dspy-codex-auth 0.2.0__tar.gz → 0.3.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: dspy-codex-auth
3
- Version: 0.2.0
3
+ Version: 0.3.0
4
4
  Summary: DSPy integration for ChatGPT Codex subscription-backed language models
5
5
  Keywords: codex,dspy,language-models,openai
6
6
  Author: hrbatra
@@ -13,8 +13,7 @@ Classifier: Programming Language :: Python :: 3
13
13
  Classifier: Programming Language :: Python :: 3.12
14
14
  Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
15
15
  Classifier: Topic :: Software Development :: Libraries :: Python Modules
16
- Requires-Dist: dspy>=3.2.0
17
- Requires-Dist: litellm>=1.82.6
16
+ Requires-Dist: dspy[litellm]>=3.5.0
18
17
  Requires-Dist: openai-codex-auth>=0.2.0
19
18
  Requires-Python: >=3.12
20
19
  Project-URL: Homepage, https://github.com/hrbatra/dspy-codex-auth
@@ -30,8 +29,9 @@ language model.
30
29
  - It uses [`openai-codex-auth`](https://github.com/hrbatra/openai-codex-auth)
31
30
  for authentication, HTTP/WebSocket requests, and streamed-response handling.
32
31
  - It installs a DSPy `LM` wrapper for `codex/...` model strings.
33
- - It converts DSPy messages and options into native Codex requests, and converts
34
- shared client responses into DSPy's output, history, and usage formats.
32
+ - Its sync and async engines convert DSPy's typed requests into native Codex
33
+ requests and return `dspy.lm15.Response` objects. DSPy owns caching, retries,
34
+ history, callbacks, and usage accounting.
35
35
 
36
36
  ## Install
37
37
 
@@ -39,6 +39,15 @@ language model.
39
39
  uv add dspy-codex-auth
40
40
  ```
41
41
 
42
+ Version 0.3 requires Python 3.12+ and DSPy 3.5+. To upgrade an existing project:
43
+
44
+ ```bash
45
+ uv add --upgrade dspy-codex-auth
46
+ ```
47
+
48
+ Restart the notebook kernel or Python process after upgrading so it loads the
49
+ new adapter and DSPy classes.
50
+
42
51
  The two packages have separate roles: `openai-codex-auth` provides the reusable
43
52
  Codex client; `dspy-codex-auth` provides its DSPy adapter. Installing the adapter
44
53
  brings in both packages. For ordinary Python without DSPy, install
@@ -77,12 +86,54 @@ lm = dspy.LM("codex/gpt-5.6-luna", cache=False)
77
86
  dspy.configure(lm=lm, adapter=dspy.JSONAdapter())
78
87
  ```
79
88
 
80
- Use `codex/<model>` for the ChatGPT Codex subscription route. This is not the
89
+ Use a model available to your ChatGPT account with `codex/<model>` for the
90
+ Codex subscription route. The examples use `gpt-5.6-sol`; older model IDs may
91
+ be rejected by the backend. This is not the
81
92
  OpenAI API key route and does not require `OPENAI_API_KEY`.
82
93
 
83
94
  `cache=False` is recommended for Codex while iterating because stale DSPy cache
84
95
  entries can preserve old empty-output responses across package upgrades.
85
96
 
97
+ ## Migrating to 0.3
98
+
99
+ Ordinary `lm("hello")`, `await lm.acall("hello")`, and DSPy modules such as
100
+ `dspy.Predict` keep their usual interface. String calls return DSPy's list of
101
+ outputs. The old `lm(messages=[...])`, `lm.forward(...)`, and
102
+ `lm.aforward(...)` interfaces are removed.
103
+
104
+ For a conversation or direct structured call, use types from **`dspy.lm15`**:
105
+
106
+ ```python
107
+ from dspy.lm15 import Config, Message, Reasoning, Request, Response
108
+
109
+ request = Request(
110
+ model=lm.model,
111
+ system="Be concise.",
112
+ messages=(Message.user("What is DSPy?"),),
113
+ config=Config(reasoning=Reasoning(effort="low", summary="detailed")),
114
+ )
115
+ response: Response = lm(request)
116
+ print(response.text)
117
+
118
+ # In a notebook or async function:
119
+ response = await lm.acall(request)
120
+ ```
121
+
122
+ `Request.model` must equal `lm.model`. Put generation settings in
123
+ `Request.config`: a typed request does **not** inherit the LM's generation
124
+ defaults or accept generation keyword arguments alongside it. String calls
125
+ continue to use constructor defaults and per-call generation overrides.
126
+ Connection and transport overrides, such as `timeout` and `codex_transport`,
127
+ can still be passed alongside either call form.
128
+
129
+ This adapter collects the complete provider response before returning it. It
130
+ does not expose token streaming through DSPy's streaming engine interface.
131
+ Stop sequences are unsupported: `stop=` on a string call or `Config.stop` on
132
+ a typed request raises `dspy.LMUnsupportedFeatureError` before provider I/O.
133
+ The completion-only engine cannot stop generation at a matching sequence.
134
+ See DSPy's [LM migration guide](https://dspy.ai/community/normalized-lm-api-migration/)
135
+ for the underlying API changes.
136
+
86
137
  ## Codex Transport Selection
87
138
 
88
139
  Codex LMs accept `codex_transport` at construction time. Its accepted values
@@ -122,11 +173,20 @@ timeout defaults to 300 seconds. Custom `api_base` values must use HTTP or
122
173
  HTTPS; the WebSocket transport converts them to WS or WSS and appends the
123
174
  `/responses` path. TLS uses Requests' CA bundle.
124
175
 
125
- The shared client owns retries. `num_retries` counts additional attempts across
126
- transports; retry delays begin at 0.5 seconds and double up to 8 seconds.
127
- Auto selection may additionally make an HTTP routing probe. An explicit numeric
128
- `timeout` also controls WebSocket idle time under the existing call/constructor
129
- precedence; HTTP supports phase-specific `httpx.Timeout` values.
176
+ DSPy owns the single retry loop. `num_retries=k` permits at most `k + 1` engine
177
+ attempts; retry delays are 1, 2, 4, ... seconds, capped at 60 seconds unless a
178
+ provider retry hint applies. The shared client receives `max_retries=0`.
179
+ Auto selection may additionally make an HTTP routing probe within an attempt.
180
+ HTTP supports phase-specific `httpx.Timeout` values. For WebSocket idle time,
181
+ the precedence is:
182
+
183
+ 1. A numeric per-call `timeout`.
184
+ 2. A per-call `codex_websocket_idle_timeout`.
185
+ 3. A numeric constructor `timeout`.
186
+ 4. The constructor `codex_websocket_idle_timeout` (300 seconds by default).
187
+
188
+ Passing `timeout=None` on a call clears the constructor timeout for that call.
189
+ Per-call overrides do not change subsequent calls' defaults.
130
190
 
131
191
  Authentication remains in the WebSocket handshake: the bearer credential is
132
192
  never placed in the `response.create` data frame. HTTP requests use originator
@@ -152,7 +212,7 @@ def configure_model(model: str, **kwargs):
152
212
  return lm
153
213
 
154
214
 
155
- codex_lm = configure_model("codex/gpt-5.5", cache=False)
215
+ codex_lm = configure_model("codex/gpt-5.6-sol", cache=False)
156
216
  api_lm = configure_model("openai/gpt-5.5", api_key="...", cache=False)
157
217
  ```
158
218
 
@@ -163,20 +223,52 @@ Pass `reasoning_effort` as usual. This package also supports
163
223
 
164
224
  ```python
165
225
  lm = dspy.LM(
166
- "codex/gpt-5.5",
226
+ "codex/gpt-5.6-sol",
167
227
  cache=False,
168
228
  reasoning_effort="medium",
169
229
  reasoning_summary="detailed",
170
230
  )
171
231
  ```
172
232
 
173
- DSPy predictions expose declared output fields. The lower-level LM history can
174
- also include a returned reasoning summary:
233
+ DSPy predictions expose declared output fields. A single-response history entry
234
+ contains the canonical response, including returned reasoning parts:
235
+
236
+ ```python
237
+ from dspy.lm15 import ThinkingPart
238
+
239
+ response = lm.history[-1]["response"]
240
+ summaries = [part.text for part in response.message.parts_of(ThinkingPart)]
241
+ ```
242
+
243
+ Read `response.usage.input_tokens`, `response.usage.output_tokens`, and
244
+ `response.usage.total_tokens` for canonical usage; an unreported counter is
245
+ `None`, not zero. DSPy's history `usage` field uses its own prompt/completion
246
+ accounting format. Cache hits do not add billed usage.
247
+
248
+ Provider-specific details are available without inspecting private objects:
175
249
 
176
250
  ```python
177
- summary = lm.history[-1]["outputs"][0].get("reasoning_content")
251
+ raw_response = response.provider_data["codex_response"]
252
+ selected_transport = response.provider_data["codex_transport"]
178
253
  ```
179
254
 
255
+ The Responses codec reports dropped or clamped settings in
256
+ `response.adaptations`. Removed Codex output-token caps are recorded there too,
257
+ so callers can inspect which requested settings were changed.
258
+
259
+ ## Copying and Saving
260
+
261
+ `lm.copy(...)`, `copy.deepcopy(lm)`, and DSPy's full-program save/load preserve
262
+ the Codex integration. Copies and saved LM state retain the resolved Codex
263
+ model and route, so changing a custom alias registration later does not
264
+ silently redirect them.
265
+
266
+ `lm.dump_state()` produces reconstructible JSON state with the auth-file path
267
+ and transport settings, omitting credentials and custom headers. Reapply any
268
+ explicit key or custom headers after loading that state. This omission applies
269
+ to the JSON state format; DSPy's full-program pickle includes runtime objects
270
+ and is not the sanitized JSON form.
271
+
180
272
  ## Fast Mode / Service Tier
181
273
 
182
274
  Codex CLI config uses `service_tier = "fast"` for Fast mode. Internally, the
@@ -186,7 +278,7 @@ spelling directly:
186
278
 
187
279
  ```python
188
280
  lm = dspy.LM(
189
- "codex/gpt-5.4",
281
+ "codex/gpt-5.6-sol",
190
282
  cache=False,
191
283
  service_tier="fast",
192
284
  reasoning_effort="low",
@@ -200,7 +292,7 @@ For raw Codex CLI calls, use `-c` config overrides:
200
292
 
201
293
  ```bash
202
294
  codex exec \
203
- -m gpt-5.4 \
295
+ -m gpt-5.6-sol \
204
296
  -c 'service_tier="fast"' \
205
297
  -c 'model_reasoning_effort="low"' \
206
298
  --json \
@@ -223,7 +315,7 @@ explicitly:
223
315
 
224
316
  ```python
225
317
  lm = dspy_codex_auth.LM(
226
- "openai/gpt-5.5",
318
+ "openai/gpt-5.6-sol",
227
319
  auth_provider="codex",
228
320
  cache=False,
229
321
  reasoning_effort="medium",
@@ -241,9 +333,9 @@ preserving provider metadata, message text, function calls, and reasoning
241
333
  summaries. This adapter converts that response into the objects DSPy expects.
242
334
  Codex requests no longer pass through LiteLLM's Responses transport.
243
335
 
244
- A provider refusal raises `openai_codex_auth.CodexError` in this adapter so it
245
- cannot become an empty DSPy prediction. Direct `CodexClient` callers receive
246
- the native refusal item in `CodexResponse.output`.
336
+ Provider failures and refusals surface through DSPy's public `LMError` family.
337
+ A refusal cannot become an empty successful prediction. Direct `CodexClient`
338
+ callers still receive the native refusal item in `CodexResponse.output`.
247
339
 
248
340
  The adapter and shared core together handle:
249
341
 
@@ -275,7 +367,7 @@ import dspy_codex_auth
275
367
 
276
368
  dspy_codex_auth.install()
277
369
 
278
- lm = dspy.LM("codex/gpt-5.5", cache=False)
370
+ lm = dspy.LM("codex/gpt-5.6-sol", cache=False)
279
371
  dspy.configure(lm=lm, adapter=dspy.JSONAdapter())
280
372
 
281
373
 
@@ -299,7 +391,7 @@ import dspy_codex_auth
299
391
  dspy_codex_auth.install()
300
392
 
301
393
  lm = dspy.LM(
302
- "codex/gpt-5.5",
394
+ "codex/gpt-5.6-sol",
303
395
  cache=False,
304
396
  reasoning_effort="medium",
305
397
  reasoning_summary="detailed",
@@ -325,7 +417,8 @@ pred = solver(
325
417
 
326
418
  print(pred.answer)
327
419
  print(pred.explanation)
328
- print(lm.history[-1]["outputs"][0].get("reasoning_content"))
420
+ response = lm.history[-1]["response"]
421
+ print(response.provider_data["codex_response"])
329
422
  ```
330
423
 
331
424
  ## Attribution
@@ -347,11 +440,13 @@ https://github.com/MaximeRivest/dspy-lm-auth/pull/2
347
440
 
348
441
  ```bash
349
442
  uv sync --dev
350
- uv run pytest
351
- uv run ruff check .
352
- uv build --no-sources
443
+ scripts/check
353
444
  ```
354
445
 
446
+ `scripts/check` runs lint, offline tests, a distribution build, and sync/async
447
+ smoke tests against an isolated installation of the built wheel. Build and
448
+ cache artifacts for this check use a temporary directory.
449
+
355
450
  The default test command skips the real-network live transport test. Run it
356
451
  explicitly only with valid Codex subscription credentials:
357
452
 
@@ -368,9 +463,12 @@ Short local release flow:
368
463
 
369
464
  ```bash
370
465
  uv version --bump patch
371
- rm -rf dist
466
+ scripts/check
372
467
  uv build --no-sources
373
468
  uv run --with twine python -m twine upload dist/*
374
469
  ```
375
470
 
376
- PyPI releases are immutable, so every update needs a new version number.
471
+ Use a minor bump for compatibility changes such as 0.3. PyPI releases are
472
+ immutable, so every update needs a new version number. Build release artifacts
473
+ in a clean `dist/` directory; see the full release instructions for GitHub,
474
+ tagging, and fresh-install verification.
@@ -6,8 +6,9 @@ language model.
6
6
  - It uses [`openai-codex-auth`](https://github.com/hrbatra/openai-codex-auth)
7
7
  for authentication, HTTP/WebSocket requests, and streamed-response handling.
8
8
  - It installs a DSPy `LM` wrapper for `codex/...` model strings.
9
- - It converts DSPy messages and options into native Codex requests, and converts
10
- shared client responses into DSPy's output, history, and usage formats.
9
+ - Its sync and async engines convert DSPy's typed requests into native Codex
10
+ requests and return `dspy.lm15.Response` objects. DSPy owns caching, retries,
11
+ history, callbacks, and usage accounting.
11
12
 
12
13
  ## Install
13
14
 
@@ -15,6 +16,15 @@ language model.
15
16
  uv add dspy-codex-auth
16
17
  ```
17
18
 
19
+ Version 0.3 requires Python 3.12+ and DSPy 3.5+. To upgrade an existing project:
20
+
21
+ ```bash
22
+ uv add --upgrade dspy-codex-auth
23
+ ```
24
+
25
+ Restart the notebook kernel or Python process after upgrading so it loads the
26
+ new adapter and DSPy classes.
27
+
18
28
  The two packages have separate roles: `openai-codex-auth` provides the reusable
19
29
  Codex client; `dspy-codex-auth` provides its DSPy adapter. Installing the adapter
20
30
  brings in both packages. For ordinary Python without DSPy, install
@@ -53,12 +63,54 @@ lm = dspy.LM("codex/gpt-5.6-luna", cache=False)
53
63
  dspy.configure(lm=lm, adapter=dspy.JSONAdapter())
54
64
  ```
55
65
 
56
- Use `codex/<model>` for the ChatGPT Codex subscription route. This is not the
66
+ Use a model available to your ChatGPT account with `codex/<model>` for the
67
+ Codex subscription route. The examples use `gpt-5.6-sol`; older model IDs may
68
+ be rejected by the backend. This is not the
57
69
  OpenAI API key route and does not require `OPENAI_API_KEY`.
58
70
 
59
71
  `cache=False` is recommended for Codex while iterating because stale DSPy cache
60
72
  entries can preserve old empty-output responses across package upgrades.
61
73
 
74
+ ## Migrating to 0.3
75
+
76
+ Ordinary `lm("hello")`, `await lm.acall("hello")`, and DSPy modules such as
77
+ `dspy.Predict` keep their usual interface. String calls return DSPy's list of
78
+ outputs. The old `lm(messages=[...])`, `lm.forward(...)`, and
79
+ `lm.aforward(...)` interfaces are removed.
80
+
81
+ For a conversation or direct structured call, use types from **`dspy.lm15`**:
82
+
83
+ ```python
84
+ from dspy.lm15 import Config, Message, Reasoning, Request, Response
85
+
86
+ request = Request(
87
+ model=lm.model,
88
+ system="Be concise.",
89
+ messages=(Message.user("What is DSPy?"),),
90
+ config=Config(reasoning=Reasoning(effort="low", summary="detailed")),
91
+ )
92
+ response: Response = lm(request)
93
+ print(response.text)
94
+
95
+ # In a notebook or async function:
96
+ response = await lm.acall(request)
97
+ ```
98
+
99
+ `Request.model` must equal `lm.model`. Put generation settings in
100
+ `Request.config`: a typed request does **not** inherit the LM's generation
101
+ defaults or accept generation keyword arguments alongside it. String calls
102
+ continue to use constructor defaults and per-call generation overrides.
103
+ Connection and transport overrides, such as `timeout` and `codex_transport`,
104
+ can still be passed alongside either call form.
105
+
106
+ This adapter collects the complete provider response before returning it. It
107
+ does not expose token streaming through DSPy's streaming engine interface.
108
+ Stop sequences are unsupported: `stop=` on a string call or `Config.stop` on
109
+ a typed request raises `dspy.LMUnsupportedFeatureError` before provider I/O.
110
+ The completion-only engine cannot stop generation at a matching sequence.
111
+ See DSPy's [LM migration guide](https://dspy.ai/community/normalized-lm-api-migration/)
112
+ for the underlying API changes.
113
+
62
114
  ## Codex Transport Selection
63
115
 
64
116
  Codex LMs accept `codex_transport` at construction time. Its accepted values
@@ -98,11 +150,20 @@ timeout defaults to 300 seconds. Custom `api_base` values must use HTTP or
98
150
  HTTPS; the WebSocket transport converts them to WS or WSS and appends the
99
151
  `/responses` path. TLS uses Requests' CA bundle.
100
152
 
101
- The shared client owns retries. `num_retries` counts additional attempts across
102
- transports; retry delays begin at 0.5 seconds and double up to 8 seconds.
103
- Auto selection may additionally make an HTTP routing probe. An explicit numeric
104
- `timeout` also controls WebSocket idle time under the existing call/constructor
105
- precedence; HTTP supports phase-specific `httpx.Timeout` values.
153
+ DSPy owns the single retry loop. `num_retries=k` permits at most `k + 1` engine
154
+ attempts; retry delays are 1, 2, 4, ... seconds, capped at 60 seconds unless a
155
+ provider retry hint applies. The shared client receives `max_retries=0`.
156
+ Auto selection may additionally make an HTTP routing probe within an attempt.
157
+ HTTP supports phase-specific `httpx.Timeout` values. For WebSocket idle time,
158
+ the precedence is:
159
+
160
+ 1. A numeric per-call `timeout`.
161
+ 2. A per-call `codex_websocket_idle_timeout`.
162
+ 3. A numeric constructor `timeout`.
163
+ 4. The constructor `codex_websocket_idle_timeout` (300 seconds by default).
164
+
165
+ Passing `timeout=None` on a call clears the constructor timeout for that call.
166
+ Per-call overrides do not change subsequent calls' defaults.
106
167
 
107
168
  Authentication remains in the WebSocket handshake: the bearer credential is
108
169
  never placed in the `response.create` data frame. HTTP requests use originator
@@ -128,7 +189,7 @@ def configure_model(model: str, **kwargs):
128
189
  return lm
129
190
 
130
191
 
131
- codex_lm = configure_model("codex/gpt-5.5", cache=False)
192
+ codex_lm = configure_model("codex/gpt-5.6-sol", cache=False)
132
193
  api_lm = configure_model("openai/gpt-5.5", api_key="...", cache=False)
133
194
  ```
134
195
 
@@ -139,20 +200,52 @@ Pass `reasoning_effort` as usual. This package also supports
139
200
 
140
201
  ```python
141
202
  lm = dspy.LM(
142
- "codex/gpt-5.5",
203
+ "codex/gpt-5.6-sol",
143
204
  cache=False,
144
205
  reasoning_effort="medium",
145
206
  reasoning_summary="detailed",
146
207
  )
147
208
  ```
148
209
 
149
- DSPy predictions expose declared output fields. The lower-level LM history can
150
- also include a returned reasoning summary:
210
+ DSPy predictions expose declared output fields. A single-response history entry
211
+ contains the canonical response, including returned reasoning parts:
212
+
213
+ ```python
214
+ from dspy.lm15 import ThinkingPart
215
+
216
+ response = lm.history[-1]["response"]
217
+ summaries = [part.text for part in response.message.parts_of(ThinkingPart)]
218
+ ```
219
+
220
+ Read `response.usage.input_tokens`, `response.usage.output_tokens`, and
221
+ `response.usage.total_tokens` for canonical usage; an unreported counter is
222
+ `None`, not zero. DSPy's history `usage` field uses its own prompt/completion
223
+ accounting format. Cache hits do not add billed usage.
224
+
225
+ Provider-specific details are available without inspecting private objects:
151
226
 
152
227
  ```python
153
- summary = lm.history[-1]["outputs"][0].get("reasoning_content")
228
+ raw_response = response.provider_data["codex_response"]
229
+ selected_transport = response.provider_data["codex_transport"]
154
230
  ```
155
231
 
232
+ The Responses codec reports dropped or clamped settings in
233
+ `response.adaptations`. Removed Codex output-token caps are recorded there too,
234
+ so callers can inspect which requested settings were changed.
235
+
236
+ ## Copying and Saving
237
+
238
+ `lm.copy(...)`, `copy.deepcopy(lm)`, and DSPy's full-program save/load preserve
239
+ the Codex integration. Copies and saved LM state retain the resolved Codex
240
+ model and route, so changing a custom alias registration later does not
241
+ silently redirect them.
242
+
243
+ `lm.dump_state()` produces reconstructible JSON state with the auth-file path
244
+ and transport settings, omitting credentials and custom headers. Reapply any
245
+ explicit key or custom headers after loading that state. This omission applies
246
+ to the JSON state format; DSPy's full-program pickle includes runtime objects
247
+ and is not the sanitized JSON form.
248
+
156
249
  ## Fast Mode / Service Tier
157
250
 
158
251
  Codex CLI config uses `service_tier = "fast"` for Fast mode. Internally, the
@@ -162,7 +255,7 @@ spelling directly:
162
255
 
163
256
  ```python
164
257
  lm = dspy.LM(
165
- "codex/gpt-5.4",
258
+ "codex/gpt-5.6-sol",
166
259
  cache=False,
167
260
  service_tier="fast",
168
261
  reasoning_effort="low",
@@ -176,7 +269,7 @@ For raw Codex CLI calls, use `-c` config overrides:
176
269
 
177
270
  ```bash
178
271
  codex exec \
179
- -m gpt-5.4 \
272
+ -m gpt-5.6-sol \
180
273
  -c 'service_tier="fast"' \
181
274
  -c 'model_reasoning_effort="low"' \
182
275
  --json \
@@ -199,7 +292,7 @@ explicitly:
199
292
 
200
293
  ```python
201
294
  lm = dspy_codex_auth.LM(
202
- "openai/gpt-5.5",
295
+ "openai/gpt-5.6-sol",
203
296
  auth_provider="codex",
204
297
  cache=False,
205
298
  reasoning_effort="medium",
@@ -217,9 +310,9 @@ preserving provider metadata, message text, function calls, and reasoning
217
310
  summaries. This adapter converts that response into the objects DSPy expects.
218
311
  Codex requests no longer pass through LiteLLM's Responses transport.
219
312
 
220
- A provider refusal raises `openai_codex_auth.CodexError` in this adapter so it
221
- cannot become an empty DSPy prediction. Direct `CodexClient` callers receive
222
- the native refusal item in `CodexResponse.output`.
313
+ Provider failures and refusals surface through DSPy's public `LMError` family.
314
+ A refusal cannot become an empty successful prediction. Direct `CodexClient`
315
+ callers still receive the native refusal item in `CodexResponse.output`.
223
316
 
224
317
  The adapter and shared core together handle:
225
318
 
@@ -251,7 +344,7 @@ import dspy_codex_auth
251
344
 
252
345
  dspy_codex_auth.install()
253
346
 
254
- lm = dspy.LM("codex/gpt-5.5", cache=False)
347
+ lm = dspy.LM("codex/gpt-5.6-sol", cache=False)
255
348
  dspy.configure(lm=lm, adapter=dspy.JSONAdapter())
256
349
 
257
350
 
@@ -275,7 +368,7 @@ import dspy_codex_auth
275
368
  dspy_codex_auth.install()
276
369
 
277
370
  lm = dspy.LM(
278
- "codex/gpt-5.5",
371
+ "codex/gpt-5.6-sol",
279
372
  cache=False,
280
373
  reasoning_effort="medium",
281
374
  reasoning_summary="detailed",
@@ -301,7 +394,8 @@ pred = solver(
301
394
 
302
395
  print(pred.answer)
303
396
  print(pred.explanation)
304
- print(lm.history[-1]["outputs"][0].get("reasoning_content"))
397
+ response = lm.history[-1]["response"]
398
+ print(response.provider_data["codex_response"])
305
399
  ```
306
400
 
307
401
  ## Attribution
@@ -323,11 +417,13 @@ https://github.com/MaximeRivest/dspy-lm-auth/pull/2
323
417
 
324
418
  ```bash
325
419
  uv sync --dev
326
- uv run pytest
327
- uv run ruff check .
328
- uv build --no-sources
420
+ scripts/check
329
421
  ```
330
422
 
423
+ `scripts/check` runs lint, offline tests, a distribution build, and sync/async
424
+ smoke tests against an isolated installation of the built wheel. Build and
425
+ cache artifacts for this check use a temporary directory.
426
+
331
427
  The default test command skips the real-network live transport test. Run it
332
428
  explicitly only with valid Codex subscription credentials:
333
429
 
@@ -344,9 +440,12 @@ Short local release flow:
344
440
 
345
441
  ```bash
346
442
  uv version --bump patch
347
- rm -rf dist
443
+ scripts/check
348
444
  uv build --no-sources
349
445
  uv run --with twine python -m twine upload dist/*
350
446
  ```
351
447
 
352
- PyPI releases are immutable, so every update needs a new version number.
448
+ Use a minor bump for compatibility changes such as 0.3. PyPI releases are
449
+ immutable, so every update needs a new version number. Build release artifacts
450
+ in a clean `dist/` directory; see the full release instructions for GitHub,
451
+ tagging, and fresh-install verification.
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "dspy-codex-auth"
3
- version = "0.2.0"
3
+ version = "0.3.0"
4
4
  description = "DSPy integration for ChatGPT Codex subscription-backed language models"
5
5
  readme = "README.md"
6
6
  authors = [
@@ -19,8 +19,7 @@ classifiers = [
19
19
  "Topic :: Software Development :: Libraries :: Python Modules",
20
20
  ]
21
21
  dependencies = [
22
- "dspy>=3.2.0",
23
- "litellm>=1.82.6",
22
+ "dspy[litellm]>=3.5.0",
24
23
  "openai-codex-auth>=0.2.0",
25
24
  ]
26
25
 
@@ -35,6 +34,7 @@ build-backend = "uv_build"
35
34
 
36
35
  [dependency-groups]
37
36
  dev = [
37
+ "coverage>=7.17.3",
38
38
  "pytest>=9.0.3",
39
39
  "ruff>=0.15.12",
40
40
  ]