Summary

Upgrading from 2026.2.14 → 2026.2.15 causes prompt cache invalidation on every turn for local model providers. The cache is trimmed/rebuilt from scratch on every message, even when no workspace files have changed and the conversation is short. Downgrading to 2026.2.14 immediately restores normal cache behavior.

Steps to reproduce

• OpenClaw: 2026.2.15 (3fe22ea) — regression confirmed; 2026.2.14 works correctly
• OS: macOS 26.3 (arm64), Mac Studio M3 Ultra 512GB
• Channel: Telegram (DM, not a group chat)
• Models tested:
  • MiniMax-M2.5 via LM Studio 0.4.2+2 (MLX backend)
  • Qwen3.5-397B-A17B via llama-server (llama.cpp built from latest main)

Reproduction

1. Install [email protected]
2. Start a fresh conversation via Telegram DM with a local model (LM Studio or llama-server)
3. Send a simple message (e.g., "hi")
4. Wait for the response
5. Send another simple message (e.g., "what are you up to?")
6. Observe: the entire prompt cache is invalidated and rebuilt from scratch

Expected behavior

The second message reuses the cached prompt prefix. Only new tokens (the user's latest message + assistant response) are processed. LM Studio shows no cache trim; llama-server shows restored context checkpoint.

Actual behavior

Every message triggers a full cache rebuild:

LM Studio (MiniMax-M2.5):

[cache_wrapper][INFO]: Trimmed 21283 tokens from the prompt cache
[minimax-m2.5] Prompt processing progress: 0.0%
[minimax-m2.5] Prompt processing progress: 2.4%
...

The trim amount is nearly identical every turn (~21,242–21,283 tokens) regardless of message content, even on a conversation under 30K tokens total. No workspace files were modified between messages.

llama-server (Qwen3.5-397B):

slot update_slots: id 3 | task 13076 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id 3 | task 13076 | prompt processing progress, n_tokens = 2048, ...

Full cache wipe (memory_seq_rm [0, end)) on every turn. Context checkpoints are erased and never restored.

Verification

# Downgrade — cache works immediately
npm install -g [email protected]

# Upgrade — cache broken again
npm install -g [email protected]

No configuration changes, no workspace file edits, no LM Studio updates between tests. The only variable is the OpenClaw version.

Suspected cause

Two changes in 2026.2.15 modify per-turn prompt assembly:

1. Group chat context injection (#14447): "Group chats: always inject group chat context (name, participants, reply guidance) into the system prompt on every turn, not just the first." — If this triggers for DM conversations or if the injected content varies per turn, it would invalidate the cache.

2. Memory-flush Current time: line (#17603, #17633): "append a Current time: line to memory-flush turns" — The changelog claims this is "without making the system prompt time-variant," but if the time line leaks into normal (non-flush) turns, it would change the prompt every second.

Code references in pi-embedded-n26FO9Pa.js:

• Line 36106: timeLine: \Current time: ${formattedTime} (${userTimezone})``
• Line 23571: Build a persistent group-chat context block that is always included in the

Workaround

Downgrade to 2026.2.14:

npm install -g [email protected]

OpenClaw version

2026.2.15

Operating system

macOS

Install method

npm install -g [email protected]

Logs, screenshots, and evidence

Impact and severity

This is a severe performance regression for all local model users. Without prompt caching:

• Every message requires full prompt processing from scratch
• A 40K-token context takes ~200 seconds to process on llama-server (vs ~5 seconds with cache)
• LM Studio/MLX similarly affected — multi-minute waits for simple replies
• Makes local models essentially unusable for multi-turn conversations

Additional information

No response