UniConvert
Local-first execution engine for AI-assisted pipelines
UniConvert is an execution engine for AI-assisted work: plans are data, steps declare inputs/outputs, and each transition passes through validators (schema, policy, capability). It integrates with LLM agents where helpful, but never conflates planning with execution authority.
Design goals:
- Local-first runs for sensitive transforms; optional remote executors for heavy jobs.
- Replayable runs: hashed artifacts + step logs for debugging and compliance.
- Composable pipelines as graphs with explicit failure domains and compensating actions.
Agent demos chain tools cheerfully; production needs idempotent steps, typed IO, and blast-radius control. Silent partial failures and non-deterministic side effects are unacceptable when AI proposes plans.
The challenge is to make agentic workflows boringly reliable: validate first, execute once, record everything.
-
Plan IR: Represent tasks as a DAG with explicit contracts (pre/post conditions).
-
Validators: Static checks (schema, dependency order) + dynamic checks (secrets, network policy).
-
Executors: Sandboxed local processes or signed remote workers; uniform telemetry.
-
Compensation: Declared rollback hooks for reversible steps; explicit manual escalation when not reversible.
- 01
Third-party APIs fail mid-DAG—handled with retries, idempotency keys, and partial replay.
- 02
Validator false positives block legitimate plans—versioned allowlists and override workflows.
- 03
Large artifacts blow storage budgets—compression tiers and external blob offload.
- 01
More ceremony than a script; necessary for production-grade agents.
- 02
Local-first prioritizes privacy/latency over elastic scale—remote executors optional.
- 03
Strict validation can slow iteration—mitigated with dev-mode profiles.
run_plan.py
Execute only after all validators pass
1 Designed
Defined plan IR, validation stages, and executor contracts for local vs remote work.
Implemented
Implemented the runner core, artifact hashing, and policy hooks.
Scrapped
Implicit tool calls directly from model output—replaced with explicit compile + approve flow.