@riddledc/riddle-proof 0.5.0 → 0.5.2
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/README.md +6 -6
- package/dist/chunk-3MHFLQKG.js +853 -0
- package/dist/{chunk-LVP22WE4.js → chunk-5GZZZ6JA.js} +5 -1
- package/dist/engine-harness.cjs +2505 -22
- package/dist/engine-harness.js +1 -1
- package/dist/index.cjs +2512 -29
- package/dist/index.js +2 -2
- package/dist/openclaw.cjs +1 -1
- package/dist/openclaw.js +1 -1
- package/dist/proof-run-core.cjs +909 -0
- package/dist/proof-run-core.d.cts +280 -0
- package/dist/proof-run-core.d.ts +280 -0
- package/dist/proof-run-core.js +48 -0
- package/dist/proof-run-engine.cjs +2499 -0
- package/dist/proof-run-engine.d.cts +677 -0
- package/dist/proof-run-engine.d.ts +677 -0
- package/dist/proof-run-engine.js +1649 -0
- package/lib/workspace-core.mjs +391 -0
- package/package.json +15 -3
- package/runtime/lib/author.py +343 -0
- package/runtime/lib/implement.py +63 -0
- package/runtime/lib/preflight.py +246 -0
- package/runtime/lib/recon.py +1048 -0
- package/runtime/lib/riddle_core_call.mjs +151 -0
- package/runtime/lib/setup.py +387 -0
- package/runtime/lib/ship.py +834 -0
- package/runtime/lib/util.py +673 -0
- package/runtime/lib/verify.py +1223 -0
- package/runtime/pipelines/riddle-proof-author.lobster +28 -0
- package/runtime/pipelines/riddle-proof-implement.lobster +26 -0
- package/runtime/pipelines/riddle-proof-recon.lobster +79 -0
- package/runtime/pipelines/riddle-proof-setup.lobster +141 -0
- package/runtime/pipelines/riddle-proof-ship.lobster +36 -0
- package/runtime/pipelines/riddle-proof-verify.lobster +74 -0
- package/runtime/tests/recon_verify_smoke.py +1198 -0
package/dist/engine-harness.cjs
CHANGED
|
@@ -5,6 +5,9 @@ var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
|
|
|
5
5
|
var __getOwnPropNames = Object.getOwnPropertyNames;
|
|
6
6
|
var __getProtoOf = Object.getPrototypeOf;
|
|
7
7
|
var __hasOwnProp = Object.prototype.hasOwnProperty;
|
|
8
|
+
var __esm = (fn, res) => function __init() {
|
|
9
|
+
return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
|
|
10
|
+
};
|
|
8
11
|
var __export = (target, all) => {
|
|
9
12
|
for (var name in all)
|
|
10
13
|
__defProp(target, name, { get: all[name], enumerable: true });
|
|
@@ -27,6 +30,2482 @@ var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__ge
|
|
|
27
30
|
));
|
|
28
31
|
var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
|
|
29
32
|
|
|
33
|
+
// src/proof-run-core.ts
|
|
34
|
+
function currentDistDir() {
|
|
35
|
+
const meta = typeof import_meta === "object" ? import_meta : {};
|
|
36
|
+
if (typeof meta.url === "string" && meta.url) {
|
|
37
|
+
return import_node_path.default.dirname((0, import_node_url.fileURLToPath)(meta.url));
|
|
38
|
+
}
|
|
39
|
+
if (typeof __dirname === "string") return __dirname;
|
|
40
|
+
return process.cwd();
|
|
41
|
+
}
|
|
42
|
+
function createWorkflowStatePath() {
|
|
43
|
+
const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[-:.TZ]/g, "").slice(0, 14);
|
|
44
|
+
return `/tmp/riddle-proof-state-${stamp}-${(0, import_node_crypto.randomUUID)().slice(0, 8)}.json`;
|
|
45
|
+
}
|
|
46
|
+
function argsPathForStatePath(statePath) {
|
|
47
|
+
const base = import_node_path.default.basename(statePath);
|
|
48
|
+
if (base.startsWith("riddle-proof-state")) {
|
|
49
|
+
return import_node_path.default.join(import_node_path.default.dirname(statePath), base.replace("riddle-proof-state", "riddle-proof-args"));
|
|
50
|
+
}
|
|
51
|
+
return import_node_path.default.join(import_node_path.default.dirname(statePath), `${base}.args.json`);
|
|
52
|
+
}
|
|
53
|
+
function isRiddleProofSkillDir(candidate) {
|
|
54
|
+
return (0, import_node_fs.existsSync)(workflowFile(candidate, "setup"));
|
|
55
|
+
}
|
|
56
|
+
function resolveRiddleProofDir(config = {}) {
|
|
57
|
+
if (config.riddleProofDir) return config.riddleProofDir;
|
|
58
|
+
const candidates = [
|
|
59
|
+
process.env.RIDDLE_PROOF_DIR,
|
|
60
|
+
...RIDDLE_PROOF_DIR_CANDIDATES
|
|
61
|
+
].filter((candidate) => Boolean(candidate));
|
|
62
|
+
return candidates.find(isRiddleProofSkillDir) || candidates.find((candidate) => (0, import_node_fs.existsSync)(candidate)) || RIDDLE_PROOF_DIR_CANDIDATES[0];
|
|
63
|
+
}
|
|
64
|
+
function resolveConfig(config = {}, params = {}) {
|
|
65
|
+
const configuredStatePath = config.statePath || "";
|
|
66
|
+
const shouldIsolateWorkflow = !params.state_path && !configuredStatePath && (params.action === "setup" || params.action === "run");
|
|
67
|
+
const statePath = params.state_path || (shouldIsolateWorkflow ? createWorkflowStatePath() : configuredStatePath || "/tmp/riddle-proof-state.json");
|
|
68
|
+
return {
|
|
69
|
+
riddleProofDir: resolveRiddleProofDir(config),
|
|
70
|
+
statePath,
|
|
71
|
+
argsPath: argsPathForStatePath(statePath),
|
|
72
|
+
defaultReviewer: config.defaultReviewer || "davisdiehl"
|
|
73
|
+
};
|
|
74
|
+
}
|
|
75
|
+
function ensureAction(action) {
|
|
76
|
+
if (["setup", "recon", "author", "implement", "verify", "ship", "status", "sync", "run"].includes(action)) {
|
|
77
|
+
return action;
|
|
78
|
+
}
|
|
79
|
+
throw new Error(`Unsupported action: ${action}`);
|
|
80
|
+
}
|
|
81
|
+
function workflowFile(riddleProofDir, action) {
|
|
82
|
+
return import_node_path.default.join(riddleProofDir, "pipelines", `riddle-proof-${action}.lobster`);
|
|
83
|
+
}
|
|
84
|
+
function buildSetupArgs(params, config) {
|
|
85
|
+
if (!params.repo) throw new Error("repo is required for setup/run");
|
|
86
|
+
if (!params.change_request) throw new Error("change_request is required for setup/run");
|
|
87
|
+
const commitMessage = (params.commit_message || params.change_request || "").trim();
|
|
88
|
+
const captureScript = (params.capture_script || "").trim();
|
|
89
|
+
const requestedReference = params.reference || (params.prod_url ? "both" : "before");
|
|
90
|
+
const reference = !params.prod_url && requestedReference !== "before" ? "before" : requestedReference;
|
|
91
|
+
if (!commitMessage) throw new Error("commit_message is required for setup/run");
|
|
92
|
+
return {
|
|
93
|
+
repo: params.repo,
|
|
94
|
+
branch: params.branch || "",
|
|
95
|
+
change_request: params.change_request,
|
|
96
|
+
commit_message: commitMessage,
|
|
97
|
+
prod_url: params.prod_url || "",
|
|
98
|
+
capture_script: captureScript,
|
|
99
|
+
success_criteria: params.success_criteria || "",
|
|
100
|
+
assertions_json: params.assertions_json || "",
|
|
101
|
+
verification_mode: params.verification_mode || "proof",
|
|
102
|
+
reference,
|
|
103
|
+
base_branch: params.base_branch || "main",
|
|
104
|
+
before_ref: params.before_ref || "",
|
|
105
|
+
allow_static_preview_fallback: params.allow_static_preview_fallback ? "true" : "",
|
|
106
|
+
context: params.context || "",
|
|
107
|
+
reviewer: params.reviewer || config.defaultReviewer,
|
|
108
|
+
mode: params.mode || "",
|
|
109
|
+
build_command: params.build_command || "npm run build",
|
|
110
|
+
build_output: params.build_output || "build",
|
|
111
|
+
server_image: params.server_image || "node:20-slim",
|
|
112
|
+
server_command: params.server_command || "npm start",
|
|
113
|
+
server_port: String(params.server_port || 3e3),
|
|
114
|
+
server_path: params.server_path || "",
|
|
115
|
+
server_path_source: params.server_path ? "user" : "",
|
|
116
|
+
use_auth: params.use_auth ? "true" : "",
|
|
117
|
+
auth_localStorage_json: params.auth_localStorage_json || "",
|
|
118
|
+
auth_cookies_json: params.auth_cookies_json || "",
|
|
119
|
+
auth_headers_json: params.auth_headers_json || "",
|
|
120
|
+
color_scheme: params.color_scheme || "",
|
|
121
|
+
wait_for_selector: params.wait_for_selector || "",
|
|
122
|
+
discord_channel: params.discord_channel || "",
|
|
123
|
+
discord_thread_id: params.discord_thread_id || "",
|
|
124
|
+
discord_message_id: params.discord_message_id || "",
|
|
125
|
+
discord_source_url: params.discord_source_url || "",
|
|
126
|
+
leave_draft: params.leave_draft ? "true" : ""
|
|
127
|
+
};
|
|
128
|
+
}
|
|
129
|
+
function readState(statePath) {
|
|
130
|
+
if (!(0, import_node_fs.existsSync)(statePath)) {
|
|
131
|
+
return null;
|
|
132
|
+
}
|
|
133
|
+
return JSON.parse((0, import_node_fs.readFileSync)(statePath, "utf-8"));
|
|
134
|
+
}
|
|
135
|
+
function writeState(statePath, state) {
|
|
136
|
+
(0, import_node_fs.writeFileSync)(statePath, JSON.stringify(state, null, 2));
|
|
137
|
+
}
|
|
138
|
+
function normalizeOptionalString(value) {
|
|
139
|
+
return typeof value === "string" ? value.trim() : void 0;
|
|
140
|
+
}
|
|
141
|
+
function knownEnvironmentIssuesFromNotes(notes) {
|
|
142
|
+
const text = notes.toLowerCase();
|
|
143
|
+
const issues = [];
|
|
144
|
+
if ((text.includes("erofs") || text.includes("read-only file system")) && text.includes("node_modules") && (text.includes(".vite-temp") || text.includes("vite.config"))) {
|
|
145
|
+
issues.push({
|
|
146
|
+
code: "vite_temp_config_erofs",
|
|
147
|
+
source: "implementation_notes",
|
|
148
|
+
severity: "environment",
|
|
149
|
+
summary: "Focused build verification hit the known Vite temp-config EROFS issue in shared node_modules."
|
|
150
|
+
});
|
|
151
|
+
}
|
|
152
|
+
return issues;
|
|
153
|
+
}
|
|
154
|
+
function guardProofEvidenceGlobalAssignments(script) {
|
|
155
|
+
return script.replace(
|
|
156
|
+
/^(\s*)(globalThis|window|self)\.__riddleProofEvidence\s*=\s*([^;\n]+);\s*$/gm,
|
|
157
|
+
(_match, indent, root, expression) => `${indent}try { if (typeof ${root} !== "undefined" && ${root}) ${root}.__riddleProofEvidence = ${expression.trim()}; } catch {}`
|
|
158
|
+
);
|
|
159
|
+
}
|
|
160
|
+
function normalizeCaptureScript(value) {
|
|
161
|
+
const script = normalizeOptionalString(value) || "";
|
|
162
|
+
return script ? guardProofEvidenceGlobalAssignments(script) : "";
|
|
163
|
+
}
|
|
164
|
+
function appendProofSummaryLine(state, line) {
|
|
165
|
+
const text = String(line || "").trim();
|
|
166
|
+
if (!text) return;
|
|
167
|
+
const existing = typeof state.proof_summary === "string" ? state.proof_summary.trim() : "";
|
|
168
|
+
if (existing.includes(text)) return;
|
|
169
|
+
state.proof_summary = existing ? `${existing}
|
|
170
|
+
${text}` : text;
|
|
171
|
+
}
|
|
172
|
+
function hasAuthoredProofPlan(state = {}) {
|
|
173
|
+
return Boolean((state?.proof_plan || "").trim()) && Boolean((state?.capture_script || "").trim());
|
|
174
|
+
}
|
|
175
|
+
function syncAuthoringState(state = {}) {
|
|
176
|
+
const reconReady = ["ready_for_proof_plan", "completed"].includes(state?.recon_status || "");
|
|
177
|
+
const reconBlocked = ["needs_agent_decision", "needs_supervisor_judgment"].includes(state?.recon_status || "");
|
|
178
|
+
const reconExhausted = state?.recon_status === "exhausted";
|
|
179
|
+
if (reconBlocked) {
|
|
180
|
+
state.author_status = "needs_recon_judgment";
|
|
181
|
+
} else if (reconExhausted) {
|
|
182
|
+
state.author_status = "recon_exhausted";
|
|
183
|
+
} else if (!reconReady) {
|
|
184
|
+
state.author_status = state.author_status || "pending_recon";
|
|
185
|
+
} else if (state.author_status === "ready" || state.proof_plan_status === "ready") {
|
|
186
|
+
state.author_status = "ready";
|
|
187
|
+
} else {
|
|
188
|
+
state.author_status = "needs_authoring";
|
|
189
|
+
}
|
|
190
|
+
if (reconBlocked) {
|
|
191
|
+
state.proof_plan_status = "needs_recon_judgment";
|
|
192
|
+
} else if (reconExhausted) {
|
|
193
|
+
state.proof_plan_status = "recon_exhausted";
|
|
194
|
+
} else if (state.author_status === "ready" || state.proof_plan_status === "ready") {
|
|
195
|
+
state.proof_plan_status = "ready";
|
|
196
|
+
} else if (reconReady) {
|
|
197
|
+
state.proof_plan_status = "needs_authoring";
|
|
198
|
+
} else {
|
|
199
|
+
state.proof_plan_status = state.proof_plan_status || "pending_recon";
|
|
200
|
+
}
|
|
201
|
+
state.author_request = state.author_request || {};
|
|
202
|
+
state.author_summary = state.author_summary || "";
|
|
203
|
+
return state;
|
|
204
|
+
}
|
|
205
|
+
function ensureStageLoopState(state = {}) {
|
|
206
|
+
if (!state || typeof state !== "object") return state;
|
|
207
|
+
syncAuthoringState(state);
|
|
208
|
+
state.explicit_stage_gate = true;
|
|
209
|
+
state.stage_attempts = state.stage_attempts || {};
|
|
210
|
+
for (const stage of WORKFLOW_STAGE_ORDER) {
|
|
211
|
+
const current = state.stage_attempts[stage] || {};
|
|
212
|
+
state.stage_attempts[stage] = {
|
|
213
|
+
count: Number(current.count || 0),
|
|
214
|
+
last_status: current.last_status || null,
|
|
215
|
+
last_checkpoint: current.last_checkpoint || null,
|
|
216
|
+
last_summary: current.last_summary || null,
|
|
217
|
+
last_attempted_at: current.last_attempted_at || null,
|
|
218
|
+
history: Array.isArray(current.history) ? current.history : []
|
|
219
|
+
};
|
|
220
|
+
}
|
|
221
|
+
state.stage_decision_request = state.stage_decision_request || {};
|
|
222
|
+
state.active_checkpoint = state.active_checkpoint || null;
|
|
223
|
+
state.active_checkpoint_stage = state.active_checkpoint_stage || null;
|
|
224
|
+
state.last_requested_advance_stage = state.last_requested_advance_stage || null;
|
|
225
|
+
return state;
|
|
226
|
+
}
|
|
227
|
+
function clearStageDecisionRequest(state = {}) {
|
|
228
|
+
ensureStageLoopState(state);
|
|
229
|
+
state.stage_decision_request = {};
|
|
230
|
+
state.active_checkpoint = null;
|
|
231
|
+
state.active_checkpoint_stage = null;
|
|
232
|
+
return state;
|
|
233
|
+
}
|
|
234
|
+
function recordStageAttempt(state, stage, entry = {}) {
|
|
235
|
+
ensureStageLoopState(state);
|
|
236
|
+
const bucket = state.stage_attempts[stage];
|
|
237
|
+
const attemptNumber = Number(bucket.count || 0) + 1;
|
|
238
|
+
const item = {
|
|
239
|
+
attempt: attemptNumber,
|
|
240
|
+
stage,
|
|
241
|
+
status: entry.status || "completed",
|
|
242
|
+
checkpoint: entry.checkpoint || null,
|
|
243
|
+
summary: entry.summary || null,
|
|
244
|
+
requested_advance_stage: entry.requestedAdvanceStage || null,
|
|
245
|
+
halted_for_approval: Boolean(entry.haltedForApproval),
|
|
246
|
+
auto_approved: Boolean(entry.autoApproved),
|
|
247
|
+
error: entry.error || null,
|
|
248
|
+
details: entry.details || {},
|
|
249
|
+
attempted_at: (/* @__PURE__ */ new Date()).toISOString()
|
|
250
|
+
};
|
|
251
|
+
bucket.count = attemptNumber;
|
|
252
|
+
bucket.last_status = item.status;
|
|
253
|
+
bucket.last_checkpoint = item.checkpoint;
|
|
254
|
+
bucket.last_summary = item.summary;
|
|
255
|
+
bucket.last_attempted_at = item.attempted_at;
|
|
256
|
+
bucket.history = [...bucket.history, item].slice(-25);
|
|
257
|
+
state.stage_attempts[stage] = bucket;
|
|
258
|
+
state.last_requested_advance_stage = entry.requestedAdvanceStage || state.last_requested_advance_stage || null;
|
|
259
|
+
return item;
|
|
260
|
+
}
|
|
261
|
+
function setStageDecisionRequest(state, request) {
|
|
262
|
+
ensureStageLoopState(state);
|
|
263
|
+
const continueWithStage = request.continueWithStage || request.recommendedAdvanceStage || null;
|
|
264
|
+
state.active_checkpoint = request.checkpoint;
|
|
265
|
+
state.active_checkpoint_stage = request.stage;
|
|
266
|
+
state.stage_decision_request = {
|
|
267
|
+
stage: request.stage,
|
|
268
|
+
checkpoint: request.checkpoint,
|
|
269
|
+
summary: request.summary,
|
|
270
|
+
next_actions: request.nextActions || [],
|
|
271
|
+
advance_options: request.advanceOptions || [],
|
|
272
|
+
recommended_advance_stage: request.recommendedAdvanceStage || null,
|
|
273
|
+
continue_from_checkpoint: Boolean(continueWithStage),
|
|
274
|
+
continue_with_stage: continueWithStage,
|
|
275
|
+
blocking: Boolean(request.blocking),
|
|
276
|
+
details: request.details || {},
|
|
277
|
+
checkpoint_contract: request.checkpointContract || null,
|
|
278
|
+
updated_at: (/* @__PURE__ */ new Date()).toISOString()
|
|
279
|
+
};
|
|
280
|
+
return state.stage_decision_request;
|
|
281
|
+
}
|
|
282
|
+
function checkpointContinueStage(state) {
|
|
283
|
+
ensureStageLoopState(state);
|
|
284
|
+
const stage = state?.stage_decision_request?.continue_with_stage || state?.stage_decision_request?.recommended_advance_stage || null;
|
|
285
|
+
return WORKFLOW_STAGE_ORDER.includes(stage) ? stage : null;
|
|
286
|
+
}
|
|
287
|
+
function invalidateVerifyEvidence(state = {}) {
|
|
288
|
+
if (!state || typeof state !== "object") {
|
|
289
|
+
return { invalidated: false };
|
|
290
|
+
}
|
|
291
|
+
const hadAfterCdn = typeof state.after_cdn === "string" && state.after_cdn.trim().length > 0;
|
|
292
|
+
const hadVerifyResults = Boolean(state.verify_results && Object.keys(state.verify_results).length > 0);
|
|
293
|
+
const hadMergeRecommendation = typeof state.merge_recommendation === "string" && state.merge_recommendation.trim().length > 0;
|
|
294
|
+
const hadProofSummary = typeof state.proof_summary === "string" && state.proof_summary.trim().length > 0;
|
|
295
|
+
const hadVerifyStatus = typeof state.verify_status === "string" && state.verify_status.trim().length > 0;
|
|
296
|
+
const hadVerifySummary = typeof state.verify_summary === "string" && state.verify_summary.trim().length > 0;
|
|
297
|
+
const hadVerifyDecisionRequest = Boolean(state.verify_decision_request && Object.keys(state.verify_decision_request).length > 0);
|
|
298
|
+
const hadEvidenceNotes = Array.isArray(state.evidence_notes) && state.evidence_notes.length > 0;
|
|
299
|
+
const hadProofAssessment = Boolean(state.proof_assessment && Object.keys(state.proof_assessment).length > 0);
|
|
300
|
+
const hadProofAssessmentRequest = Boolean(state.proof_assessment_request && Object.keys(state.proof_assessment_request).length > 0);
|
|
301
|
+
const invalidated = hadAfterCdn || hadVerifyResults || hadMergeRecommendation || hadProofSummary || hadVerifyStatus || hadVerifySummary || hadVerifyDecisionRequest || hadEvidenceNotes || hadProofAssessment || hadProofAssessmentRequest;
|
|
302
|
+
if (invalidated) {
|
|
303
|
+
state.after_cdn = "";
|
|
304
|
+
state.verify_results = {};
|
|
305
|
+
state.verify_status = "";
|
|
306
|
+
state.verify_summary = "";
|
|
307
|
+
state.verify_decision_request = {};
|
|
308
|
+
state.merge_recommendation = null;
|
|
309
|
+
state.proof_summary = "";
|
|
310
|
+
state.evidence_notes = [];
|
|
311
|
+
state.proof_assessment = {};
|
|
312
|
+
state.proof_assessment_source = null;
|
|
313
|
+
state.proof_assessment_request = {};
|
|
314
|
+
}
|
|
315
|
+
return {
|
|
316
|
+
invalidated,
|
|
317
|
+
hadAfterCdn,
|
|
318
|
+
hadVerifyResults,
|
|
319
|
+
hadMergeRecommendation,
|
|
320
|
+
hadProofSummary,
|
|
321
|
+
hadVerifyStatus,
|
|
322
|
+
hadVerifySummary,
|
|
323
|
+
hadVerifyDecisionRequest,
|
|
324
|
+
hadEvidenceNotes,
|
|
325
|
+
hadProofAssessment,
|
|
326
|
+
hadProofAssessmentRequest
|
|
327
|
+
};
|
|
328
|
+
}
|
|
329
|
+
function normalizedReference(state = {}) {
|
|
330
|
+
const reference = String(state?.requested_reference || state?.reference || "before").trim();
|
|
331
|
+
return reference || "before";
|
|
332
|
+
}
|
|
333
|
+
function normalizedProofAssessment(state = {}) {
|
|
334
|
+
const proofAssessment = state?.proof_assessment || {};
|
|
335
|
+
const source = String(proofAssessment?.source || state?.proof_assessment_source || "").trim().toLowerCase();
|
|
336
|
+
const decision = String(proofAssessment?.decision || "").trim();
|
|
337
|
+
return {
|
|
338
|
+
decision: decision || null,
|
|
339
|
+
source: source || null
|
|
340
|
+
};
|
|
341
|
+
}
|
|
342
|
+
function requiredBaselineLabelsForState(state = {}) {
|
|
343
|
+
const reference = normalizedReference(state);
|
|
344
|
+
const labels = [];
|
|
345
|
+
if (reference === "before" || reference === "both") labels.push("before");
|
|
346
|
+
if (reference === "prod" || reference === "both") labels.push("prod");
|
|
347
|
+
return labels;
|
|
348
|
+
}
|
|
349
|
+
function validateShipGate(state = {}) {
|
|
350
|
+
const reference = normalizedReference(state);
|
|
351
|
+
const prodUrl = String(state?.prod_url || "").trim();
|
|
352
|
+
const beforeCdn = String(state?.before_cdn || "").trim();
|
|
353
|
+
const prodCdn = String(state?.prod_cdn || "").trim();
|
|
354
|
+
const afterCdn = String(state?.after_cdn || "").trim();
|
|
355
|
+
const verifyStatus = String(state?.verify_status || "").trim();
|
|
356
|
+
const proofAssessment = normalizedProofAssessment(state);
|
|
357
|
+
const reasons = [];
|
|
358
|
+
if (!["before", "prod", "both"].includes(reference)) {
|
|
359
|
+
reasons.push(`reference must be before, prod, or both; got ${reference}`);
|
|
360
|
+
}
|
|
361
|
+
const requiredBaselines = requiredBaselineLabelsForState(state);
|
|
362
|
+
if (requiredBaselines.includes("before") && !beforeCdn) {
|
|
363
|
+
reasons.push("before_cdn is required before ship");
|
|
364
|
+
}
|
|
365
|
+
if (requiredBaselines.includes("prod")) {
|
|
366
|
+
if (!prodUrl) {
|
|
367
|
+
reasons.push(`prod_url is required when reference=${reference}`);
|
|
368
|
+
}
|
|
369
|
+
if (!prodCdn) {
|
|
370
|
+
reasons.push("prod_cdn is required before ship");
|
|
371
|
+
}
|
|
372
|
+
}
|
|
373
|
+
if (!afterCdn) {
|
|
374
|
+
reasons.push("after_cdn is required before ship");
|
|
375
|
+
}
|
|
376
|
+
if (verifyStatus !== "evidence_captured") {
|
|
377
|
+
reasons.push("verify_status must be evidence_captured before ship");
|
|
378
|
+
}
|
|
379
|
+
if (!["supervising_agent", "supervisor"].includes(proofAssessment.source || "")) {
|
|
380
|
+
reasons.push("proof_assessment.source must be supervising_agent before ship");
|
|
381
|
+
}
|
|
382
|
+
if (proofAssessment.decision !== "ready_to_ship") {
|
|
383
|
+
reasons.push("proof_assessment.decision must be ready_to_ship before ship");
|
|
384
|
+
}
|
|
385
|
+
return {
|
|
386
|
+
ok: reasons.length === 0,
|
|
387
|
+
reasons,
|
|
388
|
+
required_baselines: requiredBaselines,
|
|
389
|
+
evidence: {
|
|
390
|
+
reference,
|
|
391
|
+
prod_url: prodUrl || null,
|
|
392
|
+
before_cdn: beforeCdn || null,
|
|
393
|
+
prod_cdn: prodCdn || null,
|
|
394
|
+
after_cdn: afterCdn || null,
|
|
395
|
+
verify_status: verifyStatus || null,
|
|
396
|
+
proof_assessment_decision: proofAssessment.decision,
|
|
397
|
+
proof_assessment_source: proofAssessment.source
|
|
398
|
+
}
|
|
399
|
+
};
|
|
400
|
+
}
|
|
401
|
+
function buildCheckpointContract(state, request) {
|
|
402
|
+
const spec = CHECKPOINT_CONTRACT_SPECS[request.checkpoint] || {
|
|
403
|
+
purpose: request.summary
|
|
404
|
+
};
|
|
405
|
+
const continueWithStage = request.continueWithStage || request.recommendedAdvanceStage || null;
|
|
406
|
+
const payload = {
|
|
407
|
+
version: CHECKPOINT_CONTRACT_VERSION,
|
|
408
|
+
checkpoint: request.checkpoint,
|
|
409
|
+
stage: request.stage,
|
|
410
|
+
purpose: spec.purpose,
|
|
411
|
+
summary: request.summary,
|
|
412
|
+
blocking: Boolean(request.blocking),
|
|
413
|
+
next_actions: request.nextActions || [],
|
|
414
|
+
advance_options: request.advanceOptions || [],
|
|
415
|
+
accepted_inputs: spec.accepted_inputs || [],
|
|
416
|
+
response_schema: spec.response_schema || null,
|
|
417
|
+
required_state: spec.required_state || [],
|
|
418
|
+
resume: {
|
|
419
|
+
action: "run",
|
|
420
|
+
state_path: request.statePath || null,
|
|
421
|
+
continue_from_checkpoint: Boolean(continueWithStage),
|
|
422
|
+
continue_with_stage: continueWithStage
|
|
423
|
+
}
|
|
424
|
+
};
|
|
425
|
+
if (spec.include_ship_gate) {
|
|
426
|
+
payload.ship_gate = validateShipGate(state);
|
|
427
|
+
}
|
|
428
|
+
return payload;
|
|
429
|
+
}
|
|
430
|
+
function mergeStateFromParams(statePath, params) {
|
|
431
|
+
const state = readState(statePath);
|
|
432
|
+
if (!state) return null;
|
|
433
|
+
ensureStageLoopState(state);
|
|
434
|
+
const stringFields = [
|
|
435
|
+
"change_request",
|
|
436
|
+
"commit_message",
|
|
437
|
+
"prod_url",
|
|
438
|
+
"capture_script",
|
|
439
|
+
"success_criteria",
|
|
440
|
+
"assertions_json",
|
|
441
|
+
"verification_mode",
|
|
442
|
+
"base_branch",
|
|
443
|
+
"before_ref",
|
|
444
|
+
"context",
|
|
445
|
+
"reviewer",
|
|
446
|
+
"build_command",
|
|
447
|
+
"build_output",
|
|
448
|
+
"server_image",
|
|
449
|
+
"server_command",
|
|
450
|
+
"server_path",
|
|
451
|
+
"wait_for_selector",
|
|
452
|
+
"discord_channel",
|
|
453
|
+
"discord_thread_id",
|
|
454
|
+
"discord_message_id",
|
|
455
|
+
"discord_source_url",
|
|
456
|
+
"auth_localStorage_json",
|
|
457
|
+
"auth_cookies_json",
|
|
458
|
+
"auth_headers_json",
|
|
459
|
+
"proof_plan",
|
|
460
|
+
"implementation_notes"
|
|
461
|
+
];
|
|
462
|
+
for (const field of stringFields) {
|
|
463
|
+
if (params[field] !== void 0) {
|
|
464
|
+
state[field] = normalizeOptionalString(params[field]);
|
|
465
|
+
}
|
|
466
|
+
}
|
|
467
|
+
if (params.server_path !== void 0) {
|
|
468
|
+
state.server_path_source = "tool_param";
|
|
469
|
+
}
|
|
470
|
+
if (params.implementation_notes !== void 0) {
|
|
471
|
+
const issues = knownEnvironmentIssuesFromNotes(state.implementation_notes || "");
|
|
472
|
+
if (issues.length) state.implementation_environment_issues = issues;
|
|
473
|
+
}
|
|
474
|
+
if (params.reference !== void 0) state.reference = params.reference;
|
|
475
|
+
if (params.mode !== void 0) state.mode = params.mode;
|
|
476
|
+
if (params.allow_static_preview_fallback !== void 0) {
|
|
477
|
+
state.allow_static_preview_fallback = params.allow_static_preview_fallback;
|
|
478
|
+
}
|
|
479
|
+
if (params.server_port !== void 0) state.server_port = String(params.server_port);
|
|
480
|
+
if (params.color_scheme !== void 0) state.color_scheme = params.color_scheme || "";
|
|
481
|
+
if (params.use_auth !== void 0) state.use_auth = params.use_auth ? "true" : "";
|
|
482
|
+
if (params.leave_draft !== void 0) state.leave_draft = params.leave_draft ? "true" : "";
|
|
483
|
+
if (params.advance_stage !== void 0) state.last_requested_advance_stage = params.advance_stage;
|
|
484
|
+
if (params.recon_assessment_json !== void 0) {
|
|
485
|
+
const raw = normalizeOptionalString(params.recon_assessment_json) || "";
|
|
486
|
+
if (!raw) {
|
|
487
|
+
state.recon_assessment = {};
|
|
488
|
+
state.recon_assessment_source = null;
|
|
489
|
+
} else {
|
|
490
|
+
const parsed = JSON.parse(raw);
|
|
491
|
+
state.recon_assessment = {
|
|
492
|
+
...parsed,
|
|
493
|
+
source: (parsed?.source || "supervising_agent").toString()
|
|
494
|
+
};
|
|
495
|
+
state.recon_assessment_source = state.recon_assessment.source;
|
|
496
|
+
if (parsed?.baseline_understanding && typeof parsed.baseline_understanding === "object") {
|
|
497
|
+
state.recon_baseline_understanding = parsed.baseline_understanding;
|
|
498
|
+
}
|
|
499
|
+
const refined = parsed?.refined_inputs || {};
|
|
500
|
+
if (typeof refined?.server_path === "string") {
|
|
501
|
+
state.server_path = normalizeOptionalString(refined.server_path) || "";
|
|
502
|
+
state.server_path_source = "supervising_agent";
|
|
503
|
+
}
|
|
504
|
+
if (typeof refined?.wait_for_selector === "string") state.wait_for_selector = normalizeOptionalString(refined.wait_for_selector) || "";
|
|
505
|
+
if (typeof refined?.reference === "string" && refined.reference.trim()) state.reference = refined.reference.trim();
|
|
506
|
+
}
|
|
507
|
+
}
|
|
508
|
+
if (params.author_packet_json !== void 0) {
|
|
509
|
+
const raw = normalizeOptionalString(params.author_packet_json) || "";
|
|
510
|
+
if (!raw) {
|
|
511
|
+
state.supervisor_author_packet = null;
|
|
512
|
+
} else {
|
|
513
|
+
const parsed = JSON.parse(raw);
|
|
514
|
+
if (typeof parsed?.capture_script === "string") {
|
|
515
|
+
parsed.capture_script = normalizeCaptureScript(parsed.capture_script);
|
|
516
|
+
}
|
|
517
|
+
state.supervisor_author_packet = parsed;
|
|
518
|
+
if (typeof parsed?.proof_plan === "string") state.proof_plan = normalizeOptionalString(parsed.proof_plan) || "";
|
|
519
|
+
if (typeof parsed?.capture_script === "string") state.capture_script = normalizeCaptureScript(parsed.capture_script);
|
|
520
|
+
if (parsed?.baseline_understanding_used && typeof parsed.baseline_understanding_used === "object") {
|
|
521
|
+
state.author_baseline_understanding_used = parsed.baseline_understanding_used;
|
|
522
|
+
}
|
|
523
|
+
const refined = parsed?.refined_inputs || {};
|
|
524
|
+
if (typeof refined?.server_path === "string") {
|
|
525
|
+
state.server_path = normalizeOptionalString(refined.server_path) || "";
|
|
526
|
+
state.server_path_source = "supervising_agent";
|
|
527
|
+
}
|
|
528
|
+
if (typeof refined?.wait_for_selector === "string") state.wait_for_selector = normalizeOptionalString(refined.wait_for_selector) || "";
|
|
529
|
+
if (typeof refined?.reference === "string" && refined.reference.trim()) state.reference = refined.reference.trim();
|
|
530
|
+
if (typeof parsed?.confidence === "string") state.supervisor_author_confidence = normalizeOptionalString(parsed.confidence) || null;
|
|
531
|
+
if (parsed?.rationale !== void 0) state.supervisor_author_rationale = parsed.rationale;
|
|
532
|
+
if (typeof parsed?.summary === "string") state.supervisor_author_summary = normalizeOptionalString(parsed.summary) || null;
|
|
533
|
+
invalidateVerifyEvidence(state);
|
|
534
|
+
}
|
|
535
|
+
}
|
|
536
|
+
if (params.proof_assessment_json !== void 0) {
|
|
537
|
+
const raw = normalizeOptionalString(params.proof_assessment_json) || "";
|
|
538
|
+
if (!raw) {
|
|
539
|
+
state.proof_assessment = {};
|
|
540
|
+
state.proof_assessment_source = null;
|
|
541
|
+
} else {
|
|
542
|
+
const parsed = JSON.parse(raw);
|
|
543
|
+
state.proof_assessment = {
|
|
544
|
+
...parsed,
|
|
545
|
+
source: (parsed?.source || "supervising_agent").toString()
|
|
546
|
+
};
|
|
547
|
+
state.proof_assessment_source = state.proof_assessment.source;
|
|
548
|
+
if (typeof parsed?.decision === "string") {
|
|
549
|
+
state.proof_decision = parsed.decision;
|
|
550
|
+
}
|
|
551
|
+
if (typeof parsed?.summary === "string") {
|
|
552
|
+
state.proof_assessment_summary = normalizeOptionalString(parsed.summary) || null;
|
|
553
|
+
}
|
|
554
|
+
if (parsed?.decision === "ready_to_ship") {
|
|
555
|
+
state.merge_recommendation = "ready-to-ship";
|
|
556
|
+
} else if (typeof parsed?.decision === "string" && parsed.decision.trim()) {
|
|
557
|
+
state.merge_recommendation = "do-not-merge";
|
|
558
|
+
}
|
|
559
|
+
appendProofSummaryLine(state, `Supervising proof assessment: ${state.proof_assessment.decision || "unknown"}`);
|
|
560
|
+
if (state.proof_assessment_summary) {
|
|
561
|
+
appendProofSummaryLine(state, `Assessment summary: ${state.proof_assessment_summary}`);
|
|
562
|
+
}
|
|
563
|
+
const reasons = Array.isArray(parsed?.reasons) ? parsed.reasons.filter((item) => typeof item === "string" && item.trim()).slice(0, 4) : [];
|
|
564
|
+
if (reasons.length) {
|
|
565
|
+
appendProofSummaryLine(state, `Assessment reasons: ${reasons.join("; ")}`);
|
|
566
|
+
}
|
|
567
|
+
}
|
|
568
|
+
}
|
|
569
|
+
if (params.assertions_json !== void 0) {
|
|
570
|
+
const raw = normalizeOptionalString(params.assertions_json) || "";
|
|
571
|
+
if (!raw) {
|
|
572
|
+
state.parsed_assertions = null;
|
|
573
|
+
} else {
|
|
574
|
+
state.parsed_assertions = JSON.parse(raw);
|
|
575
|
+
}
|
|
576
|
+
}
|
|
577
|
+
if ((params.proof_plan !== void 0 || params.capture_script !== void 0) && hasAuthoredProofPlan(state)) {
|
|
578
|
+
state.author_summary = state.author_summary || "Proof authoring inputs were updated from tool params.";
|
|
579
|
+
}
|
|
580
|
+
syncAuthoringState(state);
|
|
581
|
+
writeState(statePath, state);
|
|
582
|
+
return state;
|
|
583
|
+
}
|
|
584
|
+
function summarizeState(state) {
|
|
585
|
+
if (!state) {
|
|
586
|
+
return {
|
|
587
|
+
stage: "missing",
|
|
588
|
+
summary: "No riddle-proof state file exists yet.",
|
|
589
|
+
state: null
|
|
590
|
+
};
|
|
591
|
+
}
|
|
592
|
+
ensureStageLoopState(state);
|
|
593
|
+
const attemptCounts = Object.fromEntries(
|
|
594
|
+
WORKFLOW_STAGE_ORDER.map((stage) => [stage, Number(state.stage_attempts?.[stage]?.count || 0)])
|
|
595
|
+
);
|
|
596
|
+
const selected = {
|
|
597
|
+
workspace_ready: Boolean(state.workspace_ready),
|
|
598
|
+
repo: state.repo || null,
|
|
599
|
+
branch: state.branch || null,
|
|
600
|
+
mode: state.mode || null,
|
|
601
|
+
reference: state.reference || null,
|
|
602
|
+
before_ref: state.before_ref || null,
|
|
603
|
+
allow_static_preview_fallback: Boolean(state.allow_static_preview_fallback),
|
|
604
|
+
commit_message: state.commit_message || null,
|
|
605
|
+
author_status: state.author_status || null,
|
|
606
|
+
author_summary: state.author_summary || null,
|
|
607
|
+
author_request: state.author_request || null,
|
|
608
|
+
proof_plan_status: state.proof_plan_status || null,
|
|
609
|
+
proof_plan: state.proof_plan || null,
|
|
610
|
+
proof_plan_request: state.proof_plan_request || null,
|
|
611
|
+
proof_profile_applied: Boolean(state.proof_profile),
|
|
612
|
+
proof_profile: state.proof_profile || null,
|
|
613
|
+
recon_status: state.recon_status || null,
|
|
614
|
+
recon_assessment: state.recon_assessment || null,
|
|
615
|
+
recon_assessment_request: state.recon_assessment_request || null,
|
|
616
|
+
recon_assessment_source: state.recon_assessment_source || null,
|
|
617
|
+
recon_decision_request: state.recon_decision_request || null,
|
|
618
|
+
recon_attempts_used: Array.isArray(state.recon_results?.attempt_history) ? state.recon_results.attempt_history.length : 0,
|
|
619
|
+
recon_attempts_max: state.recon_results?.max_attempts || null,
|
|
620
|
+
recon_hypothesis: state.recon_hypothesis || null,
|
|
621
|
+
implementation_status: state.implementation_status || null,
|
|
622
|
+
implementation_summary: state.implementation_summary || null,
|
|
623
|
+
implementation_environment_issues: state.implementation_environment_issues || [],
|
|
624
|
+
verify_status: state.verify_status || null,
|
|
625
|
+
verify_summary: state.verify_summary || null,
|
|
626
|
+
verify_decision_request: state.verify_decision_request || null,
|
|
627
|
+
proof_assessment: state.proof_assessment || null,
|
|
628
|
+
proof_assessment_request: state.proof_assessment_request || null,
|
|
629
|
+
proof_assessment_source: state.proof_assessment_source || null,
|
|
630
|
+
merge_recommendation: state.merge_recommendation || null,
|
|
631
|
+
before_cdn: state.before_cdn || null,
|
|
632
|
+
after_cdn: state.after_cdn || null,
|
|
633
|
+
prod_cdn: state.prod_cdn || null,
|
|
634
|
+
pr_url: state.pr_url || null,
|
|
635
|
+
pr_branch: state.target_branch || state.branch || null,
|
|
636
|
+
pr_state: state.pr_state || null,
|
|
637
|
+
ship_commit: state.ship_commit || null,
|
|
638
|
+
ship_remote_head: state.ship_remote_head || null,
|
|
639
|
+
merge_commit: state.merge_commit || null,
|
|
640
|
+
merged_at: state.merged_at || null,
|
|
641
|
+
ship_push: state.ship_push || null,
|
|
642
|
+
ship_report: state.ship_report || null,
|
|
643
|
+
cleanup_report: state.cleanup_report || null,
|
|
644
|
+
proof_comment_url: state.proof_comment_url || null,
|
|
645
|
+
proof_assessment_comment_url: state.proof_assessment_comment_url || null,
|
|
646
|
+
marked_ready: typeof state.marked_ready === "boolean" ? state.marked_ready : null,
|
|
647
|
+
left_draft: typeof state.left_draft === "boolean" ? state.left_draft : null,
|
|
648
|
+
ci_status: state.ci_status || null,
|
|
649
|
+
reviewer: state.reviewer || null,
|
|
650
|
+
active_checkpoint: state.active_checkpoint || null,
|
|
651
|
+
active_checkpoint_stage: state.active_checkpoint_stage || null,
|
|
652
|
+
continue_with_stage: checkpointContinueStage(state),
|
|
653
|
+
stage_decision_request: state.stage_decision_request || {},
|
|
654
|
+
stage_attempt_counts: attemptCounts,
|
|
655
|
+
explicit_stage_gate: Boolean(state.explicit_stage_gate),
|
|
656
|
+
last_requested_advance_stage: state.last_requested_advance_stage || null,
|
|
657
|
+
recon_results: state.recon_results || null,
|
|
658
|
+
verify_results: state.verify_results || null
|
|
659
|
+
};
|
|
660
|
+
const parts = [
|
|
661
|
+
state.workspace_ready ? "workspace ready" : "workspace not ready",
|
|
662
|
+
state.mode ? `mode=${state.mode}` : null,
|
|
663
|
+
state.reference ? `reference=${state.reference}` : null,
|
|
664
|
+
state.author_status ? `author=${state.author_status}` : null,
|
|
665
|
+
state.proof_plan_status ? `proof=${state.proof_plan_status}` : null,
|
|
666
|
+
state.recon_status ? `recon=${state.recon_status}` : null,
|
|
667
|
+
state.implementation_status ? `implement=${state.implementation_status}` : null,
|
|
668
|
+
state.verify_status ? `verify=${state.verify_status}` : null,
|
|
669
|
+
state.active_checkpoint ? `checkpoint=${state.active_checkpoint}` : null,
|
|
670
|
+
state.after_cdn ? "after evidence captured" : null,
|
|
671
|
+
state.pr_url ? "PR linked" : null
|
|
672
|
+
].filter(Boolean);
|
|
673
|
+
return {
|
|
674
|
+
stage: state.stage || (state.after_cdn ? "verified" : state.workspace_ready ? "setup" : "unknown"),
|
|
675
|
+
summary: parts.length ? parts.join(", ") : "State file present.",
|
|
676
|
+
state: selected
|
|
677
|
+
};
|
|
678
|
+
}
|
|
679
|
+
var import_node_fs, import_node_crypto, import_node_path, import_node_url, import_meta, WORKFLOW_STAGE_ORDER, CHECKPOINT_CONTRACT_VERSION, BUNDLED_RIDDLE_PROOF_DIR, RIDDLE_PROOF_DIR_CANDIDATES, CHECKPOINT_CONTRACT_SPECS;
|
|
680
|
+
var init_proof_run_core = __esm({
|
|
681
|
+
"src/proof-run-core.ts"() {
|
|
682
|
+
"use strict";
|
|
683
|
+
import_node_fs = require("fs");
|
|
684
|
+
import_node_crypto = require("crypto");
|
|
685
|
+
import_node_path = __toESM(require("path"), 1);
|
|
686
|
+
import_node_url = require("url");
|
|
687
|
+
import_meta = {};
|
|
688
|
+
WORKFLOW_STAGE_ORDER = ["setup", "recon", "author", "implement", "verify", "ship"];
|
|
689
|
+
CHECKPOINT_CONTRACT_VERSION = "riddle-proof-run.checkpoint.v1";
|
|
690
|
+
BUNDLED_RIDDLE_PROOF_DIR = import_node_path.default.resolve(
|
|
691
|
+
currentDistDir(),
|
|
692
|
+
"..",
|
|
693
|
+
"runtime"
|
|
694
|
+
);
|
|
695
|
+
RIDDLE_PROOF_DIR_CANDIDATES = [
|
|
696
|
+
BUNDLED_RIDDLE_PROOF_DIR
|
|
697
|
+
];
|
|
698
|
+
CHECKPOINT_CONTRACT_SPECS = {
|
|
699
|
+
setup_review: {
|
|
700
|
+
purpose: "Inspect prepared workspace state before moving into recon."
|
|
701
|
+
},
|
|
702
|
+
recon_supervisor_judgment: {
|
|
703
|
+
purpose: "Supervising agent judges the latest recon attempt and either retries recon, promotes baselines for authoring, or escalates.",
|
|
704
|
+
accepted_inputs: [{
|
|
705
|
+
name: "recon_assessment_json",
|
|
706
|
+
type: "json",
|
|
707
|
+
required: true,
|
|
708
|
+
description: "JSON assessment with decision, summary, baseline_understanding, continue_with_stage, escalation_target, refined_inputs, and reasons."
|
|
709
|
+
}],
|
|
710
|
+
response_schema: {
|
|
711
|
+
decision: ["retry_recon", "ready_for_author", "recon_stuck"],
|
|
712
|
+
summary: "string",
|
|
713
|
+
baseline_understanding: {
|
|
714
|
+
reference: ["before", "prod", "both", "unknown"],
|
|
715
|
+
target_route: "string",
|
|
716
|
+
before_evidence_url: "string",
|
|
717
|
+
visible_before_state: "string",
|
|
718
|
+
relevant_elements: ["string"],
|
|
719
|
+
requested_change: "string",
|
|
720
|
+
proof_focus: "string",
|
|
721
|
+
stop_condition: "string",
|
|
722
|
+
quality_risks: ["string"]
|
|
723
|
+
},
|
|
724
|
+
continue_with_stage: ["recon", "author"],
|
|
725
|
+
escalation_target: ["agent", "human"],
|
|
726
|
+
refined_inputs: {
|
|
727
|
+
server_path: "string",
|
|
728
|
+
wait_for_selector: "string",
|
|
729
|
+
reference: ["before", "prod", "both"]
|
|
730
|
+
},
|
|
731
|
+
reasons: ["string"],
|
|
732
|
+
source: "supervising_agent"
|
|
733
|
+
},
|
|
734
|
+
required_state: ["recon_assessment_request", "recon_results.attempt_history"]
|
|
735
|
+
},
|
|
736
|
+
recon_review: {
|
|
737
|
+
purpose: "Recon baselines were promoted and the workflow is ready for proof authoring."
|
|
738
|
+
},
|
|
739
|
+
recon_human_escalation: {
|
|
740
|
+
purpose: "Recon is explicitly blocked for human direction after supervising-agent escalation."
|
|
741
|
+
},
|
|
742
|
+
author_supervisor_judgment: {
|
|
743
|
+
purpose: "Supervising agent authors the proof packet from recon observations.",
|
|
744
|
+
accepted_inputs: [{
|
|
745
|
+
name: "author_packet_json",
|
|
746
|
+
type: "json",
|
|
747
|
+
required: true,
|
|
748
|
+
description: "JSON proof packet with proof_plan, capture_script, optional refined_inputs, rationale, confidence, and summary."
|
|
749
|
+
}],
|
|
750
|
+
response_schema: {
|
|
751
|
+
proof_plan: "string",
|
|
752
|
+
capture_script: "string",
|
|
753
|
+
baseline_understanding_used: {
|
|
754
|
+
reference: ["before", "prod", "both", "unknown"],
|
|
755
|
+
target_route: "string",
|
|
756
|
+
before_evidence_url: "string",
|
|
757
|
+
visible_before_state: "string",
|
|
758
|
+
relevant_elements: ["string"],
|
|
759
|
+
requested_change: "string",
|
|
760
|
+
proof_focus: "string",
|
|
761
|
+
stop_condition: "string",
|
|
762
|
+
quality_risks: ["string"]
|
|
763
|
+
},
|
|
764
|
+
refined_inputs: {
|
|
765
|
+
server_path: "string",
|
|
766
|
+
wait_for_selector: "string",
|
|
767
|
+
reference: ["before", "prod", "both"]
|
|
768
|
+
},
|
|
769
|
+
rationale: ["string"],
|
|
770
|
+
confidence: "low | medium | high",
|
|
771
|
+
summary: "string"
|
|
772
|
+
},
|
|
773
|
+
required_state: ["author_request", "recon_results.baselines"]
|
|
774
|
+
},
|
|
775
|
+
author_review: {
|
|
776
|
+
purpose: "Proof packet is ready; choose whether to inspect, re-author, implement, or verify."
|
|
777
|
+
},
|
|
778
|
+
implement_changes_missing: {
|
|
779
|
+
purpose: "Implementation stage did not detect material code changes.",
|
|
780
|
+
accepted_inputs: [{
|
|
781
|
+
name: "make_code_changes",
|
|
782
|
+
type: "external_action",
|
|
783
|
+
required: true,
|
|
784
|
+
description: "Make the actual code changes in the after worktree, then resume the implement stage."
|
|
785
|
+
}]
|
|
786
|
+
},
|
|
787
|
+
implement_review: {
|
|
788
|
+
purpose: "Implementation changes were detected and stale verify evidence, if any, was invalidated."
|
|
789
|
+
},
|
|
790
|
+
implement_required: {
|
|
791
|
+
purpose: "Verify cannot run until implementation changes are recorded.",
|
|
792
|
+
accepted_inputs: [{
|
|
793
|
+
name: "make_code_changes",
|
|
794
|
+
type: "external_action",
|
|
795
|
+
required: true,
|
|
796
|
+
description: "Make the requested code change, then advance the run to implement."
|
|
797
|
+
}]
|
|
798
|
+
},
|
|
799
|
+
verify_capture_retry: {
|
|
800
|
+
purpose: "Verify capture was incomplete; the agent loop should revise capture authoring before shipping.",
|
|
801
|
+
accepted_inputs: [{
|
|
802
|
+
name: "author_packet_json",
|
|
803
|
+
type: "json",
|
|
804
|
+
description: "Optional revised proof packet if returning to author."
|
|
805
|
+
}],
|
|
806
|
+
required_state: ["verify_decision_request"]
|
|
807
|
+
},
|
|
808
|
+
verify_supervisor_judgment: {
|
|
809
|
+
purpose: "Supervising agent judges whether captured evidence proves the change is ready to ship.",
|
|
810
|
+
accepted_inputs: [{
|
|
811
|
+
name: "proof_assessment_json",
|
|
812
|
+
type: "json",
|
|
813
|
+
required: true,
|
|
814
|
+
description: "JSON assessment with decision, summary, recommended_stage, continue_with_stage, escalation_target, and reasons."
|
|
815
|
+
}],
|
|
816
|
+
response_schema: {
|
|
817
|
+
decision: ["ready_to_ship", "needs_richer_proof"],
|
|
818
|
+
summary: "string",
|
|
819
|
+
recommended_stage: ["ship", "author", "implement", "recon", "verify"],
|
|
820
|
+
continue_with_stage: ["ship", "author", "implement", "recon", "verify"],
|
|
821
|
+
escalation_target: ["agent", "human"],
|
|
822
|
+
reasons: ["string"],
|
|
823
|
+
source: "supervising_agent"
|
|
824
|
+
},
|
|
825
|
+
required_state: ["before_cdn or prod_cdn", "after_cdn", "proof_assessment_request"],
|
|
826
|
+
include_ship_gate: true
|
|
827
|
+
},
|
|
828
|
+
verify_human_escalation: {
|
|
829
|
+
purpose: "Proof loop is explicitly escalated to the human after supervising-agent judgment."
|
|
830
|
+
},
|
|
831
|
+
verify_agent_retry: {
|
|
832
|
+
purpose: "Supervising agent judged the proof insufficient and kept the workflow inside the internal loop.",
|
|
833
|
+
accepted_inputs: [{
|
|
834
|
+
name: "author_packet_json",
|
|
835
|
+
type: "json",
|
|
836
|
+
description: "Optional revised proof packet if the checkpoint resumes to author."
|
|
837
|
+
}]
|
|
838
|
+
},
|
|
839
|
+
verify_ship_ready: {
|
|
840
|
+
purpose: "Proof assessment is ready_to_ship and the next continuation should enter ship.",
|
|
841
|
+
include_ship_gate: true
|
|
842
|
+
},
|
|
843
|
+
verify_required: {
|
|
844
|
+
purpose: "Ship is blocked until verify captures usable after evidence.",
|
|
845
|
+
include_ship_gate: true
|
|
846
|
+
},
|
|
847
|
+
verify_supervisor_judgment_required: {
|
|
848
|
+
purpose: "Ship is blocked until the supervising agent judges the current evidence ready_to_ship.",
|
|
849
|
+
accepted_inputs: [{
|
|
850
|
+
name: "proof_assessment_json",
|
|
851
|
+
type: "json",
|
|
852
|
+
required: true,
|
|
853
|
+
description: "JSON assessment that must use decision=ready_to_ship before ship can continue."
|
|
854
|
+
}],
|
|
855
|
+
include_ship_gate: true
|
|
856
|
+
},
|
|
857
|
+
ship_gate_blocked: {
|
|
858
|
+
purpose: "Ship is blocked because required baseline, after evidence, or supervising-agent proof approval is missing.",
|
|
859
|
+
include_ship_gate: true
|
|
860
|
+
},
|
|
861
|
+
ship_review: {
|
|
862
|
+
purpose: "Ship completed; inspect the PR, proof comment, and CI status.",
|
|
863
|
+
include_ship_gate: true
|
|
864
|
+
}
|
|
865
|
+
};
|
|
866
|
+
}
|
|
867
|
+
});
|
|
868
|
+
|
|
869
|
+
// src/proof-run-engine.ts
|
|
870
|
+
var proof_run_engine_exports = {};
|
|
871
|
+
__export(proof_run_engine_exports, {
|
|
872
|
+
createRiddleProofEngine: () => createRiddleProofEngine,
|
|
873
|
+
executeWorkflow: () => executeWorkflow
|
|
874
|
+
});
|
|
875
|
+
function snapshotFor(statePath) {
|
|
876
|
+
return summarizeState(readState(statePath));
|
|
877
|
+
}
|
|
878
|
+
function authorReady(state) {
|
|
879
|
+
return state?.author_status === "ready" || state?.proof_plan_status === "ready";
|
|
880
|
+
}
|
|
881
|
+
function implementationReady(state) {
|
|
882
|
+
return ["changes_detected", "completed"].includes(state?.implementation_status || "");
|
|
883
|
+
}
|
|
884
|
+
function stageAfterAuthor(state) {
|
|
885
|
+
return implementationReady(state) ? "verify" : "implement";
|
|
886
|
+
}
|
|
887
|
+
function latestReconAttempt(state) {
|
|
888
|
+
const history = Array.isArray(state?.recon_results?.attempt_history) ? state.recon_results.attempt_history : [];
|
|
889
|
+
return history.length ? history[history.length - 1] : null;
|
|
890
|
+
}
|
|
891
|
+
function latestReconCapturedBaselines(state) {
|
|
892
|
+
const latest = latestReconAttempt(state);
|
|
893
|
+
if (latest?.captured_baselines && typeof latest.captured_baselines === "object") return latest.captured_baselines;
|
|
894
|
+
if (latest?.baselines && typeof latest.baselines === "object") return latest.baselines;
|
|
895
|
+
return state?.recon_results?.baselines || {};
|
|
896
|
+
}
|
|
897
|
+
function requiredReconBaselineLabels(state) {
|
|
898
|
+
const labels = [];
|
|
899
|
+
const reference = state?.requested_reference || state?.reference || "before";
|
|
900
|
+
if (reference === "before" || reference === "both") labels.push("before");
|
|
901
|
+
if ((reference === "prod" || reference === "both") && String(state?.prod_url || "").trim()) labels.push("prod");
|
|
902
|
+
return labels;
|
|
903
|
+
}
|
|
904
|
+
function latestReconHasRequiredBaselines(state) {
|
|
905
|
+
const baselines = latestReconCapturedBaselines(state);
|
|
906
|
+
return requiredReconBaselineLabels(state).every((label) => Boolean((baselines?.[label]?.url || "").trim()));
|
|
907
|
+
}
|
|
908
|
+
function hasReconBaselineUnderstanding(state) {
|
|
909
|
+
const understanding = state?.recon_assessment?.baseline_understanding || state?.recon_baseline_understanding || {};
|
|
910
|
+
return Boolean(
|
|
911
|
+
String(understanding?.visible_before_state || "").trim() && String(understanding?.requested_change || "").trim() && String(understanding?.proof_focus || "").trim() && String(understanding?.stop_condition || "").trim()
|
|
912
|
+
);
|
|
913
|
+
}
|
|
914
|
+
function promoteLatestReconBaselines(state) {
|
|
915
|
+
const baselines = latestReconCapturedBaselines(state);
|
|
916
|
+
state.recon_results = state.recon_results || {};
|
|
917
|
+
state.recon_results.baselines = baselines;
|
|
918
|
+
state.recon_results.selected_attempt = latestReconAttempt(state) || {};
|
|
919
|
+
state.before_cdn = (baselines?.before?.url || "").trim();
|
|
920
|
+
state.prod_cdn = (baselines?.prod?.url || "").trim();
|
|
921
|
+
return baselines;
|
|
922
|
+
}
|
|
923
|
+
function hasSupervisorReconAssessment(state) {
|
|
924
|
+
const reconAssessment2 = state?.recon_assessment || {};
|
|
925
|
+
const source = String(reconAssessment2?.source || state?.recon_assessment_source || "").trim().toLowerCase();
|
|
926
|
+
if (!reconAssessment2?.decision) return false;
|
|
927
|
+
return source === "supervising_agent" || source === "supervisor";
|
|
928
|
+
}
|
|
929
|
+
function reconAssessment(state) {
|
|
930
|
+
const assessment = state?.recon_assessment || {};
|
|
931
|
+
const decision = assessment?.decision || null;
|
|
932
|
+
const continueWithStage = assessment?.continue_with_stage || assessment?.recommended_stage || (decision === "ready_for_author" ? "author" : "recon");
|
|
933
|
+
return {
|
|
934
|
+
decision,
|
|
935
|
+
summary: assessment?.summary || state?.recon_assessment_request?.summary || state?.recon_summary || null,
|
|
936
|
+
recommendedStage: assessment?.recommended_stage || continueWithStage || null,
|
|
937
|
+
continueWithStage: continueWithStage || null,
|
|
938
|
+
escalationTarget: assessment?.escalation_target || "agent",
|
|
939
|
+
reasons: Array.isArray(assessment?.reasons) ? assessment.reasons : [],
|
|
940
|
+
raw: assessment,
|
|
941
|
+
source: String(assessment?.source || state?.recon_assessment_source || "").trim() || null
|
|
942
|
+
};
|
|
943
|
+
}
|
|
944
|
+
function updateState(statePath, mutate) {
|
|
945
|
+
const state = readState(statePath) || {};
|
|
946
|
+
mutate(state);
|
|
947
|
+
writeState(statePath, state);
|
|
948
|
+
return state;
|
|
949
|
+
}
|
|
950
|
+
function hasSupervisorProofAssessment(state) {
|
|
951
|
+
const proofAssessment = state?.proof_assessment || {};
|
|
952
|
+
const source = String(proofAssessment?.source || state?.proof_assessment_source || "").trim().toLowerCase();
|
|
953
|
+
if (!proofAssessment?.decision) return false;
|
|
954
|
+
return source === "supervising_agent" || source === "supervisor";
|
|
955
|
+
}
|
|
956
|
+
function verifyAssessment(state) {
|
|
957
|
+
const proofAssessment = state?.proof_assessment || {};
|
|
958
|
+
const verifyDecision = state?.verify_decision_request || {};
|
|
959
|
+
if (hasSupervisorProofAssessment(state)) {
|
|
960
|
+
return {
|
|
961
|
+
decision: proofAssessment?.decision || null,
|
|
962
|
+
summary: proofAssessment?.summary || verifyDecision?.summary || null,
|
|
963
|
+
recommendedStage: proofAssessment?.continue_with_stage || proofAssessment?.recommended_stage || verifyDecision?.continue_with_stage || verifyDecision?.recommended_stage || null,
|
|
964
|
+
continueWithStage: proofAssessment?.continue_with_stage || verifyDecision?.continue_with_stage || proofAssessment?.recommended_stage || verifyDecision?.recommended_stage || null,
|
|
965
|
+
escalationTarget: proofAssessment?.escalation_target || "agent",
|
|
966
|
+
reasons: Array.isArray(proofAssessment?.reasons) ? proofAssessment.reasons : [],
|
|
967
|
+
raw: proofAssessment,
|
|
968
|
+
source: "supervising_agent"
|
|
969
|
+
};
|
|
970
|
+
}
|
|
971
|
+
if (state?.verify_status === "capture_incomplete") {
|
|
972
|
+
return {
|
|
973
|
+
decision: verifyDecision?.capture_quality?.decision || "revise_capture",
|
|
974
|
+
summary: verifyDecision?.summary || "Verify needs another internal capture iteration before the evidence can be judged.",
|
|
975
|
+
recommendedStage: verifyDecision?.continue_with_stage || verifyDecision?.recommended_stage || "author",
|
|
976
|
+
continueWithStage: verifyDecision?.continue_with_stage || verifyDecision?.recommended_stage || "author",
|
|
977
|
+
escalationTarget: "agent",
|
|
978
|
+
reasons: Array.isArray(verifyDecision?.capture_quality?.reasons) ? verifyDecision.capture_quality.reasons : [],
|
|
979
|
+
raw: verifyDecision?.capture_quality || verifyDecision,
|
|
980
|
+
source: "workflow_capture"
|
|
981
|
+
};
|
|
982
|
+
}
|
|
983
|
+
return {
|
|
984
|
+
decision: null,
|
|
985
|
+
summary: verifyDecision?.summary || "Verify captured evidence and is waiting for supervising-agent proof assessment.",
|
|
986
|
+
recommendedStage: null,
|
|
987
|
+
continueWithStage: null,
|
|
988
|
+
escalationTarget: "agent",
|
|
989
|
+
reasons: [],
|
|
990
|
+
raw: proofAssessment,
|
|
991
|
+
source: "awaiting_supervisor"
|
|
992
|
+
};
|
|
993
|
+
}
|
|
994
|
+
function nonConvergenceSignals(state, assessment = verifyAssessment(state)) {
|
|
995
|
+
const verifyAttempts = Number(state?.stage_attempts?.verify?.count || 0);
|
|
996
|
+
const authorAttempts = Number(state?.stage_attempts?.author?.count || 0);
|
|
997
|
+
const reconAttempts = Number(state?.stage_attempts?.recon?.count || 0);
|
|
998
|
+
const continueStage = assessment.continueWithStage || assessment.recommendedStage || null;
|
|
999
|
+
return {
|
|
1000
|
+
verifyAttempts,
|
|
1001
|
+
authorAttempts,
|
|
1002
|
+
reconAttempts,
|
|
1003
|
+
continueStage,
|
|
1004
|
+
warning: verifyAttempts >= 4 || continueStage === "author" && verifyAttempts >= 2 && authorAttempts >= 2 || continueStage === "recon" && verifyAttempts >= 2 && reconAttempts >= 2 || continueStage === "implement" && verifyAttempts >= 2
|
|
1005
|
+
};
|
|
1006
|
+
}
|
|
1007
|
+
function shouldEscalateVerifyToHuman(_state, assessment = verifyAssessment(_state)) {
|
|
1008
|
+
return assessment.escalationTarget === "human";
|
|
1009
|
+
}
|
|
1010
|
+
function recommendedAdvanceStage(state) {
|
|
1011
|
+
if (!state?.workspace_ready) return "setup";
|
|
1012
|
+
if (!state?.recon_results || ["needs_agent_decision", "needs_supervisor_judgment"].includes(state?.recon_status || "")) return "recon";
|
|
1013
|
+
if (!authorReady(state)) return "author";
|
|
1014
|
+
if (!implementationReady(state)) return "implement";
|
|
1015
|
+
if (state?.verify_status === "capture_incomplete") return verifyAssessment(state).continueWithStage || verifyAssessment(state).recommendedStage || "author";
|
|
1016
|
+
if (state?.verify_status === "evidence_captured") return verifyAssessment(state).continueWithStage || verifyAssessment(state).recommendedStage;
|
|
1017
|
+
if (!(state?.after_cdn || "").trim()) return "verify";
|
|
1018
|
+
return null;
|
|
1019
|
+
}
|
|
1020
|
+
function normalizeStageRequest(state, requestedAdvanceStage) {
|
|
1021
|
+
if (requestedAdvanceStage) return requestedAdvanceStage;
|
|
1022
|
+
if (!state?.workspace_ready) return null;
|
|
1023
|
+
if (!state?.recon_results || ["needs_agent_decision", "needs_supervisor_judgment"].includes(state?.recon_status || "")) return "recon";
|
|
1024
|
+
return null;
|
|
1025
|
+
}
|
|
1026
|
+
function stringValue(value) {
|
|
1027
|
+
return typeof value === "string" && value.trim() ? value.trim() : "";
|
|
1028
|
+
}
|
|
1029
|
+
function commandResult(command, args, cwd, timeout = 6e4) {
|
|
1030
|
+
try {
|
|
1031
|
+
return {
|
|
1032
|
+
ok: true,
|
|
1033
|
+
stdout: (0, import_node_child_process.execFileSync)(command, args, { cwd, encoding: "utf-8", timeout, stdio: ["ignore", "pipe", "pipe"] }),
|
|
1034
|
+
stderr: ""
|
|
1035
|
+
};
|
|
1036
|
+
} catch (error) {
|
|
1037
|
+
return {
|
|
1038
|
+
ok: false,
|
|
1039
|
+
stdout: String(error?.stdout || ""),
|
|
1040
|
+
stderr: String(error?.stderr || error?.message || "")
|
|
1041
|
+
};
|
|
1042
|
+
}
|
|
1043
|
+
}
|
|
1044
|
+
function repoDirForSync(state) {
|
|
1045
|
+
const candidates = [
|
|
1046
|
+
state?.repo_dir,
|
|
1047
|
+
state?.after_worktree,
|
|
1048
|
+
state?.before_worktree
|
|
1049
|
+
].map(stringValue).filter(Boolean);
|
|
1050
|
+
return candidates.find((candidate) => (0, import_node_fs2.existsSync)(import_node_path2.default.join(candidate, ".git"))) || "";
|
|
1051
|
+
}
|
|
1052
|
+
function parseWorktreeList(output) {
|
|
1053
|
+
const entries = [];
|
|
1054
|
+
let current = {};
|
|
1055
|
+
for (const line of output.split(/\r?\n/)) {
|
|
1056
|
+
if (!line.trim()) {
|
|
1057
|
+
if (current.worktree) entries.push(current);
|
|
1058
|
+
current = {};
|
|
1059
|
+
continue;
|
|
1060
|
+
}
|
|
1061
|
+
const [key, ...rest] = line.split(" ");
|
|
1062
|
+
const value = rest.join(" ").trim();
|
|
1063
|
+
if (key === "worktree" || key === "HEAD" || key === "branch" || key === "detached") current[key] = value;
|
|
1064
|
+
}
|
|
1065
|
+
if (current.worktree) entries.push(current);
|
|
1066
|
+
return entries;
|
|
1067
|
+
}
|
|
1068
|
+
function gitStdout(cwd, args, timeout = 6e4) {
|
|
1069
|
+
const result = commandResult("git", args, cwd, timeout);
|
|
1070
|
+
return result.ok ? result.stdout.trim() : "";
|
|
1071
|
+
}
|
|
1072
|
+
function shortBranch(ref) {
|
|
1073
|
+
return ref.startsWith("refs/heads/") ? ref.slice("refs/heads/".length) : ref;
|
|
1074
|
+
}
|
|
1075
|
+
function safeInteger(value) {
|
|
1076
|
+
const parsed = Number.parseInt(value, 10);
|
|
1077
|
+
return Number.isFinite(parsed) ? parsed : null;
|
|
1078
|
+
}
|
|
1079
|
+
function baseCheckoutReport(repoDir, baseBranch, updateRequested, updateAllowed) {
|
|
1080
|
+
const remoteRef = `origin/${baseBranch}`;
|
|
1081
|
+
const report = {
|
|
1082
|
+
requested: updateRequested,
|
|
1083
|
+
repo_dir: repoDir,
|
|
1084
|
+
base_branch: baseBranch,
|
|
1085
|
+
remote_ref: remoteRef,
|
|
1086
|
+
updated: false
|
|
1087
|
+
};
|
|
1088
|
+
const listed = commandResult("git", ["worktree", "list", "--porcelain"], repoDir, 6e4);
|
|
1089
|
+
if (!listed.ok) {
|
|
1090
|
+
report.update_skipped = "worktree_list_failed";
|
|
1091
|
+
report.error = listed.stderr.slice(0, 300);
|
|
1092
|
+
return report;
|
|
1093
|
+
}
|
|
1094
|
+
const worktrees = parseWorktreeList(listed.stdout);
|
|
1095
|
+
const baseRef = `refs/heads/${baseBranch}`;
|
|
1096
|
+
const selected = worktrees.find((entry) => entry.branch === baseRef) || worktrees.find((entry) => import_node_path2.default.resolve(entry.worktree || "") === import_node_path2.default.resolve(repoDir) && shortBranch(entry.branch || "") === baseBranch);
|
|
1097
|
+
if (!selected?.worktree) {
|
|
1098
|
+
report.worktrees_seen = worktrees.map((entry) => ({
|
|
1099
|
+
path: entry.worktree || null,
|
|
1100
|
+
branch: entry.branch ? shortBranch(entry.branch) : null,
|
|
1101
|
+
detached: Boolean(entry.detached)
|
|
1102
|
+
}));
|
|
1103
|
+
report.update_skipped = "base_worktree_not_found";
|
|
1104
|
+
return report;
|
|
1105
|
+
}
|
|
1106
|
+
const baseDir = selected.worktree;
|
|
1107
|
+
const branch = shortBranch(selected.branch || "");
|
|
1108
|
+
const status = commandResult("git", ["status", "--porcelain"], baseDir, 6e4);
|
|
1109
|
+
const clean = status.ok && !status.stdout.trim();
|
|
1110
|
+
const localHead = gitStdout(baseDir, ["rev-parse", "HEAD"]);
|
|
1111
|
+
const remoteHead = gitStdout(baseDir, ["rev-parse", "--verify", remoteRef]);
|
|
1112
|
+
const counts = remoteHead ? gitStdout(baseDir, ["rev-list", "--left-right", "--count", `HEAD...${remoteRef}`]) : "";
|
|
1113
|
+
const [aheadRaw, behindRaw] = counts.split(/\s+/);
|
|
1114
|
+
Object.assign(report, {
|
|
1115
|
+
base_worktree: baseDir,
|
|
1116
|
+
branch: branch || null,
|
|
1117
|
+
clean,
|
|
1118
|
+
local_head: localHead || null,
|
|
1119
|
+
remote_head: remoteHead || null,
|
|
1120
|
+
ahead: safeInteger(aheadRaw || ""),
|
|
1121
|
+
behind: safeInteger(behindRaw || "")
|
|
1122
|
+
});
|
|
1123
|
+
if (!updateRequested) {
|
|
1124
|
+
report.update_skipped = "update_not_requested";
|
|
1125
|
+
return report;
|
|
1126
|
+
}
|
|
1127
|
+
if (!updateAllowed) {
|
|
1128
|
+
report.update_skipped = "fetch_failed";
|
|
1129
|
+
return report;
|
|
1130
|
+
}
|
|
1131
|
+
if (branch !== baseBranch) {
|
|
1132
|
+
report.update_skipped = "base_worktree_not_on_base_branch";
|
|
1133
|
+
return report;
|
|
1134
|
+
}
|
|
1135
|
+
if (!status.ok) {
|
|
1136
|
+
report.update_skipped = "status_failed";
|
|
1137
|
+
report.status_error = status.stderr.slice(0, 300);
|
|
1138
|
+
return report;
|
|
1139
|
+
}
|
|
1140
|
+
if (!clean) {
|
|
1141
|
+
report.update_skipped = "base_worktree_dirty";
|
|
1142
|
+
return report;
|
|
1143
|
+
}
|
|
1144
|
+
if (!remoteHead) {
|
|
1145
|
+
report.update_skipped = "remote_ref_missing";
|
|
1146
|
+
return report;
|
|
1147
|
+
}
|
|
1148
|
+
if (localHead && localHead === remoteHead) {
|
|
1149
|
+
report.update_skipped = "already_current";
|
|
1150
|
+
return report;
|
|
1151
|
+
}
|
|
1152
|
+
const merge = commandResult("git", ["merge", "--ff-only", remoteRef], baseDir, 12e4);
|
|
1153
|
+
if (!merge.ok) {
|
|
1154
|
+
report.update_skipped = "fast_forward_failed";
|
|
1155
|
+
report.update_error = merge.stderr.slice(0, 500);
|
|
1156
|
+
return report;
|
|
1157
|
+
}
|
|
1158
|
+
const updatedHead = gitStdout(baseDir, ["rev-parse", "HEAD"]);
|
|
1159
|
+
const updatedCounts = gitStdout(baseDir, ["rev-list", "--left-right", "--count", `HEAD...${remoteRef}`]);
|
|
1160
|
+
const [updatedAheadRaw, updatedBehindRaw] = updatedCounts.split(/\s+/);
|
|
1161
|
+
report.updated = true;
|
|
1162
|
+
report.local_head = updatedHead || report.local_head;
|
|
1163
|
+
report.ahead = safeInteger(updatedAheadRaw || "");
|
|
1164
|
+
report.behind = safeInteger(updatedBehindRaw || "");
|
|
1165
|
+
report.update_summary = merge.stdout.trim().slice(0, 500);
|
|
1166
|
+
return report;
|
|
1167
|
+
}
|
|
1168
|
+
function normalizeGhPrStatus(value) {
|
|
1169
|
+
const status = stringValue(value).toLowerCase();
|
|
1170
|
+
if (status === "merged") return "merged";
|
|
1171
|
+
if (status === "open") return "open";
|
|
1172
|
+
if (status === "closed") return "closed";
|
|
1173
|
+
return status || "unknown";
|
|
1174
|
+
}
|
|
1175
|
+
function prRefFromState(state) {
|
|
1176
|
+
return stringValue(state?.pr_number) || stringValue(state?.pr_url);
|
|
1177
|
+
}
|
|
1178
|
+
function prNumberFromUrl(url) {
|
|
1179
|
+
const match = url.match(/\/pull\/(\d+)(?:$|[?#])/);
|
|
1180
|
+
return match?.[1] || "";
|
|
1181
|
+
}
|
|
1182
|
+
function normalizePrState(raw, state, checkedAt = (/* @__PURE__ */ new Date()).toISOString()) {
|
|
1183
|
+
const mergeCommit = typeof raw?.mergeCommit === "object" && raw.mergeCommit ? stringValue(raw.mergeCommit.oid) : stringValue(raw?.mergeCommit);
|
|
1184
|
+
const url = stringValue(raw?.url) || stringValue(state?.pr_url);
|
|
1185
|
+
return {
|
|
1186
|
+
status: normalizeGhPrStatus(raw?.state),
|
|
1187
|
+
pr_url: url || null,
|
|
1188
|
+
pr_number: String(raw?.number || state?.pr_number || prNumberFromUrl(url) || ""),
|
|
1189
|
+
repo: stringValue(state?.repo) || null,
|
|
1190
|
+
head_branch: stringValue(raw?.headRefName) || stringValue(state?.target_branch) || stringValue(state?.branch) || null,
|
|
1191
|
+
base_branch: stringValue(raw?.baseRefName) || stringValue(state?.base_branch) || "main",
|
|
1192
|
+
merge_commit: mergeCommit || null,
|
|
1193
|
+
merged_at: stringValue(raw?.mergedAt) || null,
|
|
1194
|
+
closed_at: stringValue(raw?.closedAt) || null,
|
|
1195
|
+
checked_at: checkedAt,
|
|
1196
|
+
source: "gh"
|
|
1197
|
+
};
|
|
1198
|
+
}
|
|
1199
|
+
function cleanupMergedProofRun(state, repoDir, params, prState) {
|
|
1200
|
+
const cleanup = {
|
|
1201
|
+
requested: params.cleanup_merged_pr !== false,
|
|
1202
|
+
fetch_base: params.fetch_base !== false,
|
|
1203
|
+
update_base_checkout: params.update_base_checkout !== false,
|
|
1204
|
+
repo_dir: repoDir,
|
|
1205
|
+
worktrees_removed: [],
|
|
1206
|
+
worktree_remove_errors: [],
|
|
1207
|
+
branches_deleted: [],
|
|
1208
|
+
branch_delete_errors: [],
|
|
1209
|
+
pruned: false
|
|
1210
|
+
};
|
|
1211
|
+
const baseBranch = stringValue(prState.base_branch) || stringValue(state?.base_branch) || "main";
|
|
1212
|
+
let fetchedBase = params.fetch_base === false;
|
|
1213
|
+
if (params.fetch_base !== false && baseBranch) {
|
|
1214
|
+
const fetch = commandResult("git", ["fetch", "origin", baseBranch], repoDir, 12e4);
|
|
1215
|
+
cleanup.fetch = fetch.ok ? { ok: true, base_branch: baseBranch } : { ok: false, base_branch: baseBranch, error: fetch.stderr.slice(0, 300) };
|
|
1216
|
+
if (fetch.ok) {
|
|
1217
|
+
fetchedBase = true;
|
|
1218
|
+
state.base_synced_at = (/* @__PURE__ */ new Date()).toISOString();
|
|
1219
|
+
state.base_branch = baseBranch;
|
|
1220
|
+
}
|
|
1221
|
+
}
|
|
1222
|
+
cleanup.base_checkout = baseCheckoutReport(repoDir, baseBranch, params.update_base_checkout !== false, fetchedBase);
|
|
1223
|
+
if (params.cleanup_merged_pr === false) {
|
|
1224
|
+
cleanup.skipped = "cleanup_disabled";
|
|
1225
|
+
return cleanup;
|
|
1226
|
+
}
|
|
1227
|
+
const removed = [];
|
|
1228
|
+
const removeErrors = [];
|
|
1229
|
+
for (const candidate of [state?.before_worktree, state?.after_worktree].map(stringValue).filter(Boolean)) {
|
|
1230
|
+
if (!(0, import_node_fs2.existsSync)(candidate) || import_node_path2.default.resolve(candidate) === import_node_path2.default.resolve(repoDir)) continue;
|
|
1231
|
+
const remove = commandResult("git", ["worktree", "remove", "--force", candidate], repoDir, 12e4);
|
|
1232
|
+
if (remove.ok) {
|
|
1233
|
+
removed.push(candidate);
|
|
1234
|
+
} else {
|
|
1235
|
+
removeErrors.push({ path: candidate, error: remove.stderr.slice(0, 300) });
|
|
1236
|
+
}
|
|
1237
|
+
}
|
|
1238
|
+
cleanup.worktrees_removed = removed;
|
|
1239
|
+
cleanup.worktree_remove_errors = removeErrors;
|
|
1240
|
+
const afterBranch = stringValue(state?.after_worktree_branch);
|
|
1241
|
+
if (afterBranch.startsWith("riddle-proof/")) {
|
|
1242
|
+
const deleted = commandResult("git", ["branch", "-D", afterBranch], repoDir, 6e4);
|
|
1243
|
+
if (deleted.ok) {
|
|
1244
|
+
cleanup.branches_deleted = [afterBranch];
|
|
1245
|
+
} else {
|
|
1246
|
+
cleanup.branch_delete_errors = [{ branch: afterBranch, error: deleted.stderr.slice(0, 300) }];
|
|
1247
|
+
}
|
|
1248
|
+
}
|
|
1249
|
+
const prune = commandResult("git", ["worktree", "prune"], repoDir, 6e4);
|
|
1250
|
+
cleanup.pruned = prune.ok;
|
|
1251
|
+
if (!prune.ok) cleanup.prune_error = prune.stderr.slice(0, 300);
|
|
1252
|
+
return cleanup;
|
|
1253
|
+
}
|
|
1254
|
+
function syncPrLifecycle(statePath, params) {
|
|
1255
|
+
const state = readState(statePath);
|
|
1256
|
+
if (!state) {
|
|
1257
|
+
return {
|
|
1258
|
+
ok: false,
|
|
1259
|
+
action: "sync",
|
|
1260
|
+
state_path: statePath,
|
|
1261
|
+
checkpoint: "pr_sync_not_found",
|
|
1262
|
+
summary: "No readable Riddle Proof state exists at state_path.",
|
|
1263
|
+
state: null,
|
|
1264
|
+
nextAction: "Check the wrapper state_path or run riddle_proof_status first."
|
|
1265
|
+
};
|
|
1266
|
+
}
|
|
1267
|
+
const repoDir = repoDirForSync(state);
|
|
1268
|
+
const prRef = prRefFromState(state);
|
|
1269
|
+
if (!repoDir || !prRef) {
|
|
1270
|
+
const missingPr = !prRef;
|
|
1271
|
+
const orphanSummary = "Riddle Proof state exists, but this run is not recoverable through PR sync because no PR URL or PR number was linked before it stopped.";
|
|
1272
|
+
const orphanNextAction = "Treat this as an orphaned proof run: update the base checkout directly if needed, then clean stale proof worktrees or riddle-proof/* branches outside normal PR sync.";
|
|
1273
|
+
const prState2 = {
|
|
1274
|
+
status: missingPr ? "orphaned" : "unavailable",
|
|
1275
|
+
pr_url: state.pr_url || null,
|
|
1276
|
+
pr_number: String(state.pr_number || prNumberFromUrl(stringValue(state.pr_url)) || ""),
|
|
1277
|
+
repo: state.repo || null,
|
|
1278
|
+
head_branch: state.target_branch || state.branch || null,
|
|
1279
|
+
base_branch: state.base_branch || "main",
|
|
1280
|
+
checked_at: (/* @__PURE__ */ new Date()).toISOString(),
|
|
1281
|
+
source: repoDir ? "state" : "local_state",
|
|
1282
|
+
sync_recoverable: !missingPr,
|
|
1283
|
+
sync_blocker: missingPr ? "missing_pr_linkage" : "missing_local_repo",
|
|
1284
|
+
next_action: missingPr ? orphanNextAction : "State has a PR but no readable local git repo/worktree; restore repo access and rerun sync."
|
|
1285
|
+
};
|
|
1286
|
+
state.pr_state = prState2;
|
|
1287
|
+
if (missingPr) state.pr_sync_summary = orphanSummary;
|
|
1288
|
+
writeState(statePath, state);
|
|
1289
|
+
return {
|
|
1290
|
+
ok: false,
|
|
1291
|
+
action: "sync",
|
|
1292
|
+
state_path: statePath,
|
|
1293
|
+
checkpoint: missingPr ? "pr_sync_no_pr" : "pr_sync_unavailable",
|
|
1294
|
+
summary: missingPr ? orphanSummary : prState2.next_action,
|
|
1295
|
+
state: summarizeState(state).state,
|
|
1296
|
+
pr_state: prState2,
|
|
1297
|
+
nextAction: prState2.next_action
|
|
1298
|
+
};
|
|
1299
|
+
}
|
|
1300
|
+
const viewed = commandResult("gh", ["pr", "view", prRef, "--json", "state,mergedAt,closedAt,mergeCommit,headRefName,baseRefName,url,number"], repoDir, 6e4);
|
|
1301
|
+
if (!viewed.ok) {
|
|
1302
|
+
const prState2 = {
|
|
1303
|
+
status: "unavailable",
|
|
1304
|
+
pr_url: state.pr_url || null,
|
|
1305
|
+
pr_number: String(state.pr_number || prNumberFromUrl(stringValue(state.pr_url)) || ""),
|
|
1306
|
+
repo: state.repo || null,
|
|
1307
|
+
head_branch: state.target_branch || state.branch || null,
|
|
1308
|
+
base_branch: state.base_branch || "main",
|
|
1309
|
+
checked_at: (/* @__PURE__ */ new Date()).toISOString(),
|
|
1310
|
+
source: "gh",
|
|
1311
|
+
next_action: "GitHub PR state is unavailable; fix gh auth/repo access and rerun riddle_proof_sync."
|
|
1312
|
+
};
|
|
1313
|
+
state.pr_state = prState2;
|
|
1314
|
+
state.cleanup_report = { requested: params.cleanup_merged_pr !== false, skipped: "pr_state_unavailable", error: viewed.stderr.slice(0, 300) };
|
|
1315
|
+
writeState(statePath, state);
|
|
1316
|
+
return {
|
|
1317
|
+
ok: false,
|
|
1318
|
+
action: "sync",
|
|
1319
|
+
state_path: statePath,
|
|
1320
|
+
checkpoint: "pr_sync_unavailable",
|
|
1321
|
+
summary: prState2.next_action,
|
|
1322
|
+
state: summarizeState(state).state,
|
|
1323
|
+
pr_state: prState2,
|
|
1324
|
+
cleanup: state.cleanup_report,
|
|
1325
|
+
nextAction: prState2.next_action
|
|
1326
|
+
};
|
|
1327
|
+
}
|
|
1328
|
+
let rawPr;
|
|
1329
|
+
try {
|
|
1330
|
+
rawPr = JSON.parse(viewed.stdout);
|
|
1331
|
+
} catch {
|
|
1332
|
+
rawPr = {};
|
|
1333
|
+
}
|
|
1334
|
+
const prState = normalizePrState(rawPr, state);
|
|
1335
|
+
let cleanup = null;
|
|
1336
|
+
let checkpoint = "pr_sync_open";
|
|
1337
|
+
let ok = true;
|
|
1338
|
+
let summary = "PR is still open; no merge cleanup was performed.";
|
|
1339
|
+
let nextAction = "Wait for the PR to merge, then rerun riddle_proof_sync.";
|
|
1340
|
+
if (prState.status === "merged") {
|
|
1341
|
+
cleanup = cleanupMergedProofRun(state, repoDir, params, prState);
|
|
1342
|
+
prState.cleanup = cleanup;
|
|
1343
|
+
prState.next_action = "The PR is merged; sync recorded proof cleanup and the local base checkout refresh status.";
|
|
1344
|
+
state.finalized = true;
|
|
1345
|
+
state.merge_commit = prState.merge_commit || state.merge_commit || "";
|
|
1346
|
+
state.merged_at = prState.merged_at || state.merged_at || "";
|
|
1347
|
+
state.cleanup_report = cleanup;
|
|
1348
|
+
checkpoint = "pr_sync_merged";
|
|
1349
|
+
summary = "PR is merged and Riddle Proof state has been reconciled.";
|
|
1350
|
+
nextAction = "Start the next proof run from the recorded base checkout; inspect cleanup.base_checkout only if it reports a skipped or failed fast-forward.";
|
|
1351
|
+
} else if (prState.status === "closed") {
|
|
1352
|
+
prState.next_action = "The PR is closed without a merge; inspect the PR before reusing or deleting the branch.";
|
|
1353
|
+
checkpoint = "pr_sync_closed";
|
|
1354
|
+
summary = "PR is closed without a merge; no merge cleanup was performed.";
|
|
1355
|
+
nextAction = prState.next_action;
|
|
1356
|
+
} else if (prState.status !== "open") {
|
|
1357
|
+
ok = false;
|
|
1358
|
+
prState.next_action = "PR state was not recognized; inspect gh pr view output and rerun sync.";
|
|
1359
|
+
checkpoint = "pr_sync_unavailable";
|
|
1360
|
+
summary = prState.next_action;
|
|
1361
|
+
nextAction = prState.next_action;
|
|
1362
|
+
}
|
|
1363
|
+
state.pr_state = prState;
|
|
1364
|
+
state.pr_url = prState.pr_url || state.pr_url;
|
|
1365
|
+
state.pr_number = prState.pr_number || state.pr_number;
|
|
1366
|
+
state.target_branch = prState.head_branch || state.target_branch || state.branch;
|
|
1367
|
+
state.branch = prState.head_branch || state.branch;
|
|
1368
|
+
state.base_branch = prState.base_branch || state.base_branch;
|
|
1369
|
+
writeState(statePath, state);
|
|
1370
|
+
const snapshot = summarizeState(state);
|
|
1371
|
+
return {
|
|
1372
|
+
ok,
|
|
1373
|
+
action: "sync",
|
|
1374
|
+
state_path: statePath,
|
|
1375
|
+
checkpoint,
|
|
1376
|
+
summary,
|
|
1377
|
+
state: snapshot.state,
|
|
1378
|
+
pr_state: prState,
|
|
1379
|
+
cleanup,
|
|
1380
|
+
nextAction
|
|
1381
|
+
};
|
|
1382
|
+
}
|
|
1383
|
+
async function executeWorkflow(params, pluginConfig, resolvedConfig) {
|
|
1384
|
+
const config = resolvedConfig || resolveConfig(pluginConfig, params);
|
|
1385
|
+
const action = ensureAction(params.action);
|
|
1386
|
+
if (!(0, import_node_fs2.existsSync)(config.riddleProofDir)) {
|
|
1387
|
+
throw new Error(`riddle-proof runtime directory not found: ${config.riddleProofDir}`);
|
|
1388
|
+
}
|
|
1389
|
+
if (action === "status") {
|
|
1390
|
+
return {
|
|
1391
|
+
state_path: config.statePath,
|
|
1392
|
+
...summarizeState(readState(config.statePath))
|
|
1393
|
+
};
|
|
1394
|
+
}
|
|
1395
|
+
if (action === "sync") {
|
|
1396
|
+
return syncPrLifecycle(config.statePath, params);
|
|
1397
|
+
}
|
|
1398
|
+
const stateKey = import_node_path2.default.basename(config.statePath).replace(/[^A-Za-z0-9_.-]/g, "-");
|
|
1399
|
+
const lobsterStateDir = import_node_path2.default.join(import_node_path2.default.dirname(config.statePath), "riddle-proof-lobster-state", stateKey);
|
|
1400
|
+
(0, import_node_fs2.mkdirSync)(lobsterStateDir, { recursive: true });
|
|
1401
|
+
const env = {
|
|
1402
|
+
...process.env,
|
|
1403
|
+
RIDDLE_PROOF_DIR: config.riddleProofDir,
|
|
1404
|
+
RIDDLE_PROOF_STATE_FILE: config.statePath,
|
|
1405
|
+
RIDDLE_PROOF_ARGS_FILE: config.argsPath,
|
|
1406
|
+
LOBSTER_STATE_DIR: lobsterStateDir
|
|
1407
|
+
};
|
|
1408
|
+
const lobsterCommand = process.env.RIDDLE_PROOF_LOBSTER_COMMAND || "lobster";
|
|
1409
|
+
const lobsterPrefix = process.env.RIDDLE_PROOF_LOBSTER_SCRIPT ? [process.env.RIDDLE_PROOF_LOBSTER_SCRIPT] : [];
|
|
1410
|
+
const runOne = (step) => {
|
|
1411
|
+
const args = step === "setup" ? buildSetupArgs(params, config) : {};
|
|
1412
|
+
let output;
|
|
1413
|
+
try {
|
|
1414
|
+
output = JSON.parse(
|
|
1415
|
+
(0, import_node_child_process.execFileSync)(lobsterCommand, [...lobsterPrefix, "run", "--file", workflowFile(config.riddleProofDir, step), "--args-json", JSON.stringify(args)], {
|
|
1416
|
+
encoding: "utf-8",
|
|
1417
|
+
env
|
|
1418
|
+
})
|
|
1419
|
+
);
|
|
1420
|
+
} catch (error) {
|
|
1421
|
+
return {
|
|
1422
|
+
ok: false,
|
|
1423
|
+
step,
|
|
1424
|
+
error: error?.message || String(error),
|
|
1425
|
+
stdout: String(error?.stdout || ""),
|
|
1426
|
+
stderr: String(error?.stderr || "")
|
|
1427
|
+
};
|
|
1428
|
+
}
|
|
1429
|
+
if (output?.status === "needs_approval") {
|
|
1430
|
+
if (!params.auto_approve) {
|
|
1431
|
+
return {
|
|
1432
|
+
ok: false,
|
|
1433
|
+
haltedForApproval: true,
|
|
1434
|
+
step,
|
|
1435
|
+
approval: output.requiresApproval || null,
|
|
1436
|
+
raw: output
|
|
1437
|
+
};
|
|
1438
|
+
}
|
|
1439
|
+
const token = output?.requiresApproval?.resumeToken;
|
|
1440
|
+
if (!token) throw new Error(`${step} requested approval without a resume token.`);
|
|
1441
|
+
const resumed = JSON.parse(
|
|
1442
|
+
(0, import_node_child_process.execFileSync)(lobsterCommand, [...lobsterPrefix, "resume", "--token", token, "--approve", "yes"], {
|
|
1443
|
+
encoding: "utf-8",
|
|
1444
|
+
env
|
|
1445
|
+
})
|
|
1446
|
+
);
|
|
1447
|
+
return {
|
|
1448
|
+
ok: resumed?.ok !== false,
|
|
1449
|
+
step,
|
|
1450
|
+
autoApproved: true,
|
|
1451
|
+
raw: resumed
|
|
1452
|
+
};
|
|
1453
|
+
}
|
|
1454
|
+
return {
|
|
1455
|
+
ok: output?.ok !== false,
|
|
1456
|
+
step,
|
|
1457
|
+
raw: output
|
|
1458
|
+
};
|
|
1459
|
+
};
|
|
1460
|
+
let effectiveAdvanceStage = params.advance_stage || null;
|
|
1461
|
+
const recordAttempt = (stage, status, summary, extra = {}) => {
|
|
1462
|
+
updateState(config.statePath, (state) => {
|
|
1463
|
+
recordStageAttempt(state, stage, {
|
|
1464
|
+
status,
|
|
1465
|
+
summary,
|
|
1466
|
+
checkpoint: extra.checkpoint || null,
|
|
1467
|
+
requestedAdvanceStage: effectiveAdvanceStage || null,
|
|
1468
|
+
haltedForApproval: extra.haltedForApproval,
|
|
1469
|
+
autoApproved: extra.autoApproved,
|
|
1470
|
+
error: extra.error || null,
|
|
1471
|
+
details: extra.details || {}
|
|
1472
|
+
});
|
|
1473
|
+
});
|
|
1474
|
+
};
|
|
1475
|
+
const checkpoint = (stage, name, summary, extra = {}) => {
|
|
1476
|
+
const decision = updateState(config.statePath, (state) => {
|
|
1477
|
+
const checkpointContract = buildCheckpointContract(state, {
|
|
1478
|
+
statePath: config.statePath,
|
|
1479
|
+
stage,
|
|
1480
|
+
checkpoint: name,
|
|
1481
|
+
summary,
|
|
1482
|
+
nextActions: extra.nextActions,
|
|
1483
|
+
advanceOptions: extra.advanceOptions,
|
|
1484
|
+
recommendedAdvanceStage: extra.recommendedAdvanceStage,
|
|
1485
|
+
continueWithStage: extra.continueWithStage,
|
|
1486
|
+
blocking: extra.blocking
|
|
1487
|
+
});
|
|
1488
|
+
setStageDecisionRequest(state, {
|
|
1489
|
+
stage,
|
|
1490
|
+
checkpoint: name,
|
|
1491
|
+
summary,
|
|
1492
|
+
nextActions: extra.nextActions,
|
|
1493
|
+
advanceOptions: extra.advanceOptions,
|
|
1494
|
+
recommendedAdvanceStage: extra.recommendedAdvanceStage,
|
|
1495
|
+
continueWithStage: extra.continueWithStage,
|
|
1496
|
+
blocking: extra.blocking,
|
|
1497
|
+
details: extra.details,
|
|
1498
|
+
checkpointContract
|
|
1499
|
+
});
|
|
1500
|
+
}).stage_decision_request;
|
|
1501
|
+
const snapshot2 = snapshotFor(config.statePath);
|
|
1502
|
+
return {
|
|
1503
|
+
ok: extra.ok ?? true,
|
|
1504
|
+
action,
|
|
1505
|
+
state_path: config.statePath,
|
|
1506
|
+
stage: snapshot2.stage,
|
|
1507
|
+
checkpoint: name,
|
|
1508
|
+
summary,
|
|
1509
|
+
state: snapshot2.state,
|
|
1510
|
+
decisionRequest: decision,
|
|
1511
|
+
checkpointContract: decision?.checkpoint_contract || null,
|
|
1512
|
+
...extra
|
|
1513
|
+
};
|
|
1514
|
+
};
|
|
1515
|
+
const primaryShipGateNextAction = (shipGate) => {
|
|
1516
|
+
const reasons = shipGate.reasons || [];
|
|
1517
|
+
if (reasons.some((reason) => reason.includes("proof_assessment"))) {
|
|
1518
|
+
return "resume with riddle_proof_review using decision=ready_to_ship only after the screenshots and semantic evidence visibly prove the request; otherwise choose needs_implementation or needs_richer_proof";
|
|
1519
|
+
}
|
|
1520
|
+
if (reasons.some((reason) => reason.includes("after_cdn") || reason.includes("verify_status"))) {
|
|
1521
|
+
return "rerun verify with stronger proof framing so after evidence is captured before shipping";
|
|
1522
|
+
}
|
|
1523
|
+
if (reasons.some((reason) => reason.includes("before_cdn") || reason.includes("prod_cdn") || reason.includes("prod_url"))) {
|
|
1524
|
+
return "return to recon and capture the missing required baseline before shipping";
|
|
1525
|
+
}
|
|
1526
|
+
return "inspect the ship gate details, repair the missing invariant, then resume the run";
|
|
1527
|
+
};
|
|
1528
|
+
const shipGateBlocked = (state, executed, details = {}) => {
|
|
1529
|
+
const shipGate = validateShipGate(state);
|
|
1530
|
+
const nextAction = primaryShipGateNextAction(shipGate);
|
|
1531
|
+
return checkpoint(
|
|
1532
|
+
"verify",
|
|
1533
|
+
"ship_gate_blocked",
|
|
1534
|
+
`Ship is blocked until the proof bundle satisfies the hard ship gate. Next action: ${nextAction}.`,
|
|
1535
|
+
{
|
|
1536
|
+
ok: false,
|
|
1537
|
+
nextActions: ["inspect_ship_gate", "advance_run_to_verify", "supply_proof_assessment_json", "return_to_recon_if_baseline_is_missing"],
|
|
1538
|
+
advanceOptions: ["verify", "author", "implement", "recon"],
|
|
1539
|
+
recommendedAdvanceStage: "verify",
|
|
1540
|
+
continueWithStage: "verify",
|
|
1541
|
+
blocking: true,
|
|
1542
|
+
details: { ...details, shipGate, next_action: nextAction, executed },
|
|
1543
|
+
nextAction,
|
|
1544
|
+
shipGate,
|
|
1545
|
+
verifyStatus: state?.verify_status || null,
|
|
1546
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
1547
|
+
afterCdn: state?.after_cdn || null,
|
|
1548
|
+
proofAssessment: state?.proof_assessment || null,
|
|
1549
|
+
proofAssessmentRequest: state?.proof_assessment_request || null,
|
|
1550
|
+
executed
|
|
1551
|
+
}
|
|
1552
|
+
);
|
|
1553
|
+
};
|
|
1554
|
+
const failedRun = (stage, summary, res, extra = {}) => {
|
|
1555
|
+
recordAttempt(stage, res?.haltedForApproval ? "approval_required" : "failed", summary, {
|
|
1556
|
+
checkpoint: extra.checkpoint || null,
|
|
1557
|
+
haltedForApproval: res?.haltedForApproval || false,
|
|
1558
|
+
autoApproved: res?.autoApproved || false,
|
|
1559
|
+
error: res?.error || null,
|
|
1560
|
+
details: extra.details
|
|
1561
|
+
});
|
|
1562
|
+
const snapshot2 = snapshotFor(config.statePath);
|
|
1563
|
+
return {
|
|
1564
|
+
ok: false,
|
|
1565
|
+
action,
|
|
1566
|
+
state_path: config.statePath,
|
|
1567
|
+
stage: snapshot2.stage,
|
|
1568
|
+
summary,
|
|
1569
|
+
state: snapshot2.state,
|
|
1570
|
+
approval: res?.approval || null,
|
|
1571
|
+
error: res?.error || null,
|
|
1572
|
+
checkpoint: extra.checkpoint || null,
|
|
1573
|
+
...extra
|
|
1574
|
+
};
|
|
1575
|
+
};
|
|
1576
|
+
if (action !== "setup") {
|
|
1577
|
+
mergeStateFromParams(config.statePath, params);
|
|
1578
|
+
}
|
|
1579
|
+
if (action === "run") {
|
|
1580
|
+
const executed = [];
|
|
1581
|
+
let state = readState(config.statePath);
|
|
1582
|
+
if (!state || !state.workspace_ready || params.advance_stage === "setup") {
|
|
1583
|
+
const setupRes = runOne("setup");
|
|
1584
|
+
executed.push({ step: "setup", ok: setupRes.ok, haltedForApproval: setupRes.haltedForApproval || false, autoApproved: setupRes.autoApproved || false });
|
|
1585
|
+
if (!setupRes.ok || setupRes.haltedForApproval) {
|
|
1586
|
+
return failedRun("setup", setupRes.haltedForApproval ? "setup halted for approval" : "setup failed", setupRes, {
|
|
1587
|
+
checkpoint: "setup_blocked"
|
|
1588
|
+
});
|
|
1589
|
+
}
|
|
1590
|
+
recordAttempt("setup", "completed", "Setup completed and state/worktrees are ready.", {
|
|
1591
|
+
checkpoint: params.advance_stage === "setup" ? "setup_review" : null,
|
|
1592
|
+
autoApproved: setupRes.autoApproved || false
|
|
1593
|
+
});
|
|
1594
|
+
state = readState(config.statePath);
|
|
1595
|
+
if (params.advance_stage === "setup") {
|
|
1596
|
+
return checkpoint(
|
|
1597
|
+
"setup",
|
|
1598
|
+
"setup_review",
|
|
1599
|
+
"Setup completed. Inspect the prepared workspace and explicitly advance to recon when ready.",
|
|
1600
|
+
{
|
|
1601
|
+
nextActions: ["inspect_setup_state", "advance_run_to_recon"],
|
|
1602
|
+
advanceOptions: ["recon", "setup"],
|
|
1603
|
+
recommendedAdvanceStage: "recon",
|
|
1604
|
+
details: { executed },
|
|
1605
|
+
executed
|
|
1606
|
+
}
|
|
1607
|
+
);
|
|
1608
|
+
}
|
|
1609
|
+
}
|
|
1610
|
+
state = readState(config.statePath);
|
|
1611
|
+
const continuedStage = params.continue_from_checkpoint ? checkpointContinueStage(state) : null;
|
|
1612
|
+
if (params.continue_from_checkpoint && !continuedStage) {
|
|
1613
|
+
const recommended = recommendedAdvanceStage(state);
|
|
1614
|
+
return checkpoint(
|
|
1615
|
+
state?.active_checkpoint_stage || recommended || "recon",
|
|
1616
|
+
"continue_unavailable",
|
|
1617
|
+
"This run call asked to continue from a checkpoint, but the current state has no resumable checkpoint. Inspect status or set advance_stage explicitly.",
|
|
1618
|
+
{
|
|
1619
|
+
ok: false,
|
|
1620
|
+
nextActions: ["inspect_state", "set_advance_stage", "resume_run"],
|
|
1621
|
+
advanceOptions: ["recon", "author", "implement", "verify", "ship"],
|
|
1622
|
+
recommendedAdvanceStage: null,
|
|
1623
|
+
blocking: true,
|
|
1624
|
+
details: {
|
|
1625
|
+
executed,
|
|
1626
|
+
activeCheckpoint: state?.active_checkpoint || null,
|
|
1627
|
+
suggestedAdvanceStage: recommended || null
|
|
1628
|
+
},
|
|
1629
|
+
suggestedAdvanceStage: recommended || null,
|
|
1630
|
+
executed
|
|
1631
|
+
}
|
|
1632
|
+
);
|
|
1633
|
+
}
|
|
1634
|
+
effectiveAdvanceStage = params.advance_stage || continuedStage || null;
|
|
1635
|
+
if (effectiveAdvanceStage) {
|
|
1636
|
+
updateState(config.statePath, (state2) => {
|
|
1637
|
+
clearStageDecisionRequest(state2);
|
|
1638
|
+
state2.last_requested_advance_stage = effectiveAdvanceStage;
|
|
1639
|
+
});
|
|
1640
|
+
state = readState(config.statePath);
|
|
1641
|
+
}
|
|
1642
|
+
let requestedStage = normalizeStageRequest(state, effectiveAdvanceStage);
|
|
1643
|
+
const reconCheckpointActive = ["needs_agent_decision", "needs_supervisor_judgment"].includes(state?.recon_status || "") || state?.active_checkpoint === "recon_supervisor_judgment";
|
|
1644
|
+
if (requestedStage === "recon" && reconCheckpointActive) {
|
|
1645
|
+
const latestAttempt = latestReconAttempt(state);
|
|
1646
|
+
const latestCapturedBaselines = latestReconCapturedBaselines(state);
|
|
1647
|
+
const latestAssessment = reconAssessment(state);
|
|
1648
|
+
const reconAssessmentRequest = state?.recon_assessment_request || state?.recon_decision_request || null;
|
|
1649
|
+
const reconDetails = {
|
|
1650
|
+
executed,
|
|
1651
|
+
latestAttempt,
|
|
1652
|
+
latestCapturedBaselines,
|
|
1653
|
+
reconAssessmentRequest,
|
|
1654
|
+
reconAssessment: latestAssessment.raw
|
|
1655
|
+
};
|
|
1656
|
+
if (!hasSupervisorReconAssessment(state)) {
|
|
1657
|
+
return checkpoint(
|
|
1658
|
+
"recon",
|
|
1659
|
+
"recon_supervisor_judgment",
|
|
1660
|
+
"Recon gathered route hints, candidate paths, baseline captures, and observations. The supervising agent should now judge whether the latest baseline is trustworthy, whether recon should retry/reframe, and whether recon is done.",
|
|
1661
|
+
{
|
|
1662
|
+
nextActions: ["inspect_recon_packet", "supply_recon_assessment_json", "continue_internal_loop_with_checkpoint"],
|
|
1663
|
+
advanceOptions: ["recon", "author"],
|
|
1664
|
+
recommendedAdvanceStage: "recon",
|
|
1665
|
+
continueWithStage: "recon",
|
|
1666
|
+
blocking: false,
|
|
1667
|
+
details: reconDetails,
|
|
1668
|
+
reconAssessmentRequest,
|
|
1669
|
+
reconDecisionRequest: state?.recon_decision_request || null,
|
|
1670
|
+
executed
|
|
1671
|
+
}
|
|
1672
|
+
);
|
|
1673
|
+
}
|
|
1674
|
+
if (latestAssessment.decision === "recon_stuck" && latestAssessment.escalationTarget === "human") {
|
|
1675
|
+
const summary = latestAssessment.summary || "The supervising agent concluded recon is genuinely stuck and should escalate to the human.";
|
|
1676
|
+
recordAttempt("recon", "escalated", summary, {
|
|
1677
|
+
checkpoint: "recon_human_escalation",
|
|
1678
|
+
details: reconDetails
|
|
1679
|
+
});
|
|
1680
|
+
return checkpoint(
|
|
1681
|
+
"recon",
|
|
1682
|
+
"recon_human_escalation",
|
|
1683
|
+
summary,
|
|
1684
|
+
{
|
|
1685
|
+
ok: false,
|
|
1686
|
+
nextActions: ["inspect_recon_history", "summarize_failed_baselines", "ask_human_for_direction"],
|
|
1687
|
+
advanceOptions: ["recon", "author"],
|
|
1688
|
+
recommendedAdvanceStage: null,
|
|
1689
|
+
continueWithStage: null,
|
|
1690
|
+
blocking: true,
|
|
1691
|
+
details: reconDetails,
|
|
1692
|
+
reconAssessment: latestAssessment.raw,
|
|
1693
|
+
reconAssessmentRequest,
|
|
1694
|
+
executed
|
|
1695
|
+
}
|
|
1696
|
+
);
|
|
1697
|
+
}
|
|
1698
|
+
if ((latestAssessment.decision === "ready_for_author" || latestAssessment.continueWithStage === "author") && latestReconHasRequiredBaselines(state) && hasReconBaselineUnderstanding(state)) {
|
|
1699
|
+
updateState(config.statePath, (currentState) => {
|
|
1700
|
+
promoteLatestReconBaselines(currentState);
|
|
1701
|
+
currentState.recon_status = "ready_for_proof_plan";
|
|
1702
|
+
currentState.recon_results = currentState.recon_results || {};
|
|
1703
|
+
currentState.recon_results.status = "ready_for_proof_plan";
|
|
1704
|
+
currentState.recon_assessment_request = {};
|
|
1705
|
+
currentState.recon_decision_request = {};
|
|
1706
|
+
if ((currentState.proof_plan || "").trim() && (currentState.capture_script || "").trim()) {
|
|
1707
|
+
currentState.author_status = "ready";
|
|
1708
|
+
currentState.proof_plan_status = "ready";
|
|
1709
|
+
} else if (!authorReady(currentState)) {
|
|
1710
|
+
currentState.author_status = "needs_authoring";
|
|
1711
|
+
currentState.proof_plan_status = "needs_authoring";
|
|
1712
|
+
}
|
|
1713
|
+
});
|
|
1714
|
+
state = readState(config.statePath);
|
|
1715
|
+
const approvedSummary = latestAssessment.summary || "The supervising agent approved the latest recon baseline and selected the route for proof authoring.";
|
|
1716
|
+
if (params.advance_stage === "recon") {
|
|
1717
|
+
recordAttempt("recon", "completed", approvedSummary, {
|
|
1718
|
+
checkpoint: "recon_review",
|
|
1719
|
+
details: {
|
|
1720
|
+
...reconDetails,
|
|
1721
|
+
promotedBaselines: latestReconCapturedBaselines(state)
|
|
1722
|
+
}
|
|
1723
|
+
});
|
|
1724
|
+
return checkpoint(
|
|
1725
|
+
"recon",
|
|
1726
|
+
"recon_review",
|
|
1727
|
+
approvedSummary,
|
|
1728
|
+
{
|
|
1729
|
+
nextActions: ["inspect_recon_baseline", "continue_internal_loop_with_checkpoint", "advance_run_to_author"],
|
|
1730
|
+
advanceOptions: ["author", "recon", "implement"],
|
|
1731
|
+
recommendedAdvanceStage: "author",
|
|
1732
|
+
continueWithStage: "author",
|
|
1733
|
+
blocking: false,
|
|
1734
|
+
details: {
|
|
1735
|
+
...reconDetails,
|
|
1736
|
+
promotedBaselines: latestReconCapturedBaselines(state)
|
|
1737
|
+
},
|
|
1738
|
+
reconAssessment: latestAssessment.raw,
|
|
1739
|
+
executed
|
|
1740
|
+
}
|
|
1741
|
+
);
|
|
1742
|
+
}
|
|
1743
|
+
recordAttempt("recon", "completed", approvedSummary, {
|
|
1744
|
+
checkpoint: "recon_auto_continue",
|
|
1745
|
+
details: {
|
|
1746
|
+
...reconDetails,
|
|
1747
|
+
promotedBaselines: latestReconCapturedBaselines(state)
|
|
1748
|
+
}
|
|
1749
|
+
});
|
|
1750
|
+
effectiveAdvanceStage = "author";
|
|
1751
|
+
updateState(config.statePath, (currentState) => {
|
|
1752
|
+
currentState.last_requested_advance_stage = "author";
|
|
1753
|
+
});
|
|
1754
|
+
state = readState(config.statePath);
|
|
1755
|
+
requestedStage = normalizeStageRequest(state, effectiveAdvanceStage);
|
|
1756
|
+
} else if (latestAssessment.decision === "ready_for_author") {
|
|
1757
|
+
const missingUnderstanding = latestReconHasRequiredBaselines(state) && !hasReconBaselineUnderstanding(state);
|
|
1758
|
+
return checkpoint(
|
|
1759
|
+
"recon",
|
|
1760
|
+
"recon_supervisor_judgment",
|
|
1761
|
+
missingUnderstanding ? "The supervising agent tried to approve recon, but did not provide a concrete baseline_understanding. The before evidence must be understood before proof authoring or code edits begin." : "The supervising agent tried to approve recon, but the latest attempt is still missing one or more required baseline screenshots. Retry recon with a better plan or declare the loop genuinely stuck.",
|
|
1762
|
+
{
|
|
1763
|
+
ok: false,
|
|
1764
|
+
nextActions: ["inspect_recon_packet", "refine_recon_plan", "continue_internal_loop_with_checkpoint"],
|
|
1765
|
+
advanceOptions: ["recon", "author"],
|
|
1766
|
+
recommendedAdvanceStage: "recon",
|
|
1767
|
+
continueWithStage: "recon",
|
|
1768
|
+
blocking: false,
|
|
1769
|
+
details: reconDetails,
|
|
1770
|
+
reconAssessment: latestAssessment.raw,
|
|
1771
|
+
reconAssessmentRequest,
|
|
1772
|
+
executed
|
|
1773
|
+
}
|
|
1774
|
+
);
|
|
1775
|
+
} else {
|
|
1776
|
+
updateState(config.statePath, (currentState) => {
|
|
1777
|
+
currentState.recon_status = "";
|
|
1778
|
+
currentState.recon_assessment = {};
|
|
1779
|
+
currentState.recon_assessment_source = null;
|
|
1780
|
+
currentState.recon_assessment_request = {};
|
|
1781
|
+
currentState.recon_decision_request = {};
|
|
1782
|
+
currentState.before_cdn = "";
|
|
1783
|
+
currentState.prod_cdn = "";
|
|
1784
|
+
currentState.recon_results = currentState.recon_results || {};
|
|
1785
|
+
currentState.recon_results.baselines = {};
|
|
1786
|
+
currentState.recon_results.selected_attempt = {};
|
|
1787
|
+
currentState.recon_results.status = "retry_requested";
|
|
1788
|
+
});
|
|
1789
|
+
state = readState(config.statePath);
|
|
1790
|
+
}
|
|
1791
|
+
}
|
|
1792
|
+
if (!state?.recon_results || state?.stage === "setup" || state?.stage === "preflight" || ["needs_agent_decision", "needs_supervisor_judgment"].includes(state?.recon_status || "") || requestedStage === "recon") {
|
|
1793
|
+
const reconRes = runOne("recon");
|
|
1794
|
+
executed.push({ step: "recon", ok: reconRes.ok, haltedForApproval: reconRes.haltedForApproval || false, autoApproved: reconRes.autoApproved || false });
|
|
1795
|
+
if (!reconRes.ok || reconRes.haltedForApproval) {
|
|
1796
|
+
return failedRun("recon", reconRes.haltedForApproval ? "recon halted for approval" : "recon failed", reconRes, {
|
|
1797
|
+
checkpoint: "recon_failed",
|
|
1798
|
+
details: { executed },
|
|
1799
|
+
executed
|
|
1800
|
+
});
|
|
1801
|
+
}
|
|
1802
|
+
state = readState(config.statePath);
|
|
1803
|
+
if (["needs_agent_decision", "needs_supervisor_judgment"].includes(state?.recon_status || "")) {
|
|
1804
|
+
const reconAssessmentRequest = state?.recon_assessment_request || state?.recon_decision_request || null;
|
|
1805
|
+
const summary = "Recon gathered route hints, candidate paths, baseline captures, and observations. The supervising agent should now judge whether the latest baseline is trustworthy, whether recon should retry/reframe, and whether recon is done.";
|
|
1806
|
+
const reconDetails = {
|
|
1807
|
+
executed,
|
|
1808
|
+
latestAttempt: latestReconAttempt(state),
|
|
1809
|
+
latestCapturedBaselines: latestReconCapturedBaselines(state),
|
|
1810
|
+
reconAssessmentRequest
|
|
1811
|
+
};
|
|
1812
|
+
recordAttempt("recon", "checkpoint", summary, {
|
|
1813
|
+
autoApproved: reconRes.autoApproved || false,
|
|
1814
|
+
checkpoint: "recon_supervisor_judgment",
|
|
1815
|
+
details: reconDetails
|
|
1816
|
+
});
|
|
1817
|
+
return checkpoint(
|
|
1818
|
+
"recon",
|
|
1819
|
+
"recon_supervisor_judgment",
|
|
1820
|
+
summary,
|
|
1821
|
+
{
|
|
1822
|
+
nextActions: ["inspect_recon_packet", "supply_recon_assessment_json", "continue_internal_loop_with_checkpoint"],
|
|
1823
|
+
advanceOptions: ["recon", "author"],
|
|
1824
|
+
recommendedAdvanceStage: "recon",
|
|
1825
|
+
continueWithStage: "recon",
|
|
1826
|
+
blocking: false,
|
|
1827
|
+
details: reconDetails,
|
|
1828
|
+
reconAssessmentRequest,
|
|
1829
|
+
reconDecisionRequest: state?.recon_decision_request || null,
|
|
1830
|
+
executed
|
|
1831
|
+
}
|
|
1832
|
+
);
|
|
1833
|
+
}
|
|
1834
|
+
recordAttempt("recon", "completed", "Recon completed and promoted an approved baseline context.", {
|
|
1835
|
+
autoApproved: reconRes.autoApproved || false,
|
|
1836
|
+
details: { executed }
|
|
1837
|
+
});
|
|
1838
|
+
}
|
|
1839
|
+
state = readState(config.statePath);
|
|
1840
|
+
if (!authorReady(state) || effectiveAdvanceStage === "author") {
|
|
1841
|
+
const authorRes = runOne("author");
|
|
1842
|
+
executed.push({ step: "author", ok: authorRes.ok, haltedForApproval: authorRes.haltedForApproval || false, autoApproved: authorRes.autoApproved || false });
|
|
1843
|
+
if (!authorRes.ok || authorRes.haltedForApproval) {
|
|
1844
|
+
return failedRun("author", authorRes.haltedForApproval ? "author halted for approval" : "author failed", authorRes, {
|
|
1845
|
+
checkpoint: "author_failed",
|
|
1846
|
+
details: { executed },
|
|
1847
|
+
executed
|
|
1848
|
+
});
|
|
1849
|
+
}
|
|
1850
|
+
state = readState(config.statePath);
|
|
1851
|
+
if (!authorReady(state)) {
|
|
1852
|
+
recordAttempt("author", "checkpoint", "Author prepared a supervisor judgment request instead of delegating proof authoring to an internal model.", {
|
|
1853
|
+
autoApproved: authorRes.autoApproved || false,
|
|
1854
|
+
checkpoint: "author_supervisor_judgment",
|
|
1855
|
+
details: {
|
|
1856
|
+
executed,
|
|
1857
|
+
authorSummary: state?.author_summary || null,
|
|
1858
|
+
authorRequest: state?.author_request || null,
|
|
1859
|
+
serverPath: state?.server_path || null,
|
|
1860
|
+
waitForSelector: state?.wait_for_selector || null
|
|
1861
|
+
}
|
|
1862
|
+
});
|
|
1863
|
+
return checkpoint(
|
|
1864
|
+
"author",
|
|
1865
|
+
"author_supervisor_judgment",
|
|
1866
|
+
"Author distilled recon into a proof-authoring request. The supervising agent should supply the proof packet, then resume the workflow.",
|
|
1867
|
+
{
|
|
1868
|
+
nextActions: ["inspect_author_request", "supply_author_packet_json_or_proof_plan", "continue_internal_loop_with_checkpoint"],
|
|
1869
|
+
advanceOptions: ["author", "recon", "implement", "verify"],
|
|
1870
|
+
recommendedAdvanceStage: "author",
|
|
1871
|
+
continueWithStage: "author",
|
|
1872
|
+
details: {
|
|
1873
|
+
executed,
|
|
1874
|
+
authorSummary: state?.author_summary || null,
|
|
1875
|
+
authorRequest: state?.author_request || null,
|
|
1876
|
+
proofPlanDraft: state?.author_request?.fallback_defaults?.proof_plan || null,
|
|
1877
|
+
captureScriptDraft: state?.author_request?.fallback_defaults?.capture_script || null,
|
|
1878
|
+
serverPathDraft: state?.author_request?.fallback_defaults?.server_path || null,
|
|
1879
|
+
waitForSelectorDraft: state?.author_request?.fallback_defaults?.wait_for_selector || null
|
|
1880
|
+
},
|
|
1881
|
+
authorSummary: state?.author_summary || null,
|
|
1882
|
+
authorRequest: state?.author_request || null,
|
|
1883
|
+
proofPlanDraft: state?.author_request?.fallback_defaults?.proof_plan || null,
|
|
1884
|
+
captureScriptDraft: state?.author_request?.fallback_defaults?.capture_script || null,
|
|
1885
|
+
serverPathDraft: state?.author_request?.fallback_defaults?.server_path || null,
|
|
1886
|
+
waitForSelectorDraft: state?.author_request?.fallback_defaults?.wait_for_selector || null,
|
|
1887
|
+
executed
|
|
1888
|
+
}
|
|
1889
|
+
);
|
|
1890
|
+
}
|
|
1891
|
+
const authorNextStage = stageAfterAuthor(state);
|
|
1892
|
+
const explicitAuthorDebug = params.advance_stage === "author";
|
|
1893
|
+
recordAttempt("author", "completed", "Author applied the supervising agent's proof packet to recon observations.", {
|
|
1894
|
+
autoApproved: authorRes.autoApproved || false,
|
|
1895
|
+
checkpoint: explicitAuthorDebug ? "author_review" : "author_auto_continue",
|
|
1896
|
+
details: {
|
|
1897
|
+
executed,
|
|
1898
|
+
authorSummary: state?.author_summary || null,
|
|
1899
|
+
authorModel: state?.author_model || null,
|
|
1900
|
+
authorRuntimeModelHint: state?.author_runtime_model_hint || null,
|
|
1901
|
+
serverPath: state?.server_path || null,
|
|
1902
|
+
waitForSelector: state?.wait_for_selector || null
|
|
1903
|
+
}
|
|
1904
|
+
});
|
|
1905
|
+
if (explicitAuthorDebug) {
|
|
1906
|
+
return checkpoint(
|
|
1907
|
+
"author",
|
|
1908
|
+
"author_review",
|
|
1909
|
+
authorNextStage === "verify" ? "Author applied the supervising agent's proof packet. Because implementation is already recorded, you can continue straight into verify." : "Author applied the supervising agent's proof packet. Inspect it if needed, then continue into implement.",
|
|
1910
|
+
{
|
|
1911
|
+
nextActions: authorNextStage === "verify" ? ["inspect_proof_packet", "advance_run_to_verify", "rerun_author"] : ["inspect_proof_packet", "advance_run_to_implement", "rerun_author"],
|
|
1912
|
+
advanceOptions: authorNextStage === "verify" ? ["author", "verify", "recon"] : ["author", "implement", "recon"],
|
|
1913
|
+
recommendedAdvanceStage: authorNextStage,
|
|
1914
|
+
continueWithStage: authorNextStage,
|
|
1915
|
+
details: {
|
|
1916
|
+
executed,
|
|
1917
|
+
authorSummary: state?.author_summary || null,
|
|
1918
|
+
authorModel: state?.author_model || null,
|
|
1919
|
+
authorRuntimeModelHint: state?.author_runtime_model_hint || null,
|
|
1920
|
+
proofPlan: state?.proof_plan || null,
|
|
1921
|
+
serverPath: state?.server_path || null,
|
|
1922
|
+
waitForSelector: state?.wait_for_selector || null
|
|
1923
|
+
},
|
|
1924
|
+
authorSummary: state?.author_summary || null,
|
|
1925
|
+
authorModel: state?.author_model || null,
|
|
1926
|
+
authorRuntimeModelHint: state?.author_runtime_model_hint || null,
|
|
1927
|
+
proofPlan: state?.proof_plan || null,
|
|
1928
|
+
serverPath: state?.server_path || null,
|
|
1929
|
+
waitForSelector: state?.wait_for_selector || null,
|
|
1930
|
+
executed
|
|
1931
|
+
}
|
|
1932
|
+
);
|
|
1933
|
+
}
|
|
1934
|
+
effectiveAdvanceStage = authorNextStage;
|
|
1935
|
+
updateState(config.statePath, (currentState) => {
|
|
1936
|
+
currentState.last_requested_advance_stage = authorNextStage;
|
|
1937
|
+
});
|
|
1938
|
+
state = readState(config.statePath);
|
|
1939
|
+
}
|
|
1940
|
+
if (!effectiveAdvanceStage) {
|
|
1941
|
+
const recommended = recommendedAdvanceStage(state);
|
|
1942
|
+
return checkpoint(
|
|
1943
|
+
recommended || "implement",
|
|
1944
|
+
"awaiting_stage_advance",
|
|
1945
|
+
"Proof authoring is ready. The wrapper will not guess the next stage from here, explicitly choose whether to revisit recon/author, validate implementation, capture verify evidence, or ship.",
|
|
1946
|
+
{
|
|
1947
|
+
nextActions: ["inspect_state", "set_advance_stage", "resume_run"],
|
|
1948
|
+
advanceOptions: ["recon", "author", "implement", "verify", "ship"],
|
|
1949
|
+
recommendedAdvanceStage: recommended,
|
|
1950
|
+
details: { executed },
|
|
1951
|
+
executed
|
|
1952
|
+
}
|
|
1953
|
+
);
|
|
1954
|
+
}
|
|
1955
|
+
if (effectiveAdvanceStage === "implement") {
|
|
1956
|
+
const implementRes = runOne("implement");
|
|
1957
|
+
executed.push({ step: "implement", ok: implementRes.ok, haltedForApproval: implementRes.haltedForApproval || false, autoApproved: implementRes.autoApproved || false });
|
|
1958
|
+
if (implementRes.haltedForApproval) {
|
|
1959
|
+
return failedRun("implement", "implement halted for approval", implementRes, {
|
|
1960
|
+
checkpoint: "implement_blocked",
|
|
1961
|
+
details: { executed },
|
|
1962
|
+
executed
|
|
1963
|
+
});
|
|
1964
|
+
}
|
|
1965
|
+
if (!implementRes.ok) {
|
|
1966
|
+
const implementError = `${implementRes.error || ""}
|
|
1967
|
+
${implementRes.stdout || ""}
|
|
1968
|
+
${implementRes.stderr || ""}`;
|
|
1969
|
+
if (implementError.includes("No implementation detected")) {
|
|
1970
|
+
recordAttempt("implement", "checkpoint", "Implementation checkpoint found no material code changes yet.", {
|
|
1971
|
+
checkpoint: "implement_changes_missing",
|
|
1972
|
+
error: implementRes.error || null,
|
|
1973
|
+
details: { executed }
|
|
1974
|
+
});
|
|
1975
|
+
return checkpoint(
|
|
1976
|
+
"implement",
|
|
1977
|
+
"implement_changes_missing",
|
|
1978
|
+
"Proof plan is ready, but code changes are not recorded yet. Make the implementation changes on the after worktree, then resume run.",
|
|
1979
|
+
{
|
|
1980
|
+
nextActions: ["make_code_changes", "rerun_implement"],
|
|
1981
|
+
advanceOptions: ["implement", "author", "recon"],
|
|
1982
|
+
recommendedAdvanceStage: "implement",
|
|
1983
|
+
blocking: true,
|
|
1984
|
+
details: { executed },
|
|
1985
|
+
executed
|
|
1986
|
+
}
|
|
1987
|
+
);
|
|
1988
|
+
}
|
|
1989
|
+
return failedRun("implement", "implement failed", implementRes, {
|
|
1990
|
+
checkpoint: "implement_failed",
|
|
1991
|
+
details: { executed },
|
|
1992
|
+
executed
|
|
1993
|
+
});
|
|
1994
|
+
}
|
|
1995
|
+
let invalidatedVerifyEvidence2 = false;
|
|
1996
|
+
updateState(config.statePath, (state2) => {
|
|
1997
|
+
invalidatedVerifyEvidence2 = invalidateVerifyEvidence(state2).invalidated;
|
|
1998
|
+
});
|
|
1999
|
+
recordAttempt("implement", "completed", "Implementation checkpoint recorded code changes on the after worktree.", {
|
|
2000
|
+
autoApproved: implementRes.autoApproved || false,
|
|
2001
|
+
checkpoint: "implement_review",
|
|
2002
|
+
details: { executed, invalidatedVerifyEvidence: invalidatedVerifyEvidence2 }
|
|
2003
|
+
});
|
|
2004
|
+
return checkpoint(
|
|
2005
|
+
"implement",
|
|
2006
|
+
"implement_review",
|
|
2007
|
+
invalidatedVerifyEvidence2 ? "Implementation changes were detected and prior verify evidence was invalidated. Inspect the branch diff or notes, then explicitly choose whether to iterate implementation again or advance to verify." : "Implementation changes were detected. Inspect the branch diff or notes, then explicitly choose whether to iterate implementation again or advance to verify.",
|
|
2008
|
+
{
|
|
2009
|
+
nextActions: ["inspect_branch_diff", "rerun_implement", "advance_run_to_verify"],
|
|
2010
|
+
advanceOptions: ["implement", "author", "verify", "recon"],
|
|
2011
|
+
recommendedAdvanceStage: "verify",
|
|
2012
|
+
details: {
|
|
2013
|
+
executed,
|
|
2014
|
+
implementationSummary: readState(config.statePath)?.implementation_summary || null,
|
|
2015
|
+
invalidatedVerifyEvidence: invalidatedVerifyEvidence2
|
|
2016
|
+
},
|
|
2017
|
+
implementationSummary: readState(config.statePath)?.implementation_summary || null,
|
|
2018
|
+
invalidatedVerifyEvidence: invalidatedVerifyEvidence2,
|
|
2019
|
+
executed
|
|
2020
|
+
}
|
|
2021
|
+
);
|
|
2022
|
+
}
|
|
2023
|
+
if (effectiveAdvanceStage === "verify") {
|
|
2024
|
+
state = readState(config.statePath);
|
|
2025
|
+
if (!["changes_detected", "completed"].includes(state?.implementation_status || "")) {
|
|
2026
|
+
return checkpoint(
|
|
2027
|
+
"implement",
|
|
2028
|
+
"implement_required",
|
|
2029
|
+
"Verify is blocked until implementation has been recorded. Run the implement stage after making code changes, then resume verify.",
|
|
2030
|
+
{
|
|
2031
|
+
ok: false,
|
|
2032
|
+
nextActions: ["make_code_changes", "advance_run_to_implement"],
|
|
2033
|
+
advanceOptions: ["implement", "author", "recon"],
|
|
2034
|
+
recommendedAdvanceStage: "implement",
|
|
2035
|
+
continueWithStage: "implement",
|
|
2036
|
+
blocking: true,
|
|
2037
|
+
details: { executed },
|
|
2038
|
+
executed
|
|
2039
|
+
}
|
|
2040
|
+
);
|
|
2041
|
+
}
|
|
2042
|
+
const hasIncomingProofAssessment = typeof params.proof_assessment_json === "string" && params.proof_assessment_json.trim().length > 0;
|
|
2043
|
+
const canReuseVerifyEvidence = (params.advance_stage !== "verify" || hasIncomingProofAssessment) && state?.verify_status === "evidence_captured" && Boolean((state?.after_cdn || "").trim()) && (state?.active_checkpoint === "verify_supervisor_judgment" || hasSupervisorProofAssessment(state));
|
|
2044
|
+
let verifyRes = { ok: true, step: "verify", reusedEvidence: canReuseVerifyEvidence };
|
|
2045
|
+
if (!canReuseVerifyEvidence) {
|
|
2046
|
+
verifyRes = runOne("verify");
|
|
2047
|
+
executed.push({ step: "verify", ok: verifyRes.ok, haltedForApproval: verifyRes.haltedForApproval || false, autoApproved: verifyRes.autoApproved || false });
|
|
2048
|
+
if (!verifyRes.ok || verifyRes.haltedForApproval) {
|
|
2049
|
+
return failedRun("verify", verifyRes.haltedForApproval ? "verify halted for approval" : "verify failed", verifyRes, {
|
|
2050
|
+
checkpoint: "verify_failed",
|
|
2051
|
+
details: { executed },
|
|
2052
|
+
executed
|
|
2053
|
+
});
|
|
2054
|
+
}
|
|
2055
|
+
} else {
|
|
2056
|
+
executed.push({ step: "verify", ok: true, reusedEvidence: true, haltedForApproval: false, autoApproved: false });
|
|
2057
|
+
}
|
|
2058
|
+
state = readState(config.statePath);
|
|
2059
|
+
const verifyStatus = state?.verify_status || ((state?.after_cdn || "").trim() ? "evidence_captured" : "capture_incomplete");
|
|
2060
|
+
const verifyDecisionRequest = state?.verify_decision_request || null;
|
|
2061
|
+
const verifySummary = state?.verify_summary || state?.proof_summary || null;
|
|
2062
|
+
const proofAssessment = verifyAssessment(state);
|
|
2063
|
+
const convergenceSignals = nonConvergenceSignals(state, proofAssessment);
|
|
2064
|
+
const verifyRecommendedStage = proofAssessment.recommendedStage || null;
|
|
2065
|
+
const verifyContinueWithStage = shouldEscalateVerifyToHuman(state, proofAssessment) ? null : proofAssessment.continueWithStage || verifyRecommendedStage || null;
|
|
2066
|
+
const verifyDetails = {
|
|
2067
|
+
executed,
|
|
2068
|
+
verifyStatus,
|
|
2069
|
+
verifySummary,
|
|
2070
|
+
afterCdn: state?.after_cdn || null,
|
|
2071
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2072
|
+
verifyDecisionRequest,
|
|
2073
|
+
proofAssessment: proofAssessment.raw,
|
|
2074
|
+
proofAssessmentSource: proofAssessment.source || null,
|
|
2075
|
+
proofAssessmentRequest: state?.proof_assessment_request || null,
|
|
2076
|
+
verifyRecommendedStage,
|
|
2077
|
+
verifyContinueWithStage,
|
|
2078
|
+
convergenceSignals
|
|
2079
|
+
};
|
|
2080
|
+
if (verifyStatus !== "evidence_captured") {
|
|
2081
|
+
if ((verifyContinueWithStage || verifyRecommendedStage || "author") === "author") {
|
|
2082
|
+
updateState(config.statePath, (currentState) => {
|
|
2083
|
+
currentState.author_status = "needs_authoring";
|
|
2084
|
+
currentState.proof_plan_status = "needs_authoring";
|
|
2085
|
+
currentState.supervisor_author_packet = null;
|
|
2086
|
+
});
|
|
2087
|
+
state = readState(config.statePath);
|
|
2088
|
+
}
|
|
2089
|
+
const checkpointName = "verify_capture_retry";
|
|
2090
|
+
const summary = "Verify ran, but the proof packet still needs internal capture-plan work before it should ship.";
|
|
2091
|
+
recordAttempt("verify", "checkpoint", summary, {
|
|
2092
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2093
|
+
checkpoint: checkpointName,
|
|
2094
|
+
details: verifyDetails
|
|
2095
|
+
});
|
|
2096
|
+
return checkpoint(
|
|
2097
|
+
"verify",
|
|
2098
|
+
checkpointName,
|
|
2099
|
+
summary,
|
|
2100
|
+
{
|
|
2101
|
+
ok: true,
|
|
2102
|
+
nextActions: ["inspect_after_capture", "continue_internal_loop_with_checkpoint", "return_to_recon_if_baseline_is_wrong"],
|
|
2103
|
+
advanceOptions: ["author", "verify", "implement", "recon"],
|
|
2104
|
+
recommendedAdvanceStage: verifyRecommendedStage || "author",
|
|
2105
|
+
continueWithStage: verifyContinueWithStage || "author",
|
|
2106
|
+
blocking: false,
|
|
2107
|
+
details: verifyDetails,
|
|
2108
|
+
verifyStatus,
|
|
2109
|
+
verifySummary,
|
|
2110
|
+
afterCdn: state?.after_cdn || null,
|
|
2111
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2112
|
+
verifyDecisionRequest,
|
|
2113
|
+
proofAssessment: proofAssessment.raw,
|
|
2114
|
+
executed
|
|
2115
|
+
}
|
|
2116
|
+
);
|
|
2117
|
+
}
|
|
2118
|
+
if (!hasSupervisorProofAssessment(state)) {
|
|
2119
|
+
const summary = "Verify captured usable evidence. The supervising agent should now assess whether the proof supports ship or more internal iteration, then resume the workflow with proof_assessment_json.";
|
|
2120
|
+
recordAttempt("verify", "checkpoint", summary, {
|
|
2121
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2122
|
+
checkpoint: "verify_supervisor_judgment",
|
|
2123
|
+
details: verifyDetails
|
|
2124
|
+
});
|
|
2125
|
+
return checkpoint(
|
|
2126
|
+
"verify",
|
|
2127
|
+
"verify_supervisor_judgment",
|
|
2128
|
+
summary,
|
|
2129
|
+
{
|
|
2130
|
+
nextActions: ["inspect_evidence", "author_proof_assessment_json", "continue_internal_loop_with_checkpoint"],
|
|
2131
|
+
advanceOptions: ["verify", "author", "implement", "recon", "ship"],
|
|
2132
|
+
recommendedAdvanceStage: "verify",
|
|
2133
|
+
continueWithStage: "verify",
|
|
2134
|
+
blocking: false,
|
|
2135
|
+
details: verifyDetails,
|
|
2136
|
+
verifyStatus,
|
|
2137
|
+
verifySummary,
|
|
2138
|
+
afterCdn: state?.after_cdn || null,
|
|
2139
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2140
|
+
verifyDecisionRequest,
|
|
2141
|
+
proofAssessmentRequest: state?.proof_assessment_request || null,
|
|
2142
|
+
executed
|
|
2143
|
+
}
|
|
2144
|
+
);
|
|
2145
|
+
}
|
|
2146
|
+
const shouldEscalate = shouldEscalateVerifyToHuman(state, proofAssessment);
|
|
2147
|
+
if (shouldEscalate) {
|
|
2148
|
+
const summary = "The supervising agent concluded the workflow hit a real wall and explicitly escalated the proof loop to the human.";
|
|
2149
|
+
recordAttempt("verify", "escalated", summary, {
|
|
2150
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2151
|
+
checkpoint: "verify_human_escalation",
|
|
2152
|
+
details: verifyDetails
|
|
2153
|
+
});
|
|
2154
|
+
return checkpoint(
|
|
2155
|
+
"verify",
|
|
2156
|
+
"verify_human_escalation",
|
|
2157
|
+
summary,
|
|
2158
|
+
{
|
|
2159
|
+
ok: false,
|
|
2160
|
+
nextActions: ["inspect_retry_history", "summarize_internal_loop", "ask_human_for_direction"],
|
|
2161
|
+
advanceOptions: ["author", "implement", "ship", "verify", "recon"],
|
|
2162
|
+
recommendedAdvanceStage: null,
|
|
2163
|
+
continueWithStage: null,
|
|
2164
|
+
blocking: true,
|
|
2165
|
+
details: verifyDetails,
|
|
2166
|
+
verifyStatus,
|
|
2167
|
+
verifySummary,
|
|
2168
|
+
afterCdn: state?.after_cdn || null,
|
|
2169
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2170
|
+
verifyDecisionRequest,
|
|
2171
|
+
proofAssessment: proofAssessment.raw,
|
|
2172
|
+
executed
|
|
2173
|
+
}
|
|
2174
|
+
);
|
|
2175
|
+
}
|
|
2176
|
+
const shouldAutoShip = verifyContinueWithStage === "ship" && (params.ship_after_verify || params.continue_from_checkpoint || params.advance_stage !== "verify");
|
|
2177
|
+
if (shouldAutoShip) {
|
|
2178
|
+
const shipGate = validateShipGate(state);
|
|
2179
|
+
if (!shipGate.ok) {
|
|
2180
|
+
recordAttempt("verify", "checkpoint", "Verify cannot continue into ship because the hard ship gate is missing required evidence or approval.", {
|
|
2181
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2182
|
+
checkpoint: "ship_gate_blocked",
|
|
2183
|
+
details: { ...verifyDetails, shipGate }
|
|
2184
|
+
});
|
|
2185
|
+
return shipGateBlocked(state, executed, verifyDetails);
|
|
2186
|
+
}
|
|
2187
|
+
recordAttempt("verify", "checkpoint", "Verify captured a strong proof packet and is continuing directly into ship.", {
|
|
2188
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2189
|
+
checkpoint: "verify_then_ship",
|
|
2190
|
+
details: { ...verifyDetails, shipGate }
|
|
2191
|
+
});
|
|
2192
|
+
const shipRes = runOne("ship");
|
|
2193
|
+
executed.push({ step: "ship", ok: shipRes.ok, haltedForApproval: shipRes.haltedForApproval || false, autoApproved: shipRes.autoApproved || false });
|
|
2194
|
+
if (!shipRes.ok || shipRes.haltedForApproval) {
|
|
2195
|
+
const shipNextAction = shipRes?.error && String(shipRes.error).includes("temporary proof branch") ? "product bug: ship resolved a temporary proof branch; resolve the PR head branch before retrying ship" : "inspect the ship error, confirm the PR head branch and verified commit, then retry ship";
|
|
2196
|
+
return failedRun("ship", shipRes.haltedForApproval ? "ship halted for approval" : "ship failed", shipRes, {
|
|
2197
|
+
checkpoint: "ship_failed",
|
|
2198
|
+
details: { executed, next_action: shipNextAction },
|
|
2199
|
+
nextAction: shipNextAction,
|
|
2200
|
+
executed
|
|
2201
|
+
});
|
|
2202
|
+
}
|
|
2203
|
+
recordAttempt("ship", "completed", "Ship updated the PR and posted proof artifacts after the supervising agent judged the proof strong enough.", {
|
|
2204
|
+
autoApproved: shipRes.autoApproved || false,
|
|
2205
|
+
checkpoint: "ship_review",
|
|
2206
|
+
details: { executed }
|
|
2207
|
+
});
|
|
2208
|
+
const snapshot2 = snapshotFor(config.statePath);
|
|
2209
|
+
const summary = "The supervising agent judged the proof strong enough, so the workflow shipped automatically and left the PR as the main human review surface.";
|
|
2210
|
+
const finalState = readState(config.statePath);
|
|
2211
|
+
const shipReport = finalState?.ship_report || snapshot2.state?.ship_report || null;
|
|
2212
|
+
return {
|
|
2213
|
+
ok: true,
|
|
2214
|
+
action,
|
|
2215
|
+
state_path: config.statePath,
|
|
2216
|
+
stage: snapshot2.stage,
|
|
2217
|
+
checkpoint: "ship_review",
|
|
2218
|
+
summary,
|
|
2219
|
+
state: snapshot2.state,
|
|
2220
|
+
shipReport,
|
|
2221
|
+
checkpointContract: buildCheckpointContract(readState(config.statePath), {
|
|
2222
|
+
statePath: config.statePath,
|
|
2223
|
+
stage: "ship",
|
|
2224
|
+
checkpoint: "ship_review",
|
|
2225
|
+
summary,
|
|
2226
|
+
nextActions: ["inspect_pr", "rerun_ship_if_needed"],
|
|
2227
|
+
advanceOptions: ["ship", "verify", "author", "implement"],
|
|
2228
|
+
recommendedAdvanceStage: "ship"
|
|
2229
|
+
}),
|
|
2230
|
+
executed
|
|
2231
|
+
};
|
|
2232
|
+
}
|
|
2233
|
+
if (proofAssessment.decision === "ready_to_ship") {
|
|
2234
|
+
const shipGate = validateShipGate(state);
|
|
2235
|
+
if (!shipGate.ok) {
|
|
2236
|
+
recordAttempt("verify", "checkpoint", "Verify cannot mark ship ready because the hard ship gate is missing required evidence or approval.", {
|
|
2237
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2238
|
+
checkpoint: "ship_gate_blocked",
|
|
2239
|
+
details: { ...verifyDetails, shipGate }
|
|
2240
|
+
});
|
|
2241
|
+
return shipGateBlocked(state, executed, verifyDetails);
|
|
2242
|
+
}
|
|
2243
|
+
recordAttempt("verify", "checkpoint", "Verify captured a strong proof packet and is ready to continue into ship.", {
|
|
2244
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2245
|
+
checkpoint: "verify_ship_ready",
|
|
2246
|
+
details: { ...verifyDetails, shipGate }
|
|
2247
|
+
});
|
|
2248
|
+
return checkpoint(
|
|
2249
|
+
"verify",
|
|
2250
|
+
"verify_ship_ready",
|
|
2251
|
+
"The supervising agent judged the proof strong enough to continue into ship.",
|
|
2252
|
+
{
|
|
2253
|
+
nextActions: ["inspect_evidence", "continue_internal_loop_with_checkpoint", "advance_run_to_ship_if_you_need_manual_control"],
|
|
2254
|
+
advanceOptions: ["ship", "verify", "author", "implement", "recon"],
|
|
2255
|
+
recommendedAdvanceStage: "ship",
|
|
2256
|
+
continueWithStage: "ship",
|
|
2257
|
+
blocking: false,
|
|
2258
|
+
details: { ...verifyDetails, shipGate },
|
|
2259
|
+
shipGate,
|
|
2260
|
+
verifyStatus,
|
|
2261
|
+
verifySummary,
|
|
2262
|
+
afterCdn: state?.after_cdn || null,
|
|
2263
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2264
|
+
verifyDecisionRequest,
|
|
2265
|
+
proofAssessment: proofAssessment.raw,
|
|
2266
|
+
executed
|
|
2267
|
+
}
|
|
2268
|
+
);
|
|
2269
|
+
}
|
|
2270
|
+
if (verifyContinueWithStage === "author") {
|
|
2271
|
+
updateState(config.statePath, (currentState) => {
|
|
2272
|
+
currentState.author_status = "needs_authoring";
|
|
2273
|
+
currentState.proof_plan_status = "needs_authoring";
|
|
2274
|
+
currentState.supervisor_author_packet = null;
|
|
2275
|
+
});
|
|
2276
|
+
state = readState(config.statePath);
|
|
2277
|
+
}
|
|
2278
|
+
const unresolvedSummary = convergenceSignals.warning ? "The supervising agent kept the workflow in the internal loop, but retry history suggests it may not be converging yet. Keep iterating internally or explicitly escalate with escalation_target=human when you conclude it is genuinely stuck." : "The supervising agent judged that the workflow should keep iterating internally before it ships.";
|
|
2279
|
+
recordAttempt("verify", "checkpoint", unresolvedSummary, {
|
|
2280
|
+
autoApproved: verifyRes.autoApproved || false,
|
|
2281
|
+
checkpoint: "verify_agent_retry",
|
|
2282
|
+
details: verifyDetails
|
|
2283
|
+
});
|
|
2284
|
+
return checkpoint(
|
|
2285
|
+
"verify",
|
|
2286
|
+
"verify_agent_retry",
|
|
2287
|
+
unresolvedSummary,
|
|
2288
|
+
{
|
|
2289
|
+
ok: true,
|
|
2290
|
+
nextActions: convergenceSignals.warning ? ["inspect_retry_history", "decide_whether_to_keep_iterating_or_escalate", "continue_internal_loop_with_checkpoint"] : ["inspect_proof_assessment", "continue_internal_loop_with_checkpoint", "return_to_implement_if_fix_failed"],
|
|
2291
|
+
advanceOptions: ["author", "implement", "ship", "verify", "recon"],
|
|
2292
|
+
recommendedAdvanceStage: verifyRecommendedStage,
|
|
2293
|
+
continueWithStage: verifyContinueWithStage,
|
|
2294
|
+
blocking: false,
|
|
2295
|
+
details: verifyDetails,
|
|
2296
|
+
verifyStatus,
|
|
2297
|
+
verifySummary,
|
|
2298
|
+
afterCdn: state?.after_cdn || null,
|
|
2299
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2300
|
+
verifyDecisionRequest,
|
|
2301
|
+
proofAssessment: proofAssessment.raw,
|
|
2302
|
+
executed
|
|
2303
|
+
}
|
|
2304
|
+
);
|
|
2305
|
+
}
|
|
2306
|
+
if (effectiveAdvanceStage === "ship") {
|
|
2307
|
+
state = readState(config.statePath);
|
|
2308
|
+
const shipAssessment = verifyAssessment(state);
|
|
2309
|
+
const shipGate = validateShipGate(state);
|
|
2310
|
+
if (state?.verify_status !== "evidence_captured") {
|
|
2311
|
+
return checkpoint(
|
|
2312
|
+
"verify",
|
|
2313
|
+
"verify_required",
|
|
2314
|
+
"Ship is blocked until verify has captured a usable proof packet. Run verify, inspect the evidence, then explicitly advance to ship only if the proof supports success.",
|
|
2315
|
+
{
|
|
2316
|
+
ok: false,
|
|
2317
|
+
nextActions: ["advance_run_to_verify", "inspect_verify_state"],
|
|
2318
|
+
advanceOptions: ["verify", "author", "implement", "recon"],
|
|
2319
|
+
recommendedAdvanceStage: "verify",
|
|
2320
|
+
continueWithStage: "verify",
|
|
2321
|
+
blocking: true,
|
|
2322
|
+
details: {
|
|
2323
|
+
executed,
|
|
2324
|
+
shipGate,
|
|
2325
|
+
verifyStatus: state?.verify_status || null,
|
|
2326
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2327
|
+
afterCdn: state?.after_cdn || null
|
|
2328
|
+
},
|
|
2329
|
+
shipGate,
|
|
2330
|
+
verifyStatus: state?.verify_status || null,
|
|
2331
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2332
|
+
afterCdn: state?.after_cdn || null,
|
|
2333
|
+
executed
|
|
2334
|
+
}
|
|
2335
|
+
);
|
|
2336
|
+
}
|
|
2337
|
+
if (!hasSupervisorProofAssessment(state) || shipAssessment.decision !== "ready_to_ship") {
|
|
2338
|
+
return checkpoint(
|
|
2339
|
+
"verify",
|
|
2340
|
+
"verify_supervisor_judgment_required",
|
|
2341
|
+
"Ship is blocked until the supervising agent judges the current proof packet as ready_to_ship.",
|
|
2342
|
+
{
|
|
2343
|
+
ok: false,
|
|
2344
|
+
nextActions: ["inspect_evidence", "supply_proof_assessment_json", "continue_internal_loop_with_checkpoint"],
|
|
2345
|
+
advanceOptions: ["verify", "author", "implement", "recon", "ship"],
|
|
2346
|
+
recommendedAdvanceStage: "verify",
|
|
2347
|
+
continueWithStage: "verify",
|
|
2348
|
+
blocking: true,
|
|
2349
|
+
details: {
|
|
2350
|
+
executed,
|
|
2351
|
+
shipGate,
|
|
2352
|
+
verifyStatus: state?.verify_status || null,
|
|
2353
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2354
|
+
afterCdn: state?.after_cdn || null,
|
|
2355
|
+
proofAssessment: state?.proof_assessment || null,
|
|
2356
|
+
proofAssessmentRequest: state?.proof_assessment_request || null
|
|
2357
|
+
},
|
|
2358
|
+
shipGate,
|
|
2359
|
+
verifyStatus: state?.verify_status || null,
|
|
2360
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2361
|
+
afterCdn: state?.after_cdn || null,
|
|
2362
|
+
proofAssessment: state?.proof_assessment || null,
|
|
2363
|
+
proofAssessmentRequest: state?.proof_assessment_request || null,
|
|
2364
|
+
executed
|
|
2365
|
+
}
|
|
2366
|
+
);
|
|
2367
|
+
}
|
|
2368
|
+
if (!shipGate.ok) {
|
|
2369
|
+
return shipGateBlocked(state, executed, { shipAssessment: shipAssessment.raw });
|
|
2370
|
+
}
|
|
2371
|
+
const shipRes = runOne("ship");
|
|
2372
|
+
executed.push({ step: "ship", ok: shipRes.ok, haltedForApproval: shipRes.haltedForApproval || false, autoApproved: shipRes.autoApproved || false });
|
|
2373
|
+
if (!shipRes.ok || shipRes.haltedForApproval) {
|
|
2374
|
+
return failedRun("ship", shipRes.haltedForApproval ? "ship halted for approval" : "ship failed", shipRes, {
|
|
2375
|
+
checkpoint: "ship_failed",
|
|
2376
|
+
details: { executed },
|
|
2377
|
+
executed
|
|
2378
|
+
});
|
|
2379
|
+
}
|
|
2380
|
+
recordAttempt("ship", "completed", "Ship updated the PR and posted proof artifacts.", {
|
|
2381
|
+
autoApproved: shipRes.autoApproved || false,
|
|
2382
|
+
checkpoint: "ship_review",
|
|
2383
|
+
details: { executed }
|
|
2384
|
+
});
|
|
2385
|
+
return checkpoint(
|
|
2386
|
+
"ship",
|
|
2387
|
+
"ship_review",
|
|
2388
|
+
"Ship completed. Review the PR, proof comment, and cleanup results. Re-run ship if you need to refresh the PR after more changes.",
|
|
2389
|
+
{
|
|
2390
|
+
nextActions: ["inspect_pr", "rerun_ship_if_needed"],
|
|
2391
|
+
advanceOptions: ["ship", "verify", "author", "implement"],
|
|
2392
|
+
recommendedAdvanceStage: "ship",
|
|
2393
|
+
details: {
|
|
2394
|
+
executed,
|
|
2395
|
+
prUrl: readState(config.statePath)?.pr_url || null
|
|
2396
|
+
},
|
|
2397
|
+
prUrl: readState(config.statePath)?.pr_url || null,
|
|
2398
|
+
executed
|
|
2399
|
+
}
|
|
2400
|
+
);
|
|
2401
|
+
}
|
|
2402
|
+
}
|
|
2403
|
+
if (action === "ship") {
|
|
2404
|
+
const state = readState(config.statePath);
|
|
2405
|
+
const shipGate = validateShipGate(state);
|
|
2406
|
+
if (state?.verify_status !== "evidence_captured") {
|
|
2407
|
+
return checkpoint(
|
|
2408
|
+
"verify",
|
|
2409
|
+
"verify_required",
|
|
2410
|
+
"Ship is blocked until verify has captured a usable proof packet. Run verify, inspect the evidence, then ship only if the proof supports success.",
|
|
2411
|
+
{
|
|
2412
|
+
ok: false,
|
|
2413
|
+
nextActions: ["run_verify", "inspect_verify_state"],
|
|
2414
|
+
advanceOptions: ["verify", "author", "implement", "recon"],
|
|
2415
|
+
recommendedAdvanceStage: "verify",
|
|
2416
|
+
continueWithStage: "verify",
|
|
2417
|
+
blocking: true,
|
|
2418
|
+
details: { shipGate },
|
|
2419
|
+
shipGate,
|
|
2420
|
+
verifyStatus: state?.verify_status || null,
|
|
2421
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2422
|
+
afterCdn: state?.after_cdn || null
|
|
2423
|
+
}
|
|
2424
|
+
);
|
|
2425
|
+
}
|
|
2426
|
+
if (!hasSupervisorProofAssessment(state) || verifyAssessment(state).decision !== "ready_to_ship") {
|
|
2427
|
+
return checkpoint(
|
|
2428
|
+
"verify",
|
|
2429
|
+
"verify_supervisor_judgment_required",
|
|
2430
|
+
"Ship is blocked until the supervising agent judges the current proof packet as ready_to_ship.",
|
|
2431
|
+
{
|
|
2432
|
+
ok: false,
|
|
2433
|
+
nextActions: ["inspect_evidence", "supply_proof_assessment_json", "rerun_ship"],
|
|
2434
|
+
advanceOptions: ["verify", "author", "implement", "recon", "ship"],
|
|
2435
|
+
recommendedAdvanceStage: "verify",
|
|
2436
|
+
continueWithStage: "verify",
|
|
2437
|
+
blocking: true,
|
|
2438
|
+
details: { shipGate },
|
|
2439
|
+
shipGate,
|
|
2440
|
+
verifyStatus: state?.verify_status || null,
|
|
2441
|
+
mergeRecommendation: state?.merge_recommendation || null,
|
|
2442
|
+
afterCdn: state?.after_cdn || null,
|
|
2443
|
+
proofAssessment: state?.proof_assessment || null,
|
|
2444
|
+
proofAssessmentRequest: state?.proof_assessment_request || null
|
|
2445
|
+
}
|
|
2446
|
+
);
|
|
2447
|
+
}
|
|
2448
|
+
if (!shipGate.ok) {
|
|
2449
|
+
return shipGateBlocked(state, [], {});
|
|
2450
|
+
}
|
|
2451
|
+
}
|
|
2452
|
+
const single = runOne(action);
|
|
2453
|
+
if (!single.ok || single.haltedForApproval) {
|
|
2454
|
+
return failedRun(action, single.haltedForApproval ? `${action} halted for approval` : `${action} failed`, single, {
|
|
2455
|
+
checkpoint: `${action}_failed`
|
|
2456
|
+
});
|
|
2457
|
+
}
|
|
2458
|
+
let invalidatedVerifyEvidence = false;
|
|
2459
|
+
updateState(config.statePath, (state) => {
|
|
2460
|
+
if (action === "implement") {
|
|
2461
|
+
invalidatedVerifyEvidence = invalidateVerifyEvidence(state).invalidated;
|
|
2462
|
+
}
|
|
2463
|
+
clearStageDecisionRequest(state);
|
|
2464
|
+
});
|
|
2465
|
+
const singleSummary = action === "implement" && invalidatedVerifyEvidence ? "implement completed and invalidated prior verify evidence" : `${action} completed`;
|
|
2466
|
+
recordAttempt(action, "completed", singleSummary, {
|
|
2467
|
+
autoApproved: single.autoApproved || false,
|
|
2468
|
+
details: action === "implement" ? { invalidatedVerifyEvidence } : {}
|
|
2469
|
+
});
|
|
2470
|
+
const snapshot = snapshotFor(config.statePath);
|
|
2471
|
+
return {
|
|
2472
|
+
ok: true,
|
|
2473
|
+
action,
|
|
2474
|
+
state_path: config.statePath,
|
|
2475
|
+
stage: snapshot.stage,
|
|
2476
|
+
summary: singleSummary,
|
|
2477
|
+
state: snapshot.state,
|
|
2478
|
+
approval: null,
|
|
2479
|
+
autoApproved: single.autoApproved || false,
|
|
2480
|
+
error: null
|
|
2481
|
+
};
|
|
2482
|
+
}
|
|
2483
|
+
function createRiddleProofEngine(pluginConfig = {}) {
|
|
2484
|
+
return {
|
|
2485
|
+
execute(params) {
|
|
2486
|
+
const config = resolveConfig(pluginConfig, params);
|
|
2487
|
+
return executeWorkflow(params, pluginConfig, config);
|
|
2488
|
+
},
|
|
2489
|
+
status(statePath) {
|
|
2490
|
+
const config = resolveConfig(pluginConfig, { action: "status", state_path: statePath });
|
|
2491
|
+
return executeWorkflow({ action: "status", state_path: statePath }, pluginConfig, config);
|
|
2492
|
+
},
|
|
2493
|
+
resolveConfig(params = {}) {
|
|
2494
|
+
return resolveConfig(pluginConfig, params);
|
|
2495
|
+
}
|
|
2496
|
+
};
|
|
2497
|
+
}
|
|
2498
|
+
var import_node_child_process, import_node_fs2, import_node_path2;
|
|
2499
|
+
var init_proof_run_engine = __esm({
|
|
2500
|
+
"src/proof-run-engine.ts"() {
|
|
2501
|
+
"use strict";
|
|
2502
|
+
import_node_child_process = require("child_process");
|
|
2503
|
+
import_node_fs2 = require("fs");
|
|
2504
|
+
import_node_path2 = __toESM(require("path"), 1);
|
|
2505
|
+
init_proof_run_core();
|
|
2506
|
+
}
|
|
2507
|
+
});
|
|
2508
|
+
|
|
30
2509
|
// src/engine-harness.ts
|
|
31
2510
|
var engine_harness_exports = {};
|
|
32
2511
|
__export(engine_harness_exports, {
|
|
@@ -35,10 +2514,10 @@ __export(engine_harness_exports, {
|
|
|
35
2514
|
runRiddleProofEngineHarness: () => runRiddleProofEngineHarness
|
|
36
2515
|
});
|
|
37
2516
|
module.exports = __toCommonJS(engine_harness_exports);
|
|
38
|
-
var
|
|
39
|
-
var
|
|
40
|
-
var
|
|
41
|
-
var
|
|
2517
|
+
var import_node_child_process2 = require("child_process");
|
|
2518
|
+
var import_node_fs3 = require("fs");
|
|
2519
|
+
var import_node_path3 = __toESM(require("path"), 1);
|
|
2520
|
+
var import_node_crypto2 = __toESM(require("crypto"), 1);
|
|
42
2521
|
|
|
43
2522
|
// src/result.ts
|
|
44
2523
|
function isSuccessfulStatus(status) {
|
|
@@ -420,22 +2899,22 @@ function timestamp2() {
|
|
|
420
2899
|
}
|
|
421
2900
|
function createHarnessStatePath(stateDir) {
|
|
422
2901
|
const stamp = timestamp2().replace(/\D/g, "").slice(0, 14) || "unknown";
|
|
423
|
-
return
|
|
2902
|
+
return import_node_path3.default.join(stateDir, `riddle-proof-run-${stamp}-${import_node_crypto2.default.randomUUID().slice(0, 8)}.json`);
|
|
424
2903
|
}
|
|
425
2904
|
function ensureParent(filePath) {
|
|
426
|
-
(0,
|
|
2905
|
+
(0, import_node_fs3.mkdirSync)(import_node_path3.default.dirname(filePath), { recursive: true });
|
|
427
2906
|
}
|
|
428
2907
|
function readJson(filePath) {
|
|
429
|
-
if (!filePath || !(0,
|
|
2908
|
+
if (!filePath || !(0, import_node_fs3.existsSync)(filePath)) return null;
|
|
430
2909
|
try {
|
|
431
|
-
return JSON.parse((0,
|
|
2910
|
+
return JSON.parse((0, import_node_fs3.readFileSync)(filePath, "utf-8"));
|
|
432
2911
|
} catch {
|
|
433
2912
|
return null;
|
|
434
2913
|
}
|
|
435
2914
|
}
|
|
436
2915
|
function writeJson(filePath, payload) {
|
|
437
2916
|
ensureParent(filePath);
|
|
438
|
-
(0,
|
|
2917
|
+
(0, import_node_fs3.writeFileSync)(filePath, JSON.stringify(payload, null, 2) + "\n");
|
|
439
2918
|
}
|
|
440
2919
|
function loadRunState(input) {
|
|
441
2920
|
if (input.state) return input.state;
|
|
@@ -474,9 +2953,9 @@ function workdirFromState(state) {
|
|
|
474
2953
|
return nonEmptyString(state?.after_worktree) || nonEmptyString(state?.worktree_path) || null;
|
|
475
2954
|
}
|
|
476
2955
|
function hasGitDiff(workdir) {
|
|
477
|
-
if (!workdir || !(0,
|
|
2956
|
+
if (!workdir || !(0, import_node_fs3.existsSync)(workdir)) return false;
|
|
478
2957
|
try {
|
|
479
|
-
const status = (0,
|
|
2958
|
+
const status = (0, import_node_child_process2.execFileSync)("git", ["status", "--porcelain"], {
|
|
480
2959
|
cwd: workdir,
|
|
481
2960
|
encoding: "utf-8",
|
|
482
2961
|
timeout: 1e4
|
|
@@ -487,18 +2966,18 @@ function hasGitDiff(workdir) {
|
|
|
487
2966
|
}
|
|
488
2967
|
}
|
|
489
2968
|
function removeEmptyToolArtifacts(workdir) {
|
|
490
|
-
if (!workdir || !(0,
|
|
491
|
-
const artifactPath =
|
|
492
|
-
if (!(0,
|
|
2969
|
+
if (!workdir || !(0, import_node_fs3.existsSync)(workdir)) return [];
|
|
2970
|
+
const artifactPath = import_node_path3.default.join(workdir, ".codex");
|
|
2971
|
+
if (!(0, import_node_fs3.existsSync)(artifactPath)) return [];
|
|
493
2972
|
try {
|
|
494
|
-
const status = (0,
|
|
2973
|
+
const status = (0, import_node_child_process2.execFileSync)("git", ["status", "--porcelain", "--", ".codex"], {
|
|
495
2974
|
cwd: workdir,
|
|
496
2975
|
encoding: "utf-8",
|
|
497
2976
|
timeout: 1e4
|
|
498
2977
|
}).trim();
|
|
499
|
-
const stat = (0,
|
|
2978
|
+
const stat = (0, import_node_fs3.statSync)(artifactPath);
|
|
500
2979
|
if (status.startsWith("?? ") && stat.isFile() && stat.size === 0) {
|
|
501
|
-
(0,
|
|
2980
|
+
(0, import_node_fs3.unlinkSync)(artifactPath);
|
|
502
2981
|
return [".codex"];
|
|
503
2982
|
}
|
|
504
2983
|
} catch {
|
|
@@ -580,12 +3059,12 @@ function initialRunParams(request, input, state) {
|
|
|
580
3059
|
function effectiveShipMode(request, config) {
|
|
581
3060
|
return request.ship_mode || config?.defaultShipMode || "ship";
|
|
582
3061
|
}
|
|
583
|
-
function
|
|
3062
|
+
function checkpointContinueStage2(result) {
|
|
584
3063
|
const resume = recordValue(result.checkpointContract?.resume);
|
|
585
3064
|
return nonEmptyString(resume?.continue_with_stage);
|
|
586
3065
|
}
|
|
587
3066
|
function recommendedContinuation(result) {
|
|
588
|
-
const continueStage =
|
|
3067
|
+
const continueStage = checkpointContinueStage2(result);
|
|
589
3068
|
if (!continueStage) return null;
|
|
590
3069
|
return {
|
|
591
3070
|
action: "run",
|
|
@@ -729,7 +3208,11 @@ async function resolveEngine(input) {
|
|
|
729
3208
|
if (input.engine) return input.engine;
|
|
730
3209
|
const moduleUrl = input.config?.riddleEngineModuleUrl;
|
|
731
3210
|
if (!moduleUrl) {
|
|
732
|
-
|
|
3211
|
+
const mod2 = await Promise.resolve().then(() => (init_proof_run_engine(), proof_run_engine_exports));
|
|
3212
|
+
return mod2.createRiddleProofEngine({
|
|
3213
|
+
riddleProofDir: input.config?.riddleProofDir,
|
|
3214
|
+
defaultReviewer: input.config?.defaultReviewer
|
|
3215
|
+
});
|
|
733
3216
|
}
|
|
734
3217
|
const mod = await import(moduleUrl);
|
|
735
3218
|
if (typeof mod.createRiddleProofEngine !== "function") {
|
|
@@ -746,7 +3229,7 @@ async function handleImplementation(request, state, result, agent) {
|
|
|
746
3229
|
state.worktree_path = workdir || state.worktree_path;
|
|
747
3230
|
state.branch = nonEmptyString(context.fullRiddleState?.branch) || state.branch;
|
|
748
3231
|
persist(state);
|
|
749
|
-
if (!workdir || !(0,
|
|
3232
|
+
if (!workdir || !(0, import_node_fs3.existsSync)(workdir)) {
|
|
750
3233
|
return {
|
|
751
3234
|
blocker: {
|
|
752
3235
|
code: "implementation_worktree_missing",
|
|
@@ -891,7 +3374,7 @@ async function routeCheckpoint(request, state, result, agent, input) {
|
|
|
891
3374
|
next: { ...baseContinuation(result), recon_assessment_json: jsonParam(assessment.payload) }
|
|
892
3375
|
};
|
|
893
3376
|
}
|
|
894
|
-
const continueStage =
|
|
3377
|
+
const continueStage = checkpointContinueStage2(result);
|
|
895
3378
|
const checkpointContinuesToAuthor = continueStage === "author";
|
|
896
3379
|
if (checkpoint === "author_supervisor_judgment" || checkpoint === "verify_capture_retry" || checkpoint === "verify_agent_retry" && checkpointContinuesToAuthor) {
|
|
897
3380
|
const packet = await agent.authorProofPacket(context);
|