ltcai 1.6.0 → 2.0.0

package/docs/MULTI_AGENT_RUNTIME.md

@@ -0,0 +1,410 @@
+# Lattice AI Multi-Agent Runtime 2.0
+
+The Multi-Agent Runtime is the **orchestration layer** introduced in v2.0.0. It sits
+*above* the v1.x single-agent state machine ([`AgentRuntime`](../latticeai/core/agent.py))
+and coordinates a pipeline of named **roles** that hand off work to one another,
+retry on a failing review, and emit a fully observable timeline.
+
+- **Source of truth:** `latticeai/core/multi_agent.py`
+- **HTTP surface:** `latticeai/api/agents.py`
+- **Persistence / Knowledge Graph integration:** `latticeai/core/workspace_os.py`
+  (`WorkspaceOSStore.record_agent_run`)
+
+```python
+MULTI_AGENT_VERSION = "2.0.0"
+```
+
+## How it relates to the v1 single-agent runtime
+
+v1.x shipped a single-agent state machine — `AgentRuntime` driving
+`PLAN → EXECUTE → VERIFY → DONE` (with `ROLLBACK` / `FAILED` recovery paths) over an
+injected `AgentDeps` port. That runtime is unchanged.
+
+v2.0 adds the *orchestration* layer on top: instead of one agent looping through
+internal phases, the `MultiAgentOrchestrator` drives a **pipeline of distinct roles**
+(researcher, planner, executor, reviewer, release) that hand off to one another and
+can rewind on a failing review. The two layers are complementary — the v2
+orchestrator coordinates roles; an individual role's runner could itself be backed by
+the v1 `AgentRuntime` loop or an LLM, but that is an implementation choice behind the
+injected runner port.
+
+> **Compatibility.** The Multi-Agent Runtime is purely additive. The v1
+> `AgentRuntime` state machine and its `/agent` endpoints are untouched, and the v2
+> API is namespaced under `/agents` (plural) so it never collides with the existing
+> single-agent `/agent` routes. Existing v1.x data is preserved; new runs are appended
+> to workspace state and the Knowledge Graph alongside it.
+
+## Built-in roles (`AGENT_ROLES`)
+
+The runtime defines five built-in roles. Each role id matches an entry in
+`latticeai.core.workspace_os.DEFAULT_AGENTS`, so orchestrated runs reference the same
+agents that already appear in the Workspace.
+
+```python
+AGENT_ROLES = ("researcher", "planner", "executor", "reviewer", "release")
+
+ROLE_AGENT_IDS = {
+    "researcher": "agent:researcher",
+    "planner":    "agent:planner",
+    "executor":   "agent:executor",
+    "reviewer":   "agent:reviewer",
+    "release":    "agent:release",
+}
+```
+
+| Role | Agent id | Responsibility |
+| --- | --- | --- |
+| `researcher` | `agent:researcher` | Gathers relevant context (workspace memory / graph) via an injected `context_provider`. |
+| `planner` | `agent:planner` | Decomposes the goal into ordered, inspectable steps. |
+| `executor` | `agent:executor` | Carries out steps; may drive an injected workflow / plugin runner. |
+| `reviewer` | `agent:reviewer` | Judges the result and returns a `pass` / `retry` verdict. |
+| `release` | `agent:release` | Finalizes / packages the outcome (optional). |
+
+The `agent:*` ids correspond one-to-one with `DEFAULT_AGENTS` in `workspace_os`
+(`agent:planner`, `agent:executor`, `agent:reviewer`, `agent:researcher`,
+`agent:release`), which is why an orchestrated run can record relationships against
+agents that the Workspace already knows about.
+
+### The core pipeline
+
+`researcher` and `release` are optional stages — they run only when explicitly
+requested. A quick, default run is therefore three stages:
+
+```python
+CORE_PIPELINE = ("planner", "executor", "reviewer")
+```
+
+When `MultiAgentOrchestrator.run(...)` is called without an explicit `roles` list, it
+uses `CORE_PIPELINE`. Any roles passed in are filtered to the set of known
+`AGENT_ROLES`; if nothing valid remains, it falls back to `CORE_PIPELINE`.
+
+## Orchestration: `MultiAgentOrchestrator.run`
+
+```python
+def run(
+    self,
+    goal: str,
+    *,
+    user_email: Optional[str] = None,
+    workspace_id: Optional[str] = None,
+    inputs: Optional[Dict[str, Any]] = None,
+    roles: Optional[List[str]] = None,
+    max_retries: int = 2,
+) -> AgentRunResult:
+    ...
+```
+
+`run` walks the resolved pipeline, threading a single `OrchestrationContext` through
+every stage. As it goes it appends two kinds of events to an observable timeline:
+
+- **`role` events** — emitted by `_run_role` for each stage, recording the role, its
+  `agent_id`, status (`ok` / `error`), the raw runner result, and start/end timestamps.
+- **`handoff` events** — emitted via `OrchestrationContext.handoff(frm, to, note)`
+  whenever control passes from one role to the next (and on a retry rewind).
+
+The timeline is also bracketed by a `start` event (carrying the goal and resolved
+pipeline) and an `end` event (carrying the final status and retry count).
+
+### Retry: reviewer rewinds to executor
+
+The pipeline is ordinarily linear, but the **reviewer can rewind** the pipeline to the
+executor. After the reviewer runs, if its verdict is `retry` and the retry budget is
+not yet exhausted, the orchestrator:
+
+1. increments `ctx.retries`,
+2. emits a `handoff("reviewer", "executor", note="retry #N: <reason>")`,
+3. resets the pipeline index back to the executor stage, and
+4. re-runs executor → reviewer.
+
+This repeats until the verdict is `pass` or `ctx.retries` reaches `max_retries`.
+
+```text
+planner → executor → reviewer ──pass──▶ (continue / end)
+                        │
+                        └──retry (and retries < max_retries)──▶ executor (rewind)
+```
+
+### Final status
+
+The terminal status is derived from the final reviewer verdict and retry count:
+
+| Condition | `status` |
+| --- | --- |
+| Final verdict `pass`, no retries | `ok` |
+| Final verdict `pass`, after one or more retries | `retried_ok` |
+| Final verdict not `pass` (retries exhausted) | `failed` |
+
+### `OrchestrationContext`
+
+The mutable carrier threaded through every stage:
+
+```python
+@dataclass
+class OrchestrationContext:
+    goal: str
+    user_email: Optional[str] = None
+    workspace_id: Optional[str] = None
+    inputs: Dict[str, Any] = field(default_factory=dict)
+    plan: List[Dict[str, Any]] = field(default_factory=list)
+    research: List[str] = field(default_factory=list)
+    executed: List[Dict[str, Any]] = field(default_factory=list)
+    review: Dict[str, Any] = field(default_factory=dict)
+    timeline: List[Dict[str, Any]] = field(default_factory=list)
+    retries: int = 0
+    output: str = ""
+
+    def handoff(self, frm: str, to: str, note: str = "") -> None: ...
+```
+
+### `AgentRunResult`
+
+`run` returns an `AgentRunResult`, which exposes `as_dict()` for serialization:
+
+```python
+@dataclass
+class AgentRunResult:
+    agent_id: str
+    status: str  # ok | failed | retried_ok
+    output: str
+    timeline: List[Dict[str, Any]]
+    plan: List[Dict[str, Any]]
+    review: Dict[str, Any]
+    roles_run: List[str]
+    retries: int = 0
+```
+
+## The role runner is an injected port
+
+Like the v1 runtime, the orchestrator is **pure logic over an injected `role_runner`
+port**. It runs with no LLM and no server:
+
+```python
+class MultiAgentOrchestrator:
+    def __init__(
+        self,
+        role_runner: Optional[Callable[[str, OrchestrationContext], Dict[str, Any]]] = None,
+    ):
+        self.role_runner = role_runner or default_role_runner()
+```
+
+The runner is a single callable `(role: str, ctx: OrchestrationContext) -> Dict[str, Any]`.
+The orchestration logic — pipeline walking, handoffs, retry rewind, timeline emission,
+status derivation — does not depend on *how* a role does its work. This means a
+**production deployment can swap in an LLM-backed runner without touching the
+orchestration layer**: implement the same callable signature, pass it to the
+constructor, and the pipeline behaves identically while individual roles gain
+model-backed reasoning.
+
+If a runner raises, `_run_role` captures the exception into the `role` event as
+`status: "error"` with an `{"error": ...}` result, rather than crashing the run.
+
+## The default runner is deterministic and useful
+
+`default_role_runner` builds a **dependency-free, deterministic** runner that
+implements every built-in role with real (non-LLM) behavior. This is what makes
+"agent runs can execute workflows / plugins" true in the community edition without
+requiring a model.
+
+```python
+def default_role_runner(
+    *,
+    workflow_runner: Optional[Callable[..., Any]] = None,
+    plugin_runner: Optional[Callable[..., Any]] = None,
+    context_provider: Optional[Callable[[str], List[str]]] = None,
+) -> Callable[[str, OrchestrationContext], Dict[str, Any]]:
+    ...
+```
+
+Behavior by role:
+
+- **`researcher`** — calls the injected `context_provider(goal)` (workspace memory) to
+  pull relevant context into `ctx.research`; returns the count and the first items.
+- **`planner`** — decomposes the goal into ordered steps. If `inputs["steps"]` is a
+  non-empty list, each entry becomes a planned step; otherwise it produces a default
+  three-step plan (`Analyze` / `Execute` / `Verify the result`). Steps are written to
+  `ctx.plan`.
+- **`executor`** — iterates the plan and marks each step `done`. A step may request a
+  workflow or plugin run; when the corresponding runner is injected, the executor
+  drives it (see below). Results are written to `ctx.executed` and a summary line to
+  `ctx.output`.
+- **`reviewer`** — passes only if `ctx.executed` is non-empty and *every* executed step
+  has `status == "done"`; otherwise it returns `retry`. The verdict shape is:
+
+  ```json
+  {
+    "verdict": "pass",
+    "reason": "all steps completed",
+    "confidence": 0.9
+  }
+  ```
+
+  A failing review yields `{"verdict": "retry", "reason": "no steps executed", "confidence": 0.3}`.
+- **`release`** — sets/keeps `ctx.output` and returns `{"released": true, "summary": ...}`.
+
+An unrecognized role returns `{"role": role, "status": "noop"}`.
+
+### Agent → workflow and agent → plugin integration
+
+The executor role is where Lattice's cross-feature integration happens. When the
+`default_role_runner` is built with a `workflow_runner` and/or `plugin_runner`, a plan
+step can drive them:
+
+- A step's `workflow` (or `inputs["workflow"]`) is run via `workflow_runner(wf, ctx)`
+  on the first step (`index == 0`), capturing the result under `workflow_result` (or
+  `workflow_error` on failure).
+- A step's `plugin` is run via `plugin_runner(pl, ctx)`, capturing `plugin_result` (or
+  `plugin_error`).
+
+This is the **agent → workflow** and **agent → plugin** seam: an orchestrated agent run
+can actually execute Workflows and Plugins, in the community edition, with no model
+required.
+
+## Persistence and Knowledge Graph
+
+After a run completes, the API persists it via
+`WorkspaceOSStore.record_agent_run`, which both ingests a Knowledge Graph node and
+records a Workspace timeline event:
+
+```python
+def record_agent_run(
+    self,
+    *,
+    agent_id: str,
+    status: str,
+    input_text: str,
+    output_text: str,
+    user_email: Optional[str],
+    timeline: Optional[List[Dict[str, Any]]] = None,
+    relationships: Optional[List[str]] = None,
+    graph: Any = None,
+    workspace_id: Optional[str] = None,
+) -> Dict[str, Any]:
+    ...
+```
+
+What it does:
+
+- Builds a run record (`id`, `agent_id`, `status`, `input`, `output_preview` truncated
+  to 1000 chars, `user_email`, scoped `workspace_id`, `relationships`, `timeline`,
+  `created_at`).
+- When a `graph` is supplied, calls `graph.ingest_event("AgentRun", ...)` and stores the
+  returned `graph_node_id` on the run (capturing `graph_error` if ingest fails).
+- Appends the run to workspace state (retaining the most recent 300) and records an
+  `agent` / `agent_run` timeline event.
+
+Runs are workspace-scoped; `list_agents(workspace_id=...)` returns the registered
+`agents` plus the most recent runs for that scope.
+
+## HTTP API
+
+The router is created by `create_agents_router(...)` in `latticeai/api/agents.py`. All
+paths live under `/agents` (plural).
+
+> **Compatibility.** `/agents` does **not** collide with the existing single-agent
+> `/agent` endpoints — the trailing `s` keeps the v2 namespace fully separate, so v1
+> clients continue to work unchanged.
+
+### `GET /agents` — UI page
+
+Requires an authenticated user. Serves the `agents.html` Multi-Agent UI from the static
+directory; returns `404` if the UI is not available or not found.
+
+### `GET /agents/api/roles`
+
+Requires an authenticated user. Lists the built-in roles and the default pipeline.
+
+```json
+{
+  "roles": [
+    {"role": "researcher", "agent_id": "agent:researcher"},
+    {"role": "planner",    "agent_id": "agent:planner"},
+    {"role": "executor",   "agent_id": "agent:executor"},
+    {"role": "reviewer",   "agent_id": "agent:reviewer"},
+    {"role": "release",    "agent_id": "agent:release"}
+  ],
+  "default_pipeline": ["planner", "executor", "reviewer"]
+}
+```
+
+### `GET /agents/api/runs`
+
+Requires an authenticated user; reads are scoped via `gate_read`. Returns the registered
+agents plus recent runs for the resolved workspace scope (the result of
+`store.list_agents(workspace_id=scope)`).
+
+```json
+{
+  "agents": [ { "id": "agent:planner", "name": "Planner", "...": "..." } ],
+  "runs":   [ { "id": "agent-run-...", "agent_id": "agent:executor", "status": "ok", "...": "..." } ]
+}
+```
+
+### `POST /agents/api/run`
+
+Requires an authenticated user; writes are scoped via `gate_write`. Runs the
+orchestrator and persists the result.
+
+**Request body** (`AgentRunRequest`):
+
+```json
+{
+  "goal": "Summarize the open incidents and draft a status update",
+  "roles": ["researcher", "planner", "executor", "reviewer"],
+  "inputs": { "steps": ["Collect incidents", "Draft update"] },
+  "max_retries": 2
+}
+```
+
+| Field | Type | Default | Notes |
+| --- | --- | --- | --- |
+| `goal` | string | — | Required; a `400` is returned if blank. |
+| `roles` | string[] | `[]` | Empty means the default `CORE_PIPELINE`. Unknown roles are filtered out. |
+| `inputs` | object | `{}` | Passed through to the runner (e.g. `steps`, `workflow`). |
+| `max_retries` | int | `2` | Clamped server-side to the range `0`–`5`. |
+
+The endpoint resolves an orchestrator via the injected `orchestrator_factory(user, scope)`,
+runs it, records the run with `store.record_agent_run(...)` (passing the
+`workspace_graph()` for KG ingest and the run's `roles_run` as `relationships`), and
+appends a `multi_agent_run` audit event.
+
+**Response:**
+
+```json
+{
+  "run": {
+    "id": "agent-run-…",
+    "agent_id": "agent:executor",
+    "status": "ok",
+    "input": "Summarize the open incidents and draft a status update",
+    "output_preview": "Completed 2 planned step(s) for: …",
+    "relationships": ["agent:researcher", "agent:planner", "agent:executor", "agent:reviewer"],
+    "timeline": [ { "event": "start", "...": "..." } ],
+    "graph_node_id": "…",
+    "created_at": "…"
+  },
+  "result": {
+    "agent_id": "agent:executor",
+    "status": "ok",
+    "output": "Completed 2 planned step(s) for: …",
+    "timeline": [ "…" ],
+    "plan": [ { "index": 0, "description": "Collect incidents", "status": "done" } ],
+    "review": { "verdict": "pass", "reason": "all steps completed", "confidence": 0.9 },
+    "roles_run": ["researcher", "planner", "executor", "reviewer"],
+    "retries": 0
+  }
+}
+```
+
+## Timeline event reference
+
+The orchestration timeline is a flat, ordered list of event objects. Event types:
+
+| `event` | Emitted by | Key fields |
+| --- | --- | --- |
+| `start` | `run` (before the pipeline) | `goal`, `pipeline`, `timestamp` |
+| `role` | `_run_role` (per stage) | `role`, `agent_id`, `status`, `result`, `started_at`, `timestamp` |
+| `handoff` | `OrchestrationContext.handoff` | `from`, `to`, `note`, `timestamp` |
+| `end` | `run` (after the pipeline) | `status`, `retries`, `timestamp` |
+
+This timeline is returned on the run result and persisted with the run, so it drops
+straight into the Workspace timeline and Knowledge Graph.