@claudexor/orchestrator 3.1.0 → 3.1.1

package/dist/orchestrator.js

@@ -12,11 +12,13 @@ import { HarnessRunSpec, ModeKind as ModeKindSchema, SCHEMA_VERSION, TRUST_FULL_
 import { globalConfigDir, loadConfig, trustConfigPath } from "@claudexor/config";
 import { AnswerAssembly, CLAUDEXOR_ARTIFACT_DIR, CLAUDEXOR_BROWSER_ARTIFACT_SUBDIR, HarnessUnavailableError, summarizeDiffPaths as diffStats, withInactivityWatchdog, } from "@claudexor/core";
 import { assertRouteModelsAllowed } from "./modelGovernance.js";
+import { governRouteEffort } from "./effortGovernance.js";
 import { RequestRequirementsResolver } from "./requestRequirements.js";
+import { buildRevisePrompt } from "./revisePrompt.js";
 import { cancelledResult, failTerminally, guardAnnouncedRun, writeFailure, } from "./runTerminals.js";
 import { budgetFailureRecord, classifyBudgetFailure } from "./budgetFailure.js";
 import { assertOutputSchemaCompiles, finalizeStructuredOutput } from "./structuredOutput.js";
-import { transientRetryDelayMs, promptWithEngineConstraints, sleep, redactHarnessEvent, harnessEventPayload, formatFindings, renderSummary, observeBudgetSignals, rotateOnStall, recordCleanAttemptMetrics, envInheritance, transientRetryPolicy, reviewerTimeoutMs, harnessInactivityTimeoutMs, observeAuthSwitch, emitPrimaryDivergence, emitPoolDegraded, deliveryRefusalFailure, writeRaceDeliveryDecision, } from "./runSupport.js";
+import { transientRetryDelayMs, promptWithEngineConstraints, sleep, redactHarnessEvent, harnessEventPayload, safeErrorMessage, renderSummary, observeBudgetSignals, rotateOnStall, recordCleanAttemptMetrics, envInheritance, transientRetryPolicy, reviewerTimeoutMs, harnessInactivityTimeoutMs, observeAuthSwitch, emitPrimaryDivergence, emitPoolDegraded, deliveryRefusalFailure, writeRaceDeliveryDecision, } from "./runSupport.js";
 import { candidateStatusInRouteContext, resolveReadOnlyRouteContext, } from "./routeContext.js";
 import { resolveAutoReviewerPanel, resolveExplicitReviewerPanel } from "./reviewerPanel.js";
 import { buildContinuation, } from "./continuity.js";
@@ -25,7 +27,7 @@ import { runDiffReview } from "./diffReview.js";
 import { rawScoutBundle, resolveDeepScanSynthesis, } from "./deepScanReducer.js";
 import { classifyAdapterThrow, createAttemptTelemetry, observeAttemptTelemetry, setAttemptOutcome, telemetrySummary, toolWarnings, unrecoveredToolErrors, webUnsatisfied, } from "./attemptTelemetry.js";
 import { dominantHarnessFailureCategory, harnessFailureNextActions } from "./harnessFailure.js";
-import { finalizeAttempt, readOnlyNoSuccessTerminal, resolveWorkReportEnvelope, unwrapWorkReportEnvelope, } from "./attemptFinalize.js";
+import { finalizeAttempt, readOnlyNoSuccessTerminal, resolveWorkReportEnvelope, unrecoveredToolErrorFailure, unwrapWorkReportEnvelope, webEvidenceFailure, } from "./attemptFinalize.js";
 import { buildContinuationPacket, decideContinuation, synthesizeContinuationRequest, } from "./continuation.js";
 import { interactionChannelFor } from "./interaction.js";
 import { gateSpecsFromContract, renderTestsEvidence, resolveContractGates, } from "./contract-gates.js";
@@ -40,6 +42,7 @@ import { arbitrate } from "@claudexor/arbitration";
 import { buildSynthesisPlan, decideSynthesis } from "@claudexor/synthesis";
 import { attemptCostEvidence, attemptUsageCostSettlement, BudgetLedger, isBudgetTerminal, explainRanking, loadHarnessMetrics, promptFingerprint, unknownCostSettlement, rankHarnesses, reviewUsageCostSettlement, } from "@claudexor/budget";
 import { readTextSafe, appendLine, assertNoInlineSecretValues, containsSecretLikeToken, DELEGATION_ENV, hashJson, newId, noProjectRepoRoot, nowIso, redactSecrets, safeInvoke, sha256, userConfigDir, writeText, } from "@claudexor/util";
+import { assertWriteIsolation } from "./write-isolation.js";
 /** A routed candidate adapter plus its manifest capabilities and user settings. */
 /** The two access profiles that map to codex `danger-full-access` / an
  * unsandboxed lane — the only ones under which a full-access-requiring MCP
@@ -134,13 +137,15 @@ export class Orchestrator {
             throw new Error(`unknown mode: ${String(resolved.mode)}`);
         }
         const mode = parsedMode.data;
-        // denyPaths is enforced by the post-diff policy gate BEFORE delivery, which
-        // only exists on envelope/isolated runs — an in-place run mutates the live
-        // tree directly, so the gate could not contain a violation. Refuse loudly
-        // rather than accept a knob the engine cannot honor (INV-023).
-        if ((resolved.denyPaths?.length ?? 0) > 0 && resolved.inPlace === true) {
-            throw new Error("denyPaths requires an isolated/envelope run: the post-diff policy gate blocks a violating patch before delivery, which an in-place run cannot guarantee; drop --deny-path or run isolated");
-        }
+        const projectProtectedPaths = mode === "agent" ? this.projectConfig(resolved.repoRoot).constraints.protected_paths : [];
+        assertWriteIsolation({
+            mode,
+            protectedPaths: projectProtectedPaths,
+            denyPaths: resolved.denyPaths,
+            inPlace: resolved.inPlace,
+            repoRoot: resolved.repoRoot,
+            executionRoot: this.execRootOf(resolved),
+        });
         // outputSchema constrains the run's final ANSWER. It is honored exactly
         // where a final answer is delivered (agent race incl. synthesis, and ask);
         // every other strategy refuses loudly rather than carrying a contract the
@@ -204,7 +209,7 @@ export class Orchestrator {
         // writer ran, which is the leak this closes).
         (runId) => this.routingRationaleByRun.delete(runId));
     }
-    async resolveReviewers(cwd, runAuthPreference) {
+    async resolveReviewers(cwd, runAuthPreference, onIgnoredSetting) {
         if (this.deps.reviewers)
             return this.deps.reviewers;
         if (this.deps.reviewerPanel && this.deps.reviewerPanel.length > 0) {
@@ -215,10 +220,8 @@ export class Orchestrator {
             registry: this.deps.registry,
             harnessSettings: this.config(cwd)?.global.harnesses ?? {},
             authPreferenceFor: (id) => this.authPreferenceForHarness(cwd, id, runAuthPreference),
-        }, {
-            reviewerModels: this.deps.reviewerModels,
-            reviewerEfforts: this.deps.reviewerEfforts,
-        });
+            onIgnoredSetting,
+        }, { reviewerModels: this.deps.reviewerModels, reviewerEfforts: this.deps.reviewerEfforts });
     }
     /**
      * Resolve reviewers INSIDE a strategy, after run-dir creation: an explicit
@@ -229,7 +232,9 @@ export class Orchestrator {
      */
     async resolveReviewersWithArtifacts(input, log, store, paths, runId, taskId, mode) {
         try {
-            return { reviewers: await this.resolveReviewers(input.repoRoot, input.authPreference) };
+            // Auto-panel dropped knobs (reviewerEfforts) → ignored-settings channel (QA-070):
+            const warn = (d) => void log.emit("review.preflight", { ignored_settings: [d] });
+            return { reviewers: await this.resolveReviewers(input.repoRoot, input.authPreference, warn) };
         }
         catch (err) {
             const message = safeErrorMessage(err);
@@ -917,13 +922,13 @@ export class Orchestrator {
                 intent,
                 qualityTiers: config.routing.quality_tiers,
                 ledger: routeLedger,
+                now: Date.now(), // ONE instant for the sort AND the rationale below
             };
             const ranked = rankHarnesses(remaining, routeCtx)
                 .map((candidate) => byId.get(candidate.harnessId))
                 .filter((candidate) => Boolean(candidate));
-            // QA-034: record the typed rationale ONCE at pool ordering (run evidence,
-            // not an event). Axis-aligned with rankHarnesses above so the persisted
-            // reason can never disagree with the order actually taken.
+            // QA-034: the rationale is run evidence recorded ONCE at pool ordering,
+            // pinned to routeCtx.now so it cannot disagree with the order just taken.
             if (runId)
                 this.routingRationaleByRun.set(runId, explainRanking(remaining, routeCtx));
             ordered = ranked;
@@ -1067,7 +1072,7 @@ export class Orchestrator {
             projectCommands: cfg.tests?.commands ?? [],
         });
         const commands = resolvedGates.commands;
-        const protectedPaths = [];
+        const protectedPaths = [...new Set(cfg.constraints.protected_paths)];
         const autoProtectedPaths = resolvedGates.autoProtectedPaths;
         const protectedPathApprovals = [
             ...new Map([...(input.protectedPathApprovals ?? [])].map((approval) => [approval.path, approval])).values(),
@@ -1278,17 +1283,16 @@ export class Orchestrator {
         // per-harness map, which beats the per-harness settings default. There is
         // no run-global model.
         const model = overrideModel ?? contract.routing_models[routed.adapter.id] ?? s?.defaultModel ?? null;
-        // Effort disclosure (INV-105): a requested effort on a harness with no
-        // declared ladder is DISCLOSED as ignored, never silently dropped.
-        // Harness-scoped resolution mirrors the model line above: the contract's
-        // FROZEN per-lane effort (QA-035) is authoritative so Exact Retry replays it
-        // without re-reading settings; a per-attempt `effortHint` (or settings
-        // default) applies only to a lane the contract did not freeze.
-        let effort = contract.routing_efforts[routed.adapter.id] ?? effortHint ?? s?.effort ?? null;
-        if (effort && routed.effortLevels.length === 0) {
-            ignored.push(`effort=${effort} (manifest capabilities.effort_levels is empty for ${routed.adapter.id})`);
-            effort = null;
-        }
+        // Effort disclosure (INV-105) against the harness's advertised ladder. This
+        // gate only DISCLOSES an unplaceable level; the clamp belongs to the adapter,
+        // which resolves against the catalog for the profile env the child runs in
+        // (the manifest here is the DEFAULT account's — see effortGovernance.ts). The
+        // contract's FROZEN per-lane effort (QA-035) wins so Exact Retry replays it
+        // without re-reading settings; `effortHint`/settings apply only to an unfrozen lane.
+        const governed = governRouteEffort(contract.routing_efforts[routed.adapter.id] ?? effortHint ?? s?.effort ?? null, { id: routed.adapter.id, ...routed });
+        const effort = governed.effort;
+        if (governed.ignored)
+            ignored.push(governed.ignored);
         return {
             model,
             effort,
@@ -1721,7 +1725,7 @@ export class Orchestrator {
         }
         const attemptStreamEndedMs = Date.now();
         if (webUnsatisfied(telemetry)) {
-            errors.push(`web evidence unsatisfied: ${telemetry.web.errorSummary ?? (telemetry.web.attempted ? "web tool failed without verified recovery" : "web evidence required but never attempted")}`);
+            errors.push(webEvidenceFailure(telemetry.web));
         }
         const diff = await wsm.diff(envelope);
         // D-16: un-nest {work_report, output} so answer.md persists the OUTPUT, not the envelope.
@@ -3436,7 +3440,7 @@ export class Orchestrator {
                     : "";
                 const prompt = attempt === 1
                     ? input.prompt
-                    : `${input.prompt}\n\nThe previous attempt did not converge. Address these review findings (verify each against the code; fix valid ones, rebut invalid ones with evidence):\n${formatFindings(lastFindings)}${runtimeErrors}`;
+                    : buildRevisePrompt(input.prompt, lastFindings, runtimeErrors);
                 // Loop detection (budget router): the 3rd identical repair prompt means
                 // findings/errors are not changing — stop burning paid attempts.
                 const fingerprint = promptFingerprint(prompt);
@@ -4208,20 +4212,22 @@ export class Orchestrator {
             input.signal?.removeEventListener("abort", onAbort);
             ledger.settle(lease.lease?.lease_id ?? "", attemptUsageCostSettlement(cost, costEstimated, attemptId, adapter.id, telemetry.authMode, telemetry.usageCost));
         }
-        const unrecovered = unrecoveredToolErrors(telemetry);
-        const webBlocked = webUnsatisfied(telemetry);
-        if (!harnessError && webBlocked) {
-            harnessError = `web evidence unsatisfied: ${telemetry.web.errorSummary ?? (telemetry.web.attempted ? "web tool failed without verified recovery" : "web evidence required but never attempted")}`;
-        }
-        if (!harnessError && unrecovered.length > 0) {
-            const first = unrecovered[0];
-            harnessError = `${first.tool} failed without recovery: ${first.summary}`;
-        }
         // D-16: unwrap and require PLAN TEXT — a plan with no text is not delivered.
+        // The unwrap runs BEFORE the error axes: the deliverable it yields is what
+        // decides whether an unrecovered tool error is fatal (explorer parity).
         const planUnwrapped = unwrapWorkReportEnvelope(answer.machineText() ?? "", planWorkMode, {
             sideToolReport: telemetry.sideToolWorkReport ?? undefined,
         });
         const planText = redactSecrets(planUnwrapped.deliverable).trim();
+        const unrecovered = unrecoveredToolErrors(telemetry);
+        const webBlocked = webUnsatisfied(telemetry);
+        if (!harnessError && webBlocked) {
+            harnessError = webEvidenceFailure(telemetry.web);
+        }
+        // INV-043/INV-044, explorer parity: a DELIVERED plan keeps an unrecovered
+        // non-web tool error as warning evidence instead of discarding the plan (see
+        // the helper). Web keeps its hard gate above; the finalizer outranks both.
+        harnessError ??= unrecoveredToolErrorFailure(unrecovered, planText.length > 0);
         const planFinalized = finalizeAttempt({
             deliverableEvidence: planText.length > 0,
             harnessErrored: harnessError !== null && !webBlocked,
@@ -5112,12 +5118,9 @@ export class Orchestrator {
             const webBlocked = webUnsatisfied(telemetry);
             const reportPresent = report.length > 0;
             if (!harnessError && webBlocked) {
-                harnessError = `web evidence unsatisfied: ${telemetry.web.errorSummary ?? (telemetry.web.attempted ? "web tool failed without verified recovery" : "web evidence required but never attempted")}`;
-            }
-            if (!harnessError && unrecovered.length > 0 && !reportPresent) {
-                const first = unrecovered[0];
-                harnessError = `${first.tool} failed without recovery: ${first.summary}`;
+                harnessError = webEvidenceFailure(telemetry.web);
             }
+            harnessError ??= unrecoveredToolErrorFailure(unrecovered, reportPresent);
             const roFinalized = finalizeAttempt({
                 deliverableEvidence: reportPresent,
                 harnessErrored: harnessError !== null && !webBlocked,
@@ -5696,7 +5699,4 @@ function assertNoSecretLikeTokens(label, text) {
         throw new Error(`${label} contains secret-like token; refusing to persist artifact`);
     }
 }
-function safeErrorMessage(err) {
-    return redactSecrets(err instanceof Error ? err.message : String(err));
-}
 //# sourceMappingURL=orchestrator.js.map