atris 3.74.20 → 3.74.21

package/commands/task.js

@@ -32,7 +32,7 @@ const {
   runVerifyCommand,
   DENIED_TAGS,
 } = require('../lib/auto-accept-certified');
-const { evaluateAcceptVerify } = require('../lib/accept-verify-gate');
+const { evaluateAcceptVerify, nonFalsifyingIssue } = require('../lib/accept-verify-gate');
 const { checkedCodeSha, runVerifyAtJobCode } = require('../lib/job-code-verify');
 const { proveTaskLanding } = require('../lib/landing-proof');
 const {
@@ -10089,7 +10089,7 @@ function certifyVerifyCandidate(task) {
     // The evidence extractor can keep trailing prose ("git diff --check. Evidence
     // inspected: ..."); try the full clause, then the first sentence of it.
     for (const candidate of [raw, raw.split(/\.(?:\s|$)/)[0]].map((c) => String(c || '').trim())) {
-      if (candidate && isAutoCertifyVerifyCommandAllowed(candidate) && !verifyIsOnlyFileChecks(candidate)) return candidate;
+      if (candidate && !nonFalsifyingIssue(candidate) && isAutoCertifyVerifyCommandAllowed(candidate) && !verifyIsOnlyFileChecks(candidate)) return candidate;
     }
   }
   return null;
@@ -10702,6 +10702,10 @@ function evaluateSweepAutoAccept(task, root) {
   if (denied) return { eligible: false, ref, reason: denied };
   const proof = autoAcceptSweepLatestProof(task);
   if (!proof) return { eligible: false, ref, reason: 'no_proof' };
+  const storedVerify = typeof metadata.verify === 'string' ? metadata.verify.trim() : '';
+  if (storedVerify && nonFalsifyingIssue(storedVerify)) {
+    return { eligible: false, ref, reason: 'non_falsifying_verify', verify: storedVerify };
+  }
   const policyGate = candidatePolicyGate(task, { executeDetectors: true });
   if (!policyGate.ok) return { ...policyGate, ref };
 
@@ -10712,7 +10716,6 @@ function evaluateSweepAutoAccept(task, root) {
   // executed) in the proof, but the older receipt-path check below could not
   // find file evidence for it once `ready --verify` stopped writing a
   // receipt file and started embedding the executed result into proof text.
-  const storedVerify = typeof metadata.verify === 'string' ? metadata.verify.trim() : '';
   if (storedVerify) {
     const result = runVerifyCommand(storedVerify, root);
     if (!result.ok) return { eligible: false, ref, reason: result.reason, verify: storedVerify };
@@ -10770,6 +10773,12 @@ function evaluateSweepAutoAccept(task, root) {
       }
       return { eligible: false, ref, reason: result.reason, verify: derived };
     }
+    // Certification filters these out, but the sweep should explain why a
+    // legacy proof with only whitespace evidence cannot approve the work.
+    const proofCommands = taskReviewEvidenceCommands(proof);
+    if (proofCommands.length && proofCommands.every(command => nonFalsifyingIssue(command))) {
+      return { eligible: false, ref, reason: 'non_falsifying_verify' };
+    }
   }
 
   return { eligible: false, ref, reason: verifier.reason, evidence: verifier.evidence };
@@ -10783,6 +10792,14 @@ function cmdSweep(args) {
       'atris task sweep currently requires --auto-accept for the explicit verified accept policy',
     );
   }
+  // Sweep has no dry-run implementation: every successful invocation can
+  // write accepts, even when an unrecognized --dry-run flag is supplied.
+  if (agentProofOnlyMode()) {
+    failAgentProofOnly(
+      'atris task sweep',
+      'agents cannot accept tasks or award xp. leave proof in review for a human to accept or revise.',
+    );
+  }
   const actor = String(flag(args, '--as') || 'orb-autoaccept');
   const parsedReward = parseAcceptReward(flag(args, '--reward'));
   if (!parsedReward.ok) {

package/lib/auto-accept-certified.js

@@ -10,6 +10,7 @@ const { computeTrustTier, taskTrustSubject } = require('./trust-tiers');
 const { rewriteWorktreeVerify, missingOnMain } = require('./verify-worktree-rewrite');
 const { policyLessonsForFiles, readPolicyLessons } = require('./policy-lessons');
 const { skillEvalGate } = require('./skill-eval-gate');
+const escapeRegExp = require('./escape-regexp');
 
 const AGENT_CERTIFICATION_REVIEW_PASSES = 2;
 
@@ -1074,9 +1075,10 @@ function declaredProtectedLane(task) {
 // work and would hold the whole queue.
 const PROTECTED_LANE_TERMS = [
   ['billing', ['stripe', 'checkout url', 'checkout_url', 'invoice', 'paywall', 'subscription', 'wallet', 'credit card', 'pricing tier']],
+  ['money', ['money', 'refund', 'reimburs', 'expense', 'payment', 'cash', 'bank', 'venmo', 'zelle', 'tax']],
   ['security', ['session minting', 'mints a', 'service token', 'api key', 'credential', 'secret key', 'csrf', 'access token']],
   ['deploy', ['render deploy', 'production deploy', 'deploy to prod', 'migration apply']],
-  ['customer', ['send email to', 'customer email', 'outbound email', 'notify users']],
+  ['customer', ['send email to', 'customer email', 'outbound email', 'notify users', 'imessage', 'text message', 'send a text', 'reply to', 'outreach', 'cold email', 'send to', 'dm ']],
 ];
 
 function sniffedProtectedLane(task) {
@@ -1086,7 +1088,7 @@ function sniffedProtectedLane(task) {
     .toLowerCase();
   if (!haystack.trim()) return null;
   for (const [lane, terms] of PROTECTED_LANE_TERMS) {
-    const term = terms.find((t) => haystack.includes(t));
+    const term = terms.find((t) => new RegExp(`\\b${escapeRegExp(t)}`).test(haystack));
     if (term) return { lane, term };
   }
   return null;
@@ -1216,6 +1218,12 @@ function evaluateAutoAccept(task, options = {}) {
     };
   }
 
+  // Load lazily: the human accept gate reuses this module's parser and runner.
+  const verifyCommand = resolveTaskVerify(task).verify;
+  if (verifyCommand && require('./accept-verify-gate').nonFalsifyingIssue(verifyCommand)) {
+    return { eligible: false, ref, reason: 'non_falsifying_verify', verify: verifyCommand };
+  }
+
   // accept-all: the protected lanes above are the only human gate. No
   // certification, pass-count, reviewer, or proof-quality bar, but work
   // is never marked done against evidence it isn't: a proof naming an

package/lib/context-gatherer.js

@@ -155,6 +155,8 @@ function createStarterTask(root, answer) {
       title,
       status: task && task.status || 'claimed',
       claimed_by: (task && task.claimed_by) || who,
+      workspace_root: workspaceRoot,
+      metadata: task && task.metadata,
     };
   } catch (error) {
     return {

package/lib/first-minute.js

@@ -6,10 +6,12 @@ const path = require('path');
 const { stateHome } = require('./state-home');
 const { spawnSync } = require('child_process');
 const { isGenericScratchRoot } = require('./scratch-root');
+const { resolveDefaultVerifier } = require('./default-verifier');
+const { nonFalsifyingIssue } = require('./accept-verify-gate');
 
 const FIRST_USE_COMMAND = 'atris "help me choose the first useful step for this project"';
 const FIRST_TALK_ASK = 'what do you want here';
-const CLAIMED_KEEP_WORKING_VERIFY = 'git diff --check';
+const ADD_TEST_COMMAND = 'atris "add a test for the claimed task"';
 const LATER_NOTE_FILE = '.atris-later';
 const BARE_ATRIS_FLAGS = new Set(['--yes', '-y', '--json', '--verbose']);
 const NOT_A_PERSON = new Set([
@@ -237,13 +239,17 @@ function loadLatestRecap(root = process.cwd()) {
 
 function claimedKeepWorkingCommand(task) {
   const ref = taskRef(task);
+  const recorded = task && task.metadata && task.metadata.verify;
+  const available = recorded || (task && task.workspace_root && resolveDefaultVerifier(task.workspace_root));
+  if (!available || nonFalsifyingIssue(available)) return ADD_TEST_COMMAND;
   return ref
-    ? `atris task ready ${ref} --verify "${CLAIMED_KEEP_WORKING_VERIFY}"`
-    : `atris task ready --verify "${CLAIMED_KEEP_WORKING_VERIFY}"`;
+    ? `atris task ready ${ref} --verify "${available}"`
+    : `atris task ready --verify "${available}"`;
 }
 
 function isClaimedKeepWorkingCommand(command) {
-  return /^atris task ready\b/.test(String(command || ''));
+  return /^atris task ready\b/.test(String(command || ''))
+    || command === ADD_TEST_COMMAND;
 }
 
 function taskCommand(task, person) {
@@ -709,7 +715,7 @@ function firstTalkNext({ starter = null, person = '', folder = 'this folder' } =
   const room = folder || 'this folder';
   if (starter && starter.display_id) {
     return taskCommand({
-      display_id: starter.display_id,
+      ...starter,
       status: starter.status || 'claimed',
     }, person);
   }

package/lib/voice-gate.js

@@ -250,6 +250,7 @@ const LANDING_REASON_SENTENCES = {
   receipt_verifier_failed: 'a saved receipt failed its re-check, fix what it names and re-certify',
   slop_gate: 'the touched prose breaks a deterministic taste rule, fix the named lines and re-certify',
   strict_verify_missing: 'no recorded check command to re-run, add one and re-certify',
+  non_falsifying_verify: 'its recorded check cannot prove the work is correct, add a real check and re-certify',
   untagged_protected_lane_text: 'the description reads like protected-lane work without the tag, tag it or decide it yourself',
   verification_pending: 'the recorded check has not been re-run yet, the next hourly pass runs it',
   verifier_is_builder: 'the re-check actor built this row, another actor must re-check',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "atris",
-  "version": "3.74.20",
+  "version": "3.74.21",
   "description": "you say what you want in plain words. atris builds it, checks it, and shows you proof.",
   "mcpName": "io.github.atrislabs/atris",
   "main": "bin/atris.js",