Chronos Enables Code Agents to Reason over Software Evolution

Xin Yin; Yiang Zhang; Zhiyuan Peng; Chao Ni; Zhe Cui; Xiaohua Xin

cs.SEOct 8, 2026

Chronos Enables Code Agents to Reason over Software Evolution

Authors: Xin Yin, Yiang Zhang, Zhiyuan Peng, Chao Ni, Zhe Cui, Xiaohua Xin

Organizations: Zhejiang University, China · National Industrial Information Security Development Research Center, China

Abstract

Historical pull requests record the design decisions, compatibility constraints, and implementation patterns behind a codebase's current state. Experience relevant to a new task can span related changes whose descriptions emphasize different concerns. We introduce Chronos, a test-time framework that makes this connected history available to large language model (LLM)-based code agents. Chronos distills merged pull requests into structured experience cards and connects them through a typed graph of code-level, developer-intent, and organizational relations. Semantic search identifies entry cards, and weighted multi-hop expansion retrieves connected changes for selective reading. The same memory guides candidate generation and patch selection: a patch-focused change agent and a validation-strategy agent each develop a patch, and an evolution steward consults history to select between them. On SWE-Bench Verified, the full workflow improves SWE-Agent across all six evaluated LLM backbones, raising the mean resolution rate from 69.2% to 72.9% and reaching 79.8% with MiniMax M2.5. With the same backbone, it raises resolution rates from 48.3% to 51.7% on SWE-Bench Pro and from 41.0% to 43.5% on FEA-Bench Lite. Both experience-guided single-agent variants also outperform the base agent. In a human evaluation on 100 tasks with ten cards retrieved per task, graph-grounded retrieval increases the mean number of useful cards from 1.24 to 2.87 over flat semantic retrieval. These results demonstrate the value of PR relations for retrieving useful repository experience and of the evaluated workflows for applying that experience during patch generation and selection.

Figures & tables

Figure 1. Resolution rates of representative agents across three benchmarks. Labels identify the backbone and scaffold of each configuration; Tables 4 and 5 give our backbone-specific comparisons. Each system submits one final patch per task. Colors distinguish Chronos, open-scaffold systems, and proprietary systems. Comparison of Chronos with open-scaffold and proprietary code agents on three software-engineering benchmarks. Colors identify the three system groups, and the bars display resolution rates with one final patch submitted per task.

Figure 2. Overview of Chronos. PR-level experience cards are connected by a typed evolution graph. Search and read tools expose this memory to candidate-generating agents and an evolution steward that selects the final patch. Two-stage Chronos framework. Historical pull requests are distilled into experience cards and connected in a typed evolution graph. Task-time agents retrieve historical experience, generate candidate changes, and select a final patch through an evolution steward.

Tier	Edge type (forward / reverse)	Criterion	Cov.	Weight
T1	precedes / succeeds	File-level Jaccard overlap $\geq 0.5$ between PRs.	$\sim$ 62%	1.0–1.5
	extends / extended_by	A later PR continues development on files newly introduced by an earlier PR.	$\sim$ 13%	1.0
T2	follows / followed_by	Same author, time gap $\leq$ 60 days, and directory Jaccard $\geq 0.5$ .	$\sim$ 73%	0.70–1.05
	backport_of / backported_to	Backport cues detected from title, body, commit messages, or base branch.	$\sim$ 5%	0.70
	references_to / referenced_by	Explicit #N citation in PR body or commit messages.	$\sim$ 46%	0.70
T3	same_reviewers	Reviewer overlap (Jaccard $\geq 0.5$ ).	$\sim$ 28%	0.30–0.45

Table 1. Software evolution graph edge taxonomy. Coverage reports the fraction of nodes with at least one outgoing edge of each type. Weights are assigned to tiers based on evidential strength.

Candidate outputs	Selection action
Two identical, non-empty patches	Retain the shared patch directly.
One non-empty patch	Retain the available patch directly.
Two empty patches	Submit an empty patch.
Two distinct, non-empty patches	Randomize their X/Y order and invoke the steward.

Table 2. Routing of candidate patches before final submission. Equality is checked after stripping surrounding whitespace.

Benchmark	Tasks	Repos.	Task emphasis
SWE-Bench Verified	500	12	Human-verified issue resolution
SWE-Bench Pro (public)	731	11	Longer-horizon repository-level changes
FEA-Bench Lite	200	48	New feature implementation

Table 3. Benchmark subsets used in the evaluation, with task and repository counts.

Panel of Table 4 — Table 4. Resolution rates (%) on SWE-Bench Verified. $\Delta$ is the gain in percentage points.

Benchmark	SWE-Agent	Chronos	$\Delta$
SWE-Bench Pro	48.3	51.7	$+3.4$
FEA-Bench Lite	41.0	43.5	$+2.5$

Table 5. Resolution rates (%) on the transfer benchmarks using MiniMax M2.5.

\Delta

is the gain in percentage points.

Configuration	Rate (%)	$\Delta$
Chronos (SWE-Agent)	79.8	-
Chronos (mini-SWE-Agent)	79.2	-
w/o validation-strategy agent	78.4	$-1.4$
w/o patch-focused change agent	77.8	$-2.0$
w/o software-evolution graph	79.0	$-0.8$
SWE-Agent	76.4	$-3.4$

Table 6. Component and scaffold comparisons on SWE-Bench Verified with MiniMax M2.5. Single-agent variants retain graph retrieval and produce one patch without steward selection.

\Delta

is relative to full Chronos using the same scaffold.

Retrieval Method	Useful	Hit
Random	0.18	14%
Embedding	1.24	68%
Chronos	2.87	89%

Table 7. Human evaluation of retrieval quality on 100 randomly sampled SWE-Bench Verified instances. Useful is the average number of useful cards among the top-10 retrieved; Hit is the fraction of instances with at least one useful experience in the top-10.

Figure 3. Per-edge-type ablation. Bars show the decrease in Useful (average useful cards among the top-10 retrieved) after removing each relation; larger drops indicate a greater loss of useful cards under the tested configuration. Vertical bar chart grouped by relation tier. The decrease in Useful is 0.60 for precedes and 0.44 for extends in Tier 1; 0.28 for follows, 0.40 for references, and 0.22 for backport in Tier 2; and 0.18 for same reviewers, 0.08 for same discussants, 0.14 for same milestone, 0.10 for similar labels, 0.12 for mentions in comments, and 0.06 for depends on in Tier 3.

Explore similar work

CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents
Aug 6, 2026Wuya Chen, Yihao yang, Yang Cao +1AI Coding Agents LLM Agent Workflow Optimization
1Netease Guangzhou AI Lab · 2Independent Researcher
Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments
Jul 30, 2026Haomin Qi, Xingliang Wang, Xuanqi Gao +9Code Generation Benchmarks Agent Benchmarks
1Microsoft · 3Zhejiang University · 4Xi’an Jiaotong University +1
RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations
May 25, 2026Hanyu Li, Yichi Zhang, Speed Zhu +3AI Coding Agents Software Engineering Benchmarks
Beijing University of Posts and Telecommunications · Tsinghua University · Tencent

Aug 6, 2026, cs.SEJ/K move · Enter open · S save

CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents

Wuya Chen, Yihao yang, Yang Cao, Yue Lin

1Netease Guangzhou AI Lab · 2Independent Researcher

AI Coding Agents LLM Agent Workflow Optimization RL for LLM Agents

Open paper AskPDF arXiv

Modern LLM coding agents such as Claude Code and OpenHands share a common inefficiency: they spend much of their token budget finding the file to patch, rather than patching it. On SWE-Bench Verified, a 30B OpenHands agent averages 23 rounds and 631K tokens per resolved issue, with many calls spent on grep, glob, and view_file during repository exploration. We introduce CodeGrep, a 14B retrieval agent trained end-to-end with GRPO to issue multi-turn parallel grep, glob, and read tool calls and return candidate files to a frozen downstream coding agent. On all 500 SWE-Bench Verified instances, CodeGrep preserves resolve rate while substantially improving efficiency: 27.0% versus 25.8% for the no-retrieval baseline, with 15% fewer rounds and 19% fewer tokens on resolved instances. Across retrievers, downstream utility follows a precision threshold: BM25 with precision 0.375 degrades the agent, Jina with precision 0.445 is neutral, and CodeGrep with precision 0.677 crosses the threshold at which retrieval begins to reduce rollout cost. To enable this study, we mine supervision from 67K open-source agent trajectories using CATM and build a Git-worktree environment for multi-turn agent RL. In our setting, applying the efficiency signal at the advantage layer rather than the reward layer reduces KL drift and translates cleanly into downstream efficiency. We will release the model, training pipeline, RL environment, and evaluation harnesses.

Abstract

Explore similar work

CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents

Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments

RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations