Repository navigation
Depend on Instructor as a range and make every dependency floor true - #310
Merged
Merged
Conversation
`instructor.core.client.Instructor` relied on the submodule being reachable as an attribute of the top-level package. Instructor 1.17.0 no longer exposes `instructor.core` that way, so attribute access raises AttributeError, while `instructor.Instructor` and `instructor.AsyncInstructor` are the same classes on 1.14.5 and 1.17.0. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The exact pin dates from 2a53b60 (instructor 1.9.2, upstream issue 567-labs/instructor#1796) and was bumped by hand ever since, which left every user stuck on whatever version we last touched. Full suite passes on 1.17.0 (618 passed, 3 skipped) after the client class rename in the previous commit. Example projects and the legacy requirements.txt move to the same range so the workspace resolves. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The MCP schema transformer passes both __config__ and __base__ to create_model. pydantic allowed that combination in 2.11.4 (pydantic PR #11714); 2.11.3 and below raise PydanticUserError create-model-config-base. Resolving every direct dependency at its declared lowest version put pydantic 2.11.0 in the environment and failed 150 MCP tests with that error, so the old floor was never true. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The MCP connectors import mcp.client.streamable_http, which first shipped in mcp 1.8.0 (May 2025). Resolving every direct dependency at its declared lowest version installed mcp 1.6.0 and five connector test modules failed to import. With mcp 1.8.0, pydantic 2.11.4, openai 2.0.0, litellm 1.53.1 and instructor 1.17.0 the suite passes. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ency Declared floors were never exercised: pydantic 2.11.0 and mcp 1.6.0 both failed the suite although pyproject allowed them. This job resolves the runtime dependencies with uv's lowest-direct strategy, installs current test tooling on top, and runs the suite. Locally that is instructor 1.17.0, openai 3.3.0, pydantic 2.11.4, mcp 1.8.0, litellm 1.53.1: 618 passed, 3 skipped. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…dule Instructor 1.15.3 moved its response-model wrapper to instructor.v2.core.function_calls. BaseIOSchema only exempted the two older module names from the docstring rule, so on Instructor 1.17.0 every agent.run() and run_async() raised "must have a non-empty docstring" before the request was sent. Streaming survived because Partial models carry from_streaming_response. The mocked suite never saw it; a new test module drives a real Instructor client through an httpx MockTransport and fails on the old check (3 of 4 cases). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Instructor 1.17 forwards a caller's strict setting into
model_from_chunks, and the streaming patch passed its own strict=None
on top, so run_stream with model_api_parameters={"strict": True}
raised "got multiple values for keyword argument 'strict'". The patch
now only supplies the default when the caller did not.
Covered by test_run_stream_honors_a_caller_strict_setting.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Instructor 1.14.5 used stop_after_attempt(max_retries); 1.17.0 uses stop_after_attempt(max_retries + 1) (v2/core/retry.py:275). The framework does not set max_retries, but the deployment and error-handling guides pass it through model_api_parameters. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Eigenwise
force-pushed
the
chore/instructor-range
branch
from
October 18, 2026 20:08
00879f9 to
1d8ef5a
Compare
This was referenced Oct 19, 2026
Merged
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Targets main after the 2.12.1 release. Includes the released patch fixes and six real-Instructor transport regressions.
Why
instructor==1.14.5was an exact pin bumped by hand since 1.9.2 (upstream 567-labs/instructor#1796: an import path removed in a patch release). Every user was stuck on whatever version we last touched. The mocked suite passed on 1.17.0 after five import renames, but a real-transport probe showed two runtime breaks the mocks hide, and testing the declared floors turned up two dishonest bounds.What
instructor>=1.17.0,<2in the root project, every example and the legacyrequirements.txt; the framework referencesinstructor.Instructor/instructor.AsyncInstructorfrom the package root (same classes on 1.14.5 and 1.17.0;instructor.coreis not an attribute path on 1.17.0).instructor.v2.core.function_calls;BaseIOSchemaonly exempted the two older module names from the docstring rule, so everyagent.run()/run_async()raisedValueError: ... must have a non-empty docstringbefore the request was sent. The check now accepts any Instructor-generated class.strictintomodel_from_chunks; the streaming patch passed its ownstrict=Noneon top, sorun_streamwithmodel_api_parameters={"strict": True}raised "multiple values for keyword argument 'strict'". The patch now usessetdefault.tests/agents/test_instructor_transport.pydrives a real Instructor client throughhttpx.MockTransport(run, run_async, run_stream, run_stream with strict). Without the two fixes 3 of 4 fail.pydantic>=2.11.4: the MCP schema transformer passes__config__with__base__tocreate_model; allowed since pydantic 2.11.4 (pydantic#11714), 2.11.3 raisescreate-model-config-base. Verified on both versions.mcp>=1.8.0,<2:mcp.client.streamable_httpfirst shipped in 1.8.0; on 1.6.0 five connector test modules fail to import.lowest-versions: resolves runtime deps withuv pip compile --resolution lowest-direct, installs current pytest tooling, runs the suite..release/2.13.0.mddraft; codebase map updated.Verification
Final integrated branch, after the 2.12.1 fixes:
Black and Flake8 pass across all four directories. No live provider calls were used.
Known, not in this PR
max_retriescounts one more attempt on 1.17: 1.14.5 usedstop_after_attempt(max_retries), 1.17.0 usesstop_after_attempt(max_retries + 1). The framework never sets it; users passing it throughmodel_api_parametersget one extra call. Noted in the 2.13.0 release notes.Mode.TOOLS_STRICT(referenced inatomic_agent.py) andMode.GENAI_TOOLS(two examples) emit DeprecationWarnings on 1.17 ("removed in v3.0"); fold into the mode-derivation work (OpenAI-compat: AtomicAgent Mode.JSON on factory and AgentConfig, not default tools mode #282/fix(agents): derive AgentConfig mode from client and warn on mismatch #293).litellm
<2caps openai<3, so installs with litellm stay on openai 2.x even though Instructor 1.17 allows 3.x.The streaming monkeypatch's original reason (trailing-strings partial parsing) is implemented upstream since 1.14; dropping the patch is possible but changes validation behavior and needs its own change.
🤖 Generated with Claude Code