Meta: small UX patches to make `codex exec` cleaner for downstream harnesses (cdx exo)

[team-wcv]

Context

A downstream Codex wrapper called cdx exo (https://github.com/team-wcv/CliTools, exo-tui.mjs and friends) embeds a custom Ink/React TUI that drives codex exec and codex app-server over JSON-RPC for cluster-aware local model serving. Building it surfaced a handful of small "papercut" issues in the upstream codex harness — things the TUI currently works around with stderr regexes or ad-hoc client-side filtering, but which would be cleaner to fix at the source so other harnesses benefit.

This meta-issue tracks the seven patches we extracted from the cdx exo workarounds and filed individually. Each is small, scoped, and most ship with a reference branch on a fork. None depend on each other; they can be triaged independently.

Patches

#	Title	Issue	Reference branch	Status
1	Add codex exec --quiet to drop the header chrome	#22047	team-wcv/codex@feat/exec-quiet-flag	Closed not_planned; maintainer suggested --json instead. We replied with the wrapper-vs-parser context; happy to drop or revisit.
2	Gate benign rate-limit telemetry behind notifications.rate_limit_telemetry	#22054	team-wcv/codex@feat/rate-limit-telemetry-gate	Open, awaiting triage.
3	Demote failed to record rollout items from error! to warn!	#22055	team-wcv/codex@feat/rollout-persistence-log-demote	Open, awaiting triage.
4	Collapse MCP startup notifications into a single mcpServer/startupStatus/completed summary	#22059	team-wcv/codex@feat/mcp-startup-notifications-collapse	Open, awaiting triage.
5	Local-provider reasoning-summary heuristic + per-model model_overrides	#22061	team-wcv/codex@feat/local-provider-reasoning-summaries	Open, awaiting triage.
6	model_providers.<name>.discovery_url for cluster-aware base-URL refresh	#22063	team-wcv/codex@feat/provider-discovery-url	Open, awaiting triage.
7	codex exec resume <missing-uuid> falls back to thread/start instead of erroring	#22064	team-wcv/codex@feat/thread-resume-fallback	Open, awaiting triage.

Why each one matters (one line each)

• Add codex exec --quiet to suppress the config summary banner for parent harnesses #22047 — wrappers that pass-through human-readable codex output need to suppress the config banner. Closed; documented the wrapper use case for context.
• Gate benign account/rateLimits/updated telemetry behind notifications.rate_limit_telemetry #22054 — rate-limit telemetry is benign for users that never hit a limit; routes a notification per turn and shows up as noise.
• Demote failed to record rollout items from error to warn so harness stderr stops looking like a fatal failure #22055 — transient rollout-persistence races during shutdown spam scary ERROR lines that aren't actionable.
• Collapse MCP startup notifications into a single aggregate readiness summary #22059 — N MCP servers ⇒ N+ startup notifications during steady-state startup; one aggregate readiness summary is enough.
• Make reasoning summaries opt-in via a flag (with per-model override) for self-hosted reasoning models #22061 — local providers that ship reasoning models (DeepSeek-R1, QwQ, etc.) silently drop reasoning-summary support unless users hand-edit a global flag.
• Add model_providers.<name>.discovery_url for cluster-aware base-URL refresh #22063 — multi-host inference clusters need to refresh the active node's base URL without restarting codex.
• codex exec resume <missing-uuid> surfaces noisy "thread <uuid> not found"; fall back to thread/start instead #22064 — codex exec resume <uuid> after a pruned rollout / wrong \$CODEX_HOME surfaces a confusing error string instead of just starting a new thread.

Bigger picture

The cdx exo TUI's existence is mostly about cluster-aware local model serving and inline UX shortcuts that aren't a fit for upstream codex (e.g. live cluster topology, model-catalog overlay, in-process status badges). What I'd like out of this batch is just to remove the surface-level noise and config papercuts that force harnesses to filter stderr or hand-edit configs.

Happy to:

• split anything that's too coarse-grained.
• drop patches you don't want.
• iterate on any of the reference branches based on feedback.
• coordinate landing if multiple patches touch shared protocol/config surface.

I noticed docs/contributing.md says external contributions are by invitation; that's why each patch is in an issue with a reference branch rather than a PR. If any of these are wanted, very happy to open the corresponding PR(s).