fix: detect, warn, and block file re-read/search loops after context compression#705
Merged
Merged
Conversation
When context compression summarizes conversation history, the agent loses track of which files it already read and re-reads them in a loop. Users report the agent reading the same files endlessly without writing. Root cause: context compression is lossy — file contents and read history are lost in the summary. After compression, the model thinks it hasn't examined the files yet and reads them again. Fix (two-part): 1. Track file reads per task in file_tools.py. When the same file region is read again, include a _warning in the response telling the model to stop re-reading and use existing information. 2. After context compression, inject a structured message listing all files already read in the session with explicit "do NOT re-read" instruction, preserving read history across compression boundaries. Adds 16 tests covering warning detection, task isolation, summary accuracy, tracker cleanup, and compression history injection.
…ted todos - Block file reads after 3+ re-reads of same region (no content returned) - Track search_files calls and block repeated identical searches - Filter completed/cancelled todos from post-compression injection to prevent agent from re-doing finished work - Add 10 new tests covering all three fixes
Completed/cancelled items are now filtered from format_for_injection() output. Update the existing test to verify active items appear and completed items are excluded.
Combine read/search loop detection with main's redact_sensitive_text and truncation hint features. Add tracker reset to TestSearchHints to prevent cross-test state leakage.
_FakeReadResult and _FakeSearchResult now expose the attributes that read_file_tool/search_tool access after the redact_sensitive_text integration from main.
teknium1
added a commit
that referenced
this pull request
Mar 10, 2026
…ds, fix bugs Follow-up to PR #705 (merged from 0xbyt4). Addresses several issues: 1. CONSECUTIVE-ONLY TRACKING: Redesigned the read/search tracker to only warn/block on truly consecutive identical calls. Any other tool call in between (write, patch, terminal, etc.) resets the counter via notify_other_tool_call(), called from handle_function_call() in model_tools.py. This prevents false blocks in read→edit→verify flows. 2. THRESHOLD ADJUSTMENT: Warn on 3rd consecutive (was 2nd), block on 4th+ consecutive (was 3rd+). Gives the model more room before intervening. 3. TUPLE UNPACKING BUG: Fixed get_read_files_summary() which crashed on search keys (5-tuple) when trying to unpack as 3-tuple. Now uses a separate read_history set that only tracks file reads. 4. WEB_EXTRACT DOCSTRING: Reverted incorrect removal of 'title' from web_extract return docs in code_execution_tool.py — the field IS returned by web_tools.py. 5. TESTS: Rewrote test_read_loop_detection.py (35 tests) to cover consecutive-only behavior, notify_other_tool_call, interleaved read/search, and summary-unaffected-by-searches.
Contributor
|
Merged! Thanks for the contribution @0xbyt4 — the read-loop detection and todo injection filtering are great additions. I pushed a follow-up commit (a458b53) on top with several improvements:
|
Sarge-Reaper
added a commit
to Sarge-Reaper/hermes-agent
that referenced
this pull request
Mar 24, 2026
Applied Karpathy's autoresearch pattern to autonomously optimize the context compressor. 50 experiments run, 8 improvements kept. - _SUMMARY_RATIO 0.20 → 0.30 (more budget for summaries) - _MIN_SUMMARY_TOKENS 2000 → 500 (no inflation on short conversations) - _MAX_SUMMARY_TOKENS 8000 → 4000 (tighter cap) - _DEFAULT_TAIL_TOKEN_BUDGET 20000 → 8000 (more aggressive compression) - Truncation 3000 → 4500 chars (retains more tool output) - Regex file path pre-extraction with "MUST appear in summary" - Template restructured: Relevant Files + Critical Context moved up - MANDATORY PRESERVATION RULES added to both prompts Addresses NousResearch#705, NousResearch#1273, and context drift from lossy summarization. Score improved 3.6% (0.6346 → 0.6572).
4 tasks
angelburgosrosado
pushed a commit
to angelburgosrosado/hermes-agent
that referenced
this pull request
Apr 27, 2026
…search loops after context compression Authored by 0xbyt4. Adds read/search loop detection, file history injection after compression, and todo filtering for active items only.
angelburgosrosado
pushed a commit
to angelburgosrosado/hermes-agent
that referenced
this pull request
Apr 27, 2026
…ds, fix bugs Follow-up to PR NousResearch#705 (merged from 0xbyt4). Addresses several issues: 1. CONSECUTIVE-ONLY TRACKING: Redesigned the read/search tracker to only warn/block on truly consecutive identical calls. Any other tool call in between (write, patch, terminal, etc.) resets the counter via notify_other_tool_call(), called from handle_function_call() in model_tools.py. This prevents false blocks in read→edit→verify flows. 2. THRESHOLD ADJUSTMENT: Warn on 3rd consecutive (was 2nd), block on 4th+ consecutive (was 3rd+). Gives the model more room before intervening. 3. TUPLE UNPACKING BUG: Fixed get_read_files_summary() which crashed on search keys (5-tuple) when trying to unpack as 3-tuple. Now uses a separate read_history set that only tracks file reads. 4. WEB_EXTRACT DOCSTRING: Reverted incorrect removal of 'title' from web_extract return docs in code_execution_tool.py — the field IS returned by web_tools.py. 5. TESTS: Rewrote test_read_loop_detection.py (35 tests) to cover consecutive-only behavior, notify_other_tool_call, interleaved read/search, and summary-unaffected-by-searches.
02356abc
pushed a commit
to 02356abc/hermes-agent
that referenced
this pull request
May 14, 2026
…search loops after context compression Authored by 0xbyt4. Adds read/search loop detection, file history injection after compression, and todo filtering for active items only.
02356abc
pushed a commit
to 02356abc/hermes-agent
that referenced
this pull request
May 14, 2026
…ds, fix bugs Follow-up to PR NousResearch#705 (merged from 0xbyt4). Addresses several issues: 1. CONSECUTIVE-ONLY TRACKING: Redesigned the read/search tracker to only warn/block on truly consecutive identical calls. Any other tool call in between (write, patch, terminal, etc.) resets the counter via notify_other_tool_call(), called from handle_function_call() in model_tools.py. This prevents false blocks in read→edit→verify flows. 2. THRESHOLD ADJUSTMENT: Warn on 3rd consecutive (was 2nd), block on 4th+ consecutive (was 3rd+). Gives the model more room before intervening. 3. TUPLE UNPACKING BUG: Fixed get_read_files_summary() which crashed on search keys (5-tuple) when trying to unpack as 3-tuple. Now uses a separate read_history set that only tracks file reads. 4. WEB_EXTRACT DOCSTRING: Reverted incorrect removal of 'title' from web_extract return docs in code_execution_tool.py — the field IS returned by web_tools.py. 5. TESTS: Rewrote test_read_loop_detection.py (35 tests) to cover consecutive-only behavior, notify_other_tool_call, interleaved read/search, and summary-unaffected-by-searches.
olympus-terminal
pushed a commit
to olympus-terminal/hermes-agent
that referenced
this pull request
May 16, 2026
…search loops after context compression Authored by 0xbyt4. Adds read/search loop detection, file history injection after compression, and todo filtering for active items only.
olympus-terminal
pushed a commit
to olympus-terminal/hermes-agent
that referenced
this pull request
May 16, 2026
…ds, fix bugs Follow-up to PR NousResearch#705 (merged from 0xbyt4). Addresses several issues: 1. CONSECUTIVE-ONLY TRACKING: Redesigned the read/search tracker to only warn/block on truly consecutive identical calls. Any other tool call in between (write, patch, terminal, etc.) resets the counter via notify_other_tool_call(), called from handle_function_call() in model_tools.py. This prevents false blocks in read→edit→verify flows. 2. THRESHOLD ADJUSTMENT: Warn on 3rd consecutive (was 2nd), block on 4th+ consecutive (was 3rd+). Gives the model more room before intervening. 3. TUPLE UNPACKING BUG: Fixed get_read_files_summary() which crashed on search keys (5-tuple) when trying to unpack as 3-tuple. Now uses a separate read_history set that only tracks file reads. 4. WEB_EXTRACT DOCSTRING: Reverted incorrect removal of 'title' from web_extract return docs in code_execution_tool.py — the field IS returned by web_tools.py. 5. TESTS: Rewrote test_read_loop_detection.py (35 tests) to cover consecutive-only behavior, notify_other_tool_call, interleaved read/search, and summary-unaffected-by-searches.
nnnet
added a commit
to nnnet/hermes-agent
that referenced
this pull request
May 20, 2026
Wraps `mc_task_update {status: assigned, retry_count++}` with policy
guards so an mc-pm-chief can retry a failed/blocked sub-task without
looping forever:
- Capped by HERMES_MC_RETRY_COUNT env (default 3); explicit
`max_retries` arg overrides per-call.
- Refuses to retry tasks not in {failed, blocked}.
- Optional `expected_assignee` guard refuses retry if operator
reassigned the task to a different agent during failure handling.
- Returns `alert: true` when budget is exhausted so the chief can
surface a kanban_block or TG escalation instead of silently giving
up.
Combined with the Phase 3 webhook filter and Phase 4 cost summary, a
mc-pm-chief can now run a complete retry-then-escalate loop:
if task_status == 'failed':
out = mc_task_retry(task_id=X, expected_assignee=A)
if out.get('alert'):
mc_task_comment(task_id=X, content='retry budget exhausted')
kanban_block(reason='max retries hit, needs operator')
Tests: 86/86 (76 + 10 new) — validates id/max_retries/assignee,
budget-exhausted alert shape, env default vs explicit override,
reassignment refusal, dispatch-kwargs regression.
Closes Phase 5 of plans/hermes-mc-combo.md. All 5 phases of the combo
plan now implemented; only upstream PR NousResearch#705 (pnpm migration in
builderz-labs/mission-control) remains pending review.
1 task
Egavasyug
pushed a commit
to Egavasyug/hermes-agent
that referenced
this pull request
Jun 10, 2026
…search loops after context compression Authored by 0xbyt4. Adds read/search loop detection, file history injection after compression, and todo filtering for active items only.
Egavasyug
pushed a commit
to Egavasyug/hermes-agent
that referenced
this pull request
Jun 10, 2026
…ds, fix bugs Follow-up to PR NousResearch#705 (merged from 0xbyt4). Addresses several issues: 1. CONSECUTIVE-ONLY TRACKING: Redesigned the read/search tracker to only warn/block on truly consecutive identical calls. Any other tool call in between (write, patch, terminal, etc.) resets the counter via notify_other_tool_call(), called from handle_function_call() in model_tools.py. This prevents false blocks in read→edit→verify flows. 2. THRESHOLD ADJUSTMENT: Warn on 3rd consecutive (was 2nd), block on 4th+ consecutive (was 3rd+). Gives the model more room before intervening. 3. TUPLE UNPACKING BUG: Fixed get_read_files_summary() which crashed on search keys (5-tuple) when trying to unpack as 3-tuple. Now uses a separate read_history set that only tracks file reads. 4. WEB_EXTRACT DOCSTRING: Reverted incorrect removal of 'title' from web_extract return docs in code_execution_tool.py — the field IS returned by web_tools.py. 5. TESTS: Rewrote test_read_loop_detection.py (35 tests) to cover consecutive-only behavior, notify_other_tool_call, interleaved read/search, and summary-unaffected-by-searches.
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
Fixes the issue where the agent gets stuck in an infinite reading loop after context compression, re-reading the same files endlessly without writing or responding.
Root cause: Context compression summarizes conversation history but loses track of which files were already read. After compression, the model thinks it hasn't examined the files yet and reads them again. Combined with todo re-injection of completed items, this creates an infinite loop.
Fix (multi-layered):
Read tracking with escalation (
tools/file_tools.py): Track file reads per task. 2nd read returns a soft warning with content. 3rd+ read blocks — returns error with no content, forcing the model to stop.Search tracking (
tools/file_tools.py): Same mechanism forsearch_files— identical searches are warned then blocked after 3 repeats.File history injection (
run_agent.py): After context compression, inject a structured message listing all files already read with "do NOT re-read" instruction.Todo re-injection filtering (
tools/todo_tool.py):format_for_injection()now filters out completed/cancelled todos. Only pending/in_progress items are re-injected after compression, preventing the model from re-doing finished work.Design decisions:
_read_tracker_lockTest plan