Skip to content

chore(fullsend): use rhdh-fullsend-code and allowlist claude.exe - #4496

Merged
kadel merged 1 commit into
redhat-developer:mainfrom
durandom:chore/fullsend-rhdh-image-claude-exe
Aug 31, 2026
Merged

kadel merged 1 commit into
redhat-developer:mainfrom
durandom:chore/fullsend-rhdh-image-claude-exe

Conversation

@durandom

Copy link
Copy Markdown
Member

Summary

  • Pin code/fix/review harnesses to ghcr.io/redhat-developer/rhdh-fullsend-code:latest.
  • Allowlist claude.exe (Claude Code 2.1+ bun compile) so Vertex STS is not policy_denied.
  • Add a review harness/policy so review agents use the same image and binary allowlist.

Proven

Same fix as rhdh-agentic #124 / #125:

Why now

rhdh-plugins code agents are already failing with policy_denied on:
https://github.com/redhat-developer/rhdh-plugins/actions/runs/33389059456

**/claude does not match claude.exe, which is what Claude Code 2.1+ uses to call sts.googleapis.com for Vertex.

Made with Cursor

Pin code/fix/review to ghcr.io/redhat-developer/rhdh-fullsend-code:latest and allow Vertex STS from Claude Code's bun binary (claude.exe).

Co-authored-by: Cursor <cursoragent@cursor.com>
@durandom
durandom requested review from a team as code owners August 31, 2026 14:50
@sonarqubecloud

Copy link
Copy Markdown

@kadel

kadel commented Aug 31, 2026

Copy link
Copy Markdown
Member

/lgtm

@openshift-ci openshift-ci Bot added the lgtm label Aug 31, 2026
@kadel
kadel merged commit 4494c96 into redhat-developer:main Aug 31, 2026
21 checks passed
subhashkhileri added a commit to subhashkhileri/rhdh-plugins that referenced this pull request Sep 17, 2026
Our code/fix/review harnesses pinned fullsend-ai/agents@4bbe4f50e, set
once in redhat-developer#4496 and never bumped — 794 commits behind main. Move to the
agents@v0 tag (7ab05564b, 794 commits later) to pick up two fixes
relevant to the /fs-review timeout on PR redhat-developer#4808 (run 35097265739):

- fullsend-ai/agents#1116 (PR redhat-developer#1181): sonnet-tier review sub-agents now
  request the bare `sonnet` alias, pinned via
  ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-4-6 in env/gcp-vertex.env,
  instead of the hardcoded `claude-sonnet-4-6@default` literal our pin
  still carries. That literal was falling through the CLI's own alias
  resolution to claude-sonnet-4 (Sonnet 4.5 not available on our Vertex
  project), landing on a base model with a 120k output-tokens/min quota
  vs. 300k for sonnet-4-6 — three parallel sonnet sub-agents blew
  through it and all failed with 429 RESOURCE_EXHAUSTED.
- review harness timeout_minutes raised 20 -> 45, which independently
  would have absorbed the degraded run instead of hard-timing-out.

Scope: base: URL + content sha256 only. Our rhdh/policies/*.yaml
overrides are left as-is for now — upstream moved network policy from
policy: to composable providers:/openshell: profiles: in the interim
(ADR 0065), which likely makes the claude.exe allowlist in our policy
fork redundant (profiles/fullsend-vertex-ai.yaml already allowlists it
upstream), but retiring that fork is deferred to a follow-up so this
change stays a minimal, low-risk version bump.
subhashkhileri added a commit to subhashkhileri/rhdh-plugins that referenced this pull request Sep 17, 2026
Our code/fix/review harnesses pinned fullsend-ai/agents@4bbe4f50e, set
once in redhat-developer#4496 and never bumped — 794 commits behind main. Move to the
agents@v0 tag (7ab05564b, 794 commits later) to pick up two fixes
relevant to the /fs-review timeout on PR redhat-developer#4808 (run 35097265739):

- fullsend-ai/agents#1116 (PR redhat-developer#1181): sonnet-tier review sub-agents now
  request the bare `sonnet` alias, pinned via
  ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-4-6 in env/gcp-vertex.env,
  instead of the hardcoded `claude-sonnet-4-6@default` literal our pin
  still carries. That literal was falling through the CLI's own alias
  resolution to claude-sonnet-4 (Sonnet 4.5 not available on our Vertex
  project), landing on a base model with a 120k output-tokens/min quota
  vs. 300k for sonnet-4-6 — three parallel sonnet sub-agents blew
  through it and all failed with 429 RESOURCE_EXHAUSTED.
- review harness timeout_minutes raised 20 -> 45, which independently
  would have absorbed the degraded run instead of hard-timing-out.

Scope: base: URL + content sha256 only. Our rhdh/policies/*.yaml
overrides are left as-is for now — upstream moved network policy from
policy: to composable providers:/openshell: profiles: in the interim
(ADR 0065), which likely makes the claude.exe allowlist in our policy
fork redundant (profiles/fullsend-vertex-ai.yaml already allowlists it
upstream), but retiring that fork is deferred to a follow-up so this
change stays a minimal, low-risk version bump.
durandom pushed a commit that referenced this pull request Sep 17, 2026
…4843)

Our code/fix/review harnesses pinned fullsend-ai/agents@4bbe4f50e, set
once in #4496 and never bumped — 794 commits behind main. Move to the
agents@v0 tag (7ab05564b, 794 commits later) to pick up two fixes
relevant to the /fs-review timeout on PR #4808 (run 35097265739):

- fullsend-ai/agents#1116 (PR #1181): sonnet-tier review sub-agents now
  request the bare `sonnet` alias, pinned via
  ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-4-6 in env/gcp-vertex.env,
  instead of the hardcoded `claude-sonnet-4-6@default` literal our pin
  still carries. That literal was falling through the CLI's own alias
  resolution to claude-sonnet-4 (Sonnet 4.5 not available on our Vertex
  project), landing on a base model with a 120k output-tokens/min quota
  vs. 300k for sonnet-4-6 — three parallel sonnet sub-agents blew
  through it and all failed with 429 RESOURCE_EXHAUSTED.
- review harness timeout_minutes raised 20 -> 45, which independently
  would have absorbed the degraded run instead of hard-timing-out.

Scope: base: URL + content sha256 only. Our rhdh/policies/*.yaml
overrides are left as-is for now — upstream moved network policy from
policy: to composable providers:/openshell: profiles: in the interim
(ADR 0065), which likely makes the claude.exe allowlist in our policy
fork redundant (profiles/fullsend-vertex-ai.yaml already allowlists it
upstream), but retiring that fork is deferred to a follow-up so this
change stays a minimal, low-risk version bump.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants