feat: add 'docker agent debug skills' command

[dgageot]

This adds a mirror of the existing docker agent debug toolsets command, providing introspection into the skills wired into each agent. Users can now run docker agent debug skills to see which skills are available to an agent, including their descriptions and whether they run in a forked context.

The implementation adds a skills subcommand to the debug command that looks up the *skillstool.ToolSet among each agent's toolsets and prints the skill name, a [forked] marker for context: fork skills, and the skill description. When an agent has no skills, it emits a clear empty message.

New tests cover both agents with no skills and agents with locally-discovered skills (both plain and forked variants), ensuring the feature works across different configurations.

[docker-agent]

Assessment: 🟢 APPROVE

The docker agent debug skills command is a clean, well-structured addition that mirrors the existing debug toolsets command. The implementation correctly:

• Walks agent toolsets and type-asserts to *skillstool.ToolSet using tools.As
• Handles the nil/empty case with a clear "No skills" message
• Annotates fork-context skills with the [forked] marker
• Properly tracks telemetry with TrackCommand/TrackCommandError following the established pattern
• Defers stopToolSets for cleanup, consistent with other debug subcommands

Tests cover both the empty-skills path (no_tools.yaml) and the local-discovery path (plain + forked skills), and the new skills_local.yaml fixture follows the same structure as existing fixtures.

No bugs found in the changed code.

[aheritier]

Test - single comment on debug.go

[aheritier]

Test comment on Skills() guard line

[aheritier]

Test pos 1

[aheritier]

test

[aheritier]

Test pos 2

[aheritier]

test

[aheritier]

Test pos 3

[aheritier]

test

[aheritier]

Test pos 4

[aheritier]

test

[aheritier]

Test pos 5

[aheritier]

test

[aheritier]

Test pos 6

[aheritier]

test

[aheritier]

Test pos 7

[aheritier]

test

[aheritier]

Test line-based comment

[aheritier]

Test comment on line 7

[aheritier]

Test line 180

[aheritier]

Test

[aheritier]

Test line 185

[aheritier]

Test

[aheritier]

The implementation is correct, idiomatic, and follows existing conventions closely. tools.As[T] correctly handles wrapped toolsets, telemetry hooks match every other debug subcommand exactly, and t.Setenv is correctly used without t.Parallel() in the local-skills test. Two minor suggestions worth addressing before merge:

1. Skills() called twice — see inline comment on runDebugSkillsCommand.

2. skills_local.yaml missing version: field — see inline comment on the new fixture.

No blocking issues. Everything else (alphabetical ordering, stopToolSets defer, variable rename tools → agentTools) is well done.

[aheritier]

Skills() is called twice: once in the nil || len(...) guard on the line above, and again here to bind loadedSkills. The method returns the struct's internal slice so both calls are safe, but it's redundant. Capturing the result once makes intent clearer and removes the double call:

var loadedSkills []skills.Skill
if skillsToolset != nil {
    loadedSkills = skillsToolset.Skills()
}
if len(loadedSkills) == 0 {
    out.Printf("No skills for %s\n", agent.Name())
    continue
}

[aheritier]

Every other fixture in e2e/testdata/ includes a version: field (e.g. version: "2"). The loader defaults gracefully via cmp.Or, so this works, but the omission is inconsistent and may surprise future contributors. Consider adding version: "9" (current latest) to match the rest.