Skip to content

Latest commit

 

History

History
102 lines (58 loc) · 7.48 KB

File metadata and controls

102 lines (58 loc) · 7.48 KB

Known Gaps — v0.2.1 (updated 2026-04-20, final)

An honest list of what is not verified in this bundle. Read this before claiming it "works."

✅ Closed in this update

All 10 skills cleanly measured

The previously-contamination-blocked docs-as-definition-of-done baseline is now captured via scripts/capture-clean-baseline.py — direct LLM API call, no Claude Code, no agent harness, no ambient CLAUDE.md. Raw run in tests/fixtures/docs-as-definition-of-done/runs/20260420T165000Z-clean/.

Δloss_bundle = 3.30 absolute (mean per skill) across all 10. Target ≥ 2.0 met with margin. Per-skill table in tests/README.md.

Tier-3.9 discovery run captured

New E2E run at tests/e2e/scenario-01-refactor/runs/20260420T164505Z/. Skills were installed at ~/.claude/skills/ rather than injected into the prompt. The subagent discovered them at runtime via the trigger-phrase table in using-ldd, read the SKILL.md files from disk, and applied the discipline. 7/7 rubric items satisfied, loop closed at k=1 of 5.

Finding: the subagent's Skill tool did NOT auto-discover ~/.claude/skills/ (returned Unknown skill). Fallback via direct Read worked. This is closer to tier-4 than tier-3.5 but still not the pure-plugin-install path — see tests/e2e/scenario-01-refactor/results.md for the honest caveats.

N=3 distribution demonstrated on one skill

tests/fixtures/root-cause-by-layer/runs/20260420T165603Z-clean-N3/ — 3 independent RED captures via capture-clean-baseline.py at temperature 0.8. All 3 produced type-tolerance-shim variants — same failure mode, tight distribution (stddev ≈ 0.5 on violation count). Proof that the distribution-sampling infrastructure works; a proper N≥10 distribution is cheap to run ($0.01–0.05 per call × 10 runs × 10 skills ≈ $1–5 total).

Scenario-design bias partially addressed

Second scenario added for root-cause-by-layer (tests/fixtures/root-cause-by-layer/scenario-2/) — different domain (rate-limiter precondition contract vs. notifier boundary leak), same skill. Two scenarios from one author reduces single-point-of-failure risk without eliminating author bias.

docs/ldd/ structure — drift resistance

Methodology text now lives in exactly one place: docs/ldd/. Task-specific compressed MDs (debugging.md, design-decisions.md, refactor.md, refinement.md, release.md, incident.md, method-maintenance.md) with task-types.md as the dispatch table. Prevents methodology drift between README / skill bodies / user-project docs. User-project CLAUDE.md references task-types.md rather than copying methodology inline.

⚠️ Still open (honestly)

Tier-3.9 E2E suite captured for v0.3.1 — 6/6 scenarios green

Six subagent-dispatched runs with real tool access (file IO, Bash, git) on fictional /tmp projects captured in tests/e2e/v031-runs/20260420T195255Z/:

  1. Inner-loop failing test → 5-skill cascade, layer-4/5 fix, 1 commit
  2. Refinement of C+ doc → 7-defect gradient, empty-gradient stop at iter 2 3a. Architect conservative → 11/11 rubric, novelty ≤ 1, existing-stack scaffold 3b. Architect standard → 10/10 rubric, baseline behavior (unchanged from v0.3.0) 3c. Architect inventive → acknowledgment flow fired, 2-candidate relaxation, PRIOR_ART.md + EXPERIMENT.md + fallback, 7/7 inventive rubric
  3. Negative (inventive-in-config) → warning emitted, no ack flow, downgraded to standard, 10/10 standard rubric

All observed behaviors matched specified behaviors. See tests/e2e/v031-runs/20260420T195255Z/README.md for per-scenario artifact paths.

Real tier-4 plugin-install E2E still pending

The tier-3.9 run showed that the subagent's Skill tool in my build environment does not auto-discover ~/.claude/skills/. Whether this is a subagent-harness limitation or a general-adopter issue is not yet known. Real tier-4 requires /plugin install in a live Claude Code session, which can only be tested by an adopter.

How to close: follow the adopter guide in tests/e2e/scenario-01-refactor/results.md.

Distribution at N≥10 per skill not run

N=3 was demonstrated on one skill; all other skills are point estimates. A proper distribution claim needs N≥10 across multiple models.

How to close: community runs scripts/capture-clean-baseline.py N times per fixture. Infrastructure is in place. Estimated cost: $1–5 for the whole bundle at current pricing. This is an adopter task, not a developer task, because distribution needs independent re-runs.

Author scenario-design bias remains

Two scenarios for root-cause-by-layer; one scenario each for the other 9 skills. All authored by the skill author. Cannot be closed without community contributions.

How to close: PRs adding community-authored scenarios. Template: clone an existing fixtures/<skill>/scenario.md, adapt to your domain, run through capture-clean-baseline.py, publish.

GREEN side not re-captured via clean API

RED baselines are now clean (all 10 skills); GREEN responses still come from in-session subagent runs where ambient methodology may have contributed. The GREEN side defines the compliance upper bound, so ambient contamination there reduces measured Δloss (RED moves toward GREEN, not away) — the measured values are therefore lower bounds on the skill's true effect. Re-capturing GREEN via the clean API with the skill body prepended would firm up the upper bound but is cosmetic at this point.

Distribution gaps (unchanged across v0.2.x)

Claude Code auto-trigger behavior

The README "Using LDD" section documents the LDD: buzzword as the guaranteed-activation path. Automatic description-based triggering works on paper but has not been verified in a live Claude Code session. If an adopter installs and finds auto-trigger unreliable, the buzzword path is the fallback.

Codex personal-skills path uncertain

AGENTS.md suggests ~/.agents/skills/. Exact path varies by Codex version.

Gemini CLI extension schema may have drifted

gemini-extension.json mirrors superpowers' format. Not re-verified against current Gemini docs.

Aider · Cursor · Copilot CLI · Continue.dev

Documented via reference-or-inline. No captured test runs on any.

Summary

What v0.2.1 measures:

  • All 10 skills cleanly measured. Δloss_bundle = 3.30 absolute across all 10, target ≥ 2.0 met
  • Tier-3.9 E2E capture (skills discovered at runtime, not prompt-injected), 7/7 rubric
  • Inter-reviewer variance sampling (±2 per skill, direction 100%)
  • N=3 distribution demo on one skill (tight, stddev ≈ 0.5)
  • Two scenarios for one skill (partial bias reduction)
  • scripts/capture-clean-baseline.py portable across OpenRouter / OpenAI / Anthropic

What v0.2.1 does not prove:

  • Real tier-4 (live plugin install) — adopter task
  • N≥10 distributions per skill — adopter task (infrastructure in place)
  • Scenario-author neutrality — community task
  • Live auto-trigger on Claude Code / Codex / Gemini — adopter task