@deepseek-ai/dsh-command-workgraph

@deepseek-ai/dsh-command-workgraph

English | 中文

Human-facing /graph control over ctx.workGraph. The plugin registers one global command through ctx.commands, so every composed command adapter discovers and executes it without a model turn. The grammar and rendering are faithful ports of jxca-cli's /graph (the work-graph spec owns the contracts).

Command contract

Input	Result
/graph or /graph status	Render the per-node state tree: glyphs, waits, rounds, token spend, the budget line, pending discoveries, and the pause reason.
/graph show	Render the layered box-drawing DAG (status glyphs and legend), degrading to the status tree when the packing cannot fit the width budget.
/graph [--budget ]	Plan and run the objective as a dependency graph; a trailing own-token --budget sets the token budget.
/graph pause	Pause the live episode; the command waits for bounded child settlement (the per-child await budget) before returning.
/graph resume [--budget ]	Resume a paused/blocked graph; a top-up re-enters a budget-limited graph, while a plain resume on budget_limited prints the top-up hint.
/graph retry [node]	Retry one failure chain by node id, or (bare) every failed chain as ONE union batch — a shared final blocked by sibling failures refuses any single-root reset.
/graph clear	Clear the graph and its durable tombstone; a cleared graph cannot resurrect.

Control words are case-insensitive only when they occupy the complete input. ANY input starting with resume or retry resolves to that command and NEVER falls through to set — a typo'd top-up must not silently replace a resumable budget-limited graph. Only a TRAILING, standalone --budget flag with a final all-digit positive token is consumed; anything else stays part of the objective. Expected domain rejections (already set, not retryable, budget hint, unknown node) become stable direct command errors; unexpected implementation failures still reject dispatch so adapters can report them as command failures.

Composition

The producer injects commands and workGraph. A custom app mounts their owners plus this plugin:

- id: commands
  name: '@deepseek-ai/dsh-commands'
- id: workgraph
  name: '@deepseek-ai/dsh-workgraph-scheduler'
- id: command-workgraph
  name: '@deepseek-ai/dsh-command-workgraph'

The scheduler provider owns the durable graph; the command is a pure adapter over ctx.workGraph.

Model Experience

Human /graph control

What the model sees

The slash input, mutations, and direct status/error output are absent from model requests. The engine records every accepted transition as workgraph/change; the GUI (issue 08) renders the live DAG from those events. Presentation text is never logged.

Token effect

Reading status, mutating a graph, or receiving a direct command error adds no model tokens. The graph's worker and verifier children spend their own tokens, charged to the graph budget through the session usage records.

KV Cache effect

Command discovery, mutations, and direct output do not affect the cache. Child requests follow their own fresh sessions.

Known Limitations and Deferred Work

• Blocking set — /graph runs the episode to its first settlement (complete, paused, or blocked) before returning, holding the agent's command loop; the parent conversation still spends no model turn. A detached start remains future work.
• Plain-text interaction only — the generic command registry has no modal forms; inline --budget flags and explicit control words keep intent deterministic across adapters.
• No continuous status widget — bare /graph is the portable observation API; the live DAG view lands with issue 08.