Bug: Sudden Context Overflow from Large Tool Results (no preflight guard)

Summary

OpenClaw can exceed model context in a single step when a tool returns a very large payload.
After that, requests fail repeatedly with:

• prompt is too long: ... > 200000
• input length and max_tokens exceed context limit: ... + 32000 > 200000

In several cases, no automatic compaction happened before the next request, so sessions got stuck in a failure loop.

Environment

• OpenClaw: 2026.1.30
• Model: anthropic/claude-sonnet-4-20250514 and anthropic/claude-opus-4-5
• Context window: 200000

What happened

Large single toolResult messages were appended directly to session transcript and immediately pushed next prompt over limit.

Concrete examples (from local transcripts)

1. gateway config.schema produced a huge payload:

• transcript: sessions/d3a0a84a-ab0f-4597-9de4-8353a67acfec.jsonl
• tool call line: 1402 (gateway {"action":"config.schema"})
• tool result line: 1403 (~560k chars JSON line)
• immediate error: line 1404 (prompt is too long: 205607 tokens > 200000)

2. Grepping session logs with tail ... ~/.openclaw/agents/main/sessions/*.jsonl produced huge outputs:

• transcript: sessions/d3a0a84a-ab0f-4597-9de4-8353a67acfec.jsonl
• tool calls: lines 1857, 1869
• tool results: lines 1858 (~400k chars), 1870 (~409k chars)
• immediate failures: lines 1859, 1871, 1874, 1877

3. Same pattern in another session:

• transcript: sessions/276d664b-06c9-4a1a-80ad-562972d25beb.jsonl
• huge tool results on lines 32 and 34 (~416k chars each)
• repeated context-limit errors after that

4. Repeated failures at high context without successful recovery:

• transcript: sessions/7bff962a-ac31-44f3-9c17-300507453484.jsonl
• repeated errors on lines 955, 958, 961, 964, 967, 970
• message: input length and max_tokens exceed context limit: 173629 + 32000 > 200000

Expected behavior

Before every model request, OpenClaw should guarantee:

• estimated_input_tokens + requested_max_tokens <= model_context_limit

If not true, OpenClaw should automatically do one or more of:

1. trigger compaction preflight,
2. reduce max_tokens dynamically,
3. truncate/summarize newest oversized tool results,
4. retry automatically with safe budget.

The session should not enter a repeated hard-failure loop.

Suggested fixes

1. Preflight context budget check before each LLM call, based on current transcript + planned max_tokens.
2. Hard output cap per tool result inserted into transcript (character and token budget).
3. Auto-summarize oversized tool outputs (store full output externally/log file, keep summary in context).
4. Guardrails for high-risk commands that can return giant transcript chunks (e.g. tail ~/.openclaw/agents/main/sessions/*.jsonl).
5. Fallback retry policy:

• on input + max_tokens error: reduce max_tokens and retry,
• if still too large: force compaction and retry once.

6. Optional warning mode when any single toolResult exceeds configured threshold (e.g. 10k/20k tokens).

Why this matters

This causes:

• session stalls,
• 5x+ token usage spikes (many failed retries),
• broken automation loops until manual intervention.