@creait/dsh-research-mode

Deep research as an agent mode for DeepSeek Harness — a fixed, reviewed research loop that plans, researches in adaptive parallel rounds, synthesises a cited report and then argues with it. It exists only inside its own preset.

Why this exists

dsh already ships every piece of a research fan-out: a workflow engine, a subagent seam, web_search, and — with a fetch provider mounted — web_fetch. What it does not ship is the loop.

You can ask the model to write one, with the workflow tool. The problem is that a research script authored fresh on every call re-earns the same structural mistakes on every call, and all of them are invisible from the outside: the report still arrives, still reads well, and is quietly less than it claims.

So the loop ships here as a constant — lib/script.js, reviewed once — and the model supplies parameters to it.

The design (planner → adaptive rounds driven by the researchers' own declared gaps → synthesis → adversarial review) is ported from dsh-deep-research by omdsh-dev (MIT), which got the structure right. Four things are fixed:

#	Upstream	Here
1	Questions still queued when the round cap hit were dropped: not researched, not counted, not mentioned.	The planner is told its budget up front, and every question the budget never reached leaves the loop by name — into the result, into the report's own "what this does not cover" section, and into the coverage footer.
2	Harvested leads were capped at the round width and the rest discarded, so a round where every researcher found something urgent lost most of it.	Every high-priority lead is queued. If the budget ends first it is reported as deferred, never dropped.
3	The seen set was rebuilt inside the round loop, so a gap three consecutive researchers reported got researched three times.	One dedup set for the whole run.
4	Supplying your own questions skipped planning entirely — no scope, no dimensions, no coverage audit.	Supplied questions are mandatory and used verbatim, and the planner does its job on top of them: it names the scope they imply and records what they miss.

It is also written in English throughout, which the original is not.

The mode

A mode is a dsh agent preset: a composition mounted per agent, not per process. Choosing Research mode in the picker gets you an agent with

• deep_research — the loop, as one tool call
• web_search + web_fetch — which the loop's own child agents inherit, because they are spawned onto this same composition
• file tools — to read source material and write the report somewhere it survives the session
• todo, ask-user, and compaction tuned so a report-sized tool result is not pruned to death
• skills — skill-filesystem + tool-skill, over the same default roots as the shipped modes, so a source that only answers to a particular method can be written down once and reused. Only skill descriptions sit in context; a body is loaded when the agent calls the skill tool.

and, deliberately, no shell, no workflow tool, no plan mode.

Every other mode is untouched. deep_research is not in their tool catalog and none of the guidance above is in their context window — which is the whole point of shipping this as a mode rather than as a plugin that registers a tool everywhere.

Install

▼
bash
复制代码
dsh plugin --profile web add @creait/dsh-research-mode

Restart dsh — the boot manifest is assembled at startup. The plugin copies its preset into <dsh home>/.agent-presets/research/, and Research mode appears in the picker.

That directory is writable and the copy is yours to edit. The installer records what it wrote and will not overwrite a preset you have changed; on upgrade it leaves your version alone and says so in the log.

Prerequisite: a fetch provider

The preset mounts tool-web with fetch: true. dsh implements web_fetch in full but ships no fetch provider, so without one every fetch call fails and the researchers are capped at search snippets — which is worth considerably less than being able to open the page.

@creait/dsh-web-fetch is that provider:

▼
yaml
复制代码
- id: web
  config:
    fetchProvider: local

- id: tool-web
  config:
    fetch: true

- insert:
    - id: web-fetch
      name: '@creait/dsh-web-fetch'

If you are knowingly running without one, set fetch: false in the preset.

The preset also raises fetchMaxOutputChars from its 200,000 default to 400,000. tool-web applies that cap to the source before turndown converts it, not to the markdown that comes out, so on HTML it budgets pre-cleanup bytes — most of which the converter discards. At the default a long article reaches the researcher as its first third, flagged truncated, and a lead gets spent chasing the rest of a page it already had. Turndown compresses 5-10x, so the rendered markdown still lands well under the raised cap; the cost is roughly 10k tokens on the few pages large enough to need it. A text/plain body skips the converter and can arrive at the full cap.

The loop

▼
text
复制代码
Plan        one agent, structured output: scope, dimensions, sub-questions,
            and the planner's own declared coverage gaps
Research    up to `width` agents per round, up to `rounds` rounds. Each returns
            confirmed claims (with sources), uncertainties (with reasons), and
            gaps (with priority). High-priority gaps are queued as the next
            round's questions, behind the originals — breadth before depth.
Synthesize  one agent, given every finding and the coverage accounting, writing
            the report in `language` for `audience`
Review      one agent checking the report against its own evidence for
            fabricated citations, overclaiming, contradictions and coverage
            honesty; a second applying the critique

Structured output at the plan and research stages is load-bearing. A sub-agent that returns prose has to be parsed, and a parser over model prose is where a research pipeline quietly starts inventing things.

Every stage fails soft except the two that cannot: a dead planner and a run where every researcher failed both raise, because there is nothing to write from. A single failed researcher is recorded as an unanswered question and the run continues.

Calling it

▼
text
复制代码
deep_research({ topic: "…" })

Argument	Default	Meaning
topic	required	The question the report has to answer. The planner sees this string and nothing else about the conversation, so the constraints belong in it.
questions	—	Sub-questions that must be researched, used verbatim and first. The planner still audits them.
rounds	3	Adaptive rounds. Depth on what the research turns up.
width	4	Questions researched in parallel per round. Breadth. A width pinned in the composer overrides whatever the model passes.
audience	—	Who the report is for, when it changes what belongs in it.
language	English	Report language.
review	true	The adversarial pass.

rounds and width are clamped to the preset's maxRounds / maxWidth (6 and 8), which the model cannot argue past.

Pinning the width

Research-mode sessions get a width control in the composer, beside the access mode. Leave it empty and the model picks a width per call, as above. Type a number and that number wins — it replaces the width argument the model passed, and is still clamped by maxWidth.

The pin exists because width is the one research parameter the model cannot reason about. It knows how broad the topic is; it does not know anything about the deployment. Pinning it makes the brief a deployment decision rather than a per-call guess.

It is not the capacity mechanism — see Width is not concurrency. Pin the width to shape the report; capacity is gen-limit's job, not this one's.

The value is a deployment-wide setting, not a per-session one: it persists in ~/.dsh/settings.yaml under the dsh-research-mode namespace and applies to every research run until changed. The tool reads it at call time, so a change takes effect on the next call without a restart.

Namespace	dsh-research-mode
Key	width — 0 means no pin
Route	GET/POST /api/dsh-research-mode/config, loopback only

The control renders only in sessions running the research preset — including on the new-session screen, as soon as Research mode is picked from the hero chip, so the width can be set before the first message.

Width is not concurrency

width is how many questions a round takes, not how many researchers run at the same moment. The script hands the engine the whole round at once, and what decides how many of them generate simultaneously is not in this plugin at all.

It is @creait/dsh-gen-limit, which caps concurrent generating sessions per provider/model — and every subagent is its own session (dsh-subagent mints a fresh SessionId per child), so a round of eight researchers wants eight slots and gets however many the limit allows. The rest wait: at capacity gen-limit queues a request FIFO rather than refusing it.

So any width works, at any capacity. A round of eight against a limit of three runs three, then three, then two — slower than a machine that could take eight, and identical in output. Width goes back to meaning what it reads like it means.

Set the limit in gen-limit's settings page, per provider and model. That is the only place it lives:

Namespace	dsh-gen-limit
Key	limits[].max — concurrent generating sessions; -1 is unlimited
Also	queueTimeoutMs (how long a request waits before it is refused instead), maxQueued (how deep the line may get)

Why not the workflow engine's maxConcurrentAgents

The engine has its own per-run cap and this preset deliberately leaves it unset. It is the wrong instrument for this:

• It is per run. WorkflowExecution is constructed per workflow run, so two research runs at a cap of 3 are 6 concurrent generations. It cannot express a machine-wide ceiling because it has never seen the other runs.
• It cannot see your chat. An ordinary session's turns never touch the workflow engine, so they are invisible to it and it to them.
• It duplicates a number you already set. Your deployment's capacity is declared once, in gen-limit. Writing it again in a preset file that ships with this plugin means two copies to disagree the day you change one.

Set it only if you are running research without gen-limit installed. Unset, it derives from CPU count (min(16, cores - 2)), which measures the wrong machine — what runs out is generation slots on the backend, not cores here.

Retry still earns its place, but for a narrower case than before. With gen-limit queueing, a request at capacity waits instead of failing, so the retry waterfall is no longer what keeps a wide round alive. It covers what is left: a wait that runs out (queueTimeoutMs), a queue too deep to join (maxQueued), and the ordinary transport failures that have nothing to do with capacity. Keep GEN_CAPACITY_EXCEEDED in retryableCodes — it is now rare rather than routine, and it means the backend has been saturated for a sustained period.

The coverage block

The result carries its own audit, and it is rendered under the report rather than left in a structured field nobody reads:

▼
text
复制代码
Researched 9 of 12 planned questions over 3 rounds at a width of 4, plus 2 follow-ups surfaced during the run.
Evidence: 41 sourced claims, 6 of them low-confidence, 8 recorded uncertainties.

Never answered (3) — the report above does not cover these:
- What does the enterprise tier actually cost at 500 seats?
- …

This is the part that makes the numbers honest. A report that arrives without them reads as complete coverage of its topic, and the agent relaying it has no way to know it is not.

Config

Set on the research-mode-tool row inside the preset:

Key	Default	Meaning
rounds	3	Default rounds when the call names none
width	4	Default parallel width
maxRounds	6	Ceiling the model cannot exceed
maxWidth	8	Ceiling the model cannot exceed
review	true	Run the adversarial pass by default
language	English	Default report language
toolName	deep_research	Rename the tool

Two entry points

Specifier	Plane	Injects	Does
@creait/dsh-research-mode	roster (profile bundle)	nothing required; reaches for settings and webServer through scoped injects	Installs the preset, registers the dsh-research-mode settings namespace, serves the width-pin route, and ships the composer control. Registers no tool, no prompt, no command — nothing the model can see.
@creait/dsh-research-mode/tool	agent (inside the preset)	tools, workflowEngine	Registers deep_research.

They are separate modules rather than one behind a config flag because Cordis inject is all-required and gates loading. The Web surface disables workflow-worker-thread on the host plane and each preset mounts its own, so a roster row declaring workflowEngine would wait forever for a service that composition never publishes — and the preset would never install, so the mode would never appear.

For the same reason the tool row must sit inside the group carrying isolate: { workflowEngine: true }, alongside the workflow-worker-thread row. The realm is entry-local and invisible to siblings outside the group.

To put deep_research in another mode, copy that group into its preset.

Tests

▼
bash
复制代码
node --test test/*.test.js

The loop is tested the way the engine runs it — a node:vm context with the same six globals (agent, parallel, pipeline, phase, log, args) and the same (async () => { … })() wrapper — with agent answering from a script instead of a model. Each of the four fixes above has a test named after the flaw it pins down, alongside the failure paths, the tool's clamping and result shaping, the coverage renderer, and the installer's refusal to overwrite a locally edited preset.

Licence

MIT. The loop's design is ported from dsh-deep-research (MIT, Copyright (c) 2026 dsh2026), with thanks. See LICENSE for the full acknowledgement.