omnigateway 0.13.7 → 0.14.1

package/README.md

@@ -174,6 +174,18 @@ pair is unlimited. Windows slide, `tokens` and `spend` are debited on
 completion, and limits are editable afterwards with `omni keys limits`; the
 semantics are in [docs/operations.md](docs/operations.md#key-limits).
 
+One model can be held to a stricter ceiling than the rest of the key's traffic
+with `--model-limit <model>:<dimension>:<window>=<value>` (or
+`<model>:concurrency=<value>`) — `omni keys create --label ci --limit
+requests:1m=600 --model-limit opus:requests:1m=20`. The model is the one the
+request names, matched exactly; an override stacks with the key-wide limits, so
+a request for that model must pass both. `omni keys limits <id>
+--unset-model-limit opus:requests:1m` removes a slot, and the last slot takes
+the override with it. **Rolling back** to a build without per-model limits
+refuses every key that carries an override (the older build cannot parse the
+stored limits, so the key answers 500 until its overrides are removed); keys
+without one are unaffected.
+
 A key can also carry an expiry, which is optional and off by default — a key
 with none works until it is revoked, and no upgrade ever attaches one:
 
@@ -402,6 +414,17 @@ Worth knowing before you deploy it:
   Anthropic thinking fields (including beta controls) pass through only to
   Anthropic; supply any required beta header yourself. Hidden adaptive display
   omits generated Responses reasoning summaries, without inventing a default.
+- **An OpenAI OAuth (ChatGPT) target presents itself as Codex, and answers like
+  it.** Requests on that leg match Codex 0.163.0's TUI field by field, which
+  changes model behaviour for every client routed there, Claude Code included:
+  one tool call per turn (`parallel_tool_calls: false`), terser answers
+  (`text.verbosity: "low"`), and a request that names no reasoning effort —
+  including adaptive thinking with no effort — runs at the model's Codex catalog
+  default (`low` or `medium`) rather than `medium`. A client's own `text`,
+  `reasoning` or tool choice still wins field by field. OpenAI API-key targets
+  are unaffected. Credentials connected before this pick up their per-credential
+  identity on their next token refresh. Details:
+  [the fingerprint design](docs/superpowers/specs/2026-10-16-codex-0163-fingerprint-design.md).
 - Not in scope: semantic caching, billing.
 
 What is missing on purpose, what is missing for now, and what is designed but not