dsh-subagent-flowtext

中文

dsh-subagent-flowtext routes every DeepSeek Harness user task directly to the FlowText Agent Gateway in Obsidian by default. FlowText owns the complete agent loop—planning, search, reads, writes, tools, and finalization—while DSH retains only its session/UI shell and receives the final answer. The one-shot SubagentProvider remains available as an optional tool when direct mode is disabled.

Requirements

• Node.js ^22.19.0 or >=24
• DeepSeek Harness with @deepseek-ai/dsh-subagent
• Obsidian desktop with FlowText Agent Gateway enabled

Install

Install the package into the Harness Profile that should own the provider:

dsh plugin --profile web add dsh-subagent-flowtext

Before the npm registry release, install directly from the public GitHub repository:

dsh plugin --profile web add github:ytmaps/dsh-subagent-flowtext

The repository includes reviewed, precompiled dist entry points, so a GitHub-source or marketplace installation does not execute a dependency build script and needs no pnpm allowBuilds grant. Pin a commit when a deployment must not follow later changes to main.

The bundle enables direct mode and retains both the Provider and subagent_flowtext tool. No environment variable, token copy, or Profile edit is required. After restarting Harness, every task creates a FlowText task directly instead of asking a DSH model whether to invoke a tool. The first task shows an “Allow DeepSeek Harness to connect?” confirmation in FlowText. One approval stores the credential in a local mode-0600 file under the DSH credentials directory; later starts and token repair are automatic.

DSH still performs one minimal request cycle to claim the user message and record the final answer. That cycle does not invoke the DeepSeek model or execute DSH tools. Only the latest real user message is sent to FlowText; DSH system prompts, assistant history, plugin context, tool catalogs, and tool results are not forwarded.

Remove any hand-written subagent-flowtext or tool-subagent-flowtext Profile override rows left by an older release so they do not shadow the bundle defaults.

Repository maintainers should follow PUBLISHING.md for the first npm publication and subsequent tokenless GitHub Actions releases.

Marketplace discovery

DeepSeek Harness currently uses the GitHub dsh-plugin topic as its official community-discovery convention; its built-in Plugins settings page is an inventory of already installed Loader entries, not an official download marketplace. Independent community marketplaces can index this repository and install it directly from GitHub with the source command above. A marketplace listing is not a DeepSeek security endorsement; review this package and its requested FlowText policy before installation.

Configuration

Field	Default	Meaning
providerName	flowtext	Name registered in the Harness provider registry.
directMode	true	Force every DSH agent request through a complete FlowText task; set false for optional tool mode.
directProvider	flowtext-direct	DSH LLM route used by direct mode.
directModel	flowtext-agent	Display and validation model id used by direct mode.
baseUrl	http://127.0.0.1:27124/flowtext-agent/v1	FlowText v1 endpoint. Non-loopback or HTTPS endpoints are rejected before the token can be sent.
autoPair	true	Ask FlowText for interactive local authorization when no stored credential exists.
credentialPath	$DSH_HOME/credentials/dsh-subagent-flowtext.json	Optional credential-file override; the default file uses mode 0600.
clientName	DeepSeek Harness	Client name shown in the FlowText pairing prompt.
token	unset	Legacy explicit Bearer token; normal users do not configure it.
clientId	deepseek-harness	Stable owner used by FlowText task recovery.
modelId	unset	Optional model configured in FlowText for every run.
activePath	unset	Optional vault-relative active note. No parent cwd is converted into an Obsidian path.
contextPaths	[]	Vault-relative task context paths.
policy	{}	Per-run authority request. FlowText server settings can only narrow it.
runOptions	{}	FlowText options such as thinkingEnabled.
approvalDecision	deny	Unattended answer to FlowText approvals: deny, once, or session.
requestTimeoutMs	30000	Normal request timeout.
longPollMs	25000	Event long-poll duration, at most 30 seconds.
maxResponseBytes	2097152	Maximum complete Gateway response.
maxPromptBytes	1048576	Maximum UTF-8 child prompt.
maxAnswerBytes	1048576	Maximum UTF-8 result returned to the parent.

policy is sent as task data, not as authority minted by the model. The Provider instance fixes it in deployment configuration; FlowText intersects it with its own settings and enforces the result at tool exposure and action execution.

Lifecycle and failure behavior

start() publishes only after FlowText accepts an idempotent task identified by clientId + requestId. The run long-polls incremental events and reads authoritative task snapshots. Parent cancellation and dispose() both request remote cancellation and wait for local result settlement; disposal is idempotent.

The current Gateway return channel cannot round-trip a structured FlowText clarification through DSH and resume the same task, so clarification cancels the remote task and returns an error. Approval requests use the configured unattended decision, which defaults to deny. Network, protocol, timeout, restart, and oversized-answer failures return bounded diagnostics without tokens, request bodies, file contents, or raw server payloads.

The direct adapter fixes the DSH model-request retry count at zero so a complete FlowText task containing writes cannot be executed twice by the Harness retry layer. Gateway clientId + requestId, authoritative snapshots, and cancellation still protect the single remote task.

Automatic pairing accepts loopback clients only, rejects browser-originated requests, and requires an explicit approval in Obsidian. Credentials never enter the Profile, repository, shell history, or model context.

The Provider advertises no output-schema, depth-limit, tool-filter, persona, or parent-context inheritance capabilities. It intentionally does not implement prepareContinuable; FlowText owns the remote task and Harness owns the one-shot run.

Model Experience

Direct task

What the model sees

FlowText receives only the latest real user message plus deployment-fixed model, context, run options, and policy. It does not receive DSH system prompts, history, plugin context, tool catalogs, tool results, or the parent filesystem cwd.

Token effect

The FlowText model pays for the complete Agent loop. DSH makes no task-execution model call.

KV Cache effect

Independent of the parent request cache. Reuse is controlled by FlowText's model adapter, instructions, tools, and retained task context.

DSH result

What the model sees

DSH receives only the final FlowText answer. A non-completed task returns a bounded safe diagnostic. FlowText plans, actions, observations, task ids, events, logs, tokens, and Gateway payloads are not copied into the DSH Session.

Token effect

Only the retained final answer or error enters the DSH conversation log.

KV Cache effect

Append-only: the result is appended after the reusable parent prefix and does not rewrite prior messages.

Known Limitations and Deferred Work

• One fresh FlowText task per DSH user message; DSH history is not automatically converted into a continued FlowText session.
• Text prompts and final text answers only; image and structured-output delegation are rejected.
• Clarifications fail closed because the one-shot provider interface has no user-question channel.
• Intermediate FlowText progress is consumed for wakeups but is not projected into the parent transcript or UI.
• approvalDecision: once|session is unattended authority and must be selected explicitly; deny is the default.
• CLI authority remains a high-risk FlowText deployment choice and may exceed fine-grained vault path restrictions.