Jesse-njx/dsh-docker

dsh-docker

Typed, guarded container control for DSH — structured Docker access that is hard to destroy by accident.

dsh-docker is a plugin bundle for DeepSeek Harness. It wraps the docker CLI through DSH's ctx.shell seam and gives your agent a typed, project-aware, guarded surface for containers and compose stacks — every tool returns structured JSON (await tools.docker_ps(...) gets real objects, never scraped prose), destructive operations require human sign-off through the approval gate, and stop dev-api means your local compose service, not a coincidentally-named container on the machine.

┌──────────────┐   docker ps --format json    ┌──────────────┐
│  agent tool  │ ───────────────────────────▶ │  ctx.shell   │ ──▶ docker CLI
│  call        │                              │  (sandbox/   │
└──────┬───────┘                              │   remote)    │
       │ classify()                           └──────────────┘
       ▼
┌──────────────┐   destructive?  ┌──────────────────────┐
│ pre-execute  │ ──────────────▶ │ approval gate         │  allowed-once → run
│ policy       │  (guarded)      │ (ctx.approval,        │  rejected/unavailable
└──────┬───────┘                 │  fails closed)        │  → refused
       │ safe / token recorded   └──────────────────────┘
       ▼
┌──────────────┐   hard-destructive without a token? → refused (monotonic)
│ ctx.tools    │
│ .guard()     │   backstop: no later listener can undo a denial
└──────────────┘

Flagship demo — debug a failing integration test

Your agent runs a failing test against docker-compose.yml:

$ docker_ps            → { containers: [{ name: "app-api-1", state: "running", ... }] }
$ docker_logs { container: "app-api-1", tail: 200 }
  → { lines: [...], truncated: true }        # capped pull never reads as complete
$ docker_compose_ps   → { project: "app", services: [{ name: "db", state: "running" }] }

The agent reads freely. The moment it tries something destructive — docker_rm -f on the running API container, docker_rmi on an image a container still uses, docker system prune -a, or docker compose down -v — the call is paused and routed to the human approval gate:

⏸ docker_rmi "app-api:latest" — image in use by container app-api-1
   [approve]  [reject]

No approval channel, or a denied request? The operation is refused. execReadOnly is on by default, so docker_exec without write/interactive never prompts, and a writeful exec is treated as destructive. And even if some other plugin's policy said "allow", the monotonic guard still refuses the hard-destructive set unless this call carries the dsh-docker approval token.

Install

# From the npm registry once published, or straight from this repository:
dsh plugin --profile web add @dsh-docker/bundle          # npm (when published)
dsh plugin --profile web add github:Jesse-njx/dsh-docker  # or: straight from GitHub

Install into whatever profile your agents run under (web for the desktop UI, headless for CLI sessions). The bundle mounts the tools, the policy, the opt-in health context, and the web status renderer.

Tools

Every tool is a defineTool with typed parameters and a structured output schema — Code Mode sees await tools.<name>(...) returning the canonical JSON value, never rendered prose.

Containers

Images

Compose

Project-aware targeting

Every tool resolves the compose project once per call (a pure, unit-tested ProjectResolver):

1. Walk up from the tool's working directory for the first composeFiles match (docker-compose.yml, compose.yaml, config-ordered).
2. Derive the project name (directory basename, lowercased) and pass -p <project> -f <file> to every compose call; docker_ps/start/stop default filters scope to --filter label=com.docker.compose.project=<project>.
3. Every tool accepts an explicit project: string override; with no compose file and no override, container-scoped tools operate globally but the project-aware behaviors are off.

A bare container name resolves within the project first, falling back to a global match only when unambiguous. Ambiguous refs return an error listing candidates instead of guessing.

Guardrails

Two layers, both in the tools pipeline:

• tools/pre-execute policy — classify(toolName, args, resolved) (a pure function, unit-tested without a daemon) judges each call's intended effect. Guarded ops route a one-shot approval request through ctx.approval (approval/request waterfall); approval absent fails closed to unavailable ⇒ refused. A guarded op not covered by the configured approval globs is refused outright.
• ctx.tools.guard() backstop — a monotonic deny that a later listener cannot undo. Even if another plugin's pre-execute said "allow", the guard refuses the hard-destructive set unless this call carries the dsh-docker approval token.

Read-only docker_exec is free by default; execReadOnly: false flips that for operators who accept the risk.

Service-health context (opt-in)

plugins:
  dsh-docker:
    healthContext: { enabled: true, maxServices: 12 }

When enabled, each pre-step injects one compact line per detected compose service — dev-api ▲up dev-db ▲up worker ▼exited(1) — as durable appended context the next model request sees (not a wake-up). Built from a single docker_compose_ps read, TTL-cached, capped at maxServices, and skipped entirely when no compose project is detected.

Status renderer (web client)

The client half registers a replayable conversation node for docker_ps / docker_compose_ps / docker_logs output: one stable node per tool call (keyed by callId), a table collapsed by default and expandable on click, with a log pane whose truncated flag is surfaced so a capped pull never reads as complete. Replay-safe: the renderer consumes only the durable tool/result presentation metadata, never the execution-local value.

Config

plugins:
  dsh-docker:
    enabled: true
    composeFiles: [docker-compose.yml, compose.yaml]
    approval: ["rmi:*", "rm:*", "prune:*", "compose down -v", "exec:*"]
    execReadOnly: true
    timeoutMs: 30000
    healthContext: { enabled: false, maxServices: 12 }

The same schema registers as the dsh-docker user-settings section, so a live edit reaches the very next tool call.

Non-goals (v0.1)

Docker contexts / remote-host management UI; building images from a Dockerfile with any DSL (raw docker build passthrough is fine — no wrapping); swarm / k8s; container runtime introspection beyond docker inspect; log streaming / follow; registry auth flows. Each is a v0.2+ question, not a config flag away.

Testing

• Unit (no daemon): arg-schema acceptance/rejection per tool; ProjectResolver fixtures (compose-file discovery, project-name derivation, ambiguous-ref handling); the classify() policy matrix — every destructive op asserts guarded, every safe op asserts safe, the execReadOnly toggle flips docker_exec; the approval-glob matcher.
• Guard (no daemon): a mock approval gate that always denies — rmi in-use, rm running, prune --all, compose down -v, and writeful exec are refused; an approving mock — they proceed exactly once; a preempting "allow" from another listener is refused by the backstop.
• Integration (feature-detected): probes for a live daemon (docker version) and, when present, compose up/ps/down a tiny alpine sleep stack with a mock approval gate — typed rows match reality and down -v requires approval. Skips cleanly (marked skipped, not failed) when no daemon is available.

pnpm install
pnpm typecheck && pnpm build && pnpm test

Model Experience

Request context and condition

What the model sees

The 14 tool schemas (names, descriptions, parameters) flow into system-prompt assembly through the tool registry, exactly like every other registered tool; there is no additional fixed prompt prose owned by this package. When healthContext.enabled is on and a compose project is detected in the agent's working directory, each pre-step injects a source: { kind: 'plugin', plugin: 'dsh-docker' } user message:

[dsh-docker] <project> services: dev-api ▲up  dev-db ▲up  worker ▼exited(1)

Token effect

The tool schemas are fixed prompt tokens owned by the registry. The health line is conditional: bounded to maxServices lines plus a +N more tail, present only while the feature is enabled and a compose project is detected.

KV Cache effect

Append-only in the sense that each step appends a fresh health message that the next request consumes and discards; the message text is data-dependent (service states change), so a stable prefix is preserved but the tail invalidates reuse whenever a service state changes. Package-owned changes that can invalidate reuse: none at runtime — a reinstall/upgrade of the bundle changes the schemas, which is a normal dependency-change invalidation.

Known Limitations and Deferred Work

• docker_ps/docker_compose_ps need the docker CLI to support --format json (Docker ≥ 25 for ps, compose v2 for compose ps). Older CLIs surface a clear error rather than scraped text; the docker inspect batch is the adapter where the ps Labels field is absent.
• Failed tool calls render no status table — presentationMeta is computed from successful canonical values only, so a failed docker_ps shows the standard error card, not a table. The durable log carries no canonical value by design.
• docker_prune cannot report what the engine reclaimed without scraping human prose; the canonical output carries the request and success, and the engine's reclaimed-space line lives in the rendered text only.
• docker_exec --interactive output may carry TTY framing (CRLF/ANSI); the canonical value returns it verbatim rather than laundering it.