@riddledc/riddle-proof 0.5.45 → 0.5.46
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 +17 -0
- package/dist/checkpoint.cjs +204 -1
- package/dist/checkpoint.d.cts +16 -1
- package/dist/checkpoint.d.ts +16 -1
- package/dist/checkpoint.js +4 -2
- package/dist/{chunk-W7VTDN4T.js → chunk-DUFDZJOF.js} +1 -0
- package/dist/{chunk-7S7O3NKF.js → chunk-JTNMEH57.js} +6 -1
- package/dist/{chunk-MRSYJMF4.js → chunk-L26NTZOU.js} +2 -2
- package/dist/chunk-NOBFZDZG.js +754 -0
- package/dist/chunk-PLSGW2GI.js +161 -0
- package/dist/{chunk-LXE5YUYY.js → chunk-R6SCWJCI.js} +204 -2
- package/dist/{chunk-CHRYLX6N.js → chunk-X3AQ2WUM.js} +128 -15
- package/dist/cli.cjs +6129 -0
- package/dist/cli.d.cts +1 -0
- package/dist/cli.d.ts +1 -0
- package/dist/cli.js +160 -0
- package/dist/codex-exec-agent.cjs +790 -0
- package/dist/codex-exec-agent.d.cts +91 -0
- package/dist/codex-exec-agent.d.ts +91 -0
- package/dist/codex-exec-agent.js +10 -0
- package/dist/engine-harness.cjs +740 -279
- package/dist/engine-harness.js +5 -4
- package/dist/index.cjs +1469 -247
- package/dist/index.d.cts +4 -2
- package/dist/index.d.ts +4 -2
- package/dist/index.js +33 -17
- package/dist/openclaw.js +4 -2
- package/dist/proof-run-engine.d.cts +3 -3
- package/dist/proof-run-engine.d.ts +3 -3
- package/dist/result.cjs +1 -0
- package/dist/result.js +1 -1
- package/dist/run-card.cjs +212 -0
- package/dist/run-card.d.cts +10 -0
- package/dist/run-card.d.ts +10 -0
- package/dist/run-card.js +10 -0
- package/dist/runner.cjs +2 -0
- package/dist/runner.js +5 -3
- package/dist/state.cjs +174 -5
- package/dist/state.d.cts +2 -1
- package/dist/state.d.ts +2 -1
- package/dist/state.js +4 -2
- package/dist/types.d.cts +63 -2
- package/dist/types.d.ts +63 -2
- package/package.json +15 -2
package/dist/index.cjs
CHANGED
|
@@ -2761,6 +2761,7 @@ __export(index_exports, {
|
|
|
2761
2761
|
RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION: () => RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION,
|
|
2762
2762
|
RIDDLE_PROOF_PLAYABILITY_ASSESSMENT_VERSION: () => RIDDLE_PROOF_PLAYABILITY_ASSESSMENT_VERSION,
|
|
2763
2763
|
RIDDLE_PROOF_PLAYABILITY_VERSION: () => RIDDLE_PROOF_PLAYABILITY_VERSION,
|
|
2764
|
+
RIDDLE_PROOF_RUN_CARD_VERSION: () => RIDDLE_PROOF_RUN_CARD_VERSION,
|
|
2764
2765
|
RIDDLE_PROOF_RUN_STATE_VERSION: () => RIDDLE_PROOF_RUN_STATE_VERSION,
|
|
2765
2766
|
RIDDLE_PROOF_VISUAL_SESSION_FINGERPRINT_VERSION: () => RIDDLE_PROOF_VISUAL_SESSION_FINGERPRINT_VERSION,
|
|
2766
2767
|
RIDDLE_PROOF_VISUAL_SESSION_VERSION: () => RIDDLE_PROOF_VISUAL_SESSION_VERSION,
|
|
@@ -2774,13 +2775,17 @@ __export(index_exports, {
|
|
|
2774
2775
|
buildAuthorCheckpointPacket: () => buildAuthorCheckpointPacket,
|
|
2775
2776
|
buildCheckpointPacketForEngineResult: () => buildCheckpointPacketForEngineResult,
|
|
2776
2777
|
buildProofAssessmentCheckpointPacket: () => buildProofAssessmentCheckpointPacket,
|
|
2778
|
+
buildStageCheckpointPacket: () => buildStageCheckpointPacket,
|
|
2777
2779
|
buildVisualProofSession: () => buildVisualProofSession,
|
|
2778
2780
|
checkpointResponseIdentity: () => checkpointResponseIdentity,
|
|
2779
2781
|
checkpointSummaryFromState: () => checkpointSummaryFromState,
|
|
2780
2782
|
compactRecord: () => compactRecord,
|
|
2781
2783
|
compareVisualProofSessionFingerprint: () => compareVisualProofSessionFingerprint,
|
|
2782
2784
|
createCaptureDiagnostic: () => createCaptureDiagnostic,
|
|
2785
|
+
createCodexExecAgentAdapter: () => createCodexExecAgentAdapter,
|
|
2786
|
+
createCodexExecJsonRunner: () => createCodexExecJsonRunner,
|
|
2783
2787
|
createDisabledRiddleProofAgentAdapter: () => createDisabledRiddleProofAgentAdapter,
|
|
2788
|
+
createRiddleProofRunCard: () => createRiddleProofRunCard,
|
|
2784
2789
|
createRunResult: () => createRunResult,
|
|
2785
2790
|
createRunState: () => createRunState,
|
|
2786
2791
|
createRunStatusSnapshot: () => createRunStatusSnapshot,
|
|
@@ -2800,6 +2805,7 @@ __export(index_exports, {
|
|
|
2800
2805
|
readRiddleProofRunStatus: () => readRiddleProofRunStatus,
|
|
2801
2806
|
recordValue: () => recordValue,
|
|
2802
2807
|
redactForProofDiagnostics: () => redactForProofDiagnostics,
|
|
2808
|
+
runCodexExecAgentDoctor: () => runCodexExecAgentDoctor,
|
|
2803
2809
|
runRiddleProof: () => runRiddleProof,
|
|
2804
2810
|
runRiddleProofEngineHarness: () => runRiddleProofEngineHarness,
|
|
2805
2811
|
setRunStatus: () => setRunStatus,
|
|
@@ -3018,248 +3024,18 @@ function createRunResult(input) {
|
|
|
3018
3024
|
checkpoint_summary: state.checkpoint_summary,
|
|
3019
3025
|
state_paths: state.state_paths,
|
|
3020
3026
|
proof_contract: state.proof_contract,
|
|
3027
|
+
run_card: state.run_card,
|
|
3021
3028
|
proof_session: state.proof_session,
|
|
3022
3029
|
evidence_bundle: input.evidence_bundle,
|
|
3023
3030
|
raw: input.raw
|
|
3024
3031
|
});
|
|
3025
3032
|
}
|
|
3026
3033
|
|
|
3027
|
-
// src/state.ts
|
|
3028
|
-
var RIDDLE_PROOF_RUN_STATE_VERSION = "riddle-proof.run-state.v1";
|
|
3029
|
-
function timestamp() {
|
|
3030
|
-
return (/* @__PURE__ */ new Date()).toISOString();
|
|
3031
|
-
}
|
|
3032
|
-
function createRunId(createdAt) {
|
|
3033
|
-
const stamp = createdAt.replace(/\D/g, "").slice(0, 14) || "unknown";
|
|
3034
|
-
const entropy = Math.random().toString(36).slice(2, 8) || "run";
|
|
3035
|
-
return `rp_${stamp}_${entropy}`;
|
|
3036
|
-
}
|
|
3037
|
-
function elapsedMs(start, end) {
|
|
3038
|
-
const startMs = start ? Date.parse(start) : NaN;
|
|
3039
|
-
const endMs = end ? Date.parse(end) : NaN;
|
|
3040
|
-
if (!Number.isFinite(startMs) || !Number.isFinite(endMs)) return void 0;
|
|
3041
|
-
return Math.max(0, endMs - startMs);
|
|
3042
|
-
}
|
|
3043
|
-
function normalizePrStatus2(value) {
|
|
3044
|
-
const status = nonEmptyString(value)?.toLowerCase();
|
|
3045
|
-
if (status === "merged") return "merged";
|
|
3046
|
-
if (status === "open") return "open";
|
|
3047
|
-
if (status === "closed") return "closed";
|
|
3048
|
-
if (status === "not_found" || status === "not-found") return "not_found";
|
|
3049
|
-
if (status === "unavailable") return "unavailable";
|
|
3050
|
-
return status || "unknown";
|
|
3051
|
-
}
|
|
3052
|
-
function normalizePrLifecycleState(input, checkedAt = timestamp()) {
|
|
3053
|
-
const value = recordValue(input);
|
|
3054
|
-
if (!value) return void 0;
|
|
3055
|
-
const cleanup = recordValue(value.cleanup);
|
|
3056
|
-
const mergeCommit = nonEmptyString(value.merge_commit) || nonEmptyString(value.mergeCommit) || nonEmptyString(recordValue(value.mergeCommit)?.oid);
|
|
3057
|
-
return compactRecord({
|
|
3058
|
-
status: normalizePrStatus2(value.status || value.state),
|
|
3059
|
-
pr_url: nonEmptyString(value.pr_url) || nonEmptyString(value.prUrl) || nonEmptyString(value.url),
|
|
3060
|
-
pr_number: nonEmptyString(value.pr_number) || nonEmptyString(value.prNumber) || (typeof value.number === "number" ? String(value.number) : void 0),
|
|
3061
|
-
repo: nonEmptyString(value.repo) || nonEmptyString(value.repository),
|
|
3062
|
-
head_branch: nonEmptyString(value.head_branch) || nonEmptyString(value.headBranch) || nonEmptyString(value.headRefName),
|
|
3063
|
-
base_branch: nonEmptyString(value.base_branch) || nonEmptyString(value.baseBranch) || nonEmptyString(value.baseRefName),
|
|
3064
|
-
merge_commit: mergeCommit,
|
|
3065
|
-
merged_at: nonEmptyString(value.merged_at) || nonEmptyString(value.mergedAt),
|
|
3066
|
-
closed_at: nonEmptyString(value.closed_at) || nonEmptyString(value.closedAt),
|
|
3067
|
-
checked_at: nonEmptyString(value.checked_at) || nonEmptyString(value.checkedAt) || checkedAt,
|
|
3068
|
-
source: nonEmptyString(value.source),
|
|
3069
|
-
next_action: nonEmptyString(value.next_action) || nonEmptyString(value.nextAction),
|
|
3070
|
-
cleanup: cleanup && Object.keys(cleanup).length ? cleanup : void 0
|
|
3071
|
-
});
|
|
3072
|
-
}
|
|
3073
|
-
function normalizeIntegrationContext(input, fallbackSource) {
|
|
3074
|
-
const value = recordValue(input);
|
|
3075
|
-
if (!value) {
|
|
3076
|
-
return fallbackSource ? { source: fallbackSource } : void 0;
|
|
3077
|
-
}
|
|
3078
|
-
const metadata = recordValue(value.metadata);
|
|
3079
|
-
return compactRecord({
|
|
3080
|
-
source: nonEmptyString(value.source) || fallbackSource,
|
|
3081
|
-
channel_id: nonEmptyString(value.channel_id),
|
|
3082
|
-
thread_id: nonEmptyString(value.thread_id),
|
|
3083
|
-
message_id: nonEmptyString(value.message_id),
|
|
3084
|
-
source_url: nonEmptyString(value.source_url),
|
|
3085
|
-
metadata: metadata && Object.keys(metadata).length ? metadata : void 0
|
|
3086
|
-
});
|
|
3087
|
-
}
|
|
3088
|
-
function normalizeRunParams(input) {
|
|
3089
|
-
return compactRecord({
|
|
3090
|
-
repo: input.repo,
|
|
3091
|
-
branch: input.branch,
|
|
3092
|
-
change_request: input.change_request,
|
|
3093
|
-
commit_message: input.commit_message,
|
|
3094
|
-
prod_url: input.prod_url,
|
|
3095
|
-
capture_script: input.capture_script,
|
|
3096
|
-
success_criteria: input.success_criteria,
|
|
3097
|
-
assertions: input.assertions,
|
|
3098
|
-
verification_mode: input.verification_mode,
|
|
3099
|
-
resume_session: input.resume_session,
|
|
3100
|
-
target_image_url: input.target_image_url,
|
|
3101
|
-
target_image_hash: input.target_image_hash,
|
|
3102
|
-
viewport_matrix: input.viewport_matrix,
|
|
3103
|
-
deterministic_setup: input.deterministic_setup,
|
|
3104
|
-
reference: input.reference,
|
|
3105
|
-
base_branch: input.base_branch,
|
|
3106
|
-
before_ref: input.before_ref,
|
|
3107
|
-
allow_static_preview_fallback: input.allow_static_preview_fallback,
|
|
3108
|
-
context: input.context,
|
|
3109
|
-
reviewer: input.reviewer,
|
|
3110
|
-
mode: input.mode,
|
|
3111
|
-
build_command: input.build_command,
|
|
3112
|
-
build_output: input.build_output,
|
|
3113
|
-
server_image: input.server_image,
|
|
3114
|
-
server_command: input.server_command,
|
|
3115
|
-
server_port: input.server_port,
|
|
3116
|
-
server_path: input.server_path,
|
|
3117
|
-
use_auth: input.use_auth,
|
|
3118
|
-
auth_localStorage_json: input.auth_localStorage_json,
|
|
3119
|
-
auth_cookies_json: input.auth_cookies_json,
|
|
3120
|
-
auth_headers_json: input.auth_headers_json,
|
|
3121
|
-
color_scheme: input.color_scheme,
|
|
3122
|
-
wait_for_selector: input.wait_for_selector,
|
|
3123
|
-
ship_mode: input.ship_mode,
|
|
3124
|
-
leave_draft: input.leave_draft,
|
|
3125
|
-
engine_state_path: input.engine_state_path,
|
|
3126
|
-
harness_state_path: input.harness_state_path,
|
|
3127
|
-
max_iterations: input.max_iterations,
|
|
3128
|
-
auto_approve: input.auto_approve,
|
|
3129
|
-
dry_run: input.dry_run,
|
|
3130
|
-
integration_context: normalizeIntegrationContext(input.integration_context)
|
|
3131
|
-
});
|
|
3132
|
-
}
|
|
3133
|
-
function createRunState(input) {
|
|
3134
|
-
const createdAt = input.created_at || timestamp();
|
|
3135
|
-
return compactRecord({
|
|
3136
|
-
version: RIDDLE_PROOF_RUN_STATE_VERSION,
|
|
3137
|
-
run_id: input.run_id || createRunId(createdAt),
|
|
3138
|
-
state_path: input.state_path,
|
|
3139
|
-
worktree_path: input.worktree_path,
|
|
3140
|
-
branch: input.branch || input.request.branch,
|
|
3141
|
-
current_stage: input.current_stage ?? null,
|
|
3142
|
-
stage_started_at: input.stage_started_at ?? null,
|
|
3143
|
-
status: input.status || "running",
|
|
3144
|
-
created_at: createdAt,
|
|
3145
|
-
updated_at: input.updated_at || createdAt,
|
|
3146
|
-
request: normalizeRunParams(input.request),
|
|
3147
|
-
iterations: input.iterations ?? 0,
|
|
3148
|
-
last_checkpoint: input.last_checkpoint ?? null,
|
|
3149
|
-
checkpoint_packet: input.checkpoint_packet,
|
|
3150
|
-
checkpoint_summary: input.checkpoint_summary,
|
|
3151
|
-
state_paths: input.state_paths,
|
|
3152
|
-
proof_contract: input.proof_contract,
|
|
3153
|
-
checkpoint_history: input.checkpoint_history,
|
|
3154
|
-
events: input.events ? [...input.events] : []
|
|
3155
|
-
});
|
|
3156
|
-
}
|
|
3157
|
-
function appendRunEvent(state, input) {
|
|
3158
|
-
const event = {
|
|
3159
|
-
ts: input.ts || timestamp(),
|
|
3160
|
-
kind: input.kind,
|
|
3161
|
-
checkpoint: input.checkpoint,
|
|
3162
|
-
stage: input.stage,
|
|
3163
|
-
summary: input.summary,
|
|
3164
|
-
details: input.details
|
|
3165
|
-
};
|
|
3166
|
-
state.events.push(compactRecord({
|
|
3167
|
-
ts: event.ts,
|
|
3168
|
-
kind: event.kind,
|
|
3169
|
-
checkpoint: event.checkpoint,
|
|
3170
|
-
stage: event.stage,
|
|
3171
|
-
summary: event.summary,
|
|
3172
|
-
details: event.details
|
|
3173
|
-
}));
|
|
3174
|
-
if (input.checkpoint !== void 0) state.last_checkpoint = input.checkpoint;
|
|
3175
|
-
if (input.stage !== void 0) {
|
|
3176
|
-
if (state.current_stage !== input.stage) state.stage_started_at = event.ts;
|
|
3177
|
-
state.current_stage = input.stage;
|
|
3178
|
-
}
|
|
3179
|
-
state.updated_at = event.ts;
|
|
3180
|
-
return state;
|
|
3181
|
-
}
|
|
3182
|
-
function appendStageHeartbeat(state, input) {
|
|
3183
|
-
const at = input.ts || timestamp();
|
|
3184
|
-
return appendRunEvent(state, {
|
|
3185
|
-
ts: at,
|
|
3186
|
-
kind: "stage.heartbeat",
|
|
3187
|
-
checkpoint: input.checkpoint || `${input.stage}_heartbeat`,
|
|
3188
|
-
stage: input.stage,
|
|
3189
|
-
summary: input.summary || `${input.stage} stage is active.`,
|
|
3190
|
-
details: compactRecord({
|
|
3191
|
-
elapsed_ms: elapsedMs(state.created_at, at),
|
|
3192
|
-
stage_elapsed_ms: elapsedMs(state.stage_started_at, at),
|
|
3193
|
-
wait_reason: input.wait_reason,
|
|
3194
|
-
blocker: input.blocker,
|
|
3195
|
-
...input.details
|
|
3196
|
-
})
|
|
3197
|
-
});
|
|
3198
|
-
}
|
|
3199
|
-
function createRunStatusSnapshot(state, at = timestamp()) {
|
|
3200
|
-
const latestEvent = state.events[state.events.length - 1];
|
|
3201
|
-
const runId = state.run_id || "unknown";
|
|
3202
|
-
return compactRecord({
|
|
3203
|
-
run_id: runId,
|
|
3204
|
-
status: state.status,
|
|
3205
|
-
is_terminal: isTerminalStatus(state.status),
|
|
3206
|
-
monitor_should_continue: !isTerminalStatus(state.status),
|
|
3207
|
-
current_stage: state.current_stage ?? null,
|
|
3208
|
-
state_path: state.state_path ?? null,
|
|
3209
|
-
worktree_path: state.worktree_path ?? null,
|
|
3210
|
-
branch: state.branch ?? null,
|
|
3211
|
-
pr_url: state.pr_url ?? null,
|
|
3212
|
-
pr_branch: state.pr_branch ?? null,
|
|
3213
|
-
pr_state: state.pr_state,
|
|
3214
|
-
ci_status: state.ci_status,
|
|
3215
|
-
ship_commit: state.ship_commit,
|
|
3216
|
-
ship_remote_head: state.ship_remote_head,
|
|
3217
|
-
merge_commit: state.merge_commit,
|
|
3218
|
-
merged_at: state.merged_at,
|
|
3219
|
-
proof_comment_url: state.proof_comment_url,
|
|
3220
|
-
cleanup_report: state.cleanup_report,
|
|
3221
|
-
iterations: state.iterations,
|
|
3222
|
-
last_checkpoint: state.last_checkpoint ?? null,
|
|
3223
|
-
updated_at: state.updated_at,
|
|
3224
|
-
elapsed_ms: elapsedMs(state.created_at, at),
|
|
3225
|
-
stage_elapsed_ms: elapsedMs(state.stage_started_at, at),
|
|
3226
|
-
blocker: state.blocker,
|
|
3227
|
-
checkpoint_packet: state.checkpoint_packet,
|
|
3228
|
-
checkpoint_summary: state.checkpoint_summary,
|
|
3229
|
-
state_paths: state.state_paths,
|
|
3230
|
-
proof_contract: state.proof_contract,
|
|
3231
|
-
latest_event: latestEvent
|
|
3232
|
-
});
|
|
3233
|
-
}
|
|
3234
|
-
function setRunStatus(state, status, at = timestamp()) {
|
|
3235
|
-
state.status = status;
|
|
3236
|
-
state.ok = status !== "blocked" && status !== "failed";
|
|
3237
|
-
state.updated_at = at;
|
|
3238
|
-
return state;
|
|
3239
|
-
}
|
|
3240
|
-
function applyPrLifecycleState(state, input, at = timestamp()) {
|
|
3241
|
-
const prState = normalizePrLifecycleState(input, at);
|
|
3242
|
-
if (!prState) return state;
|
|
3243
|
-
state.pr_state = prState;
|
|
3244
|
-
if (prState.pr_url) state.pr_url = prState.pr_url;
|
|
3245
|
-
if (prState.head_branch) state.pr_branch = prState.head_branch;
|
|
3246
|
-
if (prState.merge_commit) state.merge_commit = prState.merge_commit;
|
|
3247
|
-
if (prState.merged_at) state.merged_at = prState.merged_at;
|
|
3248
|
-
if (prState.cleanup) state.cleanup_report = prState.cleanup;
|
|
3249
|
-
if (prState.status === "merged") {
|
|
3250
|
-
state.finalized = true;
|
|
3251
|
-
state.status = "completed";
|
|
3252
|
-
state.ok = true;
|
|
3253
|
-
}
|
|
3254
|
-
state.updated_at = at;
|
|
3255
|
-
return state;
|
|
3256
|
-
}
|
|
3257
|
-
|
|
3258
3034
|
// src/checkpoint.ts
|
|
3259
3035
|
var import_node_crypto = __toESM(require("crypto"), 1);
|
|
3260
3036
|
var RIDDLE_PROOF_CHECKPOINT_PACKET_VERSION = "riddle-proof.checkpoint.v1";
|
|
3261
3037
|
var RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION = "riddle-proof.checkpoint_response.v1";
|
|
3262
|
-
function
|
|
3038
|
+
function timestamp() {
|
|
3263
3039
|
return (/* @__PURE__ */ new Date()).toISOString();
|
|
3264
3040
|
}
|
|
3265
3041
|
function jsonCloneRecord(value) {
|
|
@@ -3373,6 +3149,102 @@ function responseSchemaForProofAssessmentPacket() {
|
|
|
3373
3149
|
}
|
|
3374
3150
|
};
|
|
3375
3151
|
}
|
|
3152
|
+
function responseSchemaForReconPacket() {
|
|
3153
|
+
return {
|
|
3154
|
+
type: "object",
|
|
3155
|
+
required: ["version", "run_id", "checkpoint", "decision", "summary", "created_at"],
|
|
3156
|
+
additionalProperties: false,
|
|
3157
|
+
properties: {
|
|
3158
|
+
version: { const: RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION },
|
|
3159
|
+
run_id: { type: "string" },
|
|
3160
|
+
checkpoint: { type: "string" },
|
|
3161
|
+
resume_token: { type: "string" },
|
|
3162
|
+
decision: {
|
|
3163
|
+
type: "string",
|
|
3164
|
+
enum: ["ready_for_author", "retry_recon", "recon_stuck", "needs_recon", "blocked", "human_review"]
|
|
3165
|
+
},
|
|
3166
|
+
summary: { type: "string" },
|
|
3167
|
+
payload: {
|
|
3168
|
+
type: "object",
|
|
3169
|
+
description: "Optional recon assessment details such as baseline_understanding, refined_inputs, reasons, or diagnostic blocker context."
|
|
3170
|
+
},
|
|
3171
|
+
reasons: { type: "array", items: { type: "string" } },
|
|
3172
|
+
continue_with_stage: { type: "string", enum: ["recon", "author"] },
|
|
3173
|
+
source: {
|
|
3174
|
+
type: "object",
|
|
3175
|
+
properties: {
|
|
3176
|
+
kind: { type: "string" },
|
|
3177
|
+
session_id: { type: "string" },
|
|
3178
|
+
user_id: { type: "string" }
|
|
3179
|
+
}
|
|
3180
|
+
},
|
|
3181
|
+
created_at: { type: "string" }
|
|
3182
|
+
}
|
|
3183
|
+
};
|
|
3184
|
+
}
|
|
3185
|
+
function responseSchemaForImplementationPacket() {
|
|
3186
|
+
return {
|
|
3187
|
+
type: "object",
|
|
3188
|
+
required: ["version", "run_id", "checkpoint", "decision", "summary", "created_at"],
|
|
3189
|
+
additionalProperties: false,
|
|
3190
|
+
properties: {
|
|
3191
|
+
version: { const: RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION },
|
|
3192
|
+
run_id: { type: "string" },
|
|
3193
|
+
checkpoint: { type: "string" },
|
|
3194
|
+
resume_token: { type: "string" },
|
|
3195
|
+
decision: {
|
|
3196
|
+
type: "string",
|
|
3197
|
+
enum: ["implementation_complete", "needs_author", "needs_recon", "blocked", "human_review"]
|
|
3198
|
+
},
|
|
3199
|
+
summary: { type: "string" },
|
|
3200
|
+
payload: {
|
|
3201
|
+
type: "object",
|
|
3202
|
+
description: "Implementation details such as changed_files, tests_run, and implementation_notes. The changed worktree must contain a real git diff before verify can advance."
|
|
3203
|
+
},
|
|
3204
|
+
reasons: { type: "array", items: { type: "string" } },
|
|
3205
|
+
continue_with_stage: { type: "string", enum: ["implement", "verify", "author", "recon"] },
|
|
3206
|
+
source: {
|
|
3207
|
+
type: "object",
|
|
3208
|
+
properties: {
|
|
3209
|
+
kind: { type: "string" },
|
|
3210
|
+
session_id: { type: "string" },
|
|
3211
|
+
user_id: { type: "string" }
|
|
3212
|
+
}
|
|
3213
|
+
},
|
|
3214
|
+
created_at: { type: "string" }
|
|
3215
|
+
}
|
|
3216
|
+
};
|
|
3217
|
+
}
|
|
3218
|
+
function responseSchemaForAdvancePacket(stage) {
|
|
3219
|
+
return {
|
|
3220
|
+
type: "object",
|
|
3221
|
+
required: ["version", "run_id", "checkpoint", "decision", "summary", "created_at"],
|
|
3222
|
+
additionalProperties: false,
|
|
3223
|
+
properties: {
|
|
3224
|
+
version: { const: RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION },
|
|
3225
|
+
run_id: { type: "string" },
|
|
3226
|
+
checkpoint: { type: "string" },
|
|
3227
|
+
resume_token: { type: "string" },
|
|
3228
|
+
decision: {
|
|
3229
|
+
type: "string",
|
|
3230
|
+
enum: ["continue_stage", "retry_stage", "needs_recon", "needs_implementation", "blocked", "human_review"]
|
|
3231
|
+
},
|
|
3232
|
+
summary: { type: "string" },
|
|
3233
|
+
payload: { type: "object" },
|
|
3234
|
+
reasons: { type: "array", items: { type: "string" } },
|
|
3235
|
+
continue_with_stage: { type: "string", enum: ["recon", "author", "implement", "verify", "ship", stage] },
|
|
3236
|
+
source: {
|
|
3237
|
+
type: "object",
|
|
3238
|
+
properties: {
|
|
3239
|
+
kind: { type: "string" },
|
|
3240
|
+
session_id: { type: "string" },
|
|
3241
|
+
user_id: { type: "string" }
|
|
3242
|
+
}
|
|
3243
|
+
},
|
|
3244
|
+
created_at: { type: "string" }
|
|
3245
|
+
}
|
|
3246
|
+
};
|
|
3247
|
+
}
|
|
3376
3248
|
function resumeTokenFor(input) {
|
|
3377
3249
|
const hash = import_node_crypto.default.createHash("sha256").update(JSON.stringify(input)).digest("hex").slice(0, 24);
|
|
3378
3250
|
return `rpchk_${hash}`;
|
|
@@ -3400,6 +3272,106 @@ function artifactsFromState(state) {
|
|
|
3400
3272
|
}
|
|
3401
3273
|
return artifacts.slice(0, 16);
|
|
3402
3274
|
}
|
|
3275
|
+
function stageFromCheckpoint(checkpoint) {
|
|
3276
|
+
if (checkpoint.startsWith("recon_")) return "recon";
|
|
3277
|
+
if (checkpoint.startsWith("author_")) return "author";
|
|
3278
|
+
if (checkpoint.startsWith("implement_")) return "implement";
|
|
3279
|
+
if (checkpoint.startsWith("verify_")) return "verify";
|
|
3280
|
+
if (checkpoint.startsWith("ship_")) return "ship";
|
|
3281
|
+
if (checkpoint.startsWith("pr_sync_")) return "notify";
|
|
3282
|
+
if (checkpoint.includes("capture")) return "prove";
|
|
3283
|
+
return "setup";
|
|
3284
|
+
}
|
|
3285
|
+
function allowedDecisionsForStage(stage, checkpoint) {
|
|
3286
|
+
if (stage === "recon") return ["ready_for_author", "retry_recon", "recon_stuck", "blocked", "human_review"];
|
|
3287
|
+
if (stage === "implement") return ["implementation_complete", "needs_author", "needs_recon", "blocked", "human_review"];
|
|
3288
|
+
if (stage === "ship") return ["continue_stage", "needs_implementation", "needs_recon", "blocked", "human_review"];
|
|
3289
|
+
if (checkpoint === "awaiting_stage_advance") return ["continue_stage", "retry_stage", "blocked", "human_review"];
|
|
3290
|
+
return ["continue_stage", "needs_recon", "needs_implementation", "blocked", "human_review"];
|
|
3291
|
+
}
|
|
3292
|
+
function packetKindForStage(stage, checkpoint) {
|
|
3293
|
+
if (stage === "recon") return "assess_recon";
|
|
3294
|
+
if (stage === "implement") return "implement_change";
|
|
3295
|
+
if (checkpoint.includes("human")) return "human_review";
|
|
3296
|
+
return "advance_decision";
|
|
3297
|
+
}
|
|
3298
|
+
function responseSchemaForStage(stage, checkpoint) {
|
|
3299
|
+
if (stage === "recon") return responseSchemaForReconPacket();
|
|
3300
|
+
if (stage === "implement") return responseSchemaForImplementationPacket();
|
|
3301
|
+
return responseSchemaForAdvancePacket(stage);
|
|
3302
|
+
}
|
|
3303
|
+
function questionForStage(stage, checkpoint) {
|
|
3304
|
+
if (stage === "recon") {
|
|
3305
|
+
return "Assess the baseline/recon evidence. Return ready_for_author only when the baseline is trustworthy; otherwise choose retry_recon, recon_stuck, blocked, or human_review.";
|
|
3306
|
+
}
|
|
3307
|
+
if (stage === "implement") {
|
|
3308
|
+
return "Implement the requested change in the after worktree, leave a real git diff, then return implementation_complete with changed_files/tests_run/implementation_notes. Choose blocked or human_review if implementation cannot honestly advance.";
|
|
3309
|
+
}
|
|
3310
|
+
if (stage === "ship") {
|
|
3311
|
+
return "Assess whether the ship gate can continue. Return continue_stage only if proof, PR, and policy gates are satisfied; otherwise route to the appropriate earlier stage or block with a concrete reason.";
|
|
3312
|
+
}
|
|
3313
|
+
return "Choose the next Riddle Proof stage for this durable run, or block with a concrete reason if the run cannot honestly advance.";
|
|
3314
|
+
}
|
|
3315
|
+
function buildStageCheckpointPacket(input) {
|
|
3316
|
+
const checkpoint = nonEmptyString(input.engineResult.checkpoint) || "stage_checkpoint";
|
|
3317
|
+
const stage = nonEmptyString(input.engineResult.stage) || stageFromCheckpoint(checkpoint);
|
|
3318
|
+
const runId = input.runState.run_id || "unknown";
|
|
3319
|
+
const fullState = input.fullRiddleState || {};
|
|
3320
|
+
const decisionDetails = recordValue(input.engineResult.decisionRequest?.details);
|
|
3321
|
+
const checkpointContract = recordValue(input.engineResult.checkpointContract);
|
|
3322
|
+
const summary = nonEmptyString(input.engineResult.summary) || nonEmptyString(fullState.stage_summary) || `${stage} checkpoint needs a supervising decision.`;
|
|
3323
|
+
const kind = packetKindForStage(stage, checkpoint);
|
|
3324
|
+
return {
|
|
3325
|
+
version: RIDDLE_PROOF_CHECKPOINT_PACKET_VERSION,
|
|
3326
|
+
run_id: runId,
|
|
3327
|
+
state_path: input.runState.state_path,
|
|
3328
|
+
stage,
|
|
3329
|
+
checkpoint,
|
|
3330
|
+
kind,
|
|
3331
|
+
summary,
|
|
3332
|
+
question: questionForStage(stage, checkpoint),
|
|
3333
|
+
change_request: input.request.change_request || nonEmptyString(fullState.change_request) || "",
|
|
3334
|
+
context: input.request.context,
|
|
3335
|
+
artifacts: artifactsFromState(fullState),
|
|
3336
|
+
state_excerpt: compactRecord({
|
|
3337
|
+
repo: input.request.repo || fullState.repo,
|
|
3338
|
+
branch: input.request.branch || fullState.branch,
|
|
3339
|
+
after_worktree: fullState.after_worktree || input.runState.worktree_path,
|
|
3340
|
+
verification_mode: input.request.verification_mode || fullState.verification_mode,
|
|
3341
|
+
reference: input.request.reference || fullState.reference,
|
|
3342
|
+
server_path: fullState.server_path,
|
|
3343
|
+
wait_for_selector: fullState.wait_for_selector,
|
|
3344
|
+
recon_status: fullState.recon_status,
|
|
3345
|
+
author_status: fullState.author_status,
|
|
3346
|
+
implementation_status: fullState.implementation_status,
|
|
3347
|
+
verify_status: fullState.verify_status,
|
|
3348
|
+
stage_decision_request: jsonCloneRecord(fullState.stage_decision_request)
|
|
3349
|
+
}),
|
|
3350
|
+
evidence_excerpt: compactRecord({
|
|
3351
|
+
before_cdn: fullState.before_cdn || null,
|
|
3352
|
+
prod_cdn: fullState.prod_cdn || null,
|
|
3353
|
+
after_cdn: fullState.after_cdn || null,
|
|
3354
|
+
recon_results: jsonCloneRecord(fullState.recon_results),
|
|
3355
|
+
checkpoint_contract: jsonCloneRecord(checkpointContract),
|
|
3356
|
+
decision_details: jsonCloneRecord(decisionDetails)
|
|
3357
|
+
}),
|
|
3358
|
+
allowed_decisions: allowedDecisionsForStage(stage, checkpoint),
|
|
3359
|
+
response_schema: responseSchemaForStage(stage, checkpoint),
|
|
3360
|
+
routing_hint: {
|
|
3361
|
+
suggested_role: stage === "implement" ? "builder_agent" : checkpoint.includes("human") ? "human" : "main_agent",
|
|
3362
|
+
visibility: input.visibility || "quiet",
|
|
3363
|
+
urgency: stage === "ship" ? "high" : "normal",
|
|
3364
|
+
can_auto_answer: input.visibility !== "manual"
|
|
3365
|
+
},
|
|
3366
|
+
resume_token: resumeTokenFor({
|
|
3367
|
+
runId,
|
|
3368
|
+
statePath: input.engineResult.state_path || input.request.engine_state_path || null,
|
|
3369
|
+
checkpoint,
|
|
3370
|
+
stage
|
|
3371
|
+
}),
|
|
3372
|
+
created_at: input.created_at || timestamp()
|
|
3373
|
+
};
|
|
3374
|
+
}
|
|
3403
3375
|
function buildAuthorCheckpointPacket(input) {
|
|
3404
3376
|
const checkpoint = nonEmptyString(input.engineResult.checkpoint) || "author_supervisor_judgment";
|
|
3405
3377
|
const stage = "author";
|
|
@@ -3453,7 +3425,7 @@ function buildAuthorCheckpointPacket(input) {
|
|
|
3453
3425
|
checkpoint,
|
|
3454
3426
|
stage
|
|
3455
3427
|
}),
|
|
3456
|
-
created_at: input.created_at ||
|
|
3428
|
+
created_at: input.created_at || timestamp()
|
|
3457
3429
|
};
|
|
3458
3430
|
}
|
|
3459
3431
|
function visualDeltaFromState(fullState) {
|
|
@@ -3502,16 +3474,16 @@ function buildProofAssessmentCheckpointPacket(input) {
|
|
|
3502
3474
|
const visualDelta = visualDeltaFromState(fullState);
|
|
3503
3475
|
const verificationMode = nonEmptyString(input.request.verification_mode) || nonEmptyString(fullState.verification_mode) || nonEmptyString(bundle?.verification_mode) || "proof";
|
|
3504
3476
|
const visualDeltaRequired = requiredSignals?.visual_delta === true || verificationModeRequiresVisualDelta(verificationMode);
|
|
3505
|
-
const
|
|
3477
|
+
const evidenceIssueCode2 = visualDeltaIssueCode(visualDelta, visualDeltaRequired);
|
|
3506
3478
|
const summary = nonEmptyString(input.engineResult.summary) || nonEmptyString(fullState.verify_summary) || "Verify captured evidence and needs a supervising proof assessment.";
|
|
3507
|
-
const recoveryHint =
|
|
3479
|
+
const recoveryHint = evidenceIssueCode2 ? "Required visual_delta evidence is incomplete. Keep this same run in verify/evidence recovery with decision=revise_capture and continue_with_stage=verify unless the evidence proves an implementation or recon problem." : "Assess whether the current artifacts prove the requested change, then choose the next stage.";
|
|
3508
3480
|
return {
|
|
3509
3481
|
version: RIDDLE_PROOF_CHECKPOINT_PACKET_VERSION,
|
|
3510
3482
|
run_id: runId,
|
|
3511
3483
|
state_path: input.runState.state_path,
|
|
3512
3484
|
stage,
|
|
3513
3485
|
checkpoint,
|
|
3514
|
-
kind:
|
|
3486
|
+
kind: evidenceIssueCode2 ? "recover_evidence" : "assess_proof",
|
|
3515
3487
|
summary,
|
|
3516
3488
|
question: `Assess the current Riddle Proof evidence. ${recoveryHint} Return a CheckpointResponse using one allowed decision.`,
|
|
3517
3489
|
change_request: input.request.change_request || nonEmptyString(fullState.change_request) || "",
|
|
@@ -3536,7 +3508,7 @@ function buildProofAssessmentCheckpointPacket(input) {
|
|
|
3536
3508
|
visual_delta_required: visualDeltaRequired,
|
|
3537
3509
|
visual_delta_ready: visualDelta?.status === "measured" && visualDelta?.passed === true,
|
|
3538
3510
|
visual_delta: jsonCloneRecord(visualDelta),
|
|
3539
|
-
evidence_issue_code:
|
|
3511
|
+
evidence_issue_code: evidenceIssueCode2,
|
|
3540
3512
|
proof_assessment_request: jsonCloneRecord(proofAssessmentRequest),
|
|
3541
3513
|
verify_decision_request: jsonCloneRecord(fullState.verify_decision_request),
|
|
3542
3514
|
checkpoint_contract: jsonCloneRecord(input.engineResult.checkpointContract)
|
|
@@ -3553,9 +3525,9 @@ function buildProofAssessmentCheckpointPacket(input) {
|
|
|
3553
3525
|
],
|
|
3554
3526
|
response_schema: responseSchemaForProofAssessmentPacket(),
|
|
3555
3527
|
routing_hint: {
|
|
3556
|
-
suggested_role:
|
|
3528
|
+
suggested_role: evidenceIssueCode2 ? "proof_judge" : "proof_judge",
|
|
3557
3529
|
visibility: input.visibility || "quiet",
|
|
3558
|
-
urgency:
|
|
3530
|
+
urgency: evidenceIssueCode2 ? "high" : "normal",
|
|
3559
3531
|
can_auto_answer: input.visibility !== "manual"
|
|
3560
3532
|
},
|
|
3561
3533
|
resume_token: resumeTokenFor({
|
|
@@ -3564,7 +3536,7 @@ function buildProofAssessmentCheckpointPacket(input) {
|
|
|
3564
3536
|
checkpoint,
|
|
3565
3537
|
stage
|
|
3566
3538
|
}),
|
|
3567
|
-
created_at: input.created_at ||
|
|
3539
|
+
created_at: input.created_at || timestamp()
|
|
3568
3540
|
};
|
|
3569
3541
|
}
|
|
3570
3542
|
function buildCheckpointPacketForEngineResult(input) {
|
|
@@ -3572,7 +3544,12 @@ function buildCheckpointPacketForEngineResult(input) {
|
|
|
3572
3544
|
if (checkpoint === "verify_supervisor_judgment" || checkpoint === "verify_supervisor_judgment_required" || checkpoint === "verify_human_escalation") {
|
|
3573
3545
|
return buildProofAssessmentCheckpointPacket(input);
|
|
3574
3546
|
}
|
|
3575
|
-
|
|
3547
|
+
const resume = recordValue(input.engineResult.checkpointContract?.resume);
|
|
3548
|
+
const continueStage = nonEmptyString(resume?.continue_with_stage);
|
|
3549
|
+
if (checkpoint === "author_supervisor_judgment" || checkpoint === "verify_capture_retry" || checkpoint === "verify_agent_retry" && continueStage === "author") {
|
|
3550
|
+
return buildAuthorCheckpointPacket(input);
|
|
3551
|
+
}
|
|
3552
|
+
return buildStageCheckpointPacket(input);
|
|
3576
3553
|
}
|
|
3577
3554
|
function normalizeCheckpointResponse(value) {
|
|
3578
3555
|
const record = recordValue(value);
|
|
@@ -3596,7 +3573,7 @@ function normalizeCheckpointResponse(value) {
|
|
|
3596
3573
|
reasons: Array.isArray(record.reasons) ? record.reasons.filter((item) => typeof item === "string") : void 0,
|
|
3597
3574
|
continue_with_stage: nonEmptyString(record.continue_with_stage),
|
|
3598
3575
|
source: jsonCloneRecord(record.source),
|
|
3599
|
-
created_at: nonEmptyString(record.created_at) ||
|
|
3576
|
+
created_at: nonEmptyString(record.created_at) || timestamp()
|
|
3600
3577
|
});
|
|
3601
3578
|
}
|
|
3602
3579
|
function checkpointSummaryFromState(state, engineStatePath2) {
|
|
@@ -3685,9 +3662,389 @@ function proofContractFromAuthorCheckpointResponse(response, packet, payload) {
|
|
|
3685
3662
|
rationale: jsonCloneValue(payload.rationale),
|
|
3686
3663
|
verdict_dimensions: jsonCloneRecord(payload.verdict_dimensions),
|
|
3687
3664
|
payload: jsonCloneRecord(payload),
|
|
3688
|
-
created_at:
|
|
3665
|
+
created_at: timestamp()
|
|
3666
|
+
});
|
|
3667
|
+
}
|
|
3668
|
+
|
|
3669
|
+
// src/run-card.ts
|
|
3670
|
+
var RIDDLE_PROOF_RUN_CARD_VERSION = "riddle-proof.run-card.v1";
|
|
3671
|
+
function elapsedMs(start, end) {
|
|
3672
|
+
const startMs = start ? Date.parse(start) : NaN;
|
|
3673
|
+
const endMs = end ? Date.parse(end) : NaN;
|
|
3674
|
+
if (!Number.isFinite(startMs) || !Number.isFinite(endMs)) return void 0;
|
|
3675
|
+
return Math.max(0, endMs - startMs);
|
|
3676
|
+
}
|
|
3677
|
+
function jsonCloneRecord2(value) {
|
|
3678
|
+
const record = recordValue(value);
|
|
3679
|
+
if (!record) return void 0;
|
|
3680
|
+
try {
|
|
3681
|
+
return JSON.parse(JSON.stringify(record));
|
|
3682
|
+
} catch {
|
|
3683
|
+
return { ...record };
|
|
3684
|
+
}
|
|
3685
|
+
}
|
|
3686
|
+
function compactText2(value, limit = 600) {
|
|
3687
|
+
const text = nonEmptyString(value);
|
|
3688
|
+
if (!text) return void 0;
|
|
3689
|
+
return text.length <= limit ? text : `${text.slice(0, limit - 20).trimEnd()}...`;
|
|
3690
|
+
}
|
|
3691
|
+
function visualDeltaFrom(input) {
|
|
3692
|
+
const fullState = input.fullRiddleState || {};
|
|
3693
|
+
const bundle = recordValue(fullState.evidence_bundle);
|
|
3694
|
+
const after = recordValue(bundle?.after);
|
|
3695
|
+
const afterDelta = recordValue(after?.visual_delta);
|
|
3696
|
+
if (afterDelta && Object.keys(afterDelta).length) return afterDelta;
|
|
3697
|
+
const requestDelta = recordValue(recordValue(fullState.proof_assessment_request)?.visual_delta);
|
|
3698
|
+
if (requestDelta && Object.keys(requestDelta).length) return requestDelta;
|
|
3699
|
+
const packetDelta = recordValue(input.runState.checkpoint_packet?.evidence_excerpt?.visual_delta);
|
|
3700
|
+
return packetDelta && Object.keys(packetDelta).length ? packetDelta : void 0;
|
|
3701
|
+
}
|
|
3702
|
+
function artifactsFrom(input) {
|
|
3703
|
+
const packetArtifacts = input.runState.checkpoint_packet?.artifacts;
|
|
3704
|
+
if (packetArtifacts?.length) return packetArtifacts.slice(0, 12);
|
|
3705
|
+
const fullState = input.fullRiddleState || {};
|
|
3706
|
+
const artifacts = [];
|
|
3707
|
+
for (const role of ["before", "prod", "after"]) {
|
|
3708
|
+
const url = nonEmptyString(fullState[`${role}_cdn`]) || nonEmptyString(input.runState[`${role}_artifact_url`]);
|
|
3709
|
+
if (url) artifacts.push({ role, url, name: `${role}.png`, mime_type: "image/png" });
|
|
3710
|
+
}
|
|
3711
|
+
return artifacts.length ? artifacts : void 0;
|
|
3712
|
+
}
|
|
3713
|
+
function ownerFor(state) {
|
|
3714
|
+
if (state.status === "awaiting_checkpoint") {
|
|
3715
|
+
const role = nonEmptyString(state.checkpoint_packet?.routing_hint?.suggested_role);
|
|
3716
|
+
return role || "supervising_agent";
|
|
3717
|
+
}
|
|
3718
|
+
if (state.status === "running") return "engine";
|
|
3719
|
+
if (state.status === "blocked" || state.status === "failed") return "human_or_operator";
|
|
3720
|
+
return "none";
|
|
3721
|
+
}
|
|
3722
|
+
function actionFor(state) {
|
|
3723
|
+
if (state.status === "awaiting_checkpoint") return state.checkpoint_packet?.question || "Answer the pending checkpoint.";
|
|
3724
|
+
if (state.status === "running") return "Continue the current Riddle Proof stage.";
|
|
3725
|
+
if (state.status === "blocked" || state.status === "failed") return state.blocker?.message || "Inspect the blocker and decide whether a new run or infrastructure repair is needed.";
|
|
3726
|
+
if (state.status === "ready_to_ship") return "Ship is held by policy; review PR/ship policy before advancing.";
|
|
3727
|
+
return "No next action; run is terminal.";
|
|
3728
|
+
}
|
|
3729
|
+
function evidenceIssueCode(input) {
|
|
3730
|
+
const packetIssue = nonEmptyString(input.runState.checkpoint_packet?.evidence_excerpt?.evidence_issue_code);
|
|
3731
|
+
if (packetIssue) return packetIssue;
|
|
3732
|
+
const assessmentIssue = nonEmptyString(recordValue(input.fullRiddleState?.proof_assessment)?.evidence_issue_code);
|
|
3733
|
+
if (assessmentIssue) return assessmentIssue;
|
|
3734
|
+
const delta = visualDeltaFrom(input);
|
|
3735
|
+
const status = nonEmptyString(delta?.status);
|
|
3736
|
+
if (status === "unmeasured") {
|
|
3737
|
+
const reason = `${nonEmptyString(delta?.reason) || ""}
|
|
3738
|
+
${input.runState.blocker?.message || ""}`.toLowerCase();
|
|
3739
|
+
return reason.includes("fetch") || reason.includes("allowlist") || reason.includes("registered domain") || reason.includes("high risk") || reason.includes("comparator") ? "comparator_fetch_blocked" : "visual_delta_unmeasured";
|
|
3740
|
+
}
|
|
3741
|
+
if (status === "measured" && delta?.passed === false) return "semantic_proof_failed";
|
|
3742
|
+
return void 0;
|
|
3743
|
+
}
|
|
3744
|
+
function createRiddleProofRunCard(state, input = {}) {
|
|
3745
|
+
const at = input.at || (/* @__PURE__ */ new Date()).toISOString();
|
|
3746
|
+
const fullState = input.fullRiddleState || {};
|
|
3747
|
+
const packet = state.checkpoint_packet;
|
|
3748
|
+
const latestEvent = state.events[state.events.length - 1];
|
|
3749
|
+
const bundle = recordValue(fullState.evidence_bundle);
|
|
3750
|
+
const artifactContract = jsonCloneRecord2(packet?.artifact_contract) || jsonCloneRecord2(recordValue(fullState.proof_assessment_request)?.artifact_contract) || jsonCloneRecord2(recordValue(bundle?.artifact_contract)) || jsonCloneRecord2(state.proof_contract?.artifact_contract);
|
|
3751
|
+
const required = jsonCloneRecord2(recordValue(artifactContract)?.required);
|
|
3752
|
+
const statePaths = input.state_paths || state.state_paths || statePathsForRunState(state);
|
|
3753
|
+
const visualDelta = visualDeltaFrom({ fullRiddleState: fullState, runState: state });
|
|
3754
|
+
const artifacts = artifactsFrom({ fullRiddleState: fullState, runState: state });
|
|
3755
|
+
return {
|
|
3756
|
+
version: RIDDLE_PROOF_RUN_CARD_VERSION,
|
|
3757
|
+
run_id: state.run_id || "unknown",
|
|
3758
|
+
status: state.status,
|
|
3759
|
+
goal: compactRecord({
|
|
3760
|
+
repo: state.request.repo || nonEmptyString(fullState.repo),
|
|
3761
|
+
branch: state.request.branch || nonEmptyString(fullState.branch),
|
|
3762
|
+
change_request: state.request.change_request || nonEmptyString(fullState.change_request),
|
|
3763
|
+
verification_mode: state.request.verification_mode || nonEmptyString(fullState.verification_mode),
|
|
3764
|
+
success_criteria: state.request.success_criteria
|
|
3765
|
+
}),
|
|
3766
|
+
durable_state: compactRecord({
|
|
3767
|
+
...statePaths,
|
|
3768
|
+
worktree_path: state.worktree_path || nonEmptyString(fullState.after_worktree) || null,
|
|
3769
|
+
branch: state.branch || nonEmptyString(fullState.branch) || null
|
|
3770
|
+
}),
|
|
3771
|
+
current_phase: compactRecord({
|
|
3772
|
+
stage: state.current_stage ?? null,
|
|
3773
|
+
checkpoint: state.last_checkpoint ?? packet?.checkpoint ?? null,
|
|
3774
|
+
latest_event: latestEvent?.kind ?? null,
|
|
3775
|
+
elapsed_ms: elapsedMs(state.created_at, at),
|
|
3776
|
+
stage_elapsed_ms: elapsedMs(state.stage_started_at, at),
|
|
3777
|
+
iterations: state.iterations
|
|
3778
|
+
}),
|
|
3779
|
+
owner_next_action: compactRecord({
|
|
3780
|
+
owner: ownerFor(state),
|
|
3781
|
+
action: actionFor(state),
|
|
3782
|
+
checkpoint_kind: packet?.kind || null,
|
|
3783
|
+
allowed_decisions: packet?.allowed_decisions,
|
|
3784
|
+
retryable: state.status === "running" || state.status === "awaiting_checkpoint",
|
|
3785
|
+
reason: state.blocker?.code || packet?.summary || latestEvent?.summary || null
|
|
3786
|
+
}),
|
|
3787
|
+
evidence_contract: compactRecord({
|
|
3788
|
+
verification_mode: state.request.verification_mode || nonEmptyString(fullState.verification_mode),
|
|
3789
|
+
required,
|
|
3790
|
+
artifact_contract: artifactContract,
|
|
3791
|
+
proof_plan: compactText2(state.proof_contract?.proof_plan || fullState.proof_plan, 600),
|
|
3792
|
+
stop_condition: compactText2(state.proof_contract?.stop_condition, 400)
|
|
3793
|
+
}),
|
|
3794
|
+
latest_evidence: compactRecord({
|
|
3795
|
+
before_url: nonEmptyString(fullState.before_cdn) || state.before_artifact_url || null,
|
|
3796
|
+
prod_url: nonEmptyString(fullState.prod_cdn) || state.prod_artifact_url || null,
|
|
3797
|
+
after_url: nonEmptyString(fullState.after_cdn) || state.after_artifact_url || null,
|
|
3798
|
+
visual_delta: visualDelta || null,
|
|
3799
|
+
evidence_issue_code: evidenceIssueCode({ fullRiddleState: fullState, runState: state }) || null,
|
|
3800
|
+
proof_evidence_present: fullState.proof_evidence_present === true || Boolean(bundle?.proof_evidence || bundle?.proof_evidence_sample),
|
|
3801
|
+
artifacts
|
|
3802
|
+
}),
|
|
3803
|
+
stop_condition: compactRecord({
|
|
3804
|
+
status: state.status,
|
|
3805
|
+
terminal: isTerminalStatus(state.status),
|
|
3806
|
+
blocker_code: state.blocker?.code || null,
|
|
3807
|
+
blocker_message: state.blocker?.message || null,
|
|
3808
|
+
proof_decision: state.proof_decision,
|
|
3809
|
+
merge_recommendation: state.merge_recommendation,
|
|
3810
|
+
monitor_should_continue: !isTerminalStatus(state.status)
|
|
3811
|
+
}),
|
|
3812
|
+
updated_at: state.updated_at
|
|
3813
|
+
};
|
|
3814
|
+
}
|
|
3815
|
+
|
|
3816
|
+
// src/state.ts
|
|
3817
|
+
var RIDDLE_PROOF_RUN_STATE_VERSION = "riddle-proof.run-state.v1";
|
|
3818
|
+
function timestamp2() {
|
|
3819
|
+
return (/* @__PURE__ */ new Date()).toISOString();
|
|
3820
|
+
}
|
|
3821
|
+
function createRunId(createdAt) {
|
|
3822
|
+
const stamp = createdAt.replace(/\D/g, "").slice(0, 14) || "unknown";
|
|
3823
|
+
const entropy = Math.random().toString(36).slice(2, 8) || "run";
|
|
3824
|
+
return `rp_${stamp}_${entropy}`;
|
|
3825
|
+
}
|
|
3826
|
+
function elapsedMs2(start, end) {
|
|
3827
|
+
const startMs = start ? Date.parse(start) : NaN;
|
|
3828
|
+
const endMs = end ? Date.parse(end) : NaN;
|
|
3829
|
+
if (!Number.isFinite(startMs) || !Number.isFinite(endMs)) return void 0;
|
|
3830
|
+
return Math.max(0, endMs - startMs);
|
|
3831
|
+
}
|
|
3832
|
+
function normalizePrStatus2(value) {
|
|
3833
|
+
const status = nonEmptyString(value)?.toLowerCase();
|
|
3834
|
+
if (status === "merged") return "merged";
|
|
3835
|
+
if (status === "open") return "open";
|
|
3836
|
+
if (status === "closed") return "closed";
|
|
3837
|
+
if (status === "not_found" || status === "not-found") return "not_found";
|
|
3838
|
+
if (status === "unavailable") return "unavailable";
|
|
3839
|
+
return status || "unknown";
|
|
3840
|
+
}
|
|
3841
|
+
function normalizePrLifecycleState(input, checkedAt = timestamp2()) {
|
|
3842
|
+
const value = recordValue(input);
|
|
3843
|
+
if (!value) return void 0;
|
|
3844
|
+
const cleanup = recordValue(value.cleanup);
|
|
3845
|
+
const mergeCommit = nonEmptyString(value.merge_commit) || nonEmptyString(value.mergeCommit) || nonEmptyString(recordValue(value.mergeCommit)?.oid);
|
|
3846
|
+
return compactRecord({
|
|
3847
|
+
status: normalizePrStatus2(value.status || value.state),
|
|
3848
|
+
pr_url: nonEmptyString(value.pr_url) || nonEmptyString(value.prUrl) || nonEmptyString(value.url),
|
|
3849
|
+
pr_number: nonEmptyString(value.pr_number) || nonEmptyString(value.prNumber) || (typeof value.number === "number" ? String(value.number) : void 0),
|
|
3850
|
+
repo: nonEmptyString(value.repo) || nonEmptyString(value.repository),
|
|
3851
|
+
head_branch: nonEmptyString(value.head_branch) || nonEmptyString(value.headBranch) || nonEmptyString(value.headRefName),
|
|
3852
|
+
base_branch: nonEmptyString(value.base_branch) || nonEmptyString(value.baseBranch) || nonEmptyString(value.baseRefName),
|
|
3853
|
+
merge_commit: mergeCommit,
|
|
3854
|
+
merged_at: nonEmptyString(value.merged_at) || nonEmptyString(value.mergedAt),
|
|
3855
|
+
closed_at: nonEmptyString(value.closed_at) || nonEmptyString(value.closedAt),
|
|
3856
|
+
checked_at: nonEmptyString(value.checked_at) || nonEmptyString(value.checkedAt) || checkedAt,
|
|
3857
|
+
source: nonEmptyString(value.source),
|
|
3858
|
+
next_action: nonEmptyString(value.next_action) || nonEmptyString(value.nextAction),
|
|
3859
|
+
cleanup: cleanup && Object.keys(cleanup).length ? cleanup : void 0
|
|
3860
|
+
});
|
|
3861
|
+
}
|
|
3862
|
+
function normalizeIntegrationContext(input, fallbackSource) {
|
|
3863
|
+
const value = recordValue(input);
|
|
3864
|
+
if (!value) {
|
|
3865
|
+
return fallbackSource ? { source: fallbackSource } : void 0;
|
|
3866
|
+
}
|
|
3867
|
+
const metadata = recordValue(value.metadata);
|
|
3868
|
+
return compactRecord({
|
|
3869
|
+
source: nonEmptyString(value.source) || fallbackSource,
|
|
3870
|
+
channel_id: nonEmptyString(value.channel_id),
|
|
3871
|
+
thread_id: nonEmptyString(value.thread_id),
|
|
3872
|
+
message_id: nonEmptyString(value.message_id),
|
|
3873
|
+
source_url: nonEmptyString(value.source_url),
|
|
3874
|
+
metadata: metadata && Object.keys(metadata).length ? metadata : void 0
|
|
3875
|
+
});
|
|
3876
|
+
}
|
|
3877
|
+
function normalizeRunParams(input) {
|
|
3878
|
+
return compactRecord({
|
|
3879
|
+
repo: input.repo,
|
|
3880
|
+
branch: input.branch,
|
|
3881
|
+
change_request: input.change_request,
|
|
3882
|
+
commit_message: input.commit_message,
|
|
3883
|
+
prod_url: input.prod_url,
|
|
3884
|
+
capture_script: input.capture_script,
|
|
3885
|
+
success_criteria: input.success_criteria,
|
|
3886
|
+
assertions: input.assertions,
|
|
3887
|
+
verification_mode: input.verification_mode,
|
|
3888
|
+
resume_session: input.resume_session,
|
|
3889
|
+
target_image_url: input.target_image_url,
|
|
3890
|
+
target_image_hash: input.target_image_hash,
|
|
3891
|
+
viewport_matrix: input.viewport_matrix,
|
|
3892
|
+
deterministic_setup: input.deterministic_setup,
|
|
3893
|
+
reference: input.reference,
|
|
3894
|
+
base_branch: input.base_branch,
|
|
3895
|
+
before_ref: input.before_ref,
|
|
3896
|
+
allow_static_preview_fallback: input.allow_static_preview_fallback,
|
|
3897
|
+
context: input.context,
|
|
3898
|
+
reviewer: input.reviewer,
|
|
3899
|
+
mode: input.mode,
|
|
3900
|
+
build_command: input.build_command,
|
|
3901
|
+
build_output: input.build_output,
|
|
3902
|
+
server_image: input.server_image,
|
|
3903
|
+
server_command: input.server_command,
|
|
3904
|
+
server_port: input.server_port,
|
|
3905
|
+
server_path: input.server_path,
|
|
3906
|
+
use_auth: input.use_auth,
|
|
3907
|
+
auth_localStorage_json: input.auth_localStorage_json,
|
|
3908
|
+
auth_cookies_json: input.auth_cookies_json,
|
|
3909
|
+
auth_headers_json: input.auth_headers_json,
|
|
3910
|
+
color_scheme: input.color_scheme,
|
|
3911
|
+
wait_for_selector: input.wait_for_selector,
|
|
3912
|
+
ship_mode: input.ship_mode,
|
|
3913
|
+
leave_draft: input.leave_draft,
|
|
3914
|
+
engine_state_path: input.engine_state_path,
|
|
3915
|
+
harness_state_path: input.harness_state_path,
|
|
3916
|
+
max_iterations: input.max_iterations,
|
|
3917
|
+
auto_approve: input.auto_approve,
|
|
3918
|
+
dry_run: input.dry_run,
|
|
3919
|
+
integration_context: normalizeIntegrationContext(input.integration_context)
|
|
3920
|
+
});
|
|
3921
|
+
}
|
|
3922
|
+
function createRunState(input) {
|
|
3923
|
+
const createdAt = input.created_at || timestamp2();
|
|
3924
|
+
return compactRecord({
|
|
3925
|
+
version: RIDDLE_PROOF_RUN_STATE_VERSION,
|
|
3926
|
+
run_id: input.run_id || createRunId(createdAt),
|
|
3927
|
+
state_path: input.state_path,
|
|
3928
|
+
worktree_path: input.worktree_path,
|
|
3929
|
+
branch: input.branch || input.request.branch,
|
|
3930
|
+
current_stage: input.current_stage ?? null,
|
|
3931
|
+
stage_started_at: input.stage_started_at ?? null,
|
|
3932
|
+
status: input.status || "running",
|
|
3933
|
+
created_at: createdAt,
|
|
3934
|
+
updated_at: input.updated_at || createdAt,
|
|
3935
|
+
request: normalizeRunParams(input.request),
|
|
3936
|
+
iterations: input.iterations ?? 0,
|
|
3937
|
+
last_checkpoint: input.last_checkpoint ?? null,
|
|
3938
|
+
checkpoint_packet: input.checkpoint_packet,
|
|
3939
|
+
checkpoint_summary: input.checkpoint_summary,
|
|
3940
|
+
state_paths: input.state_paths,
|
|
3941
|
+
proof_contract: input.proof_contract,
|
|
3942
|
+
run_card: input.run_card,
|
|
3943
|
+
checkpoint_history: input.checkpoint_history,
|
|
3944
|
+
events: input.events ? [...input.events] : []
|
|
3945
|
+
});
|
|
3946
|
+
}
|
|
3947
|
+
function appendRunEvent(state, input) {
|
|
3948
|
+
const event = {
|
|
3949
|
+
ts: input.ts || timestamp2(),
|
|
3950
|
+
kind: input.kind,
|
|
3951
|
+
checkpoint: input.checkpoint,
|
|
3952
|
+
stage: input.stage,
|
|
3953
|
+
summary: input.summary,
|
|
3954
|
+
details: input.details
|
|
3955
|
+
};
|
|
3956
|
+
state.events.push(compactRecord({
|
|
3957
|
+
ts: event.ts,
|
|
3958
|
+
kind: event.kind,
|
|
3959
|
+
checkpoint: event.checkpoint,
|
|
3960
|
+
stage: event.stage,
|
|
3961
|
+
summary: event.summary,
|
|
3962
|
+
details: event.details
|
|
3963
|
+
}));
|
|
3964
|
+
if (input.checkpoint !== void 0) state.last_checkpoint = input.checkpoint;
|
|
3965
|
+
if (input.stage !== void 0) {
|
|
3966
|
+
if (state.current_stage !== input.stage) state.stage_started_at = event.ts;
|
|
3967
|
+
state.current_stage = input.stage;
|
|
3968
|
+
}
|
|
3969
|
+
state.updated_at = event.ts;
|
|
3970
|
+
return state;
|
|
3971
|
+
}
|
|
3972
|
+
function appendStageHeartbeat(state, input) {
|
|
3973
|
+
const at = input.ts || timestamp2();
|
|
3974
|
+
return appendRunEvent(state, {
|
|
3975
|
+
ts: at,
|
|
3976
|
+
kind: "stage.heartbeat",
|
|
3977
|
+
checkpoint: input.checkpoint || `${input.stage}_heartbeat`,
|
|
3978
|
+
stage: input.stage,
|
|
3979
|
+
summary: input.summary || `${input.stage} stage is active.`,
|
|
3980
|
+
details: compactRecord({
|
|
3981
|
+
elapsed_ms: elapsedMs2(state.created_at, at),
|
|
3982
|
+
stage_elapsed_ms: elapsedMs2(state.stage_started_at, at),
|
|
3983
|
+
wait_reason: input.wait_reason,
|
|
3984
|
+
blocker: input.blocker,
|
|
3985
|
+
...input.details
|
|
3986
|
+
})
|
|
3987
|
+
});
|
|
3988
|
+
}
|
|
3989
|
+
function createRunStatusSnapshot(state, at = timestamp2()) {
|
|
3990
|
+
const latestEvent = state.events[state.events.length - 1];
|
|
3991
|
+
const runId = state.run_id || "unknown";
|
|
3992
|
+
return compactRecord({
|
|
3993
|
+
run_id: runId,
|
|
3994
|
+
status: state.status,
|
|
3995
|
+
is_terminal: isTerminalStatus(state.status),
|
|
3996
|
+
monitor_should_continue: !isTerminalStatus(state.status),
|
|
3997
|
+
current_stage: state.current_stage ?? null,
|
|
3998
|
+
state_path: state.state_path ?? null,
|
|
3999
|
+
worktree_path: state.worktree_path ?? null,
|
|
4000
|
+
branch: state.branch ?? null,
|
|
4001
|
+
pr_url: state.pr_url ?? null,
|
|
4002
|
+
pr_branch: state.pr_branch ?? null,
|
|
4003
|
+
pr_state: state.pr_state,
|
|
4004
|
+
ci_status: state.ci_status,
|
|
4005
|
+
ship_commit: state.ship_commit,
|
|
4006
|
+
ship_remote_head: state.ship_remote_head,
|
|
4007
|
+
merge_commit: state.merge_commit,
|
|
4008
|
+
merged_at: state.merged_at,
|
|
4009
|
+
proof_comment_url: state.proof_comment_url,
|
|
4010
|
+
cleanup_report: state.cleanup_report,
|
|
4011
|
+
iterations: state.iterations,
|
|
4012
|
+
last_checkpoint: state.last_checkpoint ?? null,
|
|
4013
|
+
updated_at: state.updated_at,
|
|
4014
|
+
elapsed_ms: elapsedMs2(state.created_at, at),
|
|
4015
|
+
stage_elapsed_ms: elapsedMs2(state.stage_started_at, at),
|
|
4016
|
+
blocker: state.blocker,
|
|
4017
|
+
checkpoint_packet: state.checkpoint_packet,
|
|
4018
|
+
checkpoint_summary: state.checkpoint_summary,
|
|
4019
|
+
state_paths: state.state_paths,
|
|
4020
|
+
proof_contract: state.proof_contract,
|
|
4021
|
+
run_card: createRiddleProofRunCard(state, { at }),
|
|
4022
|
+
latest_event: latestEvent
|
|
3689
4023
|
});
|
|
3690
4024
|
}
|
|
4025
|
+
function setRunStatus(state, status, at = timestamp2()) {
|
|
4026
|
+
state.status = status;
|
|
4027
|
+
state.ok = status !== "blocked" && status !== "failed";
|
|
4028
|
+
state.updated_at = at;
|
|
4029
|
+
return state;
|
|
4030
|
+
}
|
|
4031
|
+
function applyPrLifecycleState(state, input, at = timestamp2()) {
|
|
4032
|
+
const prState = normalizePrLifecycleState(input, at);
|
|
4033
|
+
if (!prState) return state;
|
|
4034
|
+
state.pr_state = prState;
|
|
4035
|
+
if (prState.pr_url) state.pr_url = prState.pr_url;
|
|
4036
|
+
if (prState.head_branch) state.pr_branch = prState.head_branch;
|
|
4037
|
+
if (prState.merge_commit) state.merge_commit = prState.merge_commit;
|
|
4038
|
+
if (prState.merged_at) state.merged_at = prState.merged_at;
|
|
4039
|
+
if (prState.cleanup) state.cleanup_report = prState.cleanup;
|
|
4040
|
+
if (prState.status === "merged") {
|
|
4041
|
+
state.finalized = true;
|
|
4042
|
+
state.status = "completed";
|
|
4043
|
+
state.ok = true;
|
|
4044
|
+
}
|
|
4045
|
+
state.updated_at = at;
|
|
4046
|
+
return state;
|
|
4047
|
+
}
|
|
3691
4048
|
|
|
3692
4049
|
// src/runner.ts
|
|
3693
4050
|
function errorDetails(error) {
|
|
@@ -4252,6 +4609,10 @@ function shouldPreserveFinalizedRunState(filePath, incoming) {
|
|
|
4252
4609
|
function persist(state) {
|
|
4253
4610
|
state.state_paths = statePathsForRunState(state);
|
|
4254
4611
|
state.checkpoint_summary = checkpointSummaryFromState(state);
|
|
4612
|
+
state.run_card = createRiddleProofRunCard(state, {
|
|
4613
|
+
fullRiddleState: readJson(state.request.engine_state_path),
|
|
4614
|
+
state_paths: state.state_paths
|
|
4615
|
+
});
|
|
4255
4616
|
if (!state.state_path) return;
|
|
4256
4617
|
if (shouldPreserveFinalizedRunState(state.state_path, state)) return;
|
|
4257
4618
|
writeJson(state.state_path, state);
|
|
@@ -4331,7 +4692,7 @@ function removeEmptyToolArtifacts(workdir) {
|
|
|
4331
4692
|
}
|
|
4332
4693
|
return [];
|
|
4333
4694
|
}
|
|
4334
|
-
function
|
|
4695
|
+
function stageFromCheckpoint2(result) {
|
|
4335
4696
|
const explicitStage = nonEmptyString(result.stage);
|
|
4336
4697
|
if (explicitStage) return explicitStage;
|
|
4337
4698
|
const checkpoint = String(result.checkpoint || "");
|
|
@@ -4476,6 +4837,41 @@ function proofAssessmentPayloadFromCheckpointResponse(response) {
|
|
|
4476
4837
|
checkpoint_response_created_at: response.created_at
|
|
4477
4838
|
});
|
|
4478
4839
|
}
|
|
4840
|
+
function reconAssessmentPayloadFromCheckpointResponse(response) {
|
|
4841
|
+
if (![
|
|
4842
|
+
"ready_for_author",
|
|
4843
|
+
"retry_recon",
|
|
4844
|
+
"recon_stuck",
|
|
4845
|
+
"needs_recon"
|
|
4846
|
+
].includes(response.decision)) {
|
|
4847
|
+
return null;
|
|
4848
|
+
}
|
|
4849
|
+
const payload = recordValue(response.payload) || {};
|
|
4850
|
+
const decision = response.decision === "needs_recon" ? "retry_recon" : response.decision;
|
|
4851
|
+
const continueStage = response.continue_with_stage || nonEmptyString(payload.continue_with_stage) || (decision === "ready_for_author" ? "author" : "recon");
|
|
4852
|
+
return compactRecord({
|
|
4853
|
+
...payload,
|
|
4854
|
+
decision,
|
|
4855
|
+
summary: response.summary,
|
|
4856
|
+
continue_with_stage: continueStage,
|
|
4857
|
+
escalation_target: nonEmptyString(payload.escalation_target) || (decision === "recon_stuck" ? "human" : "agent"),
|
|
4858
|
+
reasons: Array.isArray(response.reasons) ? response.reasons : Array.isArray(payload.reasons) ? payload.reasons : [],
|
|
4859
|
+
source: "supervising_agent",
|
|
4860
|
+
checkpoint_response_source: response.source || null,
|
|
4861
|
+
checkpoint_response_created_at: response.created_at
|
|
4862
|
+
});
|
|
4863
|
+
}
|
|
4864
|
+
function implementationNotesFromCheckpointResponse(response) {
|
|
4865
|
+
const payload = recordValue(response.payload) || {};
|
|
4866
|
+
const notes = nonEmptyString(payload.implementation_notes) || nonEmptyString(payload.implementationNotes) || nonEmptyString(payload.summary) || response.summary;
|
|
4867
|
+
const changedFiles = Array.isArray(payload.changed_files) ? payload.changed_files.filter((item) => typeof item === "string") : [];
|
|
4868
|
+
const testsRun = Array.isArray(payload.tests_run) ? payload.tests_run.filter((item) => typeof item === "string") : [];
|
|
4869
|
+
return [
|
|
4870
|
+
notes,
|
|
4871
|
+
changedFiles.length ? `changed_files=${changedFiles.join(", ")}` : "",
|
|
4872
|
+
testsRun.length ? `tests_run=${testsRun.join(", ")}` : ""
|
|
4873
|
+
].filter(Boolean).join("\n");
|
|
4874
|
+
}
|
|
4479
4875
|
function proofAssessmentVisualBlocker(state, payload) {
|
|
4480
4876
|
if (!proofAssessmentRequestsShip(payload)) return null;
|
|
4481
4877
|
const source = nonEmptyString(payload.source) || "supervising_agent";
|
|
@@ -4564,7 +4960,7 @@ function engineFailureBlocker(result, checkpoint) {
|
|
|
4564
4960
|
}
|
|
4565
4961
|
function terminalResult(state, status, result, summary, raw = {}) {
|
|
4566
4962
|
if (result) {
|
|
4567
|
-
const terminalStage =
|
|
4963
|
+
const terminalStage = stageFromCheckpoint2(result);
|
|
4568
4964
|
if (terminalStage !== "setup") state.current_stage = terminalStage;
|
|
4569
4965
|
}
|
|
4570
4966
|
setRunStatus(state, status);
|
|
@@ -4590,7 +4986,7 @@ function terminalResult(state, status, result, summary, raw = {}) {
|
|
|
4590
4986
|
function blockerResult(state, result, blocker) {
|
|
4591
4987
|
state.blocker = blocker;
|
|
4592
4988
|
const blockerStage = nonEmptyString(recordValue(blocker.details)?.stage);
|
|
4593
|
-
const stage = blockerStage ||
|
|
4989
|
+
const stage = blockerStage || stageFromCheckpoint2(result || { checkpoint: blocker.checkpoint || void 0 });
|
|
4594
4990
|
recordEvent(state, {
|
|
4595
4991
|
kind: "run.blocked",
|
|
4596
4992
|
checkpoint: blocker.checkpoint || result?.checkpoint || null,
|
|
@@ -4777,6 +5173,67 @@ function checkpointResponseContinuation(state, value) {
|
|
|
4777
5173
|
appendCheckpointResponse(state, response);
|
|
4778
5174
|
return { next: { ...base, author_packet_json: jsonParam(payload) } };
|
|
4779
5175
|
}
|
|
5176
|
+
if (packet.kind === "assess_recon" || packet.stage === "recon") {
|
|
5177
|
+
const assessment = reconAssessmentPayloadFromCheckpointResponse(response);
|
|
5178
|
+
if (assessment) {
|
|
5179
|
+
appendCheckpointResponse(state, response);
|
|
5180
|
+
return { next: { ...base, recon_assessment_json: jsonParam(assessment) } };
|
|
5181
|
+
}
|
|
5182
|
+
if (response.decision === "blocked" || response.decision === "human_review") {
|
|
5183
|
+
appendCheckpointResponse(state, response, { clear_packet: false });
|
|
5184
|
+
return {
|
|
5185
|
+
blocker: {
|
|
5186
|
+
code: `checkpoint_response_${response.decision}`,
|
|
5187
|
+
checkpoint: packet.checkpoint,
|
|
5188
|
+
message: response.summary || `Checkpoint response stopped recon with decision=${response.decision}.`,
|
|
5189
|
+
details: { stage: packet.stage, response }
|
|
5190
|
+
}
|
|
5191
|
+
};
|
|
5192
|
+
}
|
|
5193
|
+
}
|
|
5194
|
+
if (packet.kind === "implement_change" || packet.stage === "implement") {
|
|
5195
|
+
if (response.decision === "implementation_complete") {
|
|
5196
|
+
const workdir = nonEmptyString(packet.state_excerpt?.after_worktree) || state.worktree_path;
|
|
5197
|
+
if (workdir) state.worktree_path = workdir;
|
|
5198
|
+
if (!hasGitDiff(workdir)) {
|
|
5199
|
+
return {
|
|
5200
|
+
blocker: {
|
|
5201
|
+
code: "implementation_diff_missing",
|
|
5202
|
+
checkpoint: packet.checkpoint,
|
|
5203
|
+
message: "Checkpoint response claimed implementation_complete, but the after worktree has no detectable git diff.",
|
|
5204
|
+
details: { stage: packet.stage, worktree_path: workdir || null, response }
|
|
5205
|
+
}
|
|
5206
|
+
};
|
|
5207
|
+
}
|
|
5208
|
+
appendCheckpointResponse(state, response);
|
|
5209
|
+
return {
|
|
5210
|
+
next: {
|
|
5211
|
+
...base,
|
|
5212
|
+
advance_stage: "implement",
|
|
5213
|
+
implementation_notes: implementationNotesFromCheckpointResponse(response)
|
|
5214
|
+
}
|
|
5215
|
+
};
|
|
5216
|
+
}
|
|
5217
|
+
if (response.decision === "needs_author") {
|
|
5218
|
+
appendCheckpointResponse(state, response);
|
|
5219
|
+
return { next: { ...base, advance_stage: "author" } };
|
|
5220
|
+
}
|
|
5221
|
+
if (response.decision === "needs_recon") {
|
|
5222
|
+
appendCheckpointResponse(state, response);
|
|
5223
|
+
return { next: { ...base, advance_stage: "recon" } };
|
|
5224
|
+
}
|
|
5225
|
+
if (response.decision === "blocked" || response.decision === "human_review") {
|
|
5226
|
+
appendCheckpointResponse(state, response, { clear_packet: false });
|
|
5227
|
+
return {
|
|
5228
|
+
blocker: {
|
|
5229
|
+
code: `checkpoint_response_${response.decision}`,
|
|
5230
|
+
checkpoint: packet.checkpoint,
|
|
5231
|
+
message: response.summary || `Checkpoint response stopped implementation with decision=${response.decision}.`,
|
|
5232
|
+
details: { stage: packet.stage, response }
|
|
5233
|
+
}
|
|
5234
|
+
};
|
|
5235
|
+
}
|
|
5236
|
+
}
|
|
4780
5237
|
if (response.decision === "needs_recon") {
|
|
4781
5238
|
appendCheckpointResponse(state, response);
|
|
4782
5239
|
if (packet.kind === "assess_proof" || packet.kind === "recover_evidence" || packet.stage === "verify") {
|
|
@@ -5044,6 +5501,9 @@ async function routeCheckpoint(request, state, result, agent, input) {
|
|
|
5044
5501
|
};
|
|
5045
5502
|
}
|
|
5046
5503
|
if (checkpoint === "recon_supervisor_judgment") {
|
|
5504
|
+
if (input.checkpoint_mode === "yield") {
|
|
5505
|
+
return { terminal: checkpointAwaitingResult(state, result, input.checkpoint_visibility) };
|
|
5506
|
+
}
|
|
5047
5507
|
const assessment = await agent.assessRecon(context);
|
|
5048
5508
|
const blocker = requirePayload("recon_assessment", assessment, state, result);
|
|
5049
5509
|
if (blocker) return { blocker };
|
|
@@ -5079,6 +5539,13 @@ async function routeCheckpoint(request, state, result, agent, input) {
|
|
|
5079
5539
|
};
|
|
5080
5540
|
}
|
|
5081
5541
|
if (checkpoint === "implement_changes_missing" || checkpoint === "implement_required" || checkpoint === "verify_agent_retry" && continueStage === "implement") {
|
|
5542
|
+
if (input.checkpoint_mode === "yield") {
|
|
5543
|
+
const fullState = context.fullRiddleState || {};
|
|
5544
|
+
state.worktree_path = workdirFromState(fullState) || state.worktree_path;
|
|
5545
|
+
state.branch = nonEmptyString(fullState.branch) || state.branch;
|
|
5546
|
+
persist(state);
|
|
5547
|
+
return { terminal: checkpointAwaitingResult(state, result, input.checkpoint_visibility) };
|
|
5548
|
+
}
|
|
5082
5549
|
return handleImplementation(request, state, result, agent);
|
|
5083
5550
|
}
|
|
5084
5551
|
if (checkpoint === "implement_review") {
|
|
@@ -5291,7 +5758,7 @@ async function runRiddleProofEngineHarness(input) {
|
|
|
5291
5758
|
const engineState = engineStatePath(result, state);
|
|
5292
5759
|
if (engineState) state.request.engine_state_path = engineState;
|
|
5293
5760
|
state.last_checkpoint = result.checkpoint || state.last_checkpoint || null;
|
|
5294
|
-
const resultStage =
|
|
5761
|
+
const resultStage = stageFromCheckpoint2(result);
|
|
5295
5762
|
stageIterations[resultStage] = (stageIterations[resultStage] || 0) + 1;
|
|
5296
5763
|
heartbeat(state, {
|
|
5297
5764
|
stage: resultStage,
|
|
@@ -5358,6 +5825,755 @@ async function runRiddleProofEngineHarness(input) {
|
|
|
5358
5825
|
});
|
|
5359
5826
|
}
|
|
5360
5827
|
|
|
5828
|
+
// src/codex-exec-agent.ts
|
|
5829
|
+
var import_node_child_process3 = require("child_process");
|
|
5830
|
+
var import_node_fs4 = require("fs");
|
|
5831
|
+
var import_node_os = __toESM(require("os"), 1);
|
|
5832
|
+
var import_node_path4 = __toESM(require("path"), 1);
|
|
5833
|
+
var REFINED_INPUTS_SCHEMA = {
|
|
5834
|
+
type: "object",
|
|
5835
|
+
additionalProperties: false,
|
|
5836
|
+
required: ["server_path", "wait_for_selector", "reference"],
|
|
5837
|
+
properties: {
|
|
5838
|
+
server_path: { type: ["string", "null"] },
|
|
5839
|
+
wait_for_selector: { type: ["string", "null"] },
|
|
5840
|
+
reference: { enum: ["before", "prod", "both", null] }
|
|
5841
|
+
}
|
|
5842
|
+
};
|
|
5843
|
+
var BASELINE_UNDERSTANDING_SCHEMA = {
|
|
5844
|
+
type: "object",
|
|
5845
|
+
additionalProperties: false,
|
|
5846
|
+
required: [
|
|
5847
|
+
"reference",
|
|
5848
|
+
"target_route",
|
|
5849
|
+
"before_evidence_url",
|
|
5850
|
+
"visible_before_state",
|
|
5851
|
+
"relevant_elements",
|
|
5852
|
+
"requested_change",
|
|
5853
|
+
"proof_focus",
|
|
5854
|
+
"stop_condition",
|
|
5855
|
+
"quality_risks"
|
|
5856
|
+
],
|
|
5857
|
+
properties: {
|
|
5858
|
+
reference: { type: "string", enum: ["before", "prod", "both", "unknown"] },
|
|
5859
|
+
target_route: { type: "string" },
|
|
5860
|
+
before_evidence_url: { type: "string" },
|
|
5861
|
+
visible_before_state: { type: "string" },
|
|
5862
|
+
relevant_elements: { type: "array", items: { type: "string" } },
|
|
5863
|
+
requested_change: { type: "string" },
|
|
5864
|
+
proof_focus: { type: "string" },
|
|
5865
|
+
stop_condition: { type: "string" },
|
|
5866
|
+
quality_risks: { type: "array", items: { type: "string" } }
|
|
5867
|
+
}
|
|
5868
|
+
};
|
|
5869
|
+
var RECON_SCHEMA = {
|
|
5870
|
+
type: "object",
|
|
5871
|
+
additionalProperties: false,
|
|
5872
|
+
required: [
|
|
5873
|
+
"decision",
|
|
5874
|
+
"summary",
|
|
5875
|
+
"baseline_understanding",
|
|
5876
|
+
"continue_with_stage",
|
|
5877
|
+
"escalation_target",
|
|
5878
|
+
"refined_inputs",
|
|
5879
|
+
"reasons",
|
|
5880
|
+
"source"
|
|
5881
|
+
],
|
|
5882
|
+
properties: {
|
|
5883
|
+
decision: { type: "string", enum: ["retry_recon", "ready_for_author", "recon_stuck"] },
|
|
5884
|
+
summary: { type: "string" },
|
|
5885
|
+
baseline_understanding: BASELINE_UNDERSTANDING_SCHEMA,
|
|
5886
|
+
continue_with_stage: { type: "string", enum: ["recon", "author"] },
|
|
5887
|
+
escalation_target: { type: "string", enum: ["agent", "human"] },
|
|
5888
|
+
refined_inputs: REFINED_INPUTS_SCHEMA,
|
|
5889
|
+
reasons: { type: "array", items: { type: "string" } },
|
|
5890
|
+
source: { type: "string", enum: ["supervising_agent"] }
|
|
5891
|
+
}
|
|
5892
|
+
};
|
|
5893
|
+
var AUTHOR_SCHEMA = {
|
|
5894
|
+
type: "object",
|
|
5895
|
+
additionalProperties: false,
|
|
5896
|
+
required: [
|
|
5897
|
+
"proof_plan",
|
|
5898
|
+
"capture_script",
|
|
5899
|
+
"baseline_understanding_used",
|
|
5900
|
+
"refined_inputs",
|
|
5901
|
+
"rationale",
|
|
5902
|
+
"confidence",
|
|
5903
|
+
"summary"
|
|
5904
|
+
],
|
|
5905
|
+
properties: {
|
|
5906
|
+
proof_plan: { type: "string" },
|
|
5907
|
+
capture_script: { type: "string" },
|
|
5908
|
+
baseline_understanding_used: BASELINE_UNDERSTANDING_SCHEMA,
|
|
5909
|
+
refined_inputs: REFINED_INPUTS_SCHEMA,
|
|
5910
|
+
rationale: { type: "array", items: { type: "string" } },
|
|
5911
|
+
confidence: { type: "string", enum: ["low", "medium", "high"] },
|
|
5912
|
+
summary: { type: "string" }
|
|
5913
|
+
}
|
|
5914
|
+
};
|
|
5915
|
+
var IMPLEMENT_SCHEMA = {
|
|
5916
|
+
type: "object",
|
|
5917
|
+
additionalProperties: false,
|
|
5918
|
+
required: ["summary", "implementation_notes", "changed_files", "tests_run", "blockers"],
|
|
5919
|
+
properties: {
|
|
5920
|
+
summary: { type: "string" },
|
|
5921
|
+
implementation_notes: { type: "string" },
|
|
5922
|
+
changed_files: { type: "array", items: { type: "string" } },
|
|
5923
|
+
tests_run: { type: "array", items: { type: "string" } },
|
|
5924
|
+
blockers: { type: "array", items: { type: "string" } }
|
|
5925
|
+
}
|
|
5926
|
+
};
|
|
5927
|
+
var PROOF_SCHEMA = {
|
|
5928
|
+
type: "object",
|
|
5929
|
+
additionalProperties: false,
|
|
5930
|
+
required: [
|
|
5931
|
+
"decision",
|
|
5932
|
+
"summary",
|
|
5933
|
+
"recommended_stage",
|
|
5934
|
+
"continue_with_stage",
|
|
5935
|
+
"escalation_target",
|
|
5936
|
+
"reasons",
|
|
5937
|
+
"source"
|
|
5938
|
+
],
|
|
5939
|
+
properties: {
|
|
5940
|
+
decision: {
|
|
5941
|
+
type: "string",
|
|
5942
|
+
enum: [
|
|
5943
|
+
"ready_to_ship",
|
|
5944
|
+
"needs_richer_proof",
|
|
5945
|
+
"revise_capture",
|
|
5946
|
+
"needs_recon",
|
|
5947
|
+
"needs_implementation"
|
|
5948
|
+
]
|
|
5949
|
+
},
|
|
5950
|
+
summary: { type: "string" },
|
|
5951
|
+
recommended_stage: { type: "string", enum: ["ship", "author", "implement", "recon", "verify"] },
|
|
5952
|
+
continue_with_stage: { type: "string", enum: ["ship", "author", "implement", "recon", "verify"] },
|
|
5953
|
+
escalation_target: { type: "string", enum: ["agent", "human"] },
|
|
5954
|
+
reasons: { type: "array", items: { type: "string" } },
|
|
5955
|
+
source: { type: "string", enum: ["supervising_agent"] }
|
|
5956
|
+
}
|
|
5957
|
+
};
|
|
5958
|
+
var PROMPT_STRING_LIMIT = 2e3;
|
|
5959
|
+
var PROMPT_ARRAY_LIMIT = 12;
|
|
5960
|
+
var PROMPT_OBJECT_KEY_LIMIT = 70;
|
|
5961
|
+
var PROMPT_BLOCK_LIMIT = 12e4;
|
|
5962
|
+
var PROMPT_KEY_PRIORITY = [
|
|
5963
|
+
"ok",
|
|
5964
|
+
"status",
|
|
5965
|
+
"stage",
|
|
5966
|
+
"checkpoint",
|
|
5967
|
+
"summary",
|
|
5968
|
+
"state_path",
|
|
5969
|
+
"repo",
|
|
5970
|
+
"branch",
|
|
5971
|
+
"change_request",
|
|
5972
|
+
"context",
|
|
5973
|
+
"success_criteria",
|
|
5974
|
+
"verification_mode",
|
|
5975
|
+
"reference",
|
|
5976
|
+
"server_path",
|
|
5977
|
+
"wait_for_selector",
|
|
5978
|
+
"before_cdn",
|
|
5979
|
+
"prod_cdn",
|
|
5980
|
+
"after_cdn",
|
|
5981
|
+
"before_baseline",
|
|
5982
|
+
"prod_baseline",
|
|
5983
|
+
"recon_assessment",
|
|
5984
|
+
"baseline_understanding",
|
|
5985
|
+
"supervisor_author_packet",
|
|
5986
|
+
"proof_plan",
|
|
5987
|
+
"capture_script",
|
|
5988
|
+
"implementation_status",
|
|
5989
|
+
"implementation_summary",
|
|
5990
|
+
"implementation_notes",
|
|
5991
|
+
"changed_files",
|
|
5992
|
+
"verify_status",
|
|
5993
|
+
"verify_summary",
|
|
5994
|
+
"proof_assessment",
|
|
5995
|
+
"proof_assessment_request",
|
|
5996
|
+
"semantic_context",
|
|
5997
|
+
"visual_delta",
|
|
5998
|
+
"proof_evidence_present",
|
|
5999
|
+
"proof_evidence_sample",
|
|
6000
|
+
"artifacts",
|
|
6001
|
+
"assertions",
|
|
6002
|
+
"checkpointContract",
|
|
6003
|
+
"shipGate",
|
|
6004
|
+
"last_error",
|
|
6005
|
+
"errors",
|
|
6006
|
+
"events"
|
|
6007
|
+
];
|
|
6008
|
+
var PROMPT_PRIORITY_INDEX = new Map(PROMPT_KEY_PRIORITY.map((key, index) => [key, index]));
|
|
6009
|
+
function truncatePromptString(value, limit = PROMPT_STRING_LIMIT) {
|
|
6010
|
+
if (value.length <= limit) return value;
|
|
6011
|
+
return `${value.slice(0, limit)}
|
|
6012
|
+
...[truncated ${value.length - limit} chars]`;
|
|
6013
|
+
}
|
|
6014
|
+
function compactPromptValue(value, depth = 0, key = "") {
|
|
6015
|
+
if (value === null || typeof value === "boolean" || typeof value === "number") return value;
|
|
6016
|
+
if (typeof value === "string") {
|
|
6017
|
+
const lowerKey = key.toLowerCase();
|
|
6018
|
+
const looksLikeUrl = /^https?:\/\//.test(value);
|
|
6019
|
+
if (!looksLikeUrl && (lowerKey.includes("base64") || lowerKey.includes("data_url") || lowerKey.includes("screenshot_blob"))) {
|
|
6020
|
+
return `[omitted ${value.length} chars from ${key || "large artifact"}]`;
|
|
6021
|
+
}
|
|
6022
|
+
const limit = looksLikeUrl ? 1e3 : depth <= 1 ? PROMPT_STRING_LIMIT : 1200;
|
|
6023
|
+
return truncatePromptString(value, limit);
|
|
6024
|
+
}
|
|
6025
|
+
if (Array.isArray(value)) {
|
|
6026
|
+
const lowerKey = key.toLowerCase();
|
|
6027
|
+
const items = lowerKey.includes("event") || lowerKey.includes("histor") || lowerKey.includes("retry") ? value.slice(-PROMPT_ARRAY_LIMIT) : value.slice(0, PROMPT_ARRAY_LIMIT);
|
|
6028
|
+
const compacted = items.map((item) => compactPromptValue(item, depth + 1, key));
|
|
6029
|
+
if (value.length > items.length) {
|
|
6030
|
+
const omitted = value.length - items.length;
|
|
6031
|
+
return lowerKey.includes("event") || lowerKey.includes("histor") || lowerKey.includes("retry") ? [{ omittedEarlierItems: omitted }, ...compacted] : [...compacted, { omittedItems: omitted }];
|
|
6032
|
+
}
|
|
6033
|
+
return compacted;
|
|
6034
|
+
}
|
|
6035
|
+
if (!value || typeof value !== "object") return String(value);
|
|
6036
|
+
const entries = Object.entries(value).sort(([left], [right]) => {
|
|
6037
|
+
const leftPriority = PROMPT_PRIORITY_INDEX.get(left) ?? 1e3;
|
|
6038
|
+
const rightPriority = PROMPT_PRIORITY_INDEX.get(right) ?? 1e3;
|
|
6039
|
+
if (leftPriority !== rightPriority) return leftPriority - rightPriority;
|
|
6040
|
+
return left.localeCompare(right);
|
|
6041
|
+
});
|
|
6042
|
+
const result = {};
|
|
6043
|
+
for (const [entryKey, entryValue] of entries.slice(0, PROMPT_OBJECT_KEY_LIMIT)) {
|
|
6044
|
+
result[entryKey] = compactPromptValue(entryValue, depth + 1, entryKey);
|
|
6045
|
+
}
|
|
6046
|
+
if (entries.length > PROMPT_OBJECT_KEY_LIMIT) {
|
|
6047
|
+
result.__omitted_keys = entries.length - PROMPT_OBJECT_KEY_LIMIT;
|
|
6048
|
+
}
|
|
6049
|
+
return result;
|
|
6050
|
+
}
|
|
6051
|
+
function jsonBlock(label, value) {
|
|
6052
|
+
let json = JSON.stringify(compactPromptValue(value), null, 2);
|
|
6053
|
+
if (json.length > PROMPT_BLOCK_LIMIT) {
|
|
6054
|
+
json = `${json.slice(0, PROMPT_BLOCK_LIMIT)}
|
|
6055
|
+
...[truncated ${json.length - PROMPT_BLOCK_LIMIT} chars from compacted ${label}]`;
|
|
6056
|
+
}
|
|
6057
|
+
return `${label}:
|
|
6058
|
+
${json}`;
|
|
6059
|
+
}
|
|
6060
|
+
function resolveWorkdir(context, fallback = "/tmp") {
|
|
6061
|
+
const after = typeof context.fullRiddleState?.after_worktree === "string" ? context.fullRiddleState.after_worktree.trim() : "";
|
|
6062
|
+
return after || fallback;
|
|
6063
|
+
}
|
|
6064
|
+
function basePrompt(context, role) {
|
|
6065
|
+
return [
|
|
6066
|
+
role,
|
|
6067
|
+
"",
|
|
6068
|
+
"You are the supervising Codex worker inside the Riddle Proof harness.",
|
|
6069
|
+
"Return only JSON matching the provided output schema.",
|
|
6070
|
+
"Do not ask the human to manually continue. If blocked, encode the blocker in the JSON fields allowed by the schema.",
|
|
6071
|
+
"Large raw artifacts are summarized before this prompt. Use preserved state summaries, URLs, file paths, and proof evidence samples instead of expecting full raw logs or binary artifacts inline.",
|
|
6072
|
+
"",
|
|
6073
|
+
jsonBlock("Original request", context.request),
|
|
6074
|
+
jsonBlock("Riddle checkpoint result", context.engineResult),
|
|
6075
|
+
jsonBlock("Full riddle state", context.fullRiddleState || {})
|
|
6076
|
+
].join("\n");
|
|
6077
|
+
}
|
|
6078
|
+
function schemaRequiredKeys(schema) {
|
|
6079
|
+
const required = schema?.required;
|
|
6080
|
+
return Array.isArray(required) ? required.filter((key) => typeof key === "string" && key.length > 0) : [];
|
|
6081
|
+
}
|
|
6082
|
+
function isSchemaShapedObject(value, schema) {
|
|
6083
|
+
if (!value || typeof value !== "object" || Array.isArray(value)) return false;
|
|
6084
|
+
const required = schemaRequiredKeys(schema);
|
|
6085
|
+
return required.every((key) => Object.prototype.hasOwnProperty.call(value, key));
|
|
6086
|
+
}
|
|
6087
|
+
function parseJsonCandidate(raw, schema) {
|
|
6088
|
+
try {
|
|
6089
|
+
const parsed = JSON.parse(raw);
|
|
6090
|
+
return isSchemaShapedObject(parsed, schema) ? parsed : null;
|
|
6091
|
+
} catch {
|
|
6092
|
+
return null;
|
|
6093
|
+
}
|
|
6094
|
+
}
|
|
6095
|
+
function extractJsonObjectCandidates(raw) {
|
|
6096
|
+
const candidates = [];
|
|
6097
|
+
let depth = 0;
|
|
6098
|
+
let start = -1;
|
|
6099
|
+
let inString = false;
|
|
6100
|
+
let escaped = false;
|
|
6101
|
+
for (let index = 0; index < raw.length; index += 1) {
|
|
6102
|
+
const char = raw[index];
|
|
6103
|
+
if (inString) {
|
|
6104
|
+
if (escaped) {
|
|
6105
|
+
escaped = false;
|
|
6106
|
+
} else if (char === "\\") {
|
|
6107
|
+
escaped = true;
|
|
6108
|
+
} else if (char === '"') {
|
|
6109
|
+
inString = false;
|
|
6110
|
+
}
|
|
6111
|
+
continue;
|
|
6112
|
+
}
|
|
6113
|
+
if (char === '"') {
|
|
6114
|
+
inString = true;
|
|
6115
|
+
continue;
|
|
6116
|
+
}
|
|
6117
|
+
if (char === "{") {
|
|
6118
|
+
if (depth === 0) start = index;
|
|
6119
|
+
depth += 1;
|
|
6120
|
+
continue;
|
|
6121
|
+
}
|
|
6122
|
+
if (char === "}" && depth > 0) {
|
|
6123
|
+
depth -= 1;
|
|
6124
|
+
if (depth === 0 && start >= 0) {
|
|
6125
|
+
candidates.push(raw.slice(start, index + 1));
|
|
6126
|
+
start = -1;
|
|
6127
|
+
}
|
|
6128
|
+
}
|
|
6129
|
+
}
|
|
6130
|
+
return candidates;
|
|
6131
|
+
}
|
|
6132
|
+
function parseJsonObject(raw, schema) {
|
|
6133
|
+
const trimmed = raw.trim();
|
|
6134
|
+
if (!trimmed) return null;
|
|
6135
|
+
const whole = parseJsonCandidate(trimmed, schema);
|
|
6136
|
+
if (whole) return whole;
|
|
6137
|
+
const lines = trimmed.split(/\r?\n/).map((line) => line.trim()).filter(Boolean);
|
|
6138
|
+
for (let index = lines.length - 1; index >= 0; index -= 1) {
|
|
6139
|
+
const parsed = parseJsonCandidate(lines[index], schema);
|
|
6140
|
+
if (parsed) return parsed;
|
|
6141
|
+
}
|
|
6142
|
+
const candidates = extractJsonObjectCandidates(trimmed);
|
|
6143
|
+
for (let index = candidates.length - 1; index >= 0; index -= 1) {
|
|
6144
|
+
const parsed = parseJsonCandidate(candidates[index], schema);
|
|
6145
|
+
if (parsed) return parsed;
|
|
6146
|
+
}
|
|
6147
|
+
return null;
|
|
6148
|
+
}
|
|
6149
|
+
function isHarnessVerificationOnlyBlocker(blocker) {
|
|
6150
|
+
const text = blocker.toLowerCase();
|
|
6151
|
+
return (text.includes("erofs") || text.includes("read-only file system")) && text.includes("node_modules") && (text.includes(".vite-temp") || text.includes("vite.config"));
|
|
6152
|
+
}
|
|
6153
|
+
function createCodexExecJsonRunner(config = {}) {
|
|
6154
|
+
return (request) => {
|
|
6155
|
+
if (!request.workdir || !(0, import_node_fs4.existsSync)(request.workdir)) {
|
|
6156
|
+
return {
|
|
6157
|
+
ok: false,
|
|
6158
|
+
blocker: {
|
|
6159
|
+
code: "codex_workdir_missing",
|
|
6160
|
+
message: `Codex workdir does not exist for ${request.purpose}.`,
|
|
6161
|
+
details: { workdir: request.workdir }
|
|
6162
|
+
}
|
|
6163
|
+
};
|
|
6164
|
+
}
|
|
6165
|
+
const tmpDir = (0, import_node_fs4.mkdtempSync)(import_node_path4.default.join(import_node_os.default.tmpdir(), "riddle-proof-codex-"));
|
|
6166
|
+
const schemaPath = import_node_path4.default.join(tmpDir, "schema.json");
|
|
6167
|
+
const lastMessagePath = import_node_path4.default.join(tmpDir, "last-message.json");
|
|
6168
|
+
(0, import_node_fs4.writeFileSync)(schemaPath, JSON.stringify(request.schema, null, 2));
|
|
6169
|
+
const args = [
|
|
6170
|
+
"exec",
|
|
6171
|
+
"--json",
|
|
6172
|
+
"--output-schema",
|
|
6173
|
+
schemaPath,
|
|
6174
|
+
"--output-last-message",
|
|
6175
|
+
lastMessagePath,
|
|
6176
|
+
"--cd",
|
|
6177
|
+
request.workdir,
|
|
6178
|
+
"--sandbox",
|
|
6179
|
+
config.codexSandbox || "workspace-write",
|
|
6180
|
+
"--skip-git-repo-check"
|
|
6181
|
+
];
|
|
6182
|
+
if (config.codexFullAuto !== false) args.push("--full-auto");
|
|
6183
|
+
if (config.codexModel) args.push("-m", config.codexModel);
|
|
6184
|
+
args.push("-");
|
|
6185
|
+
const env = { ...process.env };
|
|
6186
|
+
if (config.codexHome) env.CODEX_HOME = config.codexHome;
|
|
6187
|
+
delete env.OPENAI_API_KEY;
|
|
6188
|
+
try {
|
|
6189
|
+
const proc = (0, import_node_child_process3.spawnSync)(config.codexCommand || "codex", args, {
|
|
6190
|
+
input: request.prompt,
|
|
6191
|
+
encoding: "utf-8",
|
|
6192
|
+
timeout: Number(config.codexTimeoutMs || 6e5),
|
|
6193
|
+
maxBuffer: 10 * 1024 * 1024,
|
|
6194
|
+
env
|
|
6195
|
+
});
|
|
6196
|
+
if (proc.error) {
|
|
6197
|
+
const timedOut = proc.error.code === "ETIMEDOUT";
|
|
6198
|
+
return {
|
|
6199
|
+
ok: false,
|
|
6200
|
+
stdout: proc.stdout || "",
|
|
6201
|
+
stderr: proc.stderr || "",
|
|
6202
|
+
blocker: {
|
|
6203
|
+
code: timedOut ? "codex_timeout" : "codex_exec_error",
|
|
6204
|
+
message: timedOut ? `Codex timed out during ${request.purpose}.` : `Codex failed to start or complete ${request.purpose}.`,
|
|
6205
|
+
details: { error: proc.error.message }
|
|
6206
|
+
}
|
|
6207
|
+
};
|
|
6208
|
+
}
|
|
6209
|
+
if (proc.status !== 0) {
|
|
6210
|
+
return {
|
|
6211
|
+
ok: false,
|
|
6212
|
+
stdout: proc.stdout || "",
|
|
6213
|
+
stderr: proc.stderr || "",
|
|
6214
|
+
blocker: {
|
|
6215
|
+
code: "codex_nonzero_exit",
|
|
6216
|
+
message: `Codex exited with status ${proc.status} during ${request.purpose}.`,
|
|
6217
|
+
details: { stdout: proc.stdout || "", stderr: proc.stderr || "" }
|
|
6218
|
+
}
|
|
6219
|
+
};
|
|
6220
|
+
}
|
|
6221
|
+
const finalText = (0, import_node_fs4.existsSync)(lastMessagePath) ? (0, import_node_fs4.readFileSync)(lastMessagePath, "utf-8") : String(proc.stdout || "");
|
|
6222
|
+
const parsed = parseJsonObject(finalText, request.schema);
|
|
6223
|
+
if (!parsed) {
|
|
6224
|
+
return {
|
|
6225
|
+
ok: false,
|
|
6226
|
+
stdout: proc.stdout || "",
|
|
6227
|
+
stderr: proc.stderr || "",
|
|
6228
|
+
blocker: {
|
|
6229
|
+
code: "codex_invalid_json",
|
|
6230
|
+
message: `Codex completed ${request.purpose}, but did not return valid JSON.`,
|
|
6231
|
+
details: { finalText, stdout: proc.stdout || "", stderr: proc.stderr || "" }
|
|
6232
|
+
}
|
|
6233
|
+
};
|
|
6234
|
+
}
|
|
6235
|
+
return {
|
|
6236
|
+
ok: true,
|
|
6237
|
+
json: parsed,
|
|
6238
|
+
stdout: proc.stdout || "",
|
|
6239
|
+
stderr: proc.stderr || ""
|
|
6240
|
+
};
|
|
6241
|
+
} finally {
|
|
6242
|
+
(0, import_node_fs4.rmSync)(tmpDir, { recursive: true, force: true });
|
|
6243
|
+
}
|
|
6244
|
+
};
|
|
6245
|
+
}
|
|
6246
|
+
async function callRunner(runner, request) {
|
|
6247
|
+
return runner(request);
|
|
6248
|
+
}
|
|
6249
|
+
function payloadOrBlocker(raw, checkpoint) {
|
|
6250
|
+
if (!raw.ok || !raw.json) {
|
|
6251
|
+
const blocker = raw.blocker || {
|
|
6252
|
+
code: "codex_runner_failed",
|
|
6253
|
+
message: "Codex runner failed without a detailed blocker.",
|
|
6254
|
+
details: { stdout: raw.stdout || "", stderr: raw.stderr || "" }
|
|
6255
|
+
};
|
|
6256
|
+
return {
|
|
6257
|
+
ok: false,
|
|
6258
|
+
blocker: {
|
|
6259
|
+
...blocker,
|
|
6260
|
+
checkpoint
|
|
6261
|
+
}
|
|
6262
|
+
};
|
|
6263
|
+
}
|
|
6264
|
+
return {
|
|
6265
|
+
ok: true,
|
|
6266
|
+
payload: raw.json,
|
|
6267
|
+
summary: typeof raw.json.summary === "string" ? raw.json.summary : void 0
|
|
6268
|
+
};
|
|
6269
|
+
}
|
|
6270
|
+
function stringArray(value) {
|
|
6271
|
+
return Array.isArray(value) ? value.filter((item) => typeof item === "string") : [];
|
|
6272
|
+
}
|
|
6273
|
+
function parseGitStatusPaths2(status) {
|
|
6274
|
+
return status.split(/\r?\n/).map((line) => line.trimEnd()).filter(Boolean).map((line) => {
|
|
6275
|
+
let item = line.length > 3 ? line.slice(3) : line;
|
|
6276
|
+
if (item.includes(" -> ")) item = item.split(" -> ").pop() || item;
|
|
6277
|
+
return item.trim();
|
|
6278
|
+
}).filter(Boolean);
|
|
6279
|
+
}
|
|
6280
|
+
function isToolNoisePath2(filePath) {
|
|
6281
|
+
return filePath === ".codex" || filePath.startsWith(".codex/") || filePath === ".oc-smoke" || filePath.startsWith(".oc-smoke/");
|
|
6282
|
+
}
|
|
6283
|
+
function hasGitDiff2(workdir) {
|
|
6284
|
+
try {
|
|
6285
|
+
const status = git(["status", "--porcelain"], workdir);
|
|
6286
|
+
return parseGitStatusPaths2(status).some((filePath) => !isToolNoisePath2(filePath));
|
|
6287
|
+
} catch {
|
|
6288
|
+
return false;
|
|
6289
|
+
}
|
|
6290
|
+
}
|
|
6291
|
+
function buildImplementationPrompt(context, extraLines = []) {
|
|
6292
|
+
return [
|
|
6293
|
+
basePrompt(context, "Implement the requested code change in the after worktree."),
|
|
6294
|
+
"Make the code changes directly in this repository.",
|
|
6295
|
+
"Use the recon-approved baseline understanding in the state to decide exactly what prior UI/state is being changed.",
|
|
6296
|
+
"Run focused checks when practical.",
|
|
6297
|
+
"If a focused check is blocked only by the harness sandbox writing Vite temp config into a shared/symlinked node_modules path (for example EROFS on node_modules/.vite-temp/vite.config...), do not treat that as an implementation blocker when the requested git diff exists. Record it in tests_run or implementation_notes instead so the harness can advance to its own verify stage.",
|
|
6298
|
+
"Leave a real git diff in the after worktree. Do not commit or push.",
|
|
6299
|
+
"Before you return success, run git status --short and git diff --name-only in this repository and verify they show the intended change.",
|
|
6300
|
+
"Do not return success with empty changed_files or a clean git status. If you cannot create the requested diff, explain why in blockers instead.",
|
|
6301
|
+
"changed_files should match the real git diff as closely as practical.",
|
|
6302
|
+
"Return changed_files, implementation_notes, tests_run, and blockers if any.",
|
|
6303
|
+
"Use empty arrays for tests_run or blockers when none apply.",
|
|
6304
|
+
...extraLines
|
|
6305
|
+
].join("\n");
|
|
6306
|
+
}
|
|
6307
|
+
function createCodexExecAgentAdapter(config = {}, runner = createCodexExecJsonRunner(config)) {
|
|
6308
|
+
return {
|
|
6309
|
+
async assessRecon(context) {
|
|
6310
|
+
const raw = await callRunner(runner, {
|
|
6311
|
+
purpose: "recon assessment",
|
|
6312
|
+
workdir: resolveWorkdir(context),
|
|
6313
|
+
schema: RECON_SCHEMA,
|
|
6314
|
+
prompt: [
|
|
6315
|
+
basePrompt(context, "Judge the latest recon packet."),
|
|
6316
|
+
"Decide whether recon should retry, continue to author, or escalate to the human.",
|
|
6317
|
+
"This is the first intelligent before-evidence gate. It runs before proof authoring and before any code edits.",
|
|
6318
|
+
"Prefer ready_for_author only when the baseline evidence is trustworthy for the requested change.",
|
|
6319
|
+
"Do not approve recon just because telemetry_ready is true or a screenshot URL exists.",
|
|
6320
|
+
"Inspect the baseline evidence in the state: screenshot URLs, structured pageState details, visible_text_sample, headings, buttons, links, canvas_count, large_visible_elements, observed_path, and route candidates.",
|
|
6321
|
+
"Fill baseline_understanding with concrete observations about what the before/prod evidence shows, what exact user request it anchors, what proof should focus on, and what stop condition would satisfy the request.",
|
|
6322
|
+
"If you cannot write a specific baseline_understanding from the current evidence, choose retry_recon or recon_stuck; do not choose ready_for_author.",
|
|
6323
|
+
"Reject blank, banner-only, app-shell-only, loading-only, generic landing, source-import path, or wrong-feature baselines; retry recon with a better server_path or selector instead.",
|
|
6324
|
+
"Your summary must say what is visibly present in the approved baseline or why the current baseline is not good enough.",
|
|
6325
|
+
"Always include refined_inputs; use null values for server_path, wait_for_selector, or reference when no refinement is needed."
|
|
6326
|
+
].join("\n")
|
|
6327
|
+
});
|
|
6328
|
+
return payloadOrBlocker(raw, context.checkpoint);
|
|
6329
|
+
},
|
|
6330
|
+
async authorProofPacket(context) {
|
|
6331
|
+
const raw = await callRunner(runner, {
|
|
6332
|
+
purpose: "proof packet authoring",
|
|
6333
|
+
workdir: resolveWorkdir(context),
|
|
6334
|
+
schema: AUTHOR_SCHEMA,
|
|
6335
|
+
prompt: [
|
|
6336
|
+
basePrompt(context, "Author the proof packet."),
|
|
6337
|
+
"Write a proof_plan and capture_script that will verify the exact user-facing change.",
|
|
6338
|
+
"Use recon_assessment.baseline_understanding as the source of truth. Do not author a proof plan unless it names the observed before state and the requested delta from that state.",
|
|
6339
|
+
"Use the recon-approved route and baseline context; make the plan name the concrete target, expected before state, expected after state, and stop condition.",
|
|
6340
|
+
"Choose the evidence modality from verification_mode and success_criteria: screenshots for visual/UI proof, interactions plus screenshots for interaction proof, structured metrics/logs/JSON/audio analysis for non-visual proof.",
|
|
6341
|
+
"For playable/gameplay proof, treat screenshots as supporting artifacts only: start the game, send keyboard or pointer input, measure state before/after, measure non-HUD canvas/playfield pixel deltas across time, and return playability evidence with version riddle-proof.playability.v1.",
|
|
6342
|
+
"For structured proof, collect meaningful measurements inside page.evaluate and set window.__riddleProofEvidence in that browser page context to a JSON-serializable object. Screenshots are optional supporting context for data/audio/log/metric/custom modes.",
|
|
6343
|
+
"Do not assign globalThis.__riddleProofEvidence, window.__riddleProofEvidence, or self.__riddleProofEvidence outside page.evaluate; the Riddle worker context may not expose those globals safely.",
|
|
6344
|
+
"Do not call Playwright page.* APIs inside page.evaluate; page.evaluate runs in the browser page, while page.waitForFunction, page.waitForSelector, page.click, and saveScreenshot belong in the outer capture script.",
|
|
6345
|
+
"When page.evaluate needs data from the outer Playwright script, pass exactly one serializable argument object; do not pass multiple positional arguments.",
|
|
6346
|
+
"When setting Playwright wait timeouts, use the correct signature: await page.waitForFunction(fn, undefined, { timeout: 60000 }). Do not write page.waitForFunction(fn, { timeout: 60000 }); that passes the object as the page argument and leaves the default timeout in place.",
|
|
6347
|
+
"For audio/data/metric proof, return and persist the same evidence object: const evidence = await page.evaluate(async () => { /* browser measurements */ window.__riddleProofEvidence = { ... }; return window.__riddleProofEvidence; }); Then assert evidence fields before saving screenshots.",
|
|
6348
|
+
"If a capture attempt errors or proof_evidence is absent, revise the capture_script to make structured evidence easier to collect; do not only add more screenshots for non-visual proof.",
|
|
6349
|
+
"When checking visible copy, normalize text before exact matching: const normalizedText = (document.body?.innerText || '').replace(/\\s+/g, ' ').trim(); Avoid raw innerText.includes(exact sentence) because browser line wrapping can split copy.",
|
|
6350
|
+
"Prefer success-oriented evidence booleans such as newCopyVisible: true and oldCopyAbsent: true. If a positive assertion is false while screenshots/text samples look right, fix the capture script instead of leaving contradictory proof evidence.",
|
|
6351
|
+
"Include a short matchedSnippet or normalizedTextSample when proofing copy so assertion mismatches are diagnosable.",
|
|
6352
|
+
"For visual/UI proof, include saveScreenshot('after-proof') exactly once.",
|
|
6353
|
+
"Avoid generic proof language. The packet should be specific enough that verify can tell whether the requested change actually happened.",
|
|
6354
|
+
"Echo the baseline understanding you used in baseline_understanding_used so later stages can detect drift.",
|
|
6355
|
+
"Use refined_inputs for server_path, wait_for_selector, or reference when useful; use null values when no refinement is needed."
|
|
6356
|
+
].join("\n")
|
|
6357
|
+
});
|
|
6358
|
+
return payloadOrBlocker(raw, context.checkpoint);
|
|
6359
|
+
},
|
|
6360
|
+
async implementChange(context) {
|
|
6361
|
+
if (!context.workdir || !(0, import_node_fs4.existsSync)(context.workdir)) {
|
|
6362
|
+
return {
|
|
6363
|
+
ok: false,
|
|
6364
|
+
blocker: {
|
|
6365
|
+
code: "implementation_workdir_missing",
|
|
6366
|
+
checkpoint: context.checkpoint,
|
|
6367
|
+
message: "The Riddle Proof state does not include an after worktree that exists on disk.",
|
|
6368
|
+
details: { workdir: context.workdir || null }
|
|
6369
|
+
}
|
|
6370
|
+
};
|
|
6371
|
+
}
|
|
6372
|
+
const workdir = context.workdir;
|
|
6373
|
+
const attemptSummaries = [];
|
|
6374
|
+
const runImplementationAttempt = async (purpose, extraLines = []) => {
|
|
6375
|
+
const raw = await callRunner(runner, {
|
|
6376
|
+
purpose,
|
|
6377
|
+
workdir,
|
|
6378
|
+
schema: IMPLEMENT_SCHEMA,
|
|
6379
|
+
prompt: buildImplementationPrompt(context, extraLines)
|
|
6380
|
+
});
|
|
6381
|
+
if (!raw.ok || !raw.json) {
|
|
6382
|
+
return {
|
|
6383
|
+
ok: false,
|
|
6384
|
+
payload: payloadOrBlocker(raw, context.checkpoint)
|
|
6385
|
+
};
|
|
6386
|
+
}
|
|
6387
|
+
const changedFiles = stringArray(raw.json.changed_files);
|
|
6388
|
+
const blockers = stringArray(raw.json.blockers);
|
|
6389
|
+
const testsRun = stringArray(raw.json.tests_run);
|
|
6390
|
+
const softVerificationBlockers = changedFiles.length ? blockers.filter(isHarnessVerificationOnlyBlocker) : [];
|
|
6391
|
+
const hardBlockers = blockers.filter((item) => !softVerificationBlockers.includes(item));
|
|
6392
|
+
const implementationNotesRaw = typeof raw.json.implementation_notes === "string" ? raw.json.implementation_notes : "";
|
|
6393
|
+
const summary = typeof raw.json.summary === "string" ? raw.json.summary : "";
|
|
6394
|
+
const implementationNotes = [
|
|
6395
|
+
implementationNotesRaw,
|
|
6396
|
+
...softVerificationBlockers.map((item) => `Harness verification note: ${item}`)
|
|
6397
|
+
].filter(Boolean).join("\n");
|
|
6398
|
+
const agentDetails = {
|
|
6399
|
+
agent_purpose: purpose,
|
|
6400
|
+
agent_summary: summary,
|
|
6401
|
+
agent_changed_files: changedFiles,
|
|
6402
|
+
agent_tests_run: testsRun,
|
|
6403
|
+
agent_blockers: blockers
|
|
6404
|
+
};
|
|
6405
|
+
attemptSummaries.push({
|
|
6406
|
+
purpose,
|
|
6407
|
+
summary,
|
|
6408
|
+
changed_files: changedFiles,
|
|
6409
|
+
tests_run: testsRun,
|
|
6410
|
+
blockers
|
|
6411
|
+
});
|
|
6412
|
+
if (hardBlockers.length) {
|
|
6413
|
+
return {
|
|
6414
|
+
ok: false,
|
|
6415
|
+
payload: {
|
|
6416
|
+
ok: false,
|
|
6417
|
+
blocker: {
|
|
6418
|
+
code: "codex_implementation_blocked",
|
|
6419
|
+
checkpoint: context.checkpoint,
|
|
6420
|
+
message: summary || "Codex reported implementation blockers.",
|
|
6421
|
+
details: {
|
|
6422
|
+
blockers: hardBlockers,
|
|
6423
|
+
changedFiles,
|
|
6424
|
+
testsRun,
|
|
6425
|
+
implementationNotes: implementationNotesRaw,
|
|
6426
|
+
...agentDetails
|
|
6427
|
+
}
|
|
6428
|
+
}
|
|
6429
|
+
}
|
|
6430
|
+
};
|
|
6431
|
+
}
|
|
6432
|
+
return {
|
|
6433
|
+
ok: true,
|
|
6434
|
+
summary,
|
|
6435
|
+
changedFiles,
|
|
6436
|
+
testsRun,
|
|
6437
|
+
implementationNotes,
|
|
6438
|
+
implementationNotesRaw,
|
|
6439
|
+
softVerificationBlockers,
|
|
6440
|
+
hardBlockers,
|
|
6441
|
+
agentDetails
|
|
6442
|
+
};
|
|
6443
|
+
};
|
|
6444
|
+
let attempt = await runImplementationAttempt("implementation");
|
|
6445
|
+
if (!attempt.ok) return attempt.payload;
|
|
6446
|
+
let diffDetected = hasGitDiff2(workdir);
|
|
6447
|
+
if (!diffDetected) {
|
|
6448
|
+
attempt = await runImplementationAttempt("implementation retry", [
|
|
6449
|
+
"The previous implementation attempt returned without a detectable git diff in this repository.",
|
|
6450
|
+
`Previous summary: ${attempt.summary || "(none)"}`,
|
|
6451
|
+
`Previous changed_files: ${JSON.stringify(attempt.changedFiles)}`,
|
|
6452
|
+
"Before you return success this time, run git status --short and git diff --name-only. If they are still empty, do not return success; either keep editing or explain the blocker."
|
|
6453
|
+
]);
|
|
6454
|
+
if (!attempt.ok) return attempt.payload;
|
|
6455
|
+
diffDetected = hasGitDiff2(workdir);
|
|
6456
|
+
}
|
|
6457
|
+
return {
|
|
6458
|
+
ok: true,
|
|
6459
|
+
summary: attempt.summary || void 0,
|
|
6460
|
+
implementationNotes: attempt.implementationNotes || void 0,
|
|
6461
|
+
changedFiles: attempt.changedFiles,
|
|
6462
|
+
testsRun: attempt.testsRun,
|
|
6463
|
+
details: {
|
|
6464
|
+
...attempt.agentDetails,
|
|
6465
|
+
implementation_notes: attempt.implementationNotes || attempt.implementationNotesRaw || "",
|
|
6466
|
+
soft_verification_blockers: attempt.softVerificationBlockers,
|
|
6467
|
+
hard_blocker_count: attempt.hardBlockers.length,
|
|
6468
|
+
retry_attempted: attemptSummaries.length > 1,
|
|
6469
|
+
attempt_count: attemptSummaries.length,
|
|
6470
|
+
attempt_summaries: attemptSummaries,
|
|
6471
|
+
post_agent_diff_detected: diffDetected
|
|
6472
|
+
}
|
|
6473
|
+
};
|
|
6474
|
+
},
|
|
6475
|
+
async assessProof(context) {
|
|
6476
|
+
const raw = await callRunner(runner, {
|
|
6477
|
+
purpose: "proof assessment",
|
|
6478
|
+
workdir: resolveWorkdir(context),
|
|
6479
|
+
schema: PROOF_SCHEMA,
|
|
6480
|
+
prompt: [
|
|
6481
|
+
basePrompt(context, "Judge the proof bundle."),
|
|
6482
|
+
"Decide whether the evidence is ready_to_ship or which internal stage should run next.",
|
|
6483
|
+
"Only use ready_to_ship when the before/prod and after evidence actually prove the requested change.",
|
|
6484
|
+
"Do not assume screenshots are required for every verification_mode. For data/audio/log/metric/custom proof, inspect the structured evidence bundle, proof_evidence_sample, artifacts, assertions, and success criteria directly.",
|
|
6485
|
+
"Use semantic_context.route, headings, buttons, and text anchors to ground route/content judgment before calling screenshot evidence wrong-route or unrelated.",
|
|
6486
|
+
"If structured proof evidence contains failed positive assertions, such as newCopyVisible: false or hasExpectedText: false, do not choose ready_to_ship until the evidence is reconciled or a richer proof is produced.",
|
|
6487
|
+
"For visual/UI proof, screenshots and route/page-state quality still matter, but capture success alone is not proof.",
|
|
6488
|
+
"For playable/gameplay proof, do not choose ready_to_ship unless playability evidence shows accepted input, state/time progression, and playfield/canvas pixel motion; a static generated scene is not playable proof.",
|
|
6489
|
+
"For visual/UI polish, reject changes that are technically different but not legible to a reviewer. If visual_delta.status is measured and visual_delta.passed is false, choose needs_implementation when the visible result is wrong, or needs_richer_proof only when non-visual proof artifacts are insufficient.",
|
|
6490
|
+
"If required visual_delta is unmeasured, missing, or not_applicable, choose revise_capture with recommended_stage=verify and continue_with_stage=verify so the same run stays in evidence/comparison recovery; do not choose ready_to_ship or generic needs_richer_proof for that missing metric.",
|
|
6491
|
+
"Do not ship if the baseline is blank, shell-only, generic, or unrelated to the requested feature.",
|
|
6492
|
+
"Your summary must name the concrete change, the target that was tested, what changed between baseline and after evidence, and why the stop condition is satisfied."
|
|
6493
|
+
].join("\n")
|
|
6494
|
+
});
|
|
6495
|
+
return payloadOrBlocker(raw, context.checkpoint);
|
|
6496
|
+
}
|
|
6497
|
+
};
|
|
6498
|
+
}
|
|
6499
|
+
function git(args, workdir) {
|
|
6500
|
+
return (0, import_node_child_process3.execFileSync)("git", args, {
|
|
6501
|
+
cwd: workdir,
|
|
6502
|
+
encoding: "utf-8",
|
|
6503
|
+
timeout: 1e4
|
|
6504
|
+
});
|
|
6505
|
+
}
|
|
6506
|
+
async function runCodexExecAgentDoctor(config = {}, runner = createCodexExecJsonRunner(config)) {
|
|
6507
|
+
const workdir = (0, import_node_fs4.mkdtempSync)(import_node_path4.default.join(import_node_os.default.tmpdir(), "riddle-proof-codex-doctor-"));
|
|
6508
|
+
const targetPath = import_node_path4.default.join(workdir, "target.txt");
|
|
6509
|
+
(0, import_node_fs4.writeFileSync)(targetPath, "color=red\n");
|
|
6510
|
+
try {
|
|
6511
|
+
git(["init", "-b", "main"], workdir);
|
|
6512
|
+
git(["config", "user.email", "doctor@example.com"], workdir);
|
|
6513
|
+
git(["config", "user.name", "Riddle Proof Doctor"], workdir);
|
|
6514
|
+
git(["add", "target.txt"], workdir);
|
|
6515
|
+
git(["commit", "-m", "initial doctor fixture"], workdir);
|
|
6516
|
+
} catch (error) {
|
|
6517
|
+
return {
|
|
6518
|
+
ok: false,
|
|
6519
|
+
status: "blocked",
|
|
6520
|
+
blocker: {
|
|
6521
|
+
code: "doctor_git_setup_failed",
|
|
6522
|
+
message: "Could not create the temporary doctor git fixture.",
|
|
6523
|
+
details: { error: error instanceof Error ? error.message : String(error), workdir }
|
|
6524
|
+
}
|
|
6525
|
+
};
|
|
6526
|
+
}
|
|
6527
|
+
const raw = await callRunner(runner, {
|
|
6528
|
+
purpose: "codex exec doctor implementation",
|
|
6529
|
+
workdir,
|
|
6530
|
+
schema: IMPLEMENT_SCHEMA,
|
|
6531
|
+
prompt: [
|
|
6532
|
+
"You are running a Riddle Proof Codex exec doctor check.",
|
|
6533
|
+
"In this temporary git repository, edit target.txt so it says exactly: color=blue",
|
|
6534
|
+
"Do not commit. Leave a git diff.",
|
|
6535
|
+
"Return JSON matching the schema with changed_files containing target.txt.",
|
|
6536
|
+
"Use empty arrays for tests_run and blockers."
|
|
6537
|
+
].join("\n")
|
|
6538
|
+
});
|
|
6539
|
+
if (!raw.ok || !raw.json) {
|
|
6540
|
+
return {
|
|
6541
|
+
ok: false,
|
|
6542
|
+
status: "blocked",
|
|
6543
|
+
blocker: {
|
|
6544
|
+
...raw.blocker || {
|
|
6545
|
+
code: "doctor_codex_failed",
|
|
6546
|
+
message: "Codex doctor runner failed.",
|
|
6547
|
+
details: { stdout: raw.stdout || "", stderr: raw.stderr || "" }
|
|
6548
|
+
},
|
|
6549
|
+
details: {
|
|
6550
|
+
...raw.blocker?.details || {},
|
|
6551
|
+
workdir
|
|
6552
|
+
}
|
|
6553
|
+
}
|
|
6554
|
+
};
|
|
6555
|
+
}
|
|
6556
|
+
const content = (0, import_node_fs4.readFileSync)(targetPath, "utf-8");
|
|
6557
|
+
const status = git(["status", "--porcelain"], workdir);
|
|
6558
|
+
const changedFiles = stringArray(raw.json.changed_files);
|
|
6559
|
+
const ok = content.trim() === "color=blue" && status.includes("target.txt");
|
|
6560
|
+
return {
|
|
6561
|
+
ok,
|
|
6562
|
+
status: ok ? "passed" : "blocked",
|
|
6563
|
+
workdir,
|
|
6564
|
+
changedFiles,
|
|
6565
|
+
implementationSummary: raw.json.summary || null,
|
|
6566
|
+
implementationNotes: raw.json.implementation_notes || null,
|
|
6567
|
+
gitStatus: status,
|
|
6568
|
+
targetContent: content,
|
|
6569
|
+
blocker: ok ? null : {
|
|
6570
|
+
code: "doctor_diff_missing_or_wrong",
|
|
6571
|
+
message: "Codex doctor completed, but target.txt was not changed to color=blue with a git diff.",
|
|
6572
|
+
details: { workdir, changedFiles, gitStatus: status, targetContent: content }
|
|
6573
|
+
}
|
|
6574
|
+
};
|
|
6575
|
+
}
|
|
6576
|
+
|
|
5361
6577
|
// src/diagnostics.ts
|
|
5362
6578
|
var RIDDLE_PROOF_CAPTURE_DIAGNOSTIC_VERSION = "riddle-proof.capture-diagnostic.v1";
|
|
5363
6579
|
var DEFAULT_DIAGNOSTIC_STRING_LIMIT = 2e3;
|
|
@@ -5975,6 +7191,7 @@ function parseJson(value) {
|
|
|
5975
7191
|
RIDDLE_PROOF_CHECKPOINT_RESPONSE_VERSION,
|
|
5976
7192
|
RIDDLE_PROOF_PLAYABILITY_ASSESSMENT_VERSION,
|
|
5977
7193
|
RIDDLE_PROOF_PLAYABILITY_VERSION,
|
|
7194
|
+
RIDDLE_PROOF_RUN_CARD_VERSION,
|
|
5978
7195
|
RIDDLE_PROOF_RUN_STATE_VERSION,
|
|
5979
7196
|
RIDDLE_PROOF_VISUAL_SESSION_FINGERPRINT_VERSION,
|
|
5980
7197
|
RIDDLE_PROOF_VISUAL_SESSION_VERSION,
|
|
@@ -5988,13 +7205,17 @@ function parseJson(value) {
|
|
|
5988
7205
|
buildAuthorCheckpointPacket,
|
|
5989
7206
|
buildCheckpointPacketForEngineResult,
|
|
5990
7207
|
buildProofAssessmentCheckpointPacket,
|
|
7208
|
+
buildStageCheckpointPacket,
|
|
5991
7209
|
buildVisualProofSession,
|
|
5992
7210
|
checkpointResponseIdentity,
|
|
5993
7211
|
checkpointSummaryFromState,
|
|
5994
7212
|
compactRecord,
|
|
5995
7213
|
compareVisualProofSessionFingerprint,
|
|
5996
7214
|
createCaptureDiagnostic,
|
|
7215
|
+
createCodexExecAgentAdapter,
|
|
7216
|
+
createCodexExecJsonRunner,
|
|
5997
7217
|
createDisabledRiddleProofAgentAdapter,
|
|
7218
|
+
createRiddleProofRunCard,
|
|
5998
7219
|
createRunResult,
|
|
5999
7220
|
createRunState,
|
|
6000
7221
|
createRunStatusSnapshot,
|
|
@@ -6014,6 +7235,7 @@ function parseJson(value) {
|
|
|
6014
7235
|
readRiddleProofRunStatus,
|
|
6015
7236
|
recordValue,
|
|
6016
7237
|
redactForProofDiagnostics,
|
|
7238
|
+
runCodexExecAgentDoctor,
|
|
6017
7239
|
runRiddleProof,
|
|
6018
7240
|
runRiddleProofEngineHarness,
|
|
6019
7241
|
setRunStatus,
|