@su-record/vibe 3.2.6 → 3.2.7

package/skills/vibe.capability-loop/references/escalate.md

@@ -0,0 +1,86 @@
+# Step 4.5 — ESCALATE (VERIFY 실패 시)
+
+> vibe.capability-loop SKILL.md Step 4 의 VERIFY 가 **실패했을 때만** 로드한다.
+> VERIFY 가 통과한 호출(정상 경로)은 이 파일을 읽지 않는다.
+
+### Step 4.5: ESCALATE — When VERIFY Fails
+
+> **Problem**: The built capability didn't actually prevent the failure. This usually means the initial diagnosis was wrong (picked `Tool` when it needed `Guardrail`), or the failure has multiple missing capabilities.
+>
+> **Do NOT silently proceed** — a sub-standard capability log pollutes `.vibe/capabilities-log.md` and the failure will recur.
+
+**Escalation loop:**
+
+```python
+tried = [current_diagnosis.category]  # e.g., ["Tool"]
+
+while True:
+    # Re-diagnose excluding already-tried categories
+    next_diagnosis = diagnose(failure, exclude=tried)
+
+    if next_diagnosis is None:
+        # All 5 categories (Tool/Guardrail/Abstraction/Documentation/Feedback) exhausted
+        escalate_to_user(failure, tried)
+        break
+
+    if next_diagnosis.category in tried:
+        # Stuck: diagnose keeps returning the same category
+        escalate_to_user(failure, tried)
+        break
+
+    tried.append(next_diagnosis.category)
+    capability = build(next_diagnosis)
+
+    if verify(capability, failure):
+        persist(capability)
+        return  # Success — go to Step 5
+
+    # Still failing — next iteration
+```
+
+**User escalation prompt (interactive mode):**
+
+```
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+⚠️ CAPABILITY LOOP STUCK
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+Failure: {original failure description}
+
+Tried capabilities (all failed to prevent the failure):
+  ❌ Tool: {what was built, why it didn't work}
+  ❌ Guardrail: {what was built, why it didn't work}
+  ❌ Documentation: {what was built, why it didn't work}
+
+Automated diagnosis has run out of angles. This failure may require
+human judgment (process issue, cross-category solution, or external factor).
+
+How would you like to proceed?
+  1. Suggest a different angle (e.g., "this is a process issue", "needs Tool+Guardrail combination")
+     → Attempt custom approach per user instruction, then enter next verify
+  2. "manual" — resolve this failure via manual intervention, end capability loop
+     (record "escalated to manual" in capabilities-log.md)
+  3. "abort" — give up, record failure only
+     (record "diagnosis exhausted" in capabilities-log.md, do not halt the rest of the workflow)
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+```
+
+**ultrawork mode exception:**
+
+```python
+if ultrawork_mode:
+    # Skip user prompt: try all 5 categories in sequence, record final state
+    all_tried_exhausted = exhaust_all_categories(failure, tried)
+    record_failure_to_log(
+        status="diagnosis_exhausted",
+        tried=all_tried_exhausted,
+        failure=failure
+    )
+    return  # Proceed without blocking downstream workflow
+```
+
+**Rollback of failed builds:**
+
+- Each failed capability build should be rolled back before trying the next category (unless it's non-destructive documentation).
+- For code additions (tool/guardrail/abstraction): `git checkout -- {files}` or delete created files.
+- For docs-only additions: leave in place (low risk) but note in escalation prompt.

package/skills/vibe.chub-usage/SKILL.md

@@ -1,140 +1,11 @@
 ---
 name: vibe.chub-usage
+description: Compatibility alias for chub-based current API documentation lookup. Use when callers invoke vibe.chub-usage directly.
 invocation: [auto]
 tier: optional
-description: "Context Hub (chub) — fetch vetted, up-to-date API documentation. Write accurate code based on the latest docs instead of training data when working with external APIs/SDKs."
 triggers: [chub, context hub, API docs, latest API, deprecated API, SDK documentation, api reference, 최신 문서]
 priority: 60
+user-invocable: true
 ---
 
-# Context Hub (chub) Usage
-
-A skill for fetching vetted, up-to-date documentation before writing external API/SDK code.
-Solves the knowledge cutoff problem inherent in training data.
-
-## Why?
-
-| Problem | Solution |
-|---------|----------|
-| Relying on training data → using deprecated APIs | chub get → code based on vetted latest docs |
-| Web search → noisy results | chub search → curated docs only |
-| Repeating the same mistakes every session | chub annotate → accumulated learnings |
-
-## When to Use
-
-| Situation | Example |
-|-----------|---------|
-| Writing external API code | "Integrate Stripe payments" |
-| Checking latest SDK version | "Call the latest OpenAI model" |
-| Need official documentation | "Set up Supabase auth" |
-| Preventing deprecated patterns | "Firebase v10 migration" |
-
-## Workflow
-
-```
-Request to write external API/SDK code
-    ↓
-Step 0: Check if chub is installed (auto-install if not)
-    ↓
-Step 1: chub search "<library name>"
-    ↓
-Step 2: chub get <id> --lang ts
-    ↓
-Step 3: Write code based on the docs
-    ↓
-Step 4: chub annotate when a gotcha is discovered
-```
-
-## Step 0 — Auto-install
-
-**Always perform this step before running the skill.**
-
-```bash
-# 1. Check if chub exists
-which chub || command -v chub
-```
-
-If chub is not found, attempt automatic installation:
-
-```bash
-npm install -g @aisuite/chub
-```
-
-If installation fails, fall back to running via `npx @aisuite/chub`:
-
-```bash
-# search example
-npx @aisuite/chub search "stripe"
-# get example
-npx @aisuite/chub get stripe/api --lang ts
-```
-
-If `npx` also fails, fall back to context7 or Web Search (see Fallback Chain).
-
-## Usage
-
-### Step 1 — Search for docs
-
-```bash
-chub search "stripe"
-chub search "openai"
-chub search ""           # View full list
-```
-
-### Step 2 — Fetch latest docs
-
-```bash
-chub get stripe/api --lang ts
-chub get openai/chat --lang py
-chub get supabase/auth --lang js
-```
-
-### Step 3 — Write code based on the docs
-
-Write accurate code based on the fetched documentation.
-**Never rely on training data. Docs first, code second.**
-
-### Step 4 — Record learnings
-
-Gotchas, workarounds, and version issues discovered during work:
-
-```bash
-chub annotate stripe/api "pg parameter is required for Korean payments"
-chub annotate openai/chat "tool_calls in streaming comes as delta"
-chub annotate firebase/auth "getAuth() import path changed in v10"
-```
-
-Annotations are stored locally and automatically included the next time you run `chub get`.
-
-## Implementation Pattern (Subagent)
-
-Run via subagent to prevent context bloat:
-
-```
-Task tool call:
-- subagent_type: Explore
-- model: haiku
-- prompt: "Run `chub search <library>` then `chub get <id> --lang <lang>` to fetch latest API documentation for [topic]. Return only the relevant API usage examples, key changes from previous versions, and any annotations."
-```
-
-The subagent handles the chub calls and returns only a summary — keeping the main context clean.
-
-## Supported APIs (1,000+)
-
-OpenAI, Anthropic, Stripe, Firebase, Supabase, Vercel, AWS S3, Cloudflare Workers, Auth0, Clerk, and more.
-
-```bash
-chub search    # Run without arguments to view the full list
-```
-
-## Fallback Chain
-
-```
-which chub fails
-    ↓
-npm install -g @aisuite/chub (auto attempt)
-    ↓
-On failure: npx @aisuite/chub <command> (temporary execution)
-    ↓
-If npx also fails: context7 or Web Search fallback
-```
+Load `vibe.documentation-provider` and force the chub provider while preserving all arguments.