Workflows calling agent commands — are two extension mechanisms in one run indended or something to be optimized? #4557
markuswondrak
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
The built-in
speckitworkflow orchestrates the SDD cycle by dispatching the stock commands ascommandsteps (workflows/speckit/workflow.yml). Acommandstep does not execute a command itself: it builds the native invocation (/speckit.plan …) and passes that prompt to the integration CLI (src/specify_cli/workflows/steps/command/__init__.py:197-261,src/specify_cli/integrations/base.py:351-413). The agent then loads the full command template.That template contains its own extension protocol.
Pre-Execution ChecksandMandatory Post-Execution Hooksread.specify/extensions.ymland dispatch further commands (templates/commands/plan.md:27-58,:74-106). A singleworkflow runtherefore has two extension mechanisms active:slotsteps (docs/reference/workflows.md:118-121,:289-321)Within one run both are present, and the engine only makes step ordering deterministic — execution inside a
commandstep remains agent-dispatched, the same model whose observed invocation rate is documented in #4200.The layering itself is documented (#4461: extensions provide commands + hooks, workflows orchestrate), and the runtime-hook route was settled in #2104 and #3473. My question would be:
This is a question about direction, not a design request. If black-box dispatch of commands is the intended stable end state, that answers it as well.
Drafted with AI assistance; the examples and line references were reviewed by me.
All reactions