Loading ToolBench
ChatGPT thinks. Codex works. Use ChatGPT as the planning brain while keeping the Codex harness.
16 tools with complete schemas and descriptions, but significant gaps in clarity and LLM-optimization. Tool names follow verb_noun pattern well (omnigent_submit_task, zcode_enqueue_task). Descriptions are present but often verbose (200-400 chars) and domain-specific jargon-heavy, making them less actionable for LLMs. Parameters have types and descriptions, but many lack concrete constraints (enums, ranges, patterns). Error handling is mentioned but recovery guidance is sparse. Security redaction is implemented (safeOutput, credential filtering) but not documented in tool descriptions. Output schemas are not explicitly documented.
Get a policy-grounded recommendation on which provider and model to use based on current Quanta quota telemetry, task type, and risk level. Returns decision=route with a primary recommendation plus a fallback chain, or decision=blocked when every quota pool is exhausted/unavailable (no forced recommendation is made). GLM and DeepSeek are telemetry-only: they are surfaced in usage data but are not auto-scheduled targets.
Read local Quanta telemetry for available AI agents and quota pools (Codex, Antigravity Gemini, Antigravity Claude/GPT shared pool, GLM, DeepSeek, Muse). All secrets, emails, and credentials are strictly redacted.
Request cancellation of a known non-terminal Omnigent task. This only appends CANCEL_REQUESTED to the task's control record — it never writes terminal receipts; the Omnigent coordinator owns COMPLETED/FAILED/CANCELLED. Repeated requests are idempotent.
Read an Omnigent task's status and, once the turn has settled, its bounded assistant output. Includes the harness, worktree branch, model identity, token usage and any task error.
Descriptions are verbose (250-350 chars) and domain-jargon-heavy ('Omnigent meta-harness', 'Z2C control plane', 'worktree branch'). LLMs struggle with specialized terminology without context. Descriptions should be 50-200 chars and explain WHEN to use the tool, not internal architecture.
Output schemas are not documented. Tools return structured data (status, task_id, output, error) but LLMs cannot see what fields to expect. Add explicit 'Returns' sections documenting field names, types, and when they are present.
Enum values are not always documented in descriptions. 'effort' and 'speed_profile' in omnigent_submit_task reference OMNIGENT_EFFORTS and OMNIGENT_SPEED_PROFILES constants but do not list valid values. LLMs cannot see the constant definitions, they need the enum values in the description.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Check the local Omnigent orchestration backend: server health, execution host, harness readiness (codex, glm/oh-my-pi) and registered provider agents. Note: gemini is not an Omnigent provider on this deployment — submit Gemini tasks via submit_codex_task(provider=gemini), which drives the local agy CLI with Google OAuth. Call this before submitting Omnigent tasks.
Dispatch a coding task to the local Omnigent meta-harness. provider selects the execution stack: codex (OpenAI Codex) or glm (oh-my-pi on Z.AI GLM). The provider is pinned — an unavailable provider fails the task and is never silently substituted. never silently substituted. Optionally creates a dedicated git worktree branch; the task never writes outside it when one is created. Pass session_id of a previous result to continue that session instead of creating a new one.
Request cancellation of a known non-terminal governed ZCode scheduled-queue task. This only appends CANCEL_REQUESTED to control.jsonl — it never writes terminal receipts; the ZCode coordinator owns COMPLETED/FAILED/CANCELLED. Repeated requests are idempotent.
Append one governed task to the fixed governed ZCode scheduled queue. network must remain false. The task_id must be unique across the queue and its receipts. Credential-like instructions (token/password/cookie/api_key/client_secret/private key) are rejected. The ZCode governed coordinator claims tasks while its claim window is open; terminal receipts are written only by that coordinator.
Read one governed ZCode scheduled-queue task: merged status (queued → cancel_requested → running → completed/failed/cancelled), its receipts and any recorded error. receipts.jsonl is the lifecycle truth.
List governed ZCode scheduled-queue tasks in queue order with merged lifecycle status, the bounded worker state cache, and a control_plane health layer (QUEUE_ROOT_MISSING, COORDINATOR_NOT_RUNNING, OUTSIDE_CLAIM_WINDOW, ZCODE_DESKTOP_UNAVAILABLE, AUTH_NOT_ATTESTED, WRONG_PROVIDER, READY). Use zcode_get_task for one task's receipts.
Request cancellation of a known non-terminal native ZCode task. This only appends CANCEL_REQUESTED to the task's control record — it never writes terminal receipts; the ZCode coordinator owns COMPLETED/FAILED/CANCELLED. Repeated requests are idempotent.
Read a native ZCode task's status, output, and lifecycle. Includes workspace binding verification, provider identity, and any task error.
Read and attest the exact native session workspace, Desktop-managed provider and GLM model binding. Fails closed on any mismatch. No scheduled-queue scheduling.
Run a fixed server-owned readonly plan probe to verify native durable idempotency, exact Desktop GLM binding, replay, conflict, and unchanged C2C queue/writer state. Creates one native task and cancels it only when execution.cancel is authorized. Returns bounded evidence only.
Health and Desktop GLM binding identity of the independent Z2C control plane, scoped to the authorized workspace_id. Reports the binding observed via Z2C's native exact-session read for the workspace's current session; UNKNOWN when no observable session exists. Status is informational and never blocks task creation — admission identity is proven per task. Read-only; no fallback.
Dispatch a realtime native ZCode task through Z2C in an authorized governed workspace enabled via ZCODE_NATIVE_ALLOWED_WORKSPACES. Honors the shared workspace queue pause/freeze and writer slot. Z2C admits the task only after observing the sanctioned Desktop-managed binding (builtin:zai-start-plan/GLM-5.3-Flash) on the exact created session, and the returned task binding is re-verified here — fails closed, never falls back to the scheduled queue.
Error recovery guidance is missing. Tools mention failure modes ('fails closed', 'never silently substituted') but do not tell LLMs what to do next. E.g., omnigent_submit_task should say 'If provider unavailable, call omnigent_status() to check harness readiness.'
Security redaction (credential filtering, path redaction) is implemented in code (safeOutput, redactLocalPaths) but not documented in tool descriptions. LLMs do not know that sensitive data is stripped, they may avoid passing legitimate context or distrust the tool.