dsh-team-v3

dsh-team-v3

Team presets for DeepSeek Harness AgentTeams: three ready rosters (audit6, impl5, lite2) with seed task DAGs and per-role prompts, plus a captain doctrine section — so a fresh session opens the right team without pasting a prompt file.

This package configures the profiles of @nanmicoder/dsh-agent-teams (the engine that ships the agent_teams_* tools, the /agent-teams command and the Web approval panel). The stock @deepseek-ai/dsh-experimental-agent-team has no template mechanism, so presets belong here.

Install

dsh plugin --profile <profile> add @nanmicoder/dsh-agent-teams dsh-team-v3
# or: bash install.sh <profile>   /   .\install.ps1 -Profile <profile>

Restart Harness once (profile config is composed at boot), then:

agent_teams_create({ name: "audit-x", description: "<goal>", profile: "audit6", approval: "required" })

Rosters

profile	members	seed DAG	wall clock
audit6	primary-a, primary-b, context-scout, adversary, reproducer, reviewer	a ∥ b ∥ c ∥ d → e(a,b) → f(a,b,c,d)	25 min
impl5	architect, implementer, breaker, verifier, reviewer	arch → build ∥ break → verify → review	50 min
lite2	implementer, breaker	captain designs the DAG	20 min

Every member carries the report format (≤6 lines: scoped conclusion / ≤5 evidence items / absolute artifact path / "cannot confirm" field) plus a role patch; every seed task carries criteria, prohibitions, artifact path (evidence/<member>/report.md, ≤80 lines) and a stop rule.

Config

id: team-v3 in cordis.patch.yml: enabled, order (default 118), profiles, wallClockAudit / wallClockImpl / wallClockLite. Edit profiles.yml, then node tools/build.mjs.

Verify

node tools/verify.mjs   # parse → apply like dsh-app-boot → both schemas → render the injected section

Limits

Seed tasks are work kind: the engine's task schema has no kind/acceptance/verify, so quality gates are upgraded by the captain after approval. A team is only ever created as a staged plan awaiting approval — zero-click auto-spawn has no public API and is not worth its cost.

MIT