Everything below is on the dev branch unless marked Partly built or
Vision.
| Feature | Detail |
|---|
| YAML workflow definitions (schema 2.0) | Validated and compiled by azhi validate: graph checks, ref type-checking against output schemas, tool argument and projection checks, CEL checks |
| Eleven node types | tool, script, agent, retrieve, condition, parallel, loop, subworkflow, approval, report, notify |
| Durable runs on Temporal | Retries by error class, cancellation, worker_offline waits |
| Schedules | Cron in IANA timezones; a published version’s trigger.schedule becomes its schedule |
| Parallel runs | Several runs of the same workflow, each harness step in its own fresh checkout; at most 2 runs per repository by default |
| Loops and subworkflows | A workflow can call another; nesting stops at four levels |
| Entry points | CLI, REST API, schedules, Mission Control. Web chat and webhooks are Vision |
| Feature | Detail |
|---|
| Run plan | Capability marks, policy coverage, taint paths, missing grants, blockers |
| Taint rule | Ungated write after untrusted data fails to compile |
| Action ledger | Exactly-once writes with reconciliation |
| Context manifest | What the model saw, every turn |
| Approvals | Approve or reject in the CLI or Mission Control |
| Signed packages | Per-publisher keys, worker trust policies |
| Feature | Detail |
|---|
| Executors | Model agent, OpenCode, Claude Agent SDK (Codex experimental) |
| Providers | Anthropic, OpenAI, GitHub Copilot, ChatGPT plan. See models |
| Validated output | Output schema with two repairs |
| Harness builder | Per-step agent prompt, command, skills, read-only tools and MCP servers for OpenCode reviewers |
| Isolated workspaces | Fresh git checkout per step, isolated HOME, deleted after |
| Result confidence | A verifier on a different model family tries to refute findings |
| Feature | Detail |
|---|
| Usage and cost | Per turn, with null for unknown |
| Budgets and spend limits | Per step, per workflow, per workspace, with alerts |
| Copilot allowance | azhi copilot quota and the run page show AI credits |
| Feature | Detail |
|---|
| Datasets | Hybrid retrieval with numbered citations to immutable excerpts |
| MCP and the Library | Stdio and Streamable HTTP MCP, reusable agents, skills and commands, live marketplaces |
| Tool gateway | Projection, observation records, ledger |
| Feature | Detail |
|---|
| Mission Control | Web app: runs, approvals, alerts, spend, schedules, workers, secrets |
| Build with chat | Describe a workflow; get a compiled draft |
azhi doctor | Database, Temporal, workers online, interpreter builds |
| Audit log | Who changed secrets, published, approved, changed schedules or trust policies |
