Delegate one stage of work to DSH (starts a model request)
dsh_submit_taskSubmits a bounded development stage by validating the workspace, creating an isolated task directory, starting an SDK runtime, and returning task/session IDs immediately; use wait/result tools.
Instructions
Validates the workspace against the allowed roots, creates an isolated task directory and DSH_HOME, starts a private SDK runtime process, sends the stage prompt and returns immediately with taskId/sessionId/state. It does NOT wait for the work to finish: use dsh_wait_task and dsh_get_result. May create or modify files inside the workspace.
Tool call template
Replace each placeholder with a value matching the input schema. Only required parameters are included; add optional parameters as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {
"workspace": "<workspace>",
"goal": "<goal>"
},
"name": "dsh_submit_task"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| goal | Yes | What this stage must produce. | |
| context | No | Background the model needs. | |
| budgetMs | No | Per-stage wall-clock budget. Default 300000ms when omitted; explicit shorter/longer values are preserved. For high effort complex tasks consider 600000ms or more. Expiry stops this task SDK process and marks timed_out. | |
| maxTokens | No | Per-response output budget, including model reasoning. Default 16384, maximum 32768. Explicit short budgets are preserved, not raised to a floor; values below 2048 get a warning. Output exhaustion is incomplete, not acceptance. | |
| workspace | Yes | Absolute path to the task workspace. Must be inside the bridge's allowed roots. | |
| acceptance | No | How Codex will judge this stage. | |
| autoRepair | No | Opt-in same-session repair on failed host verification. Requires protected acceptance files. Stops on success, unchanged failures, limits, permission errors or cancellation. | |
| constraints | No | Rules the model must respect. | |
| autoContinue | No | Bounded same-live-session continuation ONLY after a real provider max_tokens/length turn end. Omission is disabled; deliveryPreset complete enables maxAttempts=2, noProgressLimit=2, totalBudgetMs=3600000. Explicit fields win. Storage/preview truncation, cancel, permission/credential failures, a dead runtime, shutdown and verifier-integrity failures never trigger it. | |
| deliveryPreset | No | Complete delivery: high/32768/30min, 10min verification, up to 3 automatic repair stages. Explicit fields override defaults. Requires an independent command and protected acceptance files. | |
| deliverySource | No | Explicit 1.0.7 source scope for reproducible complete/patch delivery. Omit for the legacy observed-changes export. | |
| reasoningEffort | No | Default medium for work. Use explicit low for short diagnostic probes and high for difficult tasks. none is a legacy alias of off; on GOAT off/none maps to low with an explicit warning, without changing provider. | |
| executionProfile | No | Optional execution profile. Omit for the legacy no-preset overlay. "native" mounts the minimal agent preset with native tool schemas; "ptc" presents model tools through run_code and is an explicit opt-in only. Any other value is refused before the task directory is created. | |
| verificationCommand | No | Optional argv array the BRIDGE runs itself after the stage ends, in the workspace. No shell is involved. Its real exit code is recorded separately from the model report. | |
| verificationTimeoutMs | No | ||
| protectedVerificationFiles | No | Existing acceptance files, including checker dependencies and fixtures. Bridge pins real paths and SHA-256 before the model starts and checks before/after verification. |