@riddledc/riddle-proof 0.5.48 → 0.5.50

package/dist/cli.cjs

@@ -3650,6 +3650,68 @@ function compactText2(value, limit = 600) {
   if (!text) return void 0;
   return text.length <= limit ? text : `${text.slice(0, limit - 20).trimEnd()}...`;
 }
+function numericValue(value) {
+  const number = typeof value === "number" ? value : typeof value === "string" ? Number(value) : NaN;
+  return Number.isFinite(number) ? number : void 0;
+}
+function eventDetails(event) {
+  return recordValue(recordValue(event)?.details) || {};
+}
+function collectRunnerMetrics(details) {
+  const metrics = [];
+  const adapterDetails = recordValue(details.adapter_details);
+  const direct = recordValue(adapterDetails?.runner_metrics);
+  if (direct) metrics.push(direct);
+  const attempts = Array.isArray(adapterDetails?.attempt_summaries) ? adapterDetails.attempt_summaries : [];
+  for (const attempt of attempts) {
+    const attemptMetrics = recordValue(recordValue(attempt)?.runner_metrics);
+    if (attemptMetrics) metrics.push(attemptMetrics);
+  }
+  return metrics;
+}
+function observabilityFrom(state) {
+  const engineEvents = state.events.filter((event) => event.kind === "engine.result").map((event) => ({ event, details: eventDetails(event) }));
+  const agentEvents = state.events.filter((event) => event.kind.startsWith("agent.")).map((event) => ({ event, details: eventDetails(event) })).filter(({ details }) => numericValue(details.duration_ms) !== void 0);
+  const retryEvents = state.events.filter((event) => /retry|recovery/i.test(event.kind) || /retry|recovery/i.test(event.summary || "")).slice(-8);
+  const runnerMetrics2 = agentEvents.flatMap(({ details }) => collectRunnerMetrics(details));
+  const promptChars = runnerMetrics2.map((metrics) => numericValue(metrics.prompt_chars)).filter((value) => value !== void 0);
+  const sum = (values) => values.reduce((total, value) => total + (value ?? 0), 0);
+  const recentEngineTimings = engineEvents.slice(-8).map(({ event, details }) => compactRecord({
+    checkpoint: event.checkpoint || null,
+    stage: event.stage || null,
+    duration_ms: numericValue(details.duration_ms),
+    ok: details.ok ?? null
+  }));
+  const recentAgentTimings = agentEvents.slice(-8).map(({ event, details }) => {
+    const metrics = collectRunnerMetrics(details);
+    const localPromptChars = metrics.map((item) => numericValue(item.prompt_chars)).filter((value) => value !== void 0);
+    return compactRecord({
+      kind: event.kind,
+      checkpoint: event.checkpoint || null,
+      stage: event.stage || null,
+      duration_ms: numericValue(details.duration_ms),
+      prompt_chars: localPromptChars.length ? Math.max(...localPromptChars) : void 0,
+      attempt_count: numericValue(recordValue(details.adapter_details)?.attempt_count)
+    });
+  });
+  return compactRecord({
+    engine_call_count: engineEvents.length,
+    agent_call_count: agentEvents.length,
+    engine_total_ms: sum(engineEvents.map(({ details }) => numericValue(details.duration_ms))),
+    agent_total_ms: sum(agentEvents.map(({ details }) => numericValue(details.duration_ms))),
+    max_agent_prompt_chars: promptChars.length ? Math.max(...promptChars) : void 0,
+    total_agent_prompt_chars: promptChars.length ? sum(promptChars) : void 0,
+    recent_engine_timings: recentEngineTimings.length ? recentEngineTimings : void 0,
+    recent_agent_timings: recentAgentTimings.length ? recentAgentTimings : void 0,
+    retry_event_count: retryEvents.length,
+    recent_retry_reasons: retryEvents.length ? retryEvents.map((event) => compactRecord({
+      kind: event.kind,
+      checkpoint: event.checkpoint || null,
+      stage: event.stage || null,
+      summary: compactText2(event.summary, 240)
+    })) : void 0
+  });
+}
 function visualDeltaFrom(input) {
   const fullState = input.fullRiddleState || {};
   const bundle = recordValue(fullState.evidence_bundle);
@@ -3762,6 +3824,7 @@ function createRiddleProofRunCard(state, input = {}) {
       proof_evidence_present: fullState.proof_evidence_present === true || Boolean(bundle?.proof_evidence || bundle?.proof_evidence_sample),
       artifacts
     }),
+    observability: observabilityFrom(state),
     stop_condition: compactRecord({
       status: state.status,
       terminal: isTerminalStatus(state.status),
@@ -4816,7 +4879,10 @@ async function handleImplementation(request, state, result, agent) {
       worktree_path: workdir || null
     }
   });
+  const implementationStartedAt = timestamp3();
+  const implementationStartedMs = Date.now();
   const implementation = await agent.implementChange({ ...context, workdir });
+  const implementationDurationMs = Date.now() - implementationStartedMs;
   if (implementation.blocker || implementation.ok === false) {
     recordEvent(state, {
       kind: "agent.implementation.blocked",
@@ -4825,6 +4891,8 @@ async function handleImplementation(request, state, result, agent) {
       summary: implementation.summary || implementation.blocker?.message || "Implementation adapter reported a blocker.",
       details: compactRecord({
         worktree_path: workdir || null,
+        started_at: implementationStartedAt,
+        duration_ms: implementationDurationMs,
         changed_files: implementation.changedFiles || [],
         tests_run: implementation.testsRun || [],
         implementation_notes: implementation.implementationNotes || null,
@@ -4850,6 +4918,8 @@ async function handleImplementation(request, state, result, agent) {
       summary: implementation.summary || "Implementation adapter returned without leaving a detectable git diff.",
       details: compactRecord({
         worktree_path: workdir || null,
+        started_at: implementationStartedAt,
+        duration_ms: implementationDurationMs,
         changed_files: implementation.changedFiles || [],
         tests_run: implementation.testsRun || [],
         implementation_notes: implementation.implementationNotes || null,
@@ -4864,7 +4934,8 @@ async function handleImplementation(request, state, result, agent) {
       details: compactRecord({
         worktree_path: workdir || null,
         checkpoint: result.checkpoint || null,
-        next_stage: "implement"
+        next_stage: "implement",
+        previous_duration_ms: implementationDurationMs
       })
     });
     return {
@@ -4883,6 +4954,8 @@ async function handleImplementation(request, state, result, agent) {
     summary: implementation.summary || "Implementation adapter reported code changes.",
     details: {
       worktree_path: workdir || null,
+      started_at: implementationStartedAt,
+      duration_ms: implementationDurationMs,
       diffDetected,
       changed_files: implementation.changedFiles || [],
       tests_run: implementation.testsRun || [],
@@ -4999,15 +5072,44 @@ async function routeCheckpoint(request, state, result, agent, input) {
     if (input.checkpoint_mode === "yield") {
       return { terminal: checkpointAwaitingResult(state, result, input.checkpoint_visibility) };
     }
+    const startedAt = timestamp3();
+    const startedMs = Date.now();
+    recordEvent(state, {
+      kind: "agent.recon_assessment.started",
+      checkpoint,
+      stage: "recon",
+      summary: "Recon assessment agent started.",
+      details: { started_at: startedAt }
+    });
     const assessment = await agent.assessRecon(context);
+    const durationMs = Date.now() - startedMs;
     const blocker = requirePayload("recon_assessment", assessment, state, result);
-    if (blocker) return { blocker };
+    if (blocker) {
+      recordEvent(state, {
+        kind: "agent.recon_assessment.blocked",
+        checkpoint,
+        stage: "recon",
+        summary: blocker.message,
+        details: compactRecord({
+          started_at: startedAt,
+          duration_ms: durationMs,
+          blocker,
+          adapter_details: assessment.details || null
+        })
+      });
+      return { blocker };
+    }
     recordEvent(state, {
       kind: "agent.recon_assessment.completed",
       checkpoint,
       stage: "recon",
       summary: assessment.summary,
-      details: { payload: assessment.payload }
+      details: compactRecord({
+        payload: assessment.payload,
+        started_at: startedAt,
+        duration_ms: durationMs,
+        adapter_details: assessment.details || null
+      })
     });
     return {
       next: { ...baseContinuation(result), recon_assessment_json: jsonParam(assessment.payload) }
@@ -5019,15 +5121,44 @@ async function routeCheckpoint(request, state, result, agent, input) {
     if (input.checkpoint_mode === "yield") {
       return { terminal: checkpointAwaitingResult(state, result, input.checkpoint_visibility) };
     }
+    const startedAt = timestamp3();
+    const startedMs = Date.now();
+    recordEvent(state, {
+      kind: "agent.author_packet.started",
+      checkpoint,
+      stage: "author",
+      summary: "Proof authoring agent started.",
+      details: { started_at: startedAt }
+    });
     const packet = await agent.authorProofPacket(context);
+    const durationMs = Date.now() - startedMs;
     const blocker = requirePayload("author_packet", packet, state, result);
-    if (blocker) return { blocker };
+    if (blocker) {
+      recordEvent(state, {
+        kind: "agent.author_packet.blocked",
+        checkpoint,
+        stage: "author",
+        summary: blocker.message,
+        details: compactRecord({
+          started_at: startedAt,
+          duration_ms: durationMs,
+          blocker,
+          adapter_details: packet.details || null
+        })
+      });
+      return { blocker };
+    }
     recordEvent(state, {
       kind: "agent.author_packet.completed",
       checkpoint,
       stage: "author",
       summary: packet.summary,
-      details: { payload: packet.payload }
+      details: compactRecord({
+        payload: packet.payload,
+        started_at: startedAt,
+        duration_ms: durationMs,
+        adapter_details: packet.details || null
+      })
     });
     return {
       next: { ...baseContinuation(result), author_packet_json: jsonParam(packet.payload) }
@@ -5050,9 +5181,31 @@ async function routeCheckpoint(request, state, result, agent, input) {
     if (input.checkpoint_mode === "yield") {
       return { terminal: checkpointAwaitingResult(state, result, input.checkpoint_visibility) };
     }
+    const startedAt = timestamp3();
+    const startedMs = Date.now();
+    recordEvent(state, {
+      kind: "agent.proof_assessment.started",
+      checkpoint,
+      stage: "verify",
+      summary: "Proof assessment agent started.",
+      details: { started_at: startedAt }
+    });
     const assessment = await agent.assessProof(context);
+    const durationMs = Date.now() - startedMs;
     const blocker = requirePayload("proof_assessment", assessment, state, result);
     if (blocker) {
+      recordEvent(state, {
+        kind: "agent.proof_assessment.blocked",
+        checkpoint,
+        stage: "verify",
+        summary: blocker.message,
+        details: compactRecord({
+          started_at: startedAt,
+          duration_ms: durationMs,
+          blocker,
+          adapter_details: assessment.details || null
+        })
+      });
       if (blocker.code === "main_agent_proof_review_required") {
         recordEvent(state, {
           kind: "checkpoint.packet.requested",
@@ -5071,7 +5224,12 @@ async function routeCheckpoint(request, state, result, agent, input) {
       checkpoint,
       stage: "verify",
       summary: assessment.summary,
-      details: { payload }
+      details: compactRecord({
+        payload,
+        started_at: startedAt,
+        duration_ms: durationMs,
+        adapter_details: assessment.details || null
+      })
     });
     const visualBlocker = proofAssessmentVisualBlocker({
       ...context.fullRiddleState || {},
@@ -5092,7 +5250,8 @@ async function routeCheckpoint(request, state, result, agent, input) {
           recovery_stage: "verify",
           evidence_issue_code: recoveryAssessment.evidence_issue_code || null,
           visual_delta: recoveryAssessment.visual_delta || null,
-          proof_assessment: recoveryAssessment
+          proof_assessment: recoveryAssessment,
+          agent_duration_ms: durationMs
         })
       });
       return { next: proofAssessmentContinuation(result, recoveryAssessment) };
@@ -5450,10 +5609,11 @@ var PROOF_SCHEMA = {
     source: { type: "string", enum: ["supervising_agent"] }
   }
 };
-var PROMPT_STRING_LIMIT = 2e3;
-var PROMPT_ARRAY_LIMIT = 12;
-var PROMPT_OBJECT_KEY_LIMIT = 70;
-var PROMPT_BLOCK_LIMIT = 12e4;
+var PROMPT_STRING_LIMIT = 1e3;
+var PROMPT_ARRAY_LIMIT = 8;
+var PROMPT_OBJECT_KEY_LIMIT = 50;
+var PROMPT_BLOCK_LIMIT = 16e3;
+var PROMPT_TOTAL_LIMIT = 58e3;
 var PROMPT_KEY_PRIORITY = [
   "ok",
   "status",
@@ -5475,6 +5635,16 @@ var PROMPT_KEY_PRIORITY = [
   "after_cdn",
   "before_baseline",
   "prod_baseline",
+  "route_hints",
+  "route_candidates",
+  "keyword_hits",
+  "observations",
+  "latest_attempt",
+  "attempt_history",
+  "current_plan",
+  "plan_history",
+  "decision_history",
+  "refined_inputs",
   "recon_assessment",
   "baseline_understanding",
   "supervisor_author_packet",
@@ -5498,6 +5668,7 @@ var PROMPT_KEY_PRIORITY = [
   "shipGate",
   "last_error",
   "errors",
+  "runtime_events",
   "events"
 ];
 var PROMPT_PRIORITY_INDEX = new Map(PROMPT_KEY_PRIORITY.map((key, index) => [key, index]));
@@ -5514,7 +5685,7 @@ function compactPromptValue(value, depth = 0, key = "") {
     if (!looksLikeUrl && (lowerKey.includes("base64") || lowerKey.includes("data_url") || lowerKey.includes("screenshot_blob"))) {
       return `[omitted ${value.length} chars from ${key || "large artifact"}]`;
     }
-    const limit = looksLikeUrl ? 1e3 : depth <= 1 ? PROMPT_STRING_LIMIT : 1200;
+    const limit = looksLikeUrl ? 1e3 : depth <= 1 ? PROMPT_STRING_LIMIT : 900;
     return truncatePromptString(value, limit);
   }
   if (Array.isArray(value)) {
@@ -5557,7 +5728,7 @@ function resolveWorkdir(context, fallback = "/tmp") {
   return after || fallback;
 }
 function basePrompt(context, role) {
-  return [
+  const prompt = [
     role,
     "",
     "You are the supervising Codex worker inside the Riddle Proof harness.",
@@ -5569,6 +5740,9 @@ function basePrompt(context, role) {
     jsonBlock("Riddle checkpoint result", context.engineResult),
     jsonBlock("Full riddle state", context.fullRiddleState || {})
   ].join("\n");
+  if (prompt.length <= PROMPT_TOTAL_LIMIT) return prompt;
+  return `${prompt.slice(0, PROMPT_TOTAL_LIMIT).trimEnd()}
+...[truncated ${prompt.length - PROMPT_TOTAL_LIMIT} chars from total prompt]`;
 }
 function schemaRequiredKeys(schema) {
   const required = schema?.required;
@@ -5645,11 +5819,39 @@ function isHarnessVerificationOnlyBlocker(blocker) {
   const text = blocker.toLowerCase();
   return (text.includes("erofs") || text.includes("read-only file system")) && text.includes("node_modules") && (text.includes(".vite-temp") || text.includes("vite.config"));
 }
+function runnerMetrics(input) {
+  const schemaText = JSON.stringify(input.request.schema);
+  const finishedAt = (/* @__PURE__ */ new Date()).toISOString();
+  return compactRecord({
+    purpose: input.request.purpose,
+    workdir: input.request.workdir,
+    started_at: input.startedAt,
+    finished_at: finishedAt,
+    duration_ms: Date.now() - input.startedMs,
+    prompt_chars: input.request.prompt.length,
+    prompt_lines: input.request.prompt.split(/\r?\n/).length,
+    schema_chars: schemaText.length,
+    stdout_chars: (input.stdout || "").length,
+    stderr_chars: (input.stderr || "").length,
+    final_message_chars: (input.finalText || "").length,
+    exit_status: input.status ?? null,
+    timed_out: input.timedOut || false,
+    error_code: input.errorCode,
+    codex_command: input.config.codexCommand || "codex",
+    codex_model: input.config.codexModel,
+    codex_sandbox: input.config.codexSandbox || "workspace-write",
+    codex_full_auto: input.config.codexFullAuto !== false,
+    timeout_ms: Number(input.config.codexTimeoutMs || 6e5)
+  });
+}
 function createCodexExecJsonRunner(config = {}) {
   return (request) => {
+    const startedAt = (/* @__PURE__ */ new Date()).toISOString();
+    const startedMs = Date.now();
     if (!request.workdir || !(0, import_node_fs4.existsSync)(request.workdir)) {
       return {
         ok: false,
+        metrics: runnerMetrics({ request, config, startedAt, startedMs, errorCode: "workdir_missing" }),
         blocker: {
           code: "codex_workdir_missing",
           message: `Codex workdir does not exist for ${request.purpose}.`,
@@ -5694,6 +5896,17 @@ function createCodexExecJsonRunner(config = {}) {
           ok: false,
           stdout: proc.stdout || "",
           stderr: proc.stderr || "",
+          metrics: runnerMetrics({
+            request,
+            config,
+            startedAt,
+            startedMs,
+            stdout: proc.stdout || "",
+            stderr: proc.stderr || "",
+            status: proc.status,
+            timedOut,
+            errorCode: proc.error.code || "spawn_error"
+          }),
           blocker: {
             code: timedOut ? "codex_timeout" : "codex_exec_error",
             message: timedOut ? `Codex timed out during ${request.purpose}.` : `Codex failed to start or complete ${request.purpose}.`,
@@ -5706,6 +5919,16 @@ function createCodexExecJsonRunner(config = {}) {
           ok: false,
           stdout: proc.stdout || "",
           stderr: proc.stderr || "",
+          metrics: runnerMetrics({
+            request,
+            config,
+            startedAt,
+            startedMs,
+            stdout: proc.stdout || "",
+            stderr: proc.stderr || "",
+            status: proc.status,
+            errorCode: "nonzero_exit"
+          }),
           blocker: {
             code: "codex_nonzero_exit",
             message: `Codex exited with status ${proc.status} during ${request.purpose}.`,
@@ -5720,6 +5943,17 @@ function createCodexExecJsonRunner(config = {}) {
           ok: false,
           stdout: proc.stdout || "",
           stderr: proc.stderr || "",
+          metrics: runnerMetrics({
+            request,
+            config,
+            startedAt,
+            startedMs,
+            stdout: proc.stdout || "",
+            stderr: proc.stderr || "",
+            finalText,
+            status: proc.status,
+            errorCode: "invalid_json"
+          }),
           blocker: {
             code: "codex_invalid_json",
             message: `Codex completed ${request.purpose}, but did not return valid JSON.`,
@@ -5731,7 +5965,17 @@ function createCodexExecJsonRunner(config = {}) {
         ok: true,
         json: parsed,
         stdout: proc.stdout || "",
-        stderr: proc.stderr || ""
+        stderr: proc.stderr || "",
+        metrics: runnerMetrics({
+          request,
+          config,
+          startedAt,
+          startedMs,
+          stdout: proc.stdout || "",
+          stderr: proc.stderr || "",
+          finalText,
+          status: proc.status
+        })
       };
     } finally {
       (0, import_node_fs4.rmSync)(tmpDir, { recursive: true, force: true });
@@ -5752,14 +5996,21 @@ function payloadOrBlocker(raw, checkpoint) {
       ok: false,
       blocker: {
         ...blocker,
-        checkpoint
+        checkpoint,
+        details: {
+          ...blocker.details || {},
+          runner_metrics: raw.metrics || null
+        }
       }
     };
   }
   return {
     ok: true,
     payload: raw.json,
-    summary: typeof raw.json.summary === "string" ? raw.json.summary : void 0
+    summary: typeof raw.json.summary === "string" ? raw.json.summary : void 0,
+    details: compactRecord({
+      runner_metrics: raw.metrics || null
+    })
   };
 }
 function stringArray(value) {
@@ -5895,14 +6146,16 @@ function createCodexExecAgentAdapter(config = {}, runner = createCodexExecJsonRu
           agent_summary: summary,
           agent_changed_files: changedFiles,
           agent_tests_run: testsRun,
-          agent_blockers: blockers
+          agent_blockers: blockers,
+          runner_metrics: raw.metrics || null
         };
         attemptSummaries.push({
           purpose,
           summary,
           changed_files: changedFiles,
           tests_run: testsRun,
-          blockers
+          blockers,
+          runner_metrics: raw.metrics || null
         });
         if (hardBlockers.length) {
           return {

package/dist/cli.js

@@ -3,15 +3,15 @@ import {
   createDisabledRiddleProofAgentAdapter,
   readRiddleProofRunStatus,
   runRiddleProofEngineHarness
-} from "./chunk-2RS4PBYK.js";
+} from "./chunk-722PW3X3.js";
 import "./chunk-4ASMX4R6.js";
 import "./chunk-JFQXAJH2.js";
 import {
   createCodexExecAgentAdapter,
   runCodexExecAgentDoctor
-} from "./chunk-NOBFZDZG.js";
-import "./chunk-JOXTKWX6.js";
-import "./chunk-PLSGW2GI.js";
+} from "./chunk-3266V3MO.js";
+import "./chunk-U7Q3RB5D.js";
+import "./chunk-CI2F66EE.js";
 import "./chunk-R6SCWJCI.js";
 import "./chunk-DUFDZJOF.js";