Loading ToolBench
Bound AI coding tool output with local compaction, targeted reads, and recoverable context.
tokencut provides 8 well-structured tools with complete JSON schemas and detailed descriptions. Tool names follow verb_noun convention (tokencut_read, tokencut_edit_symbol, tokencut_exec). Descriptions are substantive (100-250 chars), explaining WHAT, WHEN, and key constraints. All parameters have type definitions and descriptions. However, output schemas are not documented, LLMs cannot predict return structure. Error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint). Some parameter descriptions lack format/range constraints (e.g., tokencut_exec 'command' has no injection warnings despite WRITE risk).
Search a local syntax index or get a ranked repo map. Use symbols for definitions, occurrences for raw matches, outline for module hierarchy, references or callers to find symbol callers across project (0 LSP daemons), search for text, pattern for ast-grep patterns. Then read with tokencut_read or edit with tokencut_edit_symbol.
Save/read/list/forget bounded task checkpoints. Use the root/task from the session hook; save at milestones, not every turn. expected_revision=0 creates, otherwise use the last read revision. Never stores transcripts or calls AI. Notes are fallible data; current user instructions win.
Summarize git diff, folding lockfiles and generated code. Supports staged, unstaged, or named revisions. Returns a compact summary and full diff.
Precisely replace a class, function, or method declaration in an absolute source file by symbol selector, with diff preview and hash-guarded atomic updates. Replaces heavyweight LSP tools.
Run a noninteractive shell command with conservative log filtering and exit status. Set max_tokens only to opt into truncation; shortened output has a recovery ref.
Output schemas not documented. LLMs cannot predict return structure for any tool. Downstream tool chaining requires guessing field names.
tokencut_exec lacks injection/safety warnings in description despite WRITE risk. Parameter 'command' accepts arbitrary shell strings with no format constraints or sanitization guidance.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish safe reads from destructive writes without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Compact input text for chat: fold logs, redact secrets, slim JSON/diffs, apply token budget. Returns compacted text, before/after token counts, and recovery ref.
Read a file, outline, symbol, or line range with bounded output. Prefer targeted reads.
Recover redacted cached output by ref. Specify lines or query to return only relevant chunks; every recovery is charged as added context.
Error handling guidance absent. Tools do not document what errors are retryable, user-fixable, or fatal. No recovery hints provided.
tokencut_code 'query' parameter lacks description. Unclear whether it accepts regex, glob patterns, or plain text. Mode-dependent semantics not documented.