GPTsAgent

Canonical development workspace for GPTsAgent.

GPTsAgent is an open project for building a disciplined GPT Builder wrapper around ChatGPT.com sandbox work. The current flagship package is Sandbox File Operator: a Custom GPT configuration that makes uploaded file and ZIP archive work more agent-like while staying honest about sandbox boundaries.

Current package version: v0.3.1.

This repository is a public pilot. It is intended to be cloneable, reviewable, and easy for contributors to improve without confusing it with a private host-level project.

The current operating model includes response shape discipline, long-running repository governance, and private-evidence-to-public distillation: concise direct answers for simple or prompt-only requests, handoff/archive reconciliation, authority-chain checks, stale metadata handling, CI/platform limitation classification, production-smoke evidence separation, bounded deep-update decomposition, and public-safe extraction of reusable patterns from maintainer/contributor-provided private evidence.

Download the full working directory as a ZIP:

https://github.com/GPTsAgent/GPTsAgent/archive/refs/heads/main.zip

Or clone it:

git clone https://github.com/GPTsAgent/GPTsAgent.git
cd GPTsAgent
python3 scripts/validate_workspace.py

No dependency installation is required for the core checks. Python 3.10+ and Git are enough. make targets are provided as shortcuts, but they are optional.

GPTsAgent intentionally keeps the public surface small:

The 20-file configuration and system instructions live inside this repository as config/ and instructions/.

• Public open-source pilot for ChatGPT.com Custom GPT configuration work.
• Validated as a 20-file Knowledge pack plus local release tooling.
• Builder and publishing choices should be rechecked at release time because OpenAI docs and GPT Builder constraints can change.

From the user's perspective, the deployed GPT is a static configuration package. The 20 Knowledge files and the GPT Builder Instructions are uploaded by a maintainer or admin during a release, then used as versioned runtime guidance inside ChatGPT.com.

End users do not edit those 20 files from inside the GPT session. When behavior needs to change, the maintainer updates the repository package, regenerates the Instructions export, and publishes a new reviewed release.

Sandbox File Operator does not collect, access, learn from, or reuse other users' private sessions. GPTsAgent maintainers cannot see other users' Sandbox File Operator chats through the GPT.

Normal chats do not update Knowledge, Instructions, future behavior, or this public repository. A maintainer or contributor may intentionally use voluntarily provided private evidence for a public-safe repository change, but only as a separate reviewed task with private details removed.

• Public Telegram group for discussion and coordination: https://t.me/GPTsAgentChat
• Use Telegram for quick, sanitized discussion, idea triage, and contributor coordination.
• Use GitHub Issues for tracked bugs, safety reports, and documentable change requests.
• Use the Contribution Idea issue template when a Telegram discussion should become tracked work.
• Use Pull Requests for actual changes.
• Do not post secrets, private keys, cookies, sessions, tokens, or private data in Telegram or GitHub.

Pick one small lane:

1. Fork GPTsAgent/GPTsAgent.
2. Create a branch.
3. Run python3 scripts/check_open_prs.py and avoid overlapping open PRs.
4. Change files in config/, instructions/, profile/, docs/, or scripts/.
5. Run python3 scripts/validate_workspace.py.
6. Open a Pull Request and use the checklist.

The maintainer reviews the PR, confirms validation, and merges accepted changes into the canonical project.

Use this lane for normal local validation. It does not require GitHub, DNS, live GPT Builder access, external package installation, or repository secrets.

python3 scripts/validate_workspace.py
python3 scripts/validate_eval_fixtures.py
python3 scripts/extract_instructions.py
python3 scripts/compile_prompt.py --check --compare-current
python3 scripts/validate_schemas.py
python3 scripts/generate_archive_fixtures.py
python3 scripts/validate_archive_fixtures.py
python3 scripts/build_release_zip.py
python3 scripts/validate_release_artifacts.py
python3 scripts/prepublish_audit.py --local
make check-local
make prepublish-audit-local

Use this lane only when GitHub API access is available. It is isolated from check-local.

python3 scripts/check_open_prs.py --soft --allow-dependency-workflow-overlap
python3 scripts/sync_labels.py --dry-run --soft
make check-online

python3 scripts/build_release_zip.py
python3 scripts/validate_release_artifacts.py
bash scripts/snapshot_release.sh

Local validators do not prove live GPT behavior. Run manual Preview with docs/GPT-BUILDER-PREVIEW-RUNBOOK.md, docs/PREVIEW-TEST-MATRIX.md, eval/preview-scenarios.jsonl, and eval/knowledge-retrieval-probes.jsonl. Mark this lane NOT VERIFIED unless it was actually run.

scripts/validate_workspace.py checks:

• config/ contains exactly the expected 20 Markdown files.
• instructions/SYSTEM-INSTRUCTIONS.txt matches the Instructions block in config/GPT-BUILDER-CONFIG.md.
• the profile README exists and names the project correctly.
• contributor/community surfaces are present and aligned.
• repo-local AGENTS.md and the open PR checker are present.
• public-safe agent operating patterns are present.
• boundary/state contract is present and aligned with repository zones.
• runtime sandbox capability evidence doctrine is present and aligned with sandbox limitation language.
• eval fixtures, Preview matrix, and threat model are present.
• the private-reference and private-to-public distillation policies are present and connected to eval coverage.
• release ZIP artifacts can be validated after packaging.
• prepublish audit can combine PR coordination, validation, release checks, private-marker scans, secret-like scans, and git-history scan.
• schema examples, mode catalog references, Knowledge retrieval probe structure, and archive fixtures can be validated locally.
• issue-template labels are defined in .github/labels.json.
• no obvious secret-like token was introduced.

1. Use the Copy-Ready Builder Field Map in config/GPT-BUILDER-CONFIG.md as the install sheet.
2. Copy the contents of instructions/SYSTEM-INSTRUCTIONS.txt into GPT Builder Instructions.
3. Upload exactly the 20 Markdown files from config/ as Knowledge.
4. Choose the current available GPT Builder model from the live UI and record it.
5. Enable file uploads and Code Interpreter/Data Analysis.
6. Keep Apps off for the public GPT Store pilot.
7. Keep Actions off unless a real backend, OpenAPI schema, authentication model, privacy policy, logging, rate limits, and tests exist.
8. Run the Preview tests using docs/GPT-BUILDER-PREVIEW-RUNBOOK.md, config/EVALUATION-CHECKLIST.md, docs/PREVIEW-TEST-MATRIX.md, and docs/PILOT-LAUNCH-CHECKLIST.md.

Contributions should come through Pull Requests against GPTsAgent/GPTsAgent. A contributor should be able to say: "I made a focused change, validation passes, please review." The maintainer can then review the diff, run the same checks, request changes, or merge.

GPTsAgent does not give ChatGPT direct access to a user's local computer, private repositories, CI, cloud account, or production host. It is an operating layer for uploaded files, sandbox artifacts, and scoped tools inside ChatGPT.com.

Files intentionally uploaded into the user's ChatGPT session are valid session inputs. Sandbox File Operator may inspect relevant uploaded project files, docs, configs, logs, metadata, and environment-like files for the user's private task. The safety boundary is disclosure and transfer: do not unnecessarily expose, copy, publish, commit, or reuse sensitive/private values outside the user's intended task.

Private archives, projects, transcripts, and handoff reports can be used only as temporary pattern input under docs/REFERENCE-DISTILLATION-POLICY.md and docs/PRIVATE-TO-PUBLIC-DISTILLATION.md. Public files must not preserve private names, paths, examples, logs, source text, or provenance.