mirror of
https://github.com/deepseek-ai/deepseek-harness
synced 2026-08-15 21:04:50 +00:00
89 lines
4.5 KiB
Markdown
89 lines
4.5 KiB
Markdown
# @deepseek-ai/dsh-plan-mode
|
|
|
|
Logged, per-agent plan collaboration state with deployment-owned guidance, a direct `/plan [message]` entry command, and the reviewed `exit_plan_mode` exit. Plan mode is soft guidance; sandbox mode and approval policy remain independent enforcement axes.
|
|
|
|
## Durable state
|
|
|
|
`plan/mode` (`{ active: boolean }`) is a log-only, whole-value-replace `SessionEventMap` member. `foldPlanMode(events)` returns the last logged value or `false`, so resume, fork, and compaction recover plan state directly from the session log. UIs observe committed flips through `session/event`.
|
|
|
|
`ctx.planMode.set(agent, active)` records a pending selection and flushes it inside the next turn boundary. `get(agent)` returns `{ active, pending? }`, separating the logged state shaping the current step from a user's optimistic selection. Prompt submission, ordinary continuation, and request-recovery retry are all covered; a changed user selection contributes one `context/message` notice when the last logged request header described the other state.
|
|
|
|
## Model and human surfaces
|
|
|
|
While active, `plan:policy` renders the configured `section`. The plugin always registers `exit_plan_mode`, keeping tool schemas stable across the transition; its execute path accepts only active plan mode and leaves it only after an exact user approval through `ctx.userInteraction`.
|
|
|
|
When `ctx.commands` is composed, the package registers `/plan [message]`. The command selects plan mode first. A non-empty argument is then submitted through `agent.steer()`, so it becomes the next step's ordinary logged user message under plan guidance; bare `/plan` only changes state.
|
|
|
|
ACP is an adapter, not the owner of this vocabulary: it advertises the fixed wire ids `default` and `plan`, maps `session/set_mode` to the boolean service, and translates committed `plan/mode` events back to `current_mode_update`.
|
|
|
|
## Configuration
|
|
|
|
```yaml
|
|
- id: plan-mode
|
|
name: '@deepseek-ai/dsh-plan-mode'
|
|
config:
|
|
section: |
|
|
You are in plan mode. Explore and design before presenting the complete
|
|
plan through exit_plan_mode.
|
|
```
|
|
|
|
`section` is required and non-empty. Unknown keys fail at load. The package does not accept arbitrary named modes, tool filters, sandbox settings, or approval policy.
|
|
|
|
Design: [plan-mode Agent Note](../../../.agents/notes/implemented/feature/2026-07-07-plan-mode.md) and [plan-specific state simplification](../../../.agents/notes/implemented/simplification/2026-07-22-plan-specific-collaboration-state.md).
|
|
|
|
## Model Experience
|
|
|
|
### Plan policy system prompt
|
|
|
|
#### What the model sees
|
|
|
|
While plan mode is active, the model sees the deployment's exact `section` text at prompt order 50; inactive mode contributes no text.
|
|
|
|
##### Configuration example
|
|
|
|
```markdown
|
|
You are in plan mode. Explore and design before presenting the complete plan through exit_plan_mode.
|
|
```
|
|
|
|
#### Token effect
|
|
|
|
Inactive mode adds no tokens; active mode adds the configured section to every request.
|
|
|
|
#### KV Cache effect
|
|
|
|
The section is stable within plan mode, but entering or leaving changes the system prompt from order 50 onward.
|
|
|
|
### Optional command message
|
|
|
|
#### What the model sees
|
|
|
|
`/plan` and its terminal result stay outside model history; a non-empty suffix becomes one trimmed user text block through `agent.steer()` after plan mode is selected.
|
|
|
|
#### Token effect
|
|
|
|
The suffix costs the same history tokens as submitting that text separately; a bare command adds none.
|
|
|
|
#### KV Cache effect
|
|
|
|
The user block is append-only conversation growth, while entering plan mode also changes the earlier policy section.
|
|
|
|
### Exit tool schema and review exchange
|
|
|
|
#### What the model sees
|
|
|
|
The [`exit_plan_mode` schema](../../../docs/tool-catalog.md#deepseek-aidsh-plan-mode) remains available in both states; execution outside plan mode fails, while an approved in-mode review returns the exit result and rejection returns feedback.
|
|
|
|
#### Token effect
|
|
|
|
The stable schema is paid according to ToolRegistry mode, and each plan argument and review result remains in conversation history.
|
|
|
|
#### KV Cache effect
|
|
|
|
Mode transitions do not change the tool catalog; plan arguments and review results extend the conversation normally.
|
|
|
|
## Known Limitations and Deferred Work
|
|
|
|
- Plan mode guides rather than enforces; deployments needing a hard boundary must combine independent sandbox and approval controls.
|
|
- A pending selection made while idle is lost if the process exits before the next boundary, so the UI must reapply it.
|
|
- Forked agents inherit logged plan state, while newly spawned agents begin inactive; there is no creation-time plan option.
|