gxy-tool-bot

Because once upon a time Björn said to Danielle: Wouldn't an Agentic Galaxy Tool Bot be cool? :smile:

So this is an agentic bot that generates Galaxy tool wrappers from user requests, powered by LLM APIs and orchestrated through GitHub Actions.

Overview

gxy-tool-bot is a Python library + companion GitHub Actions workflows that automate the creation of Galaxy tool wrappers. It is not a tool repo itself — it's a library consumed by repos that house Galaxy tools.

Current Status

This repository is under active development. The bot can plan tool wrappers, generate XML/macros/test data, and open PRs automatically. It has been tested end-to-end on real tool requests. Expect rough edges and iterate on configuration as needed.

How it works

1. Request: Users file a GitHub issue using the "Tool Request" issue template (structured fields for tool name, description, links, contact).
2. Plan: The tool-request label triggers a workflow. An agent researches the tool (bioconda, GitHub, publications, web) and posts a plan as an issue comment with a plan-ready label.
3. Review: A maintainer reviews the plan and adds the ready-to-implement label.
4. Generate: An agent generates the tool XML, macros, and test data, then opens a PR with a pr-opened label on the issue. If the agent gives up or crashes, a generation-failed label is applied instead. If generation completes but validation finds issues, the PR is still created with the validation errors noted in the description — apply the address-feedback label to have the bot attempt fixes.

Setup (for consuming repos)

1. Install the bot

The bot can be installed from PyPI or directly from GitHub:

pip install gxy-tool-bot              # from PyPI
pip install git+https://github.com/d-callan/gxy-tool-bot.git  # latest from GitHub

2. Create a config file

Create .gxy-tool-bot.yml in the repo root:

api:
  base_url: https://openrouter.ai/api/v1   # or https://api.openai.com/v1
  model: z-ai/glm-5.2                       # or gpt-4o, etc.
  fallback_models:                          # optional; tried in order if primary fails
    - deepseek-ai/deepseek-r1
    - gpt-4o
  max_tool_iterations: 25
  temperature_plan: 0.4
  temperature_generate: 0.2
  max_context_chars: 100000
  max_validation_retries: 3
  read_timeout: 600                         # seconds; increase for slow models

exemplars:
  - url: https://raw.githubusercontent.com/galaxyproject/tools-iuc/main/tools/bcftools/bcftools_view.xml
    macros: https://raw.githubusercontent.com/galaxyproject/tools-iuc/main/tools/bcftools/macros.xml
    shed_yml: https://raw.githubusercontent.com/galaxyproject/tools-iuc/main/tools/bcftools/.shed.yml
  - url: https://raw.githubusercontent.com/galaxyproject/tools-iuc/main/tools/seqtk/seqtk_seq.xml
    macros: https://raw.githubusercontent.com/galaxyproject/tools-iuc/main/tools/seqtk/macros.xml
    shed_yml: https://raw.githubusercontent.com/galaxyproject/tools-iuc/main/tools/seqtk/.shed.yml

# repo is optional — falls back to GITHUB_REPOSITORY env var (set automatically in GitHub Actions)
repo: your-org/your-repo

allowed_maintainers:
  - your-github-handle

3. Create GitHub labels

Create these labels in the repo (Settings → Labels):

Label	Color	Purpose
tool-request	#0075ca	Applied automatically by issue template; triggers planning
plan-ready	#a2eeef	Applied by bot after plan is posted
ready-to-implement	#0e8a16	Applied by maintainer to approve plan; triggers generation
pr-opened	#1d76db	Applied by bot after PR is created
generation-failed	#b60205	Applied by bot if generation or planning fails
retry-plan	#fbca04	Applied by user/maintainer to re-trigger planning after a failure
retry-generate	#fbca04	Applied by user/maintainer to re-trigger generation after a failure
address-feedback	#5319e7	Applied to a PR to have the bot address review comments and CI failures

4. Add the issue template

Copy examples/issue-template-tool-request.yml from this repo into your repo's .github/ISSUE_TEMPLATE/tool-request.yml. The template auto-applies the tool-request label so the planning workflow triggers automatically.

5. Add workflow files

Copy the workflow templates from the workflows/ directory in this repo into your repo's .github/workflows/:

• on-tool-request.yml → .github/workflows/gxy-on-tool-request.yml — triggers on new issues with tool-request label or when retry-plan label is added; runs the planner
• on-ready-to-implement.yml → .github/workflows/gxy-on-ready-to-implement.yml — triggers when ready-to-implement or retry-generate label is added; runs the generator and opens a PR
• on-pr-feedback.yml → .github/workflows/gxy-on-pr-feedback.yml — triggers when address-feedback label is added to a PR; reads review comments and CI failures, pushes fixes as new commits

CI artifact assumption: The feedback workflow fetches CI failure details from GitHub Actions artifacts. This assumes the CI workflow uploads failure reports as artifacts (e.g. lint reports as .txt files, planemo test results as .json), following the same conventions as the tools-iuc repo's pr.yaml workflow. If your repo uses a different CI setup that doesn't upload artifacts on failure, the bot will not be able to include CI failure details in its feedback context.

Validation flag file: When generation or feedback addressing completes but validation finds issues, the CLI writes a .validation-failed flag file to GITHUB_WORKSPACE (a built-in GitHub Actions env var). The workflows check for this file to post a comment on the PR noting that validation issues exist. The file is cleaned up after use and is never committed.

Both workflows install the bot from PyPI (pinned to a specific version):

- run: pip install gxy-tool-bot==0.1.0

6. Add repo secrets

Go to Settings → Secrets and variables → Actions → New repository secret:

Secret	Value	Used by
GXY_TOOL_BOT_API_KEY	Your LLM API key (e.g. OpenRouter key)	Plan + generate workflows

The GITHUB_TOKEN (automatically provided by GitHub Actions) is used for issue comments, labels, and PR creation — no extra secret needed.

7. Enable Issues

If the repo is a fork, Issues may be disabled by default. Enable them under Settings → General → Features → check "Issues".

8. Enable Actions

Make sure Actions are enabled: Settings → Actions → General → "Allow all actions and reusable workflows".

9. Test it

1. Go to Issues → New Issue → "Tool Request" template
2. Fill in the fields and submit
3. Check the Actions tab — the planning workflow should run
4. After the plan is posted, add the ready-to-implement label
5. The generation workflow should run and open a PR

Philosophy & Design Decisions

Run in GitHub Actions, not a black box

The bot runs entirely in GitHub Actions. This keeps agent traces, plans, generated files, and CI results publicly inspectable. Every run produces logs and artifacts you can review after the fact — to diagnose why a specific tool was written a certain way, or to identify patterns to improve the bot. The Actions logs capture tool calls, context size per iteration, and validation results; combined with file-based artifacts (generated files, test data) and summary comments posted to issues and PRs, this provides enough transparency to understand what the bot did and why.

Validate files, don't over-prompt

Best practices don't go in the system prompt — they go in a validation loop that inspects generated files for common mistakes (missing test data, bare output labels, Cheetah in macros, etc.) and instructs the agent to fix them in a retry loop. This keeps prompts concise (every line costs tokens on every LLM call) and prevents unnecessary bloat that might distract or confuse the agent, while still catching errors structurally. See DEVELOPMENT.md for the full tiered guidance on where to put conventions.

Incremental feedback, not infinite loops

The validation loop only runs a limited number of times before forcing the result back to human hands — a commit is pushed to a PR that a maintainer can review and provide further direction on. The maintainer adds the address-feedback label to trigger another loop with CI failures and review comments incorporated. This keeps the bot autonomous for the common case but avoids burning tokens indefinitely on a problem it can't solve alone.

Let planemo CI catch the rest

Things planemo already checks (XML well-formedness, shed metadata, duplicated output labels) are deliberately not duplicated in validation. The CI workflow reports these failures and the feedback flow picks them up on the next iteration. Only add a validation check if the bot is consistently making a specific mistake that wastes tokens and maintainer time.

Give the agent a way out

The agent has a give_up tool that lets it stop and explain why it can't proceed — open assumptions, unresolved questions, or fundamental issues with the request. A tool request might produce a recommendation not to make a tool rather than a plan to make one. A feedback request might result in push back with no commit until something the agent flags is resolved. This prevents forcing the agent into action when it doesn't have enough information, which would produce low-quality output. Better to surface the problem to a human than to generate a confident but wrong tool wrapper.

Development

See DEVELOPMENT.md for guidance on the codebase structure, where to add new conventions (prompts vs. validation vs. let CI catch it), and how the generation and feedback flows are organized.

License

MIT