2026 10 04 api sync

CHANGELOG.md

@@ -1,5 +1,30 @@
 # Changelog
 
+## [1.6.0] - 2026-10-04
+
+### Changed
+
+- Sync the bundled OpenAPI spec, adding 10 paths and 18 schemas. The typed models gain the new response fields, including `attachments` on a run and on each step
+- Move `LATEST_API_VERSION` to `2026-10-03`
+
+### Added
+
+- Add `list_cloud_drives()`, `get_cloud_drive()`, `update_cloud_drive()`, `disconnect_cloud_drive()`, `delete_cloud_drive()`, `list_cloud_drive_providers()`, `get_agents_using_cloud_drive()` and `list_cloud_drive_rejections()` for cloud-drive connections. The list methods return the items on either response shape
+- Add `list_source_contents()` and `get_source_content_status()` to read the indexing status of a source's content, keyed by the `content_version_id` the upload methods return
+- Add `list_embedding_models()` and `list_reranker_models()`, with the pricing and defaults that sit beside each list
+- Add `ApiVersion` members for `2026-08-03`, `2026-08-21`, `2026-09-28`, `2026-09-30` and `2026-10-03`, so each can be selected without `allow_unknown_api_version`
+- Add a `param_style` argument to `paginate()` for endpoints that page by `offset` rather than `page`
+
+### Fixed
+
+- Read the body of an error response in the streaming methods. A 422 from `run_streaming_agent_and_wait()` escaped as `httpx.ResponseNotRead`, a 422 from `run_streaming_agent()` lost its field-level detail, and every streaming error had an empty `response_text`
+- Raise `SeclaiAPIStatusError` on a 422 whose `detail` is a plain string, from every method including the typed ones and the uploads. Decoding it as field-level validation raised `ValueError` from inside the generated models
+- Apply the unknown-version guard to a `Seclai-Version` passed in a per-request `headers` argument, on `request()` and the streaming methods. Any value was sent
+- Replace a header case-insensitively when a per-request `headers` argument or the auth layer supplies one the client already set. Both spellings were sent
+- Yield the items from `paginate()` when the endpoint answers with a bare array, or under `data` while a per-resource `items_key` was given. It yielded nothing in both cases, and now raises `SeclaiError` on a shape it cannot read
+- Copy `default_headers` at construction. Mutating the mapping afterwards put an unvalidated `Seclai-Version` on the wire
+- Reject an empty `Seclai-Version` in `default_headers`. It was read as absent, sent anyway, and suppressed `api_version`
+
 ## [1.5.0] - 2026-07-27
 
 ### Changed
@@ -158,6 +183,7 @@ _Stable release. Packaging, CI, and documentation deployment only; no API change
 
 _Initial release._
 
+[1.6.0]: https://github.com/seclai/seclai-python/releases/tag/1.6.0
 [1.5.0]: https://github.com/seclai/seclai-python/releases/tag/1.5.0
 [1.4.0]: https://github.com/seclai/seclai-python/releases/tag/1.4.0
 [1.3.0]: https://github.com/seclai/seclai-python/releases/tag/1.3.0

README.md

@@ -174,7 +174,9 @@ reshape responses, and this client would decode them incorrectly rather than
 reject them. Upgrade the package to adopt a new version, or pass
 `allow_unknown_api_version=True` if you have to move first and accept that risk.
 
-The guard only covers the header. An account pinned server-side can still be
+The guard covers the header however it is supplied — `api_version`,
+`default_headers`, or a per-request `headers` argument — and nothing else. An
+account pinned server-side can still be
 newer than this release — `get_api_version()` reports the `effective_version` the
 request resolved to, and comparing it against `LATEST_API_VERSION` is how you
 detect the gap.
@@ -184,8 +186,6 @@ of being ignored, and list endpoints move to the canonical
 `{"data": [...], "pagination": {...}}` envelope. The affected methods read both
 shapes, so they keep working either way — but the metadata moves:
 
-| Method | Before | From 2026-07-27 |
-| --- | --- | --- |
 | Method | Before | From 2026-07-27 | Legacy paging |
 | --- | --- | --- | --- |
 | `list_evaluation_criteria_page()` | bare list | `data` + `pagination` | none — returns everything |
@@ -210,8 +210,25 @@ then removed once the canonical envelope is the default.
 
 The two evaluation endpoints are **unpaginated** on the legacy shape — they
 ignore `page`/`limit` and return everything — so a paginate-until-empty loop over
-them only terminates once you have opted in. The other four paginate on either
-shape.
+them only terminates once you have opted in. `get_generation_tiers()` takes no
+paging arguments at all and always returns the full set. The remaining three
+paginate on either shape.
+
+**Later versions.** Each is cumulative, and none changes a response shape this
+client decodes:
+
+| Version | What it changes |
+| --- | --- |
+| `2026-08-03` | `create_memory_bank()` and `update_memory_bank()` reject a non-zero `max_age_days` with a 400, and an omitted `retention_days` on create resolves per bank type |
+| `2026-08-21` | `create_source()` rejects an embedding dimension its embedder does not support with a 400 — `list_embedding_models()` reports the supported ones |
+| `2026-09-28` | Agent-definition writes use the current file-list grammar: an omitted `attachments` keeps the stored list and `[]` means no files |
+| `2026-09-30` | A run's and a step's `output`, and a step's `input`, are the text rather than a JSON manifest; files are in `attachments` on every version |
+| `2026-10-03` | A new LLM step written without `attachments` takes its parent's files, and a new retrieval step's matched media are its files |
+
+The cloud-drive listings and `list_embedding_models()` / `list_reranker_models()`
+follow the `2026-07-27` envelope rule as well. The cloud-drive methods return
+the items on either shape; read the two model listings with
+`unwrap_items(result, "models")`.
 
 ## Resources
 
@@ -282,9 +299,9 @@ search = client.search_agent_runs({"query": "test"})
 # Fetch run details (optionally with step outputs)
 detail = client.get_agent_run("run_id", include_step_outputs=True)
 
-# Cancel or delete
+# Cancel an in-flight or queued run. `delete_agent_run()` is the same
+# operation on the same endpoint, returning a typed model instead of a dict.
 client.cancel_agent_run("run_id")
-client.delete_agent_run("run_id")
 ```
 
 ### Streaming
@@ -366,8 +383,8 @@ with response:
 steps = client.generate_agent_steps("agent_id", {"user_input": "Build a RAG pipeline"})
 config = client.generate_step_config("agent_id", {"step_type": "llm", "user_input": "..."})
 
-# Conversation history
-history = client.get_agent_ai_conversation_history("agent_id")
+# Conversation history — step_type is required by the API
+history = client.get_agent_ai_conversation_history("agent_id", step_type="llm")
 client.mark_agent_ai_suggestion("agent_id", "conversation_id", {"accepted": True})
 ```
 
@@ -448,6 +465,33 @@ client.update_source("source_id", {"name": "Updated"})
 client.delete_source("source_id")
 ```
 
+Indexing status of a source's content, keyed by the `content_version_id` the
+upload methods return:
+
+```python
+failed = client.list_source_contents("source_id", status="failed")
+batch = client.list_source_contents(
+    "source_id", content_version_ids=["cv_1", "cv_2"]
+)
+one = client.get_source_content_status("source_id", "cv_1")
+```
+
+### Cloud drives
+
+```python
+providers = client.list_cloud_drive_providers()
+drives = client.list_cloud_drives()
+drive = client.get_cloud_drive("connection_id")
+client.update_cloud_drive("connection_id", {"name": "Contracts"})
+
+# Which agents depend on it, and which files it skipped and why
+agents = client.get_agents_using_cloud_drive("connection_id")
+skipped = client.list_cloud_drive_rejections("connection_id", limit=20)
+
+client.disconnect_cloud_drive("connection_id")  # keeps the connection
+client.delete_cloud_drive("connection_id")
+```
+
 ### File uploads
 
 Upload a file to a source (max 200 MiB):
@@ -657,6 +701,10 @@ print(removed.get("cleanup_note"))  # set when the domain was Seclai-managed
 # Media-generation quality tiers (fast/balanced/thorough) and what each resolves to
 tiers = client.get_generation_tiers()
 
+# Embedding and reranker models, with their pricing
+embedders = unwrap_items(client.list_embedding_models(), "models")
+rerankers = unwrap_items(client.list_reranker_models(), "models")
+
 alerts = client.list_model_alerts()
 client.mark_model_alert_read("alert_id")
 client.mark_all_model_alerts_read()
@@ -719,8 +767,14 @@ All list methods accept `page` and `limit` parameters. For auto-pagination acros
 for agent in client.paginate("GET", "/agents"):
     print(agent["name"])
 
-# With a custom items key
-for alert in client.paginate("GET", "/alerts", items_key="items"):
+# With a per-resource items key; `data` is read first, so this works on
+# either response shape
+for config in client.paginate("GET", "/alerts/configs", items_key="configs"):
+    print(config["id"])
+
+# Endpoints that declare `offset` rather than `page`
+for alert in client.paginate("GET", "/models/alerts", items_key="alerts",
+                             param_style="offset"):
     print(alert["id"])
 ```