fix(file-tools): escalate to BLOCKED on repeated read_file dedup stubs - #16382
Merged
Conversation
read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes #15759
Contributor
🚨 CRITICAL Supply Chain Risk DetectedThis PR contains a pattern that has been used in real supply chain attacks. A maintainer must review the flagged code carefully before merging. 🚨 CRITICAL: Install-hook file added or modifiedThese files can execute code during package installation or interpreter startup. Files: Scanner only fires on high-signal indicators: .pth files, base64+exec/eval combos, subprocess with encoded commands, or install-hook files. Low-signal warnings were removed intentionally — if you're seeing this comment, the finding is worth inspecting. |
alt-glitch
added
type/bug
Collaborator
4 tasks
cluricaun28
referenced
this pull request
in cluricaun28/Logos
Apr 28, 2026
…s (#16382) read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes #15759
donald131
pushed a commit
to donald131/hermes-agent
that referenced
this pull request
May 2, 2026
NousResearch#16382) read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes NousResearch#15759
02356abc
pushed a commit
to 02356abc/hermes-agent
that referenced
this pull request
May 14, 2026
NousResearch#16382) read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes NousResearch#15759
dannyJ848
pushed a commit
to dannyJ848/hermes-agent
that referenced
this pull request
May 17, 2026
NousResearch#16382) read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes NousResearch#15759
gweeteve
pushed a commit
to gweeteve/hermes-agent
that referenced
this pull request
Jun 2, 2026
NousResearch#16382) read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes NousResearch#15759
waefrebeorn
pushed a commit
to waefrebeorn/slermes
that referenced
this pull request
Jul 2, 2026
NousResearch#16382) read_file's dedup path returned a lightweight stub on re-reads of an unchanged file, then returned early — so the consecutive-read loop guard (hard block at count>=4) at the bottom of read_file_tool never ran for stub-looped calls. Weaker tool-following models (local Qwen3.6 variants in the reported case) ignore the passive 'refer to earlier result' hint and hammer the same read_file call until iteration budget runs out. Track per-key stub returns in task_data['dedup_hits'] and, on the second stub for the same (path, offset, limit), return a hard BLOCKED error mirroring the wording the real-read path already uses. A real read, an intervening non-read tool call (notify_other_tool_call), or reset_file_dedup (on context compression) all clear the counter so the guard never stays engaged longer than the actual loop. Closes NousResearch#15759
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
2 participants
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.
Summary
Kills an infinite
read_fileloop that weaker tool-following models (local Qwen3.6 variants in the reported case) hit when they ignore the dedup stub's passive 'refer to earlier result' hint.Root cause: the dedup path (
tools/file_tools.pyL449) returned the stub and then returned early, so the consecutive-read loop guard at L552 (hard block at count>=4) never ran for stub-looped calls. Reporter's analysis was correct.Changes
tools/file_tools.py: track per-key stub hits intask_data['dedup_hits']; on the 2nd stub for the same(path, offset, limit)return a hard BLOCKED error mirroring the real-read path's wording. Counter clears on a real read of that key, onnotify_other_tool_call(intervening non-read tool breaks the loop), and onreset_file_dedup(context compression). Cap wired into_cap_read_tracker_datalike the other dicts.tests/tools/test_file_read_guards.py: 6 new tests covering 3rd-call block, sustained block on further hammering, block cleared by real file modification, counter cleared by an intervening non-read tool call, per-range independence, andreset_file_dedupclearing hits.Validation
tests/tools/test_file_read_guards.pytest_file_state_registry,test_file_staleness)Closes #15759