Bug type
Regression (worked before, now fails)
Beta release blocker
No
Summary
Reproducing consistently on 2026.4.11 in sessions with large build/deploy workflows (15-30+ consecutive tool calls: exec, read, write, edit).
Previously reported in: #24800, #5433 — both closed, appears to be a regression.
Steps to reproduce
Start a fresh OpenClaw session with compaction.mode: safeguard and anthropic/claude-sonnet-4-6 as primary model
Run a heavy build/deploy workflow involving 15-30+ consecutive tool calls (exec, read, write, edit) without user messages between them
Monitor context via /status or session_status tool
Observe context climbing past 100% of model limit with Compactions: 0
Expected behavior
Compaction fires proactively when context approaches the limit (e.g. at contextWindow - reserveTokensFloor), before the API returns an overflow error.
Actual behavior
Compaction never triggers during tool-use loops. Context grew from ~150k to 287k (143% over the 200k limit) with zero compactions. Manual wake event partially resolved it but auto-compaction never fired on its own.
OpenClaw version
2026.4.11
Operating system
macOS 15.2 (Darwin 25.2.0, arm64)
Install method
No response
Model
anthropic/claude-sonnet-4-6
Provider / routing chain
anthropic plugin, direct API key (no proxy)
Additional provider/model setup details
compaction.mode: safeguard
reserveTokensFloor: unset initially, set to 20000 as workaround
compaction.model: anthropic/claude-sonnet-4-6
compaction.memoryFlush.enabled: true
Logs, screenshots, and evidence
Impact and severity
High — heavy tool-use sessions (build/deploy workflows) become progressively slower and eventually require a manual session reset. Affects any workflow with sustained tool-use loops.
Additional information
Previously reported in #24800 and #5433, both closed. Appears to be a regression in 2026.4.11. Workaround: setting reserveTokensFloor: 20000 reduces severity but does not fix the root cause (reactive vs proactive check in run.ts).
Bug type
Regression (worked before, now fails)
Beta release blocker
No
Summary
Reproducing consistently on 2026.4.11 in sessions with large build/deploy workflows (15-30+ consecutive tool calls: exec, read, write, edit).
Previously reported in: #24800, #5433 — both closed, appears to be a regression.
Steps to reproduce
Start a fresh OpenClaw session with compaction.mode: safeguard and anthropic/claude-sonnet-4-6 as primary model
Run a heavy build/deploy workflow involving 15-30+ consecutive tool calls (exec, read, write, edit) without user messages between them
Monitor context via /status or session_status tool
Observe context climbing past 100% of model limit with Compactions: 0
Expected behavior
Compaction fires proactively when context approaches the limit (e.g. at contextWindow - reserveTokensFloor), before the API returns an overflow error.
Actual behavior
Compaction never triggers during tool-use loops. Context grew from ~150k to 287k (143% over the 200k limit) with zero compactions. Manual wake event partially resolved it but auto-compaction never fired on its own.
OpenClaw version
2026.4.11
Operating system
macOS 15.2 (Darwin 25.2.0, arm64)
Install method
No response
Model
anthropic/claude-sonnet-4-6
Provider / routing chain
anthropic plugin, direct API key (no proxy)
Additional provider/model setup details
compaction.mode: safeguard
reserveTokensFloor: unset initially, set to 20000 as workaround
compaction.model: anthropic/claude-sonnet-4-6
compaction.memoryFlush.enabled: true
Logs, screenshots, and evidence
Impact and severity
High — heavy tool-use sessions (build/deploy workflows) become progressively slower and eventually require a manual session reset. Affects any workflow with sustained tool-use loops.
Additional information
Previously reported in #24800 and #5433, both closed. Appears to be a regression in 2026.4.11. Workaround: setting reserveTokensFloor: 20000 reduces severity but does not fix the root cause (reactive vs proactive check in run.ts).