SDK Migration Guide

Migrating from 0.x to 1.0

Version 1.0 aligns the Python SDK's response types with the API wire format. Response models are now Pydantic BaseModel instances instead of dataclasses, and their fields match the JSON returned by the Subconscious API 1:1.

Summary of breaking changes

• Usage is now a flat Pydantic model (inputtokens, outputtokens, durationms). The old models and platformtools fields are removed.
• ModelUsage and PlatformToolUsage are deleted entirely.
• ReasoningNode is removed. Use ReasoningTask instead.
• ReasoningTask.subtask is renamed to subtasks.
• ReasoningTask.tooluse is now Optional[AgentToolUse] instead of List[Any].
• RunResult.reasoning is now Optional[List[ReasoningTask]] instead of Optional[ReasoningNode].
• Run has a new error: Optional[RunError] field for failed runs.
• Run, RunResult, Usage, RunError, ReasoningTask, and AgentToolUse are Pydantic BaseModel instead of @dataclass.
• Engine literal expanded: added "tim-edge", "tim-oss-local", "tim-1.5", "tim-gpt-heavy-tc".

Usage — before and after

# 0.x — broken; models/platform_tools were raw dicts, not dataclass instances
run.usage.models[0].input_tokens
run.usage.platform_tools[0].tool_id

# 1.0 — flat, matches API wire format
run.usage.input_tokens    # int
run.usage.output_tokens   # int
run.usage.duration_ms     # Optional[int]

ReasoningTask — before and after

# 0.x
from subconscious import ReasoningNode
node.subtask       # List[ReasoningNode]
node.tooluse       # List[Any]

# 1.0 — ReasoningNode is removed
from subconscious import ReasoningTask
task.subtasks      # Optional[List[ReasoningTask]]
task.tooluse       # Optional[AgentToolUse]
task.tooluse.tool_name       # str
task.tooluse.tool_call_id    # Optional[str]
task.tooluse.parameters      # Dict[str, Any]
task.tooluse.tool_result     # Optional[Any]

RunResult.reasoning — before and after

# 0.x — single node
run.result.reasoning  # Optional[ReasoningNode]

# 1.0 — list of tasks (matches API)
run.result.reasoning  # Optional[List[ReasoningTask]]
for task in run.result.reasoning or []:
    print(task.title, task.thought)

Run.error — new field

# 1.0 — error details on failed runs
if run.status == "failed" and run.error:
    print(f"Error {run.error.code}: {run.error.message}")

Removed exports

# 0.x — these imports will fail in 1.0
from subconscious import ModelUsage       # REMOVED
from subconscious import PlatformToolUsage  # REMOVED
from subconscious import ReasoningNode    # REMOVED

# 1.0 — new exports
from subconscious import AgentToolUse
from subconscious import RunError
from subconscious import ReasoningTask

Dataclass to Pydantic

Attribute access is unchanged (run.runid, run.usage.inputtokens), but isinstance checks and construction patterns differ:

# 0.x
from dataclasses import is_dataclass
assert is_dataclass(run)

# 1.0
from pydantic import BaseModel
assert isinstance(run, BaseModel)

# Construction still works the same way:
run = Run(run_id="run_abc", status="queued")

Engine literal

# 0.x
Engine = Literal["tim", "tim-claude", "tim-claude-heavy"]

# 1.0
Engine = Literal[
    "tim", "tim-edge", "tim-claude", "tim-claude-heavy",
    "tim-oss-local", "tim-1.5", "tim-gpt-heavy-tc",
]

Quick find-and-replace patterns

reasoning_format removed

The reasoningformat input field (aliased as reasoningFormat on the wire) is gone. If you were shaping the reasoning trace with a JSON/Pydantic schema, fold that guidance into your instructions or into answerformat instead — the agent's final output is the contract, and the reasoning trace is best treated as a read-only byproduct.

# Before
client.run(
    engine='tim',
    input={
        'instructions': '...',
        'answerFormat': AnswerSchema,
        'reasoningFormat': ReasoningSchema,
    },
)

# After
client.run(
    engine='tim',
    input={
        'instructions': '...',
        'answerFormat': AnswerSchema,
    },
)

Requests that still include reasoningFormat are rejected by the API.

RunResult.parsed_answer

result.answer is always a str on the wire, even when answerFormat is supplied — the API JSON-encodes the structured value. The SDK now attaches a parsed_answer companion field on every response that runs through the client (run, get, wait, cancel), populated via a best-effort json.loads of answer.

from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int

run = client.run(
    engine='tim',
    input={
        'instructions': 'return JSON for a person',
        'answerFormat': Person,
    },
    options={'await_completion': True},
)

run.result.answer         # '{"name":"ada","age":36}'  (raw string)
run.result.parsed_answer  # {'name': 'ada', 'age': 36}  (decoded dict)

parsedanswer is typed as Any — validate with your Pydantic model of choice (Person.modelvalidate(run.result.parsed_answer)) if you want a typed instance. It is None when answer is empty or not valid JSON.