@askalf/dario 6.2.0 → 6.3.0

package/docs/integrations/codex-cli.md

@@ -0,0 +1,94 @@
+# Codex CLI on your Claude plan
+
+OpenAI's Codex CLI runs on a Claude subscription through dario, with its
+tools, since 6.3. The other way round — Claude Code on a ChatGPT plan — has
+worked since 5.5.89. Both on one machine, one `dario proxy`, is the whole
+point of the project in two panes.
+
+## Why this needed a new endpoint
+
+Codex CLI 0.154 removed `wire_api = "chat"` for custom providers
+([openai/codex discussion 7782](https://github.com/openai/codex/discussions/7782)):
+a provider has to speak the **Responses API** or it cannot be used at all. So
+dario now has `POST /v1/responses`. The request is translated once at the front
+door into the Messages shape every other dario path already serves — pool,
+template, failover, mid-stream continuation — and everything written back is
+translated at the write boundary. For a ChatGPT-subscription model the body
+goes through to the codex backend untouched instead (it speaks Responses
+natively), which is what keeps Codex's newest request features working there:
+`additional_tools` input items, `custom` tools, `reasoning.context`,
+`include`.
+
+The OpenAI Agents SDK and anything else that speaks Responses gets the same
+endpoint.
+
+## Setup
+
+```bash
+dario proxy                                   # your Claude plan, port 3456
+```
+
+`~/.codex/config.toml`:
+
+```toml
+model = "claude-opus-5"
+model_provider = "dario"
+
+[model_providers.dario]
+name = "dario"
+base_url = "http://127.0.0.1:3456/v1"
+env_key = "DARIO_API_KEY"
+wire_api = "responses"
+```
+
+`export DARIO_API_KEY=dario` (any value until you set one on the proxy; then it
+has to match), and `codex` runs on the Claude pool. `-m claude-sonnet-5` /
+`-m claude-opus-5` / any `claude-*` id; `codex exec` works the same way.
+
+With a ChatGPT account attached as well (`dario add altman`),
+`-m gpt-5.6-sol` on the same provider block goes to that plan through dario —
+pooling, failover and `x-dario-*` headers included.
+
+Codex prints `Model metadata for claude-opus-5 not found. Defaulting to fallback
+metadata` for any model it does not ship metadata for. Harmless: it falls back
+to its classic request profile (top-level `tools` and `instructions`), which is
+exactly the one the translation reads.
+
+## What runs, and what does not
+
+Verified on 2026-09-12 with Codex CLI 0.154.0 on a Claude Max plan:
+
+- a plain turn;
+- the full agent loop — `exec_command` called by Claude, executed by Codex,
+  the `function_call_output` returned, Claude answering from it — with the
+  17 KB Codex system prompt cached on the Claude side (98–99% cache reads
+  from the second turn);
+- the same loop on a ChatGPT plan through the passthrough.
+
+Translated on the Claude pool: `instructions` and `developer` messages
+(hoisted to the system prompt, in order), `message` items with `input_text`,
+`output_text` and `input_image` parts (data-URL and https images),
+`function_call` / `function_call_output` (tool_use / tool_result, call ids
+preserved), `function` tools (and `namespace` groups, flattened),
+`tool_choice` including `required` and a named function,
+`parallel_tool_calls: false`, `max_output_tokens`, `temperature`, `top_p`,
+`reasoning.effort` (as dario's own `model:high` effort spelling). Back out:
+`message`, `function_call` and `reasoning` items, the full Responses event
+sequence with sequence numbers, usage in OpenAI terms (cached prefix inside
+`input_tokens`, reported again under `cached_tokens`), `incomplete` on
+`max_tokens`, `response.failed` on an upstream error.
+
+Dropped, with a line at `--verbose`: hosted tool types the pool cannot run
+(`web_search`, `file_search`, `mcp`, …), `custom` freeform tools,
+`reasoning` items on the way in (OpenAI's encrypted content — the pool does
+not need it back), `text.format`. `previous_response_id` on the Claude pool is a 400 naming the
+field (dario is stateless there; send the full input each turn, which Codex
+does); on a ChatGPT-subscription model it is forwarded untouched to a backend
+that keeps state, `store` as you sent it or omitted for the backend's default.
+
+Not built: a buffered (non-streaming) response from a ChatGPT-subscription
+model on this route — the backend streams, and folding a Responses stream into
+a response object is not written yet; the answer is a 400 naming `stream`.
+Mid-stream continuation runs under this route on the Claude pool (the
+translated request is an ordinary Anthropic-shape request) and not on the
+passthrough.

package/docs/midstream-continuation.md

@@ -134,6 +134,32 @@ is never closed with a synthetic `end_turn`; only the resume's own
 | `--pool-fallback=…` | where the second hop goes; without an entry for the other provider a stream gets the same-model resume only |
 
 On by default: it only ever acts where the alternative is a broken stream.
+`dario doctor` reports which hops this host can take:
+
+```
+[ OK ]  Continuation  on: a dying stream resumes on the same model, then on gpt-5.6-sol → claude-sonnet-5 (two hops)
+[ OK ]  Continuation  on: a dying stream resumes on the same model only — add --pool-fallback for a second hop on the other subscription
+[INFO]  Continuation  off — a stream that dies mid-answer ends truncated (unset DARIO_MIDSTREAM_CONTINUE / drop --no-midstream-continue)
+```
+
+## Seeing it happen
+
+Nothing about a healthy stream shows the feature, so there is a tap that
+kills one on purpose:
+
+```bash
+DARIO_CHAOS_CUT_AFTER=300 dario proxy
+```
+
+The first streamed answer dies after 300 characters — the upstream socket is
+cut from dario's side, exactly the failure a real reset produces — and the
+continuation finishes it. Point any client at the proxy, ask for something
+long, and watch the answer keep going past the cut; a raw `curl -N` shows the
+seam comment. `DARIO_CHAOS_CUT_STREAMS=3` cuts the first three instead of one.
+The tap spares resumes, so it shows the first hop — the same model finishing
+its own answer; the other subscription takes over only when that model cannot
+serve the resume. dario warns loudly at startup while the tap is set; it is a
+demo and test affordance, never a default.
 
 ## How it was proven
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@askalf/dario",
-  "version": "6.2.0",
+  "version": "6.3.0",
   "description": "Use your Claude and ChatGPT subscriptions in Cursor, Cline, Aider, Claude Code and the Agent SDK — at subscription pricing, not per-token API bills. One local Anthropic + OpenAI-compatible endpoint: either plan answers either wire shape, with automatic failover when one hits its limit.",
   "type": "module",
   "bin": {