
|
|
2349
|
-
liveRoutingEnabled: false,
|
|
2350
|
-
budgetEnforced: false,
|
|
2351
|
-
globalAutonomyReady: false,
|
|
2352
|
-
globalAutonomyNoShip: true,
|
|
2353
|
-
bodyStored: false,
|
|
2354
|
-
promptBodiesStored: false,
|
|
2355
|
-
outputBodiesStored: false,
|
|
2356
|
-
};
|
|
2357
|
-
if (hasForbiddenBodyKeys(gate)) throw new Error("autonomous completion gate would store forbidden plaintext body keys");
|
|
2358
|
-
return gate;
|
|
2359
|
-
}
|
|
2360
|
-
|
|
2361
|
-
function buildAutonomousReadOnlySmokeFinalReport(validation: Record<string, unknown>, factoryRunRef: Record<string, unknown>, oracleReview: Record<string, unknown>): string {
|
|
2362
|
-
const blockers = Array.isArray(validation.blockers) ? validation.blockers.map((blocker) => `- ${String(blocker)}`).join("\n") : "";
|
|
2363
|
-
return [
|
|
2364
|
-
"# Autonomous Read-Only Smoke Report",
|
|
2365
|
-
"",
|
|
2366
|
-
`Run ID: ${String(validation.runId ?? "unknown")}`,
|
|
2367
|
-
`Status: ${String(validation.status ?? "unknown")}`,
|
|
2368
|
-
`Smoke no-ship: ${String(validation.no_ship === true)}`,
|
|
2369
|
-
`Global autonomy no-ship: ${String(validation.globalAutonomyNoShip === true)}`,
|
|
2370
|
-
"",
|
|
2371
|
-
"## Factory smoke",
|
|
2372
|
-
"",
|
|
2373
|
-
`- Selected factory: ${String(factoryRunRef.selectedFactory ?? "none")}`,
|
|
2374
|
-
`- Factory run id: ${String(factoryRunRef.factoryRunId ?? "none")}`,
|
|
2375
|
-
`- Factory status: ${String(factoryRunRef.status ?? "unknown")}`,
|
|
2376
|
-
`- Deterministic execution: ${String(factoryRunRef.execution === "deterministic")}`,
|
|
2377
|
-
"",
|
|
2378
|
-
"## Structural oracle",
|
|
2379
|
-
"",
|
|
2380
|
-
`- Verdict: ${String(oracleReview.verdict ?? "unknown")}`,
|
|
2381
|
-
`- No-ship for this smoke: ${String(oracleReview.no_ship === true)}`,
|
|
2382
|
-
`- Live oracle dispatched: ${String(oracleReview.liveOracleDispatched === true)}`,
|
|
2383
|
-
"",
|
|
2384
|
-
"## Blockers",
|
|
2385
|
-
"",
|
|
2386
|
-
blockers || "- None for this read-only smoke slice.",
|
|
2387
|
-
"",
|
|
2388
|
-
"## Evidence refs",
|
|
2389
|
-
"",
|
|
2390
|
-
"- spec-gate.json",
|
|
2391
|
-
"- context-scope.json",
|
|
2392
|
-
"- context-pack.json",
|
|
2393
|
-
"- runtime-gates.json",
|
|
2394
|
-
"- model-routing-plan.json",
|
|
2395
|
-
"- model-routing-proof-plan.json",
|
|
2396
|
-
"- factory-selection.json",
|
|
2397
|
-
"- factory-run-ref.json",
|
|
2398
|
-
"- oracle-review.json",
|
|
2399
|
-
"- promotion-plan.json",
|
|
2400
|
-
"- promotion-proof-plan.json",
|
|
2401
|
-
"- scheduler-plan.json",
|
|
2402
|
-
"- scheduler-proof-plan.json",
|
|
2403
|
-
"- mission-control-plan.json",
|
|
2404
|
-
"- mission-control-proof-plan.json",
|
|
2405
|
-
"- sandbox-apply-plan.json",
|
|
2406
|
-
"- strict-budget-proof-plan.json",
|
|
2407
|
-
"- current-source-fingerprint.json",
|
|
2408
|
-
"- final-e2e-proof-plan.json",
|
|
2409
|
-
"- final-no-ship-oracle.json",
|
|
2410
|
-
"- completion-gate.json",
|
|
2411
|
-
"- validation.json",
|
|
2412
|
-
"- SMOKE_AUTONOMY_PASSED.sentinel when status=smoke_autonomy_passed",
|
|
2413
|
-
"",
|
|
2414
|
-
"Compliance: Phase 4A read-only deterministic smoke evidence only; global autonomy remains no-ship.",
|
|
2415
|
-
"",
|
|
2416
|
-
].join("\n");
|
|
2417
|
-
}
|
|
2418
|
-
|
|
2419
|
-
export function validateAutonomousRuntimeDryRunArtifacts(repoRoot: string, runId: string): Record<string, unknown> {
|
|
2420
|
-
const safeRunId = safeFileStem(runId);
|
|
2421
|
-
const runDir = join(repoRoot, "reports", "autonomous-runs", safeRunId);
|
|
2422
|
-
const requiredArtifacts = ["spec-gate.json", "context-scope.json", "context-lookup.json", "context-pack.json", "runtime-gates.json", "model-routing-plan.json", "run-graph.json", "factory-selection.json", "proof-plan.json", "dry-run-report.json", "validation.json", "final-report.md", "DRY_RUN_READY.sentinel"];
|
|
2423
|
-
const jsonArtifacts = requiredArtifacts.filter((name) => name.endsWith(".json"));
|
|
2424
|
-
const parsedArtifacts: Record<string, Record<string, unknown> | unknown[]> = {};
|
|
2425
|
-
const artifactMetadata = requiredArtifacts.map((name) => {
|
|
2426
|
-
const artifactPath = join(runDir, name);
|
|
2427
|
-
const present = existsSync(artifactPath);
|
|
2428
|
-
if (!present) return { name, path: `reports/autonomous-runs/${safeRunId}/${name}`, present, bodyStored: false };
|
|
2429
|
-
if (!name.endsWith(".json")) {
|
|
2430
|
-
const raw = readFileSync(artifactPath, "utf8");
|
|
2431
|
-
return { name, path: `reports/autonomous-runs/${safeRunId}/${name}`, present, hash: sha256(raw), bodyStored: false };
|
|
2432
|
-
}
|
|
2433
|
-
const read = readJsonArtifact(artifactPath);
|
|
2434
|
-
if (read.parsed) parsedArtifacts[name] = read.parsed;
|
|
2435
|
-
return { name, path: `reports/autonomous-runs/${safeRunId}/${name}`, present, hash: read.hash, schema: isRecord(read.parsed) ? read.parsed.schema : undefined, error: read.error, bodyStored: false };
|
|
2436
|
-
});
|
|
2437
|
-
const artifact = (name: string): Record<string, unknown> => isRecord(parsedArtifacts[name]) ? parsedArtifacts[name] as Record<string, unknown> : {};
|
|
2438
|
-
const arrayArtifact = (name: string): unknown[] => Array.isArray(parsedArtifacts[name]) ? parsedArtifacts[name] as unknown[] : [];
|
|
2439
|
-
const report = artifact("dry-run-report.json");
|
|
2440
|
-
const validation = artifact("validation.json");
|
|
2441
|
-
const specGate = artifact("spec-gate.json");
|
|
2442
|
-
const contextScope = artifact("context-scope.json");
|
|
2443
|
-
const contextLookups = arrayArtifact("context-lookup.json");
|
|
2444
|
-
const contextPack = artifact("context-pack.json");
|
|
2445
|
-
const runtimeGates = artifact("runtime-gates.json");
|
|
2446
|
-
const modelRoutingPlan = artifact("model-routing-plan.json");
|
|
2447
|
-
const runGraph = artifact("run-graph.json");
|
|
2448
|
-
const factorySelection = artifact("factory-selection.json");
|
|
2449
|
-
const proofPlan = artifact("proof-plan.json");
|
|
2450
|
-
const proofStages = Array.isArray(proofPlan.stages) ? proofPlan.stages.filter(isRecord) : [];
|
|
2451
|
-
const graphEdges = Array.isArray(runGraph.edges) ? runGraph.edges.filter(isRecord) : [];
|
|
2452
|
-
const graphNodes = Array.isArray(runGraph.nodes) ? runGraph.nodes.filter(isRecord) : [];
|
|
2453
|
-
const modelRoutes = Array.isArray(modelRoutingPlan.routes) ? modelRoutingPlan.routes.filter(isRecord) : [];
|
|
2454
|
-
const lookupPlan = isRecord(report.contextPlan) && isRecord(report.contextPlan.lookupPlan) ? report.contextPlan.lookupPlan : {};
|
|
2455
|
-
const factoryReadiness = isRecord(factorySelection.factoryReadiness) ? factorySelection.factoryReadiness : {};
|
|
2456
|
-
const jsonBodyFree = jsonArtifacts.every((name) => {
|
|
2457
|
-
const parsed = parsedArtifacts[name];
|
|
2458
|
-
return parsed !== undefined && !hasForbiddenBodyKeys(parsed);
|
|
2459
|
-
});
|
|
2460
|
-
const checks = [
|
|
2461
|
-
{ name: "required_artifacts_present", passed: artifactMetadata.every((item) => item.present === true), detail: { requiredArtifacts } },
|
|
2462
|
-
{ name: "json_artifacts_parse", passed: artifactMetadata.filter((item) => item.name.endsWith(".json")).every((item) => !item.error), detail: { jsonArtifacts } },
|
|
2463
|
-
{ name: "artifact_schemas", passed: specGate.schema === "zob.autonomous-spec-gate.v1" && contextScope.schema === "zob.context-scope.v1" && contextPack.schema === "zob.context-pack.v1" && runtimeGates.schema === "zob.autonomous-runtime-gates.v1" && modelRoutingPlan.schema === "zob.autonomous-model-routing-plan.v1" && runGraph.schema === "zob.autonomous-run-graph.v1" && factorySelection.schema === "zob.autonomous-factory-selection.v1" && proofPlan.schema === "zob.autonomous-proof-plan.v1" && report.schema === "zob.autonomous-runtime-dry-run.v1" && validation.schema === "zob.autonomous-runtime-dry-run-validation.v1", detail: "expected autonomous dry-run artifact schemas" },
|
|
2464
|
-
{ name: "spec_gate_scope_locked", passed: specGate.specLocked === true && specGate.allowedPathsRequired === true && specGate.pathGatePassed === true && specGate.applyPolicyRequired === true && specGate.applyPolicyProvided === true && specGate.budgetProfileRequired === true && specGate.budgetProfileProvided === true && specGate.autonomousStrictBudgetRequired === true && specGate.autonomousStrictBudgetSatisfied === true && Array.isArray(specGate.allowedPaths) && specGate.allowedPaths.length > 0 && Array.isArray(specGate.forbiddenPaths) && specGate.forbiddenPaths.length > 0, detail: "autonomous spec lock requires safe bounded allowed_paths, forbidden_paths, explicit apply_policy, and explicit strict_requested budget_profile" },
|
|
2465
|
-
{ name: "dry_run_safety_flags", passed: report.noExecution === true && report.childDispatchAllowed === false && report.daemonStarted === false && report.productionWritesPerformed === false && report.autoApply === false && report.networkAccessed === false && report.globalAutonomyReady === false && report.globalAutonomyNoShip === true, detail: "report safety posture" },
|
|
2466
|
-
{ name: "validation_safety_flags", passed: validation.passed === true && validation.noExecution === true && validation.childDispatchAllowed === false && validation.networkAccessed === false && validation.globalAutonomyNoShip === true && validation.sentinel === "DRY_RUN_READY.sentinel", detail: "validation safety posture" },
|
|
2467
|
-
{ name: "context_lookup_pack_cited_bounded", passed: contextLookups.length > 0 && contextLookups.every((lookup) => isRecord(lookup) && lookup.schema === "zob.brain-lookup-result.v1" && lookup.queryStored === false && lookup.citationRequired === true) && Array.isArray(contextPack.citations) && contextPack.citations.length > 0 && isRecord(contextPack.loadingRules) && contextPack.loadingRules.boundedContextOnly === true && contextPack.loadingRules.agentLoadsEntireCorpus === false, detail: { lookupCount: contextLookups.length, citationCount: Array.isArray(contextPack.citations) ? contextPack.citations.length : 0 } },
|
|
2468
|
-
{ name: "gbrain_disabled", passed: lookupPlan.gbrainImportEnabled === false && lookupPlan.gbrainEmbedEnabled === false && lookupPlan.gbrainSyncEnabled === false && lookupPlan.gbrainWriteEnabled === false, detail: "P0 dry-run does not import/embed/sync/write GBrain" },
|
|
2469
|
-
{ name: "runtime_gates_no_execution", passed: runtimeGates.passed === true && isRecord(runtimeGates.autonomousStrictBudgetGate) && runtimeGates.autonomousStrictBudgetGate.strictRequested === true && runtimeGates.autonomousStrictBudgetGate.strictEnabled === false && runtimeGates.autonomousStrictBudgetGate.budgetEnforced === false && runtimeGates.noExecution === true && runtimeGates.childDispatchAllowed === false && runtimeGates.globalBudgetEnforced === false && runtimeGates.globalModelRoutingEnabled === false && runtimeGates.daemonStarted === false && runtimeGates.productionWritesPerformed === false && runtimeGates.autoApply === false && runtimeGates.networkAccessed === false, detail: "runtime gates require strict budget intent while remaining disabled/proposal-only" },
|
|
2470
|
-
{ name: "model_routing_plan_safe", passed: modelRoutingPlan.routingPlanReady === true && modelRoutingPlan.liveRoutingEnabled === false && modelRoutingPlan.globalLiveRoutingEnabled === false && modelRoutingPlan.modelRouterUsed === false && modelRoutingPlan.routingApplied === false && modelRoutingPlan.childDispatchAllowed === false && modelRoutingPlan.noExecution === true && modelRoutes.some((route) => route.stage === "context_reuse_scout" && route.recommendedModelClass === "cheap_scout") && modelRoutes.some((route) => route.oracleCritical === true && route.recommendedModelClass === "strong_oracle") && modelRoutes.every((route) => route.noExecution === true && route.modelRouterUsed === false && route.routingApplied === false && route.childDispatchAllowed === false), detail: { routeCount: modelRoutes.length, failedChecks: modelRoutingPlan.failedChecks } },
|
|
2471
|
-
{ name: "run_graph_parent_owned_no_dispatch", passed: runGraph.status === "dry_run_graph_ready" && runGraph.parentOwned === true && runGraph.noExecution === true && runGraph.childDispatchAllowed === false && graphNodes.some((node) => node.id === "model_routing_plan") && graphNodes.some((node) => node.id === "registered_factory_current_source_proof") && graphEdges.length > 0 && graphEdges.every((edge) => edge.parentOwned === true && edge.dispatchAllowed === false), detail: { nodes: graphNodes.length, edges: graphEdges.length } },
|
|
2472
|
-
{ name: "factory_selection_requires_current_source_proof", passed: factorySelection.currentSourceProofRequired === true && typeof factoryReadiness.registeredBatchReady === "boolean" && factorySelection.proofBeforeExecutionRequired === (factoryReadiness.registeredBatchReady !== true), detail: { selectedFactory: factorySelection.selectedFactory, registeredBatchReady: factoryReadiness.registeredBatchReady } },
|
|
2473
|
-
{ name: "proof_plan_has_required_gates", passed: ["context_lookup_and_pack", "runtime_gates_preflight", "model_routing_plan", "registered_factory_current_source_proof", "smoke_oracle", "pilot_oracle", "batch"].every((name) => proofStages.some((stage) => stage.name === name)) && proofStages.every((stage) => stage.name === "spec_lock" || stage.dispatchAllowed === false || stage.dispatchAllowed === undefined), detail: { stages: proofStages.map((stage) => stage.name) } },
|
|
2474
|
-
{ name: "json_artifacts_body_free", passed: jsonBodyFree, detail: "no forbidden plaintext body keys in JSON artifacts" },
|
|
2475
|
-
];
|
|
2476
|
-
const failedChecks = checks.filter((check) => !check.passed).map((check) => check.name);
|
|
2477
|
-
const result = {
|
|
2478
|
-
schema: "zob.autonomous-runtime-dry-run-artifact-validation.v1",
|
|
2479
|
-
runId: safeRunId,
|
|
2480
|
-
valid: failedChecks.length === 0,
|
|
2481
|
-
failedChecks,
|
|
2482
|
-
checks,
|
|
2483
|
-
artifacts: artifactMetadata,
|
|
2484
|
-
noExecution: true,
|
|
2485
|
-
childDispatchAllowed: false,
|
|
2486
|
-
networkAccessed: false,
|
|
2487
|
-
globalAutonomyReady: false,
|
|
2488
|
-
globalAutonomyNoShip: true,
|
|
2489
|
-
bodyStored: false,
|
|
2490
|
-
promptBodiesStored: false,
|
|
2491
|
-
outputBodiesStored: false,
|
|
2492
|
-
generatedAt: new Date().toISOString(),
|
|
2493
|
-
};
|
|
2494
|
-
if (hasForbiddenBodyKeys(result)) throw new Error("autonomous dry-run artifact validation would store forbidden plaintext body keys");
|
|
2495
|
-
return result;
|
|
2496
|
-
}
|
|
2497
|
-
|
|
2498
|
-
export function writeAutonomousRuntimeDryRunReport(repoRoot: string, input: AutonomousRuntimeDryRunInput): Record<string, unknown> {
|
|
2499
|
-
const report = buildAutonomousRuntimeDryRun(repoRoot, input);
|
|
2500
|
-
const runId = String(report.runId);
|
|
2501
|
-
const safeRunId = safeFileStem(runId);
|
|
2502
|
-
const runDir = join(repoRoot, "reports", "autonomous-runs", safeRunId);
|
|
2503
|
-
mkdirSync(runDir, { recursive: true });
|
|
2504
|
-
const specGatePath = join(runDir, "spec-gate.json");
|
|
2505
|
-
const contextScopePath = join(runDir, "context-scope.json");
|
|
2506
|
-
const contextLookupPath = join(runDir, "context-lookup.json");
|
|
2507
|
-
const contextPackPath = join(runDir, "context-pack.json");
|
|
2508
|
-
const runtimeGatesPath = join(runDir, "runtime-gates.json");
|
|
2509
|
-
const modelRoutingPlanPath = join(runDir, "model-routing-plan.json");
|
|
2510
|
-
const runGraphPath = join(runDir, "run-graph.json");
|
|
2511
|
-
const factorySelectionPath = join(runDir, "factory-selection.json");
|
|
2512
|
-
const proofPlanPath = join(runDir, "proof-plan.json");
|
|
2513
|
-
const reportPath = join(runDir, "dry-run-report.json");
|
|
2514
|
-
const validationPath = join(runDir, "validation.json");
|
|
2515
|
-
const finalReportPath = join(runDir, "final-report.md");
|
|
2516
|
-
writeFileSync(specGatePath, `${JSON.stringify(report.specGate, null, 2)}\n`);
|
|
2517
|
-
const contextPlan = isRecord(report.contextPlan) ? report.contextPlan : {};
|
|
2518
|
-
const contextScope = isRecord(contextPlan.contextScope) ? contextPlan.contextScope : {};
|
|
2519
|
-
writeFileSync(contextScopePath, `${JSON.stringify(contextScope, null, 2)}\n`);
|
|
2520
|
-
writeFileSync(contextLookupPath, `${JSON.stringify(contextPlan.lookupResults ?? [], null, 2)}\n`);
|
|
2521
|
-
writeFileSync(contextPackPath, `${JSON.stringify(contextPlan.contextPack ?? {}, null, 2)}\n`);
|
|
2522
|
-
writeFileSync(runtimeGatesPath, `${JSON.stringify(report.runtimeGates, null, 2)}\n`);
|
|
2523
|
-
writeFileSync(modelRoutingPlanPath, `${JSON.stringify(report.modelRoutingPlan, null, 2)}\n`);
|
|
2524
|
-
writeFileSync(runGraphPath, `${JSON.stringify(report.runGraph, null, 2)}\n`);
|
|
2525
|
-
writeFileSync(factorySelectionPath, `${JSON.stringify(report.factorySelection, null, 2)}\n`);
|
|
2526
|
-
writeFileSync(proofPlanPath, `${JSON.stringify(report.proofPlan, null, 2)}\n`);
|
|
2527
|
-
writeFileSync(reportPath, `${JSON.stringify(report, null, 2)}\n`);
|
|
2528
|
-
const validation = buildAutonomousRuntimeDryRunValidation(report);
|
|
2529
|
-
writeFileSync(validationPath, `${JSON.stringify(validation, null, 2)}\n`);
|
|
2530
|
-
writeFileSync(finalReportPath, buildAutonomousRuntimeDryRunFinalReport(report));
|
|
2531
|
-
if (report.status === "dry_run_plan_ready") writeFileSync(join(runDir, "DRY_RUN_READY.sentinel"), "dry-run-ready\n");
|
|
2532
|
-
return {
|
|
2533
|
-
...report,
|
|
2534
|
-
specGatePath: `reports/autonomous-runs/${safeRunId}/spec-gate.json`,
|
|
2535
|
-
contextScopePath: `reports/autonomous-runs/${safeRunId}/context-scope.json`,
|
|
2536
|
-
contextLookupPath: `reports/autonomous-runs/${safeRunId}/context-lookup.json`,
|
|
2537
|
-
contextPackPath: `reports/autonomous-runs/${safeRunId}/context-pack.json`,
|
|
2538
|
-
runtimeGatesPath: `reports/autonomous-runs/${safeRunId}/runtime-gates.json`,
|
|
2539
|
-
modelRoutingPlanPath: `reports/autonomous-runs/${safeRunId}/model-routing-plan.json`,
|
|
2540
|
-
runGraphPath: `reports/autonomous-runs/${safeRunId}/run-graph.json`,
|
|
2541
|
-
factorySelectionPath: `reports/autonomous-runs/${safeRunId}/factory-selection.json`,
|
|
2542
|
-
proofPlanPath: `reports/autonomous-runs/${safeRunId}/proof-plan.json`,
|
|
2543
|
-
reportPath: `reports/autonomous-runs/${safeRunId}/dry-run-report.json`,
|
|
2544
|
-
validationPath: `reports/autonomous-runs/${safeRunId}/validation.json`,
|
|
2545
|
-
finalReportPath: `reports/autonomous-runs/${safeRunId}/final-report.md`,
|
|
2546
|
-
sentinelPath: report.status === "dry_run_plan_ready" ? `reports/autonomous-runs/${safeRunId}/DRY_RUN_READY.sentinel` : undefined,
|
|
2547
|
-
};
|
|
2548
|
-
}
|
|
2549
|
-
|
|
2550
|
-
export function writeAutonomousReadOnlySmokeRunReport(repoRoot: string, input: AutonomousReadOnlySmokeRunInput): Record<string, unknown> {
|
|
2551
|
-
const runId = safeFileStem(input.runId ?? `autonomous-readonly-smoke-${sha256(input.userNeed || "missing-spec").slice(0, 12)}`);
|
|
2552
|
-
const safeRunId = safeFileStem(runId);
|
|
2553
|
-
const runDir = join(repoRoot, "reports", "autonomous-runs", safeRunId);
|
|
2554
|
-
const smokeAutonomySentinelPath = join(runDir, "SMOKE_AUTONOMY_PASSED.sentinel");
|
|
2555
|
-
if (existsSync(smokeAutonomySentinelPath)) throw new Error(`Autonomous smoke run already passed; choose a fresh run_id to avoid stale sentinel reuse: ${runId}`);
|
|
2556
|
-
const dryRun = writeAutonomousRuntimeDryRunReport(repoRoot, {
|
|
2557
|
-
...input,
|
|
2558
|
-
runId,
|
|
2559
|
-
applyPolicy: input.applyPolicy ?? "no_apply",
|
|
2560
|
-
budgetProfile: input.budgetProfile ?? "strict_requested",
|
|
2561
|
-
});
|
|
2562
|
-
const factorySelection = isRecord(dryRun.factorySelection) ? dryRun.factorySelection : {};
|
|
2563
|
-
const selectedFactory = typeof factorySelection.selectedFactory === "string" ? factorySelection.selectedFactory : undefined;
|
|
2564
|
-
const manifestPath = selectedFactory ? `.pi/factories/${selectedFactory}/smoke-manifest.json` : undefined;
|
|
2565
|
-
const factoryRunId = safeFileStem(input.factoryRunId ?? `autonomous-smoke-${safeRunId}`);
|
|
2566
|
-
const factoryRunAlreadyExists = existsSync(join(repoRoot, "reports", "factory-runs", factoryRunId));
|
|
2567
|
-
const blockers = [
|
|
2568
|
-
...(dryRun.status === "dry_run_plan_ready" ? [] : ["dry_run_not_ready"]),
|
|
2569
|
-
...(isRecord(dryRun.validation) && Array.isArray(dryRun.validation.blockers) ? dryRun.validation.blockers.filter((blocker): blocker is string => typeof blocker === "string") : []),
|
|
2570
|
-
...(isRecord(dryRun.specGate) && dryRun.specGate.applyPolicy === "no_apply" ? [] : ["apply_policy_must_be_no_apply_for_readonly_smoke"]),
|
|
2571
|
-
...(isRecord(dryRun.specGate) && dryRun.specGate.budgetProfile === "strict_requested" ? [] : ["strict_requested_budget_required_for_readonly_smoke"]),
|
|
2572
|
-
...(isRecord(dryRun.runtimeGates) && dryRun.runtimeGates.childDispatchAllowed === false && dryRun.runtimeGates.daemonStarted === false && dryRun.runtimeGates.productionWritesPerformed === false && dryRun.runtimeGates.autoApply === false ? [] : ["runtime_gates_not_readonly_safe"]),
|
|
2573
|
-
...(isRecord(dryRun.modelRoutingPlan) && dryRun.modelRoutingPlan.routingPlanReady === true && dryRun.modelRoutingPlan.liveRoutingEnabled === false && dryRun.modelRoutingPlan.childDispatchAllowed === false ? [] : ["model_routing_plan_not_readonly_safe"]),
|
|
2574
|
-
...(selectedFactory ? [] : ["factory_selection_missing"]),
|
|
2575
|
-
...(factorySelection.selectionStatus === "existing_factory_selected" ? [] : ["selected_factory_must_be_existing_for_readonly_smoke"]),
|
|
2576
|
-
...(manifestPath && existsSync(join(repoRoot, manifestPath)) ? [] : ["selected_factory_smoke_manifest_missing"]),
|
|
2577
|
-
...(factoryRunAlreadyExists ? ["factory_run_id_already_exists"] : []),
|
|
2578
|
-
];
|
|
2579
|
-
const shouldRunFactory = blockers.length === 0 && selectedFactory !== undefined && manifestPath !== undefined;
|
|
2580
|
-
const factoryRunResult = shouldRunFactory ? runFactoryRun(repoRoot, {
|
|
2581
|
-
factory: selectedFactory,
|
|
2582
|
-
input_manifest: manifestPath,
|
|
2583
|
-
run_id: factoryRunId,
|
|
2584
|
-
mode: "smoke",
|
|
2585
|
-
max_items: 1,
|
|
2586
|
-
execution: "deterministic",
|
|
2587
|
-
budget: { strictRequested: true, strictEnabled: false, maxRuns: 1, estimatedRuns: 1, maxParallelChildren: 1, estimatedParallelChildren: 1 },
|
|
2588
|
-
model_routing: { enabled: false, risk: input.risk ?? "medium", contextTokens: input.maxContextTokens },
|
|
2589
|
-
}) : undefined;
|
|
2590
|
-
const factoryRunDir = join(repoRoot, "reports", "factory-runs", factoryRunId);
|
|
2591
|
-
const factoryValidationPath = join(factoryRunDir, "validation.json");
|
|
2592
|
-
const factoryValidationRead = readJsonArtifact(factoryValidationPath);
|
|
2593
|
-
const factoryValidation = isRecord(factoryValidationRead.parsed) ? factoryValidationRead.parsed : {};
|
|
2594
|
-
const phaseSentinelPresent = existsSync(join(factoryRunDir, "SMOKE_PASSED.sentinel"));
|
|
2595
|
-
const doneSentinelPresent = existsSync(join(factoryRunDir, "DONE.sentinel"));
|
|
2596
|
-
const factoryRunRef = {
|
|
2597
|
-
schema: "zob.autonomous-readonly-smoke-factory-run-ref.v1",
|
|
2598
|
-
runId,
|
|
2599
|
-
selectedFactory,
|
|
2600
|
-
manifestPath,
|
|
2601
|
-
factoryRunId,
|
|
2602
|
-
factoryRunPath: relativeFactoryRunPath(factoryRunId),
|
|
2603
|
-
status: factoryRunResult?.status ?? "not_started",
|
|
2604
|
-
processed: factoryRunResult?.processed ?? 0,
|
|
2605
|
-
failed: factoryRunResult?.failed ?? 0,
|
|
2606
|
-
execution: "deterministic",
|
|
2607
|
-
mode: "smoke",
|
|
2608
|
-
reportsOnlyWrites: true,
|
|
2609
|
-
productionWritesPerformed: false,
|
|
2610
|
-
autoApply: false,
|
|
2611
|
-
childDispatchAllowed: false,
|
|
2612
|
-
liveChildDispatches: 0,
|
|
2613
|
-
daemonStarted: false,
|
|
2614
|
-
phaseSentinel: "SMOKE_PASSED.sentinel",
|
|
2615
|
-
phaseSentinelPresent,
|
|
2616
|
-
doneSentinelPresent,
|
|
2617
|
-
validationPath: relativeFactoryRunPath(factoryRunId, "validation.json"),
|
|
2618
|
-
validationHash: factoryValidationRead.hash,
|
|
2619
|
-
artifactHashes: {
|
|
2620
|
-
validation: factoryValidationRead.hash,
|
|
2621
|
-
smokeSentinel: artifactHashIfPresent(join(factoryRunDir, "SMOKE_PASSED.sentinel")),
|
|
2622
|
-
doneSentinel: artifactHashIfPresent(join(factoryRunDir, "DONE.sentinel")),
|
|
2623
|
-
telemetry: artifactHashIfPresent(join(factoryRunDir, "telemetry.json")),
|
|
2624
|
-
agenticPlan: artifactHashIfPresent(join(factoryRunDir, "agentic-plan.json")),
|
|
2625
|
-
},
|
|
2626
|
-
artifacts: factoryRunResult?.artifacts ?? [],
|
|
2627
|
-
errors: factoryRunResult?.errors ?? [],
|
|
2628
|
-
bodyStored: false,
|
|
2629
|
-
promptBodiesStored: false,
|
|
2630
|
-
outputBodiesStored: false,
|
|
2631
|
-
};
|
|
2632
|
-
const oracleChecks = [
|
|
2633
|
-
{ name: "dry_run_ready", passed: dryRun.status === "dry_run_plan_ready" },
|
|
2634
|
-
{ name: "context_pack_cited", passed: isRecord(dryRun.contextPlan) && dryRun.contextPlan.contextPackValid === true },
|
|
2635
|
-
{ name: "runtime_gates_readonly", passed: isRecord(dryRun.runtimeGates) && dryRun.runtimeGates.childDispatchAllowed === false && dryRun.runtimeGates.daemonStarted === false && dryRun.runtimeGates.productionWritesPerformed === false && dryRun.runtimeGates.autoApply === false },
|
|
2636
|
-
{ name: "model_routing_plan_readonly", passed: isRecord(dryRun.modelRoutingPlan) && dryRun.modelRoutingPlan.routingPlanReady === true && dryRun.modelRoutingPlan.liveRoutingEnabled === false && dryRun.modelRoutingPlan.childDispatchAllowed === false },
|
|
2637
|
-
{ name: "factory_selection_existing", passed: factorySelection.selectionStatus === "existing_factory_selected" && Boolean(selectedFactory) },
|
|
2638
|
-
{ name: "factory_smoke_done", passed: factoryRunResult?.status === "done" && factoryRunResult.processed === 1 && factoryRunResult.failed === 0 },
|
|
2639
|
-
{ name: "factory_validation_passed", passed: factoryValidation.status === "passed" },
|
|
2640
|
-
{ name: "smoke_sentinel_present", passed: phaseSentinelPresent },
|
|
2641
|
-
{ name: "done_sentinel_present", passed: doneSentinelPresent },
|
|
2642
|
-
];
|
|
2643
|
-
const structuralOraclePassed = blockers.length === 0 && oracleChecks.every((check) => check.passed === true);
|
|
2644
|
-
const oracleReview = {
|
|
2645
|
-
schema: "zob.autonomous-readonly-smoke-oracle-review.v1",
|
|
2646
|
-
runId,
|
|
2647
|
-
oracleType: "deterministic_structural",
|
|
2648
|
-
verdict: structuralOraclePassed ? "PASS" : "FAIL",
|
|
2649
|
-
no_ship: !structuralOraclePassed,
|
|
2650
|
-
liveOracleDispatched: false,
|
|
2651
|
-
evidenceChecked: true,
|
|
2652
|
-
checks: oracleChecks,
|
|
2653
|
-
failedChecks: oracleChecks.filter((check) => check.passed !== true).map((check) => check.name),
|
|
2654
|
-
evidenceRefs: [
|
|
2655
|
-
`reports/autonomous-runs/${safeRunId}/spec-gate.json`,
|
|
2656
|
-
`reports/autonomous-runs/${safeRunId}/context-pack.json`,
|
|
2657
|
-
`reports/autonomous-runs/${safeRunId}/runtime-gates.json`,
|
|
2658
|
-
`reports/autonomous-runs/${safeRunId}/model-routing-plan.json`,
|
|
2659
|
-
`reports/autonomous-runs/${safeRunId}/factory-selection.json`,
|
|
2660
|
-
`reports/autonomous-runs/${safeRunId}/factory-run-ref.json`,
|
|
2661
|
-
relativeFactoryRunPath(factoryRunId, "validation.json"),
|
|
2662
|
-
relativeFactoryRunPath(factoryRunId, "SMOKE_PASSED.sentinel"),
|
|
2663
|
-
relativeFactoryRunPath(factoryRunId, "DONE.sentinel"),
|
|
2664
|
-
],
|
|
2665
|
-
globalAutonomyReady: false,
|
|
2666
|
-
globalAutonomyNoShip: true,
|
|
2667
|
-
bodyStored: false,
|
|
2668
|
-
promptBodiesStored: false,
|
|
2669
|
-
outputBodiesStored: false,
|
|
2670
|
-
};
|
|
2671
|
-
const validation = {
|
|
2672
|
-
schema: "zob.autonomous-readonly-smoke-validation.v1",
|
|
2673
|
-
runId,
|
|
2674
|
-
status: structuralOraclePassed ? "smoke_autonomy_passed" : "blocked",
|
|
2675
|
-
passed: structuralOraclePassed,
|
|
2676
|
-
no_ship: !structuralOraclePassed,
|
|
2677
|
-
smokeRunNoShip: !structuralOraclePassed,
|
|
2678
|
-
globalAutonomyReady: false,
|
|
2679
|
-
globalAutonomyNoShip: true,
|
|
2680
|
-
blockers: [...blockers, ...oracleChecks.filter((check) => check.passed !== true).map((check) => check.name)],
|
|
2681
|
-
warnings: ["phase_4a_deterministic_structural_oracle_only", "global_autonomy_no_ship", "reports_only_writes"],
|
|
2682
|
-
dryRunReady: dryRun.status === "dry_run_plan_ready",
|
|
2683
|
-
selectedFactory,
|
|
2684
|
-
factoryRunId,
|
|
2685
|
-
factoryRunStatus: factoryRunResult?.status ?? "not_started",
|
|
2686
|
-
oracleVerdict: oracleReview.verdict,
|
|
2687
|
-
oracleNoShip: oracleReview.no_ship,
|
|
2688
|
-
reportsOnlyWrites: true,
|
|
2689
|
-
deterministicExecution: true,
|
|
2690
|
-
childDispatchAllowed: false,
|
|
2691
|
-
liveChildDispatches: 0,
|
|
2692
|
-
daemonStarted: false,
|
|
2693
|
-
productionWritesPerformed: false,
|
|
2694
|
-
autoApply: false,
|
|
2695
|
-
liveRoutingEnabled: false,
|
|
2696
|
-
globalLiveRoutingEnabled: false,
|
|
2697
|
-
sentinel: structuralOraclePassed ? "SMOKE_AUTONOMY_PASSED.sentinel" : undefined,
|
|
2698
|
-
bodyStored: false,
|
|
2699
|
-
promptBodiesStored: false,
|
|
2700
|
-
outputBodiesStored: false,
|
|
2701
|
-
generatedAt: new Date().toISOString(),
|
|
2702
|
-
};
|
|
2703
|
-
const promotionPlan = buildAutonomousPromotionPlan({ runId, selectedFactory, factoryRunRef, oracleReview, validation });
|
|
2704
|
-
const promotionProofPlan = buildAutonomousPromotionProofPlan({ runId, promotionPlan, factoryRunRef, oracleReview, validation });
|
|
2705
|
-
const schedulerPlan = buildAutonomousSchedulerPlan(repoRoot, { runId, promotionPlan, validation });
|
|
2706
|
-
const schedulerProofPlan = buildAutonomousSchedulerProofPlan({ runId, schedulerPlan, validation });
|
|
2707
|
-
const missionControlPlan = buildAutonomousMissionControlPlan(repoRoot, { runId, schedulerPlan, validation });
|
|
2708
|
-
const strictBudgetProofPlan = buildAutonomousStrictBudgetProofPlan({ runId, runtimeGates: isRecord(dryRun.runtimeGates) ? dryRun.runtimeGates : {}, validation });
|
|
2709
|
-
const modelRoutingProofPlan = buildAutonomousModelRoutingProofPlan({ runId, modelRoutingPlan: isRecord(dryRun.modelRoutingPlan) ? dryRun.modelRoutingPlan : {}, validation });
|
|
2710
|
-
const missionControlProofPlan = buildAutonomousMissionControlProofPlan({ runId, missionControlPlan, schedulerProofPlan, modelRoutingProofPlan, validation });
|
|
2711
|
-
const sandboxApplyPlan = buildAutonomousSandboxApplyPlan(repoRoot, { runId, missionControlPlan, validation });
|
|
2712
|
-
const currentSourceFingerprint = buildAutonomousCurrentSourceFingerprint(repoRoot, { runId, factorySelection });
|
|
2713
|
-
const finalE2EProofPlan = buildAutonomousFinalE2EProofPlan(repoRoot, {
|
|
2714
|
-
runId,
|
|
2715
|
-
runtimeGates: isRecord(dryRun.runtimeGates) ? dryRun.runtimeGates : {},
|
|
2716
|
-
strictBudgetProofPlan,
|
|
2717
|
-
modelRoutingProofPlan,
|
|
2718
|
-
modelRoutingPlan: isRecord(dryRun.modelRoutingPlan) ? dryRun.modelRoutingPlan : {},
|
|
2719
|
-
factorySelection,
|
|
2720
|
-
factoryRunRef,
|
|
2721
|
-
oracleReview,
|
|
2722
|
-
promotionPlan,
|
|
2723
|
-
promotionProofPlan,
|
|
2724
|
-
schedulerPlan,
|
|
2725
|
-
schedulerProofPlan,
|
|
2726
|
-
missionControlPlan,
|
|
2727
|
-
missionControlProofPlan,
|
|
2728
|
-
sandboxApplyPlan,
|
|
2729
|
-
currentSourceFingerprint,
|
|
2730
|
-
validation,
|
|
2731
|
-
});
|
|
2732
|
-
const finalNoShipOracle = buildAutonomousFinalNoShipOracle({ runId, finalE2EProofPlan, currentSourceFingerprint });
|
|
2733
|
-
const currentSourceFingerprintFreshness = validateAutonomousCurrentSourceFingerprintFreshness(repoRoot, currentSourceFingerprint, factorySelection);
|
|
2734
|
-
const completionGate = buildAutonomousCompletionGate({ runId, finalE2EProofPlan, finalNoShipOracle, currentSourceFingerprintFreshness });
|
|
2735
|
-
if (hasForbiddenBodyKeys(factoryRunRef) || hasForbiddenBodyKeys(oracleReview) || hasForbiddenBodyKeys(promotionPlan) || hasForbiddenBodyKeys(promotionProofPlan) || hasForbiddenBodyKeys(schedulerPlan) || hasForbiddenBodyKeys(schedulerProofPlan) || hasForbiddenBodyKeys(missionControlPlan) || hasForbiddenBodyKeys(missionControlProofPlan) || hasForbiddenBodyKeys(sandboxApplyPlan) || hasForbiddenBodyKeys(strictBudgetProofPlan) || hasForbiddenBodyKeys(modelRoutingProofPlan) || hasForbiddenBodyKeys(currentSourceFingerprint) || hasForbiddenBodyKeys(finalE2EProofPlan) || hasForbiddenBodyKeys(finalNoShipOracle) || hasForbiddenBodyKeys(completionGate) || hasForbiddenBodyKeys(validation)) throw new Error("autonomous readonly smoke artifacts would store forbidden plaintext body keys");
|
|
2736
|
-
const factoryRunRefPath = join(runDir, "factory-run-ref.json");
|
|
2737
|
-
const oracleReviewPath = join(runDir, "oracle-review.json");
|
|
2738
|
-
const promotionPlanPath = join(runDir, "promotion-plan.json");
|
|
2739
|
-
const promotionProofPlanPath = join(runDir, "promotion-proof-plan.json");
|
|
2740
|
-
const schedulerPlanPath = join(runDir, "scheduler-plan.json");
|
|
2741
|
-
const schedulerProofPlanPath = join(runDir, "scheduler-proof-plan.json");
|
|
2742
|
-
const missionControlPlanPath = join(runDir, "mission-control-plan.json");
|
|
2743
|
-
const missionControlProofPlanPath = join(runDir, "mission-control-proof-plan.json");
|
|
2744
|
-
const sandboxApplyPlanPath = join(runDir, "sandbox-apply-plan.json");
|
|
2745
|
-
const strictBudgetProofPlanPath = join(runDir, "strict-budget-proof-plan.json");
|
|
2746
|
-
const modelRoutingProofPlanPath = join(runDir, "model-routing-proof-plan.json");
|
|
2747
|
-
const currentSourceFingerprintPath = join(runDir, "current-source-fingerprint.json");
|
|
2748
|
-
const finalE2EProofPlanPath = join(runDir, "final-e2e-proof-plan.json");
|
|
2749
|
-
const finalNoShipOraclePath = join(runDir, "final-no-ship-oracle.json");
|
|
2750
|
-
const completionGatePath = join(runDir, "completion-gate.json");
|
|
2751
|
-
const validationPath = join(runDir, "validation.json");
|
|
2752
|
-
const finalReportPath = join(runDir, "final-report.md");
|
|
2753
|
-
writeFileSync(factoryRunRefPath, `${JSON.stringify(factoryRunRef, null, 2)}\n`);
|
|
2754
|
-
writeFileSync(oracleReviewPath, `${JSON.stringify(oracleReview, null, 2)}\n`);
|
|
2755
|
-
writeFileSync(promotionPlanPath, `${JSON.stringify(promotionPlan, null, 2)}\n`);
|
|
2756
|
-
writeFileSync(promotionProofPlanPath, `${JSON.stringify(promotionProofPlan, null, 2)}\n`);
|
|
2757
|
-
writeFileSync(schedulerPlanPath, `${JSON.stringify(schedulerPlan, null, 2)}\n`);
|
|
2758
|
-
writeFileSync(schedulerProofPlanPath, `${JSON.stringify(schedulerProofPlan, null, 2)}\n`);
|
|
2759
|
-
writeFileSync(missionControlPlanPath, `${JSON.stringify(missionControlPlan, null, 2)}\n`);
|
|
2760
|
-
writeFileSync(missionControlProofPlanPath, `${JSON.stringify(missionControlProofPlan, null, 2)}\n`);
|
|
2761
|
-
writeFileSync(sandboxApplyPlanPath, `${JSON.stringify(sandboxApplyPlan, null, 2)}\n`);
|
|
2762
|
-
writeFileSync(strictBudgetProofPlanPath, `${JSON.stringify(strictBudgetProofPlan, null, 2)}\n`);
|
|
2763
|
-
writeFileSync(modelRoutingProofPlanPath, `${JSON.stringify(modelRoutingProofPlan, null, 2)}\n`);
|
|
2764
|
-
writeFileSync(currentSourceFingerprintPath, `${JSON.stringify(currentSourceFingerprint, null, 2)}\n`);
|
|
2765
|
-
writeFileSync(finalE2EProofPlanPath, `${JSON.stringify(finalE2EProofPlan, null, 2)}\n`);
|
|
2766
|
-
writeFileSync(finalNoShipOraclePath, `${JSON.stringify(finalNoShipOracle, null, 2)}\n`);
|
|
2767
|
-
writeFileSync(completionGatePath, `${JSON.stringify(completionGate, null, 2)}\n`);
|
|
2768
|
-
writeFileSync(validationPath, `${JSON.stringify(validation, null, 2)}\n`);
|
|
2769
|
-
writeFileSync(finalReportPath, buildAutonomousReadOnlySmokeFinalReport(validation, factoryRunRef, oracleReview));
|
|
2770
|
-
if (structuralOraclePassed) writeFileSync(smokeAutonomySentinelPath, "smoke-autonomy-passed\n");
|
|
2771
|
-
return {
|
|
2772
|
-
...dryRun,
|
|
2773
|
-
status: validation.status,
|
|
2774
|
-
no_ship: validation.no_ship,
|
|
2775
|
-
smokeRunNoShip: validation.smokeRunNoShip,
|
|
2776
|
-
globalAutonomyReady: false,
|
|
2777
|
-
globalAutonomyNoShip: true,
|
|
2778
|
-
factoryRunRef,
|
|
2779
|
-
oracleReview,
|
|
2780
|
-
promotionPlan,
|
|
2781
|
-
promotionProofPlan,
|
|
2782
|
-
schedulerPlan,
|
|
2783
|
-
schedulerProofPlan,
|
|
2784
|
-
missionControlPlan,
|
|
2785
|
-
missionControlProofPlan,
|
|
2786
|
-
sandboxApplyPlan,
|
|
2787
|
-
strictBudgetProofPlan,
|
|
2788
|
-
modelRoutingProofPlan,
|
|
2789
|
-
currentSourceFingerprint,
|
|
2790
|
-
finalE2EProofPlan,
|
|
2791
|
-
finalNoShipOracle,
|
|
2792
|
-
completionGate,
|
|
2793
|
-
validation,
|
|
2794
|
-
factoryRunRefPath: `reports/autonomous-runs/${safeRunId}/factory-run-ref.json`,
|
|
2795
|
-
oracleReviewPath: `reports/autonomous-runs/${safeRunId}/oracle-review.json`,
|
|
2796
|
-
promotionPlanPath: `reports/autonomous-runs/${safeRunId}/promotion-plan.json`,
|
|
2797
|
-
promotionProofPlanPath: `reports/autonomous-runs/${safeRunId}/promotion-proof-plan.json`,
|
|
2798
|
-
schedulerPlanPath: `reports/autonomous-runs/${safeRunId}/scheduler-plan.json`,
|
|
2799
|
-
schedulerProofPlanPath: `reports/autonomous-runs/${safeRunId}/scheduler-proof-plan.json`,
|
|
2800
|
-
missionControlPlanPath: `reports/autonomous-runs/${safeRunId}/mission-control-plan.json`,
|
|
2801
|
-
missionControlProofPlanPath: `reports/autonomous-runs/${safeRunId}/mission-control-proof-plan.json`,
|
|
2802
|
-
sandboxApplyPlanPath: `reports/autonomous-runs/${safeRunId}/sandbox-apply-plan.json`,
|
|
2803
|
-
strictBudgetProofPlanPath: `reports/autonomous-runs/${safeRunId}/strict-budget-proof-plan.json`,
|
|
2804
|
-
modelRoutingProofPlanPath: `reports/autonomous-runs/${safeRunId}/model-routing-proof-plan.json`,
|
|
2805
|
-
currentSourceFingerprintPath: `reports/autonomous-runs/${safeRunId}/current-source-fingerprint.json`,
|
|
2806
|
-
finalE2EProofPlanPath: `reports/autonomous-runs/${safeRunId}/final-e2e-proof-plan.json`,
|
|
2807
|
-
finalNoShipOraclePath: `reports/autonomous-runs/${safeRunId}/final-no-ship-oracle.json`,
|
|
2808
|
-
completionGatePath: `reports/autonomous-runs/${safeRunId}/completion-gate.json`,
|
|
2809
|
-
validationPath: `reports/autonomous-runs/${safeRunId}/validation.json`,
|
|
2810
|
-
finalReportPath: `reports/autonomous-runs/${safeRunId}/final-report.md`,
|
|
2811
|
-
smokeSentinelPath: structuralOraclePassed ? `reports/autonomous-runs/${safeRunId}/SMOKE_AUTONOMY_PASSED.sentinel` : undefined,
|
|
2812
|
-
};
|
|
2813
|
-
}
|
|
2814
|
-
|
|
2815
|
-
export function validateAutonomousReadOnlySmokeRunArtifacts(repoRoot: string, runId: string): Record<string, unknown> {
|
|
2816
|
-
const safeRunId = safeFileStem(runId);
|
|
2817
|
-
const runDir = join(repoRoot, "reports", "autonomous-runs", safeRunId);
|
|
2818
|
-
const requiredArtifacts = ["spec-gate.json", "context-scope.json", "context-pack.json", "runtime-gates.json", "model-routing-plan.json", "model-routing-proof-plan.json", "factory-selection.json", "factory-run-ref.json", "oracle-review.json", "promotion-plan.json", "promotion-proof-plan.json", "scheduler-plan.json", "scheduler-proof-plan.json", "mission-control-plan.json", "mission-control-proof-plan.json", "sandbox-apply-plan.json", "strict-budget-proof-plan.json", "current-source-fingerprint.json", "final-e2e-proof-plan.json", "final-no-ship-oracle.json", "completion-gate.json", "validation.json", "final-report.md", "SMOKE_AUTONOMY_PASSED.sentinel"];
|
|
2819
|
-
const jsonArtifacts = requiredArtifacts.filter((name) => name.endsWith(".json"));
|
|
2820
|
-
const parsedArtifacts: Record<string, Record<string, unknown>> = {};
|
|
2821
|
-
const artifacts = requiredArtifacts.map((name) => {
|
|
2822
|
-
const artifactPath = join(runDir, name);
|
|
2823
|
-
const present = existsSync(artifactPath);
|
|
2824
|
-
if (!present) return { name, path: `reports/autonomous-runs/${safeRunId}/${name}`, present, bodyStored: false };
|
|
2825
|
-
if (!name.endsWith(".json")) return { name, path: `reports/autonomous-runs/${safeRunId}/${name}`, present, hash: artifactHashIfPresent(artifactPath), bodyStored: false };
|
|
2826
|
-
const read = readJsonArtifact(artifactPath);
|
|
2827
|
-
if (isRecord(read.parsed)) parsedArtifacts[name] = read.parsed;
|
|
2828
|
-
return { name, path: `reports/autonomous-runs/${safeRunId}/${name}`, present, hash: read.hash, schema: isRecord(read.parsed) ? read.parsed.schema : undefined, error: read.error, bodyStored: false };
|
|
2829
|
-
});
|
|
2830
|
-
const artifact = (name: string): Record<string, unknown> => parsedArtifacts[name] ?? {};
|
|
2831
|
-
const runtimeGates = artifact("runtime-gates.json");
|
|
2832
|
-
const modelRoutingPlan = artifact("model-routing-plan.json");
|
|
2833
|
-
const modelRoutingProofPlan = artifact("model-routing-proof-plan.json");
|
|
2834
|
-
const factorySelection = artifact("factory-selection.json");
|
|
2835
|
-
const factoryRunRef = artifact("factory-run-ref.json");
|
|
2836
|
-
const oracleReview = artifact("oracle-review.json");
|
|
2837
|
-
const promotionPlan = artifact("promotion-plan.json");
|
|
2838
|
-
const promotionProofPlan = artifact("promotion-proof-plan.json");
|
|
2839
|
-
const schedulerPlan = artifact("scheduler-plan.json");
|
|
2840
|
-
const schedulerProofPlan = artifact("scheduler-proof-plan.json");
|
|
2841
|
-
const missionControlPlan = artifact("mission-control-plan.json");
|
|
2842
|
-
const missionControlProofPlan = artifact("mission-control-proof-plan.json");
|
|
2843
|
-
const sandboxApplyPlan = artifact("sandbox-apply-plan.json");
|
|
2844
|
-
const strictBudgetProofPlan = artifact("strict-budget-proof-plan.json");
|
|
2845
|
-
const currentSourceFingerprint = artifact("current-source-fingerprint.json");
|
|
2846
|
-
const finalE2EProofPlan = artifact("final-e2e-proof-plan.json");
|
|
2847
|
-
const finalNoShipOracle = artifact("final-no-ship-oracle.json");
|
|
2848
|
-
const completionGate = artifact("completion-gate.json");
|
|
2849
|
-
const validation = artifact("validation.json");
|
|
2850
|
-
const currentSourceFingerprintFreshness = validateAutonomousCurrentSourceFingerprintFreshness(repoRoot, currentSourceFingerprint, factorySelection);
|
|
2851
|
-
const factoryRunId = typeof factoryRunRef.factoryRunId === "string" ? factoryRunRef.factoryRunId : "unknown";
|
|
2852
|
-
const factoryRunDir = join(repoRoot, "reports", "factory-runs", factoryRunId);
|
|
2853
|
-
const finalDoneSentinelPath = join(runDir, "DONE.sentinel");
|
|
2854
|
-
const globalReadySentinelPath = join(runDir, "GLOBAL_AUTONOMY_READY.sentinel");
|
|
2855
|
-
const finalDoneSentinelPresent = existsSync(finalDoneSentinelPath);
|
|
2856
|
-
const globalReadySentinelPresent = existsSync(globalReadySentinelPath);
|
|
2857
|
-
const checks = [
|
|
2858
|
-
{ name: "required_artifacts_present", passed: artifacts.every((item) => item.present === true), detail: { requiredArtifacts } },
|
|
2859
|
-
{ name: "json_artifacts_parse", passed: artifacts.filter((item) => item.name.endsWith(".json")).every((item) => !item.error), detail: { jsonArtifacts } },
|
|
2860
|
-
{ name: "artifact_schemas", passed: factoryRunRef.schema === "zob.autonomous-readonly-smoke-factory-run-ref.v1" && oracleReview.schema === "zob.autonomous-readonly-smoke-oracle-review.v1" && promotionPlan.schema === "zob.autonomous-promotion-plan.v1" && promotionProofPlan.schema === "zob.autonomous-promotion-proof-plan.v1" && schedulerPlan.schema === "zob.autonomous-scheduler-plan.v1" && schedulerProofPlan.schema === "zob.autonomous-scheduler-proof-plan.v1" && missionControlPlan.schema === "zob.autonomous-mission-control-plan.v1" && missionControlProofPlan.schema === "zob.autonomous-mission-control-proof-plan.v1" && sandboxApplyPlan.schema === "zob.autonomous-sandbox-apply-plan.v1" && strictBudgetProofPlan.schema === "zob.autonomous-strict-budget-proof-plan.v1" && modelRoutingProofPlan.schema === "zob.autonomous-model-routing-proof-plan.v1" && currentSourceFingerprint.schema === "zob.autonomous-current-source-fingerprint.v1" && finalE2EProofPlan.schema === "zob.autonomous-final-e2e-proof-plan.v1" && finalNoShipOracle.schema === "zob.autonomous-final-no-ship-oracle.v1" && completionGate.schema === "zob.autonomous-completion-gate.v1" && validation.schema === "zob.autonomous-readonly-smoke-validation.v1", detail: "expected Phase 4A/5B/6B/7B/8B/9A/9B/10B/11F smoke artifact schemas" },
|
|
2861
|
-
{ name: "runtime_and_model_gates_safe", passed: runtimeGates.childDispatchAllowed === false && runtimeGates.daemonStarted === false && runtimeGates.productionWritesPerformed === false && runtimeGates.autoApply === false && modelRoutingPlan.liveRoutingEnabled === false && modelRoutingPlan.globalLiveRoutingEnabled === false && modelRoutingPlan.childDispatchAllowed === false, detail: "runtime/model routing gates remain disabled" },
|
|
2862
|
-
{ name: "model_routing_proof_plan_gate_only", passed: modelRoutingProofPlan.status === "model_routing_dispatch_gate_proof_ready_global_default_blocked" && modelRoutingProofPlan.modelRoutingProofReady === true && modelRoutingProofPlan.routingDefaultDisabledProofPassed === true && modelRoutingProofPlan.routingOracleProofPassed === true && modelRoutingProofPlan.routingHighContextProofPassed === true && modelRoutingProofPlan.routingSecurityNoDowngradeProofPassed === true && modelRoutingProofPlan.selectedModelsStored === false && modelRoutingProofPlan.selectedModelHashesOnly === true && modelRoutingProofPlan.finalE2ERequirementCleared === false && modelRoutingProofPlan.no_ship === true && Array.isArray(modelRoutingProofPlan.scenarios) && modelRoutingProofPlan.scenarios.some((scenario) => isRecord(scenario) && scenario.name === "routing_oracle_applies_when_enabled" && scenario.selectedModelClass === "strong_oracle" && typeof scenario.selectedModelHash === "string" && scenario.selectedModelStored === false) && modelRoutingProofPlan.globalLiveRoutingEnabled === false && modelRoutingProofPlan.liveAutonomousRoutingApplied === false && modelRoutingProofPlan.modelRouterUsed === false && modelRoutingProofPlan.routingApplied === false && modelRoutingProofPlan.childDispatchAllowed === false && modelRoutingProofPlan.noExecution === true && modelRoutingProofPlan.globalAutonomyReady === false && modelRoutingProofPlan.globalAutonomyNoShip === true, detail: "Phase 6B proves model routing dispatch gate class selection without enabling global/live routing" },
|
|
2863
|
-
{ name: "factory_selection_existing", passed: factorySelection.selectionStatus === "existing_factory_selected" && typeof factorySelection.selectedFactory === "string", detail: { selectedFactory: factorySelection.selectedFactory } },
|
|
2864
|
-
{ name: "factory_smoke_completed", passed: factoryRunRef.status === "done" && factoryRunRef.execution === "deterministic" && factoryRunRef.phaseSentinelPresent === true && factoryRunRef.doneSentinelPresent === true && existsSync(join(factoryRunDir, "SMOKE_PASSED.sentinel")) && existsSync(join(factoryRunDir, "DONE.sentinel")), detail: { factoryRunId } },
|
|
2865
|
-
{ name: "structural_oracle_passed", passed: oracleReview.verdict === "PASS" && oracleReview.no_ship === false && oracleReview.liveOracleDispatched === false, detail: { oracleType: oracleReview.oracleType } },
|
|
2866
|
-
{ name: "promotion_plan_smoke_to_pilot_gate", passed: isRecord(promotionPlan.smokeGate) && promotionPlan.smokeGate.passed === true && isRecord(promotionPlan.pilotGate) && promotionPlan.pilotGate.preconditionsMet === true && promotionPlan.pilotGate.executionAllowed === false && promotionPlan.pilotGate.dispatchAllowed === false && isRecord(promotionPlan.batchGate) && promotionPlan.batchGate.preconditionsMet === false && promotionPlan.batchGate.batchConcurrencyCapRequired === true && promotionPlan.globalAutonomyNoShip === true, detail: "Phase 7A promotion gates are metadata-only and do not execute pilot/batch" },
|
|
2867
|
-
{ name: "promotion_proof_plan_execution_blocked", passed: promotionProofPlan.status === "pilot_batch_promotion_proof_ready_execution_blocked" && promotionProofPlan.promotionProofReady === true && promotionProofPlan.smokeToPilotPreconditionsProved === true && promotionProofPlan.pilotExecutionBlockedProofPassed === true && promotionProofPlan.batchExecutionBlockedProofPassed === true && promotionProofPlan.resumeRetryPolicyProofPassed === true && promotionProofPlan.oraclePolicyProofPassed === true && promotionProofPlan.finalE2ERequirementCleared === false && promotionProofPlan.no_ship === true && promotionProofPlan.pilotExecutionAllowed === false && promotionProofPlan.batchExecutionAllowed === false && promotionProofPlan.pilotExecuted === false && promotionProofPlan.batchExecuted === false && promotionProofPlan.childDispatchAllowed === false && promotionProofPlan.noExecutionBeyondSmoke === true && promotionProofPlan.globalAutonomyReady === false && promotionProofPlan.globalAutonomyNoShip === true, detail: "Phase 7B proves pilot/batch promotion prerequisites and blocked execution without running pilot/batch" },
|
|
2868
|
-
{ name: "scheduler_plan_disabled_bounded", passed: schedulerPlan.schedulerPlanReady === true && schedulerPlan.schedulerExecutionAllowed === false && schedulerPlan.daemonStarted === false && schedulerPlan.autoStartDaemon === false && schedulerPlan.continuousLoop === false && schedulerPlan.childDispatchAllowed === false && isRecord(schedulerPlan.killSwitch) && schedulerPlan.killSwitch.required === true && isRecord(schedulerPlan.workerPool) && schedulerPlan.workerPool.maxWorkers === 1 && isRecord(schedulerPlan.retryPolicy) && schedulerPlan.retryPolicy.retriesCapped === true && isRecord(schedulerPlan.budgetPolicy) && schedulerPlan.budgetPolicy.strictBudgetRequired === true && schedulerPlan.globalAutonomyNoShip === true, detail: "Phase 8A scheduler plan is bounded and disabled by default" },
|
|
2869
|
-
{ name: "scheduler_proof_plan_execution_blocked", passed: schedulerProofPlan.status === "scheduler_daemon_proof_ready_execution_blocked" && schedulerProofPlan.schedulerProofReady === true && schedulerProofPlan.daemonDefaultDisabledProofPassed === true && schedulerProofPlan.oneWorkerBoundedProofPassed === true && schedulerProofPlan.stopConditionsProofPassed === true && schedulerProofPlan.killSwitchRetryProofPassed === true && schedulerProofPlan.strictBudgetBeforeDispatchProofPassed === true && schedulerProofPlan.alwaysOnApprovalProofPassed === true && schedulerProofPlan.finalE2ERequirementCleared === false && schedulerProofPlan.no_ship === true && schedulerProofPlan.schedulerExecutionAllowed === false && schedulerProofPlan.daemonStarted === false && schedulerProofPlan.autoStartDaemon === false && schedulerProofPlan.continuousLoop === false && schedulerProofPlan.childDispatchAllowed === false && schedulerProofPlan.productionWritesPerformed === false && schedulerProofPlan.autoApply === false && schedulerProofPlan.noExecutionBeyondSmoke === true && schedulerProofPlan.globalAutonomyReady === false && schedulerProofPlan.globalAutonomyNoShip === true, detail: "Phase 8B proves daemon/scheduler policy and blocked execution without starting daemon" },
|
|
2870
|
-
{ name: "mission_control_plan_proposal_only", passed: missionControlPlan.missionControlPlanReady === true && missionControlPlan.transportDispatch === false && missionControlPlan.networkComsEnabled === false && missionControlPlan.directWorkerWrites === false && missionControlPlan.childDispatchAllowed === false && isRecord(missionControlPlan.commandPolicy) && missionControlPlan.commandPolicy.proposalOnly === true && missionControlPlan.commandPolicy.directWorkerWrites === false && missionControlPlan.commandPolicy.directWorkerCommandBlocked === true && isRecord(missionControlPlan.comsPolicy) && missionControlPlan.comsPolicy.topologyGuardActive === true && missionControlPlan.comsPolicy.hashOnlyLedgers === true && missionControlPlan.globalAutonomyNoShip === true, detail: "Phase 9A Mission Control/coms plan is proposal-only and transport-disabled" },
|
|
2871
|
-
{ name: "mission_control_proof_plan_execution_blocked", passed: missionControlProofPlan.status === "mission_control_comms_proof_ready_execution_blocked" && missionControlProofPlan.missionControlProofReady === true && missionControlProofPlan.proposalOnlyCommandsProofPassed === true && missionControlProofPlan.topologyHashOnlyCommsProofPassed === true && missionControlProofPlan.directWorkerCommandsBlockedProofPassed === true && missionControlProofPlan.liveTransportNetworkDisabledProofPassed === true && missionControlProofPlan.liveGlobalRoutingApprovalProofPassed === true && missionControlProofPlan.post8bSchedulerBlockedProofPassed === true && missionControlProofPlan.finalE2ERequirementCleared === false && missionControlProofPlan.no_ship === true && missionControlProofPlan.childDispatchAllowed === false && missionControlProofPlan.daemonStarted === false && missionControlProofPlan.directWorkerWrites === false && missionControlProofPlan.transportDispatch === false && missionControlProofPlan.networkComsEnabled === false && missionControlProofPlan.liveRoutingEnabled === false && missionControlProofPlan.globalLiveRoutingEnabled === false && missionControlProofPlan.productionWritesPerformed === false && missionControlProofPlan.autoApply === false && missionControlProofPlan.globalAutonomyReady === false && missionControlProofPlan.globalAutonomyNoShip === true, detail: "Phase 9B proves proposal-only commands, topology/hash-only comms, blocked direct-worker commands, disabled transport/network, and approval-required live/global routing" },
|
|
2872
|
-
{ name: "sandbox_apply_plan_metadata_only", passed: sandboxApplyPlan.sandboxApplyPlanReady === true && sandboxApplyPlan.productionWritesPerformed === false && sandboxApplyPlan.autoApply === false && sandboxApplyPlan.realApplyExecuted === false && sandboxApplyPlan.childDispatchAllowed === false && sandboxApplyPlan.globalAutonomyNoShip === true && isRecord(sandboxApplyPlan.isolatedTempWorkspace) && sandboxApplyPlan.isolatedTempWorkspace.required === true && sandboxApplyPlan.isolatedTempWorkspace.executed === true && isRecord(sandboxApplyPlan.diffGate) && sandboxApplyPlan.diffGate.diffHashRequired === true && sandboxApplyPlan.diffGate.diffHashesMatch === true && isRecord(sandboxApplyPlan.oracleDiffReview) && sandboxApplyPlan.oracleDiffReview.required === true && sandboxApplyPlan.oracleDiffReview.reviewPassed === true && isRecord(sandboxApplyPlan.rollbackPolicy) && sandboxApplyPlan.rollbackPolicy.rollbackMetadataRequired === true && sandboxApplyPlan.rollbackPolicy.rollbackPrepared === true && isRecord(sandboxApplyPlan.approvalPolicy) && sandboxApplyPlan.approvalPolicy.manualApprovalRequired === true && sandboxApplyPlan.approvalPolicy.approvedForMainWorkspaceApply === false && isRecord(sandboxApplyPlan.manualApplyPreflight) && sandboxApplyPlan.manualApplyPreflight.required === true && sandboxApplyPlan.manualApplyPreflight.preflightPassed === true && sandboxApplyPlan.manualApplyPreflight.executionAllowedByThisTool === false && sandboxApplyPlan.manualApplyPreflight.realApplyExecuted === false && isRecord(sandboxApplyPlan.mainWorkspaceApply) && sandboxApplyPlan.mainWorkspaceApply.realApplyAllowed === false && sandboxApplyPlan.mainWorkspaceApply.realApplyExecuted === false && sandboxApplyPlan.mainWorkspaceApply.productionWritesPerformed === false, detail: "Phase 10B sandbox/apply plan requires temp workspace, diff hash, oracle review, rollback, manual approval, and manual apply preflight while blocking real apply" },
|
|
2873
|
-
{ name: "strict_budget_proof_plan_gate_only", passed: strictBudgetProofPlan.status === "strict_budget_dispatch_gate_proof_ready_global_default_blocked" && strictBudgetProofPlan.strictBudgetProofReady === true && strictBudgetProofPlan.strictBudgetAllowProofPassed === true && strictBudgetProofPlan.strictBudgetBlockProofPassed === true && strictBudgetProofPlan.strictBudgetDefaultDisabledProofPassed === true && strictBudgetProofPlan.finalE2ERequirementCleared === false && strictBudgetProofPlan.no_ship === true && Array.isArray(strictBudgetProofPlan.scenarios) && strictBudgetProofPlan.scenarios.some((scenario) => isRecord(scenario) && scenario.name === "strict_gate_blocks_exceedance_pre_dispatch" && scenario.wouldBlockDispatch === true && scenario.gateChildDispatchAllowed === false && scenario.dispatchDecision === "block") && strictBudgetProofPlan.globalStrictBudgetEnabled === false && strictBudgetProofPlan.globalBudgetEnforced === false && strictBudgetProofPlan.liveAutonomousBudgetEnforced === false && strictBudgetProofPlan.budgetEnforced === false && strictBudgetProofPlan.strictEnabled === false && strictBudgetProofPlan.childDispatchAllowed === false && strictBudgetProofPlan.noExecution === true && strictBudgetProofPlan.globalAutonomyReady === false && strictBudgetProofPlan.globalAutonomyNoShip === true, detail: "Phase 5B proves strict dispatch gate allow/block/default-disabled behavior without clearing final live strict-budget requirement" },
|
|
2874
|
-
{ name: "current_source_fingerprint_hash_only", passed: currentSourceFingerprint.status === "current_source_fingerprint_captured" && currentSourceFingerprint.currentSourceFingerprintReady === true && currentSourceFingerprint.currentSourceFingerprintCaptured === true && currentSourceFingerprint.noMockCurrentSourceE2EProved === false && typeof currentSourceFingerprint.fingerprintHash === "string" && /^[a-f0-9]{64}$/.test(currentSourceFingerprint.fingerprintHash) && Array.isArray(currentSourceFingerprint.sourceFiles) && currentSourceFingerprint.sourceFiles.length > 0 && isRecord(currentSourceFingerprint.fileHashes) && Object.values(currentSourceFingerprint.fileHashes).every((hash) => typeof hash === "string" && /^[a-f0-9]{64}$/.test(hash)) && Array.isArray(currentSourceFingerprint.missingFiles) && currentSourceFingerprint.missingFiles.length === 0 && isRecord(currentSourceFingerprint.evidencePolicy) && currentSourceFingerprint.evidencePolicy.hashOnly === true && currentSourceFingerprint.evidencePolicy.sourceBodiesStored === false && isRecord(currentSourceFingerprint.safety) && currentSourceFingerprint.safety.noExecution === true && currentSourceFingerprint.safety.globalAutonomyNoShip === true, detail: "Phase 11C current-source fingerprint captures hash-only source evidence without proving final no-mock E2E" },
|
|
2875
|
-
{ name: "current_source_fingerprint_fresh", passed: currentSourceFingerprintFreshness.fresh === true && currentSourceFingerprintFreshness.fileHashesMatch === true && currentSourceFingerprintFreshness.fingerprintHashMatches === true && currentSourceFingerprintFreshness.sourceFilesMatch === true && currentSourceFingerprintFreshness.missingFilesMatch === true && currentSourceFingerprintFreshness.currentMissingFiles && Array.isArray(currentSourceFingerprintFreshness.currentMissingFiles) && currentSourceFingerprintFreshness.currentMissingFiles.length === 0, detail: "Phase 11D validator recomputes current source hashes and fails stale/tampered fingerprints" },
|
|
2876
|
-
{ name: "final_e2e_proof_plan_no_ship", passed: finalE2EProofPlan.status === "blocked_for_final_e2e_proof" && finalE2EProofPlan.finalE2EProofReady === false && finalE2EProofPlan.no_ship === true && Array.isArray(finalE2EProofPlan.blockers) && finalE2EProofPlan.blockers.includes("live_strict_budget_not_enforced") && finalE2EProofPlan.blockers.includes("live_model_routing_not_enabled") && finalE2EProofPlan.blockers.includes("pilot_execution_disabled") && finalE2EProofPlan.blockers.includes("batch_execution_disabled") && finalE2EProofPlan.blockers.includes("final_oracle_missing") && isRecord(finalE2EProofPlan.sentinelPolicy) && finalE2EProofPlan.sentinelPolicy.globalReadySentinelAllowed === false && finalE2EProofPlan.sentinelPolicy.globalReadySentinelPresent === false && isRecord(finalE2EProofPlan.currentSourceProof) && finalE2EProofPlan.currentSourceProof.currentSourceFingerprintCaptured === true && typeof finalE2EProofPlan.currentSourceProof.currentSourceFingerprintHash === "string" && finalE2EProofPlan.currentSourceProof.noMockCurrentSourceE2EProved === false && finalE2EProofPlan.globalAutonomyReady === false && finalE2EProofPlan.globalAutonomyNoShip === true && finalE2EProofPlan.childDispatchAllowed === false && finalE2EProofPlan.productionWritesPerformed === false && finalE2EProofPlan.autoApply === false, detail: "Phase 11C final E2E proof plan references current-source fingerprint while preserving no-ship" },
|
|
2877
|
-
{ name: "final_no_ship_oracle_blocks_global_autonomy", passed: finalNoShipOracle.verdict === "FAIL" && finalNoShipOracle.no_ship === true && finalNoShipOracle.finalOracleReady === false && finalNoShipOracle.finalOraclePass === false && Array.isArray(finalNoShipOracle.failedChecks) && finalNoShipOracle.failedChecks.includes("final_e2e_proof_ready") && finalNoShipOracle.failedChecks.includes("no_mock_current_source_e2e_proved") && isRecord(finalNoShipOracle.decision) && finalNoShipOracle.decision.globalAutonomyReady === false && finalNoShipOracle.decision.globalAutonomyNoShip === true && finalNoShipOracle.decision.writeGlobalReadySentinelAllowed === false && finalNoShipOracle.globalAutonomyReady === false && finalNoShipOracle.globalAutonomyNoShip === true && finalNoShipOracle.childDispatchAllowed === false && finalNoShipOracle.productionWritesPerformed === false && finalNoShipOracle.autoApply === false, detail: "Phase 11E deterministic final no-ship oracle blocks global autonomy until final E2E proof" },
|
|
2878
|
-
{ name: "completion_gate_blocks_goal_completion", passed: completionGate.status === "blocked_for_goal_completion" && completionGate.completionReady === false && completionGate.allRequirementsVerified === false && completionGate.no_ship === true && completionGate.updateGoalAllowed === false && completionGate.requiredGoalStatus === "in_progress" && Array.isArray(completionGate.failedChecks) && completionGate.failedChecks.includes("final_e2e_proof_ready") && isRecord(completionGate.decision) && completionGate.decision.updateGoalStatusCompleteAllowed === false && completionGate.decision.globalAutonomyReady === false && completionGate.decision.globalAutonomyNoShip === true && completionGate.decision.writeGlobalReadySentinelAllowed === false && completionGate.globalAutonomyReady === false && completionGate.globalAutonomyNoShip === true && completionGate.childDispatchAllowed === false && completionGate.productionWritesPerformed === false && completionGate.autoApply === false, detail: "Phase 11F completion gate blocks update_goal complete until every final requirement is verified" },
|
|
2879
|
-
{ name: "final_global_sentinels_absent", passed: finalDoneSentinelPresent === false && globalReadySentinelPresent === false, detail: { doneSentinelPresent: finalDoneSentinelPresent, globalReadySentinelPresent } },
|
|
2880
|
-
{ name: "validation_passed_scope_limited", passed: validation.passed === true && validation.status === "smoke_autonomy_passed" && validation.no_ship === false && validation.globalAutonomyReady === false && validation.globalAutonomyNoShip === true && validation.childDispatchAllowed === false && validation.productionWritesPerformed === false && validation.autoApply === false, detail: "run-scoped smoke pass without global autonomy" },
|
|
2881
|
-
{ name: "json_artifacts_body_free", passed: jsonArtifacts.every((name) => isRecord(parsedArtifacts[name]) && !hasForbiddenBodyKeys(parsedArtifacts[name])), detail: "no forbidden plaintext body keys" },
|
|
2882
|
-
];
|
|
2883
|
-
const failedChecks = checks.filter((check) => check.passed !== true).map((check) => check.name);
|
|
2884
|
-
const result = {
|
|
2885
|
-
schema: "zob.autonomous-readonly-smoke-artifact-validation.v1",
|
|
2886
|
-
runId: safeRunId,
|
|
2887
|
-
valid: failedChecks.length === 0,
|
|
2888
|
-
failedChecks,
|
|
2889
|
-
checks,
|
|
2890
|
-
artifacts,
|
|
2891
|
-
finalGlobalSentinels: {
|
|
2892
|
-
doneSentinelPresent: finalDoneSentinelPresent,
|
|
2893
|
-
globalReadySentinelPresent,
|
|
2894
|
-
doneSentinelAllowed: false,
|
|
2895
|
-
globalReadySentinelAllowed: false,
|
|
2896
|
-
},
|
|
2897
|
-
currentSourceFingerprintFreshness,
|
|
2898
|
-
globalAutonomyReady: false,
|
|
2899
|
-
globalAutonomyNoShip: true,
|
|
2900
|
-
childDispatchAllowed: false,
|
|
2901
|
-
daemonStarted: false,
|
|
2902
|
-
productionWritesPerformed: false,
|
|
2903
|
-
autoApply: false,
|
|
2904
|
-
liveRoutingEnabled: false,
|
|
2905
|
-
bodyStored: false,
|
|
2906
|
-
promptBodiesStored: false,
|
|
2907
|
-
outputBodiesStored: false,
|
|
2908
|
-
generatedAt: new Date().toISOString(),
|
|
2909
|
-
};
|
|
2910
|
-
if (hasForbiddenBodyKeys(result)) throw new Error("autonomous readonly smoke validation would store forbidden plaintext body keys");
|
|
2911
|
-
return result;
|
|
2912
|
-
}
|
|
1
|
+
export type { AutonomousApplyPolicy, AutonomousBudgetProfile, AutonomousRisk, AutonomousLevel, AutonomousRuntimeDryRunInput, AutonomousReadOnlySmokeRunInput } from "./autonomous-runtime/types.js";
|
|
2
|
+
export { buildAutonomousRuntimeDryRun, buildAutonomousRuntimeDryRunValidation, buildAutonomousRuntimeDryRunFinalReport } from "./autonomous-runtime/dry-run.js";
|
|
3
|
+
export { writeAutonomousRuntimeDryRunReport, writeAutonomousReadOnlySmokeRunReport } from "./autonomous-runtime/report-writers.js";
|
|
4
|
+
export { validateAutonomousRuntimeDryRunArtifacts, validateAutonomousReadOnlySmokeRunArtifacts } from "./autonomous-runtime/validation.js";
|