dsh-swarm
BSP coding swarm plugin for DeepSeek Harness — nested, barrier-synchronized coding swarms of 100+ concurrent agents tiered across DeepSeek V4 models.
- 1 root planner (
deepseek-v4-pro) → 2–16 squad leads (deepseek-v4-pro) → 4–16 workers each (deepseek-v4-flash, default fan-out 10×10 = 100 workers) - Five BSP phases (partition → planning → execution → synthesis → integration), each closed by a barrier with a straggler deadline — no agent reads another agent's mid-phase output
- Write isolation: one git worktree per squad + a tiered region-ownership map enforced by the sandbox policy
- Guardrail permissions: in-grant actions auto-approve; out-of-grant escalate to the root planner; dangerous/budget actions escalate to you
- Budget discipline: per-tier pricing (pro $1.32/M in, flash $0.0826/M in), 90% pre-escalation, hard mission cap
- Cache-stable prefixes: byte-identical worker prefixes per squad ride one DeepSeek prefix-cache entry — a 100-worker mission lands at roughly $10–30
- Checkpoint/resume at every barrier
Install
dsh plugin add github:Makoveli89/dsh-swarmThis runs pnpm in your DSH profile and reconciles the bundle layer: the five swarm rows (engine, context, worktree, guard, tool) join your profile's composition. It needs a profile whose base bundles provide the subagent spawn backend, the LLM route, and the token meter (the default DSH profile does).
Mount the mode
Copy the preset into your agent-presets directory:
mkdir -p "${DSH_HOME:-$HOME/.dsh}/.agent-presets/swarm"
cp preset.yml agent.cordis.yml "${DSH_HOME:-$HOME/.dsh}/.agent-presets/swarm/"Then select the DeepSeek Coding Swarm preset for a session (or mount it as your standing preset). The session agent gains the swarm tool.
Use
swarm launch — repo: /absolute/path/to/repo, brief: "implement the TODOs in src/queue"
swarm status — missionId: <id>
swarm pause | resume | cancel | continue — missionId: <id>
swarm answer — escalationId: <id>, decision: <allow | reject | raise-cap>Missions run asynchronously: after launch, monitor with swarm status and react to swarm/* events. Defaults: 10×10 fan-out, $30 hard budget cap, per- phase deadlines 5m/5m/20m/5m/5m — all overridable per launch or in the row config.
Packages
| Package | Role | ctx key |
|---|---|---|
@deepseek-ai/dsh-swarm | Service definition + engine (BSP barriers, region ownership, budget, checkpoints, swarm/* events) | ctx.swarmEngine |
@deepseek-ai/dsh-swarm-context | Cache-stable prefix compositor + role personas | ctx.swarmContext |
@deepseek-ai/dsh-swarm-worktree | Per-squad git worktree leases | ctx.swarmWorktree |
@deepseek-ai/dsh-swarm-guard | Capability grants + escalation verdicts | ctx.swarmGuard |
@deepseek-ai/dsh-tool-swarm | Model-facing mission tool | registers on ctx.tools |
Verification
Developed in the harness monorepo (PR: deepseek-ai/deepseek-harness): 117 tests, 100% per-file coverage, host typecheck and oxlint clean, and a 100-concurrent-worker load run over the barrier. Live-mission acceptance (real-model cache-hit ratio and spend) is documented in the subsystem doc there.
License
MIT. The cache-stable-prefix doctrine and planner/executor tiering are adopted from DeepSeek-Reasonix (MIT); no code was imported.