How should independent coding tasks be scheduled and brought back together?
Current stateHistorical Claude Code orchestration plugin with a retained dependency-aware implementation and tests
Approach
This work made the work order explicit before delegating it: each task names its scope, branch, test directory, expected tests, and dependencies. The coordinator prepares worker requests, tracks their outcomes, and brings completed branches back together.
Parallel implementation depends on the relationships between components: tasks wait for their prerequisites, and a failed prerequisite blocks its dependents. Once the branches are combined, the coordinator offers a full test run of the integrated result.
Why this approach
Parallel work needs an integration plan
The manifest is a concrete decomposition of the job, including what can happen independently and what must happen in order. That makes the intended dependency graph available for validation before workers start.
The retained implementation also makes the host boundary explicit: the Python coordinator returns prompts and spawn requests, while the Claude Code workflow starts the workers and reports their outcomes. The coordinator is not a separate agent runtime.
Where it stands
The coordinator validates dependency references and cycles, releases ready tasks, records worker identities and retry counts, and persists task state through atomic JSON replacement. Retry prompts include the prior failure so a retry has a concrete problem to address.
Completed task branches are merged in dependency order. The coordinator records failed merges, aborts the conflicted merge, and offers a full-suite pytest verification step. Gateway checks inspect branches, tests, and commits; worker prompts describe a test-first workflow.
This is the earlier orchestration plugin. Agent-swarm and Harness later developed their own coordination and execution mechanisms.
What the work has shown
Dependency failures block the affected work
The scheduler returns a pending task only when all its prerequisites have completed. When a task exhausts its retries, failure propagates through its dependents while independent tasks remain eligible. The test suite contains cases for dependency release, terminal failure propagation, and an independent branch surviving a partial failure.
The behavior is visible in a small state machine: pending, spawned, completed, or failed, with the worker identity and last error retained.
The integration order comes from the same task graph
The coordinator uses a topological order when merging completed branches and records each merge result. A separate verification method runs the complete pytest suite on the combined tree.
The repository preserves implementation and regression tests for these mechanics. The accompanying account describes how I used parallel assistance, but it does not provide a controlled comparison of total time and result quality against a sequential workflow.