Loading ToolBench
A local Claude and Codex collaboration bridge with durable jobs, project rooms, GPT-5.6 profiles, and native image generation.
Server has 9 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention (codex_native_*, write_*, replace_*). Descriptions range 50-250 chars, meeting baseline. However, several tools lack parameter descriptions (codex_native_health, codex_native_models have 'cwd' but no description detail), and output schemas are not documented. Error handling is minimal, no recovery guidance or actionable error messages. Tool composition is reasonable but codex_native_start combines multiple concerns (task execution, room attachment, sandbox mode). Risk annotations present but tool annotations (readOnlyHint/destructiveHint) not visible in schema.
Check the Codex Native Bridge, current Codex model catalog, image generation capability, and reasoning profiles.
List recent durable Codex Native jobs. Use after Claude restarts or when a job id was lost.
List live Codex models and supported reasoning efforts. Use before an explicit model/effort override.
Delete finished job records, logs, and generated images beyond the retention window. Running jobs are never removed. Use dryRun first to see what would go.
Append one timestamped message to an isolated shared-conversation room. Use this to preserve the real human/Claude/Codex dialogue before delegating with codex_native_start and the same roomId. Messages are durable append-only JSONL entries; this tool does not start a Codex job.
Read messages after one participant's durable cursor in an isolated shared-conversation room. By default a successful read advances only that participant's cursor through the returned entries; set peek:true to inspect unread messages without advancing it.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields (e.g., job IDs from codex_native_jobs, model names from codex_native_models). Baseline: 100% of A+ tools document return types.
Error handling lacks recovery guidance. Tools return errors but do not tell LLMs what to do next (retry, ask user, call alternative tool). Baseline: pattern:recovery-guide requires actionable error messages.
codex_native_start combines multiple concerns: task execution, room attachment, sandbox mode, model selection, and timeout. Should split into separate tools or provide clear composition guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Start a durable Codex task in the background. Returns a job id immediately. Optionally attach a shared-conversation room so unread real dialogue is injected into the Codex user turn and the final response is posted back. Poll with codex_native_status, then read codex_native_result. danger-full-access requires explicit user confirmation before passing dangerousConfirmed: true.
Replace one exact text occurrence in an existing UTF-8 project file. Refuses missing or ambiguous matches.
Create or replace one UTF-8 text file inside the current project. Paths must be project-relative.
Parameter descriptions incomplete. codex_native_health and codex_native_models have 'cwd' parameter but lack detail on what 'first allowed root' means or when to override. Baseline: 100% of A+ tool params have descriptions.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) not visible in schema definitions. Risk field present in metadata but not in MCP schema format. Prevents clients from rendering safety warnings.