Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

OpenClaw fails to parse the usage field from API responses, causing token counts to display as N/A across the Control UI, the usage page to omit token statistics, and the auto-compression mechanism to fail — resulting in context overflow.

Steps to reproduce

1. Update OpenClaw to version 2026.4.9 or later (up to 2026.4.11)
2. Open Control UI (http://127.0.0.1:18789/)
3. Navigate to Sessions view — observe the "Tokens" column shows N/A
4. Navigate to Usage page — observe only message counts are tracked, token counts are missing
5. Continue chatting in a session until context window is exceeded — observe auto-compression does not trigger and context overflow occurs

Expected behavior

• Token usage should display actual consumed token counts and percentages (e.g., "10.5k/262k (4%)"), as observed in versions prior to 2026.4.9 which showed correct usage statistics.
• The Usage page should include token consumption statistics alongside message counts.
• Auto-compression should trigger when context window approaches capacity, preventing context overflow.

Actual behavior

• Sessions view: Tokens column displays "unknown/262k (?%)" — the consumed token count shows "unknown", the percentage shows a question mark, and only the context window size (262k) appears correctly. On the web UI, tokens show as "N/A" (不适用).

│ agent:main:main                  │ direct │ 1m ago  │ mlx-qwopus3.5-27b-v3 │ unknown/262k (?)%   │
│ agent:main:dreaming-narrative…  │ direct │ 2h ago  │ mlx-qwopus3.5-27b-v3 │ unknown/262k (?)%   │

• Usage page: Only message counts are displayed; token consumption statistics are entirely missing. Previous versions included correct token usage records.
• API parsing failure: The upstream API response includes a standard-compliant usage field, but OpenClaw appears unable to correctly parse it.
• Auto-compression failure: The automatic context compression mechanism does not trigger, leading to context overflow when the context window is exceeded.
• Manual compression workaround: After manually triggering compression, token usage temporarily displays correctly, but immediately reverts to N/A upon the next chat message.

OpenClaw version

2026.4.9 (first affected) — 2026.4.11 (still affected)

Operating system

macOS 26.4.1 (arm64)

Install method

npm-global

Model

mlx-qwopus3.5-27b-v3

Provider / routing chain

openclaw -> lmstudio (local)

Additional provider/model setup details

Local model served via LM Studio; Node v25.9.0. with oprnai-compatible node

Logs, screenshots, and evidence

1. `openclaw status` output showing all sessions with "unknown/262k (?%)":


│ agent:main:main                  │ direct │ 1m ago  │ mlx-qwopus3.5-27b-v3 │ unknown/262k (?)%   │
│ agent:main:dreaming-narrative…  │ direct │ 2h ago  │ mlx-qwopus3.5-27b-v3 │ unknown/262k (?)%   │


2. Control UI web Sessions page: Tokens column displays "N/A" (不适用).
3. Control UI Usage page: Only message counts shown; token statistics absent.
4. Manual compression temporarily restores token display, but it reverts to N/A after the next message.

Impact and severity

• Affected users: Maybe all users running OpenClaw 2026.4.9+ with local LM Studio models; multiple sessions show the same pattern, indicating it is not session-specific.
• Severity: Blocks workflow — auto-compression failure causes context overflow, which can interrupt or lose ongoing sessions. Token monitoring is completely unavailable.
• Frequency: Always — every session consistently fails to display token usage and auto-compression never triggers.
• Consequence: Context overflow causes session failures and potential data loss; users cannot monitor or manage context window consumption; manual compression is only a temporary workaround.

Additional information

This is a regression: last known good version is prior to 2026.4.9, first known bad version is 2026.4.9, and the issue persists through 2026.4.11. The root cause appears to be that OpenClaw cannot correctly parse the usage field from API responses, which cascades into token display failure and auto-compression breakdown. Related issue #50196 described a different symptom (wrong calculation: "510.4k/128k") rather than "unknown/N/A" values, suggesting a distinct root cause or a new regression starting from 2026.4.9.