refine autonomous mode: review gate, time limits, stopAfterApprovalBoundary

- Fix over-aggressive auto-approval: remove review/write_paper from
  autoApproveNodes; add WritePaperGateConfig with evidence bar checks
  (baseline, quantitative results, branch score, manuscript type)
- Add meetsWritePaperBar() with 3 gate checkpoints: top-of-loop pre-exec,
  recommendation path, no-recommendation path; backtrack to
  design_experiments on failure
- Add stopAfterApprovalBoundary option to runCurrentAgentWithOptions so
  autonomous controller gets per-node control between approval gates
- Overnight runtime: 8h → 24h (maxMinutes: 1440)
- Autonomous runtime: 24h → unbounded (maxMinutes: Infinity) with
  Number.isFinite() guard
- Progress reporter: add runtimePolicy, writePaperGateBlocked,
  writePaperGateBlockers to snapshot and markdown output
- TUI/CLI: overnight '24-hour limit', autonomous 'No runtime time limit'
  and 'write_paper gated by minimum evidence bar'
- 10 new tests (30 total), all passing; 881/882 full suite pass

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: lhy0718

ISSUES.md

@@ -163,3 +163,33 @@
   - Paper-pressure consolidation jumps to `review` node with force mode; if review node encounters artifacts from a different cycle context, it may produce a review that doesn't match the latest experiment
   - Stagnation detection relies on node-level notes; if nodes don't write meaningful notes, novelty detection may under-count signals
   - `evaluateBestBranch` reads artifacts at fixed paths; if the run has branched into multiple experiment directions, only the latest artifacts are evaluated
+
+---
+
+### AM-002 — Autonomous Mode Refinement: Review Gate, Time Limits, stopAfterApprovalBoundary
+- Status: IMPLEMENTED
+- Category: feature refinement
+- Validation target: `AutonomousRunController.runAutonomous()`, `AgentOrchestrator.runCurrentAgentWithOptions()`, review/write_paper gating, time-limit policy
+- Summary: Corrected over-aggressive auto-approval in autonomous mode and adjusted time limits:
+  - **Review gate**: `review` and `write_paper` removed from `autoApproveNodes`. Review is a real structural gate; write_paper is only entered when `meetsWritePaperBar()` evidence bar is met.
+  - **WritePaperGateConfig**: New config with `requireBaselineOrComparator`, `requireQuantitativeResults`, `minBranchScore`, `blockedManuscriptTypes`. On failure, backtracks to `design_experiments`.
+  - **Three gate checkpoints**: (1) top-of-loop pre-execution check for `currentNode === "write_paper"`, (2) recommendation path check when advancing from review, (3) no-recommendation path check at review/write_paper.
+  - **stopAfterApprovalBoundary**: Added `stopAfterApprovalBoundary?: boolean` to `AgentOrchestrator.runCurrentAgentWithOptions()`. Autonomous mode uses `stopAfterApprovalBoundary: true` so the runtime returns after each approval gate, giving the controller a chance to check evidence gates between nodes.
+  - **Overnight runtime**: 8h → 24h (`maxMinutes: 1440`)
+  - **Autonomous runtime**: 24h → unbounded (`maxMinutes: Infinity`), with `Number.isFinite()` guard
+  - **Progress reporter**: Added `runtimePolicy`, `writePaperGateBlocked`, `writePaperGateBlockers` to snapshot; shown in markdown status output
+  - **TUI/CLI copy**: Updated overnight banner ("24-hour limit"), autonomous banner ("No runtime time limit", "write_paper gated by minimum evidence bar")
+- Tests: 10 new tests added (30 total), all passing:
+  - Policy limits (overnight 24h, autonomous Infinity)
+  - Gate config defaults
+  - autoApproveNodes exclusions (review, write_paper)
+  - meetsWritePaperBar: passes, blocks, no-branch
+  - Gate blocks at review node (integration)
+  - Gate blocks advance recommendation from review (integration)
+  - No time_limit stop with Infinity
+- Evidence: 881/882 tests pass (10 new). Only pre-existing `zzz_noProjectRootLeak` failure.
+- Architecture insight: Two-level approval system — runtime `resolveApprovalGate()` auto-approves nodes in "minimal" mode BEFORE the controller sees them. `stopAfterApprovalBoundary: true` is the key fix that gives the controller per-node control.
+- Risks:
+  - `stopAfterApprovalBoundary: true` means each node takes one controller iteration, making the loop slower (more iterations per cycle). Acceptable for autonomous long-running mode.
+  - If `evaluateBestBranch` misreads evidence artifacts, the gate may incorrectly block or pass write_paper
+  - `minBranchScore: 5` threshold may need tuning based on real-world evidence patterns

src/core/agents/agentOrchestrator.ts

@@ -75,7 +75,7 @@ export class AgentOrchestrator {
 
   async runCurrentAgentWithOptions(
     runId: string,
-    opts?: { abortSignal?: AbortSignal }
+    opts?: { abortSignal?: AbortSignal; stopAfterApprovalBoundary?: boolean }
   ): Promise<AgentRunResponse> {
     await this.runtime.start(runId);
     const current = await this.runStore.getRun(runId);
@@ -84,7 +84,8 @@ export class AgentOrchestrator {
     }
     await this.runtime.runUntilPause(runId, {
       abortSignal: opts?.abortSignal,
-      floorNode: current.currentNode
+      floorNode: current.currentNode,
+      stopAfterApprovalBoundary: opts?.stopAfterApprovalBoundary
     });
     const run = await this.getPersistedRunOrThrow(runId);
 

src/core/agents/autonomousProgressReporter.ts

@@ -27,6 +27,12 @@ export interface AutonomousCycleSnapshot {
   evidenceGaps?: string[];
   nextUpgradeAction?: string;
   whyContinued?: string;
+  /** Runtime policy description: "24h" or "unbounded" */
+  runtimePolicy?: string;
+  /** Whether write_paper is currently blocked by the evidence gate */
+  writePaperGateBlocked?: boolean;
+  /** Specific conditions blocking write_paper entry */
+  writePaperGateBlockers?: string[];
 }
 
 // ---------------------------------------------------------------------------
@@ -63,6 +69,7 @@ export class AutonomousProgressReporter {
       : `# Autonomous Run Status — ${run.id.slice(0, 8)}\n\n` +
         `**Topic:** ${run.topic}\n` +
         `**Mode:** ${snap.mode}\n` +
+        `**Runtime policy:** ${snap.runtimePolicy || (snap.mode === "autonomous" ? "unbounded" : "24h")}\n` +
         `**Started:** ${new Date().toISOString()}\n\n` +
         `---\n\n`;
 
@@ -89,6 +96,9 @@ export class AutonomousProgressReporter {
     lines.push(`| Field | Value |`);
     lines.push(`|-------|-------|`);
     lines.push(`| Mode | ${snap.mode} |`);
+    if (snap.runtimePolicy) {
+      lines.push(`| Runtime Policy | ${snap.runtimePolicy} |`);
+    }
     lines.push(`| Current Node | ${snap.currentNode} |`);
     lines.push(`| Status | ${snap.status} |`);
     lines.push(`| Paper Status | ${snap.paperStatus} |`);
@@ -122,6 +132,11 @@ export class AutonomousProgressReporter {
       lines.push(`| Next Upgrade Action | ${snap.nextUpgradeAction} |`);
     }
 
+    // Write-paper gate status
+    if (snap.writePaperGateBlocked != null) {
+      lines.push(`| Write-Paper Gate | ${snap.writePaperGateBlocked ? "⛔ BLOCKED" : "✅ PASSED"} |`);
+    }
+
     lines.push("");
     lines.push(`**Message:** ${snap.message}`);
 
@@ -130,6 +145,15 @@ export class AutonomousProgressReporter {
       lines.push(`**Why continued:** ${snap.whyContinued}`);
     }
 
+    // Write-paper gate blockers
+    if (snap.writePaperGateBlocked && snap.writePaperGateBlockers && snap.writePaperGateBlockers.length > 0) {
+      lines.push("");
+      lines.push("**Write-paper gate blockers (conditions not met for drafting):**");
+      for (const blocker of snap.writePaperGateBlockers) {
+        lines.push(`- ${blocker}`);
+      }
+    }
+
     if (snap.noveltySignals.length > 0) {
       lines.push("");
       lines.push("**Recent Novelty Signals:**");
@@ -169,6 +193,9 @@ export class AutonomousProgressReporter {
     lines.push(`| Field | Value |`);
     lines.push(`|-------|-------|`);
     lines.push(`| Mode | ${snap.mode} |`);
+    if (snap.runtimePolicy) {
+      lines.push(`| Runtime Policy | ${snap.runtimePolicy} |`);
+    }
     lines.push(`| Total Cycles | ${snap.cycle} |`);
     lines.push(`| Total Iterations | ${snap.iteration} |`);
     lines.push(`| Final Node | ${snap.currentNode} |`);
@@ -182,6 +209,9 @@ export class AutonomousProgressReporter {
     if (snap.paperCandidateStatus) {
       lines.push(`| Paper Candidate | ${snap.paperCandidateStatus} |`);
     }
+    if (snap.writePaperGateBlocked != null) {
+      lines.push(`| Write-Paper Gate | ${snap.writePaperGateBlocked ? "⛔ BLOCKED" : "✅ PASSED"} |`);
+    }
 
     lines.push("");
     lines.push(`**Why stopped:** ${snap.message}`);