@miller-tech/uap 2.11.1 → 2.16.2

package/dist/cli/deliver.js

@@ -278,6 +278,8 @@ import { createEscalationController, defaultEscalationLadder, DEFAULT_STAGNATION
 import { FilePracticeStore, defaultPracticePath, extractKeywords, retrievePracticesSemantic, } from '../delivery/practice.js';
 import { detectRungs, mergeRedetectedRungs, runLadder, runTieredLadder, tierOf, TIER_ORDER, demoteBaselineFailures, baselineRegressions } from '../delivery/verifier-ladder.js';
 import { recordGateEvidence, gateOutcomesFromResult } from '../delivery/gate-evidence.js';
+import { resolveAdversarialGate, runAdversarialGate } from '../delivery/adversarial-gate.js';
+import { resolveMasterPipeline, formatMasterPipelineLine } from '../delivery/master-pipeline.js';
 import { runDeployDevLadder } from '../delivery/deploy-dev-gate.js';
 import { commitPushAndWatch } from '../delivery/ci-watcher.js';
 import { snapshotTree, restoreTreeScoped, disposeSnapshot } from '../delivery/snapshot.js';
@@ -1495,6 +1497,25 @@ async function runDeliver(instruction, options) {
             console.log(chalk.cyan('⚖ acceptance judge on (verification rail; UAP_DELIVER_ACCEPTANCE=0 to disable)'));
         }
     }
+    // Master-pipeline stage readout — the Harness/Loop/Graph composition this
+    // command runs by default: worktree-isolated parallel fan-out → evidence-
+    // gated convergence loops → state-hash read dedup → adversarial red-team
+    // gate. Resolved once via delivery/master-pipeline.ts (which mirrors each
+    // stage's own env > .uap.json > default precedence) and surfaced in the
+    // startup banner and the dry-run plan, so what is printed is what the
+    // run gets.
+    const deliverCfgEarly = (() => {
+        try {
+            return (loadUapConfigRaw(projectRoot) ?? {}).deliver;
+        }
+        catch {
+            return undefined;
+        }
+    })();
+    const masterPipeline = resolveMasterPipeline(deliverCfgEarly);
+    if (!options.dryRun) {
+        console.log(chalk.cyan(`⇒ master pipeline: ${formatMasterPipelineLine(masterPipeline)}`));
+    }
     // `--optimize` turns on every convergence aid at once. Deploy queueing is
     // deliberately excluded — committing applied files is a side effect the
     // user must opt into explicitly.
@@ -1880,6 +1901,12 @@ async function runDeliver(instruction, options) {
             branch: watchCi ? currentBranch(projectRoot) : null,
             gates: rungs.map((r) => ({ id: r.id, name: r.name, required: r.required, tier: tierOf(r) })),
             selfGate: needsSelfGate,
+            masterPipeline: {
+                parallelTasks: masterPipeline.parallelTasks,
+                stateHash: masterPipeline.stateHash,
+                stateHashMinBytes: masterPipeline.stateHashMinBytes,
+                adversarialGate: masterPipeline.adversarial.enabled ? masterPipeline.adversarial.maxRounds : 0,
+            },
         };
         if (options.json) {
             console.log(JSON.stringify(summary, null, 2));
@@ -1898,6 +1925,7 @@ async function runDeliver(instruction, options) {
             console.log(`  HALO tracing: ${summary.halo ? 'on' : 'off'}`);
             console.log(`  Coordination: ${summary.coordinate ? 'on' : 'off'}`);
             console.log(`  Deploy queue on success: ${summary.deploy ? 'on' : 'off'}`);
+            console.log(`  Master pipeline: ${formatMasterPipelineLine(masterPipeline)}`);
             console.log(`  Test protection: ${summary.protectTests ? `on (${summary.protectedTestFiles} pre-existing test/oracle file(s))` : 'off'}`);
             console.log(`  IaC protection: ${summary.protectIac ? 'on (Dockerfile/compose/terraform writes blocked)' : 'off (permissive — IaC writes allowed)'}`);
             console.log(`  Guidance file: ${summary.guidanceFile ?? 'none (mission runs unattended)'}`);
@@ -2139,6 +2167,12 @@ async function runDeliver(instruction, options) {
     // sweep record real written paths here; the applier below drains them into
     // ApplyResult.filesWritten so history[].filesApplied is truthful again.
     const agenticWriteLedger = new Set();
+    // Hoisted shared protection set (HLG stage 5 hardening): the adversarial
+    // gate's breaching tests are written AFTER this t0 snapshot, so a repair
+    // could delete them unless the set grows. Both executors hold THIS
+    // reference and check membership per write, so the stage's onBreach hook
+    // can extend it before reconvergence runs.
+    const sharedProtectedFiles = options.protectTests !== false ? snapshotProtection(projectRoot).protectedFiles : new Set();
     const executor = agentic
         ? createAgenticExecutor(model, {
             projectRoot,
@@ -2149,7 +2183,7 @@ async function runDeliver(instruction, options) {
             contractFiles: contractLock,
             writeLedger: agenticWriteLedger,
             // Block oracle tampering: protected test files are read-only to the agent.
-            protectedFiles: options.protectTests !== false ? snapshotProtection(projectRoot).protectedFiles : new Set(),
+            protectedFiles: sharedProtectedFiles,
             // Block gate-config / IaC rigging in the agentic path too (it bypasses
             // the file-block applier where this protection otherwise lives).
             protectGateConfigs: options.protectTests !== false,
@@ -2171,7 +2205,7 @@ async function runDeliver(instruction, options) {
             contextTokenBudget: sessionBudget,
             contractFiles: contractLock,
             writeLedger: agenticWriteLedger,
-            protectedFiles: options.protectTests !== false ? snapshotProtection(projectRoot).protectedFiles : new Set(),
+            protectedFiles: sharedProtectedFiles,
             protectGateConfigs: options.protectTests !== false,
             allowBash: options.allowBash === true || process.env.UAP_DELIVER_ALLOW_BASH === '1',
             onEvent: (e) => console.log(chalk.dim(`    [repair r${e.round} ${e.kind}${e.tool ? `:${e.tool}` : ''}] ${e.detail ?? ''}`)),
@@ -3847,6 +3881,81 @@ async function runDeliver(instruction, options) {
             disposeSnapshot(regressSnapshot);
         }
     }
+    // Adversarial red-team gate (HLG stage 5): before the verdict is accepted,
+    // a skeptical verifier attacks the converged patch with edge-case tests
+    // designed to break it. A breach routes back into the convergence loop as
+    // ordinary gate evidence (bounded rounds); only a surviving patch proceeds
+    // to the outcome/evidence bookkeeping below. Placed AFTER keep-best so the
+    // attack targets the same tree the rollback rail settled on, and BEFORE
+    // evidence recording so the artifact reflects the post-attack state.
+    // Skipped on alreadyDelivered (no patch to attack) and on --dry-run.
+    // Default ON; escape hatch: UAP_DELIVER_ADVERSARIAL_GATE=0 or
+    // `.uap.json` deliver.adversarialGate=false — same path deliver.parallelTasks
+    // resolves through.
+    const adversarialSettings = resolveAdversarialGate(cfgRaw.deliver?.adversarialGate);
+    if (result.success && !result.alreadyDelivered && adversarialSettings.enabled && !options.dryRun) {
+        const adversarial = await runAdversarialGate({
+            instruction,
+            projectRoot,
+            rungs,
+            // Gate-time redetection: `rungs` was detected at t0, so a mission that
+            // CREATED its test gate (greenfield) would no-surface forever. Merge
+            // the fresh detection with the same policy the loop's mid-mission
+            // redetection and the combined-tree verification use.
+            redetectRungs: () => {
+                try {
+                    return mergeRedetectedRungs(rungs, detectRungs(projectRoot), loopConfig.redetectFilter);
+                }
+                catch {
+                    return rungs; // detection unavailable — attack with the t0 rungs
+                }
+            },
+            // Defense in depth for the repair: the executors' protected set was
+            // snapshotted before the breaching tests existed — add them so the
+            // repair agent structurally CANNOT delete/skip/weaken them (the gate's
+            // own post-repair verification remains the enforcing floor).
+            onBreach: (breach) => {
+                for (const rel of breach.testFiles)
+                    sharedProtectedFiles.add(rel.toLowerCase());
+            },
+            // Authoring judges text; keep it on the blind executor even when the
+            // loop's executor is agentic (same rule as the critic/judge).
+            executor: jsonBlindExecutor,
+            settings: adversarialSettings,
+            initial: result,
+            reconverge: (prompt) => new ConvergenceLoop({ ...loopConfig, baselineCheck: false, resumeFrom: undefined, onCheckpoint: undefined }, executor, seams).deliver(prompt),
+            note: (line) => console.log(chalk.dim(`  ${line}`)),
+        });
+        result = adversarial.result;
+        result.adversarial = {
+            status: adversarial.status,
+            rounds: adversarial.rounds.length,
+            authored: adversarial.authored,
+            ran: adversarial.ran,
+            failed: adversarial.failed,
+            summary: adversarial.summary,
+        };
+        if (!options.json) {
+            const color = adversarial.status === 'breached'
+                ? chalk.red
+                : adversarial.status === 'survived'
+                    ? chalk.green
+                    : chalk.dim;
+            console.log(color(`  🛡 ${adversarial.summary}`));
+        }
+    }
+    else if (result.success && !result.alreadyDelivered && !adversarialSettings.enabled) {
+        // Record the hatch on the result so the JSON outcome and evidence say the
+        // stage was deliberately off, not silently absent.
+        result.adversarial = {
+            status: 'disabled',
+            rounds: 0,
+            authored: 0,
+            ran: 0,
+            failed: 0,
+            summary: 'adversarial gate: disabled (UAP_DELIVER_ADVERSARIAL_GATE=0 or deliver.adversarialGate=false)',
+        };
+    }
     // Surface the observed stop on the result itself: the JSON output (and the
     // MCP deliver tool reading it) is what the NEXT agent decides from, and a
     // stopped run that prints as a plain failure gets relaunched instead of
@@ -3891,6 +4000,20 @@ async function runDeliver(instruction, options) {
     if (result.success) {
         try {
             const gates = gateOutcomesFromResult(result, rungs);
+            // HLG stage 5 telemetry: bind the adversarial attack outcome into the
+            // same artifact as the run's other gate evidence, so evidence consumers
+            // see what the red team did (tests authored / ran / failed / survived).
+            // 'disabled' is recorded too — an absent entry must never be ambiguous
+            // with a skipped one.
+            if (result.adversarial) {
+                gates.push({
+                    name: 'adversarial-gate',
+                    command: 'uap deliver adversarial red-team gate',
+                    exitCode: 0,
+                    outputTail: result.adversarial.summary,
+                    at: new Date().toISOString(),
+                });
+            }
             const evidencePath = recordGateEvidence(projectRoot, gates, { runId });
             if (!options.json)
                 console.log(chalk.dim(`  gate evidence: ${evidencePath}`));
@@ -3899,6 +4022,29 @@ async function runDeliver(instruction, options) {
             console.warn(chalk.yellow(`  ⚠ gate evidence not recorded: ${e.message}`));
         }
     }
+    else if (result.adversarial) {
+        // Breached runs record their adversarial evidence TOO (C-stage review):
+        // the stage ran and REJECTED the patch, so an absent entry would be
+        // ambiguous with "stage skipped". Nonzero exitCode — the enforcer only
+        // accepts 0, so this artifact can never masquerade as a pass. Fail-soft,
+        // same as the success path (a dirty post-breach tree simply warns).
+        try {
+            const evidencePath = recordGateEvidence(projectRoot, [
+                {
+                    name: 'adversarial-gate',
+                    command: 'uap deliver adversarial red-team gate',
+                    exitCode: 1,
+                    outputTail: result.adversarial.summary,
+                    at: new Date().toISOString(),
+                },
+            ], { runId });
+            if (!options.json)
+                console.log(chalk.dim(`  gate evidence: ${evidencePath}`));
+        }
+        catch (e) {
+            console.warn(chalk.yellow(`  ⚠ gate evidence not recorded: ${e.message}`));
+        }
+    }
     // Close the HALO observe→learn loop: mine the trace tail for recurring
     // failure patterns after every run — no operator step required.
     if (isHaloTracingEnabled()) {