Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
VIBE AI Teammate exhibits systemic gaps across schema completeness, parameter documentation, and error handling. While 17 tools are registered with descriptions, critical production patterns are largely absent. Schema validation is minimal, most tools declare parameters but lack type constraints, descriptions, or structured output schemas. The codebase shows tools are defined but schemas are inferred from TypeScript types rather than explicitly visible in MCP protocol format. Error handling is rudimentary (no recovery guidance or categorization). Parameter documentation is sparse, most params have one-line descriptions lacking format constraints or allowed value guidance. No evidence of idempotency guarantees, pagination support, or composition chaining. Security concerns: shell execution without sandboxing or approval gates; no visible secret injection or audit trails. This is a typical community-grade MCP server: functional but not production-ready.
Shell/execution tools (shell, run_command, vibe_execute, vibe_mission) lack approval gates, confirmation workflows, and dry-run support. Destructive operations expose the system to irreversible data loss without user consent or audit trails.
Input schemas are inferred from TypeScript types, not explicitly visible in MCP protocol format. No structured JSON Schema with type definitions, constraints (enums, patterns, min/max), or per-parameter descriptions for most tools. Schema completeness cannot be verified.
No documented output schemas. Tools return results but LLM cannot know what fields to expect, breaking downstream composition and context loss between turns. Returns appear unstructured or inferred.
Recommendations
Consolidate execution tools: merge shell, run_command, vibe_execute into a single execute_command tool with enums (mode: 'read'|'write'|'execute'), required approval gates, and structured output (exit_code, stdout, stderr, execution_time).
Consolidate search tools: merge search, search_files, grep, vibe_search into a single search_codebase tool that accepts query type enum (file_pattern|text_content|semantic) and returns paginated results (limit, offset, total_count) with file paths and line numbers.
Add explicit JSON Schema to every tool input. Document every parameter with type, description (50 - 150 chars), constraints (enum, pattern, min/max), and whether it is required. Use JSON Schema $schema: 'http://json-schema.org/draft-2020-12' format.
Document output schemas for all tools. Specify return type, fields, field types, and example structure. E.g., 'Returns { results: Array<{file_path: string, line_number: int, matched_text: string}>, total_count: int, has_more: boolean }'.
Add confirmation workflows to destructive operations (shell, edit, write_file, vibe_mission). Implement a pre-execution dry-run or explicit approval step. Return a confirmation_token that must be passed on final execute call.
Implement pagination and result limits: add limit and offset parameters to search, search_files, grep, vibe_search. Cap results at 50 items per page. Return total_count and has_more flag to guide agent paging.
Add error recovery guidance to all tools. Define error responses with error_code, user_message (actionable next step), and suggested_tool (if applicable). E.g., 'File not found in /workspace. Try search_files(pattern="*.txt") to list available files.'
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
Score history
Overall score trend
↑ 4 points across a rubric change (v1 → v2)
45/100
Scored
Grade
Overall
Spec posture
Rubric
2026-09-22
F
45
<=2025-11-25
v2
2026-03-09
F
41
-
v1
Run a shell command
searchread onlysource verified52/100
Search for text patterns across the codebase
search_filesread onlysource verified55/100
Search for files matching a pattern
shelldestructivesource verified52/100
Execute a shell command in the current directory
vibe_executedestructivesource verified40/100
Execute a specific command or tool within the Vibe environment.
vibe_get_mapread onlysource verified43/100
Get a comprehensive structural map of the repository (Aider-style).
Parameter descriptions are one-liners (avg ~15 - 20 chars) lacking format constraints, allowed values, ranges, or examples. 'Command to run' tells LLM nothing about escaping, environment, timeout, or safe patterns. Does not meet baseline of 72 chars avg.
No error recovery guidance. Tools have no documented error messages, categorization (retryable vs fatal), or next-step hints. LLM cannot infer whether to retry, ask user, or abort.
No visible pagination, limits, or result capping. Tools like 'search', 'search_files', 'grep' can return unbounded result sets, exhausting context window and degrading reasoning.
No visible composition chains. Tools do not document which fields (IDs, references) are returned, making it unclear if tool A's output feeds into tool B. Forces LLM to infer or make wasteful discovery calls.
No idempotency or side-effect declarations. 'edit', 'write_file', 'vibe_mission' modify state but do not promise safe retries. LLM cannot know if retrying will duplicate changes (e.g., double-commit, duplicate file writes).
File path parameters expose potential path traversal or symlink attacks. No documented constraints (e.g., 'must be within /workspace', reject '..' and symlinks). 'path' is ambiguous, could be absolute or relative; no guidance on safe ranges.
No visible secret injection, authentication, or audit logging. Tools like 'shell' and 'write_file' have no access control, no logging of who/when/what, and no credential masking. Production compliance risk.
Declare tool permissions and audit requirements. Each tool must state what access it needs (read:filesystem, execute:shell, etc.) and return execution metadata (who, when, what, status) for audit trails.
Add idempotency declarations: mark tools as 'idempotent' (safe to retry) or 'stateful' (side effects; do not retry without verification). Provide idempotency keys for stateful operations.
Document file path constraints. Specify that paths must be relative to /workspace, reject '..' and symlinks, and validate path sanitization. Add a path validation example: 'relative_path: string (regex: ^[a-zA-Z0-9._/-]+$, no .. or symlinks)'.
Add timeout and resource limits to all shell/command tools. Document timeout default (e.g., 30s), max memory, max output size. Return partial results if timeout occurs rather than full failure.
Separate tool concerns: 'vibe_mission' combines Plan + Execute + Verify + Commit. Split into vibe_plan (returns plan JSON), vibe_execute (runs plan step), vibe_verify (runs tests), vibe_commit (git commit). Allows agent to compose steps.
Add tool annotations (readOnlyHint, destructiveHint, idempotentHint) to MCP server response for tools that support them. Mark shell, run_command, vibe_execute, edit, write_file with destructiveHint=true.
Replace generic 'message' parameters with structured requests. E.g., approval tool should ask for target_action, risk_level (enum), and estimated_impact, not free-form 'message'.
Add batch/bulk variants for tools called in loops: add_multiple_edits, search_files_batch.
Implement secret injection pattern: no credentials in tool parameters. Use environment variables (MCP_GITHUB_TOKEN) or vault. Document this in tool descriptions and server init.