chore: release v1.21.0-beta.2

[Soju06]

Summary

• Prepares beta release v1.21.0-beta.2 for the 1.21.0 stable train.
• Updates release-managed version files only; release-please remains the stable release owner.
• Merging this PR will publish v1.21.0-beta.2 only after the release-candidate validation checklist below is completed for this exact candidate SHA.
• Synced automatically from release-please PR chore(main): release 1.21.0 #1087; no manual beta workflow dispatch is required.

Changes since v1.21.0-beta.1

• fix(proxy): preserve multi-agent v2 passthrough (fix(proxy): preserve multi-agent v2 passthrough #1189) (9bbc5cd)
• fix(proxy): stop synthesizing top-level tools for upstream Codex requests (fix(proxy): stop synthesizing top-level tools for upstream Codex requests #1200) (
  fc57c9)

Release-candidate validation

Validated candidate: aec536a

Complete these checks on the exact candidate SHA above before merging. The beta release guard and publish workflow fail while any item remains unchecked or the SHA is stale.

• Backend/unit/integration gates
• Frontend tests/build
• Wheel/package validation
• Docker/container smoke
• Live upstream/account smoke
• Live upstream/account smoke not required

Validation evidence (2026-07-11, candidate aec536a):

• CI matrix green on the candidate SHA (31 checks; only the guard pending this checklist).
• Wheel codex_lb-1.21.0b2 built (make package), fresh-venv install, codex-lb --help OK, metadata version 1.21.0b2, frontend assets verified; verify_release_version --require-channel beta → tag v1.21.0-beta.2.
• Docker image built at the SHA; container GET /health/live 200, dashboard root 200, zero tracebacks; container/image cleaned up.
• uv run pytest tests/unit -q at the SHA: 3292 passed, 3 skipped; ruff + ty (frozen) clean.
• Live upstream/account smoke PERFORMED with real Codex CLI v0.144.1 + a real ChatGPT account against a scratch instance running the identical code tree (main @ fc57c9c; the candidate differs only by release version files): gpt-5.6-sol catalog fetch (model_messages present), shell tool execution, and collaboration.spawn_agent multi-agent v2 flow all succeeded — 8/8 upstream requests success, zero 400s. This directly re-verifies the beta.1 field regressions (bug: proxy synthesizes top-level tools:[] breaking GPT-5.6 reserved tools (collaboration.spawn_agent 400) #1184, bug: multi-agent v2 passthrough gaps — subagent metadata dropped on HTTP bridge, compact affinity ignores turn-state, spawn dedupe namespace-blind #1185).

Install after publish

uvx --from "codex-lb==1.21.0b2" codex-lb
docker pull ghcr.io/Soju06/codex-lb:1.21.0-beta.2
helm install codex-lb oci://ghcr.io/soju06/charts/codex-lb --version 1.21.0-beta.2 --devel

Promotion

Merge the normal release-please PR to promote this beta-tested train to 1.21.0.