pi-smart-compact 9.6.1 → 9.6.2

package/dist/infra/ai-messages.d.ts

@@ -15,9 +15,13 @@
  * conversation buffer in native `Message[]` directly (see `phases/explore.ts`),
  * which is the preferred pattern for any new code that talks to a provider.
  */
-import type { Message } from "@earendil-works/pi-ai";
-import type { LlmMessage } from "../types.ts";
+import { type Message } from "@earendil-works/pi-ai";
+import type { LlmMessage, SessionMessageEntry } from "../types.ts";
 import type { SecretScrubber } from "../domain/scrub.ts";
+/** Preserve the host's context projection and original entry IDs, including custom and branch summaries. */
+export declare function contextMessageEntries(entries: readonly unknown[]): SessionMessageEntry[];
+/** Text transcript without the host summarizer's implicit 2,000-character tool-result cap. */
+export declare function serializeConversationText(messages: LlmMessage[]): string;
 /**
  * Upcast a raw branch entry's `message` to a `Message` for `convertToLlm`.
  *

package/dist/infra/ai-messages.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ai-messages.d.ts","sourceRoot":"","sources":["../../src/infra/ai-messages.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,uBAAuB,CAAC;AACrD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAC9C,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAC;AAEzD;;;;;;;;GAQG;AACH,wBAAgB,eAAe,CAAC,OAAO,EAAE,OAAO,GAAG,OAAO,CAEzD;AAED;;;;;;;;GAQG;AACH,wBAAgB,sBAAsB,CAAC,IAAI,EAAE,UAAU,EAAE,GAAG,OAAO,EAAE,CAIpE;AAED;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,cAAc,GAAG,UAAU,EAAE,CAE3F"}
+{"version":3,"file":"ai-messages.d.ts","sourceRoot":"","sources":["../../src/infra/ai-messages.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAe,KAAK,OAAO,EAAE,MAAM,uBAAuB,CAAC;AAClE,OAAO,KAAK,EAAE,UAAU,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAEnE,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAC;AAEzD,4GAA4G;AAC5G,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,SAAS,OAAO,EAAE,GAAG,mBAAmB,EAAE,CAKxF;AAED,8FAA8F;AAC9F,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,UAAU,EAAE,GAAG,MAAM,CAKxE;AAED;;;;;;;;GAQG;AACH,wBAAgB,eAAe,CAAC,OAAO,EAAE,OAAO,GAAG,OAAO,CAEzD;AAED;;;;;;;;GAQG;AACH,wBAAgB,sBAAsB,CAAC,IAAI,EAAE,UAAU,EAAE,GAAG,OAAO,EAAE,CAIpE;AAED;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,cAAc,GAAG,UAAU,EAAE,CAE3F"}

package/dist/phases/verify.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"verify.d.ts","sourceRoot":"","sources":["../../src/phases/verify.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EACX,KAAK,EACL,GAAG,EACH,eAAe,EAEf,MAAM,uBAAuB,CAAC;AAC/B,OAAO,KAAK,EACX,eAAe,EACf,oBAAoB,EACpB,eAAe,EACf,qBAAqB,EACrB,kBAAkB,EAClB,UAAU,EACV,MAAM,aAAa,CAAC;AAoCrB,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AACjE,MAAM,WAAW,oBAAoB;IACpC,cAAc,CAAC,EAAE,SAAS,UAAU,EAAE,CAAC;IACvC,QAAQ,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,IAAI,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC7C,mBAAmB,CAAC,EAAE,MAAM,CAAC;CAC7B;AA8SD,wBAAgB,qBAAqB,CAAC,GAAG,EAAE,eAAe,GAAG,MAAM,CAkClE;AAED,wBAAgB,0BAA0B,CACzC,MAAM,EAAE,kBAAkB,GACxB,MAAM,GAAG,IAAI,CAcf;AAED,wFAAwF;AACxF,qBAAa,qBAAsB,SAAQ,KAAK;IAC/C,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,QAAQ,EAAE,eAAe,CAAC,MAAM,CAAC,EAAE,CAAC;IAC7C,QAAQ,CAAC,KAAK,EAAE,qBAAqB,CAAC;IACtC,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAE1B,YACC,MAAM,EAAE,kBAAkB,EAC1B,YAAY,EAAE,MAAM,EACpB,KAAK,EAAE,qBAAqB,EAW5B;CACD;AAwUD,wBAAgB,4BAA4B,CAAC,GAAG,EAAE,eAAe,GAAG,OAAO,CAK1E;AAED,qGAAqG;AACrG,wBAAgB,8BAA8B,CAC7C,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,kBAAkB,EAC1B,UAAU,EAAE,oBAAoB,EAChC,UAAU,GAAE,eAAe,GAAG,IAAW,EACzC,QAAQ,GAAE,oBAAyB,EACnC,SAAS,SAAI,GACX;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,kBAAkB,CAAC;IAAC,OAAO,EAAE,eAAe,EAAE,CAAA;CAAE,CAyB7E;AAiYD,wBAAgB,aAAa,CAC5B,OAAO,EAAE,MAAM,EACf,UAAU,EAAE,oBAAoB,EAChC,UAAU,GAAE,eAAe,GAAG,IAAW,EACzC,QAAQ,GAAE,oBAAyB,GACjC,kBAAkB,CA2CpB;AAED,iHAAiH;AACjH,wBAAgB,kBAAkB,CACjC,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,eAAe,EAAE,EACvB,UAAU,EAAE,oBAAoB,EAChC,UAAU,GAAE,eAAe,GAAG,IAAW,EACzC,QAAQ,GAAE,oBAAyB,GACjC,MAAM,CA4MR;AA2CD,wBAAsB,YAAY,CACjC,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,eAAe,EAAE,EACvB,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,IAAI,EAAE;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,eAAe,CAAA;CAAE,EACnD,MAAM,CAAC,EAAE,WAAW,EACpB,QAAQ,CAAC,EAAE,oBAAoB,GAC7B,OAAO,CAAC,MAAM,CAAC,CA2DjB"}
+{"version":3,"file":"verify.d.ts","sourceRoot":"","sources":["../../src/phases/verify.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EACX,KAAK,EACL,GAAG,EACH,eAAe,EAEf,MAAM,uBAAuB,CAAC;AAC/B,OAAO,KAAK,EACX,eAAe,EACf,oBAAoB,EACpB,eAAe,EACf,qBAAqB,EACrB,kBAAkB,EAClB,UAAU,EACV,MAAM,aAAa,CAAC;AAoCrB,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AACjE,MAAM,WAAW,oBAAoB;IACpC,cAAc,CAAC,EAAE,SAAS,UAAU,EAAE,CAAC;IACvC,QAAQ,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,IAAI,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC7C,mBAAmB,CAAC,EAAE,MAAM,CAAC;CAC7B;AAiTD,wBAAgB,qBAAqB,CAAC,GAAG,EAAE,eAAe,GAAG,MAAM,CAkClE;AAED,wBAAgB,0BAA0B,CACzC,MAAM,EAAE,kBAAkB,GACxB,MAAM,GAAG,IAAI,CAcf;AAED,wFAAwF;AACxF,qBAAa,qBAAsB,SAAQ,KAAK;IAC/C,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,QAAQ,EAAE,eAAe,CAAC,MAAM,CAAC,EAAE,CAAC;IAC7C,QAAQ,CAAC,KAAK,EAAE,qBAAqB,CAAC;IACtC,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAE1B,YACC,MAAM,EAAE,kBAAkB,EAC1B,YAAY,EAAE,MAAM,EACpB,KAAK,EAAE,qBAAqB,EAW5B;CACD;AA4UD,wBAAgB,4BAA4B,CAAC,GAAG,EAAE,eAAe,GAAG,OAAO,CAK1E;AAED,qGAAqG;AACrG,wBAAgB,8BAA8B,CAC7C,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,kBAAkB,EAC1B,UAAU,EAAE,oBAAoB,EAChC,UAAU,GAAE,eAAe,GAAG,IAAW,EACzC,QAAQ,GAAE,oBAAyB,EACnC,SAAS,SAAI,GACX;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,kBAAkB,CAAC;IAAC,OAAO,EAAE,eAAe,EAAE,CAAA;CAAE,CAyB7E;AAkYD,wBAAgB,aAAa,CAC5B,OAAO,EAAE,MAAM,EACf,UAAU,EAAE,oBAAoB,EAChC,UAAU,GAAE,eAAe,GAAG,IAAW,EACzC,QAAQ,GAAE,oBAAyB,GACjC,kBAAkB,CA4CpB;AAED,iHAAiH;AACjH,wBAAgB,kBAAkB,CACjC,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,eAAe,EAAE,EACvB,UAAU,EAAE,oBAAoB,EAChC,UAAU,GAAE,eAAe,GAAG,IAAW,EACzC,QAAQ,GAAE,oBAAyB,GACjC,MAAM,CA4MR;AA2CD,wBAAsB,YAAY,CACjC,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,eAAe,EAAE,EACvB,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,IAAI,EAAE;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,eAAe,CAAA;CAAE,EACnD,MAAM,CAAC,EAAE,WAAW,EACpB,QAAQ,CAAC,EAAE,oBAAoB,GAC7B,OAAO,CAAC,MAAM,CAAC,CA2DjB"}

package/dist/provider-eval.js

@@ -156,7 +156,7 @@ function formatProviderEvaluation(report) {
 import fs2 from "fs";
 
 // src/constants.ts
-var VERSION = "9.6.1";
+var VERSION = "9.6.2";
 var SETTLED_TRIGGER_COOLDOWN_MS = 10 * 60000;
 var COMPACT_SYSTEM_PREFIX = "You are an expert conversation summarizer for a coding agent. " + "Produce structured markdown summaries. " + "Follow output format exactly. " + "Use EXACT names \u2014 never paraphrase code identifiers. " + "Trust deterministic extraction data over intuition.";
 var PROFILES = {
@@ -1815,6 +1815,229 @@ var processToolSupport = new ToolSupportCache;
 var processTokenCalibration = new TokenCalibrationStore;
 var _default = createServices();
 
+// src/domain/telemetry.ts
+function p95(values) {
+  if (!values.length)
+    return 0;
+  const sorted = [...values].sort((a, b) => a - b);
+  return sorted[Math.max(0, Math.ceil(sorted.length * 0.95) - 1)] ?? 0;
+}
+function stats(entries, damage) {
+  const evidence = entries.filter((entry) => entry.status !== "dry-run");
+  const successfulRuns = evidence.filter((entry) => entry.status === "success");
+  const quality = successfulRuns.filter((entry) => typeof entry.verificationScore === "number");
+  const appliedRunIds = new Set(successfulRuns.filter((entry) => typeof entry.runId === "string" && entry.runId.length >= 8).map((entry) => entry.runId));
+  const observedScores = new Map;
+  for (const observation of damage) {
+    if (!observation.runId || !appliedRunIds.has(observation.runId) || typeof observation.damageScore !== "number" || !Number.isFinite(observation.damageScore))
+      continue;
+    observedScores.set(observation.runId, Math.max(observedScores.get(observation.runId) ?? 0, Math.max(0, Math.min(100, observation.damageScore))));
+  }
+  const damaging = [...observedScores.values()].filter((score) => score > 0).length;
+  return {
+    runs: entries.length,
+    appliedRuns: evidence.length,
+    successRate: evidence.length ? successfulRuns.length / evidence.length : 1,
+    avgQuality: quality.length ? quality.reduce((sum, entry) => sum + (entry.verificationScore ?? 0), 0) / quality.length : null,
+    qualityCoverage: evidence.length ? quality.length / evidence.length : 0,
+    p95LatencyMs: p95(evidence.map((entry) => entry.durationMs ?? entry.avgLatency).filter(Number.isFinite)),
+    avgTokens: evidence.length ? evidence.reduce((sum, entry) => sum + entry.totalInput + entry.totalCacheHit + (entry.totalCacheWrite ?? 0) + entry.totalOutput, 0) / evidence.length : 0,
+    fallbackRate: evidence.length ? evidence.filter((entry) => entry.method === "heuristic" || Array.isArray(entry.providerRoutes) && entry.providerRoutes.some((route) => route.successes < route.calls)).length / evidence.length : 0,
+    damageRate: observedScores.size ? damaging / observedScores.size : 0,
+    damageCoverage: successfulRuns.length ? observedScores.size / successfulRuns.length : 0
+  };
+}
+function roundStats(value) {
+  return {
+    ...value,
+    successRate: Math.round(value.successRate * 1000) / 1000,
+    avgQuality: value.avgQuality == null ? null : Math.round(value.avgQuality * 10) / 10,
+    qualityCoverage: Math.round(value.qualityCoverage * 1000) / 1000,
+    p95LatencyMs: Math.round(value.p95LatencyMs),
+    avgTokens: Math.round(value.avgTokens),
+    fallbackRate: Math.round(value.fallbackRate * 1000) / 1000,
+    damageRate: Math.round(value.damageRate * 1000) / 1000,
+    damageCoverage: Math.round(value.damageCoverage * 1000) / 1000
+  };
+}
+function assessCanary(entries, damageEntries, options) {
+  const minCanaryRuns = Math.max(5, options.minCanaryRuns ?? 20);
+  const canaryEntries = entries.filter((entry) => entry.metricsSchemaVersion === 2 && entry.version === options.version && entry.releaseChannel === "canary").slice(-Math.max(100, minCanaryRuns));
+  const baselineEntries = entries.filter((entry) => entry.metricsSchemaVersion === 2 && (entry.releaseChannel ?? "stable") === "stable").slice(-(options.baselineRuns ?? Math.max(50, minCanaryRuns * 2)));
+  const baseline = stats(baselineEntries, damageEntries);
+  const canary = stats(canaryEntries, damageEntries);
+  const triggers = [];
+  const failureBaseline = 1 - baseline.successRate;
+  const failureCanary = 1 - canary.successRate;
+  if (canary.appliedRuns >= 3 && (failureCanary > 0.050001 || failureCanary - failureBaseline >= 0.050001)) {
+    triggers.push({
+      metric: "failure-rate",
+      baseline: failureBaseline,
+      canary: failureCanary,
+      threshold: failureCanary > 0.050001 ? ">5% absolute" : "+5pp regression"
+    });
+  }
+  if (canary.avgQuality != null && (canary.avgQuality < 85 || baseline.avgQuality != null && baseline.avgQuality - canary.avgQuality >= 5)) {
+    triggers.push({
+      metric: "quality",
+      baseline: baseline.avgQuality ?? 0,
+      canary: canary.avgQuality,
+      threshold: canary.avgQuality < 85 ? "<85 absolute" : "-5 points"
+    });
+  }
+  if (baseline.p95LatencyMs >= 1000 && canary.p95LatencyMs >= baseline.p95LatencyMs * 1.5) {
+    triggers.push({ metric: "latency", baseline: baseline.p95LatencyMs, canary: canary.p95LatencyMs, threshold: "+50% p95" });
+  }
+  if (baseline.avgTokens >= 1000 && canary.avgTokens >= baseline.avgTokens * 1.5) {
+    triggers.push({ metric: "tokens", baseline: baseline.avgTokens, canary: canary.avgTokens, threshold: "+50%" });
+  }
+  if (canary.fallbackRate - baseline.fallbackRate >= 0.1) {
+    triggers.push({ metric: "fallback", baseline: baseline.fallbackRate, canary: canary.fallbackRate, threshold: "+10pp" });
+  }
+  if (canary.damageRate - baseline.damageRate >= 0.1) {
+    triggers.push({ metric: "damage", baseline: baseline.damageRate, canary: canary.damageRate, threshold: "+10pp" });
+  }
+  const canarySampleAdequacy = Math.min(1, canary.appliedRuns / minCanaryRuns);
+  const baselineSampleAdequacy = Math.min(1, baseline.appliedRuns / Math.max(20, minCanaryRuns));
+  const dataConfidence = Math.round(100 * (canarySampleAdequacy * 0.25 + baselineSampleAdequacy * 0.15 + canary.qualityCoverage * canarySampleAdequacy * 0.2 + canary.damageCoverage * canarySampleAdequacy * 0.2 + baseline.damageCoverage * baselineSampleAdequacy * 0.2));
+  const reasons = [];
+  let decision = "hold";
+  if (triggers.length && canary.appliedRuns >= 3) {
+    decision = "rollback";
+    reasons.push(...triggers.map((trigger) => trigger.metric + " crossed " + trigger.threshold));
+  } else if (canary.appliedRuns < minCanaryRuns) {
+    reasons.push("need " + (minCanaryRuns - canary.appliedRuns) + " more canary runs with applied outcomes");
+  } else if (baseline.appliedRuns < Math.max(20, minCanaryRuns)) {
+    reasons.push("stable baseline is too small");
+  } else if (canary.qualityCoverage < 0.7) {
+    reasons.push("schema-v2 quality coverage is below 70%");
+  } else if (canary.damageCoverage < 0.7) {
+    reasons.push("correlated canary damage-observation coverage is below 70%");
+  } else if (baseline.damageCoverage < 0.7) {
+    reasons.push("correlated stable damage-observation coverage is below 70%");
+  } else if ((canary.avgQuality ?? 0) < 85) {
+    reasons.push("absolute verifier quality is below 85");
+  } else if (canary.successRate < 0.949999) {
+    reasons.push("absolute success rate is below 95%");
+  } else {
+    decision = "promote";
+    reasons.push("sample, absolute quality, reliability, latency, token, fallback, and damage gates passed");
+  }
+  return {
+    version: options.version,
+    decision,
+    dataConfidence,
+    baseline: roundStats(baseline),
+    canary: roundStats(canary),
+    triggers,
+    reasons
+  };
+}
+function safeMetricLabel(value, fallback) {
+  if (typeof value !== "string" || !/^[\w./:@+-]{1,160}$/.test(value))
+    return fallback;
+  return value;
+}
+var TELEMETRY_FAILURE_KINDS = new Set([
+  "cancelled",
+  "timeout",
+  "rate-limit",
+  "authentication",
+  "budget",
+  "output-limit",
+  "provider",
+  "persistence",
+  "validation",
+  "verification",
+  "yield",
+  "internal"
+]);
+function isTelemetryFailureKind(value) {
+  return typeof value === "string" && TELEMETRY_FAILURE_KINDS.has(value);
+}
+function buildPrivacySafeTelemetry(entries, damageEntries, options) {
+  const groups = new Map;
+  const failures = {};
+  for (const entry of entries) {
+    const version = safeMetricLabel(entry.version, "legacy");
+    const channel = entry.releaseChannel === "canary" ? "canary" : "stable";
+    const provider = safeMetricLabel(entry.provider, "unknown");
+    const rawModel = safeMetricLabel(entry.model, "unknown");
+    const model = rawModel.startsWith(provider + "/") ? rawModel.slice(provider.length + 1) : rawModel;
+    const key = [version, channel, provider, model].join("\x00");
+    const group = groups.get(key) ?? {
+      version,
+      channel,
+      provider,
+      model,
+      runs: 0,
+      successes: 0,
+      quality: 0,
+      qualityRuns: 0,
+      latency: 0,
+      input: 0,
+      output: 0
+    };
+    group.runs++;
+    if (entry.status === "success" || entry.status === "dry-run")
+      group.successes++;
+    if (entry.metricsSchemaVersion === 2 && typeof entry.verificationScore === "number") {
+      group.quality += entry.verificationScore;
+      group.qualityRuns++;
+    }
+    group.latency += entry.avgLatency;
+    group.input += entry.totalInput + entry.totalCacheHit + (entry.totalCacheWrite ?? 0);
+    group.output += entry.totalOutput;
+    groups.set(key, group);
+    if (isTelemetryFailureKind(entry.failureKind)) {
+      failures[entry.failureKind] = (failures[entry.failureKind] ?? 0) + 1;
+    }
+  }
+  const aggregates = [...groups.values()].map((group) => ({
+    version: group.version,
+    channel: group.channel,
+    provider: group.provider,
+    model: group.model,
+    runs: group.runs,
+    successes: group.successes,
+    avgQuality: group.qualityRuns ? Math.round(group.quality / group.qualityRuns * 10) / 10 : null,
+    avgLatencyMs: group.runs ? Math.round(group.latency / group.runs) : 0,
+    inputTokens: group.input,
+    outputTokens: group.output
+  })).sort((a, b) => b.runs - a.runs || a.version.localeCompare(b.version));
+  return {
+    generatedAt: new Date().toISOString(),
+    totalRuns: entries.length,
+    aggregates,
+    failures,
+    canary: assessCanary(entries, damageEntries, options),
+    privacy: "aggregate-only; no session ids, project ids, prompts, summaries, paths, or error text"
+  };
+}
+function formatPrivacySafeTelemetry(report) {
+  const lines = [
+    "# Smart Compact Telemetry",
+    "",
+    "Privacy: " + report.privacy + ".",
+    "",
+    "| Version | Channel | Provider/model | Runs | Success | Quality | Latency | Input | Output |",
+    "|---|---|---|---:|---:|---:|---:|---:|---:|"
+  ];
+  for (const item of report.aggregates) {
+    lines.push("| " + item.version + " | " + item.channel + " | " + item.provider + "/" + item.model + " | " + item.runs + " | " + item.successes + "/" + item.runs + " | " + (item.avgQuality == null ? "n/a" : item.avgQuality.toFixed(1)) + " | " + item.avgLatencyMs + "ms | " + item.inputTokens + " | " + item.outputTokens + " |");
+  }
+  lines.push("", "## Canary: " + report.canary.decision.toUpperCase() + " (data confidence " + report.canary.dataConfidence + "%)", "");
+  const baseline = report.canary.baseline;
+  const canary = report.canary.canary;
+  lines.push("| Gate | Stable baseline | Canary |", "|---|---:|---:|", "| Runs (total/applied) | " + baseline.runs + "/" + baseline.appliedRuns + " | " + canary.runs + "/" + canary.appliedRuns + " |", "| Success | " + Math.round(baseline.successRate * 100) + "% | " + Math.round(canary.successRate * 100) + "% |", "| Verify quality | " + (baseline.avgQuality ?? "n/a") + " | " + (canary.avgQuality ?? "n/a") + " |", "| p95 duration | " + baseline.p95LatencyMs + "ms | " + canary.p95LatencyMs + "ms |", "| Avg tokens | " + baseline.avgTokens + " | " + canary.avgTokens + " |", "| Fallback | " + Math.round(baseline.fallbackRate * 100) + "% | " + Math.round(canary.fallbackRate * 100) + "% |", "| Damage | " + Math.round(baseline.damageRate * 100) + "% | " + Math.round(canary.damageRate * 100) + "% |", "| Damage observed | " + Math.round(baseline.damageCoverage * 100) + "% | " + Math.round(canary.damageCoverage * 100) + "% |", "");
+  for (const reason of report.canary.reasons)
+    lines.push("- " + reason);
+  const failureText = Object.entries(report.failures).map(([kind, count]) => kind + "=" + count).join(", ");
+  lines.push("", "Failures: " + (failureText || "none classified"));
+  return lines.join(`
+`);
+}
+
 // src/utils/cache.ts
 var INTERNAL_PHASES = new Set([
   "explore-retry",

package/dist/provider-scenario-eval.js

@@ -5,7 +5,7 @@ var __require = import.meta.require;
 import { ModelRegistry, ModelRuntime } from "@earendil-works/pi-coding-agent";
 
 // src/constants.ts
-var VERSION = "9.6.1";
+var VERSION = "9.6.2";
 var SETTLED_TRIGGER_COOLDOWN_MS = 10 * 60000;
 var COMPACT_SYSTEM_PREFIX = "You are an expert conversation summarizer for a coding agent. " + "Produce structured markdown summaries. " + "Follow output format exactly. " + "Use EXACT names \u2014 never paraphrase code identifiers. " + "Trust deterministic extraction data over intuition.";
 var PROFILES = {
@@ -1667,6 +1667,229 @@ var processToolSupport = new ToolSupportCache;
 var processTokenCalibration = new TokenCalibrationStore;
 var _default = createServices();
 
+// src/domain/telemetry.ts
+function p95(values) {
+  if (!values.length)
+    return 0;
+  const sorted = [...values].sort((a, b) => a - b);
+  return sorted[Math.max(0, Math.ceil(sorted.length * 0.95) - 1)] ?? 0;
+}
+function stats(entries, damage) {
+  const evidence = entries.filter((entry) => entry.status !== "dry-run");
+  const successfulRuns = evidence.filter((entry) => entry.status === "success");
+  const quality = successfulRuns.filter((entry) => typeof entry.verificationScore === "number");
+  const appliedRunIds = new Set(successfulRuns.filter((entry) => typeof entry.runId === "string" && entry.runId.length >= 8).map((entry) => entry.runId));
+  const observedScores = new Map;
+  for (const observation of damage) {
+    if (!observation.runId || !appliedRunIds.has(observation.runId) || typeof observation.damageScore !== "number" || !Number.isFinite(observation.damageScore))
+      continue;
+    observedScores.set(observation.runId, Math.max(observedScores.get(observation.runId) ?? 0, Math.max(0, Math.min(100, observation.damageScore))));
+  }
+  const damaging = [...observedScores.values()].filter((score) => score > 0).length;
+  return {
+    runs: entries.length,
+    appliedRuns: evidence.length,
+    successRate: evidence.length ? successfulRuns.length / evidence.length : 1,
+    avgQuality: quality.length ? quality.reduce((sum, entry) => sum + (entry.verificationScore ?? 0), 0) / quality.length : null,
+    qualityCoverage: evidence.length ? quality.length / evidence.length : 0,
+    p95LatencyMs: p95(evidence.map((entry) => entry.durationMs ?? entry.avgLatency).filter(Number.isFinite)),
+    avgTokens: evidence.length ? evidence.reduce((sum, entry) => sum + entry.totalInput + entry.totalCacheHit + (entry.totalCacheWrite ?? 0) + entry.totalOutput, 0) / evidence.length : 0,
+    fallbackRate: evidence.length ? evidence.filter((entry) => entry.method === "heuristic" || Array.isArray(entry.providerRoutes) && entry.providerRoutes.some((route) => route.successes < route.calls)).length / evidence.length : 0,
+    damageRate: observedScores.size ? damaging / observedScores.size : 0,
+    damageCoverage: successfulRuns.length ? observedScores.size / successfulRuns.length : 0
+  };
+}
+function roundStats(value) {
+  return {
+    ...value,
+    successRate: Math.round(value.successRate * 1000) / 1000,
+    avgQuality: value.avgQuality == null ? null : Math.round(value.avgQuality * 10) / 10,
+    qualityCoverage: Math.round(value.qualityCoverage * 1000) / 1000,
+    p95LatencyMs: Math.round(value.p95LatencyMs),
+    avgTokens: Math.round(value.avgTokens),
+    fallbackRate: Math.round(value.fallbackRate * 1000) / 1000,
+    damageRate: Math.round(value.damageRate * 1000) / 1000,
+    damageCoverage: Math.round(value.damageCoverage * 1000) / 1000
+  };
+}
+function assessCanary(entries, damageEntries, options) {
+  const minCanaryRuns = Math.max(5, options.minCanaryRuns ?? 20);
+  const canaryEntries = entries.filter((entry) => entry.metricsSchemaVersion === 2 && entry.version === options.version && entry.releaseChannel === "canary").slice(-Math.max(100, minCanaryRuns));
+  const baselineEntries = entries.filter((entry) => entry.metricsSchemaVersion === 2 && (entry.releaseChannel ?? "stable") === "stable").slice(-(options.baselineRuns ?? Math.max(50, minCanaryRuns * 2)));
+  const baseline = stats(baselineEntries, damageEntries);
+  const canary = stats(canaryEntries, damageEntries);
+  const triggers = [];
+  const failureBaseline = 1 - baseline.successRate;
+  const failureCanary = 1 - canary.successRate;
+  if (canary.appliedRuns >= 3 && (failureCanary > 0.050001 || failureCanary - failureBaseline >= 0.050001)) {
+    triggers.push({
+      metric: "failure-rate",
+      baseline: failureBaseline,
+      canary: failureCanary,
+      threshold: failureCanary > 0.050001 ? ">5% absolute" : "+5pp regression"
+    });
+  }
+  if (canary.avgQuality != null && (canary.avgQuality < 85 || baseline.avgQuality != null && baseline.avgQuality - canary.avgQuality >= 5)) {
+    triggers.push({
+      metric: "quality",
+      baseline: baseline.avgQuality ?? 0,
+      canary: canary.avgQuality,
+      threshold: canary.avgQuality < 85 ? "<85 absolute" : "-5 points"
+    });
+  }
+  if (baseline.p95LatencyMs >= 1000 && canary.p95LatencyMs >= baseline.p95LatencyMs * 1.5) {
+    triggers.push({ metric: "latency", baseline: baseline.p95LatencyMs, canary: canary.p95LatencyMs, threshold: "+50% p95" });
+  }
+  if (baseline.avgTokens >= 1000 && canary.avgTokens >= baseline.avgTokens * 1.5) {
+    triggers.push({ metric: "tokens", baseline: baseline.avgTokens, canary: canary.avgTokens, threshold: "+50%" });
+  }
+  if (canary.fallbackRate - baseline.fallbackRate >= 0.1) {
+    triggers.push({ metric: "fallback", baseline: baseline.fallbackRate, canary: canary.fallbackRate, threshold: "+10pp" });
+  }
+  if (canary.damageRate - baseline.damageRate >= 0.1) {
+    triggers.push({ metric: "damage", baseline: baseline.damageRate, canary: canary.damageRate, threshold: "+10pp" });
+  }
+  const canarySampleAdequacy = Math.min(1, canary.appliedRuns / minCanaryRuns);
+  const baselineSampleAdequacy = Math.min(1, baseline.appliedRuns / Math.max(20, minCanaryRuns));
+  const dataConfidence = Math.round(100 * (canarySampleAdequacy * 0.25 + baselineSampleAdequacy * 0.15 + canary.qualityCoverage * canarySampleAdequacy * 0.2 + canary.damageCoverage * canarySampleAdequacy * 0.2 + baseline.damageCoverage * baselineSampleAdequacy * 0.2));
+  const reasons = [];
+  let decision = "hold";
+  if (triggers.length && canary.appliedRuns >= 3) {
+    decision = "rollback";
+    reasons.push(...triggers.map((trigger) => trigger.metric + " crossed " + trigger.threshold));
+  } else if (canary.appliedRuns < minCanaryRuns) {
+    reasons.push("need " + (minCanaryRuns - canary.appliedRuns) + " more canary runs with applied outcomes");
+  } else if (baseline.appliedRuns < Math.max(20, minCanaryRuns)) {
+    reasons.push("stable baseline is too small");
+  } else if (canary.qualityCoverage < 0.7) {
+    reasons.push("schema-v2 quality coverage is below 70%");
+  } else if (canary.damageCoverage < 0.7) {
+    reasons.push("correlated canary damage-observation coverage is below 70%");
+  } else if (baseline.damageCoverage < 0.7) {
+    reasons.push("correlated stable damage-observation coverage is below 70%");
+  } else if ((canary.avgQuality ?? 0) < 85) {
+    reasons.push("absolute verifier quality is below 85");
+  } else if (canary.successRate < 0.949999) {
+    reasons.push("absolute success rate is below 95%");
+  } else {
+    decision = "promote";
+    reasons.push("sample, absolute quality, reliability, latency, token, fallback, and damage gates passed");
+  }
+  return {
+    version: options.version,
+    decision,
+    dataConfidence,
+    baseline: roundStats(baseline),
+    canary: roundStats(canary),
+    triggers,
+    reasons
+  };
+}
+function safeMetricLabel(value, fallback) {
+  if (typeof value !== "string" || !/^[\w./:@+-]{1,160}$/.test(value))
+    return fallback;
+  return value;
+}
+var TELEMETRY_FAILURE_KINDS = new Set([
+  "cancelled",
+  "timeout",
+  "rate-limit",
+  "authentication",
+  "budget",
+  "output-limit",
+  "provider",
+  "persistence",
+  "validation",
+  "verification",
+  "yield",
+  "internal"
+]);
+function isTelemetryFailureKind(value) {
+  return typeof value === "string" && TELEMETRY_FAILURE_KINDS.has(value);
+}
+function buildPrivacySafeTelemetry(entries, damageEntries, options) {
+  const groups = new Map;
+  const failures = {};
+  for (const entry of entries) {
+    const version = safeMetricLabel(entry.version, "legacy");
+    const channel = entry.releaseChannel === "canary" ? "canary" : "stable";
+    const provider = safeMetricLabel(entry.provider, "unknown");
+    const rawModel = safeMetricLabel(entry.model, "unknown");
+    const model = rawModel.startsWith(provider + "/") ? rawModel.slice(provider.length + 1) : rawModel;
+    const key = [version, channel, provider, model].join("\x00");
+    const group = groups.get(key) ?? {
+      version,
+      channel,
+      provider,
+      model,
+      runs: 0,
+      successes: 0,
+      quality: 0,
+      qualityRuns: 0,
+      latency: 0,
+      input: 0,
+      output: 0
+    };
+    group.runs++;
+    if (entry.status === "success" || entry.status === "dry-run")
+      group.successes++;
+    if (entry.metricsSchemaVersion === 2 && typeof entry.verificationScore === "number") {
+      group.quality += entry.verificationScore;
+      group.qualityRuns++;
+    }
+    group.latency += entry.avgLatency;
+    group.input += entry.totalInput + entry.totalCacheHit + (entry.totalCacheWrite ?? 0);
+    group.output += entry.totalOutput;
+    groups.set(key, group);
+    if (isTelemetryFailureKind(entry.failureKind)) {
+      failures[entry.failureKind] = (failures[entry.failureKind] ?? 0) + 1;
+    }
+  }
+  const aggregates = [...groups.values()].map((group) => ({
+    version: group.version,
+    channel: group.channel,
+    provider: group.provider,
+    model: group.model,
+    runs: group.runs,
+    successes: group.successes,
+    avgQuality: group.qualityRuns ? Math.round(group.quality / group.qualityRuns * 10) / 10 : null,
+    avgLatencyMs: group.runs ? Math.round(group.latency / group.runs) : 0,
+    inputTokens: group.input,
+    outputTokens: group.output
+  })).sort((a, b) => b.runs - a.runs || a.version.localeCompare(b.version));
+  return {
+    generatedAt: new Date().toISOString(),
+    totalRuns: entries.length,
+    aggregates,
+    failures,
+    canary: assessCanary(entries, damageEntries, options),
+    privacy: "aggregate-only; no session ids, project ids, prompts, summaries, paths, or error text"
+  };
+}
+function formatPrivacySafeTelemetry(report) {
+  const lines = [
+    "# Smart Compact Telemetry",
+    "",
+    "Privacy: " + report.privacy + ".",
+    "",
+    "| Version | Channel | Provider/model | Runs | Success | Quality | Latency | Input | Output |",
+    "|---|---|---|---:|---:|---:|---:|---:|---:|"
+  ];
+  for (const item of report.aggregates) {
+    lines.push("| " + item.version + " | " + item.channel + " | " + item.provider + "/" + item.model + " | " + item.runs + " | " + item.successes + "/" + item.runs + " | " + (item.avgQuality == null ? "n/a" : item.avgQuality.toFixed(1)) + " | " + item.avgLatencyMs + "ms | " + item.inputTokens + " | " + item.outputTokens + " |");
+  }
+  lines.push("", "## Canary: " + report.canary.decision.toUpperCase() + " (data confidence " + report.canary.dataConfidence + "%)", "");
+  const baseline = report.canary.baseline;
+  const canary = report.canary.canary;
+  lines.push("| Gate | Stable baseline | Canary |", "|---|---:|---:|", "| Runs (total/applied) | " + baseline.runs + "/" + baseline.appliedRuns + " | " + canary.runs + "/" + canary.appliedRuns + " |", "| Success | " + Math.round(baseline.successRate * 100) + "% | " + Math.round(canary.successRate * 100) + "% |", "| Verify quality | " + (baseline.avgQuality ?? "n/a") + " | " + (canary.avgQuality ?? "n/a") + " |", "| p95 duration | " + baseline.p95LatencyMs + "ms | " + canary.p95LatencyMs + "ms |", "| Avg tokens | " + baseline.avgTokens + " | " + canary.avgTokens + " |", "| Fallback | " + Math.round(baseline.fallbackRate * 100) + "% | " + Math.round(canary.fallbackRate * 100) + "% |", "| Damage | " + Math.round(baseline.damageRate * 100) + "% | " + Math.round(canary.damageRate * 100) + "% |", "| Damage observed | " + Math.round(baseline.damageCoverage * 100) + "% | " + Math.round(canary.damageCoverage * 100) + "% |", "");
+  for (const reason of report.canary.reasons)
+    lines.push("- " + reason);
+  const failureText = Object.entries(report.failures).map(([kind, count]) => kind + "=" + count).join(", ");
+  lines.push("", "Failures: " + (failureText || "none classified"));
+  return lines.join(`
+`);
+}
+
 // src/utils/cache.ts
 var INTERNAL_PHASES = new Set([
   "explore-retry",
@@ -1796,8 +2019,9 @@ function hasListedPath(listed, file, display, normalizedOwners) {
   }
   return false;
 }
-function outcomeClaims(summary) {
-  return Array.from(new Set(summary.split(/\r?\n/).map((line) => line.replace(/^\s*(?:[-*+]|\d+[.)])\s+/, "").replace(/^\[[ x]\]\s+/i, "").trim()).filter((line) => line.length > 0 && !line.startsWith("#")).filter((line) => HIGH_RISK_OUTCOME_RE.test(line)).filter((line) => /\bno\s+(?:errors?|failures?)\b/i.test(line) || !NEGATED_OUTCOME_RE.test(line)))).slice(0, 12);
+function outcomeClaims(summary, pathEvidence) {
+  const pathLines = new Set(Array.from(pathEvidence, ([path2, display]) => [path2, display, "`" + path2 + "`"]).flat());
+  return Array.from(new Set(summary.split(/\r?\n/).map((line) => line.replace(/^\s*(?:[-*+]|\d+[.)])\s+/, "").replace(/^\[[ x]\]\s+/i, "").trim()).filter((line) => line.length > 0 && !line.startsWith("#") && !pathLines.has(line)).filter((line) => HIGH_RISK_OUTCOME_RE.test(line)).filter((line) => /\bno\s+(?:errors?|failures?)\b/i.test(line) || !NEGATED_OUTCOME_RE.test(line)))).slice(0, 12);
 }
 function classifyOutcomeClaim(claim) {
   const lower = claim.toLowerCase();
@@ -2031,7 +2255,7 @@ function stemToken(token) {
   return lower;
 }
 function semanticTokens(text) {
-  return (text.normalize("NFKC").match(/[\p{L}\p{N}_-]+/gu) ?? []).map(stemToken).filter((token) => token.length > 2);
+  return (text.normalize("NFKC").match(/[\p{L}\p{N}_-]+/gu) ?? []).map(stemToken).filter((token) => token.length > 2 || NEGATION_MARKERS.has(token));
 }
 var semanticShapeCache = new Map;
 var semanticFragmentCache = new Map;
@@ -2051,7 +2275,7 @@ function hasEffectiveTargetNegation(tokens, anchor) {
     if (token !== anchor)
       return false;
     const nearbyStart = Math.max(0, anchorIndex - 2);
-    const nearbyNegations = tokens.slice(nearbyStart, anchorIndex + 3).map((near, offset) => NEGATION_MARKERS.has(near) ? nearbyStart + offset : -1).filter((index) => index >= 0);
+    const nearbyNegations = tokens.slice(nearbyStart, anchorIndex + 3).map((near, offset) => NEGATION_MARKERS.has(near) && !(near === "without" && nearbyStart + offset > anchorIndex) ? nearbyStart + offset : -1).filter((index) => index >= 0);
     const governingStart = Math.max(0, anchorIndex - 3);
     const preceding = tokens.slice(governingStart, anchorIndex);
     const nearbyGuards = preceding.map((near, offset) => POLARITY_INVERTING_GUARDS.has(near) ? governingStart + offset : -1).filter((index) => index >= 0);
@@ -2100,12 +2324,13 @@ function hasSemanticEvidence(source, target) {
   });
 }
 function hasSemanticContradiction(source, target) {
+  const sourceFragments = new Set(semanticFragments(source).map((tokens) => tokens.join(" ")));
   const { sourceTokens, concepts, anchor, negative, conditional } = semanticShape(source);
   if (!anchor)
     return false;
   const required = Math.min(concepts.length, Math.max(1, Math.ceil(concepts.length * 0.6)));
   return semanticFragments(target).some((tokens) => {
-    if (!tokens.includes(anchor))
+    if (!tokens.includes(anchor) || sourceFragments.has(tokens.join(" ")))
       return false;
     const overlap = concepts.filter((concept) => tokens.includes(concept)).length;
     if (overlap < required)
@@ -2361,7 +2586,7 @@ function verifyProgressConsistency(parsed, extraction, collected, paths, accumul
     }
   }
 }
-function verifyOpenLoopsAndClaims(summary, parsed, extraction, continuity, evidence, collected, accumulator) {
+function verifyOpenLoopsAndClaims(summary, parsed, paths, extraction, continuity, evidence, collected, accumulator) {
   const unresolvedCount = collected.unresolved.length + (continuity?.openLoops.filter((loop) => loop.status !== "resolved").length ?? 0);
   if (unresolvedCount >= 1 && !findSection(parsed, "open-loops") && !summary.toLowerCase().replace(/\\/g, "/").includes("unresolved")) {
     addGap(accumulator, { kind: "missing-open-loops", unresolvedCount }, 5);
@@ -2369,7 +2594,7 @@ function verifyOpenLoopsAndClaims(summary, parsed, extraction, continuity, evide
   if (!evidence.sourceMessages)
     return;
   const tools = successfulToolEvidence(evidence.sourceMessages);
-  for (const claim of outcomeClaims(summary)) {
+  for (const claim of outcomeClaims(summary, paths.rendered)) {
     if (!successfulToolSupportsClaim(claim, tools, extraction)) {
       addGap(accumulator, { kind: "unsupported-claim", claim }, 20);
     }
@@ -2385,7 +2610,7 @@ function verifySummary(summary, extraction, continuity = null, evidence = {}) {
   verifySemanticCoverage(parsed, collected, accumulator);
   verifyFileReferences(summary, extraction, continuity, evidence, collected, paths, accumulator);
   verifyProgressConsistency(parsed, extraction, collected, paths, accumulator);
-  verifyOpenLoopsAndClaims(summary, parsed, extraction, continuity, evidence, collected, accumulator);
+  verifyOpenLoopsAndClaims(summary, parsed, paths, extraction, continuity, evidence, collected, accumulator);
   const score = Math.max(0, accumulator.score);
   return {
     ok: accumulator.gaps.length === 0 && score >= 85,

package/dist/telemetry-report.js

@@ -2,7 +2,7 @@
 var __require = import.meta.require;
 
 // src/constants.ts
-var VERSION = "9.6.1";
+var VERSION = "9.6.2";
 var SETTLED_TRIGGER_COOLDOWN_MS = 10 * 60000;
 var COMPACT_SYSTEM_PREFIX = "You are an expert conversation summarizer for a coding agent. " + "Produce structured markdown summaries. " + "Follow output format exactly. " + "Use EXACT names \u2014 never paraphrase code identifiers. " + "Trust deterministic extraction data over intuition.";
 var PROFILES = {

package/dist/types.d.ts

@@ -84,6 +84,8 @@ export interface LLMCallMetric {
     cacheWriteTokens: number;
     latencyMs: number;
     success: boolean;
+    /** Content-free provider failure category; never stores response/error text. */
+    failureKind?: TelemetryFailureKind;
     /** True when provider usage was absent or partial and local estimates were used. */
     usageEstimated?: boolean;
 }
@@ -94,6 +96,7 @@ export interface ProviderRouteMetric {
     model: string;
     calls: number;
     successes: number;
+    failures?: Partial<Record<TelemetryFailureKind, number>>;
     avgLatencyMs: number;
     inputTokens: number;
     outputTokens: number;