You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thanks. A stateless /v1/responses (the input items, instructions, tools, reasoning, and the typed streaming events) fits on top of the existing chat-completions path. It is planned after the current fixes. Which client do you need it for (Codex CLI, an SDK, something else)? That decides which parts of the streaming format matter first.
please give me initial feedback apdogod123321-ops i'm not sure its far enough along to be a draft pull request
working on a different version that starts with support for gbnf.
unless someone expresses interest in this branch it is dead
edit i'm updating this branch in parallel until i get the gbnf working.
the current Codex profile requests unsupported reasoning representations and namespaced tools.
working against the codex v160
so this is an intermediate step. just ignore this entire thing for now.
Thanks, that helps. Codex speaks only the Responses API, so it's the first client we'll test against: its streaming events and tool calls. The same endpoint should then serve Pi and Hermes too. It's planned after 0.1.38, so I'm keeping this open.
apdogod123321-ops
please check these pull requests out. #630 // this one is for you - i need your feedback to improve it #636 // this one is for me - i want feedback in general
I just know codex wont work with it and I think there are others like me that may want to use it with that and having compatibility with that will help overall.
Fixed in 0.1.39: Strata now has POST /v1/responses, OpenAI's Responses API. It is stateless (the client sends the whole conversation each time, as Codex does with store: false) and runs on the same path as /v1/chat/completions: string or item input, instructions, function tools (also Codex's namespace and custom tools), function_call / function_call_output round trips, reasoning.effort, max_output_tokens, text.format JSON schemas, and the official streaming events. The model's thinking comes back as reasoning items, and when they are sent back the conversation cache is reused.
Tested with Codex CLI 0.160.0 on Q2_0: a plain prompt, and a tool loop where Codex ran shell commands and the model answered from their output. Later turns reused about 96% of the prompt from the cache. The Codex config.toml is in docs/DETAILS.md, under "The Responses API and Codex CLI".
Not supported: previous_response_id and stored responses (nothing is stored), hosted tools such as web search (they are left out), and reasoning summaries (this model writes none; set show_raw_agent_reasoning = true in Codex to see its thinking). Pi and Hermes should work through the same endpoint. If one of them doesn't, please open an issue with the request it sends.
Activity
Niko1221 commentedon Oct 2, 2026
Thanks. A stateless
/v1/responses(theinputitems,instructions, tools, reasoning, and the typed streaming events) fits on top of the existing chat-completions path. It is planned after the current fixes. Which client do you need it for (Codex CLI, an SDK, something else)? That decides which parts of the streaming format matter first.StoyanStAtanasov commentedon Oct 3, 2026
Codex, Hermes or Pi one of those works only with responses
CC-David-CC commentedon Oct 3, 2026
https://github.com/CC-David-CC/Strata-a5500/tree/feat/responses-api-451
please give me initial feedback apdogod123321-ops i'm not sure its far enough along to be a draft pull request
working on a different version that starts with support for gbnf.
unless someone expresses interest in this branch it is dead
edit i'm updating this branch in parallel until i get the gbnf working.
the current Codex profile requests unsupported reasoning representations and namespaced tools.
working against the codex v160
so this is an intermediate step. just ignore this entire thing for now.
Niko1221 commentedon Oct 3, 2026
Thanks, that helps. Codex speaks only the Responses API, so it's the first client we'll test against: its streaming events and tool calls. The same endpoint should then serve Pi and Hermes too. It's planned after 0.1.38, so I'm keeping this open.
CC-David-CC commentedon Oct 3, 2026
apdogod123321-ops
please check these pull requests out.
#630 // this one is for you - i need your feedback to improve it
#636 // this one is for me - i want feedback in general
CC-David-CC commentedon Oct 3, 2026
working out a few more issues however, apdogod123321-ops
, i think i am almost to your goal
edit
sigh i wish just gbnf was good enough but they use json in codex. ew. hold on i'm extending the gbnf branch deeper so it can support json
edit 2
update
no i haven't pushed that live to even mine yet. working on it as fast as possible
CC-David-CC commentedon Oct 3, 2026
https://github.com/CC-David-CC/Strata-a5500/tree/work/gbnf
apdogod123321-ops,
I am calling this request complete unless you bring up any issues
apdogod123321-ops commentedon Oct 4, 2026
I just know codex wont work with it and I think there are others like me that may want to use it with that and having compatibility with that will help overall.
apdogod123321-ops commentedon Oct 4, 2026
I will try your solution when I get some time and let you know unless someone else gets to it first
serve: the OpenAI Responses API, POST /v1/responses (#451)
Engine 0.1.39: the #577 file-tier regression fixed, the OpenAI Respon…
Niko1221 commentedon Oct 4, 2026
Fixed in 0.1.39: Strata now has
POST /v1/responses, OpenAI's Responses API. It is stateless (the client sends the whole conversation each time, as Codex does withstore: false) and runs on the same path as/v1/chat/completions: string or iteminput,instructions, function tools (also Codex's namespace and custom tools), function_call / function_call_output round trips,reasoning.effort,max_output_tokens,text.formatJSON schemas, and the official streaming events. The model's thinking comes back as reasoning items, and when they are sent back the conversation cache is reused.Tested with Codex CLI 0.160.0 on Q2_0: a plain prompt, and a tool loop where Codex ran shell commands and the model answered from their output. Later turns reused about 96% of the prompt from the cache. The Codex
config.tomlis in docs/DETAILS.md, under "The Responses API and Codex CLI".Not supported:
previous_response_idand stored responses (nothing is stored), hosted tools such as web search (they are left out), and reasoning summaries (this model writes none; setshow_raw_agent_reasoning = truein Codex to see its thinking). Pi and Hermes should work through the same endpoint. If one of them doesn't, please open an issue with the request it sends.serve: the OpenAI Responses API, POST /v1/responses (#451)
Engine 0.1.39: the #577 file-tier regression fixed, the OpenAI Respon…
serve: the OpenAI Responses API, POST /v1/responses (Niko1221#451)
Engine 0.1.39: the Niko1221#577 file-tier regression fixed, the OpenA…
serve: the OpenAI Responses API, POST /v1/responses (Niko1221#451)
Engine 0.1.39: the Niko1221#577 file-tier regression fixed, the OpenA…