v2.2.0 — GPT-6 Astra, Daybreak Blue, and codex-cli 0.154.0 compatibility

Compatibility pass against codex-cli 0.154.0, and support for the models
that shipped with it.

Added

• GPT-6 Astra (gpt-6-astra) — OpenAI's most capable model, for complex,
  demanding work. Supports every effort up to ultra.
• Daybreak Blue (gpt-daybreak-blue-latest) — the frontier agentic coding
  model for broad defensive cybersecurity work. It is access-gated, so it
  appears in the catalog only on accounts approved for it, and the delegation
  skill now routes security reviews and hardening work to it when present.
• gpt-5.3-codex-spark is selectable again — see Fixed. It is the
  ultra-fast tier, for small, mechanical, fully specified edits.
• Each model's own description is carried through to codex_models. Routing
  advice now comes from the live catalog rather than from a table in this repo
  that goes stale whenever OpenAI ships something.
• Model aliases for /codex:delegate: astra, daybreak, and spark join
  sol, terra, and luna.

Fixed

• Out-of-quota runs were reported as unknown failures. codex-cli 0.154.0
  states it in prose — "You've hit your usage limit … or try again at
  <date>" — with no usage_limit_reached code and no HTTP 429. The
  classifier matched only the older machine-readable spellings, so a plain
  out-of-quota job raised a generic codex_error and dropped the advice that
  is the entire point of a typed rate-limit error: switch to a cheaper model,
  lower the effort, or wait for the reset. The reset time is now read from
  either wording, and quoted back with codex's own capitalization.
• A live model could be blocked by a stale constant. validate consulted
  the built-in deny-list before the catalog, so gpt-5.3-codex-spark was
  unreachable through this server — rejected as "not available on this
  account" while codex debug models listed it as available. The live catalog
  now outranks both DEPRECATED_MODELS and ALIAS_HINTS; those tables are
  consulted only for slugs the catalog does not know. This is the same rot the
  live catalog was introduced to solve, arriving from the other direction.
• A standalone error event is now read. Codex usually emits one alongside
  turn.failed, but not always. When it was the only structured account of a
  failure, classification fell through to stderr — a stream shared with every
  MCP server in the user's codex config, where a bystander's routine HTTP 401
  reads exactly like an expired session. A real turn.failed still outranks it.

Changed

• The fallback catalog is refreshed to the 0.154.0 generation. gpt-5.4 and
  gpt-5.4-mini have been retired from the ChatGPT-account catalog and now
  fail with a pointer to a live model instead of a bare HTTP 400.
• 323 tests, up from 292.

Verified against codex-cli 0.154.0

Re-checked rather than assumed, since these are the assumptions the server is
built on. Unchanged: codex exec still accepts --sandbox, --color,
-C/--cd, --skip-git-repo-check, --json, -o, and --output-schema;
codex exec resume still rejects --sandbox, --color, and -C/--cd, so
sandbox on a resume still goes through -c sandbox_mode=; -c values are
still parsed as TOML; and the --json event stream still emits
thread.started, turn.completed, turn.failed, and item.completed with
the item types the progress phases key off. A live end-to-end codex exec run
confirmed the stream shape rather than the help text alone.