feat: add swamp-getting-started skill with interactive onboarding

[stack72]

Summary

• Adds a new swamp-getting-started skill that auto-triggers when Claude starts in a swamp repo with no models
• Walks new users through a 5-step state machine: Goals → Create → Run → Inspect → Graduate
• Asks users what they want to automate upfront, then resolves the right model type (local search → extension search → custom extension)
• Teaches plan mode, review workflows, and contextual next steps in the graduation step
• Updates the generated CLAUDE.md to auto-invoke the skill in empty repos
• Adds the skill to BUNDLED_SKILLS so it ships with swamp repo init

Test Plan

• deno check passes
• deno lint passes
• deno fmt --check passes
• deno run test passes (80/80 repo service tests)
• npx tessl skill review passes (94% score)
• Binary compiles successfully with skill embedded
• Manual testing in fresh swamp repos

🤖 Generated with Claude Code

[github-actions]

Code Review

Blocking Issues

None. The PR is clean and well-structured.

Suggestions

1. Track naming inconsistency between SKILL.md and tracks.md: SKILL.md references "Track A", "Track B", and "Track C" throughout (State 1 On Failure, State 2 Action, State 3 Action), but references/tracks.md uses "Step 1: Search Local Types", "Step 2: Search Extensions", "Step 3: Build a Custom Extension" — with no mention of tracks by letter. The SKILL.md says "Follow the track-specific steps in references/tracks.md" but the file doesn't organize by track. Consider aligning the terminology so the AI agent following the state machine has a clear mapping between the track letters and the resolution steps.

2. Missing test for swamp-getting-started in skill_assets_test.ts: Other bundled skills have explicit inclusion tests (e.g., SkillAssets includes swamp-extension-model skill, SkillAssets includes swamp-troubleshooting skill). Adding a similar test for swamp-getting-started would follow the existing pattern and catch regressions if the skill is accidentally removed from BUNDLED_SKILLS.

3. Minor inconsistency in model detection approach: The generated CLAUDE.md template tells Claude to "check if the models/ directory contains any model YAML files", while the SKILL.md's "Before Starting" section says to run swamp model search --json. These are two different detection mechanisms. Not a problem in practice (the SKILL.md approach is more robust), but worth noting if the intent is consistency.

Overall this is a solid onboarding skill with a well-designed state machine, good pre-check logic, early exit for experienced users, and proper delegation to other skills. The TypeScript changes are minimal and correct — just registering the new skill in the bundled skills list and updating the generated CLAUDE.md template.

[github-actions]

Adversarial Review

Critical / High

None found.

Medium

1. SKILL.md: Track A/B/C terminology is referenced but never defined. State 1's On Failure says "default to Track A (shell commands)". State 3's method selection lists Track A (execute), Track B (sync/get), and Track C (extension-dependent). State 5 references "their track". However, references/tracks.md uses "Step 1", "Step 2", "Step 3" — the Track A/B/C mapping is never established anywhere. An AI consuming this skill must infer from context that Track A ≈ local shell type, Track B ≈ local non-shell type, Track C ≈ extension type. This won't break the walkthrough (the AI will figure it out), but it reduces clarity and could cause inconsistent behavior across sessions.

   Suggested fix: Either (a) define Track A/B/C explicitly in SKILL.md or tracks.md with a mapping table, or (b) replace Track A/B/C references in SKILL.md with the step terminology from tracks.md (e.g., "local shell type", "local typed model", "extension model").

Low

1. Dual detection mechanism between CLAUDE.md and SKILL.md pre-check. The generated CLAUDE.md instructs the AI to check if the models/ directory contains model YAML files, while the SKILL.md pre-check uses swamp model search --json. These are defensible as a two-layer check (CLAUDE.md triggers the skill, skill verifies authoritatively), but the inconsistency means an AI might enter the skill when models exist in the registry but not in models/, or vice versa.

2. SKILL.md pre-check doesn't handle command failure. If swamp model search --json fails (repo not initialized, corrupt config), no fallback behavior is specified. The AI would likely handle this gracefully, but explicit guidance would improve reliability.

Verdict

PASS — The two TypeScript changes are structurally correct and follow existing patterns exactly. All CLI commands referenced in the skill exist. The Track A/B/C terminology gap is worth fixing for clarity but does not block the merge since the context provides enough information for the AI to resolve the mapping.