Cairn

Preserve think blocks in multi-turn history replay

NousResearch/hermes-agent · cluster 20577 · agent

Data as of

4 open items. Fix attempted, not merged.

The short answer

1open issue
3open PRs
7items grouped
Known?
Yes. 7 related items grouped here, 1 issue already closed.
Fix event
No fix has been merged to main. 2 PRs closed without merging; see the attempts below.
Tracker
The maintainers' tracker marks this unresolved. That is separate from the fix event.
Still open
1 issue and 3 PRs.

These issues and PRs address a problem where <think> blocks are stripped from assistant message history when replaying to vLLM and custom providers, losing reasoning context across turns. Work focuses on detecting, preserving, and re-embedding thinking content during history replay for these providers.

Open issues1

  • Issue#20577[Bug]: <think> blocks stripped from assistant history on replay to vLLM/custom provider — degrades multi-turn reasoning

Open PRs3

  • PR#20607fix: preserve <think> blocks in content for vLLM/custom provider replay
  • PR#56019fix(agent): preserve Qwen3.6 / vLLM preserve_thinking history (#56004)
  • PR#56020fix(agent): replay preserved thinking for custom providers

2 PRs closed without merging

From the close comments: done elsewhere 1 · other 1

  • #20609 fix(agent): re-embed <think> blocks inline for vLLM/custom providers on history replay (#20577) · done elsewhere
  • #58380 feat(custom-provider) Support opt-in preserve thinking · other
Closed issues · 1
  • Closed#56004Thinking models lose their own reasoning between tool calls and turns on OpenAI-compatible endpoints (reasoning stripped on replay)
How this cluster was grouped
  • Link evidence (partly structural links): some members are connected by structural evidence (shared changed files, or a closing/superseded reference); the rest are joined by an automated relevance judgement over their text.
  • Members outside that structural core may belong to a different underlying problem.
  • Pairwise link density 0.26: the share of item pairs that carry a link, counting a structural link in full and a relevance-only link at 0.36. 1.00 means every pair is linked.

Pick up a Lab task

Check whether a fix covers the open issues here, with your own agent. Bring back evidence for maintainers.

Open the Lab

At a glance

Subsystem
agent
Closed issues
1
Attempts
2 PRs closed without merging

For your agent

The same facts, typed.

cluster/20577.json
cluster/20577.md