[infra P2] SKILL.md↔helper integration test layer for /idd-edit (gap exposed by #154 R1/R2)

[kiki830621]

Problem

Source: surfaced during /idd-verify --pr 159 Round 2 (M-R2-4 + DA Probe 7 + Codex Bug 3 — independently flagged by 3 reviewers)

/idd-edit skill has 19→23 test fixtures (R1→R3 series of #154) but all fixtures invoke the helper script directly, bypassing the SKILL.md orchestration layer. The integration layer between SKILL.md (prose bash for AI to follow) and idd-edit-helper.sh (extracted parser/validator) has zero test coverage.

Evidence

R2 verify findings that originated in SKILL.md↔helper integration layer (NOT helper standalone):

• R1 B1: $APPEND_BODY undefined — SKILL.md heredoc used wrong variable name vs helper's BODY_INPUT export. Fixtures didn't catch (helper alone is fine; integration was broken).
• R1 B2: $REPO vs $GITHUB_REPO — same pattern. Helper exports REPO; SKILL.md Step 6/7 referenced undefined GITHUB_REPO.
• R2 H7: SKILL.md Step 1 for-loop closed BEFORE Steps 1.5-7 → batch mode silently processed only LAST target. Pure SKILL.md prose structure bug; helper has nothing to do with it.

All 3 were "silently passing tests, broken in production" failures. The fixture gap is structural — current fixtures test helper in isolation, but the actual user-facing surface is helper-through-SKILL.md.

Open questions

1. What's the test layer? SKILL.md is prose for AI to follow, not executable bash. Options:

   • (a) Extract SKILL.md orchestration into sourceable idd-edit-orchestrate.sh that AI invokes; helper subcommands become its primitives. Fixture tests orchestrate end-to-end with mock gh.
   • (b) Generate runnable bash from SKILL.md prose (parse fenced bash blocks, dry-run them with mocked gh/git).
   • (c) Property-based testing: spec the expected variable bindings + invariants; static check SKILL.md prose against spec.
   • (d) Live integration: a "smoke test" that runs /idd-edit end-to-end via Claude Code automation against a test repo. Most accurate but heaviest infrastructure.

2. Which variables are part of the SKILL.md↔helper contract?

   • Helper exports: MODE / SCOPE_FLAG / SECTION_FLAG / REASON / BODY_INPUT / BODY_FILE / REPO / CWD / LAST / OVERRIDE_USER_CONTENT / TARGETS / RESOLVED_COMMENT_IDS
   • SKILL.md consumes: same names (must stay in sync)
   • Need a contract spec doc to prevent drift

3. Granularity:

   • Per-step integration (Step 1 → Step 1.5 → Step 2 → ...)?
   • Per-mode integration (--append / --replace / --prepend-note end-to-end)?
   • Per-flag integration?

Type

infrastructure / test framework

Priority

P2 — pattern proven 3x (R1 B1/B2 + R2 H7) and #154 still has latent integration bugs likely post-R3 since fixture gap unchanged. Until this lands, future SKILL.md edits remain vulnerable to "tests green but production broken" regression.

Trigger conditions

• Immediate: as part of [infra] IDD plugin test framework — adversarial fixtures for SKILL.md regression coverage (sister concern from #154) #156 (test framework generalization) decision OR as standalone P2 follow-up
• Or: next time SKILL.md↔helper integration breaks in production

Related

• [infra] IDD plugin test framework — adversarial fixtures for SKILL.md regression coverage (sister concern from #154) #156 — IDD plugin test framework (P2; reuse same infrastructure)
• [feature] /idd-edit runtime enforcement — proper proposal after 3-iteration bash failure (R1/R2/R3 on #153) #154 — root of R3 fix loop that exposed this gap
• [META] /idd-edit + IDD skills: bash vs alternative layer for strict flag parsing (sister concern from #154) #155 — bash vs alternative layer (orthogonal but related: if [META] /idd-edit + IDD skills: bash vs alternative layer for strict flag parsing (sister concern from #154) #155 picks Rust/Python layer, integration test pattern changes anyway)

Refs #154 #156