Background And Related Work
Last updated: 2026-07-15T0200-07:00 Source/command: primary papers, official proceedings/project documentation, official npm metadata, and the detailed reports linked below Completeness: first BOOTSTRAP coverage pass complete; RECAP artifact availability and external user-study recruitment remain unresolved
Detailed reports:
Search Log
| Date | Query/source | Purpose | Result |
|---|---|---|---|
| 2026-07-15 | SeeSoft, Evolution Matrix, CodeCity, EvoStreets, Software Cartography, History Flow, code_swarm, Ownership Map, Evolution Storylines | Verify the seven historical families | Primary metadata and author/official PDFs verified; several user-supplied dates were corrected or made precise. |
| 2026-07-15 | coding-agent trajectory analysis, replay, observability, survival, long-horizon | Find same-problem and same-claim work | RECAP is the closest process-replay system; recent trajectory and survival papers create high empirical-overlap risk. |
| 2026-07-15 | Git visual analytics, evaluation of software visualization | Find commit-only baselines and accepted protocols | Githru and the Merino et al. review imply real systems, task accuracy/time, case studies, usability, and engagement evidence. |
| 2026-07-15 | official ECharts, D3, Cytoscape.js, uPlot, Perfetto, Gource, Hercules docs and npm metadata | Bound custom implementation | Existing libraries cover charts, stable/preset graphs, large time series, trace timelines, animation, and Git burndown/coupling. |
PDF Corpus
No PDFs are committed. Eight open-access, high-risk or method-relevant papers
are retained locally under ignored docs/reference/ for full-text claim and
protocol checks. Durable claims below link public primary pages, and each active
BibTeX entry records its verification source and local-PDF availability.
Claim-Oriented Novelty Map
| Claim | Closest prior work | Same-claim risk | Novelty delta | Baselines implied | Expansion opportunity |
|---|---|---|---|---|---|
| Joining fine-grained coding activity with code evolution recovers context unavailable from chat or Git alone. | RECAP | High | Cross-vendor native histories; repository-level multi-session evolution; explicit recorded-event/candidate-actual-Git/endpoint mismatch. | RECAP position, native event table, Git-only view. | Treat mismatch and durable survival as first-class evidence, not only replay alignment. |
| Coordinated evolution views improve long-horizon process review. | Githru, RECAP, AgentSeer | Medium-high | Seven coordinated software-evolution projections over event-plus-Git data, shared stable layout and time cursor, decision-specific tasks. | Githru/Git-only, RECAP-style linear replay, event table. | Establish which review questions require a joined representation rather than claiming visual novelty alone. |
| Event granularity exposes stable coding-agent behavior patterns. | Understanding Code Agent Behaviour, Agent trajectories as programs | High | Multi-session repository growth, durable outcomes, and local system evidence rather than benchmark issue trajectories alone. | Published action taxonomies and trajectory statistics. | Link process patterns to commit/survival outcomes and review decisions. |
| Agent-associated code has measurable survival and ownership dynamics. | Will It Survive?, Code Lifespan Survival Analysis | High | Visualization and event-to-outcome provenance rather than a new survival-analysis claim. | Published survival protocols and replication packages. | Use survival as one coordinated projection and validity check; do not claim first agent-code survival study. |
| The system remains interactive across weeks or months. | Perfetto, Githru | Medium | Repository/event semantic aggregation with stable spatial views, not general trace rendering. | Perfetto trace export and raw browser benchmark. | Publish a reusable longitudinal dataset/format if privacy permits. |
Closest Work
| Work | Claim | Method/artifact | Evaluation | Relation | Gap relative to this project |
|---|---|---|---|---|---|
| RECAP (2026) | Chat or Git alone cannot reconstruct AI-assisted programming context. | VS Code extension records Copilot chat plus shadow-Git edits; web timeline replay and analyses. | 41 students, two-week project, 2,034 prompts and 8,239 edits. | Same problem and core join mechanism. | Copilot/VS Code-specific; focuses developer-AI interactions and replay, not cross-agent repository evolution, actual Git outcomes, blame, or current survival. |
| Githru (TVCG 2021) | Git metadata needs scalable visual analytics. | Commit-graph reconstruction/clustering, summaries, file hierarchy, comparisons. | Domain expert cases and controlled study with 12 developers. | Same evaluation tasks and Git setting; commit-only mechanism. | No agent events or process/outcome mismatch. |
| AgentSeer (AAAI 2026 demo) | Raw spans are insufficient for agent observability. | Temporal action and component graphs plus action-level red teaming. | Six-agent demonstration. | Same action observability, different domain goal. | No coding-repository evolution or longitudinal Git join. |
| Understanding Code Agent Behaviour (ICSE 2026 manuscript) | Success rates hide meaningful trajectory structure. | Normalized OpenHands/SWE-agent/Prometheus trajectories and manual/statistical analyses. | SWE-bench trajectories; success/failure and localization analyses. | Same event-pattern problem. | Benchmark tasks rather than days-to-months product evolution; no visualization artifact or durable history. |
| Will It Survive? (EASE 2026 manuscript) | Agent-authored code longevity differs from human code. | File/line survival over 201 repositories and 200k+ units. | Kaplan--Meier, Cox models, modification taxonomy, released package. | Same survival question. | No process events; authorship starts from agent PR labels. This project must not claim novelty for survival analysis itself. |
| Hercules | Full Git history supports fast burndown, ownership, coupling, and churn analysis. | Go analysis DAG plus plotting tools. | Open-source system and documented large-repository runs. | Same Git-derived metrics and several view families. | No native agent events; inactive release line increases reproduction risk. |
Historical Foundations
- SeeSoft maps each source line to a thin colored row and reports interaction over up to 50,000 lines.
- The Evolution Matrix represents versions and classes/files as a matrix and classifies recurring evolution shapes.
- CodeCity, EvoStreets, and Software Cartography establish the city metaphor, incrementally stable layouts, and semantically meaningful stable maps respectively.
- History Flow makes contribution survival visible; code_swarm studies organic animated histories; Software Evolution Storylines uses storyline/metro-map conventions to show developer interaction.
- Ownership Map connects repository changes to developer knowledge and responsibility.
Mandatory Baselines
| Baseline | Official artifact/version | Runnable status | Visible information | Tuning surface | Protocol | Risk | Consequence |
|---|---|---|---|---|---|---|---|
| Git-only | local Git plus official git log, diff, blame | Runnable | Commits, files, lines, authors, current survival | time range, rename threshold, first-parent/all | Same repository/time range; no agent fields | Merge/squash and rename ambiguity | If it answers process tasks equally well, the event-level claim fails. |
| Native event table | agent-session normalized JSON | Runnable after exporter | prompts, tools, paths, status, tokens, timestamps | filtering/aggregation only | Same events without coordinated views or Git join | Raw logs can leak private text | If it matches the gallery, visualization utility is unsupported. |
| Perfetto timeline | current official browser UI; legacy Trace Event JSON accepted | Runnable export; external UI | timestamped tracks and event detail | track mapping | Same event set converted without gallery-specific views | Not repo spatial/survival aware | Establishes whether custom timeline work is necessary. |
| Gource | 0.56 | Not locally installed; custom-log export is feasible | animated file tree and actors | time scale, filters, colors | Feed equivalent Git and agent-touch logs | Poster-oriented and no analysis tasks | If animation alone answers tasks, richer playback may be unnecessary. |
| Hercules | v10.7.2 | Not installed; Docker/source possible | Git burndown, ownership, couples, churn | sampling/granularity/identity map | Cross-check Git-derived aggregates on one repository | Last release 2020; optional TensorFlow plotting deps | Metric disagreement blocks interpretation until resolved. |
| RECAP position | arXiv:2605.01104 | Paper verified; public code link not found in full text | merged chat/edit replay and analysis | system-specific | Compare supported questions and information model, not fabricated runtime numbers | Artifact availability unresolved | Claims must remain differentiated even without a runnable comparison. |
Experimental Precedents And External Assets
| RQ | Accepted paper/protocol | Asset | Reusable design | Required deviation/glue |
|---|---|---|---|---|
| RQ1 | RECAP's parallel streams; Githru's Git-history abstraction | Local native sessions; Git plumbing; RECAP paper | Explicitly compare information available from each source and the join | Add observed/committed/surviving mismatch audit and rename handling. |
| RQ2 | ICSE trajectory study and published action taxonomies | SWE-bench experiment trajectories where licensing permits | Predefine action classes; compare success/failure and agents | Extend from single issue attempts to multi-session repository evolution and durable outcomes. |
| RQ3 | Merino et al. evaluation guidance; Githru controlled tasks | Real OSS repositories and recorded task answers | Accuracy/time plus case study, usability, recollection/engagement | Use process-review tasks and include event-table and Git-only baselines. |
| RQ4 | Perfetto large-trace UI and Githru graph abstraction | Synthetic scale-up from real event distributions plus full local history | Report throughput, size, latency, memory and semantic zoom | Add stable spatial layouts and linked selection benchmarks. |
| Survival projection | Will It Survive replication package; CLSA Zenodo package | Agent-labelled PR histories and line-survival protocol | Rename/refactoring-aware lineage checks and censored survival | Treat as validation/projection, not central novelty or causal agent attribution. |
Absorbable Ideas
| Source | Idea to absorb | Claim expansion | Experiment implication | Risk |
|---|---|---|---|---|
| RECAP | Shadow history preserves discarded work. | Model observed-but-not-committed changes explicitly. | Quantify each mismatch category. | Native session logs may lack exact edit snapshots. |
| Githru | Adjustable history abstraction and task-derived requirements. | Tie semantic zoom to review tasks. | Ablate stable aggregation and measure task effects. | Reimplementing its commit graph would bloat scope. |
| Software Cartography | Stable coordinates enable comparison and spatial memory. | One layout across time and views. | Compare stable versus recomputed layout. | Semantic layouts can surprise users; hierarchy-first layout is safer initially. |
| Hercules | Incremental burndown and explicit sampling/granularity. | Reuse verified Git metrics instead of inventing survival math. | Cross-check aggregates and report sampling settings. | Old artifact and heavy optional dependencies. |
| Trajectory studies | Action sequences and repeated edits distinguish agent behavior. | Connect behavior to durable repository outcomes. | Define patterns before final data inspection. | Post-hoc pattern naming would invalidate inference. |
Adjacent Communities
| Community | Why relevant | Keywords | Useful sources |
|---|---|---|---|
| VIS/TVCG/visual analytics | Coordinated views, dynamic graphs, evaluation | semantic zoom, insight evaluation, stable mental map | Githru, visual-analytics evaluation surveys |
| VISSOFT/ICSE/MSR | Software history, program comprehension, empirical protocols | evolution, repository mining, ownership, survival | Seven historical families, trajectory studies, survival papers |
| Systems observability | Large event streams and trace interaction | timeline, trace query, cross-layer correlation | Perfetto and Trace Event format |
| HCI/CS education | Naturalistic AI-assisted programming and replay studies | process replay, longitudinal classroom deployment | RECAP and History Flow |
Venue Evaluation Patterns
The strongest cross-domain pattern is a mixed evaluation: real-system case studies for open-ended discovery, controlled tasks for time/accuracy, qualitative strategy evidence, and computational scalability. The 2018 systematic review found that many software-visualization papers lacked strong evaluation and explicitly recommends real open-source systems and controlled experiments when variables can be controlled. A poster/demo-only gallery would not support the paper's target claim.
Must-Read List
- RECAP (highest same-mechanism risk).
- Githru (closest visual-analytics and evaluation precedent).
- Merino et al. evaluation review (protocol obligation).
- Understanding Code Agent Behaviour and Agent Trajectories as Programs (behavior-pattern overlap).
- Will It Survive? and CLSA (survival overlap and AST-aware line matching).
- SeeSoft, Evolution Matrix, Software Cartography, History Flow, code_swarm, Ownership Map, and Evolution Storylines (design foundations).
Novelty Verdict
- Overall same-claim risk: high if framed as replay or agent-code survival; medium if the contribution is the explicit cross-vendor event-to-durable- outcome representation plus demonstrated review utility.
- Ambitious target claim: coordinated event-plus-Git evidence enables process-review judgments that commit-only and event-only interfaces cannot reliably support over long horizons.
- Claims requiring stronger evidence: join fidelity, pattern predefinition, human review utility, and interaction scale.
- Larger opportunity: make disagreement between observed actions, committed changes, and surviving code an empirical object rather than hiding it.
- Mandatory baselines: Git-only, event table, Perfetto, Gource where applicable, Hercules metric cross-check, and a position-level RECAP comparison.
- Next action: preserve the current strong thesis, implement the minimum joined representation needed for an RQ1 real preflight, and do not claim authorship or causality from timestamps alone.
继续阅读
返回索引
AgentSight:基于 eBPF 的零侵入 LLM 智能体可观测性工具
 
上一篇 / 上一页
Supported Agents
AgentSight works with any process that makes TLS-encrypted API calls. This page covers agent-specific setup and quirks.
下一篇 / 下一页
Build From Source
Use this guide when developing AgentSight or building a local binary from the repository. If you only want to run a release binary, see the Quick Start in README.md.
- 最后更新
- 2026年7月15日
- 首次发布
- 2026年7月15日
- 贡献者
- github-actions[bot]
这个页面有帮助吗?