An agent preset whose only route to work is delegation: every task is handed to
the OpenCode CLI through the subagent_opencode tool.
The point is the absence of tools. This preset mounts no file, shell, search, or job tools, so the agent cannot read the workspace or edit it itself. It can only describe work and delegate it. That makes it a genuine test of the OpenCode subagent — a delegation that silently did nothing cannot be papered over by the orchestrator doing the work directly.
- What the agent can do
- Install
- Prerequisites
- Choosing the preset
- Configuration
- Automatic continuation on output limit
- How it differs from a normal coding agent
| Capability | Row | Why |
|---|---|---|
subagent_opencode |
@deepseek-ai/dsh-tool-subagent |
The only way to touch the workspace |
| Interrupt / observe the child | dsh-tool-subagent-control |
Orchestration without a shell |
| Plan before acting | dsh-plan-mode |
A wrong delegation is expensive |
| Ask, track, set a goal | dsh-tool-ask-user, dsh-tool-todo, dsh-tool-goal |
Reasoning aids, no side effects |
| Compaction | dsh-compaction-basic, dsh-command-compact |
Long sessions |
| Resume a cut-off reply | @sredevopsorg/dsh-agent-opencode/auto-continue |
The orchestrator's own answer can exceed the model's output cap |
Everything absent on purpose: dsh-tool-fs, dsh-tool-fs-search,
dsh-tool-bash, dsh-tool-pwsh, dsh-tool-bash-persistent, dsh-tool-jobs,
dsh-tool-skill, skill-filesystem, dsh-workflow-ptc.
dsh plugin --profile web add ./plugins/agent-opencode
Or from a pinned release tag:
dsh plugin --profile web add "github:sredevopsorg/dsh-plugins#agent-opencode-v0.2.0&path:/plugins/agent-opencode"
Restart DSH afterwards. Installing declares the preset; it does not change any existing agent.
The bundle also enables the @sredevopsorg/dsh-subagent-opencode provider, so
it needs the same things that provider needs: the OpenCode CLI, and a login for
each provider you intend to delegate to.
opencode auth login
If opencode is not on the DSH process's PATH, either point command at it
or set requirements.install: true (see below) to have the plugin install it on
first use.
Select OpenCode in the agent preset picker. Presets are per session, so the rest of your setup is untouched.
order: 30 places it after the shipped presets (standard 1, minimal 3,
sredevops-writer 20) in the roster.
The patch enables the provider with these defaults:
| Setting | Default | Meaning |
|---|---|---|
permission |
reject |
The child refuses to edit without a human. Raise to allow for preset-driven work that should change files |
requirements.install |
false |
Install OpenCode on first delegation when command does not resolve |
requirements.require |
[] |
No mandatory gates |
Override them from the profile patch by row id:
- id: subagent-opencode-provider
name: '@sredevopsorg/dsh-subagent-opencode'
config:
providerName: opencode
permission: allow
requirements:
require: ['install']
install: true
permission: reject is the honest default: an orchestrator that delegates
everything will otherwise produce children that quietly cannot edit. Set it to
allow deliberately, once you have decided this preset should change files.
When a reply reaches the model's output token cap, the Web chat ends the turn
with an "Output token limit reached" notice that suggests sending continue.
This bundle answers that cutoff for the OpenCode preset on its own:
- A turn that ends because of the output cap (
turn/endwithreason.kind === 'max-tokens') marks the session. - Once the agent is idle again, the host row queues
continueas an ordinary user message — the same prompt a human would type, recorded in the session log like any other. - The next turn either finishes normally (the loop stops) or hits the cap again (the loop continues, guarded).
The trigger is the turn's end reason, not the rendered notice text, so the behavior is identical in every UI language and in headless surfaces.
The guard stops a model that only ever emits cutoff-length replies: more than five automatic continuations inside a rolling three-minute window pause the loop. Attempts age out individually, so continuations resume as the window drains. A prompt the human queues during the cut-off turn takes precedence; the watcher then leaves the session alone.
| Setting | Default | Meaning |
|---|---|---|
presetId |
opencode |
Preset whose sessions are served |
text |
continue |
Exact text queued as the continuing user message |
maxAttempts |
5 |
Continuations allowed inside one rolling window |
windowMs |
180000 |
Window length in milliseconds |
Override any of them from the profile patch by row id:
- id: auto-continue-opencode
name: '@sredevopsorg/dsh-agent-opencode/auto-continue'
config:
maxAttempts: 10
windowMs: 600000
Disable the watcher without removing the preset by setting disabled: true on
that row.
A normal preset gives the agent the tools and lets it work. This one gives the agent a budget of delegations and nothing else.
The child does not see the conversation, so the orchestrator has to restate context with every delegation. That is a constraint worth having: it exposes whether a task can be stated well enough for another agent to complete alone, which is the same property you need for a queue, a schedule, or a handoff.