deeda Think — developer docs

Every recipe below is a complete workflow.md you can copy, validate, and dispatch. Extend any of them with blocks from Snippets; check field meanings in the Workflow Schema.

Choosing a method

The method field selects how a runtime is driven. Rules of thumb:

Method	Choose when	Trade-off
`api`	You want stateless turns, structured output, tight latency control, or server-side features (e.g. OpenAI Responses/Realtime/Operator).	You manage context/state in the workflow; no vendor-native agent loop.
`sdk`	You want the vendor’s full agentic loop — native tool calling, hooks, sessions, thinking budgets (e.g. Claude Agent SDK, OpenAI Agents SDK, GenAI SDK).	Heavier per-turn; vendor loop semantics apply.
`cli`	The capability lives in a local agent binary (Claude Code, Codex CLI, Antigravity) — local file edits, sandboxed shell, repo-scale work.	Requires local install; dispatch_path is `local`.
`mcp`	You are bridging an external MCP server’s tools into the run.	Tool-surface only.
`desktop_app`	Driving a desktop surface (voice shell, UI automation).	Desktop session required.

Concretely: research/synthesis → gemini-genai-sdk (sdk); repo implementation → anthropic-agent-sdk (cli or sdk); independent code review → openai-codex-sdk (cli); cheap structured calls → openai-responses-api (api); voice → openai-realtime-api (api) or local-voice; computer use → openai-operator (api); offline/private → local-llm (api).

Recipe: minimal single agent

The smallest valid production workflow — one runtime, fixed route:

---
id: minimal-single-agent
version: 1.0.0
agent:
  runtime_key: anthropic-agent-sdk
  provider: anthropic
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required: [fs.read, fs.write, git]
---

# Minimal Single Agent

Plan, implement, verify; record evidence before finishing.

Recipe: ordered fallback (resilient single agent)

Try Anthropic first; fall back to Codex, then Antigravity, on failure or credit exhaustion:

---
id: resilient-worker
version: 1.0.0
agent:
  runtime_key: anthropic-agent-sdk
  provider: anthropic
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 10
  routing_strategy: single_agent_routine
  single_agent:
    select: ordered
    candidates:
      - runtime_key: anthropic-agent-sdk
        method: sdk
        dispatch_path: cloud_managed
      - runtime_key: openai-codex-sdk
        method: cli
        dispatch_path: local
      - runtime_key: gemini-antigravity-cli
        method: cli
        dispatch_path: local
    fallback_on_credit_exhausted: true
budget:
  tokens: 100000
  wall_clock_minutes: 45
sandbox:
  workspace_write: true
tools:
  required: [fs.read, fs.write, git]
retry:
  max_attempts: 2
  retryable_runtime_failures: [provider_unavailable, rate_limited]
---

# Resilient Worker

Same job, three routes. Cadence advances to the next candidate on failure.

Recipe: pair programming (researcher + engineer + reviewer)

Three providers, each where it is strongest, with an isolated reviewer:

---
id: pair-research-build-review
version: 1.0.0
agent:
  runtime_key: anthropic-agent-sdk
  provider: anthropic
  role: orchestrator
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 12
pair_profiles:
  default:
    name: "Research, build, review"
    researcher:
      runtime_key: gemini-genai-sdk
      provider: gemini
      model: gemini-2.5-pro
      method: sdk
    engineer:
      runtime_key: anthropic-agent-sdk
      provider: anthropic
      model: claude-opus-4-7
      method: cli
    reviewer:
      runtime_key: openai-codex-sdk
      provider: openai
      model: gpt-5.5
      method: cli
      isolate_context: true
budget:
  tokens: 180000
  wall_clock_minutes: 90
sandbox:
  workspace_write: true
  network: [api.anthropic.com, api.openai.com, generativelanguage.googleapis.com]
tools:
  required: [fs.read, fs.write, git, probes.run]
---

# Pair: Research, Build, Review

Researcher gathers context, engineer implements, isolated reviewer judges.

Recipe: provider-diverse review panel

Gate completion on a 3-reviewer, cross-vendor consensus:

review_consensus:
  workflow_id: review-consensus
  reviewers: 3
  rule: unanimous-on-block
  diversity: [anthropic, gemini, openai]
  runtime_keys:
    anthropic: anthropic-agent-sdk
    gemini: gemini-genai-sdk
    openai: openai-codex-sdk
  timeout_ms: 5400000

Add this block to any recipe above; Cadence will not close the run until the panel agrees (any reviewer can block; a block forces rework).

Recipe: state-machine workflow

Multi-stage run where Cadence advances stages only on typed events:

state_machine:
  start: research
  stage_order: [research, plan, implement, review, signoff]
  terminal_states: [done]
  on_event:
    research_completed: plan
    plan_completed: implement
    implementation_completed: review
    review_clean: signoff
    review_blocked: implement
    signoff_completed: done
  loop_caps:
    total_turns: 12
    on_exceeded: signoff

The full worked example combining all of the above is in Workflow Markdown.

Getting Started

Concepts

Agent Harness

Provider Capabilities

Workflow Cookbook

Choosing a method

Recipe: minimal single agent

Recipe: ordered fallback (resilient single agent)

Recipe: pair programming (researcher + engineer + reviewer)

Recipe: provider-diverse review panel

Recipe: state-machine workflow

​Choosing a method

​Recipe: minimal single agent

​Recipe: ordered fallback (resilient single agent)

​Recipe: pair programming (researcher + engineer + reviewer)

​Recipe: provider-diverse review panel

​Recipe: state-machine workflow

Choosing a method

Recipe: minimal single agent

Recipe: ordered fallback (resilient single agent)

Recipe: pair programming (researcher + engineer + reviewer)

Recipe: provider-diverse review panel

Recipe: state-machine workflow