[maisim]

Move my detailed file-scoped conventions (copilot instructions) into tracked Claude skills under .claude/skills/, following the pattern proposed in upstream PR #1683:

• writing-operations: operation patterns, _inner() composition, error handling
• writing-facts: FactBase anatomy, parsing, multi-file marker pattern
• writing-operation-tests: JSON/YAML fixture schema for tests/operations/
• writing-fact-tests: JSON/YAML fixture schema for tests/facts/

Also enrich AGENTS.md with a few items previously held only locally:

• Command objects (StringCommand, QuoteString, MaskString, ...)
• Connector data (connector_data_meta / host.connector_data)
• Coverage gevent concurrency note
• Index of the new skills under "Detailed conventions"

• Pull request is based on the default branch (3.x at this time)
• Pull request includes tests for any new/updated operations/facts
• Pull request includes documentation for any new/updated operations/facts
• Tests pass (see scripts/dev-test.sh)
• Type checking & code style passes (see scripts/dev-lint.sh)

[maisim]

@Fizzadar @wowi42 @DonDebonair
To share, Just move my untracked copilot instructions files to claude skills.

we should keep all, part, or nothing, let me know

I just did a complete reread and didn't find any major mistakes. 🤞

[Fizzadar]

Couple of thoughts!

[Fizzadar]

Exceptions look like (example):

    "exception": {
        "names": ["OperationValueError"],
        "message": "'content' must be supplied when 'present' == True"
    }

[Fizzadar]

This schema + facts are duplicated in AGENTS.md as well which is likely to drift - we could point agents at each skill file? (Will non-claude harnesses handle that correctly?)

[Fizzadar]

Suggested change

	Use real command output whenever possible — copy from an actual system.
	Always use real command output — copy from an actual system.

[wowi42]

The skills need another pass before they are safe to use. Several examples do not match the test harness or the project shell quoting rules.

Issues

• .claude/skills/writing-operation-tests/SKILL.md documents invalid fixture forms. A string exception fails because the parser expects a mapping with name/names and message; file uploads serialize as arrays, not {type: "put"} objects:

  "exception": {"names": ["OperationError"], "message": "..."}
  "commands": [["upload", "local/path", "/remote/path"]]

  The existing maintainer correction is unapplied.
• .claude/skills/writing-fact-tests/SKILL.md describes arg as one argument list per call variant. Each fixture calls command() only once: lists are expanded as positional arguments, dicts as keyword arguments, and scalars as one argument. The documented semantics will produce incorrect command expectations.
• .claude/skills/writing-facts/SKILL.md advertises default = None, but the executor unconditionally calls fact.default(). A None result requires the inherited or overridden default() method, not a None attribute.
• Both implementation skills show unsafe f-string interpolation of user-controlled values, contrary to the StringCommand + QuoteString rule:

  yield f'gpg --dearmor -o "{dest}" "{src}"'
  return f"sha256sum {path} 2>/dev/null | cut -d' ' -f1"

  These examples teach shell injection and contradict the safety text in the operation skill itself.

Notes

• AGENTS.md says gevent coverage concurrency is configured in pyproject.toml, but [tool.coverage.run] has no such setting, including at the PR head.
• All three maintainer review threads remain unresolved, including the single-source/cross-harness design question. Maintainer input is still needed on that approach. The branch also conflicts with 3.x.

Done at PyCon FR.