Keep required review audits in the foreground

[gontzess]

Problem

A one-shot review can post its provisional summary, start background audit agents, then finish its main turn saying it is waiting for them. The SDK reports a successful turn, but the action has no later conversation turn to synthesize those results. The publication guard correctly rejects the still-provisional output.

An isolated diagnostic reproduced that sequence: two agents launched asynchronously with the background argument omitted, neither completed before the parent returned, and no final summary was written. Tool denials occurred but the agents recovered and continued reading; permission widening is not the demonstrated correction.

Change

Set CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1 for the Claude review step. Required subagent work returns in the foreground before the parent can finish. Subagents remain available; model, effort, prompts, tool allow-list, permissions and review limits are unchanged.

The provisional, freshness, head-binding and publication guards remain intact. No blind retry, prior-finding cap, raw public logging or approval fallback is added.

Reference: Claude Code background-task environment control.

Verification

• Existing focused suite: 134 tests passed.
• Independent source review confirmed the flag reaches the pinned SDK process.
• Tiny pinned-runtime lifecycle probe passed: two foreground agents completed and returned their results before the parent terminal result; no backgrounded agents.
• Full frozen-input replay passed with the same generated prompt, prior-finding input and source tree. Before: two background agents, zero completed before the parent ended. After: the same two general-purpose agents (background argument omitted) ran in the foreground, both returned before the parent terminal result, and the final non-provisional report plus commit-bound verdict were published.
• Both before and after used the same model, effort configuration, tool allow-list and safety gates. Permission friction remained, but the corrected run recovered and completed. This is lifecycle/publication proof, not an assertion that every review finding is correct.
• The latest replay tested exact PR head de1f60868b429bc43fbffff737afe55110433428, including the official release pin below. The 134-test count is local execution evidence.

The diagnostic proves the reproduced mechanism. Historical runs without retained raw SDK traces do not expose their exact internal tool calls.

Deployment note

The reusable workflow checks out job.workflow_sha. After merging, trigger a fresh review event so it uses the updated workflow revision; rerunning an older execution may retain the older revision. This PR does not modify consumer branches or rerun production reviews.

Official release pin

The action is now pinned to v1.0.234 at immutable commit 9171db3e57d6a3140a37ddc2ba92788584e0ead6, rather than the earlier untagged main commit. The annotated release tag resolves to that commit, which includes the previous pin and is three commits ahead.

This moves Claude Code / Agent SDK from 2.1.280 / 0.3.280 to 2.1.282 / 0.3.282. Background tasks remain disabled; the configured model, effort, prompts, permissions and review limits are unchanged. The upstream commits are the integration-test model fix and two CLI/SDK patch-version bumps.

The local 134-test suite passed again. Both live validation jobs installed Claude Code 2.1.282 and Agent SDK 0.3.282:

• Tiny lifecycle probe: two agents omitted the background argument; both started in the foreground and completed before the parent terminal result.
• Full frozen 89-finding replay: two general-purpose agents, zero background starts, two completions before parent termination; 26 parent turns. A final non-provisional report and exact-head-bound neutral COMMENT verdict were published, and publication reached completed. Baseline mode did not approve.
• Remaining limitations: 11 permission denials occurred. The model saved a condensed final report but could not append its requested “Prompt for AI agents” section. Finding counts/severity varied from the previous runtime trial; this validates lifecycle and publication, not substantive equivalence of every finding. No permissions or publication guards were weakened.

The release validation is complete; the earlier 2.1.280 proof is retained as historical evidence, not substituted for the new-runtime check.