fix(#1117): allow claude.exe binary in Vertex AI profile

[fullsend-ai-coder]

Summary

• Add **/claude.exe to the binaries allowlist in profiles/fullsend-vertex-ai.yaml
• Claude Code 2.1+ ships as a Bun-compiled binary named claude.exe (even on Linux), but the existing **/claude glob does not match it — OPA blocks the binary from reaching sts.googleapis.com, preventing Vertex AI authentication entirely
• A downstream fix was already merged for one affected repo (fix(fullsend): allow Claude Code Vertex authentication rhdh-parasol/rhdh-plugin-export-overlays#13); this updates the upstream profile to protect all repos using the default profile

Testing

• Verified YAML validity of the modified file
• Ran skillsaw --strict lint — all checks passed
• Secret scan passed on staged content

---

Closes #1117

Post-script verification

• Branch is not main/master (agent/1117-allow-claude-exe-binary)
• Secret scan passed (gitleaks — c4e059f9f2b8c48e13b3845992b9b774d41b5de7..HEAD)
• PR body secret scan passed (gitleaks — no-git)

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 2:31 PM UTC · Completed 2:43 PM UTC

Commit: ba1d46f · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $2.59

[fullsend-ai-review]

Risk Assessment: low (1/5)

Details

A single-line addition of a binary glob to a YAML allowlist by a trusted bot, with a well-triaged issue and exact scope match, presents minimal risk.

[fullsend-ai-review]

Review

Findings

Medium

• [protected-path] profiles/fullsend-vertex-ai.yaml — This PR modifies profiles/fullsend-vertex-ai.yaml, which is under the profiles/ protected path. The PR links to issue Review agent outage: opus alias resolves to policy-denied claude-opus-5 on Vertex AI #1117 and the description clearly explains the rationale (Claude Code 2.1+ ships as a Bun-compiled binary named claude.exe, but the existing **/claude glob does not match it, causing OPA to block Vertex AI authentication). Human approval is always required for protected-path changes, regardless of context.

Low

• [wildcard-usage] profiles/fullsend-vertex-ai.yaml:30 — The **/claude.exe glob uses a recursive wildcard prefix (**), matching the binary at any filesystem path. This is consistent with all other entries in the binaries list (**/claude, **/pi, **/node) and introduces no additional risk beyond the established pattern.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 7:26 AM UTC · Completed 7:33 AM UTC

Commit: ba1d46f · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $1.60

[fullsend-ai-retro]

Retro: PR #1118 — allow claude.exe binary in Vertex AI profile

Workflow quality: good. After a human corrected the original retro agent's misdiagnosis, the triage → code → review pipeline executed cleanly. The code agent made a minimal 1-line fix, and the review agent produced accurate findings with no false positives. Total agent cost: $5.22 (triage $0.95, code $1.68, review $2.59). Human approved within 3 minutes of PR creation.

Timeline

1. Aug 27–31 — Claude Code 2.1+ (Bun-compiled claude.exe) caused review agent failures on redhat-developer/rhdh-plugin-export-overlays. OPA blocked claude.exe from reaching sts.googleapis.com because the binary allowlist only had **/claude.
2. Aug 31 — Retro agent filed #1117 with an incorrect root cause: claimed the opus alias resolved to claude-opus-5 which Vertex AI denied, with high stated confidence. The actual cause was the OPA binary allowlist.
3. Sep 1 14:13 — rh-hemartin corrected the diagnosis with a detailed analysis of OPA denial logs showing the binary path claude.exe → sts.googleapis.com.
4. Sep 1 14:17–14:29 — Triage (run 33518346639) incorporated the correction. Code agent (run 33519101251) created PR fix(#1117): allow claude.exe binary in Vertex AI profile #1118 with the 1-line fix.
5. Sep 1 14:31–14:43 — Review agent (run 33519846209) correctly flagged profiles/ as a protected path, assessed risk as low (1/5).
6. Sep 2 07:24 — Merged by rh-hemartin.

Key finding: retro agent misdiagnosed the root cause

The retro agent that filed #1117 attributed policy_denied errors to Vertex AI model authorization, but the errors came from OPA's binary sandbox. No traffic ever reached Vertex AI. The retro's proposed fix (enable claude-opus-5 on the Vertex project) would not have resolved the outage. A human had to provide the correct diagnosis before agents could proceed.

Existing issue evidence

• #812 (pre-retro early-exit for retro-originated issues): This retro run is itself analyzing a PR that closes retro-originated issue Review agent outage: opus alias resolves to policy-denied claude-opus-5 on Vertex AI #1117 — exactly the low-value scenario Pre-retro should early-exit for PRs that close retro-originated issues #812 proposes skipping.
• #406 (decompose multi-layer failure modes): The retro agent's misdiagnosis is a related but distinct failure — rather than conflating layers of a single error chain, it misidentified the error source entirely (OPA sandbox vs. Vertex AI model auth). New proposal filed below to address this gap.

Proposals filed

• Retro agent should verify error-source attribution against log evidence before committing to a root cause (in fullsend-ai/agents)