Idea And Hypothesis History
Initial Narrative
Problem and stakes
Long-horizon coding agents can modify a repository across many sessions, days, and models. The resulting commits show what snapshots survived, while native agent logs show a large event stream that no reviewer can read end to end. The development rationale can therefore disappear as soon as a session ends: humans inherit code without having observed the exploration, false starts, verification gaps, or cross-file information flow that produced it.
Challenged belief
Software-evolution visualization has usually treated commits as the natural unit of observation and people as the persistent authors of a system. The project challenges the belief that simply applying existing commit-centric views to agent-authored repositories is sufficient.
Central thesis
Agent events and version-control history are complementary evidence. Joining timestamped read/edit/tool/model events with commits, file births and deaths, blame, and the current tree exposes process structure that commits cannot contain while retaining the durable outcome that agent logs cannot prove. A stable set of coordinated views can compress this joined history into a form that supports both exploratory discovery and concrete review decisions.
Proposed system
Extend agent-session with a longitudinal exporter that joins native Claude,
Codex, and Gemini sessions to Git history and the current repository state. A
static, locally served visualization gallery reuses established browser
libraries and interchange formats to provide pixel/line views, evolution
matrices, stable repository maps, animated growth replays, survival strata,
forensic hotspots and coupling, ownership/storyline views, and coordinated
long-horizon navigation. Every view states the review decision it supports.
Intended contributions
- An event-resolved longitudinal representation that preserves both agent process evidence and Git-grounded software outcomes.
- A coordinated visualization design that adapts seven established software evolution traditions to multi-session coding agents while preserving stable layouts and time navigation.
- An empirical evaluation of information loss at commit granularity, behavior-pattern observability, diagnostic usefulness, and scalability on real multi-day development histories.
Scope
The first artifact targets local coding-agent transcripts and Git repositories. It does not claim that temporal order alone proves semantic causality, that a session event proves a change was committed, or that inferred agent authorship is equivalent to signed provenance. These distinctions are explicit fields in the joined representation.
Provisional research questions
- RQ1 — Recoverability: What development-process evidence is present in agent events but absent from commits, and how reliably can events be joined to Git-grounded outcomes?
- RQ2 — Behavioral structure: Which recurring exploration, churn, verification, coupling, and survival patterns become observable at event granularity, and are they stable across sessions and agents?
- RQ3 — Review utility: Do coordinated event-plus-history views improve the accuracy and time required to answer realistic process-review questions compared with logs and commit-centric views?
- RQ4 — Long-horizon scalability: Can the representation and interactions remain responsive and interpretable across days to months and large event volumes?
Evaluation promise
The paper will report real multi-day repositories, an explicit join-quality audit, commit-only and log/table baselines, controlled diagnostic tasks, and runtime/interaction measurements. All missing values, plots, and result-dependent claims remain placeholders until complete final runs.
Belief And Principle Evolution
| Date | Prior belief/model | External evidence | Updated principle | Paper impact |
|---|---|---|---|---|
| 2026-07-15 | Commits are the default evolution unit. | Pending literature grounding and real-data audit. | Test events and commits as complementary evidence. | Defines RQ1 and the joined representation. |
| 2026-07-15 | Fine-grained replay might itself be a novel contribution. | RECAP already joins Copilot conversations with observed shadow-repository edits over a two-week project. | Replay is a baseline capability; the differentiator must be cross-vendor native events, candidate associations to actual Git outcomes, and separately labeled endpoint survival. | Narrows the thesis and makes RECAP the highest-risk closest work. |
| 2026-07-15 | Timestamped reads and edits might reveal causal coupling or agent authorship. | Native schemas and actual Git history do not provide signed identity or causal intervention evidence. | Treat read-before-edit as ordered evidence and event-to-Git links as zero/one/many candidates with confidence. | Removes causal/authorship claims and makes ambiguity visible. |
| 2026-07-15 | Path survival and line survival could share one join. | Rename/refactoring work and CLSA show that event-to-Git matching and Git-to-current lineage fail for different reasons. | Evaluate event-to-Git association separately from Git hunk-to-current-line lineage; unsupported events remain path-level. | RQ1 becomes a gate for downstream claims. |
| 2026-07-15 | A broad gallery can establish utility by inspection. | Software-visualization evaluation work and Githru use real systems and controlled tasks; a gallery alone does not show decision value. | Freeze four task-facing core combinations and treat other views as exploratory/communicative support. | RQ3 separates information content from coordinated interaction. |
Hypothesis Frontier
| ID | Parent | Prediction | Falsifier | Evidence for/against | Status | Decisive next test | Reopen condition |
|---|---|---|---|---|---|---|---|
| H1 | root | Event-plus-Git views expose review-relevant process structures that commit-only views omit. | Process questions are answered equally well from commit-only views, or event joins are too unreliable. | Pending. | leading | Build joined dataset and measure recoverability on real history. | Revisit after RQ1 final run. |
| H2 | root | Most useful longitudinal insight comes from durable Git outcomes; fine-grained agent events add little beyond animation. | Event-only structures materially improve diagnosis and remain stable across histories. | Pending. | competing | Compare matched tasks across commit, event table, and coordinated views. | Revisit after RQ3 protocol. |
| H3 | H1 | Stable layout plus semantic zoom makes weeks of events usable without discarding local detail. | Interaction latency or visual crowding prevents task completion at target scale. | Pending. | stronger generalization | Benchmark aggregated frames and task navigation at increasing history sizes. | Revisit after RQ4 final run. |
Claim Evolution
| Date | Ambitious target claim | Evidence status | Unresolved uncertainty | Next evidence program |
|---|---|---|---|---|
| 2026-07-15 | Joined event and Git history forms a practical observation layer for long-horizon agentic software evolution. | Hypothesis only. | Join fidelity, generality, human utility, and scale. | Literature grounding, implementation, and RQ1 preflight. |
| 2026-07-15 | Uncertainty-preserving cross-vendor event-to-actual-Git evidence plus endpoint survival supports review decisions unavailable from either source alone. | Literature-differentiated hypothesis; no artifact result yet. | Candidate-link accuracy, line-lineage accuracy, stable pattern replication, task effect, and tested scale. | Controlled and naturalistic RQ1 ground truth before any event-to-outcome claim. |
Rejected Or Dormant Paths
| Path | Why rejected/dormant | Raw evidence | Revisit trigger |
|---|---|---|---|
| 3D-only code city | Depth and camera motion are not yet justified by a review task. | User discussion and pending literature verification. | A task study shows unique value over stable 2D maps. |
| Replay-first novelty | RECAP already supplies high-resolution conversation/edit replay. | Closest-work audit and full-text review. | New evidence shows a materially different replay task not covered by the joined representation. |
| First agent-code survival study | Recent survival studies already analyze agent-associated code at scale. | Will It Survive? and CLSA. | The project produces a distinct validated survival construct rather than an endpoint projection. |
| Timestamp-as-causality or authorship | Temporal proximity cannot establish why a change occurred or who authored the durable Git outcome. | Native schema and Git evidence audit. | Signed provenance or an intervention-based causal protocol becomes available. |
| One association state | Pathless, ambiguous, event-unmatched, Git-unmatched, path-level, and line-level records have different semantics. | Writing consistency and meaning-preservation audits. | RQ1 shows a safe deterministic collapse within a validated stratum. |
Narrative Evolution
- 2026-07-15 — closest-work narrowing. The Initial Narrative's joined- evidence thesis remains, but replay and agent-code survival are no longer novelty targets. The central empirical object is disagreement among recorded process events, candidate actual-Git outcomes, and separately derived current- tree endpoint survival.
- 2026-07-15 — evidence boundary. The representation now distinguishes association eligibility, zero/one/many candidate Git changes, Git-side unmatched changes, optional event-to-hunk association, Git line lineage, and endpoint survival. Recorded verification actions do not establish correctness, and ordered read-before-edit does not establish causality.
- 2026-07-15 — task-facing core. Four coordinated view combinations are frozen around concrete review decisions. The remaining seven-family gallery remains a broad exploratory and communication surface and is still an explicit artifact commitment.
Continue exploring
Back to index
AgentSight: System-wide AI agent profiling and monitoring with eBPF
  
Previous
AgentSight Flamegraph Gallery
AgentSight's semantic flamegraphs connect agent intent to observable activity. Each horizontal frame adds context to the stack, while frame width represents the selected metric: token volume, elapsed time, file
Next
Implementation Status
- Worktree: /home/yunwei37/workspace/agentsight-evolution-gallery - Branch: codex/vis-gallery, based on origin/master at 926c81c44. - Existing agent-session library parses Claude, Codex, and Gemini native histories into
- Last updated
- Jul 15, 2026
- First published
- Jul 15, 2026
- Contributors
- github-actions[bot]
Was this page helpful?