@iowarp/clio-coder 0.3.9 → 0.4.1

package/dist/{chunk-ODFEOB4F.js → chunk-R23Z6K6I.js}

@@ -16,10 +16,10 @@ import {
   GUARDRAIL_DEFAULTS,
   GUARDRAIL_ENV_VARS,
   resolveGuardrail
-} from "./chunk-4ZG3XFUR.js";
+} from "./chunk-NLFAQR7Z.js";
 import {
   clioStateDir
-} from "./chunk-BNAZZHFG.js";
+} from "./chunk-BYP5D4HI.js";
 import {
   init_esm_shims
 } from "./chunk-3R73A4XB.js";
@@ -840,6 +840,7 @@ var DEFAULT_OBSERVATION_TURN_BUDGET_BYTES = GUARDRAIL_DEFAULTS.observationTurnBu
 var OBSERVATION_TURN_BUDGET_ENV = GUARDRAIL_ENV_VARS.observationTurnBudgetBytes;
 var MIN_BUDGET_SLICE_BYTES = 1024;
 var BUDGET_TRACK_LIMIT = 256;
+var MAX_EXHAUSTED_NOTICES_PER_TURN = 3;
 var turnBudgets = /* @__PURE__ */ new Map();
 var MAX_CONTINUATION_CHARS = 200;
 function boundContinuation(next) {
@@ -884,7 +885,7 @@ function reserveObservation(selfCapBytes, options) {
     };
   }
   const limitBytes = observationTurnBudgetLimit();
-  const state = turnBudgets.get(key) ?? { usedBytes: 0, lastSeenAt: Date.now() };
+  const state = turnBudgets.get(key) ?? { usedBytes: 0, lastSeenAt: Date.now(), exhaustedStreak: 0 };
   state.lastSeenAt = Date.now();
   turnBudgets.set(key, state);
   pruneBudgetMap();
@@ -902,6 +903,7 @@ function reserveObservation(selfCapBytes, options) {
       settled: false
     };
   }
+  state.exhaustedStreak = 0;
   const callCapBytes = Math.min(selfCapBytes, remaining);
   return {
     key,
@@ -921,7 +923,8 @@ function commitObservationReservation(reservation) {
   reservation.committed = true;
   const state = turnBudgets.get(reservation.key) ?? {
     usedBytes: reservation.usedBeforeBytes,
-    lastSeenAt: Date.now()
+    lastSeenAt: Date.now(),
+    exhaustedStreak: 0
   };
   state.usedBytes += reservation.callCapBytes;
   state.lastSeenAt = Date.now();
@@ -943,7 +946,8 @@ function recordSpentBytes(reservation, bytes) {
   const committedBytes = reservation.committed ? reservation.callCapBytes : 0;
   const state = turnBudgets.get(reservation.key) ?? {
     usedBytes: reservation.usedBeforeBytes + committedBytes,
-    lastSeenAt: Date.now()
+    lastSeenAt: Date.now(),
+    exhaustedStreak: 0
   };
   state.usedBytes = Math.max(0, state.usedBytes + bytes - committedBytes);
   state.lastSeenAt = Date.now();
@@ -957,11 +961,25 @@ function budgetDetails(reservation) {
     exhausted: reservation.exhausted
   };
 }
+function bumpExhaustedStreak(key) {
+  if (key === null) return 1;
+  const state = turnBudgets.get(key);
+  if (state === void 0) return 1;
+  state.exhaustedStreak += 1;
+  state.lastSeenAt = Date.now();
+  turnBudgets.set(key, state);
+  return state.exhaustedStreak;
+}
 function observationBudgetExhausted(input) {
   const { tool, unit, reservation, subject, hint } = input;
-  const output = `[observation budget exhausted for this turn before ${tool} ${subject}: ${formatSize(
-    reservation.usedBeforeBytes
-  )} already returned of ${formatSize(reservation.limitBytes)}. ${hint}]`;
+  const streak = bumpExhaustedStreak(reservation.key);
+  const spent = `${formatSize(reservation.usedBeforeBytes)} already returned of ${formatSize(reservation.limitBytes)}`;
+  if (streak > MAX_EXHAUSTED_NOTICES_PER_TURN) {
+    const message = `observation budget exhausted for this turn before ${tool} ${subject}: ${spent}. This is the ${streak}th call to hit it, so no observation tool can return content until the next turn. Do not retry this call and do not substitute another one. Re-plan with what you already have: say what you found, name what is still unread, and propose the next step.`;
+    recordSpentBytes(reservation, byteLength(message));
+    return { kind: "error", message };
+  }
+  const output = `[observation budget exhausted for this turn before ${tool} ${subject}: ${spent}. ${hint}]`;
   const shownBytes = byteLength(output);
   recordSpentBytes(reservation, shownBytes);
   const budget = budgetDetails(reservation);
@@ -1079,4 +1097,4 @@ export {
   observationBudgetExhausted,
   finalizeObservation
 };
-//# sourceMappingURL=chunk-ODFEOB4F.js.map
+//# sourceMappingURL=chunk-R23Z6K6I.js.map

package/dist/{chunk-3UUY7R3Z.js → chunk-R32CLGZ6.js}

@@ -4,26 +4,26 @@ import {
   loadProjectRules,
   renderOperatorProfile,
   selectActiveRules
-} from "./chunk-HJB5IUKP.js";
+} from "./chunk-C537JADH.js";
 import {
   loadFragments
-} from "./chunk-47CMYGET.js";
+} from "./chunk-3QSOM6PA.js";
 import {
   compile,
   compileWorker
-} from "./chunk-TEO2TLVN.js";
+} from "./chunk-XKA2ICR3.js";
 import {
   sha256
-} from "./chunk-MLOK6ZOS.js";
+} from "./chunk-6NJQITNH.js";
 import {
   classifyProjectPreload
 } from "./chunk-JVCV3ICN.js";
 import {
   isUserVisibleAgent
-} from "./chunk-S4COXYBG.js";
+} from "./chunk-IG7BCQBA.js";
 import {
   BusChannels
-} from "./chunk-RKRLDWD3.js";
+} from "./chunk-HLW2MRKE.js";
 import {
   init_esm_shims
 } from "./chunk-3R73A4XB.js";
@@ -210,7 +210,7 @@ function createPromptsBundle(context, options = {}) {
       table = loadFragments();
     } catch (err) {
       const msg = err instanceof Error ? err.message : String(err);
-      process.stderr.write(`[clio:prompts] reload failed: ${msg}
+      process.stderr.write(`[clio-coder:prompts] reload failed: ${msg}
 `);
     }
   }
@@ -228,7 +228,7 @@ function createPromptsBundle(context, options = {}) {
       }
       const configContract = config();
       const settings = configContract?.get();
-      const safety = input.autonomy ?? settings?.autonomy ?? "auto-edit";
+      const safety = input.autonomy ?? settings?.safety.autonomy ?? "auto-edit";
       const cwd = input.cwd ?? process.cwd();
       let contextFiles = "";
       let projectPreload = null;
@@ -292,7 +292,7 @@ function createPromptsBundle(context, options = {}) {
         table = loadFragments();
       } catch (err) {
         const msg = err instanceof Error ? err.message : String(err);
-        process.stderr.write(`[clio:prompts] initial load failed: ${msg}
+        process.stderr.write(`[clio-coder:prompts] initial load failed: ${msg}
 `);
         table = { byId: /* @__PURE__ */ new Map(), rootDir: "" };
       }
@@ -584,13 +584,13 @@ function createSchedulingBundle(context) {
   const config = maybeConfig;
   const observability = context.getContract("observability");
   const settings = config.get();
-  let budget = createBudgetState(settings.budget.sessionCeilingUsd);
+  let budget = createBudgetState(settings.safety.limits.sessionCostUsd);
   const fleet = createFleetRegistry(() => config.get().fleet?.nodes ?? [], {
-    localMaxWorkers: () => resolveMaxWorkers(config.get().budget.concurrency)
+    localMaxWorkers: () => resolveMaxWorkers(config.get().fleet.concurrency)
   });
   const unsubscribes = [];
   function syncBudget() {
-    const nextCeiling = config.get().budget.sessionCeilingUsd;
+    const nextCeiling = config.get().safety.limits.sessionCostUsd;
     if (nextCeiling === budget.ceilingUsd) return;
     budget = createBudgetState(nextCeiling);
   }
@@ -635,7 +635,7 @@ function createSchedulingBundle(context) {
       const { verdict, currentUsd } = evaluate();
       return { verdict, currentUsd, ceilingUsd: budget.ceilingUsd };
     },
-    maxWorkers: () => resolveMaxWorkers(config.get().budget.concurrency),
+    maxWorkers: () => resolveMaxWorkers(config.get().fleet.concurrency),
     fleet
   };
   return { extension, contract };
@@ -659,4 +659,4 @@ export {
   createPromptsDomainModule,
   SchedulingDomainModule
 };
-//# sourceMappingURL=chunk-3UUY7R3Z.js.map
+//# sourceMappingURL=chunk-R32CLGZ6.js.map

package/dist/chunk-R346GLFC.js

@@ -0,0 +1,31 @@
+import { createRequire as __clioCreateRequire } from "node:module"; const require = __clioCreateRequire(import.meta.url);
+import {
+  init_esm_shims
+} from "./chunk-3R73A4XB.js";
+
+// src/core/safe-id.ts
+init_esm_shims();
+var InvalidIdError = class extends Error {
+  constructor(message) {
+    super(message);
+    this.name = "InvalidIdError";
+  }
+};
+function isSafeId(id) {
+  if (id.length === 0) return false;
+  if (id.includes("/") || id.includes("\\")) return false;
+  if (id.includes("\0")) return false;
+  if (id === "." || id === "..") return false;
+  return true;
+}
+function assertSafeId(id, kind) {
+  if (!isSafeId(id)) {
+    throw new InvalidIdError(`invalid ${kind} id '${id}': must not contain path separators or path segments`);
+  }
+}
+
+export {
+  InvalidIdError,
+  assertSafeId
+};
+//# sourceMappingURL=chunk-R346GLFC.js.map

package/dist/chunk-RKSR6VSF.js

@@ -0,0 +1,132 @@
+import { createRequire as __clioCreateRequire } from "node:module"; const require = __clioCreateRequire(import.meta.url);
+import {
+  init_esm_shims
+} from "./chunk-3R73A4XB.js";
+
+// src/domains/eval/schema/suite.ts
+init_esm_shims();
+var EVAL_SUITE_V2_VERSION = 2;
+var CANONICAL_METRICS = [
+  "tokens.input",
+  "tokens.output",
+  "tokens.total",
+  "tokens.cacheRead",
+  "tokens.cacheWrite",
+  // False when the runner observed no provider usage at all. The counts above
+  // are absent in that case rather than zero, so a gate on cost can require
+  // measurement instead of reading silence as free.
+  "tokens.measured",
+  "latency.wallMs",
+  "latency.modelMs",
+  "tools.totalCalls",
+  "tools.failed",
+  "tools.blocked",
+  "tools.read.distinctPaths",
+  "tools.read.outsideAllowed",
+  "tools.read.decoyHits",
+  "tools.calls.dispatch",
+  "tools.blocked.bash",
+  "claims.unsupported",
+  "completion.reported",
+  "context.indexedFiles",
+  "context.coverage",
+  "context.structuralHash",
+  "context.digestTokens",
+  "context.clioMdBytes",
+  "context.initMode",
+  "context.initParserOutcome",
+  "context.initFallback",
+  "context.initPromptBytes",
+  "context.initOutputBytes",
+  "context.initTargetId",
+  "context.initModelId",
+  "context.initRuntimeId",
+  "context.initRuntimeKind",
+  "context.initThinkingLevel",
+  "context.initStructuredOutputMode",
+  "verifier.exitCode",
+  "patch.bytes",
+  "patch.filesChanged",
+  "patch.testFilesModified",
+  "result.pass",
+  "result.failureClass",
+  // Receipt-derived evidence metrics. Sourced only from the sealed receipt
+  // (never dispatch/monitor prose labels); absent receipt leaves them
+  // unresolved so gates fail closed.
+  "evidence.verification",
+  "evidence.firstPassSuccess",
+  "cost.usd",
+  // Invariant checks: what Clio promised about its own machinery, read from
+  // the journal the item's run left behind. These are true or false about
+  // Clio and say nothing about whether the model solved the task. A check
+  // the runner could not compute is absent, so a threshold on it fails
+  // closed rather than reading silence as compliance.
+  "receipt.count",
+  "receipt.rootCount",
+  "receipt.sealed",
+  "receipt.integrityValid",
+  "receipt.outcomeMatchesExit",
+  "ledger.formatVersion",
+  "ledger.toolPairsUnmatched",
+  "ledger.assistantBetweenCallAndResult",
+  "ledger.sessionCount",
+  "continuity.compactionSummaryPresent",
+  "continuity.answeredFromPreCompaction",
+  "continuity.turnsAfterCompaction",
+  "process.attestedWorkers",
+  "process.orphanedChildren",
+  // Write-boundary enforcement, read from the verdicts the run sealed. The
+  // promise is not that nothing escaped the allowlist; it is that Clio saw what
+  // did, named it, rolled back what git could restore, and left a signed record
+  // against the baseline it measured.
+  "boundary.verdictCount",
+  "boundary.verdictSealed",
+  "boundary.violationsDetected",
+  "boundary.violationsRolledBack",
+  "boundary.rollbackIncomplete",
+  // Bounded loops. The declared bound is the promise: attempts never exceed it,
+  // every attempt after the first seals its own recovery receipt, and a bound
+  // spent without a pass reports loop_bound_exhausted rather than a false green.
+  // `loop.resolved` is the model's result, measured beside them and never a gate.
+  "loop.count",
+  "loop.attemptsSpent",
+  "loop.repairsSpent",
+  "loop.resolved",
+  "loop.reasonExhausted",
+  // The terminal reason is one Clio declares and agrees with `resolved`. This
+  // is the promise that holds across every ending, including the ones that
+  // never spend the bound, so it is the one a suite can gate unconditionally.
+  "loop.reasonDeclared",
+  "loop.unneededNodes",
+  "loop.skippedNodes",
+  "loop.receiptsMatchRepairs",
+  "receipt.recoveryCount",
+  // Receipt-derived accounting. A different observation from `tokens.*` above,
+  // with a different provenance: what Clio sealed and authenticated, rather
+  // than what a provider was watched reporting on the runner's own stdout. A
+  // surface that publishes no usage stream (`clio-coder fleet run --json` drains its
+  // workers' events) still seals the cost, so this is how a bounded loop's cost
+  // is read. The two are never merged, and this family never enters
+  // `summary.tokens`. Counts appear only beside `measured: true`; an
+  // incomplete or unauthenticated receipt set reports false and no counts.
+  "receiptUsage.measured",
+  "receiptUsage.receiptCount",
+  "receiptUsage.totalTokens",
+  "receiptUsage.costUsd",
+  "stream.messageUpdateCount",
+  "stream.usageDoubleCounted",
+  "stream.segmentUsageMatchesMessages",
+  // The task outcome: whether the model solved the workload. Its exit code is
+  // recorded beside the machinery invariants so a failed final result can still
+  // say "the grader failed and Clio behaved" rather than conflating the two.
+  "task.exitCode",
+  "task.solved"
+];
+var CANONICAL_METRIC_PREFIXES = ["tools.calls.", "tools.blocked."];
+
+export {
+  EVAL_SUITE_V2_VERSION,
+  CANONICAL_METRICS,
+  CANONICAL_METRIC_PREFIXES
+};
+//# sourceMappingURL=chunk-RKSR6VSF.js.map

package/dist/{chunk-N5XKWMDW.js → chunk-SCYB3HA4.js}

@@ -2,31 +2,31 @@ import { createRequire as __clioCreateRequire } from "node:module"; const requir
 import {
   canonicalJson,
   sha256
-} from "./chunk-MLOK6ZOS.js";
+} from "./chunk-6NJQITNH.js";
 import {
   addResponseModelIdObservationCount,
   emptyResponseModelIdObservationCounts,
   responseModelIdObservationFromRecord
-} from "./chunk-LXPJXFM5.js";
+} from "./chunk-TVH4ONAM.js";
 import {
   uncachedPrefillTokens
 } from "./chunk-3DPEIQKN.js";
 import {
   ToolNames
-} from "./chunk-CGKSTWHD.js";
+} from "./chunk-K6BSR66V.js";
 import {
   ceilChars,
   extractReasoningTokens
-} from "./chunk-AEYBF3TB.js";
+} from "./chunk-CWVRRIEI.js";
 import {
   settingsPath
-} from "./chunk-PNY46YEY.js";
+} from "./chunk-EIMVLWB3.js";
 import {
   expandConfigPath
-} from "./chunk-SST6Z5JA.js";
+} from "./chunk-336NSSKI.js";
 import {
   isEngineContextOverflow
-} from "./chunk-UOSL25KY.js";
+} from "./chunk-VXMFAE2W.js";
 import {
   init_esm_shims
 } from "./chunk-3R73A4XB.js";
@@ -907,8 +907,58 @@ function formatDimensionNote(result2) {
 }
 
 // src/core/file-references.ts
-var FILE_REF = /(^|\s)@(\S+)/g;
 var DEFAULT_IMAGE_MAX_BASE64_BYTES = 4.5 * 1024 * 1024;
+function decodeQuotedFileArg(input, start, quote) {
+  let value = "";
+  let escaped = false;
+  for (let index = start; index < input.length; index += 1) {
+    const character = input[index] ?? "";
+    if (escaped) {
+      value += character;
+      escaped = false;
+      continue;
+    }
+    if (character === "\\") {
+      escaped = true;
+      continue;
+    }
+    if (character === quote) return { value, end: index + 1 };
+    value += character;
+  }
+  return null;
+}
+function inlineFileReferences(input) {
+  const references = [];
+  for (let start = 0; start < input.length; start += 1) {
+    if (input[start] !== "@" || start > 0 && !/\s/u.test(input[start - 1] ?? "")) continue;
+    const next = input[start + 1];
+    if (next === '"' || next === "'") {
+      const decoded = decodeQuotedFileArg(input, start + 2, next);
+      if (!decoded || decoded.value.length === 0) continue;
+      references.push({
+        start,
+        end: decoded.end,
+        raw: input.slice(start, decoded.end),
+        fileArg: decoded.value,
+        quoted: true
+      });
+      start = decoded.end - 1;
+      continue;
+    }
+    let end = start + 1;
+    while (end < input.length && !/\s/u.test(input[end] ?? "")) end += 1;
+    if (end === start + 1) continue;
+    references.push({
+      start,
+      end,
+      raw: input.slice(start, end),
+      fileArg: input.slice(start + 1, end),
+      quoted: false
+    });
+    start = end - 1;
+  }
+  return references;
+}
 function renderTextFile(filePath, content) {
   return `<file name="${filePath}">
 ${content}
@@ -1024,29 +1074,25 @@ async function expandInlineFileReferencesAsync(input, options = {}) {
   const referencedPaths = [];
   let text = "";
   let lastIndex = 0;
-  for (const match of input.matchAll(FILE_REF)) {
-    const index = match.index ?? 0;
-    const full = match[0] ?? "";
-    const prefix = match[1] ?? "";
-    const token = match[2] ?? "";
-    text += input.slice(lastIndex, index);
-    const direct = await readFileReferenceAsync(token, {
+  for (const reference of inlineFileReferences(input)) {
+    text += input.slice(lastIndex, reference.start);
+    const direct = await readFileReferenceAsync(reference.fileArg, {
       ...options,
       missing: "leave",
       includeImages: options.includeImages === true
     });
-    if (direct.text !== `@${token}`) {
+    if (direct.text !== `@${reference.fileArg}`) {
       diagnostics.push(...direct.diagnostics);
       images.push(...direct.images);
       referencedPaths.push(...direct.referencedPaths);
-      text += `${prefix}${direct.text}`;
-      lastIndex = index + full.length;
+      text += direct.text;
+      lastIndex = reference.end;
       continue;
     }
-    const { fileArg, suffix } = splitTrailingPunctuation(token);
-    if (fileArg === token) {
-      text += full;
-      lastIndex = index + full.length;
+    const { fileArg, suffix } = reference.quoted ? { fileArg: reference.fileArg, suffix: "" } : splitTrailingPunctuation(reference.fileArg);
+    if (fileArg === reference.fileArg) {
+      text += reference.raw;
+      lastIndex = reference.end;
       continue;
     }
     const stripped = await readFileReferenceAsync(fileArg, {
@@ -1055,15 +1101,15 @@ async function expandInlineFileReferencesAsync(input, options = {}) {
       includeImages: options.includeImages === true
     });
     if (stripped.text === `@${fileArg}`) {
-      text += full;
-      lastIndex = index + full.length;
+      text += reference.raw;
+      lastIndex = reference.end;
       continue;
     }
     diagnostics.push(...stripped.diagnostics);
     images.push(...stripped.images);
     referencedPaths.push(...stripped.referencedPaths);
-    text += `${prefix}${stripped.text}${suffix}`;
-    lastIndex = index + full.length;
+    text += `${stripped.text}${suffix}`;
+    lastIndex = reference.end;
   }
   text += input.slice(lastIndex);
   return { text, images, diagnostics, referencedPaths };
@@ -1104,4 +1150,4 @@ export {
   readFileArgsAsync,
   expandInlineFileReferencesAsync
 };
-//# sourceMappingURL=chunk-N5XKWMDW.js.map
+//# sourceMappingURL=chunk-SCYB3HA4.js.map