包工头

Foreman · dsh-agent-foreman

Parent agent dispatches, sub-agents do the work, the parent reviews.

No preset roles, no scripted phases: you decide who works and with which model.

Chinese docs: README.zh.md

The problem

DSH's built-in Agent Team lets you give each teammate a name and a prompt — but not a model. Real work is mixed: some of it needs an expensive model, some of it is a two-line edit. Inheriting the parent's model for every pair of hands is both slow and expensive.

Foreman splits it in two:

• Contractors — the crew you define in the settings page, named by trade like a real

construction site: 力工 (laborer — does the business work) and, optionally, 安全员 (safety officer — reviews and rejects). The foreman itself is the parent agent: it splits the work and signs off, and is never in the roster. Want specialties? Hire several instances (力工 · frontend, 力工 · render), each with its own model and its own prompt (the 👁 button next to the name), and its own concurrency cap: default 1, so one instance never works two orders at once; set it to 2 and it can. That prompt is prepended to every order the contractor receives. The minimal setup is a single 力工.

• Orders — the work the parent agent splits out. Each order carries a task and acceptance

criteria, and may name a contractor or rotate automatically.

The parent only splits work and reviews results. Everything in between happens in isolated child sessions, and the products come back as one shared board.

Install

dsh plugin add E:/path/to/dsh-agent-foreman

Manual (dev) install:

1. Copy the whole directory to /node_modules/dsh-agent-foreman
2. Add "dsh-agent-foreman" to dsh.profile.bundles in /package.json
3. Restart DSH

Then open Settings → 包工头 and add contractors.

Usage

Type /foreman (registered at runtime through ctx.skills.register()) followed by the work: /foreman extract GraphBlocked into shared logic; add regression tests. Semicolons or newlines split orders. The skill body walks the parent through roster → split → one dispatch → review, so no standing system prompt is needed.

Tool · Purpose foreman_roster · List contractors with route, prompt length, concurrency cap, and the current policy foreman_dispatch · Dispatch a batch of orders, wait for all of them, return the board foreman_review · Pass / fail each order; a failed one is re-dispatched with feedback foreman_status · Re-read a board; omit runId for the latest one in this session

Where members show up

There is no custom in-chat panel any more: in team mode a contractor is a real Agent Teams member, so the native Agent Teams button in the session header lists them and shows each one's model. The settings page keeps the full roster editor.

Roster notes:

• A contractor is a name, an optional prompt, a provider, a model, an effort level and a concurrency cap.

Leave the provider on "follow the current session" and the child simply runs on whatever model the parent session is using.

• To use a different model: pick the provider first — the model dropdown is derived from it and

only lists that provider's models (providers with no models never show up).

• Settings page collapses policy behind "派单策略"; the roster table keeps name / provider / model /

effort / concurrency / enabled, and 👁 opens that contractor's prompt editor.

Dispatch notes:

• orders[].task must be self-contained — the child cannot see the parent's context.
• orders[].contractor rotates through the roster when omitted; unknown names fail loudly.
• orders[].dependsOn only references orders in the same batch; dependents run serially after.
• waitSeconds caps the whole batch; unfinished orders are aborted rather than silently reported done.
• concurrency defaults to 4, capped at 8.

Policy

Field · Default · Meaning concurrency · 4 · Contractors working at once (1–8); the per-instance cap lives in the roster timeoutSeconds · 900 · Wall-clock cap per order maxRework · 1 · How many times an order may be sent back maxDepth · 1 · Whether a child may dispatch further children provider · spawn · Which subagents provider to use (spawn / fork / cursor)

Data

• Windows: %LOCALAPPDATA%/dsh-agent-foreman/
• Elsewhere: /data/dsh-agent-foreman/

roster.json (contractors + policy), runs.jsonl (one line per dispatch round). A corrupt roster never blocks startup — it is treated as empty and logged.

Security

• Every HTTP route accepts same-origin, loopback requests only; cross-site calls get 403.
• The roster stores provider/model/effort names — never credentials.
• Contractors run in their own child sessions and inherit DSH's effective sandbox.

Development

node --test test/*.test.mjs   # 26 cases: tools, engine, routes, CSRF, ledger

test/harness.test.mjs runs the real host half against a fake host, so no DSH instance is needed.

Compared to

· Agent Team (built-in) · swarm-orchestrator · Foreman Per-member model · no · no · yes Preset roles · fixed · 4 fixed roles · none — you define Shared board · yes · yes · yes Rework loop · no · limited · yes, with feedback Settings page · yes · no · yes, roster + live model catalog

Team mode (experimental)

Off by default. Tick 团队模式 under Settings → 包工头 → 派单策略 and each contractor becomes a real member of the host's agent team: the plugin spawns it with agentTeams.spawnTeammate, switches its session model with sessions.selectModel, sends the work order as a team message, then reads the delivery back from the member's session events.

Members therefore show up in the team area and can change model in their own session. Every step depends on host team services, so any failure falls back to the plain activation path and the board's teamMode field says which one ran. Known trade-offs:

• Member names must be ASCII (lower-kebab-case), so Chinese contractor names all

become worker, worker-2, … — the Chinese name and model live in the member description, which the team area shows.

• A member's first turn runs on the parent's model (the model can only be switched

after the session exists — hence the one-line handshake).

• Switching is bounded by the host's per-subagent model allow-list; values outside

it are refused and the board shows model switch: failed · reason.

• Members stay in the team after the round.

License

MIT