On 4/23/26 17:21, Jochen Theodorou wrote:
On 4/23/26 03:30, Paul King wrote:
I committed one:
https://github.com/apache/groovy/blob/master/AGENTS.md <https://
github.com/apache/groovy/blob/master/AGENTS.md>
But please make suggestions if you spot anything that should be
included initially.
You have it kind of in there, but I would add to not reformat code and
keep the formatting style of the file for new and changed code.
There is quite a bit more, but I first have to ask for permission to use
that.
we add things like this for the workflow
Global Override
Always state assumptions explicitly before substantial execution.
Always ask clarifying questions in multi-answer format with
suggested/recommended answers before non-trivial changes.
Keep rationale concise, matter of fact: explain decisions briefly
(assumptions, tradeoffs, risks, verification), not hidden or verbose
reasoning.
Treat substantial execution as any task beyond a typo-only or
comment-only trivial edit, including any file edit, mutating command,
commit, plan creation, or multi-file investigation.
Hard Stop Gate (Before Edits)
Do not edit files, run mutating commands, create plans, or create
commits until:
assumptions are stated
multi-answer clarifications are asked
user answers/approves direction
Only typo-only or comment-only trivial edits may bypass this gate.
Core Workflow
Plan first for non-trivial tasks (objectives, steps, key commands,
estimate XS/S/M/L).
Ask clarifying questions and a final refinement question on goal,
scope, or success criteria.
Execute with traceability (list key files/commands/decisions).
Run quality gate for runtime-impacting changes: lint, tests,
self-review.
Report what was run and what was skipped (with reason).
Kickoff Checklist (Before Any Tool Call)
Assumptions stated explicitly
Verification questions asked (with suggested answers)
Discovery questions asked (multi-answer with recommendation)
Refinement question asked
User confirmation received
If any item is unchecked, execution must pause.
Security And Architecture Guardrails
Preserve security posture (input validation, authn/authz, secret
handling, safe logging).
Ask before introducing potential security/performance/design
regressions.
Co-Author Calculation (At Commit Time)
Maintain agent-usage.log during task execution, one line per agent
identity:
Name <email>
At commit time, append unique trailers derived from the log:
sort -u agent-usage.log | sed 's/^/Co-authored-by: /'
Do not add co-authors for agents not used in the current task.
Transparency Additions
Add an ## Agent Trace section to PR descriptions:
agents used
files touched per agent
checks run per agent
Keep agent-usage.log under .gitignore, but attach a summarized
trace in PR description.
Add commit footers for provenance:
Generated-by: <agent/tool>
Reviewed-by: <human or agent>
Add a CI check that rejects commits missing:
valid Conventional Commits header
required JIRA ID policy
required co-author/provenance trailers
Create docs/agent-transparency.md with examples and policy exceptions.
Then Commit and PR conventions, a Quality gate (like pass all tests, do
not disable tests)
Things like the guardrails in the other post are also useful. I am not
saying we *need* all that, just giving some food for thought
bye Jochen
