v5.0.0 — MCP server, auto-instrumentation, Debug-Bench

debug-bank v5 ships the orchestration layer: programmatic access to the pattern bank, auto-pattern extraction from novel bugs, and the first benchmark that measures debugging quality, not just pass/fail.

What's new

MCP server (6 tools)

Any MCP-compatible client (Claude Code, Cursor, Codex CLI) gets the pattern bank as a live API:

• debug_classify — symptom → ranked pattern matches with confidence
• debug_pattern — full pattern content by ID (P01–P24)
• debug_search — full-text search across patterns, compositions, postmortems
• debug_record — persist debugging trajectories for compound learning
• debug_score — 6-criterion quality rubric (100 points)
• debug_dap_commands — generate DAP commands from a pattern's debugger strategy

Install: cd mcp-server && pip install -e .

Auto-instrumentation fallback

When no pattern matches, the system generates hypothesis-driven logging with #region DEBUG-Hn blocks. Each block captures the evidence needed to prove/disprove a specific hypothesis. After the fix lands, a pattern candidate gets auto-extracted. Templates ship for Python, JavaScript/TypeScript, and Go.

Debug-Bench (3 Docker scenarios)

• S01 Stale cache race
• S02 Retry storm
• S03 Silent schema drift

Each comes with a 6-criterion trajectory scorer (pattern check, reproduction, hypothesis quality, isolation, root-cause depth, fix minimality). First benchmark measuring debugging quality rather than just fix correctness.

New patterns since v3

• P20 Filler/background audio pipeline contention
• P21 Untested handler path after shared code change
• P22 Iterative fix regression (failswitch) — the 3-exchange rule
• P23 S2S native-audio model voice-quality limits
• P24 API strict-validation upgrade

Pattern bank now at 24, organized into Code Structure / Data Integrity / Configuration / Dependencies / Platform Quirks / LLM-AI.

The 4-layer model

```
Layer 4: ORCHESTRATION ← v5 (MCP server, auto-instrumentation, benchmark)
Layer 3: KNOWLEDGE ← 24 patterns, classifier, protocol, compositions
Layer 2: RUNTIME ← debugger strategies (PDB/JDB breakpoints, watch expressions)
Layer 1: STATIC ← most agents today (grep, read, guess, retry)
```

Install / upgrade

Same one-liner:

```bash
curl -O https://raw.githubusercontent.com/soleimanmansouri/debug-bank/main/CLAUDE.md
```

For MCP integration with Claude Code, add to ~/.claude/mcp_servers.json:

```json
{ "debug-bank": { "command": "debug-bank-mcp", "args": [] } }
```

MIT licensed. All core pattern files remain dependency-free markdown.