openclaw - 💡(How to fix) Fix Control UI: context token counter inflates to >4× window (claude-cli) — cacheRead summed as fresh input [1 participants]

Summary

The Control UI context/token indicator reports values far exceeding the model's context window (observed: ~800k shown against a 200k Opus 4.7 window) after a handful of cache-hit turns on the claude-cli provider path.

Version

• OpenClaw 2026.4.19-beta.1 (4862d34)
• Primary model: claude-cli/claude-opus-4-7
• Host: macOS 15.x, Node 25.4.0

Root cause

dist/status-D-R7Q_A-.js contains the formula:

totalTokens ?? input + cacheRead + cacheWrite

cacheRead represents tokens the model read from the prompt cache on a given turn — it is already part of the cached prefix, not new context being added on top. Summing input + cacheRead + cacheWrite turn after turn produces a running total rather than the actual window-resident token count.

Repro (minimal)

With a heavy system prompt + plugins (e.g. lossless-claw, openclaw-mem, memory-wiki):

openclaw agent --json --agent main -m "ping" --timeout 60

Observed usage for a 6-token user message: input: 6, output: 6, cacheRead: 20566, cacheWrite: 35458.

Four–five such turns and the UI indicator exceeds 200k despite the true live prompt staying well under the window.

Relationship to #64669

Fix #64669 landed in 2026.4.19-beta.1 for the Codex path ("stop cumulative app-server token totals from being treated as fresh context usage"). The same class of bug exists on the claude-cli path and was not addressed.

Suggested fix

Either:

• Display input + cacheWrite (cacheRead is not additive), or
• Surface cache_read_input_tokens as a separate read-only metric and exclude it from the contextTokens/percent calculation.

Impact

Operators cannot rely on the UI counter to decide whether to compact / rotate sessions; lossless-claw's contextThreshold is judged against a grossly inflated figure.