Skip to content

README.md

Latest commit

 

History

History
362 lines (276 loc) · 18.5 KB

README.md

File metadata and controls

362 lines (276 loc) · 18.5 KB
Raw

Agent Controller

Declarative runtime for AI agents. Define agents in YAML (ADL — Agent Definition Language). Run the same spec on the Pi, opencode, or codex runtime through a consistent backend interface.

Agent Controller separates agent intent (model, persona, tools, skills, MCP servers, guardrails, observability) from execution substrate. A spec that sticks to the cross-adapter feature set (persona, skills, MCP servers, guardrails) runs on the Pi, opencode, or codex adapter by changing one field — runtime.type. Specs that use Pi-only capabilities (custom Pi-extension tools like get_time, spec.extensions[], subagents) are Pi-only — opencode and codex reject them at startup. The codex adapter additionally requires model.provider: openai (anthropic and google are rejected). Remote backends, like the Kubernetes target (a schema-level skeleton today), attach via a RuntimeBinding. Full capability matrix.

agentctl run  examples/hello.yaml           # Pi adapter
agentctl run  examples/hello-opencode.yaml  # opencode adapter (needs the `opencode` CLI on PATH)
agentctl run  examples/hello-codex.yaml     # codex adapter (needs the `codex` CLI on PATH + OPENAI_API_KEY)
agentctl chat examples/hello.yaml           # interactive REPL with session persistence

Architecture

Agent Controller architecture — agentctl dispatches a compiled spec over a stdio NDJSON wire protocol to the Pi, opencode, or codex runtime adapter, which loads the local registry and calls the LLM provider

agentctl (a Go binary) compiles the ADL spec and dispatches it over a versioned stdio NDJSON wire protocol to a runtime adapter (Pi, opencode, or codex). The adapter loads the local registry — tools, extensions, skills, agents, MCP servers — and drives the session against the model provider. Backends beyond Local (Kubernetes skeleton, AgentCore) are reserved in the schema but not fully wired. See docs/architecture/overview.md for the full layer breakdown and wire protocol reference.

Quick start

Recommended (pre-built binary + npm)

# 1. Download agentctl from the latest GitHub release and put it on PATH
#    or: go install github.com/CCDevelopForFun/agent-controller/cli/cmd/agentctl@v0.7.0

# 2. Install the runtime adapter you need
npm install -g @agent-controller/runtime              # Pi  (runtime.type: local)
npm install -g @agent-controller/runtime-opencode     # opencode (runtime.type: local-opencode)
npm install -g @agent-controller/runtime-codex        # codex (runtime.type: local-codex)
#   the opencode adapter also needs the separate `opencode` CLI on PATH:  npm install -g opencode-ai
#   the codex adapter also needs the `codex` CLI on PATH and OPENAI_API_KEY set in the environment

# 3. Write a self-contained spec (no registry refs, so it runs from any dir)
cat > /tmp/hello.yaml <<'EOF'
apiVersion: agent-controller.dev/v1alpha1
kind: Agent
metadata: { name: hello }
spec:
  model: { provider: anthropic, name: claude-sonnet-4-20250514 }
  persona: { role: Helpful demo, instructions: Answer concisely. }
  task: Say hello.
  tools: []
  runtime: { type: local }
EOF

# 4. Set your API key and run
export ANTHROPIC_API_KEY=sk-ant-...
AGENT_CONTROLLER_RUNTIME="$(npm root -g)/@agent-controller/runtime/dist/index.js" \
  agentctl run /tmp/hello.yaml

Source clone

(cd runtime && npm install --ignore-scripts && npm run build)
(cd runtime-opencode && npm install --ignore-scripts && npm run build)
(cd cli && go build -o bin/agentctl ./cmd/agentctl)

export ANTHROPIC_API_KEY=sk-ant-...
./cli/bin/agentctl run examples/hello.yaml

Run from the repo root. The adapter path and manifest registry (tools/, extensions/, skills/, agents/) resolve relative to cwd. Set AGENT_CONTROLLER_RUNTIME to override the adapter path.

Commands

agentctl validate  spec.yaml                          # check ADL against JSON Schema
agentctl compile   spec.yaml                          # print resolved CompiledSpec JSON
agentctl run       spec.yaml                          # run agent, stream NDJSON events
agentctl run       spec.yaml --task "override"        # override spec.task at runtime
agentctl run       spec.yaml --resume <session-id>    # resume a prior session (Pi + opencode + codex)
agentctl run       spec.yaml --binding binding.yaml   # resolve against a RuntimeBinding
agentctl chat      spec.yaml                          # interactive REPL (v0.6)
agentctl sessions  list                               # list persisted sessions
agentctl sessions  sweep --ttl 168h                   # retire idle sessions
agentctl install   npm:<pkg>                          # install a Pi package/extension (shells to `pi install`)

# v0.7 — scheduler-task surface (parameterize input, capture output, share memory)
agentctl run       spec.yaml --input text="hello"     # ${inputs.text} interpolation in spec.task
agentctl run       spec.yaml --input text=@prev.txt   # read an input value from a file (handoff)
agentctl run       spec.yaml --output-file out.json   # capture result; validated by spec.outputSchema
agentctl run       spec.yaml --workspace ./run-42     # durable memory shared across steps

Serving agents over HTTP

agentctl serve (v0.8) wraps any ADL agent spec in a long-lived HTTP/SSE server so external clients — schedulers, UIs, integration tests — can open sessions and exchange turns without spawning a new process per request.

agentctl serve examples/hello.yaml --port 8080

Flags

Flag Default Description
--port 8080 TCP port to listen on
--in-memory off Use an in-memory session store instead of SQLite
--max-concurrent-turns 8 Max in-flight turns before returning 429
--max-sessions 1000 Max active sessions before create returns 429
--session-ttl 168h Idle session TTL swept in the background
--shutdown-grace 25s Max time to drain in-flight turns on SIGTERM

Endpoints

Method Path Description
GET /healthz Liveness probe — always 200
GET /readyz Readiness probe — 200 until draining, then 503
POST /v1/sessions Create a session; body is optional — omit entirely, send {}, or send {"inputs":{...}} for interpolation values; returns 201 {id,agentName,runtimeType,status,createdAt}
GET /v1/sessions List sessions (all statuses); add ?status=active to filter to active only
GET /v1/sessions/{id} Get a single session
DELETE /v1/sessions/{id} Delete a session (204)
POST /v1/sessions/{id}/turns Run a turn; body {"input":"..."} ; streams text/event-stream SSE

Error codes: 409 (session busy), 429 (global turn cap or max-sessions), 503 (draining).

SSE frame format: each frame is event: <type>\ndata: <json>\n\n; the stream ends on a session.ended frame.

Example

# 1. Start the server
agentctl serve examples/hello.yaml --port 8080 &

# 2. Create a session
SESSION=$(curl -s -X POST http://localhost:8080/v1/sessions \
  -H 'Content-Type: application/json' \
  -d '{}' | python3 -c "import sys,json; print(json.load(sys.stdin)['id'])")

# 3. Run a turn and stream SSE events
curl -N -X POST "http://localhost:8080/v1/sessions/${SESSION}/turns" \
  -H 'Content-Type: application/json' \
  -d '{"input":"Hello!"}' \
  --no-buffer

A complete runnable example is at examples/serve-client.sh.

Metrics and auth (OIDC / API-key middleware) are deferred to v0.8.x. TLS is expected to be terminated by a sidecar or load-balancer proxy.

What ADL can declare

Field Since What it is
model v0.0.1 Provider + model name (Anthropic / OpenAI / Google) and optional temperature
persona v0.0.1 role + instructions — prepended to the system prompt
task v0.0.1 Initial user prompt that drives the session — supports ${inputs.<key>} interpolation from --input (v0.7)
outputSchema v0.7 JSON Schema — with --output-file, the agent's reply is parsed as JSON, validated, and written
tools v0.0.1 Tool allowlist — local registry names or Pi built-ins (bash, read, edit, write)
extensions v0.0.1 Pi extension allowlist — local registry or source: npm:<pkg> for auto-install
skills v0.1.2 Markdown skill files — bodies inlined into the system prompt at session start
subagents v0.1.3 Child agents the parent can delegate to
mcpServers v0.1.5 MCP servers (stdio / streamable-http / SSE)
guardrails v0.1.8 Hallucination detector mode (block / warn / correct)
observability.tracing v0.5 Emit OTel spans to an OTLP endpoint
runtime v0.0.1 type (local-pi / local-opencode / local-codex; local is the legacy Pi alias) + optional requirements for capability matching

The full schema is at schemas/adl.v1alpha1.json. Unknown fields are rejected at compile time.

Skills

Skill bodies are inlined into the system prompt at session start — no lazy loading. Declare them by name; the runtime reads skills/<name>/SKILL.md and appends the body.

spec:
  skills:
    - name: example-time-skill   # formats all timestamps as ISO-8601 UTC
    - name: using-superpowers    # teaches the agent how to find and use skills
  tools:
    - name: bash                 # required if a skill prescribes shell commands

Vendoring an external skill is a file copy: cp SKILL.md skills/<slug>/SKILL.md.

See examples/claude-skills-with-bash.yaml and examples/claude-skills-demo.yaml.

MCP servers

MCP-registered tools surface to the model as mcp_<server>_<tool>. The runtime writes .pi/mcp.json at session start and loads the MCP extension automatically.

spec:
  mcpServers:
    - name: time-server
      transport: stdio
      command: npx
      args: ["-y", "@modelcontextprotocol/server-time"]
      lifecycle: eager   # connect at session start; default is lazy

Supported transports: stdio, streamable-http, sse. See examples/mcp-time.yaml and examples/self-contained-mcp.yaml.

Chat & sessions

agentctl chat opens an interactive REPL on any adapter (runtime.type: local / local-pi / local-opencode / local-codex). Sessions persist across process restarts via SQLite; pick up where you left off with --resume.

agentctl chat examples/hello.yaml               # start a session
agentctl chat examples/hello.yaml --resume <id> # resume a specific session
agentctl sessions sweep --ttl 168h              # expire sessions idle longer than 7d

Each session keeps full turn history, so a resumed chat continues exactly where it left off. The SQLite store lives at $XDG_DATA_HOME/agent-controller/sessions.db (default ~/.local/share/agent-controller/sessions.db).

OTel tracing

agentctl run emits one agentctl.run root span; agentctl chat emits one agentctl.run root span for the whole REPL with a chat.turn span per turn. Adapter spans (LLM calls, tool calls) nest underneath. Point the exporter at any OTLP-compatible backend:

OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 \
  agentctl run examples/tracing-demo.yaml

Enable tracing in the spec:

spec:
  observability:
    tracing: true

Works with Jaeger, Grafana Tempo, Honeycomb, BrainTrust, and any OTLP receiver. See examples/tracing-demo.yaml.

Guardrails

Three layers defend against the model writing fake tool-call XML in its response:

  1. Honesty preamble — system-prompt block telling the model real tool calls go through the runtime's tool channel.
  2. Skill body framing — each inlined skill body is wrapped with a runtime override.
  3. XML detector — scans every assistant message; behavior is configurable.
spec:
  guardrails:
    hallucinationDetector: warn   # block (default) | warn | correct
  • block — emits an error event and exits non-zero.
  • warn — scrubs the fabricated XML and lets the session complete.
  • correct — same as warn, plus one corrective re-prompt.

See examples/guardrails-block.yaml, guardrails-warn.yaml, guardrails-correct.yaml.

RuntimeBinding

RuntimeBinding maps an agent spec to a concrete execution target. At run time the CLI matches the spec's runtime.requirements against the target's advertised capabilities — unmet requirements emit a warning and the run proceeds, unless the binding sets target.strict: true, which promotes them to a hard error before the session starts.

agentctl run spec.yaml --binding examples/bindings/local-default.yaml
apiVersion: agent-controller.dev/v1alpha1
kind: RuntimeBinding
metadata: { name: local-default }
spec:
  selector:
    runtimeType: local-pi
    capabilities: { streaming: true, sandbox: false }
  target:
    type: local

See examples/bindings/ for ready-to-use specs. Full capability table: docs/architecture/harness-matrix.md.

Scheduler tasks (DAGs of agents)

Since v0.7, agentctl run is a per-step task primitive: an external scheduler (Maestro, Airflow, Temporal) owns the DAG, and agent-controller runs one agent per step. There's no workflow YAML and no in-process engine — just the three things a step needs to compose.

agentctl run classifier.yaml --input text="great product"        # 1. parameterize:  ${inputs.text} → spec.task
agentctl run classifier.yaml --input text=@/shared/prev.json     #    file handoff:   value read from a prior step
agentctl run classifier.yaml --output-file /shared/result.json   # 2. capture:       result written on a clean run
agentctl run classifier.yaml --workspace ./run-42                # 3. share memory:  durable state across steps

spec.outputSchema validates the reply as JSON before it's written, and --workspace exposes memory tools over MCP — so the same flags work on both adapters. Copy-paste configs for Maestro, Airflow, and Temporal live in examples/schedulers/.

Examples

File What it demos
examples/hello.yaml Basic agent — get_time tool, audit-log extension, example-time-skill
examples/hello-opencode.yaml Same shape on the opencode adapter (runtime.type: local-opencode)
examples/hello-codex.yaml Minimal agent on the codex adapter (runtime.type: local-codex, model.provider: openai)
examples/tracing-demo.yaml OTel tracing — run with OTEL_EXPORTER_OTLP_ENDPOINT=...
examples/mcp-time.yaml MCP via @modelcontextprotocol/server-time
examples/self-contained-mcp.yaml Same MCP agent with extensions[].source auto-install
examples/claude-skills-with-bash.yaml A skill + bash tool (agent runs a real shell command)
examples/claude-skills-demo.yaml External skills inlined into the prompt (no tools)
examples/subagent-demo.yaml Parent agent delegates to a sql-explorer child
examples/bash-allowlist.yaml Bash command allowlist via @gotgenes/pi-permission-system
examples/guardrails-{block,warn,correct}.yaml Hallucination detector modes side-by-side
examples/bindings/ RuntimeBinding specs — local Pi (local-default, local-strict) + Kubernetes (kubernetes-kind)
examples/schedulers/ agentctl run as a Maestro / Airflow / Temporal task — shared text-classifier.yaml + per-scheduler config

Repo layout

Directory Purpose
cli/ Go binary agentctl. See cli/README.md.
runtime/ Pi adapter (Node). See runtime/README.md.
runtime-opencode/ opencode adapter (Node). See runtime-opencode/README.md.
runtime-codex/ codex adapter (Node). Requires codex CLI on PATH + OPENAI_API_KEY. See runtime-codex/README.md.
schemas/ JSON Schemas (ADL v1alpha1 + manifest v1). See schemas/README.md.
tools/ Built-in tools (get_time)
extensions/ Built-in Pi extensions (audit-log, subagent)
skills/ Built-in / vendored skills (example-time-skill, using-superpowers)
agents/ Subagent persona files (sql-explorer.md)
examples/ Example ADL files
e2e/ End-to-end test harness — supports ADAPTER=pi|opencode|codex
docs/architecture/ Architecture diagrams and reference

Status

v0.7.0 — developer preview.

Tag What landed
v0.1.x ADL + Pi adapter + full original feature surface (tools, extensions, skills, subagents, MCP, guardrails)
v0.2.0 opencode adapter, multi-adapter abstraction, harness capability matrix
v0.3.x RuntimeBinding, capability matcher, self-contained npm install
v0.4.0 First remote backend — Kubernetes Pod skeleton
v0.5.x End-to-end OTel tracing (root span → adapter spans → tool spans)
v0.6.0 agentctl chat REPL, SQLite session store, per-turn OTel spans (folded into v0.7.0 — never tagged standalone)
v0.7.0 agentctl run as a scheduler task — --input/${inputs}, --output-file/outputSchema, file handoff, --workspace durable memory
v0.1.x release history
Tag What landed
v0.0.1-mvp ADL + local runner + Pi adapter + get_time + audit-log
v0.1.0 agentctl install
v0.1.1 Session resumption (--resume, sessions list)
v0.1.2 Skill kind
v0.1.3 Subagent kind
v0.1.4 install extension <name> alias
v0.1.5 MCPServer kind
v0.1.6 Self-contained extensions (spec.extensions[].source)
v0.1.7 Skill bodies active by default
v0.1.8 Hallucination guardrails
v0.1.9 Pi built-in tools (bash/read/edit/write) in ADL
v0.1.10 spec.guardrails.hallucinationDetector modes (warn/block/correct)

For next steps, see ROADMAP.md. For the full per-feature capability table, see docs/architecture/harness-matrix.md.

Contributing

See CONTRIBUTING.md for the dev setup and slice-by-slice workflow. Security issues: SECURITY.md.

License

MIT.