[Bug] Agent output enters an infinite repetition loop that runs for ~10 minutes with no circuit breaker — only a manual abort stops it

[Codingendless]

Summary

When an agent's context gets very large (long-running session plus a broad "verify everything after an upgrade" task), the model can enter a degenerate repetition loop: it streams the same sentence over and over as assistant text chunks. DSH has no repetition guard or circuit breaker, so the loop runs until max_tokens is hit or the user manually aborts. In our case it streamed about 420,000 characters of identical text over 10 full minutes.

Environment

• DSH: 0.1.0-rc.7 (via @deepseek-ai/dsh)
• Provider / model: volcengine/deepseek-v4-flash
• Session event log: DSH session JSONL (.jsonl.zstd, concatenated zstd frames)

Reproduction (what we observed in the session log)

The archived session (a plugin project workspace) had:

1. User: "DSH 升级了，我还适配吗" (DSH was upgraded, is my plugin still compatible?)
2. The agent worked through 20 normal steps (read repo files, grep DSH source under node_modules, inspect profile config) - all fine.
3. At step 21, after about 34 normal streamed chunks, the model locked into repeating one sentence:

The `register` API matches. Let me check the host-side `Service` and `webServer` APIs in rc.7.

4. That sentence was repeated across 3015 text-chunks events (about 2825 repetitions), totalling about 420,000 characters, streamed for 10 minutes (21:54:34 -> 22:04:23) until the user manually aborted the turn (aborted / user).

Evidence in the log: the step produced text-chunks events forever, with no finish chunk, no stop reason, and no assistant/message - just endless identical text deltas.

A related failure appeared in the same project's other session: turn 23 sent 467,456 input tokens and immediately hit max-tokens with 1 output token (empty text). Both failures point to context saturation plus model degeneration, with no harness-level protection.

Root cause (from code inspection)

Searching the DSH and pi-ai stack for frequency_penalty, presence_penalty, repetition_penalty, or any "N consecutive identical chunks - abort" logic returns nothing. The only "repeat" handling in dsh-agent is about status transitions and dispatcher dedup - nothing about degenerate model output.

So the only exit conditions are:

• the provider's max_tokens (not set tight enough here), or
• a manual user abort.

Neither is a reasonable guardrail for a stuck model.

Suggested fix (any of, ideally a combination)

1. Circuit breaker on repeated output: detect N consecutive identical text-chunks (or the same sentence repeated M times) and automatically abort the current step with a clear reason, similar to how other agent loops are guarded.
2. Expose repetition penalties in the provider request config (frequency_penalty / presence_penalty / repetition_penalty passthrough) so providers that support it can suppress degeneration.
3. Cap per-step output tokens by default so a stuck loop terminates in seconds rather than minutes.
4. Context compaction for very long sessions before the context approaches saturation.

Notes

• The session files are stored as concatenated zstd frames (see scanZstdFrames in dsh-session-persistence-jsonl); a single zstdDecompressSync on the whole file only decodes the first frame, so tooling that analyzes session logs must iterate frames.

[tianhao8687]

确认当前只有“重复工具调用提醒”，没有针对模型文本流退化重复的保护。这个行为不应直接写进 agent-loop；按现有架构，更适合做一个可配置的 llm/stream middleware 插件。

建议实现独立的 output-repetition guard：

1. 对 text 与 reasoning 分开维护规范化滚动窗口，检测与 chunk 边界无关的重复周期/重复 n-gram；不能只比较连续 chunk，因为 provider 可任意切片。
2. 设置 minChars、窗口大小、重复次数、单步最大字符/时长等可配置阈值，并在达到最低输出量前不触发，降低代码表格、测试日志等合法重复的误报。
3. 触发时主动结束上游 iterator，并输出唯一的 terminal error finish，例如 REPETITIVE_OUTPUT。Agent loop 对错误流会丢弃未闭合的 partial message，因此不会把 420K 坏文本存成 assistant/message；原始 chunks 仍可用于诊断。
4. 不自动重试同一请求，否则很可能再次计费并重复退化；由正常 request-error/retry policy 决定是否换模型或结束。

Provider 的 frequency/presence/repetition penalty 只能作为可选能力，DeepSeek 思考模式和许多兼容网关并不保证支持，不能替代 harness 级熔断。有限的 adapter 默认 maxTokens 与提前 compaction 也应保留，但两者只是第二道防线。

回归测试应覆盖：同一句话采用不同 chunk 切分仍触发、长重复字符/句子、合法重复代码不触发、触发后 upstream return() 被调用、只有一个 error finish、无 assistant/message 持久化，以及用户手动 abort 仍保持最高优先级。