[google/syzkaller] 927dfd: pkg/aflow: harden llm_agent context compression mode

[Aleksandr Nogikh]

Branch: refs/heads/gh-readonly-queue/master/pr-7282-4df5680b122c9a5da0c486b78325bc60a4883e2a
Home: https://github.com/google/syzkaller
Commit: 927dfd751917c0d0e6d1021127cd1daf8ce0c674
https://github.com/google/syzkaller/commit/927dfd751917c0d0e6d1021127cd1daf8ce0c674
Author: Aleksandr Nogikh <nog...@google.com>
Date: 2026-05-16 (Sat, 16 May 2026)

Changed paths:
M pkg/aflow/llm_agent.go
M pkg/aflow/llm_agent_test.go
M pkg/aflow/testdata/TestTokenCompression.llm.json
M pkg/aflow/testdata/TestTokenCompression.trajectory.json

Log Message:
-----------
pkg/aflow: harden llm_agent context compression mode

The previous compression logic calculated token limits based on the
total number of input tokens, which included the anchor prompt. For
workflows with massive initial prompts (e.g. ones with large C reproducers),
this would result in the context threshold being exceeded immediately,
causing continuous and infinite context compression loops.

Update the maybeCompressContext logic to establish an 'anchorTokens'
baseline from the first response. In subsequent turns, only evaluate
the token delta against the CompressTokens threshold, effectively
only counting tokens added after the first prompt. Update tests and
golden trajectories to reflect the new delta-based tracking behavior.



To unsubscribe from these emails, change your notification settings at https://github.com/google/syzkaller/settings/notifications

[Aleksandr Nogikh]

Branch: refs/heads/master