AI Time Machine

AI Time Machine turns real Git history into an interactive, evidence-backed timeline explaining why a codebase evolved.

Public demo: https://ai-time-machine-demo.vedheshvit.chatgpt.site

Demo video: https://youtu.be/5LGXtmJNO0U

This repository contains an OpenAI Build Week developer tool: a generated OrbitCart Git repository, Git ingestion, an evidence timeline, reference-validated GPT-5.6-in-Codex artifacts, Ask the Repo, and a visual Bug Origin Trace.

How Codex and GPT-5.6 were used

Codex supported architecture exploration, implementation, debugging, UI refinement, test development, trust-semantics review, and grounding evaluation.

GPT-5.6 Sol was used through ChatGPT-authenticated Codex to generate the committed OrbitCart causal-analysis and Ask the Repo artifacts from repository evidence. This generation happened at build time.

The hosted application does not call GPT-5.6 live. It replays an artifact only after validating its evidence digest, event IDs, commit references, and affected files. If validation fails, the application uses a clearly labeled deterministic fallback instead of displaying unsupported model output.

Real Repo Mode

Point the local tool at any Git worktree. The required flow uses only Python and Git; it does not require GitHub OAuth, an OpenAI API key, or a paid runtime call.

python3 -m app.cli analyze /path/to/repository
python3 -m app.cli analyze /path/to/repository --branch feature/my-work
python3 -m app.cli serve /path/to/repository
python3 -m app.cli serve --open
python3 -m app.cli context /path/to/repository --base main --head HEAD

analyze prints the normalized timeline JSON and supports --output. serve opens the existing interface at http://127.0.0.1:8765 with the selected repository fixed at process start. context reports changed files, commits in the range, recent commits touching those files, connected incidents or fixes, and risks that Git actually records.

For ordinary repositories, the Developer Workspace adds a visual Review my branch report and four adaptive deterministic questions. Their answers cite clickable commits and same-event files and are labeled Local evidence engine · deterministic; they are not GPT output. serve defaults to the current worktree, and --open launches the local page automatically.

Ordinary commit subjects and diffs remain confirmed Git evidence. Conventional commit classifications are marked inferred, while absent rationale and risk are shown exactly as not recorded. OrbitCart-only Ask the Repo and Bug Origin artifacts are not exposed for generic repositories.

The workflow is dogfooded on this repository: its M0-M4 commits render in chronological order, shared files connect milestones, and the foundation-to-HEAD context report is grounded in real changed files. See the M4.5 implementation brief.

Quick start

Requirements: Python 3.11+ and Git. No third-party packages are required.

git clone https://github.com/ved-devAI/AI_Time_Machine.git
cd AI_Time_Machine
python3 scripts/create_orbitcart.py
python3 -m app.server

Open http://127.0.0.1:8765.

The local app reads the generated OrbitCart Git repository at request time. The public demo is an API-free snapshot produced from that same repository during deployment. Its timeline and Codex references are checked before publication, and the UI labels hosted evidence as a verified Git snapshot rather than a live model call.

Quick test without rebuilding

Analyze the current Git worktree directly, or open it in Real Repo Mode:

python3 -m app.cli analyze .
python3 -m app.cli serve --open

These commands read the existing repository history and do not regenerate the OrbitCart demo repository.

Verify everything

Run the complete zero-dependency verification suite:

python3 scripts/verify.py

This regenerates OrbitCart, runs all application and repository tests, validates both Codex artifacts, produces a deterministic grounding scorecard, and checks the browser JavaScript and repository whitespace.

Individual commands:

python3 -m unittest discover -s tests -v
PYTHONPATH=.data/orbitcart python3 -m unittest discover -s .data/orbitcart/tests -v
python3 scripts/codex_artifact.py validate
python3 scripts/ask_repo_artifact.py validate
python3 scripts/evaluate_grounding.py

How it works

1. scripts/create_orbitcart.py creates a genuine 12-commit demo repository.
2. app/git_ingest.py reads generic commit metadata, changed files, overlap history, and branch/range context using Git.
3. scripts/codex_artifact.py exports evidence and validates a strict, reproducible GPT-5.6-in-Codex artifact.
4. app/analysis.py verifies artifact provenance, evidence digest, event IDs, commit hashes, and file references before returning an investigation.
5. app/ask_repo.py serves three reference-validated, evidence-linked repository answers.
6. app/repo_questions.py produces and validates four deterministic answers for ordinary repositories.
7. app/cli.py selects a local repository for analysis, context, or serving.
8. app/server.py exposes the selected timeline and branch review while restricting reference-validated OrbitCart artifacts to OrbitCart.
9. frontend/ renders the Developer Workspace, Ask the Repo, the timeline, and the Bug Origin Trace.

The default flow makes no paid runtime call. It replays the committed artifact generated by GPT-5.6 Sol through ChatGPT-authenticated Codex. If the artifact is missing or fails validation, the investigation remains runnable using an explicitly labeled local evidence fallback.

Reproduce the Codex artifact

python3 scripts/codex_artifact.py prepare
codex exec -m gpt-5.6-sol -s read-only \
  --output-schema artifacts/orbitcart/analysis.schema.json \
  -o .data/codex-run/analysis.json \
  "Read artifacts/orbitcart/analysis.prompt.md and perform that task."
python3 scripts/codex_artifact.py finalize .data/codex-run/analysis.json \
  --model gpt-5.6-sol
python3 scripts/codex_artifact.py validate

Ask the Repo uses the same workflow with artifacts/orbitcart/ask-repo.prompt.md, scripts/ask_repo_artifact.py, and its own strict output schema. Both committed artifacts are tied to the same OrbitCart evidence digest.

codex exec reuses ChatGPT-managed Codex authentication. The app does not need an API key to replay the reference-validated result.

See the analysis design for grounding and fallback details and the evaluation guide for the scorecard.

Public deployment

python3 scripts/create_orbitcart.py
python3 scripts/build_public_demo.py

The build writes a static client and a minimal host worker to dist/. Pushes to main run the full verification suite; production releases package the same reference-validated build for the public host. No API key or paid runtime call is required.

Release screenshots

• Desktop product overview
• Bug Origin Trace
• 390 × 844 mobile trace