Summary

The safeguard compaction mode (compaction.mode: "safeguard") excludes toolResult entries from valid cut points in findValidCutPoints(). This causes compaction to fail (producing empty fallback summaries) for sessions dominated by tool calls — exactly the pattern seen in long-running agentic tasks and cron jobs.

Problem to solve

findValidCutPoints() only accepts these entry types as valid cut points:

• user
• assistant
• bashExecution
• custom
• branchSummary
• compactionSummary

toolResult is explicitly excluded.

When a session's messages consist primarily of toolCall → toolResult pairs (with few or no user/assistant text messages), findCutPoint() returns no valid cut point. This triggers the safeguard's "no real conversation messages to summarize" branch, writing an empty fallback summary like:

## Decisions
No prior history.

## Open TODOs
None.
...

The compaction then fails to produce a meaningful summary, and the session cannot be extended via compaction.

Proposed solution

Add toolResult to the list of valid cut points in findValidCutPoints():

// src/core/compaction/compaction.ts (or equivalent location)
const validCutPointTypes = [
  'user',
  'assistant',
  'bashExecution',
  'custom',
  'branchSummary',
  'compactionSummary',
  'toolResult',  // ← Add this
];

Alternatives considered

If enabling toolResult as a cut point by default is deemed too risky, an alternative would be to make it configurable:

{
  "agents": {
    "defaults": {
      "compaction": {
        "toolResultAsCutPoint": true  // ← New config option
      }
    }
  }
}

Pros of the configurable approach:

• Allows users to opt in based on their specific use case (agentic tasks vs. simple chat)
• Provides a safe default for existing users while enabling the fix for those who need it
• Easier to roll out and test in production without risking regressions for all users

Recommended default: true for mode: "safeguard" (since safeguard is already the more aggressive mode), or false for backward compatibility.

However, the simpler fix (adding toolResult to valid cut points unconditionally) should work because:

• The compaction splitting logic already ensures tool call pairs stay together (it moves the boundary if it lands inside a tool block)
• Cutting after a toolResult naturally preserves the preceding toolCall

Impact

This affects:

• Long-running agentic tasks that execute many tool calls (read, exec, api calls) with minimal user text input
• Cron sessions that perform automated workflows with tool-heavy execution
• Any session where tool results dominate the conversation history

Even after the fix for #78300 (which added custom-message, bash, and branch-summary as anchors), sessions without explicit user/assistant text messages still fail to compact properly.

Evidence/examples

Session Analysis

A typical affected session (e.g., from E2E testing):

• 7 safeguard compaction attempts
• All compactions have firstKeptEntryId pointing to session header
• findCutPoint() returns no valid cut point
• messagesToSummarize is empty → boundary-only compaction
• No training export produced (no valid sample)

Related Issues

• [Bug]: Compaction safeguard mode produces empty fallback summaries that loop until 60s aggregate timeout #78300 — Fixed empty fallback summaries by adding custom-message, bash, branch-summary as anchors, but did not include toolResult
• compaction.mode=safeguard never triggers automatically when context engine windows messages or sets ownsCompaction #71325 — Fixed safeguard not triggering on windowed/ownsCompaction sessions, but cut point logic unchanged
• fix: compaction-safeguard causes HTTP 400 error instead of graceful skip when no conversation messages #32759 — Fixed HTTP 400 error when no messages to summarize, but root cause (missing cut points) not addressed

Rationale

1. Tool results carry semantic content: They contain the actual output of tool calls (file contents, command output, API responses) — this is often the most valuable information to preserve in a summary.

2. Tool call pairs are atomic: A toolCall and its matching toolResult should be kept together. Using toolResult as a cut point naturally preserves this pairing (the cut happens after the result, keeping the full call+result unit).

3. Consistent with existing anchors: bashExecution is already a valid cut point, and it's semantically similar to toolResult (both are execution outputs).

4. Fixes the agentic task gap: Long-running agentic tasks are exactly the use case where compaction should work, but they're currently blocked by this limitation.

Additional information

Environment

• OpenClaw version: 2026.5.5 (latest main)
• Config: compaction.mode: "safeguard"
• Affected session types: Agentic tasks, cron jobs, tool-heavy workflows

Related

• [Bug]: Compaction safeguard mode produces empty fallback summaries that loop until 60s aggregate timeout #78300 (empty fallback summaries — partial fix)
• compaction.mode=safeguard never triggers automatically when context engine windows messages or sets ownsCompaction #71325 (safeguard not triggering — different root cause)
• fix: compaction-safeguard causes HTTP 400 error instead of graceful skip when no conversation messages #32759 (HTTP 400 on empty messages — symptom, not root cause)
• [Bug] Compaction safeguard loop blocks cron lane — isolated sessions never execute #41981 (cron lane blocked by compaction loop — symptom)