chore(fullsend): allowlist claude.exe for Vertex STS

[durandom]

Summary

• Review on Skip or use lightweight retro analysis for bot-authored PRs fullsend-ai/fullsend#3317 failed because OpenShell denied /usr/lib/node_modules/@anthropic-ai/claude-code/bin/claude.exe → sts.googleapis.com:443 (policy vertex_ai). **/claude does not match claude.exe.
• Allowlist **/claude.exe on code, e2e-triage, and a forked review policy; point the review harness at the local policy.
• Pin the review harness to ghcr.io/redhat-developer/rhdh-fullsend-code:latest (code/fix/e2e already use this image). Claude Code 2.1.243 in that image is what needs claude.exe.

Proven

Same image + claude.exe allowlist succeeded on rhdh-agentic fullsend-ai/fullsend#124 / fullsend-ai/fullsend#125:

• https://github.com/redhat-developer/rhdh-agentic/pull/124
• https://github.com/redhat-developer/rhdh-agentic/pull/125
• https://github.com/redhat-developer/rhdh-agentic/actions/runs/33403845734
  • Image ghcr.io/redhat-developer/rhdh-fullsend-code:latest
  • Claude v2.1.243
  • no policy_denied

Test plan

• Re-run fullsend review on a PR (e.g. re-trigger on Skip or use lightweight retro analysis for bot-authored PRs fullsend-ai/fullsend#3317 or open a trivial PR)
• Job log shows Vertex STS succeeding (no policy_denied / claude.exe DENIED)
• Review agent posts a review comment

Evidence: https://github.com/redhat-developer/rhdh-plugin-export-overlays/actions/runs/33395911446/job/99500322474

[fullsend-ai-review]

🤖 Finished Review · ❌ Failure · Started 2:10 PM UTC · Completed 2:12 PM UTC

Commit: 993532d · View workflow run →

Runtime: claude · Model: opus → claude-opus-5

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[fullsend-ai-review]

🤖 Finished Review · ❌ Failure · Started 2:51 PM UTC · Completed 2:53 PM UTC

Commit: 7139c94 · View workflow run →

Runtime: claude · Model: opus → claude-opus-5

[kadel]

fullsend only changs, no need to wait for CI

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 3:06 PM UTC · Completed 3:19 PM UTC

Commit: 7139c94 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $6.88

[fullsend-ai-retro]

Retro: PR fullsend-ai/fullsend#3480 — allowlist claude.exe for Vertex STS

What happened

PR fullsend-ai/fullsend#3480 was a human-authored infrastructure fix (co-authored with Cursor) that addressed a sandbox policy gap: Claude Code 2.1.243+ ships its binary as claude.exe (bun compile) instead of claude, so the existing **/claude glob in sandbox network policies did not match it. This caused policy_denied errors when the binary tried to reach sts.googleapis.com:443 for Vertex AI STS token exchange.

The PR added **/claude.exe to the code.yaml and e2e-triage.yaml policies, created a new local review.yaml policy fork with the same fix, and pinned the review harness to ghcr.io/redhat-developer/rhdh-fullsend-code:latest.

Both review agent runs failed (run 1, run 2) — expected, since the review agent checked out the base branch config (which didn't yet include the fix). This is a chicken-and-egg situation: the PR fixing the review agent's sandbox policy cannot itself be reviewed by the review agent.

A human reviewer (kadel) approved and merged within ~1 hour, correctly noting these were fullsend-only changes that didn't need CI.

Key finding: upstream fix is missing

The fix only lives in this repo's local .fullsend/ policy forks. The upstream agents repo (fullsend-ai/agents) still does not have **/claude.exe in its Vertex AI profile (profiles/fullsend-vertex-ai.yaml). Any repo relying on the default fleet profile (without a local policy override) is susceptible to the same policy_denied failure when running Claude Code 2.1.243+.

Existing issue coverage

• fullsend-ai/fullsend#6795 covers the broader review agent outage on Vertex AI but attributes it to model alias resolution (opus → claude-opus-5), not the sandbox binary name mismatch. These are distinct issues: Review agent outage: opus alias resolves to policy-denied claude-opus-5 on Vertex AI fullsend-ai/agents#1117 is about model availability on the GCP project, while this PR addresses the sandbox-level binary allowlist. Both may have contributed to the failures on PR Skip or use lightweight retro analysis for bot-authored PRs fullsend-ai/fullsend#3317.
• No existing issue in fullsend-ai/agents or fullsend-ai/fullsend covers the claude.exe binary name gap in sandbox policies.

Proposals skipped (target repo not allowed)

File manually or update create_issues.allow_targets in config.yaml:

• Add **/claude.exe to Vertex AI profile binary allowlist (fullsend-ai/agents)