openclaw - 💡(How to fix) Fix Telegram plugin-state cache writes rows with `expires_at = NULL`, hits 1000-row hard cap, migrations fail

Summary

Doctor in 2026.5.26 reports:

PluginStateStoreError: Plugin state for telegram exceeds the 1000 live row limit.
Failed migrating Telegram forum topic-name cache entry -1003734409696:1230

Root cause: telegram.message-cache namespace writes rows with expires_at = NULL, so the expiry sweep (idx_plugin_state_expiry WHERE expires_at IS NOT NULL) never evicts them. They accumulate until the per-plugin 1000-row cap, then any new write (including the doctor's topic-name-cache migration) fails.

Evidence

~/.openclaw/plugin-state/state.sqlite, current live state:

Within hours of a manual prune this morning the cache refilled to 998/1000.

Sample row:

namespace=telegram.message-cache
entry_key=eafd015eb3f678ca29b14de8:default:-1003734409696:17179
value_json={"sourceMessage":{"message_id":17179, ...}}
created_at=1779886069586
expires_at=NULL

Environment

• openclaw 2026.5.26 (10ad3aa)
• node 26.0.0 (homebrew)
• macOS 25.5.0

Impact

• Doctor migrations on upgrade silently lose data (topic-name-cache entries dropped when cap is hit during the rewrite).
• Per-plugin 1000-row ceiling reached forever for any Telegram instance with more than ~1000 cached inbound messages.
• Manual SQLite prune is the only mitigation, and it refills within hours.
• topic-name-cache.<hash> namespace appears partially populated after the failed migration — only 1 row where the migration source had more.

Suggested fix direction

Either:

• Set a TTL on telegram.message-cache writes (e.g. 24h / 7d, matching how long inbound message context is realistically useful), and let pruneExpiredPluginState evict them, OR
• Cap telegram.message-cache namespace size explicitly with LRU eviction on insert, decoupled from the global 1000-row plugin cap, OR
• Raise the 1000-row cap (less ideal — same growth pattern will hit any new ceiling).

Side note

Filed alongside #87331 (Native hook relay unavailable regression). Both surfaced on the same upgrade; this one looks design-level rather than a regression.