KatalKavya96 opened a new pull request, #64029:
URL: https://github.com/apache/airflow/pull/64029
### Summary
This PR introduces a small PoC for modeling Airflow contributor workflows as
structured, executable agent skills, with context-aware command selection.
The goal is to explore a narrow vertical slice of the problem described in
#62500 , focusing on workflow semantics rather than broad coverage.
---
### What is implemented
#### 1. Structured workflow blocks (Executable-doc style)
Two workflows are defined in a markdown file using structured blocks:
- `run_targeted_tests`
- local-first (`uv run --project ... pytest`)
- Breeze fallback (`breeze run pytest`) when system dependencies are
missing
- `setup_breeze_environment`
- host-only workflow
- returns guidance when invoked inside Breeze instead of suggesting
incorrect commands
---
#### 2. Extraction pipeline
A lightweight extractor converts structured blocks into machine-readable
skill definitions:
```
Markdown → Extractor → skills.json
```
---
#### 3. Context-aware decision engine
A simple decision engine consumes `skills.json` and returns the appropriate
response based on execution context:
- detects `host` vs `breeze`
- applies local-first vs fallback logic
- prevents invalid operations across contexts
---
#### 4. Command vs Guidance response model
Instead of returning only commands, the PoC introduces a response contract:
```json
{ "mode": "command", "command": "...", "reason": "..." }
{ "mode": "guidance", "message": "...", "next_action": "..." }
```
This allows modeling workflows where the correct action is not execution,
but a context-aware transition.
Example:
- running `setup_breeze_environment` inside Breeze returns guidance instead
of suggesting `breeze shell`
---
#### 5. Tests
Unit tests cover:
- skill extraction
- context detection
- decision logic
- command vs guidance behavior
---
### Why this approach
This PoC explores whether:
> agent-facing interfaces should return structured workflow-aware responses
(command or guidance), rather than only raw commands.
This may better reflect real contributor workflows where:
- some actions are executable
- others require environment/context adjustments first
---
### Scope
This is intentionally limited to:
- 2 workflows
- basic context detection
- minimal extraction pipeline
It does not yet include:
- full contributor-doc integration
- drift detection / prek hooks
- multiple workflow categories
---
Related: #62500
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
