Summary

Fixes openclaw#54383 — Compaction fails with max_tokens: 240000 > 128000 when using Anthropic models with 1M context windows.

Root Cause

In @mariozechner/pi-coding-agent, generateSummary() calculates:

const maxTokens = Math.floor(0.8 * reserveTokens);

With reserveTokensFloor: 300000 (appropriate for 1M context), this produces max_tokens = 240000 — exceeding Anthropic's per-request output cap of 128K for both Sonnet 4.6 and Opus 4.6.

Fix

Clamp reserveTokens in src/agents/compaction.ts before passing to generateSummary():

const modelMaxTokens = params.model.maxTokens ?? 128_000;
const clampedReserveTokens = Math.min(params.reserveTokens, Math.floor(modelMaxTokens / 0.8));

This ensures the downstream max_tokens calculation (0.8 * reserveTokens) never exceeds the model's actual output limit. The fix uses model.maxTokens from the provider registry, so it's forward-compatible — if future models raise their output cap, no code change is needed.

Impact

• Before: Compaction broken for all users with Anthropic + 1M context (any reserveTokensFloor > 160K)
• After: Compaction works correctly, respecting model output limits while preserving the existing summarization quality

Testing

The fix is in the OpenClaw wrapper layer (src/agents/compaction.ts), not in the upstream pi-coding-agent package. This is the minimal, safest change — the upstream package could also benefit from the same clamp in generateSummary() itself.

Verified that:

• model.maxTokens is populated from the provider catalog (128K for Anthropic Vertex models)
• Math.floor(128000 / 0.8) = 160000, so clampedReserveTokens = min(300000, 160000) = 160000
• generateSummary then calculates Math.floor(0.8 * 160000) = 128000 ✅ (within model limit)