Repository navigation
Conversation
…y summary The summarizer sent the history to the unbound model. After a long run of tool calls, Opus 5.5 answers that request by ending its turn with no text: a CVLR author on the Solana stake program compacted three times in one run, got 27, 1 and 24 output tokens back, and resumed each time from an empty summary. Replaying the first request reproduced it every time. With the tools bound the model calls `memory` instead; with either, a closing instruction to answer in plain text produces the summary (4,126 and 4,396 tokens). - The summarizer calls the tool-bound model, so the request shares the agent's cached prefix. Each of those compactions wrote ~517k tokens to the cache instead. - Every summarization request ends with an instruction not to call a tool. - A reply with no text is retried once. If it is still empty, the history is kept and the next turn summarizes again, with a warning naming the stop reason and any tool calls. A summarization that raises is logged rather than dropped silently. - A node generator may now make more than one LLM request. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ericeil
added a commit
to Certora/AutoProver
that referenced
this pull request
Nov 6, 2026
… summary Pins graphcore at f4221b6, the head of Certora/graphcore#42 (open). Re-pin to the merge commit once it lands. The bump also brings in graphcore master's #41 (bounded grep_files output). On the cloud stake run all three compactions of the Withdrawal author returned no text (27, 1 and 24 output tokens), and the author resumed each time from an empty summary. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
The summarizer sent the history to the unbound model. After a long run of tool calls, Opus 5.5 answers a bare summarization request by ending its turn with no text. On a cloud CVLR run (Solana stake program, Withdrawal author) the author compacted three times. The three summarize calls returned 27, 1 and 24 output tokens, and each time the author resumed from
BEGIN SUMMARY: END SUMMARY. Each call also wrote ~517k tokens to the cache, since without the tool definitions the prefix differs from the agent's.Replaying the first real request, loaded from the checkpoint:
end_turn, 1 output token, empty, both timestool_use: callsmemory, no textChange
SUMMARY_ATTEMPTS = 2). If it is still empty, the history is kept and the next turn summarizes again, with a warning naming the stop reason, the tool calls and the output tokens. A summarization that raises is now logged; before, it was dropped silently._stitch_sync_impl/_stitch_async_impllet a node generator make more than one LLM request, which the retry needs. Single-request nodes behave as before.Testing
tests/test_summarizer.py: the request asks for plain text, a text reply replaces the history, an empty reply is retried, and two empty replies keep the history.🤖 Generated with Claude Code