Repository navigation
Description
Summary
Agent.acontinue_run (and the streaming variant) re-runs its continuation preparation on every retry attempt: the run is re-forked/truncated via _apply_continue_modifiers, dependencies are re-resolved, and the caller's input is re-appended by _maybe_append_input_message. With retries >= 1, a transient model error on the first attempt therefore appends the caller's instruction twice (visible in the transcript sent to the model) and re-forks the run a second time.
The sync path already does this correctly: continue_run_dispatch performs fork/truncate, dependency resolution and input append once, before _continue_run's retry loop.
Impact
- The retried model call receives the continuation instruction duplicated (transcript:
['ORIGINAL QUESTION', 'NEW INSTRUCTION', 'NEW INSTRUCTION']). - The run is re-forked from the original run on every attempt.
- Dependency resolution (
aresolve_run_dependencies) invokes caller callables — with side effects (LLM/DB calls) — once per attempt, and rewritesrun_context.dependencies; per-attempt resolution can changerun_inputafter the first attempt.
Repro
Offline FailOnceModel (fails once with a transient error, then succeeds), retries=1, delay_between_retries=0, invoked via acontinue_run with an input argument; parameterized over stream=False/True and run_response= vs run_id=+session_id:
E AssertionError: assert ['ORIGINAL QUESTION', 'NEW INSTRUCTION'] == ['ORIGINAL QUESTION', 'NEW INSTRUCTION']
E Left contains one more item: 'NEW INSTRUCTION'
The retried model call sees the instruction duplicated; the sync continue_run path with the same setup sees it once.
Suggested fix
Set a run_prepared flag: run the fork/truncate + dependency resolution + input append once, before the retry loop, in both _acontinue_run and _acontinue_run_stream — matching continue_run_dispatch. Everything downstream of the preparation stays per-attempt, so a failure during preparation still retries the preparation itself.
Tests
libs/agno/tests/unit/agent/test_continue_run_retries.py (4 variants: stream on/off × run_response/run_id+session_id): asserts a single NEW INSTRUCTION in the transcript, that the retried model call sees one instruction, and forked_from_run_id == first.run_id.
- On base: 4 failed with the duplicate instruction.
- With the fix: 4 passed; neighboring run-loop suites 180 passed; full
unit/agent1129 passed (3 pre-existing Windows tmp-file failures reproduce at base); ruff clean. An independent review proved the large diff is whitespace-only reindentation plus exactly six inserted tokens.
Activity
gunjan-l commentedon Oct 18, 2026
Hi team, I would like to work on this issue. Could you please assign it to me and add the
readylabel? I have analyzed the root cause and have a reproduction and fix ready.