kc-academy/handoff-debt

Handoff Debt Trajectories

Execution records for Handoff Debt: The Rediscovery Cost When Coding Agents Take Over Interrupted Tasks. The code is available at https://github.com/anjilab/agent-handoff-debt.

This release follows the selected 75-task SWE-bench Verified pool. It includes Qwen, Gemma, and Devstral predecessors and successors, four handoff views, and repeated takeover attempts. Each takeover is linked to its predecessor and interruption point.

Family	Recorded model identifier
Qwen	openai/Qwen/Qwen3.6-27B
Gemma	openai/google/gemma-4-31B-it
Devstral	openai/mistralai/Devstral-Small-2-24B-Instruct-2512

The validation script checks these identities against the native agent configuration in each exported execution.

Coverage

Subset	Runs
predecessors	225
main_takeovers	2,172
repeated_takeovers	320
robustness_takeovers	3,948

These counts are generated from the source files during export. main_takeovers contains Qwen-predecessor r01 executions. robustness_takeovers contains the Gemma and Devstral predecessor r01 executions. repeated_takeovers contains r02 and later attempts. These labels describe the experiment design, not train/test partitions. The export uses the current execution files. Some executions were rerun, so their per-run outcomes and metrics can differ from aggregates in the submitted paper. The release manifest reports the actual files and coverage exported here.

Loading

import json
from datasets import load_dataset

runs = load_dataset("kc-academy/handoff-debt", "runs", split="records")
events = load_dataset(
    "kc-academy/handoff-debt", "events", split="records", streaming=True
)
event = next(iter(events))
payload = json.loads(event["event_json"])

See SCHEMA.md for table fields and joins. handoffs contains the actual rendered run prompts and supplied predecessor context. checkpoints describes the interruption rules and recorded validation outcomes. Both harness event records and native OpenHands trajectory records are provided in their original order.

Processing

The export reads the selected task list and the canonical run directories. Only exact rNN attempt directories are included. Archived summary-only copies, quarantine, unselected tasks, unrelated RL experiments, model-server logs, container states, and repository workspace copies are excluded.

Metadata uses an explicit field allowlist. Credential patterns, personal home paths, and private model endpoint URLs are replaced with labeled redaction markers in exported strings. No original execution files are changed. Repository code may contain public author emails. Those locations are flagged for review. Redaction can change text used by commands or tests, so these exports are analysis records, not a guarantee of byte-for-byte replay. File checksums support integrity checks.

Limitations and use

The runs use an OpenHands-style runtime and SWE-bench Verified tasks. They include unsuccessful runs and execution errors. Missing final validation is represented explicitly and must not be interpreted as a successful outcome. These are evaluation trajectories rather than curated expert demonstrations. Using them for training may contaminate evaluation on the same SWE-bench tasks. Split by source task, not by run or handoff view, when evaluating on held-out tasks.

License and publication status

This is a mixed-source collection, not a blanket MIT dataset. MIT covers only original author material as described in LICENSE. Copied and adapted third-party content retains its upstream terms, including GPL terms where applicable. See THIRD_PARTY_NOTICES.md. Upstream license and attribution texts are included under source_licenses/ and indexed in source_licenses.json. Runtime prompt attribution is retained in OPENHANDS_LICENSE. The official SWE-bench README describes its code and data and states MIT as the project license. We rely on that declaration for benchmark material and include its notice in SWEBENCH_LICENSE. Underlying third-party material retains its own terms.

Local release validation and contact-pattern review passed. The release is ready for upload. These checks are not a guarantee that every possible type of sensitive data is absent.

Citation

@article{kc2026handoff,
  title={Handoff Debt: The Rediscovery Cost When Coding Agents Take Over Interrupted Tasks},
  author={KC, Dipesh and Budathoki, Anjila},
  journal={arXiv preprint arXiv:2606.02875},
  year={2026}
}