Skip to main content

Task Lifecycle: From Intent to Verified Code

OpenExec doesn't just "write code." It treats AI agents as managed workers in a structured pipeline. Every task you run automatically progresses through a series of stages defined by a Blueprint.

This methodical approach ensures that every change is researched, executed, and independently verified before being finalized.


The Modern Model: Blueprint Engine

OpenExec uses a Blueprint Engine to define flexible execution graphs. The default blueprint for implementation (openexec run) uses a sequence of Deterministic (code-based) and Agentic (AI-based) stages.

Standard Execution Stages

StageTypeDescription
gather_contextDeterministicScans the repository, extracts relevant symbols, and assembles a context pack.
implementAgenticThe primary AI model (Executor) generates code changes and applies them as patches.
lintDeterministicRuns the project's native linting tools (e.g., go vet, eslint, ruff).
fix_lintAgenticIf linting fails, the AI analyzes the errors and attempts to fix them automatically.
testDeterministicExecutes the relevant test suite to ensure no regressions were introduced.
fix_testsAgenticIf tests fail, the AI iterates on the implementation until the tests pass.
reviewAgenticA final review stage where the AI summarizes the work and verifies it against the original task criteria.

Legacy Model: 5-Phase Pipeline (Deprecated)

Earlier versions of OpenExec used a fixed 5-phase pipeline. While still supported for backward compatibility, new projects should use the Blueprint model.

  1. TD (Technical Design): Research and strategy formulation.
  2. IM (Implementation): Actual code modification.
  3. RV (Review): Independent quality audit.
  4. RF (Refinement): Post-review adjustments.
  5. FL (Finalize): Goal validation and state sync.

Fast-Track: Analytical Tasks

For tasks that are purely analytical (e.g., "Study the codebase", "Map orchestration boundaries"), OpenExec automatically activates a Fast-Track pipeline.

These tasks skip the implementation and testing stages entirely, moving directly from context gathering to the final summary.


Observation in the Dashboard

When a run is active, you can monitor these transitions in real-time via the Web Console:

  • Active Stage: Highlighted in the execution loop.
  • Logs: Real-time tail of agent reasoning and tool output.
  • Audit Trail: Every transition and tool call is recorded in the immutable audit vault.