The server provides 16 well-named debugging tools with complete input schemas and consistent descriptions. However, there are critical gaps: NO output schemas are documented for any tool, error handling lacks recovery guidance, parameters are minimally constrained, and there is no validation or input sanitization. Naming is verb-leading and clear (gdb_start, gdb_load, etc.), descriptions are present but generic (average ~30 chars, well below the 194-char production baseline), and schemas are properly typed but lack constraints. The server lacks idempotency documentation, permission gates, and audit logging for destructive operations like gdb_terminate. These gaps place it in the 'fair' range with room for improvement.
Attach to a running process
Show call stack
Execute a GDB command
Continue program execution
Examine memory
Execute until the current function returns
Display registers
List all active GDB sessions
No output schemas documented. LLMs cannot infer what fields tools return, forcing agents to guess the structure of responses and plan downstream tool chaining. This violates the response-shaper pattern and blocks proper error handling and result validation.
Descriptions are extremely short (average 20-30 chars, well below the 194-char production baseline). They lack WHEN to use the tool, prerequisites, or consequence visibility. For example, 'Execute a GDB command' tells LLMs nothing about what commands are valid or what happens to program state. Descriptions should be 50-200 chars and include context for tool selection.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 55 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Load a program into GDB
Load a core dump file
Step over function calls
Print value of expression
Set a breakpoint
Start a new GDB session
Step program execution
Terminate a GDB session
Destructive tools (gdb_terminate) have no confirmation mechanism, permission gates, or recovery guidance. The error handling in index.ts provides only basic error logging to console without actionable recovery hints for the agent. An LLM can terminate a session without understanding the consequence or being able to undo it.
Parameters lack constraints and validation rules. 'gdbPath' accepts any string with no format validation. 'command' in gdb_command is completely unconstrained, LLMs could pass arbitrary or malicious GDB commands. 'format' in gdb_examine has no enum (should be: 'x', 'd', 'i', 't', 'f', 's', 'a', 'c', 'o', etc.) and no description of valid values.
No input validation or error reporting. The code spawns GDB and executes commands without sanitizing or validating inputs. Path traversal attacks via 'workingDir' or 'program' are possible. Command injection via 'command' parameter is trivial. No error messages guide the LLM on what went wrong or how to recover.
No idempotency guarantees. Tools like gdb_set_breakpoint, gdb_continue, and gdb_step modify program state. If an LLM retries due to a transient error, duplicate breakpoints may be created or the program may be stepped twice. No idempotency markers or deduplication logic exist.
No audit logging or permission gates. The server logs only to console and provides no audit trail of who called what tool or when. Destructive operations (gdb_terminate, gdb_load) have no permission checks. No scope declarations (e.g., 'requires:debug', 'requires:execute') are present.
Tools operating on the same resource (gdb_step, gdb_next, gdb_continue all manipulate execution state) lack clear distinction in their descriptions. The 'instructions' boolean parameter is mentioned only once for gdb_step and gdb_next but has no explanation of what 'by instructions' means. Parameter relationships are undocumented.
Result limits not enforced or documented. gdb_backtrace and gdb_list_sessions could return unbounded results (thousands of frames, thousands of sessions). No 'limit' parameter with sensible defaults (e.g., 20-50) is present, and no pagination is offered. This can blow the context window.