Bug type
Regression (worked before, now fails)
Summary
Two related compaction issues: (1) safeguard mode actively trims context but Compactions counter in /status always stays at 0, making monitoring impossible. (2) default mode compaction triggers a large summarization API request that exceeds Google's 1M TPM limit, causing failover.
Steps to reproduce
Issue 1 (safeguard counter):
Set compaction.mode: "safeguard", contextTokens: 160000, maxHistoryShare: 0.4
Use google/gemini-3-flash-preview and have a continuous conversation
Observe /status as context grows past the limit and gets trimmed
Compactions counter remains 0 despite active trimming
Issue 2 (default TPM burst):
Switch compaction.mode to "default" with same settings
Have a conversation until context approaches the limit
Default compaction triggers a large summarization request
Request exceeds Google 1M TPM limit → 429 error → failover to backup model
Expected behavior
Safeguard: Compactions counter should increment when context is trimmed, or /status should show a separate indicator for safeguard trimming events
Default: Compaction should not send a request that exceeds TPM limits. At minimum, documentation should warn that default mode can cause TPM spikes with TPM-limited providers
Actual behavior
Safeguard: Context drops from 199k to 102k between turns (trimming works), but Compactions stays at 0 throughout. Observed during continuous conversation with no manual /compact or /new:
21:18 → 100k/160k (63%) · Compactions: 0
21:22 → 199k/160k (125%) · Compactions: 0
21:23 → 102k/160k (64%) · Compactions: 0
Default: Compaction immediately triggered a large API request that exceeded 1M TPM, caused 429 error, and forced failover to google/gemini-3.1-pro-preview
OpenClaw version
OpenClaw: 2026.3.2 (85377a2)
Operating system
Rocky Linux (Proxmox VM)
Install method
No response
Logs, screenshots, and evidence
"compaction": {
"mode": "safeguard",
"maxHistoryShare": 0.4
}
Model: google/gemini-3-flash-preview, contextTokens: 160000Impact and severity
safeguard mode is functionally working but unmonitorable
default mode is unusable with TPM-limited providers like Google AI Studio (1M TPM)
Users have no compaction mode that both works correctly and is monitorable
Additional information
Reverted to safeguard mode as workaround. Default mode caused immediate service disruption.
Bug type
Regression (worked before, now fails)
Summary
Two related compaction issues: (1) safeguard mode actively trims context but Compactions counter in /status always stays at 0, making monitoring impossible. (2) default mode compaction triggers a large summarization API request that exceeds Google's 1M TPM limit, causing failover.
Steps to reproduce
Issue 1 (safeguard counter):
Set compaction.mode: "safeguard", contextTokens: 160000, maxHistoryShare: 0.4
Use google/gemini-3-flash-preview and have a continuous conversation
Observe /status as context grows past the limit and gets trimmed
Compactions counter remains 0 despite active trimming
Issue 2 (default TPM burst):
Switch compaction.mode to "default" with same settings
Have a conversation until context approaches the limit
Default compaction triggers a large summarization request
Request exceeds Google 1M TPM limit → 429 error → failover to backup model
Expected behavior
Safeguard: Compactions counter should increment when context is trimmed, or /status should show a separate indicator for safeguard trimming events
Default: Compaction should not send a request that exceeds TPM limits. At minimum, documentation should warn that default mode can cause TPM spikes with TPM-limited providers
Actual behavior
Safeguard: Context drops from 199k to 102k between turns (trimming works), but Compactions stays at 0 throughout. Observed during continuous conversation with no manual /compact or /new:
21:18 → 100k/160k (63%) · Compactions: 0
21:22 → 199k/160k (125%) · Compactions: 0
21:23 → 102k/160k (64%) · Compactions: 0
Default: Compaction immediately triggered a large API request that exceeded 1M TPM, caused 429 error, and forced failover to google/gemini-3.1-pro-preview
OpenClaw version
OpenClaw: 2026.3.2 (85377a2)
Operating system
Rocky Linux (Proxmox VM)
Install method
No response
Logs, screenshots, and evidence
Impact and severity
safeguard mode is functionally working but unmonitorable
default mode is unusable with TPM-limited providers like Google AI Studio (1M TPM)
Users have no compaction mode that both works correctly and is monitorable
Additional information
Reverted to safeguard mode as workaround. Default mode caused immediate service disruption.