feat: register compaction retry hook to prevent cascade overflow by 1kuna · Pull Request #10220 · openclaw/openclaw

1kuna · 2026-02-06T07:12:52Z

Summary

Fixes #10613. Implements a compaction retry hook that prevents cascade overflow loops.

When auto-compaction fires during context overflow, the retry can immediately overflow again. This hook:

Intercepts the retry before it fires
Calculates a safe token budget
Downgrades the prompt to a slim one-shot version if needed
Cancels the retry if even the slim prompt won't fit

Dependencies

Requires setAutoCompactionRetryHook from pi-coding-agent (PR badlogic/pi-mono#1318).

Verification

pnpm tsgo passes clean with local pi-mono build
Tests included

greptile-apps

_{3 files reviewed, 3 comments}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-02-06T07:16:18Z

package.json

+    "@mariozechner/pi-agent-core": "^0.52.6",
+    "@mariozechner/pi-ai": "^0.52.6",
+    "@mariozechner/pi-coding-agent": "^0.52.6",
+    "@mariozechner/pi-tui": "^0.52.6",


Dependency range breaks hook gating

This PR relies on a new Pi SDK API (setAutoCompactionRetryHook), but package.json loosens all Pi deps to ^0.52.6. If the hook lands in 0.52.7+ (as described), users can end up on a Pi version that still doesn’t have the hook (or has a different shape) while the code assumes “maybe supported”. This makes the safeguard unreliable in the exact scenario it’s meant to fix.

Recommend pinning the minimum Pi versions that actually include the hook (or bump the range to ^<first-version-with-hook>).

Prompt To Fix With AI

This is a comment left during a code review. Path: package.json Line: 111:114 Comment: **Dependency range breaks hook gating** This PR relies on a *new* Pi SDK API (`setAutoCompactionRetryHook`), but `package.json` loosens all Pi deps to `^0.52.6`. If the hook lands in `0.52.7+` (as described), users can end up on a Pi version that still doesn’t have the hook (or has a different shape) while the code assumes “maybe supported”. This makes the safeguard unreliable in the exact scenario it’s meant to fix. Recommend pinning the **minimum** Pi versions that actually include the hook (or bump the range to `^<first-version-with-hook>`). How can I resolve this? If you propose a fix, please make it concise.

greptile-apps · 2026-02-06T07:16:19Z

src/agents/pi-embedded-runner/run/attempt.ts

+        const mutableSession = activeSession as unknown as {
+          _baseSystemPrompt?: string;
+          _rebuildSystemPrompt?: (toolNames: string[]) => string;
+        };
+        const previousBasePrompt = mutableSession._baseSystemPrompt;
+        const previousRebuild = mutableSession._rebuildSystemPrompt;
+        applySystemPromptOverrideToSession(activeSession, getRetrySystemPromptText());
+        restoreOneShotRetryPromptOverride = () => {
+          mutableSession._baseSystemPrompt = previousBasePrompt;
+          mutableSession._rebuildSystemPrompt = previousRebuild;
+          activeSession.agent.setSystemPrompt(previousBasePrompt ?? systemPromptText);
+        };
+      };


Prompt restore uses undefined

restoreOneShotRetryPromptOverride calls activeSession.agent.setSystemPrompt(previousBasePrompt ?? systemPromptText), but previousBasePrompt can be undefined if Pi’s internals don’t populate _baseSystemPrompt (or rename it). In that case, it resets the agent prompt to systemPromptText, but it also writes back mutableSession._baseSystemPrompt = previousBasePrompt (i.e. undefined), potentially leaving the session’s internal “base prompt” unset for subsequent _rebuildSystemPrompt calls.

This can break later prompt rebuilds (e.g. tool list changes) after a compaction event. Consider restoring to the current session base prompt when available (or only writing _baseSystemPrompt back if it was originally defined).

Prompt To Fix With AI

This is a comment left during a code review. Path: src/agents/pi-embedded-runner/run/attempt.ts Line: 521:533 Comment: **Prompt restore uses undefined** `restoreOneShotRetryPromptOverride` calls `activeSession.agent.setSystemPrompt(previousBasePrompt ?? systemPromptText)`, but `previousBasePrompt` can be `undefined` if Pi’s internals don’t populate `_baseSystemPrompt` (or rename it). In that case, it resets the agent prompt to `systemPromptText`, but it also writes back `mutableSession._baseSystemPrompt = previousBasePrompt` (i.e. `undefined`), potentially leaving the session’s internal “base prompt” unset for subsequent `_rebuildSystemPrompt` calls. This can break later prompt rebuilds (e.g. tool list changes) after a compaction event. Consider restoring to the **current** session base prompt when available (or only writing `_baseSystemPrompt` back if it was originally defined). How can I resolve this? If you propose a fix, please make it concise.

greptile-apps · 2026-02-06T07:16:20Z

src/agents/pi-embedded-subscribe.handlers.lifecycle.ts

  emitAgentEvent({
    runId: ctx.params.runId,
    stream: "compaction",
-    data: { phase: "end", willRetry },
+    data: { phase: "end", willRetry, retryCanceledMessage },
  });
  void ctx.params.onAgentEvent?.({
    stream: "compaction",
-    data: { phase: "end", willRetry },
+    data: { phase: "end", willRetry, retryCanceledMessage },
  });
+
+  if (!willRetry && retryCanceledMessage) {
+    // User-facing propagation: Pi succeeded compacting but refused to retry due to prompt sizing.
+    void ctx.params.onBlockReply?.({ text: retryCanceledMessage });
+  }


User message emitted mid-stream

handleAutoCompactionEnd calls onBlockReply immediately when retryCanceledMessage is present. This happens on the compaction event stream, not the normal assistant response lifecycle, so it can interleave with other block buffering/chunking state and produce out-of-order user-visible output.

If onBlockReply is used by external messaging channels (which must only receive final replies), this risks sending a standalone message during an in-flight run. Consider routing this through the same “final reply” path you use for other user-facing errors (or gate it to internal UIs only).

Prompt To Fix With AI

This is a comment left during a code review. Path: src/agents/pi-embedded-subscribe.handlers.lifecycle.ts Line: 54:67 Comment: **User message emitted mid-stream** `handleAutoCompactionEnd` calls `onBlockReply` immediately when `retryCanceledMessage` is present. This happens on the compaction event stream, not the normal assistant response lifecycle, so it can interleave with other block buffering/chunking state and produce out-of-order user-visible output. If `onBlockReply` is used by external messaging channels (which must only receive final replies), this risks sending a standalone message during an in-flight run. Consider routing this through the same “final reply” path you use for other user-facing errors (or gate it to internal UIs only). How can I resolve this? If you propose a fix, please make it concise.

…12889 #12309 #3594 #7483 #10094 #10368 #11317 #11359 #11649 #12022 #12432 #12676 #12711; PRs #7567 #10220 #10601 #10620 #10760 #11680 #11685 #12052 #12226 #12433 #12702 #12720 #12726 #12777)

Takhoffman · 2026-02-10T01:53:05Z

Fixed in #12988.

This will go out in the next OpenClaw release.

If you still see this after updating to the first release that includes #12988, please open a new issue with:

your OpenClaw version
channel (Telegram/Slack/etc)
the exact prompt/response that got rewritten
whether Web UI showed the full text vs the channel being rewritten
relevant logs around send/normalize (if available)

Link back here for context.

Co-authored-by: Alyx <kunaclawd@gmail.com>

…eout safety net (openclaw#9277, openclaw#7630)

Co-authored-by: Alyx <kunaclawd@gmail.com>

tsgo cannot track mutations through closures called asynchronously, so restoreOneShotRetryPromptOverride narrows to never at the finally block. Snapshot to a local const before calling to satisfy strict control-flow analysis.

tsgo narrows closure-mutated let variables to their init type (null), making them uncallable. Wrap in a { current } container object which tsgo does not narrow through property access, matching the React useRef pattern. No runtime behavior change.

openclaw-barnacle · 2026-03-08T04:08:26Z

This pull request has been automatically marked as stale due to inactivity.
Please add updates or it will be closed.

openclaw-barnacle bot added the agents Agent runtime and tooling label Feb 6, 2026

greptile-apps bot reviewed Feb 6, 2026

View reviewed changes

This was referenced Feb 6, 2026

feat: auto-compaction retry hook for embedded consumers badlogic/pi-mono#1319

Closed

Compaction retry cascade causes context overflow loop #10613

Closed

1kuna force-pushed the feat/compaction-retry-safeguard branch 2 times, most recently from e046781 to 31ca12b Compare February 7, 2026 04:30

Takhoffman self-assigned this Feb 10, 2026

Takhoffman mentioned this pull request Feb 10, 2026

Agents: scope sanitizeUserFacingText rewrites to errorContext #12988

Merged

This was referenced Feb 11, 2026

Compaction retry cascade when system prompt exceeds context budget (not resolved by #11664 or #12988) #14299

Closed

Bug still present: compaction retry cascade after overflow (not resolved by #11664) #14302

Closed

1kuna and others added 9 commits February 11, 2026 17:54

fix: guard against NaN reserveTokens in compaction safeguard

5a230ae

Co-authored-by: Alyx <kunaclawd@gmail.com>

fix(agents): use session-only lane for compaction + add lane task tim…

67eecee

…eout safety net (openclaw#9277, openclaw#7630)

Clamp lane task timeout minimum

240f080

feat: register compaction retry hook to prevent cascade overflow

b9c1270

Co-authored-by: Alyx <kunaclawd@gmail.com>

fix: compaction retry safeguard (cancel propagation, one-shot downgrade)

3c0daab

Co-authored-by: Alyx <kunaclawd@gmail.com>

test: update compaction retry hook tests

fa60323

Co-authored-by: Alyx <kunaclawd@gmail.com>

fix: address round-2 review findings

831863d

Co-authored-by: Alyx <kunaclawd@gmail.com>

1kuna force-pushed the feat/compaction-retry-safeguard branch from 31ca12b to 2ade518 Compare February 12, 2026 00:25

This was referenced Feb 12, 2026

fix(agents): use session-only lane for compaction + add lane task timeout safety net #10585

Closed

fix: guard against NaN reserveTokens in compaction safeguard #9287

Closed

thewilloftheshadow force-pushed the main branch from bfc1ccb to f92900f Compare February 15, 2026 18:46

Takhoffman removed their assignment Feb 21, 2026

openclaw-barnacle bot added the stale Marked as stale due to inactivity label Mar 8, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: register compaction retry hook to prevent cascade overflow#10220

feat: register compaction retry hook to prevent cascade overflow#10220
1kuna wants to merge 9 commits intoopenclaw:mainfrom
1kuna:feat/compaction-retry-safeguard

1kuna commented Feb 6, 2026 •

edited

Loading

Uh oh!

greptile-apps bot left a comment

Uh oh!

greptile-apps bot Feb 6, 2026

Uh oh!

greptile-apps bot Feb 6, 2026

Uh oh!

greptile-apps bot Feb 6, 2026

Uh oh!

Takhoffman commented Feb 10, 2026

Uh oh!

openclaw-barnacle bot commented Mar 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

1kuna commented Feb 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Dependencies

Verification

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Feb 6, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Feb 6, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Feb 6, 2026

Choose a reason for hiding this comment

Uh oh!

Takhoffman commented Feb 10, 2026

Uh oh!

openclaw-barnacle bot commented Mar 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

1kuna commented Feb 6, 2026 •

edited

Loading