Migrating to SDK 0.6

SDK 0.6 is a surgical break — no back-compat with 0.5.x callers or pre-0.6.0 servers. Upgrade both.

1. Minimum server version

The SDK requires tappass server ≥ 0.6.0. Older servers are refused at client construction.

# 0.5: works against any server
Agent("http://tappass", api_key="tp_x")

# 0.6: probes /health; raises TapPassConfigError if server < 0.6.0
Agent("http://tappass", api_key="tp_x")

Upgrade the server first.

2. Correlation IDs on responses

ChatResponse now has five required fields populated by the server.

# 0.5: response only has content, pipeline, usage
r = agent.chat("...")

# 0.6: correlation IDs always present
r = agent.chat("...")
r.session_id, r.task_id, r.agent_uuid, r.pipeline_uuid, r.audit_url

If you were defensively accessing these via .get(), simplify — they’re guaranteed.

3. Framework guards are context managers

CrewAI

# 0.5: mutated crew in place
from tappass.integrations.crewai import guard_crew
guarded = guard_crew(crew, constraints={...}, tappass_url=..., api_key="tp_x")
result = guarded.kickoff()   # caller drives the crew

# 0.6: context manager; session open/close is explicit
from tappass.integrations.crewai import guard_crew
with guard_crew(crew, tappass_url=..., api_key="tp_x",
                constraints={...}) as session:
    result = session.kickoff()
    print(session.id, session.audit_url)

Calling guard_crew(crew, ...) without with now raises TypeError. No shim.

LangChain

# 0.5: guard_tools only
tools = guard_tools(tools, tappass_url=..., api_key="tp_x")

# 0.6: high-level guard_agent for whole executor; guard_tools still
# available for the low-level case.
with guard_agent(executor, tappass_url=..., api_key="tp_x") as session:
    result = session.invoke({"input": "..."})

OpenAI, Google ADK, MCP, Temporal, A2A, FastAPI

Each follows the same pattern. See the runtime-governance reference for the full matrix.

4. Typed GovernanceDecision exceptions

PolicyBlockError stayed, but it’s now part of a richer tree. The server can return six different decisions:

from tappass import (
    GovernanceDecision,
    PolicyBlockError,
    RedactionApplied,
    ApprovalRequired,
    TrustTierDenied,
    BreakGlassActive,
    ToolIntegrityViolation,
)

try:
    r = agent.chat(prompt)
except PolicyBlockError as e:
    ...                       # hard block — treat as failure
except ApprovalRequired as e:
    print(e.approval_url)     # pause, resume via dashboard
except TrustTierDenied as e:
    ...                       # agent clearance too low
except GovernanceDecision as e:
    ...                       # any other decision

If you previously caught Exception or relied on HTTP response codes, rewire to these subclasses.

5. TapPass admin class stripped

The TapPass client no longer has pipelines, create_pipeline, assign, audit, etc. Those moved to the dashboard.

0.5

tp.create_pipeline("strict", steps={...})
tp.assign("my-agent", "strict")
events = tp.audit(agent_id="my-agent")

# 0.6 — removed. Use the web dashboard.
tp.health()                     # still here
tp.agent(agent_uuid)            # still here (lookup)
tp.register_agent("svc", ...)   # still here (self-registration on startup)

6. Reporter emits structured AuditEvent

Callers building custom integrations against the reporter need to enqueue AuditEvent instances instead of raw dicts.

0.5

get_reporter().enqueue(url, key, {"tool": "search", "duration_ms": 12})

# 0.6
from tappass.types import AuditEvent
from datetime import datetime, timezone

get_reporter().enqueue(url, key, AuditEvent(
    event_type="tool_call",
    session_id="", task_id=None, agent_uuid="",
    pipeline_uuid=None,
    timestamp=datetime.now(tz=timezone.utc),
    payload={"tool": "search", "duration_ms": 12, "outcome": "ok"},
))

Empty session_id/agent_uuid are filled from the active session.