| name | mysql-mcp-heal |
| description | Guide for modifying the `mysql-mcp` codebase (schemas, server instructions, proxy interceptors) to proactively prevent, mitigate, and heal LLM hallucinations for all future agents. |
Optimize Codebase and Heal Hallucinations
💎 Value Proposition
- Execute complex logic via Code Mode, reducing token usage by 70-90%.
- Build AI integrations instantly.
- Empower agents with secure database access.
- Scale operations with robust connection pooling.
- Leverage OAuth 2.1 for enterprise security.
The primary philosophy is: Automate away as many problems as possible. Do not work around hallucinations, confusion, or artificial roadblocks. Do not settle for throwing a good error message.
You must modify the underlying infrastructure. Ensure execution succeeds seamlessly. The agent should not even notice.
1. Phase 1: Determine the Healing Layer
There is a strict hierarchy of how you should heal the codebase:
- Prevent (Highest Preference - Frictionless Recovery): Automate the problem away. Modify API schemas, normalizers, or aliases. Silently catch, transform, or permit hallucinatory payloads. Execution should succeed automatically.
- Mitigate (Fallback Only): Use this only if inputs cannot be safely inferred or coerced. Intercept the input. Throw an actionable custom error. Explicitly guide the agent on immediate fixes.
Determine which of the 6 layers needs modification:
Layer 1: Harden Schemas and Instructions
Explicitly instruct the agent to prevent hallucinations. Use the tools list or initialization payload.
- Schema
.describe() Tags: When an agent hallucinates a property, update the Zod schema in src/adapters/mysql/schemas/*.ts. Add an explicit "Anti-Hallucination Hint" in the .describe() string. These surface directly in the JSON Schema that agents see on tools/list.
- Example:
.describe('Note: Pass table, not tableName.')
- Tool Descriptions: Update the
description fields on tool definition files (e.g., src/adapters/mysql/tools/*/index.ts). This is the agent's first impression of each tool.
- Server Instructions: If a structural hallucination persists, update
src/constants/server-instructions/*.md. Then run pnpm run generate:instructions (or npx tsx scripts/generate-server-instructions.ts). These alter the context payload for all agents upon connection.
Layer 2: Heal Parameter Normalization
If agents consistently hallucinate inputs, see if they can be logically inferred or normalized. If so, heal them silently via z.preprocess() wrappers or positional params. This implements the Dual-Schema Pattern.
- Parameter Aliases: If agents pass
tableName instead of table, use the preprocess-utils.ts alias resolver to automatically map the keys.
- Type Coercion: If agents pass strings (e.g.,
limit: "50") instead of numbers, use coerceNumber or similar in Zod to heal it before validation throws.
- Positional Params: If agents intuitively pass positional arguments (e.g.,
mysql.core.readQuery("SELECT 1") instead of mysql.core.readQuery({ sql: "SELECT 1" })), add or update bindings in src/codemode/api/constants/positional.ts.
Layer 3: Intercept Sandbox Proxies
When agents execute code in Code Mode (which is now multiplexed to all tools), they often make standard Javascript mistakes. Instead of cryptic V8 runtime failures, mysql-mcp uses Proxy interceptors in src/codemode/sandbox.ts to throw explicit guidance.
- Missing
await Catching: Agents frequently forget the await keyword. All RPC Promise returns are wrapped. If the agent attempts to synchronously access a property (e.g. result.success), the wrapPromise interceptor throws: "Attempted to access property 'success' on a Promise object. Did you forget to 'await' the tool call?"
- Array Method Healing: Agents often assume tools return Arrays instead of Objects (e.g.,
{ success: true, rows: [...] }). The wrapResult proxy intercepts Array methods (like .map, .filter) called directly on the returned Object and binds them dynamically to the inner array (e.g. .rows, .tables, .results).
- Failed Operation Masking: The
wrapResult proxy prevents agents from destructuring data or rows out of a failed response object, throwing an explicit error referencing the tool failure instead.
Layer 4: Manage Aliases Intuitively
Agents commonly extrapolate method names based on patterns (e.g., trying to call mysql.json.jsonExtract() instead of mysql.json.extract()). You can add aliases to gracefully handle these hallucinations.
- Adding Aliases: Add entries to the
METHOD_ALIASES constant in src/codemode/api/constants/aliases.ts (e.g. jsonExtract: 'extract').
- Safe Resolution: Because alias resolution happens at API construction time in
src/codemode/api/generator.ts rather than runtime, you do not face the same "shadowing bug" risks as other implementations. Aliases simply copy the reference of the correctly bound function.
Layer 5: Inject CLI Guardrails and Help
When an agent hallucinates CLI flags or usage patterns for mysql-mcp standalone scripts or the cli.ts commander CLI:
- Option Aliasing: Add hardcoded aliases for commonly truncated flags in commander (e.g.,
.option('--log, --audit-log')) so execution naturally succeeds.
- Dummy Command Interception: When creating dummy commands to intercept agent hallucinations, ALWAYS use
[args...] in the signature to prevent Commander from throwing a "too many arguments" error before the action can execute.
- Help Meta-Prompting: Use Commander's
addHelpText('after', ...) to inject explicit 🤖 AI AGENT INSTRUCTIONS directly into the --help output.
- Actionable Errors: Ensure that all CLI validation failures and process exits are prefixed with
🛠️ AUTONOMOUS HEALING: to prevent agent loop retries.
Layer 6: Collaborative Environment Guidelines
Agents often run into environment-level interceptors or tooling issues that throw hallucinatory errors.
- Fragile Native Tools: When attempting to heal global environment files, please avoid using the agent-native
replace_file_content tool on large, ambiguous code blocks. It uses loose heuristics that can negatively impact files.
- Safe Rewrites: Instead, heal environment files using precise
multi_replace_file_content targeting single lines. Alternatively, write a dedicated deterministic script to parse and rebuild the file safely.
- Native Tool Priority: If you trigger a system interceptor for using shell commands like
cat, grep, or ls, please switch to native tools (e.g., view_file, grep_search, list_dir). It is recommended to do this autonomously.
- Transient IDE Errors: If shifting lines via code edits triggers IDE typing errors, please verify with
tsc --noEmit before attempting complex type refactoring. It may just be a stale ESLint cache.
- Coordinator Guidelines: To mitigate unexpected structural behaviors, please use a
task.md checklist. Run a list_dir requirement before accessing unknown paths. Halt execution immediately on FAILED_FILE_NOT_FOUND rather than retrying.
- Audit Tool Payload Bloat: To prevent payload bloat, the audit tool has a strict security requirement: it MUST receive at least one filter. If an agent hallucinates a blank audit call, the tool should reject it or the schema should enforce it.
2. Phase 2: Implement the Fixes
- Update the necessary files (e.g. schemas, instructions, normalizers, aliases).
- Validate the integrity of your code changes according to the verification steps required by your current test prompt or workflow. If unspecified, run
pnpm run check to ensure full safety.
- Agent-First Validation (CRITICAL): Programmatic tests cannot verify if an AI agent is actually healed. You MUST verify the fix against an actual agent by running the usability coordinator workflows (e.g.,
test-server/test-usability/coordinator-workflow.md or equivalent test protocols in test-server/test-tools.md).
- Commit the changes locally.
3. Review Important References
- CLI Wrapper Library (
lib-agent-exec): Load the lib-agent-exec skill - For execution bridge boundaries and safety rules.
- MySQL Skill: Load the
mysql skill - For MySQL enterprise production rules, query safety, and schema configurations.
