Bug Description

Thread sessions forked from a parent DM session inherit the parent's full context. When the parent session has accumulated tokens near/above the model's context limit, newly forked thread sessions are born dead — every API call immediately returns prompt is too long: N tokens > 190000 maximum and the user gets no response.

The failure is silent on Slack — no error message is sent to the user. Messages are simply swallowed.

Steps to Reproduce

1. Use Slack DM channel with thread.inheritParent: true (default)
2. Have extended conversations in the parent DM session until it accumulates ~170k+ tokens
3. Start a new thread (Slack auto-creates a thread session forked from parent)
4. The thread session immediately hits context overflow on first API call
5. User sees no response — the error is logged but never surfaced

Observed Behavior

From gateway error log:

[session-init] forking from parent session: parentKey=agent:main:slack:direct:... → sessionKey=...thread:... parentTokens=169508
[agent/embedded] embedded run agent end: isError=true error=Context overflow: prompt too large for the model.

From Anthropic API:

{"type":"error","error":{"type":"invalid_request_error","message":"prompt is too long: 193390 tokens > 190000 maximum"}}

Key details:

• contextTokens config is set to 1,000,000 but this is a cap, not an override — model's actual limit (190k for Opus 4-6) takes precedence
• Auto-compaction attempts (up to 3) fail because the inherited context is already too large
• The error response has usage: {input: 0, output: 0} — no tokens consumed
• No error message is sent to the Slack channel — complete silent failure

Expected Behavior

1. Thread sessions should NOT inherit full parent context when parent is near/over the context limit — compact before forking, or fork with a summary only
2. Context overflow errors should be surfaced to the user on Slack (e.g., "⚠️ Context overflow — starting fresh session") instead of silently dropping the message
3. Parent sessions should auto-compact proactively before hitting the model limit, not only on overflow

Environment

• OpenClaw version: 2026.2.24 (npm)
• Model: anthropic/claude-opus-4-6 (190k context limit)
• Channel: Slack (Socket Mode)
• Config: contextTokens: 1000000, compaction.mode: "default", thread.inheritParent: true

Impact

This caused 3 separate incidents over 2 days where Slack became unresponsive. Each required manual session file deletion + gateway restart to recover. The silent failure mode means the user has no idea why the bot stopped responding.

Related Issues

• Auto-compaction not triggered during tool-use loops, leading to unrecoverable context overflow #24800 — Auto-compaction not triggered during tool-use loops
• Feature: Context overflow warning before hitting limit #13097 — Context overflow warning before hitting limit
• [Bug]: Context-overflow recovery can finalize empty instead of deterministic fallback #19550 — Context-overflow recovery can finalize empty instead of deterministic fallback