DSH Task Orchestrator

Adaptive delegation for DeepSeek Harness. The parent model remains the only orchestrator: it creates or validates a small dependency graph, selects the minimum useful number of workers, packs them by context budget, and may assign review to one ordinary worker slot.

Author: Hasan Aghayev
License: MIT

Install

Install the public GitHub package into a DSH profile:

pnpm dsh plugin --profile web add https://github.com/hasan-aghayev/dsh-task-orchestrator.git

The package declares a dsh.bundle manifest in package.json, so the command can discover and apply cordis.patch.yml automatically. Its bundle installs one package-owned delegation group containing the workflow engine and the model-facing surface: subagent, subagent_fork, send_message, interrupt_agent, and list_agents. It does not enable DSH's separate model-facing workflow tool. The standard web-profile rows stay disabled, so disabling this plugin in DSH Market disables the group and all of these tools together. Remove it with:

pnpm dsh plugin --profile web remove dsh-task-orchestrator

Restart the profile after installation if it is already running.

What it does

The plugin adds the task_orchestrate tool and enables the standard DSH delegation tools. A simple request stays on the normal path. A complex request is scored with a deterministic detector and then sent through these stages:

1. The parent orchestrator supplies a strict JSON plan, or the plugin creates a minimal deterministic graph without a planner child.
2. Each worker receives only an explicit TASK, GOAL, RELEVANT CONTEXT, CONSTRAINTS, KNOWN FACTS, FILES / CODE, DEPENDENCIES, EXPECTED OUTPUT, and DO NOT packet.
3. The scheduler starts a role only after its dependencies complete. Independent read-only roles are admitted incrementally by context tier and active-generation limits; when one role settles, the next fitting role can start without waiting for its sibling.
4. Workers return structured evidence, changed files, tests, blockers, next steps, and can request NEED_FILE, NEED_HISTORY, NEED_MORE_CONTEXT, NEED_DEPENDENCY, NEED_BUDGET, or NEED_TOOL_RESULT. Dependency reports are compacted to bounded facts before they are handed to another worker or reused for a context escalation; only NEED_MORE_CONTEXT triggers one bounded context escalation.
5. A task with role reviewer is an ordinary worker and produces the final review fields; no seventh child is created.

Supported roles are researcher, architect, backend, frontend, tester, documentation, and reviewer. The plugin uses the existing DSH subagent and workflow services and does not modify the agent loop.

The model can also call subagent for a fresh child, subagent_fork for a child that inherits completed parent turns, and list_agents, send_message, or interrupt_agent to manage continuable children. These tools are package-owned group children rather than edits to the base rows. That ownership makes the Market toggle atomic: plugin off means the group and all five model-facing tools are off, while the web profile's standard rows remain off as well.

Safe defaults

The default mode is hybrid:

• simple requests do not start child agents;
• complex read-only requests can run automatically;
• plans that require writes stop at plan-only until a human approves them;
• writes and parallel writes are disabled by default;
• workers require a fresh structured-output subagent provider;
• at most six workers can exist in one orchestration; the parent is not counted as a child, so the logical maximum is one parent plus six workers;
• oversized plans, reports, and parent notices are rejected or truncated at configured limits.

Automatic write execution should be enabled only in a profile that has its own approval and workspace policy:

config:
  mode: hybrid
  minComplexityScore: 55
  subagentProvider: spawn
  preferredWorkers: 2
  maxWorkers: 6
  maxTotalAgents: 6
  maxConcurrentAgents: 2
  allowWrites: false
  allowParallelWrites: false
  requireReview: true
  maxActiveGenerations: 2
  hardContextTokens: 98304
  priorityAgingMs: 30000
  totalContextTokens: 98304
  contextCompactionChars: 4096
  concurrencyByContext:
    - { maxContextTokens: 8192, maxActiveGenerations: 2 }
    - { maxContextTokens: 16384, maxActiveGenerations: 2 }
    - { maxContextTokens: 24576, maxActiveGenerations: 2 }
    - { maxContextTokens: 32768, maxActiveGenerations: 2 }
    - { maxContextTokens: 49152, maxActiveGenerations: 2 }
    - { maxContextTokens: 65536, maxActiveGenerations: 1 }
    - { maxContextTokens: 81920, maxActiveGenerations: 1 }
    - { maxContextTokens: 98304, maxActiveGenerations: 1 }
  minimumVramHeadroomGiB: 0.8
  parentOrchestratorOnly: true

The suggest mode always returns a plan first. off disables automatic planning but keeps the explicit tool. auto is available for deployments that intentionally permit automatic execution.

maxActiveGenerations is the deployment ceiling. concurrencyByContext can lower it for large requests, so a small worker may use more lanes on hardware that passes the corresponding benchmark. On the current RTX 3090 profile every small-worker tier is intentionally set to two and the larger tiers to one: NInfer failed its startup memory reservation at three and six, so the profile does not claim unsafe parallelism. A stream holds a lane only while its output is consumed; an agent waiting for tools or children does not hold one. Completed streams release their context budget immediately, so a queued worker can replace them while another active worker continues. The parent orchestrator starts with preferredWorkers, then admits additional ready workers as results free capacity, up to six. contextCompactionChars bounds dependency reports before escalation or handoff. priorityAgingMs raises a waiting request by one priority level after the configured interval, so a long-running worker cannot wait forever. hardContextTokens is a conservative pre-tokenization guard; NInfer remains authoritative for exact token counts. totalContextTokens is a conservative active-work budget, not a promise that requests fit concurrently in GPU memory. minimumVramHeadroomGiB documents the safety target used when selecting a deployment profile.

The package defaults are conservative (maxActiveGenerations: 2 and two lanes for small requests). To test higher small-worker concurrency, raise the global ceiling and provide matching concurrencyByContext entries in the deployment profile only after a benchmark on that hardware.

The task_orchestrate tool accepts objective, optional parent-created plan, optional planOnly, optional executeWrites, and an optional maxWorkers cap. A request-side cap can never exceed the configured deployment ceiling.

Development

pnpm install
pnpm check
pnpm test
pnpm build

The package emits JavaScript to lib/ and declaration files to lib/types/. The published package contains only the built runtime, declarations, bundle patch, and paired READMEs.

Limitations

• Workers share the profile workspace. Parallel writes remain disabled by default and the plugin does not create automatic git worktrees.
• Complexity detection uses lexical signals and may miss a short difficult request or classify a long simple request as complex.
• NInfer does not force arbitrary JSON output by itself. DSH validates plan and worker fields after each model response, but that validation does not replace human review.
• Write scopes are declared to the scheduler and worker prompts; the surrounding DSH profile remains responsible for filesystem permissions and approval policy.
• Context tiers are 8K, 16K, 24K, 32K, 49K, 65K, 81K and 96K. The local profile keeps the shared budget at 98,304 tokens and applies a two-stream safety ceiling after the RTX 3090 reservation benchmark; higher per-tier concurrency values require a separate hardware benchmark.
• Compaction is deterministic and bounded. It preserves the report fields needed for scheduling and review, but it is not a semantic summary and does not prove that NInfer restored a KV cache after a restart.

Repository

Source: https://github.com/hasan-aghayev/dsh-task-orchestrator