Durable Task
Treat chat context as working memory, not the source of truth for a long-running task.
Workflow
- Inspect only the repository context needed to understand the requested outcome.
- Freeze a compact task contract with:
- stable
task_id - objective
- in-scope work
- explicit non-goals
- constraints and authorization boundaries
- testable acceptance criteria
- stop conditions
- stable
- Prefer the bundled
hermes-durable initcommand when available. Otherwise persist the same fields under.agent/tasks/<task_id>/task.json. - Do not silently rewrite acceptance criteria to match an implementation that already exists. If the contract must change, record the change before continuing.
- Execute the smallest next step that advances one or more acceptance criteria.
- After a meaningful unit of work, persist a checkpoint with the completed steps and one concrete next action.
- Record observable evidence separately from the agent's own summary.
Boundaries
- Do not create durable state for tiny, single-step edits that can be safely completed and verified in the current session.
- Do not use the task contract to expand the user's authorization.
- Do not mark a task complete merely because an implementation command exited successfully.
- Keep provider-specific session IDs optional; the durable task must remain understandable without one.
Completion
A task is ready for completion review only when its current acceptance criteria can be mapped to fresh evidence. Use the evidence-gate skill for that review.