@bastani/atomic 0.9.21-alpha.1 → 0.9.21-alpha.2

package/dist/builtin/web-access/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.9.21-alpha.1",
+  "version": "0.9.21-alpha.2",
   "private": true,
   "description": "Atomic extension for web search, URL fetching, GitHub repo cloning, PDF/video extraction. Fork of: https://github.com/nicobailon/pi-web-access",
   "contributors": [

package/dist/builtin/workflows/CHANGELOG.md

@@ -6,6 +6,17 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ## [Unreleased]
 
+## [0.9.21-alpha.2] - 2026-09-26
+
+### Changed
+
+- `model: "auto"` routing has a chat model answer fixed questions about the complete task, then the router chooses between a short, evidence-backed list of models without seeing the task, in at most two small requests. The new `modelRouting` setting's `allowedProviders` and `excludedProviders` lists limit which providers are candidates.
+- `model: "auto"` accepts `taskNeeds` so the caller can state the task's kind of work, difficulty, mistake cost, need for images, need for a very large context and whether speed matters, and `modelConstraints.allowedModels` sets the shortlist the router chooses between. `modelConstraints.allowedProviders` and `excludedProviders` skip whole providers and replace the `modelRouting` provider settings for that call. Falling back to the current chat model is silent unless `ATOMIC_MODEL_ROUTING_DEBUG=1`.
+
+### Fixed
+
+- The builtin `goal` and `ralph` workflows no longer count a reviewer's `stop_review_loop=true` as approval when the same review calls the patch incorrect, reports `goal_oracle_satisfied: false`, or lists a blocking finding. A run whose reviewers all flag an empty implementation no longer ends `complete` or starts the pull-request stage ([#3295](https://github.com/bastani-inc/atomic/issues/3295)).
+
 ## [0.9.21-alpha.1] - 2026-09-25
 
 ### Changed

package/dist/builtin/workflows/builtin/adversarial-verification.js

@@ -1,10 +1,10 @@
 import {
   adversarial_verification_default
-} from "./chunk-jaj27rqc.js";
+} from "./chunk-a1ar035d.js";
 import"./chunk-2377r57g.js";
 import"./chunk-n1910xc4.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   adversarial_verification_default as default
 };

package/dist/builtin/workflows/builtin/{chunk-cw86srkf.js → chunk-632fxeb8.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   stableArtifactRoot
 } from "./chunk-rc5ec5xx.js";

package/dist/builtin/workflows/builtin/{chunk-w18kbz63.js → chunk-94f8zfz1.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   score_progress,
   classify_trend

package/dist/builtin/workflows/builtin/{chunk-kspvwc24.js → chunk-9cm2ckq2.js}

@@ -178,7 +178,7 @@ var SCOPE_DISCIPLINE_CONTRACT = keepContext([
 `));
 var EVIDENCE_CLOSURE_POLICY = [
   "Convergence flag (stop_review_loop):",
-  "- stop_review_loop is the single authoritative convergence signal; the harness trusts it without recomputing approval from findings, priorities, or requirements_traceability.",
+  "- stop_review_loop is the convergence signal. The harness does not recompute approval from requirements_traceability, but it does not count stop_review_loop=true as approval when the same review says the patch is incorrect, reports goal_oracle_satisfied=false, or lists a blocking finding.",
   "- Derive stop_review_loop=false while any objective-relevant blocking work remains: a P0/P1/P2 finding, a required_by_objective finding at any priority including P3, or an unproven implementation/validation requirement.",
   "- Derive stop_review_loop=true when independent verification proves implementation and validation and only non-blocking items remain: consistent_with_objective P3 items, beyond_objective/contradicts_objective observations, an authorized post-approval PR/MR/review action, or reviewer quorum. Never hold it false for those items.",
   "- If the bounded loop ends first, preserve unresolved findings and remaining work for a human rather than relabeling them."

package/dist/builtin/workflows/builtin/{chunk-jaj27rqc.js → chunk-a1ar035d.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   parse_rubric,
   normalize_criteria,

package/dist/builtin/workflows/builtin/{chunk-wh19f5ee.js → chunk-d3xhj946.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   keepContext,
   REVIEWER_CALIBRATION_RULES,
@@ -24,7 +24,7 @@ import {
   EVIDENCE_CLOSURE_POLICY,
   WORKTREE_DISCIPLINE_CONTRACT,
   REVIEW_CODE_DELTA_CONTRACT
-} from "./chunk-kspvwc24.js";
+} from "./chunk-9cm2ckq2.js";
 import {
   VERIFICATION_SCALE
 } from "./chunk-n1910xc4.js";
@@ -33,12 +33,13 @@ import {
   record_convergence,
   convergence_escalation_evidence,
   finalActionRemaining,
+  stopReviewLoopContradictions,
   consolidateFindingsBatch,
   parseFailureDiagnostics,
   reviewerFailureText,
   summarizeReviewConvergence,
   reverify_consolidated_batch
-} from "./chunk-73c5tayt.js";
+} from "./chunk-kgn5wz1c.js";
 import {
   fold_usage
 } from "./chunk-tgt0s5e5.js";
@@ -52,7 +53,7 @@ import { dirname, join } from "node:path";
 import { Type } from "typebox";
 // dist/builtin/workflows/builtin/ralph-review-gate.ts
 function reviewDecisionApproved(decision) {
-  return decision.stop_review_loop === true && decision.reviewer_error == null;
+  return decision.stop_review_loop === true && decision.reviewer_error == null && stopReviewLoopContradictions(decision).length === 0;
 }
 
 // dist/builtin/workflows/builtin/ralph-core.ts

package/dist/builtin/workflows/builtin/{chunk-crw7q5g1.js → chunk-gyjck9z8.js}

@@ -1,6 +1,6 @@
 import {
   withSteeringPropagationContext2
-} from "./chunk-kspvwc24.js";
+} from "./chunk-9cm2ckq2.js";
 
 // dist/builtin/workflows/src/authoring/workflow.ts
 import { basename } from "node:path";

package/dist/builtin/workflows/builtin/{chunk-maqv731g.js → chunk-hj8mh07f.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   keepContext,
   REVIEWER_CALIBRATION_RULES,
@@ -24,19 +24,21 @@ import {
   EVIDENCE_CLOSURE_POLICY,
   WORKTREE_DISCIPLINE_CONTRACT,
   REVIEW_CODE_DELTA_CONTRACT
-} from "./chunk-kspvwc24.js";
+} from "./chunk-9cm2ckq2.js";
 import {
   workflowArtifactDirectoryPath,
   ensureWorkflowArtifactDirectory,
   record_convergence,
   convergence_escalation_evidence,
   finalActionRemaining,
+  stopReviewLoopContradictions,
+  contradictedStopReviewLoopGap,
   consolidateFindingsBatch,
   parseFailureDiagnostics,
   reviewerFailureText,
   summarizeReviewConvergence,
   reverify_consolidated_batch
-} from "./chunk-73c5tayt.js";
+} from "./chunk-kgn5wz1c.js";
 import {
   VERIFICATION_SCALE
 } from "./chunk-n1910xc4.js";
@@ -321,7 +323,7 @@ function reduceGoalDecision(ledger, turnReviews, options) {
         ...summary,
         turn: options.turn,
         decision: "complete",
-        reason: `Reviewer quorum met: ${completeVotes}/${options.reviewQuorum} reviewers independently reported stop_review_loop=true with no reviewer execution errors.`,
+        reason: `Reviewer quorum met: ${completeVotes}/${options.reviewQuorum} reviewers independently reported stop_review_loop=true with a correct patch, a satisfied goal oracle, no blocking findings, and no reviewer execution errors.`,
         complete_votes: completeVotes,
         review_quorum: options.reviewQuorum
       }
@@ -459,7 +461,7 @@ function parsedReviewDecisionFromResult(result, reviewer) {
   };
 }
 function reviewApproved(decision) {
-  return decision.stop_review_loop === true && decision.reviewer_error == null;
+  return decision.stop_review_loop === true && decision.reviewer_error == null && stopReviewLoopContradictions(decision).length === 0;
 }
 function reviewerErrorDecision(message) {
   return {
@@ -492,6 +494,7 @@ function blockerFromReviewDecision(decision) {
 function reviewDecisionToRecord(args) {
   const blocker = blockerFromReviewDecision(args.decision);
   const approved = reviewApproved(args.decision);
+  const contradictions = args.decision.stop_review_loop === true ? stopReviewLoopContradictions(args.decision) : [];
   const hasFinalActionRemaining = args.allowFinalActionRemaining && finalActionRemaining(args.decision.requirements_traceability);
   const verificationGap = args.decision.verification_remaining.trim();
   const traceabilityGaps = args.decision.requirements_traceability.filter((entry) => entry.status !== "proven").map((entry) => `${entry.status}: ${entry.requirement} — ${entry.evidence}`);
@@ -499,7 +502,8 @@ function reviewDecisionToRecord(args) {
     ...args.decision.findings.map((finding) => `[${finding.objective_alignment}] ${finding.title}: ${finding.body}`),
     ...traceabilityGaps,
     ...approved || verificationGap.length === 0 ? [] : [verificationGap],
-    ...args.decision.reviewer_error == null ? [] : [`${args.decision.reviewer_error.kind}: ${args.decision.reviewer_error.message}`]
+    ...args.decision.reviewer_error == null ? [] : [`${args.decision.reviewer_error.kind}: ${args.decision.reviewer_error.message}`],
+    ...contradictions.length === 0 ? [] : [contradictedStopReviewLoopGap(contradictions)]
   ];
   const nextAction = approved ? hasFinalActionRemaining ? "pull-request" : "finish" : blocker === null ? "implementation" : "blocked";
   const convergenceDecision = summarizeReviewConvergence({

package/dist/builtin/workflows/builtin/{chunk-0nrgfbag.js → chunk-kb7s3gx2.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   scoring_prompt_reads,
   build_scoring_prompt

package/dist/builtin/workflows/builtin/{chunk-73c5tayt.js → chunk-kgn5wz1c.js}

@@ -10476,6 +10476,24 @@ function findingBlocksClosure(finding) {
     return true;
   return priority <= MAX_BLOCKING_PRIORITY;
 }
+function stopReviewLoopContradictions(verdict) {
+  const contradictions = [];
+  if (verdict.overall_correctness !== "patch is correct") {
+    contradictions.push(`overall_correctness is "${verdict.overall_correctness}"`);
+  }
+  if (verdict.goal_oracle_satisfied === false) {
+    contradictions.push("goal_oracle_satisfied is false");
+  }
+  for (const finding of verdict.findings) {
+    if (findingBlocksClosure(finding)) {
+      contradictions.push(`blocking finding "${finding.title}"`);
+    }
+  }
+  return contradictions;
+}
+function contradictedStopReviewLoopGap(contradictions) {
+  return `stop_review_loop=true was not counted as approval because the same review reports ${contradictions.join("; ")}`;
+}
 function findingConsolidationKey(finding) {
   const normalizedTitle = finding.title.replace(/^\s*\[P[0-3]\]\s*/iu, "").toLowerCase().replace(/\s+/gu, " ").trim();
   return `${finding.code_location?.absolute_file_path ?? ""}::${normalizedTitle}`;
@@ -10696,4 +10714,4 @@ async function reverify_consolidated_batch(ctx, input) {
   return apply_reverify_results(input.batch, audits);
 }
 
-export { workflowArtifactDirectoryPath, ensureWorkflowArtifactDirectory, createWorkflowArtifactDirectory, record_convergence, convergence_escalation_evidence, finalActionRemaining, consolidateFindingsBatch, parseFailureDiagnostics, reviewerFailureText, summarizeReviewConvergence, reverify_consolidated_batch };
+export { workflowArtifactDirectoryPath, ensureWorkflowArtifactDirectory, createWorkflowArtifactDirectory, record_convergence, convergence_escalation_evidence, finalActionRemaining, stopReviewLoopContradictions, contradictedStopReviewLoopGap, consolidateFindingsBatch, parseFailureDiagnostics, reviewerFailureText, summarizeReviewConvergence, reverify_consolidated_batch };

package/dist/builtin/workflows/builtin/{chunk-2hak3b7g.js → chunk-nkj4qhj6.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 
 // dist/builtin/workflows/builtin/open-claude-design.ts
 import { Type as Type2 } from "typebox";

package/dist/builtin/workflows/builtin/{chunk-bzdvx2xr.js → chunk-r97ndfvt.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   stableArtifactRoot
 } from "./chunk-rc5ec5xx.js";

package/dist/builtin/workflows/builtin/{chunk-vkmpdat4.js → chunk-wf876n16.js}

@@ -1,7 +1,7 @@
 import {
   workflow,
   withBuiltinContext
-} from "./chunk-crw7q5g1.js";
+} from "./chunk-gyjck9z8.js";
 import {
   parse_rubric,
   normalize_criteria,

package/dist/builtin/workflows/builtin/classify-and-act.js

@@ -1,9 +1,9 @@
 import {
   DEFAULT_ACTION_CATEGORIES2,
   classify_and_act_default
-} from "./chunk-bzdvx2xr.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+} from "./chunk-r97ndfvt.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   DEFAULT_ACTION_CATEGORIES2 as DEFAULT_ACTION_CATEGORIES,
   classify_and_act_default as default

package/dist/builtin/workflows/builtin/fan-out-and-synthesize.js

@@ -1,8 +1,8 @@
 import {
   fan_out_and_synthesize_default
-} from "./chunk-cw86srkf.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+} from "./chunk-632fxeb8.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   fan_out_and_synthesize_default as default
 };

package/dist/builtin/workflows/builtin/generate-and-filter.js

@@ -1,10 +1,10 @@
 import {
   generate_and_filter_default
-} from "./chunk-0nrgfbag.js";
+} from "./chunk-kb7s3gx2.js";
 import"./chunk-2377r57g.js";
 import"./chunk-n1910xc4.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   generate_and_filter_default as default
 };

package/dist/builtin/workflows/builtin/goal.js

@@ -1,12 +1,12 @@
 import {
   goal_default
-} from "./chunk-maqv731g.js";
-import"./chunk-73c5tayt.js";
+} from "./chunk-hj8mh07f.js";
+import"./chunk-kgn5wz1c.js";
 import"./chunk-2377r57g.js";
 import"./chunk-z0d03ktm.js";
 import"./chunk-n1910xc4.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   goal_default as default
 };

package/dist/builtin/workflows/builtin/index.js

@@ -1,36 +1,36 @@
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 import"./chunk-n1910xc4.js";
 import"./chunk-2377r57g.js";
 import {
   adversarial_verification_default
-} from "./chunk-jaj27rqc.js";
+} from "./chunk-a1ar035d.js";
 import {
   classify_and_act_default
-} from "./chunk-bzdvx2xr.js";
+} from "./chunk-r97ndfvt.js";
 import {
   fan_out_and_synthesize_default
-} from "./chunk-cw86srkf.js";
+} from "./chunk-632fxeb8.js";
 import {
   generate_and_filter_default
-} from "./chunk-0nrgfbag.js";
-import"./chunk-73c5tayt.js";
+} from "./chunk-kb7s3gx2.js";
+import"./chunk-kgn5wz1c.js";
 import"./chunk-z0d03ktm.js";
 import {
   goal_default
-} from "./chunk-maqv731g.js";
+} from "./chunk-hj8mh07f.js";
 import {
   open_claude_design_default
-} from "./chunk-2hak3b7g.js";
+} from "./chunk-nkj4qhj6.js";
 import {
   loop_until_done_default
-} from "./chunk-w18kbz63.js";
+} from "./chunk-94f8zfz1.js";
 import {
   ralph_default
-} from "./chunk-wh19f5ee.js";
+} from "./chunk-d3xhj946.js";
 import {
   tournament_default
-} from "./chunk-vkmpdat4.js";
+} from "./chunk-wf876n16.js";
 export {
   adversarial_verification_default as adversarialVerification,
   classify_and_act_default as classifyAndAct,

package/dist/builtin/workflows/builtin/loop-until-done.js

@@ -1,10 +1,10 @@
 import {
   loop_until_done_default
-} from "./chunk-w18kbz63.js";
+} from "./chunk-94f8zfz1.js";
 import"./chunk-z0d03ktm.js";
 import"./chunk-n1910xc4.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   loop_until_done_default as default
 };

package/dist/builtin/workflows/builtin/open-claude-design.js

@@ -1,8 +1,8 @@
 import {
   open_claude_design_default
-} from "./chunk-2hak3b7g.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+} from "./chunk-nkj4qhj6.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   open_claude_design_default as default
 };

package/dist/builtin/workflows/builtin/ralph.js

@@ -1,12 +1,12 @@
 import {
   ralph_default
-} from "./chunk-wh19f5ee.js";
-import"./chunk-73c5tayt.js";
+} from "./chunk-d3xhj946.js";
+import"./chunk-kgn5wz1c.js";
 import"./chunk-2377r57g.js";
 import"./chunk-z0d03ktm.js";
 import"./chunk-n1910xc4.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   ralph_default as default
 };

package/dist/builtin/workflows/builtin/steering-context.js

@@ -1,6 +1,6 @@
 import {
   withSteeringPropagationContext2
-} from "./chunk-kspvwc24.js";
+} from "./chunk-9cm2ckq2.js";
 export {
   withSteeringPropagationContext2 as withSteeringPropagationContext
 };

package/dist/builtin/workflows/builtin/tournament.js

@@ -1,10 +1,10 @@
 import {
   tournament_default
-} from "./chunk-vkmpdat4.js";
+} from "./chunk-wf876n16.js";
 import"./chunk-2377r57g.js";
 import"./chunk-n1910xc4.js";
-import"./chunk-crw7q5g1.js";
-import"./chunk-kspvwc24.js";
+import"./chunk-gyjck9z8.js";
+import"./chunk-9cm2ckq2.js";
 export {
   tournament_default as default
 };

package/dist/builtin/workflows/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/workflows",
-  "version": "0.9.21-alpha.1",
+  "version": "0.9.21-alpha.2",
   "private": true,
   "description": "Atomic extension for multi-stage workflow authoring and execution.",
   "contributors": [

package/dist/builtin/workflows/src/extension/index.bundle.mjs

@@ -94631,7 +94631,7 @@ var SCOPE_DISCIPLINE_CONTRACT = keepContext([
 `));
 var EVIDENCE_CLOSURE_POLICY = [
   "Convergence flag (stop_review_loop):",
-  "- stop_review_loop is the single authoritative convergence signal; the harness trusts it without recomputing approval from findings, priorities, or requirements_traceability.",
+  "- stop_review_loop is the convergence signal. The harness does not recompute approval from requirements_traceability, but it does not count stop_review_loop=true as approval when the same review says the patch is incorrect, reports goal_oracle_satisfied=false, or lists a blocking finding.",
   "- Derive stop_review_loop=false while any objective-relevant blocking work remains: a P0/P1/P2 finding, a required_by_objective finding at any priority including P3, or an unproven implementation/validation requirement.",
   "- Derive stop_review_loop=true when independent verification proves implementation and validation and only non-blocking items remain: consistent_with_objective P3 items, beyond_objective/contradicts_objective observations, an authorized post-approval PR/MR/review action, or reviewer quorum. Never hold it false for those items.",
   "- If the bounded loop ends first, preserve unresolved findings and remaining work for a human rather than relabeling them."
@@ -96509,6 +96509,24 @@ function findingBlocksClosure(finding) {
     return true;
   return priority <= MAX_BLOCKING_PRIORITY;
 }
+function stopReviewLoopContradictions(verdict) {
+  const contradictions = [];
+  if (verdict.overall_correctness !== "patch is correct") {
+    contradictions.push(`overall_correctness is "${verdict.overall_correctness}"`);
+  }
+  if (verdict.goal_oracle_satisfied === false) {
+    contradictions.push("goal_oracle_satisfied is false");
+  }
+  for (const finding of verdict.findings) {
+    if (findingBlocksClosure(finding)) {
+      contradictions.push(`blocking finding "${finding.title}"`);
+    }
+  }
+  return contradictions;
+}
+function contradictedStopReviewLoopGap(contradictions) {
+  return `stop_review_loop=true was not counted as approval because the same review reports ${contradictions.join("; ")}`;
+}
 function findingConsolidationKey(finding) {
   const normalizedTitle = finding.title.replace(/^\s*\[P[0-3]\]\s*/iu, "").toLowerCase().replace(/\s+/gu, " ").trim();
   return `${finding.code_location?.absolute_file_path ?? ""}::${normalizedTitle}`;
@@ -96762,7 +96780,7 @@ function reduceGoalDecision(ledger, turnReviews, options) {
         ...summary,
         turn: options.turn,
         decision: "complete",
-        reason: `Reviewer quorum met: ${completeVotes}/${options.reviewQuorum} reviewers independently reported stop_review_loop=true with no reviewer execution errors.`,
+        reason: `Reviewer quorum met: ${completeVotes}/${options.reviewQuorum} reviewers independently reported stop_review_loop=true with a correct patch, a satisfied goal oracle, no blocking findings, and no reviewer execution errors.`,
         complete_votes: completeVotes,
         review_quorum: options.reviewQuorum
       }
@@ -96900,7 +96918,7 @@ function parsedReviewDecisionFromResult(result, reviewer) {
   };
 }
 function reviewApproved(decision) {
-  return decision.stop_review_loop === true && decision.reviewer_error == null;
+  return decision.stop_review_loop === true && decision.reviewer_error == null && stopReviewLoopContradictions(decision).length === 0;
 }
 function reviewerErrorDecision(message) {
   return {
@@ -96933,6 +96951,7 @@ function blockerFromReviewDecision(decision) {
 function reviewDecisionToRecord(args) {
   const blocker = blockerFromReviewDecision(args.decision);
   const approved = reviewApproved(args.decision);
+  const contradictions = args.decision.stop_review_loop === true ? stopReviewLoopContradictions(args.decision) : [];
   const hasFinalActionRemaining = args.allowFinalActionRemaining && finalActionRemaining(args.decision.requirements_traceability);
   const verificationGap = args.decision.verification_remaining.trim();
   const traceabilityGaps = args.decision.requirements_traceability.filter((entry) => entry.status !== "proven").map((entry) => `${entry.status}: ${entry.requirement} — ${entry.evidence}`);
@@ -96940,7 +96959,8 @@ function reviewDecisionToRecord(args) {
     ...args.decision.findings.map((finding) => `[${finding.objective_alignment}] ${finding.title}: ${finding.body}`),
     ...traceabilityGaps,
     ...approved || verificationGap.length === 0 ? [] : [verificationGap],
-    ...args.decision.reviewer_error == null ? [] : [`${args.decision.reviewer_error.kind}: ${args.decision.reviewer_error.message}`]
+    ...args.decision.reviewer_error == null ? [] : [`${args.decision.reviewer_error.kind}: ${args.decision.reviewer_error.message}`],
+    ...contradictions.length === 0 ? [] : [contradictedStopReviewLoopGap(contradictions)]
   ];
   const nextAction = approved ? hasFinalActionRemaining ? "pull-request" : "finish" : blocker === null ? "implementation" : "blocked";
   const convergenceDecision = summarizeReviewConvergence({
@@ -99150,7 +99170,7 @@ import { dirname as dirname7, join as join19 } from "node:path";
 import { Type as Type19 } from "typebox";
 // dist/builtin/workflows/builtin/ralph-review-gate.ts
 function reviewDecisionApproved(decision) {
-  return decision.stop_review_loop === true && decision.reviewer_error == null;
+  return decision.stop_review_loop === true && decision.reviewer_error == null && stopReviewLoopContradictions(decision).length === 0;
 }
 
 // dist/builtin/workflows/builtin/ralph-core.ts
@@ -109393,7 +109413,9 @@ import { createStructuredOutputCapture, runCallback as runCallback2 } from "@bas
 import {
   AutoRoutingInferenceError,
   convertToLlm,
-  parseModelConstraints
+  isModelRoutingDebugEnabled,
+  parseModelConstraints,
+  setsProviders
 } from "@bastani/atomic";
 
 // dist/builtin/workflows/src/runs/shared/model-fallback-candidates.ts
@@ -110134,6 +110156,7 @@ function stripWorkflowOnlyOptions(options, defaultSessionDir, meta, pendingStage
     fallbackModels: _fallbackModels,
     fallbackThinkingLevels: _fallbackThinkingLevels,
     modelConstraints: _modelConstraints,
+    taskNeeds: _taskNeeds,
     inheritedModelConstraints: _inheritedModelConstraints,
     routerSelection: _routerSelection,
     context,
@@ -110988,7 +111011,9 @@ class StageSessionController {
             stageName: this.opts.stageName,
             constraints,
             signal: this.startupWait.signal,
-            selection: options?.routerSelection
+            selection: options?.routerSelection,
+            ...options?.taskNeeds ? { taskNeeds: options.taskNeeds } : {},
+            ...setsProviders(parseModelConstraints(options?.modelConstraints)) ? { overrideProviderSettings: true } : {}
           });
         } catch (error) {
           this.startupWait.signal.throwIfAborted();
@@ -110996,7 +111021,8 @@ class StageSessionController {
           if (!(error instanceof AutoRoutingInferenceError) || error.currentModelRoute === undefined)
             throw error;
           this.modelRoute = error.currentModelRoute;
-          this.modelCatalog?.recordWarning?.(`workflows: stage auto routing failed; running "${this.opts.stageName}" on the current chat model ${this.modelRoute.modelOverride}.`);
+          if (isModelRoutingDebugEnabled())
+            this.modelCatalog?.recordWarning?.(`workflows: stage auto routing failed; running "${this.opts.stageName}" on the current chat model ${this.modelRoute.modelOverride}.`);
         }
         this.modelRoute.assertCurrent();
         this.meta.stageOptions = { ...options, model: this.modelRoute.modelOverride };
@@ -116349,6 +116375,7 @@ function workflowModelCatalogFromContext(ctx, resolveLiveContext) {
         ctx: {
           model: current.model,
           getRouterModel: () => current.getRouterModel(),
+          ...current.getModelRouting ? { getModelRouting: () => current.getModelRouting() } : {},
           modelRegistry: {
             getAvailable: () => registry.getAvailable(),
             getAll: () => registry.getAll(),
@@ -116364,7 +116391,9 @@ function workflowModelCatalogFromContext(ctx, resolveLiveContext) {
         agent: { name: input.stageName, description: "Workflow stage" },
         constraints: input.constraints,
         signal: input.signal,
-        selection: input.selection
+        selection: input.selection,
+        ...input.taskNeeds ? { taskNeeds: input.taskNeeds } : {},
+        ...input.overrideProviderSettings ? { overrideProviderSettings: true } : {}
       });
     },
     listModels: async () => {
@@ -122051,7 +122080,7 @@ var DEFAULT_PROMPT_GUIDANCE = [
   - Before launching, share a brief duration estimate without confidence labels. Estimates are not guarantees or budgets.
   - Preserve inherited budget limits and approval gates without asking for a budget before each launch. Pass \`budget\` only for limits the user stated; omission inherits and zero disables only its field. Never convert an estimate into a cap.
   - Heartbeat cadence is 15 minutes by default. Keep that interval unless the user explicitly asks for a different cadence. A heartbeat is a periodic alignment check: re-read the objective, judge whether the run is still on goal, and continue a progressing run when no intervention is needed. Use workflow pause, resume, or quit when run control is useful, and ask the user when a decision is needed.
-  - Wrap critical parts of the \`inputs\` you pass — objective, prompt, acceptance criteria — in \`<keepContext>\` / \`</keepContext>\`. Tagged text survives compaction verbatim regardless of the compression ratio, and stages inherit it because workflows inject your inputs into their prompts. A long stage is compacted repeatedly, and compaction ranks lines individually: a verbose restated objective outlives the one line that bounds it, so a stage can end up executing a coherent but unbounded version of what you asked for. Decide this per launch rather than tagging everything — tag role constraints, prohibitions, must-hold criteria, and identifiers such as a target branch, worktree path, or issue number; leave background, quoted issue text, and reference material untagged and pass bulk through files and \`reads\`. Protected lines count against the keep target rather than raising it, so an over-tagged input makes the evidence stages need compress harder. \`model: "auto"\` stages also route on an excerpt of about 9 KB (roughly 1,200 words) that keeps only the beginning, the end, and the protected spans, so state the objective in the first or last lines of an input or inside a short tagged span, keep tagged spans to a few lines, and put bulky material in the middle.
+  - Wrap critical parts of the \`inputs\` you pass — objective, prompt, acceptance criteria — in \`<keepContext>\` / \`</keepContext>\`. Tagged text survives compaction verbatim regardless of the compression ratio, and stages inherit it because workflows inject your inputs into their prompts. A long stage is compacted repeatedly, and compaction ranks lines individually: a verbose restated objective outlives the one line that bounds it, so a stage can end up executing a coherent but unbounded version of what you asked for. Decide this per launch rather than tagging everything — tag role constraints, prohibitions, must-hold criteria, and identifiers such as a target branch, worktree path, or issue number; leave background, quoted issue text, and reference material untagged and pass bulk through files and \`reads\`. Protected lines count against the keep target rather than raising it, so an over-tagged input makes the evidence stages need compress harder.
   - Tag steering the same way. An Intercom \`send\` amendment is authoritative and stages must carry it forward, but it arrives late in an already-long session and is exactly one short message competing with the whole transcript; \`<keepContext>\` around the requirement keeps it from being compacted away before the stage acts on it.
   - In interactive chat, named workflow launches run in the background. Run \`/workflow connect <run>\` to see agents working and chat with and steer each stage. Inspection and control calls (\`status\`, \`stages\`, \`stage\`, \`transcript\`, \`answer\`, \`pause\`, \`resume\`, \`quit\`) remain available while work runs.
   - At lifecycle updates, revise the estimate only when new evidence materially changes the remaining path; at completion, report actual elapsed time against the estimate and note the main bottleneck for future scheduling. Do not add confidence labels or scores to duration estimates.
@@ -122089,7 +122118,7 @@ var DEFAULT_PROMPT_GUIDANCE = [
   - Prefer \`ctx.tool(name, args, fn)\` for workflow-owned TypeScript operations with side effects, such as filesystem writes, network mutations, and external API actions. It creates a tracked, non-attachable durable graph node before invoking \`fn\`; it may be used before, between, after, or without model stages. A completed \`ctx.tool\` call is durably checkpointed with its serializable result, so resume replays that result without rerunning \`fn\`. For checks that may fail during a bounded repair loop, opt into \`failureMode: "return"\`, branch on the typed outcome, and pass only the needed error fields to the repair stage or artifact yourself; Atomic does not inject failure evidence into later prompts. Retries do not bound a hung callback, so every \`ctx.tool\` callback that starts a child process or performs network I/O MUST set an explicit positive finite \`timeoutMs\` and forward the supplied \`signal\` to cancellable work. A timeout applies per attempt, aborts that attempt's signal, and gives retries a fresh deadline and signal. Cancellation and storage faults still throw. Keep pure computation and side-effect-free transformations as ordinary TypeScript. Do not wrap agent-stage internals or every function call indiscriminately; this rule applies to side effects orchestrated directly by the workflow definition.
   - Pass large stage context through files/artifacts and \`reads\` with an explicit \`Read the file at <path>...\` prompt rather than large \`previous\` payloads or injected session history.
   - Wrap critical stage-prompt content in \`<keepContext>\` / \`</keepContext>\`. Tagged content survives compression verbatim regardless of the compression ratio, and every line of the span including the tag lines is protected. A long-running stage will be compacted, and compaction ranks lines individually: an objective is verbose and restated while the constraint that narrows it is usually one line, so the constraint is the cheaper deletion and what survives reads as broader permission than intended. Tag role constraints that bound a stage to part of the work, acceptance criteria and immutable contracts, explicit prohibitions, and identifiers a stage must not lose such as a target branch, worktree path, or run ID. Do not tag bulk context: protected lines count against the keep target rather than raising it, so a large protected span makes the surrounding transcript compress harder — tag the constraint, not the material it applies to.
-  - Structure \`model: "auto"\` stage prompts for the model router. Routing sees an excerpt of about 9 KB (roughly 1,200 words): the beginning, the end, and every \`<keepContext>\` span, with cuts marked \`[... truncated ...]\`; the stage still receives the full prompt. Put the stage's role and objective in the first or last lines or a short tagged span, place artifact reads, issue text, and prior-stage output in the middle, and keep tagged spans brief — spans that alone exceed the excerpt are cut through the middle.
+  - For \`model: "auto"\` stages, a chat model reads the prompt to answer the routing questions. Set \`taskNeeds\` (work, difficulty, mistakeCost, needsImages, longContext, latencySensitive) with everything you can judge so routing does not have to infer it. To skip whole providers because the user asked, set \`modelConstraints.allowedProviders\` or \`excludedProviders\`, which replace the user's \`modelRouting\` provider settings for that stage. To choose the contenders yourself, list them in \`modelConstraints.allowedModels\`; prefer the user's subscription models (for example ChatGPT/Codex, GitHub Copilot, a Claude subscription, or a coding or token plan) over pay-per-token API models, unless the user asked for API models or has no subscription models.
   - Separate implementation/review acceptance from explicitly authorized final actions such as PR/MR/review creation, release tagging, deployment, or publication. Stop implementation loops once acceptance is proven and carry a remaining final action separately.
   - A model stage sees its local prompt, artifacts, tools, and reads, not the graph name or surrounding implementation. State the concrete action, evidence, and success criteria directly.`
 ];