[Team] Execute declarative read-only task DAGs

[ranveersequeira]

Parent

#377

What to build

Add the first workflow-first agentify team vertical slice: execute a declarative dependency graph of read-only research, planning, and review tasks through the durable broker from #378.

Workers exchange only declared artifacts. The workflow must remain resumable and auditable without depending on opaque provider continuation state.

User story

As a lead developer or agent, I can describe a small dependency-aware team workflow, run ready tasks in parallel, inspect status, answer a blocking question, retry a failed node, and receive a standalone HTML/JSON report.

Proposed CLI

agentify team plan "review the payment retry architecture" --out team.yaml
agentify team run team.yaml
agentify team status <run-id> --watch
agentify team answer <run-id>/<node-id> "Preserve the legacy error schema"
agentify team retry <run-id>/<node-id>
agentify team status <run-id> --open

team plan may initially emit a template or validate a caller-authored graph. Starting a paid planner agent is optional and must require explicit routing/budget inputs.

Minimal graph contract

Each graph records:

• schema version, goal, immutable base SHA, and dirty-tree policy;
• nodes with ID, task, dependencies, read-only mode, route/provider/profile request;
• declared input and output artifacts;
• optional verification rules and bounded retry policy;
• maximum parallelism and failure policy.

Each node receives only:

• its task and attempt number;
• declared upstream artifacts;
• repository guidance and relevant bounded context;
• immutable base SHA;
• required output schema and limits.

Required behavior

• Validate schema, node IDs, dependency references, and acyclicity before execution.
• Resolve the base ref to an immutable SHA.
• Schedule dependency-ready nodes up to maxParallel.
• Use broker task IDs and events rather than launching providers directly.
• Persist graph, node attempts, messages/questions, artifacts, usage, verification, and terminal status.
• Make downstream artifacts available only after the producing node succeeds and passes configured validation.
• Represent a blocking worker question as a structured artifact and waiting_input state.
• Record retries as new immutable attempts; never overwrite prior evidence.
• Produce a self-contained HTML report plus versioned JSON receipt.
• Resume a run after the CLI exits without replaying completed nodes.
• Reject write-mode nodes in this slice.

Suggested implementation seams

• Add a deep src/core/team/ scheduler/artifact/report module above the broker.
• Reuse route/profile policy from src/core/models.js.
• Reuse repository identity and local runtime paths from src/core/project-store.js.
• Use affected-test/risk operations only as optional read-only verification inputs; do not mutate repository state.
• Keep src/core/workflows.js focused on installing platform skills.

Acceptance criteria

• A three-node graph (research → plan → review) executes with correct dependency ordering.
• Independent ready nodes run in parallel up to the declared cap.
• Invalid, cyclic, missing-dependency, and duplicate-node graphs fail before provider work starts.
• Every node uses a durable broker task and records resolved provider/model/capability/policy.
• Structured artifacts pass between nodes; undeclared files or raw hidden transcripts do not.
• A blocking question pauses only the affected dependency chain and can be answered as a new immutable input.
• Retry creates a new attempt while preserving prior output, usage, and failure evidence.
• A client can reconnect and resume status without rerunning completed nodes.
• The HTML report shows the DAG, node state, provider, duration, usage provenance, artifacts, questions, retries, and errors.
• --json returns a stable versioned run receipt.
• Write-mode nodes and automatic integration are rejected.
• Deterministic scheduler tests cover concurrency, dependency failure, retry, resume, cancellation, and corrupt artifact metadata.

Out of scope

• Write-capable nodes.
• Worktree creation or integration branches.
• Dynamic graph mutation by workers.
• Automatic conflict resolution.
• Remote workers.
• Free-form shared chat rooms.
• Making paid model planning the default.

Blocked by

• Blocked by [Relay] Add a durable read-only broker through the CLI #378