PR #13321: fix(batch_runner): don't re-append per-batch completed_prompts to the final checkpoint

• Repository: NousResearch/hermes-agent
• Author: Sanjays2402
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/13321

Description (problem / solution / changelog)

Summary

Closes #13285.

batch_runner.py's incremental checkpoint path maintains completed_prompts_set as a deduplicated set of every completed prompt ID (lines 925-939, written to disk as sorted(completed_prompts_set) at line 941). The final-aggregation path, however, did:

all_completed_prompts = list(completed_prompts_set)
for batch_result in results:
    all_completed_prompts.extend(batch_result.get('completed_prompts', []))

…which re-appends every prompt that's already in the set — so the final checkpoint contains each completed ID twice. Checkpoint consumers expect completed_prompts to be the set of completed items; duplicates inflate counts and complicate resume/progress logic.

The fix

Drop the extend loop, write sorted(completed_prompts_set) directly. Matches the shape already used by the incremental write at line 941 and keeps output deterministic for easier diff/review of checkpoint files.

Diff

-        # Aggregate all batch statistics and update checkpoint
-        all_completed_prompts = list(completed_prompts_set)
+        # Aggregate all batch statistics and update checkpoint.
+        # NB: completed_prompts_set is already maintained as a dedup'd
+        # set of every completed prompt ID; we must NOT extend this list
+        # again from each batch result below, or the final checkpoint
+        # will contain duplicates.  See #13285.
         total_reasoning_stats = {...}

         for batch_result in results:
-            # Add newly completed prompts
-            all_completed_prompts.extend(batch_result.get("completed_prompts", []))
-
             # Aggregate tool stats
             ...
...
-        # Save final checkpoint (best-effort; incremental writes already happened)
+        # Save final checkpoint (best-effort; incremental writes already happened).
+        # Use the already-deduplicated set, sorted for deterministic output,
+        # matching the format written by the incremental checkpoint path above.
         try:
-            checkpoint_data["completed_prompts"] = all_completed_prompts
+            checkpoint_data["completed_prompts"] = sorted(completed_prompts_set)
             self._save_checkpoint(checkpoint_data, lock=checkpoint_lock)

1 file changed, +11, −8.

Verification

Using the reporter's minimal repro shape:

completed_prompts_set = {'p0'}
results = [{'completed_prompts': ['p0']}]
# Before: list(completed_prompts_set) + extend -> ['p0', 'p0']
# After:  sorted(completed_prompts_set)         -> ['p0']

• python3 -m py_compile batch_runner.py → clean.
• Confirmed all_completed_prompts is no longer referenced anywhere else in the file (grep returns nothing).
• Incremental checkpoint behavior at lines 925-941 is unchanged.

Linked Issue

• Closes #13285

Changed files

• batch_runner.py (modified, +11/-8)

PR #13324: fix(batch): dedupe final checkpoint completions

• Repository: NousResearch/hermes-agent
• Author: sgaofen
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/13324

Description (problem / solution / changelog)

Summary

• Fixes #13285
• Deduplicates completed prompt IDs when writing the final batch checkpoint
• Keeps incremental checkpoint recovery behavior intact while preventing final aggregation from re-adding the same prompt IDs

Root Cause

The parent process updated completed_prompts_set during incremental checkpoint writes, then initialized final aggregation from that same set and extended it again with each batch result. That made final completed_prompts contain duplicates.

Tests

• uv run --python 3.11 --extra dev pytest tests/test_batch_runner_checkpoint.py -q (14 passed)
• git diff --check -- batch_runner.py tests/test_batch_runner_checkpoint.py

Changed files

• batch_runner.py (modified, +12/-5)
• tests/test_batch_runner_checkpoint.py (modified, +15/-1)