Use Case
Researchers using a sufficiently provisioned model route should be able to run independent data-quality, statistical, reproducibility, and evidence reviews concurrently within one research workflow.
Constrained local routes should continue to work predictably without unnecessary parallel model load.
Problem
Heartwood currently runs one specialist at a time.
That is appropriate for local inference, but it makes independent research and review tasks unnecessarily slow on routes that can support concurrent work.
Parallel agents also introduce cost, cancellation, recovery, and shared-state risks that cannot be left to prompt conventions.
Solution
- Reuse OpenHands task delegation and parallel tool execution for independent, read-only specialist work.
- Keep the parent conversation and gateway as the only writers to project and session state.
- Add route capabilities for maximum concurrency and token, reported-cost, time, and model-step budgets; default to sequential execution when the route is unknown or unqualified.
- Request confirmation before starting a parallel run and show the expected worker count and budget.
- Give each specialist an isolated context and produce one parent synthesis after all successful or terminal results are known.
- Project task lineage, progress, cancellation, partial failure, and final results through the shared interface contract.
Acceptance Criteria
- Independent specialists overlap in execution on a qualified concurrent route and remain sequential on a constrained route.
- Parallel work cannot mutate project files or resolve another specialist's actions.
- Timeouts, cancellation, partial failure, restart, and retry produce one authoritative result without duplicate model or tool work.
- Terminal, browser, and notebook clients render the same task tree and budget state.
- Capable-model evaluation measures result quality, latency, and usage against the sequential baseline.
- Parallel execution remains disabled when the benchmark or route capability evidence is insufficient.
Alternatives considered
A Heartwood-specific worker or message-bus implementation would duplicate OpenHands orchestration and persistence.
Parallel project writes were rejected until OpenHands exposes a gateway-visible, restart-safe child-action contract.
Additional context
This work depends on the OpenHands upgrade in #123 and the baseline benchmark in #44.
It extends the sequential specialists delivered in #99.
The workflow definitions in #130 remain unchanged whether their advisory stages run sequentially or concurrently.
Code of Conduct
Use Case
Researchers using a sufficiently provisioned model route should be able to run independent data-quality, statistical, reproducibility, and evidence reviews concurrently within one research workflow.
Constrained local routes should continue to work predictably without unnecessary parallel model load.
Problem
Heartwood currently runs one specialist at a time.
That is appropriate for local inference, but it makes independent research and review tasks unnecessarily slow on routes that can support concurrent work.
Parallel agents also introduce cost, cancellation, recovery, and shared-state risks that cannot be left to prompt conventions.
Solution
Acceptance Criteria
Alternatives considered
A Heartwood-specific worker or message-bus implementation would duplicate OpenHands orchestration and persistence.
Parallel project writes were rejected until OpenHands exposes a gateway-visible, restart-safe child-action contract.
Additional context
This work depends on the OpenHands upgrade in #123 and the baseline benchmark in #44.
It extends the sequential specialists delivered in #99.
The workflow definitions in #130 remain unchanged whether their advisory stages run sequentially or concurrently.
Code of Conduct