Assistant repeats identical output / identical tool calls until it hits the output token limit #59567

Description

Description

In a long-running session, the assistant repeatedly re-emits the same block of
text and/or re-issues the same tool call with identical arguments, without making
any progress. The repetition does not terminate on its own; it only stops when the
model reaches the provider's max output tokens, leaving the turn truncated. The
only way to continue is to abort and re-prompt.

This happened several times within a single session. It affects both plain-text
assistant output and tool-call sequences (the same tool, same arguments, repeated
back to back).

Environment

• opencode 2.1.4 (update to 2.2.0 offered at the time of the report)
• OS: Windows 11, build 10.0.26200.9711
• Terminal: Windows Terminal (WT_SESSION set), TERM=xterm-256color
• Shell: PowerShell 5.1.26100.9711
• Provider: custom, @ai-sdk/anthropic (Anthropic-compatible endpoint)
• Model: deepseek-v4.1-flash-1017
• Plugin: superpowers (loaded from a local directory)
• Session shape: one long session, 100+ tool calls, several large file reads and
  long tool outputs

Steps to reproduce

1. On Windows, start a long session that performs many sequential tool calls,
   including large file reads and long tool outputs.
2. Ask for a series of small edits and verifications spread over several files.
3. After a number of turns, the assistant begins repeating an identical paragraph,
   or re-issuing an identical tool call with identical arguments.
4. Observe that the repetition never resolves by itself; the turn ends only when
   the output token limit is reached.

Expected behaviour

• The assistant finishes the step, then either calls the next tool or ends the turn.
• A run of identical tool invocations (same tool name + same arguments) whose
  results have not changed should be detected and broken out of, ideally with a
  visible warning, instead of silently consuming the whole output budget.

Actual behaviour

• Identical text blocks and identical tool calls repeat many times inside one
  assistant message until the output budget is exhausted.
• The turn is left truncated mid-step, so the user has to abort manually.
• Nothing is written to opencode.log: the log contains only spawning process
  entries for the tool calls, with no finish reason, no output token count, and no
  truncation marker. The failure is therefore invisible in diagnostics.

Impact

• Blocks progress: the session cannot continue without a manual abort.
• Each occurrence burns a full output budget in tokens, time and cost.
• Hard to triage: no trace in the log, and the behaviour is not tied to a single
  tool, so it is not obvious which layer (model provider, streaming, or agent loop)
  is responsible.

Additional context / suggestions

• Log finish reason (stop / length / abort) plus output token counts per
  assistant message, so output-limit truncation is diagnosable after the fact.
• Consider a guard in the agent loop for N consecutive identical tool invocations
  (same name + same arguments) or near-identical repeated text within a single
  assistant message; cut the stream and surface a warning instead of letting it run
  to the token limit.
• It is unclear whether the root cause is provider-side (repeated sampling) or the
  client re-sending/merging streamed parts incorrectly; the missing finish-reason
  logging makes it impossible to tell from the outside.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot

No response

Operating System

No response

Terminal

No response