Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

When agents.defaults.compaction.mode = "safeguard", certain sessions repeatedly trigger the safeguard's "no real conversation messages" branch and write a 144-char fallback summary (## Decisions\nNo prior history.\n...). The actual compaction LLM call then runs in parallel but is canceled after the 60s COMPACTION_RETRY_AGGREGATE_TIMEOUT_MS, and the runtime falls back to a "pre-compaction snapshot" — meaning the transcript still has full size, but a placeholder compaction marker has been recorded. On the next turn the same session re-enters the same path, creating a loop of 60s-stalled turns with empty summaries.

Switching to mode: "default" makes the symptom go away (no safeguard hook → no empty-fallback writes).

The sessions where this triggers are real working sessions with substantive user and assistant messages plus tool calls. A simulation replay of isRealConversationMessage on the same transcript reports 14/108 "real" messages — well above the threshold the safeguard's if (!hasRealSummarizable && !hasRealTurnPrefix) check should require — so something in the message set passed to the safeguard hook is being seen as "no real conversation" even though the file-on-disk transcript clearly contains real conversation.

This affects:

• Cron-triggered sessions (e.g. daily-report cron at 38K tokens)
• Newly-reset interactive sessions (/new flow)
• Long DM sessions

I have not yet identified the upstream commit that introduced the regression — containsRealConversationMessages and the 60s aggregate timer existed in 2026.4.23 (still ran fine for me there), so the divergence likely lives in how messagesToSummarize is constructed for the session_before_compact event. Happy to dig further if maintainers point at the right area.

Steps to reproduce

1. Run OpenClaw 2026.5.3-1 with agents.defaults.compaction.mode: "safeguard" and Anthropic-style provider (modelsproxy → claude-opus-4-7 in my case)
2. Either:
   • Trigger a cron task that does a few read / exec tool calls (the daily report pipeline I have hits this every morning), OR
   • Start a /new session that does a few rounds of investigation (read files, run exec, etc.)
3. Wait until preflight compaction triggers (this happens well before context limit because of truncateAfterCompaction being false in my config — I am NOT setting maxActiveTranscriptBytes)
4. Watch the gateway logs

Expected: compaction summary captures the conversation so far.
Actual: 144-char fallback summary; 60s aggregate timeout; pre-compaction snapshot used; same session re-triggers compaction on the next turn.

Expected behavior

Either:

1. Fix root cause — the safeguard hook should agree with the underlying transcript on whether real conversation exists. Real cron / interactive sessions with user prompts + assistant tool-using turns should always summarize, not fall through to "No prior history".
2. Surface the gap — if the safeguard intentionally treats some message classes as non-conversation (e.g. all toolCall blocks with empty assistant text), the fallback summary should at least note "compaction skipped — N messages not classified as conversation" rather than "No prior history", to help operators diagnose.
3. Make default the documented recommendation for users without specific safeguard requirements — currently the schema help text says "Keep 'default' unless you observe aggressive history loss near limit boundaries" but the wizard / docs are silent on the regression risk of safeguard.

Actual behavior

Logs show repeated:

[compaction-safeguard] Compaction safeguard: no real conversation messages to summarize; writing compaction boundary to suppress re-trigger loop.
[agent/embedded] compaction retry aggregate timeout (60000ms): proceeding with pre-compaction state runId=... sessionId=...
[agent/embedded] using pre-compaction snapshot: timed out during compaction runId=... sessionId=...

Recorded compaction entries in transcript files have:

{"type":"compaction","summary":"## Decisions\nNo prior history.\n\n## Open TODOs\nNone.\n\n## Constraints/Rules\nNone.\n\n## Pending user asks\nNone.\n\n## Exact identifiers\nNone captured.","tokensBefore":38085,"fromHook":true}

User-visible impact:

• Each compaction-loop turn adds ~60s of dead time
• API token usage is wasted on the aborted summarization calls
• Session history is not actually compressed — pre-compaction snapshot is restored, so the next turn faces the same pressure
• Cron pipelines that should finish in 5–10 min stretch to 15+ min

OpenClaw version

2026.5.3-1 (commit 2eae30e)

Operating system

Ubuntu 20.04 LTS, kernel 5.10.25, x64

Install method

npm install -g openclaw via tarball (proxy-bound install)

Model

claude-opus-4-7

Provider / routing chain

feishu (DM / cron) → openclaw gateway → model-provider(enterprize) → anthropic-messages

Additional provider/model setup details

• Compaction model NOT explicitly overridden — uses session default
• agents.defaults.compaction.mode: "safeguard" (was the trigger; switching to "default" resolved symptoms)
• agents.defaults.compaction.timeoutSeconds: not set, so 900s default
• agents.defaults.compaction.truncateAfterCompaction: not set (so byte-trigger is disabled, only token threshold is active)

Logs, screenshots, and evidence

Repro count over ~12 hours on 2026-05-06

For comparison, on 2026-05-05 (same install, same config) one compaction at 178K tokens did produce a real ~5KB summary on session 963e608a. So the safeguard path is not 100% broken — it's an as-yet-uncharacterized condition that some sessions hit and others don't.

Sample log excerpt

May 06 08:42:16 ... [compaction-safeguard] Compaction safeguard: no real conversation messages to summarize; writing compaction boundary to suppress re-trigger loop.
May 06 08:43:16 ... [agent/embedded] compaction retry aggregate timeout (60000ms): proceeding with pre-compaction state runId=abf1edff-9629-4b4b-9168-fe4ec4d1c41a sessionId=abf1edff-9629-4b4b-9168-fe4ec4d1c41a
May 06 08:43:16 ... [agent/embedded] using pre-compaction snapshot: timed out during compaction runId=abf1edff-9629-4b4b-9168-fe4ec4d1c41a sessionId=abf1edff-9629-4b4b-9168-fe4ec4d1c41a

Sample empty fallback summary entry from transcript

{
  "type": "compaction",
  "id": "17c34363",
  "parentId": "3aa68b51",
  "timestamp": "2026-05-06T00:42:16.544Z",
  "summary": "## Decisions\nNo prior history.\n\n## Open TODOs\nNone.\n\n## Constraints/Rules\nNone.\n\n## Pending user asks\nNone.\n\n## Exact identifiers\nNone captured.",
  "firstKeptEntryId": "1e91a7ec",
  "tokensBefore": 38085,
  "fromHook": true
}

Workaround that resolved my issue

 "compaction": {
-  "mode": "safeguard"
+  "mode": "default"
 }

After systemctl --user restart openclaw-gateway no further [compaction-safeguard] log lines appear.

Hypothesis (for maintainer triage)

compaction-successor-transcript-D4FqvNh7.js line 3585 (in 2026.5.3-1 dist):

if (!hasRealSummarizable && !hasRealTurnPrefix) {
    log.info("Compaction safeguard: no real conversation messages to summarize; writing compaction boundary to suppress re-trigger loop.");
    return { compaction: {
        summary: buildStructuredFallbackSummary(preparation.previousSummary),
        firstKeptEntryId: preparation.firstKeptEntryId,
        tokensBefore: preparation.tokensBefore
    } };
}

hasRealSummarizable is computed from stripRuntimeContextCustomMessages(preparation.messagesToSummarize). My guess: preparation.messagesToSummarize in some code paths contains only the slice of messages already split off as "to-summarize" while real conversation messages are in preparation.turnPrefixMessages or a different field, so the OR-check (!hasRealSummarizable && !hasRealTurnPrefix) should be evaluating both, but one or both views may be wrong post-strip.

A repro with [compaction-diag] debug logging enabled would pinpoint which of messagesToSummarize / turnPrefixMessages is empty when the safeguard fires.

Happy to provide the affected *.jsonl transcripts (after redaction) and a longer log window if useful.

Impact and severity

No response

Additional information

No response