continue_from.py
"""Time-travel via /continue with continue_from.
Choose a message boundary and resume from there. For a COMPLETED run this
auto-forks into a new sibling run, preserving the "1 run = 1 model loop"
contract. The source run stays intact.
Four ways to express the same intent:
- ``continue_from="end"`` -> continue from the full transcript
- ``continue_from="last_user"`` -> symbolic boundary
- ``continue_from=K`` (int) -> exact message-index boundary
- ``regenerate=True`` -> friendly sugar for last response
When to use the numeric form: when the symbolic boundaries don't land where you
want. For example, dropping the last *three* messages (a tool batch + an
assistant reply) requires a specific index — count back from
``len(run.messages)`` and pass that as ``continue_from=K``.
To discover valid indices for a run, either inspect ``run.messages`` directly
or call the checkpoint timeline endpoint (see
``../18_checkpointing/03_checkpoint_endpoints.py``).
"""
import asyncio
from agno.agent import Agent
from agno.db.sqlite import SqliteDb
from agno.models.openai import OpenAIResponses
def get_population(city: str) -> str:
"""Mock population lookup."""
data = {"Paris": "2.1M", "Tokyo": "13.9M", "Lagos": "15.3M"}
return data.get(city, "unknown")
async def main() -> None:
agent = Agent(
name="travel-agent",
model=OpenAIResponses(id="gpt-5.4"),
db=SqliteDb(
session_table="checkpoint_demo",
db_file="tmp/checkpoint_time_travel.db",
),
checkpoint="tool-batch",
tools=[get_population],
)
first = await agent.arun(input="What is the population of Paris?")
print("First run completed")
print(" run_id:", first.run_id)
print(" message count:", len(first.messages or []))
print(" content:", first.content)
print()
# Continue from the end: this is the default. Completed runs auto-fork, so
# the original Paris run is preserved while this follow-up becomes a sibling.
follow_up = await agent.acontinue_run(
run_id=first.run_id,
session_id=first.session_id,
continue_from="end",
input="Now compare that with Lagos.",
)
print("After /continue with continue_from='end' + follow-up")
print(" run_id:", follow_up.run_id, "(new sibling)")
print(" forked_from_run_id:", follow_up.forked_from_run_id)
print(" content:", follow_up.content)
print()
# Rewind to just after the last user message and ask something different.
# Completed runs auto-fork, so the original Paris run is preserved.
rewound = await agent.acontinue_run(
run_id=first.run_id,
session_id=first.session_id,
continue_from="last_user",
input="Actually, what is the population of Tokyo instead?",
)
print("After /continue with continue_from='last_user' + input='Tokyo'")
print(" run_id:", rewound.run_id, "(new sibling)")
print(" forked_from_run_id:", rewound.forked_from_run_id)
print(" message count:", len(rewound.messages or []))
print(" content:", rewound.content)
print()
# Numeric form: continue_from=K (int) addresses an exact message boundary.
# Use this when "end" / "last_user" don't land where you need to rewind
# to — for example, dropping the last tool batch in addition to the
# assistant reply.
print("Messages in first run (for picking an index):")
for i, m in enumerate(first.messages or [], start=1):
preview = (m.content or "")[:60].replace("\n", " ")
print(f" [{i}] {m.role}: {preview}")
print()
# Keep only the first message (the original user question) and resume with
# a totally different prompt. Demonstrates the K=1 boundary.
rewound_to_index = await agent.acontinue_run(
run_id=first.run_id,
session_id=first.session_id,
continue_from=1,
input="Instead, what is the population of Tokyo?",
)
print("After /continue with continue_from=1 (drop everything past msg 1)")
print(" run_id:", rewound_to_index.run_id, "(new sibling)")
print(" forked_from_run_id:", rewound_to_index.forked_from_run_id)
print(" forked_from_message_index:", rewound_to_index.forked_from_message_index)
print(" content:", rewound_to_index.content)
print()
# Verify: all paths coexist in the session.
session = agent.db.get_session(session_id=first.session_id, session_type="agent")
print(f"Runs in session: {len(session.runs or [])} (source preserved, forks added)")
if __name__ == "__main__":
asyncio.run(main())
Run the Example
1
Set up your virtual environment
uv venv --python 3.12
source .venv/bin/activate
uv venv --python 3.12
.venv\Scripts\activate
2
Install dependencies
uv pip install -U agno openai sqlalchemy
3
Export your OpenAI API key
export OPENAI_API_KEY="your_openai_api_key_here"
$Env:OPENAI_API_KEY="your_openai_api_key_here"
4
Run the example
Save the code above as
continue_from.py, then run:python continue_from.py