Skip to main content
context-engineer is an archived, copyable concierge example; it is not installed by init. Copy it from examples/init-pack/ with its dependency closure, or ask create-workflow to author a repository-specific equivalent. You hand it a vague, plain-English script and it does the context engineering for you: it classifies the ask, inspects the repo/tools/skills to draft an explicit context contract, grills you on the blocking ambiguities one question at a time, routes the work to the right skills or durable workflow, builds a backpressure gate matrix, executes (or dispatches) the work, and writes a self-contained HTML report of the whole run.
By default it pauses for you to approve the context contract before any work is executed. Clear the gate with bunx smithers-orchestrator approve RUN_ID --node approve-contract, or run autonomously with --input '{"review":false}'.

Stages

  1. classify-script: read the script and decide which task modes it touches and whether it needs durable execution.
  2. inventory-context: inspect the repo, tools, and .smithers/skills to draft a context contract (goal, non-goals, assumptions, inputs, missing inputs, constraints, risks, desired artifacts, success criteria, verification signals).
  3. grill: resolve blocking ambiguity with GrillMe, one question at a time, each with a recommended answer.
  4. route: choose how to carry out the work: single task, skills, a durable seeded workflow, or hand to a human.
  5. build-backpressure: map every success criterion to a verification method and enforcement gate.
  6. approve-contract: durable human approval gate for the whole contract (skippable via review: false).
  7. execute: a Ralph loop that does or dispatches the work, honoring the gates, until it reports done.
  8. report: a concise HTML slideshow of the script, contract, decisions, route, gates, artifacts, and unresolved items.

Inputs

Monitor it

Use this when you have a fuzzy ask and want Smithers to do the context engineering (clarify the real goal, ground it in what is actually available, add backpressure, and only then execute) instead of firing a half-specified task at a bare agent.